Linux Audio

Check our new training course

Loading...
v4.6
 
   1/*
   2 * Copyright (C) 2013, 2014 ARM Limited, All Rights Reserved.
   3 * Author: Marc Zyngier <marc.zyngier@arm.com>
   4 *
   5 * This program is free software; you can redistribute it and/or modify
   6 * it under the terms of the GNU General Public License version 2 as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it will be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
  16 */
  17
 
 
 
  18#include <linux/bitmap.h>
  19#include <linux/cpu.h>
 
  20#include <linux/delay.h>
 
  21#include <linux/interrupt.h>
 
 
 
 
  22#include <linux/log2.h>
 
  23#include <linux/mm.h>
  24#include <linux/msi.h>
  25#include <linux/of.h>
  26#include <linux/of_address.h>
  27#include <linux/of_irq.h>
  28#include <linux/of_pci.h>
  29#include <linux/of_platform.h>
  30#include <linux/percpu.h>
  31#include <linux/slab.h>
 
  32
  33#include <linux/irqchip.h>
  34#include <linux/irqchip/arm-gic-v3.h>
 
  35
  36#include <asm/cacheflush.h>
  37#include <asm/cputype.h>
  38#include <asm/exception.h>
  39
  40#include "irq-gic-common.h"
  41
  42#define ITS_FLAGS_CMDQ_NEEDS_FLUSHING		(1ULL << 0)
  43#define ITS_FLAGS_WORKAROUND_CAVIUM_22375	(1ULL << 1)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  44
  45#define RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING	(1 << 0)
  46
  47/*
  48 * Collection structure - just an ID, and a redistributor address to
  49 * ping. We use one per CPU as a bag of interrupts assigned to this
  50 * CPU.
  51 */
  52struct its_collection {
  53	u64			target_address;
  54	u16			col_id;
  55};
  56
  57/*
 
 
 
 
 
 
 
 
 
 
 
 
 
  58 * The ITS structure - contains most of the infrastructure, with the
  59 * top-level MSI domain, the command queue, the collections, and the
  60 * list of devices writing to it.
 
 
 
 
  61 */
  62struct its_node {
  63	raw_spinlock_t		lock;
 
  64	struct list_head	entry;
  65	void __iomem		*base;
  66	unsigned long		phys_base;
 
  67	struct its_cmd_block	*cmd_base;
  68	struct its_cmd_block	*cmd_write;
  69	struct {
  70		void		*base;
  71		u32		order;
  72	} tables[GITS_BASER_NR_REGS];
  73	struct its_collection	*collections;
 
 
 
 
 
 
  74	struct list_head	its_device_list;
  75	u64			flags;
  76	u32			ite_size;
 
 
 
 
  77};
  78
 
 
 
 
  79#define ITS_ITT_ALIGN		SZ_256
  80
 
 
 
 
 
 
 
 
 
 
 
 
 
  81/* Convert page order to size in bytes */
  82#define PAGE_ORDER_TO_SIZE(o)	(PAGE_SIZE << (o))
  83
  84struct event_lpi_map {
  85	unsigned long		*lpi_map;
  86	u16			*col_map;
  87	irq_hw_number_t		lpi_base;
  88	int			nr_lpis;
 
 
 
 
  89};
  90
  91/*
  92 * The ITS view of a device - belongs to an ITS, a collection, owns an
  93 * interrupt translation table, and a list of interrupts.
 
 
  94 */
  95struct its_device {
  96	struct list_head	entry;
  97	struct its_node		*its;
  98	struct event_lpi_map	event_map;
  99	void			*itt;
 100	u32			nr_ites;
 101	u32			device_id;
 
 
 
 
 
 
 
 
 
 
 
 
 
 102};
 103
 
 
 104static LIST_HEAD(its_nodes);
 105static DEFINE_SPINLOCK(its_lock);
 106static struct rdists *gic_rdists;
 
 
 
 
 
 
 
 107
 108#define gic_data_rdist()		(raw_cpu_ptr(gic_rdists->rdist))
 
 109#define gic_data_rdist_rd_base()	(gic_data_rdist()->rd_base)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 110
 111static struct its_collection *dev_event_to_col(struct its_device *its_dev,
 112					       u32 event)
 113{
 114	struct its_node *its = its_dev->its;
 115
 116	return its->collections + its_dev->event_map.col_map[event];
 117}
 118
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 119/*
 120 * ITS command descriptors - parameters to be encoded in a command
 121 * block.
 122 */
 123struct its_cmd_desc {
 124	union {
 125		struct {
 126			struct its_device *dev;
 127			u32 event_id;
 128		} its_inv_cmd;
 129
 130		struct {
 131			struct its_device *dev;
 132			u32 event_id;
 
 
 
 
 
 133		} its_int_cmd;
 134
 135		struct {
 136			struct its_device *dev;
 137			int valid;
 138		} its_mapd_cmd;
 139
 140		struct {
 141			struct its_collection *col;
 142			int valid;
 143		} its_mapc_cmd;
 144
 145		struct {
 146			struct its_device *dev;
 147			u32 phys_id;
 148			u32 event_id;
 149		} its_mapvi_cmd;
 150
 151		struct {
 152			struct its_device *dev;
 153			struct its_collection *col;
 154			u32 event_id;
 155		} its_movi_cmd;
 156
 157		struct {
 158			struct its_device *dev;
 159			u32 event_id;
 160		} its_discard_cmd;
 161
 162		struct {
 163			struct its_collection *col;
 164		} its_invall_cmd;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 165	};
 166};
 167
 168/*
 169 * The ITS command block, which is what the ITS actually parses.
 170 */
 171struct its_cmd_block {
 172	u64	raw_cmd[4];
 
 
 
 173};
 174
 175#define ITS_CMD_QUEUE_SZ		SZ_64K
 176#define ITS_CMD_QUEUE_NR_ENTRIES	(ITS_CMD_QUEUE_SZ / sizeof(struct its_cmd_block))
 177
 178typedef struct its_collection *(*its_cmd_builder_t)(struct its_cmd_block *,
 
 179						    struct its_cmd_desc *);
 180
 
 
 
 
 
 
 
 
 
 
 
 181static void its_encode_cmd(struct its_cmd_block *cmd, u8 cmd_nr)
 182{
 183	cmd->raw_cmd[0] &= ~0xffUL;
 184	cmd->raw_cmd[0] |= cmd_nr;
 185}
 186
 187static void its_encode_devid(struct its_cmd_block *cmd, u32 devid)
 188{
 189	cmd->raw_cmd[0] &= BIT_ULL(32) - 1;
 190	cmd->raw_cmd[0] |= ((u64)devid) << 32;
 191}
 192
 193static void its_encode_event_id(struct its_cmd_block *cmd, u32 id)
 194{
 195	cmd->raw_cmd[1] &= ~0xffffffffUL;
 196	cmd->raw_cmd[1] |= id;
 197}
 198
 199static void its_encode_phys_id(struct its_cmd_block *cmd, u32 phys_id)
 200{
 201	cmd->raw_cmd[1] &= 0xffffffffUL;
 202	cmd->raw_cmd[1] |= ((u64)phys_id) << 32;
 203}
 204
 205static void its_encode_size(struct its_cmd_block *cmd, u8 size)
 206{
 207	cmd->raw_cmd[1] &= ~0x1fUL;
 208	cmd->raw_cmd[1] |= size & 0x1f;
 209}
 210
 211static void its_encode_itt(struct its_cmd_block *cmd, u64 itt_addr)
 212{
 213	cmd->raw_cmd[2] &= ~0xffffffffffffUL;
 214	cmd->raw_cmd[2] |= itt_addr & 0xffffffffff00UL;
 215}
 216
 217static void its_encode_valid(struct its_cmd_block *cmd, int valid)
 218{
 219	cmd->raw_cmd[2] &= ~(1UL << 63);
 220	cmd->raw_cmd[2] |= ((u64)!!valid) << 63;
 221}
 222
 223static void its_encode_target(struct its_cmd_block *cmd, u64 target_addr)
 224{
 225	cmd->raw_cmd[2] &= ~(0xffffffffUL << 16);
 226	cmd->raw_cmd[2] |= (target_addr & (0xffffffffUL << 16));
 227}
 228
 229static void its_encode_collection(struct its_cmd_block *cmd, u16 col)
 230{
 231	cmd->raw_cmd[2] &= ~0xffffUL;
 232	cmd->raw_cmd[2] |= col;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 233}
 234
 235static inline void its_fixup_cmd(struct its_cmd_block *cmd)
 236{
 237	/* Let's fixup BE commands */
 238	cmd->raw_cmd[0] = cpu_to_le64(cmd->raw_cmd[0]);
 239	cmd->raw_cmd[1] = cpu_to_le64(cmd->raw_cmd[1]);
 240	cmd->raw_cmd[2] = cpu_to_le64(cmd->raw_cmd[2]);
 241	cmd->raw_cmd[3] = cpu_to_le64(cmd->raw_cmd[3]);
 242}
 243
 244static struct its_collection *its_build_mapd_cmd(struct its_cmd_block *cmd,
 
 245						 struct its_cmd_desc *desc)
 246{
 247	unsigned long itt_addr;
 248	u8 size = ilog2(desc->its_mapd_cmd.dev->nr_ites);
 249
 250	itt_addr = virt_to_phys(desc->its_mapd_cmd.dev->itt);
 251	itt_addr = ALIGN(itt_addr, ITS_ITT_ALIGN);
 252
 253	its_encode_cmd(cmd, GITS_CMD_MAPD);
 254	its_encode_devid(cmd, desc->its_mapd_cmd.dev->device_id);
 255	its_encode_size(cmd, size - 1);
 256	its_encode_itt(cmd, itt_addr);
 257	its_encode_valid(cmd, desc->its_mapd_cmd.valid);
 258
 259	its_fixup_cmd(cmd);
 260
 261	return NULL;
 262}
 263
 264static struct its_collection *its_build_mapc_cmd(struct its_cmd_block *cmd,
 
 265						 struct its_cmd_desc *desc)
 266{
 267	its_encode_cmd(cmd, GITS_CMD_MAPC);
 268	its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
 269	its_encode_target(cmd, desc->its_mapc_cmd.col->target_address);
 270	its_encode_valid(cmd, desc->its_mapc_cmd.valid);
 271
 272	its_fixup_cmd(cmd);
 273
 274	return desc->its_mapc_cmd.col;
 275}
 276
 277static struct its_collection *its_build_mapvi_cmd(struct its_cmd_block *cmd,
 
 278						  struct its_cmd_desc *desc)
 279{
 280	struct its_collection *col;
 281
 282	col = dev_event_to_col(desc->its_mapvi_cmd.dev,
 283			       desc->its_mapvi_cmd.event_id);
 284
 285	its_encode_cmd(cmd, GITS_CMD_MAPVI);
 286	its_encode_devid(cmd, desc->its_mapvi_cmd.dev->device_id);
 287	its_encode_event_id(cmd, desc->its_mapvi_cmd.event_id);
 288	its_encode_phys_id(cmd, desc->its_mapvi_cmd.phys_id);
 289	its_encode_collection(cmd, col->col_id);
 290
 291	its_fixup_cmd(cmd);
 292
 293	return col;
 294}
 295
 296static struct its_collection *its_build_movi_cmd(struct its_cmd_block *cmd,
 
 297						 struct its_cmd_desc *desc)
 298{
 299	struct its_collection *col;
 300
 301	col = dev_event_to_col(desc->its_movi_cmd.dev,
 302			       desc->its_movi_cmd.event_id);
 303
 304	its_encode_cmd(cmd, GITS_CMD_MOVI);
 305	its_encode_devid(cmd, desc->its_movi_cmd.dev->device_id);
 306	its_encode_event_id(cmd, desc->its_movi_cmd.event_id);
 307	its_encode_collection(cmd, desc->its_movi_cmd.col->col_id);
 308
 309	its_fixup_cmd(cmd);
 310
 311	return col;
 312}
 313
 314static struct its_collection *its_build_discard_cmd(struct its_cmd_block *cmd,
 
 315						    struct its_cmd_desc *desc)
 316{
 317	struct its_collection *col;
 318
 319	col = dev_event_to_col(desc->its_discard_cmd.dev,
 320			       desc->its_discard_cmd.event_id);
 321
 322	its_encode_cmd(cmd, GITS_CMD_DISCARD);
 323	its_encode_devid(cmd, desc->its_discard_cmd.dev->device_id);
 324	its_encode_event_id(cmd, desc->its_discard_cmd.event_id);
 325
 326	its_fixup_cmd(cmd);
 327
 328	return col;
 329}
 330
 331static struct its_collection *its_build_inv_cmd(struct its_cmd_block *cmd,
 
 332						struct its_cmd_desc *desc)
 333{
 334	struct its_collection *col;
 335
 336	col = dev_event_to_col(desc->its_inv_cmd.dev,
 337			       desc->its_inv_cmd.event_id);
 338
 339	its_encode_cmd(cmd, GITS_CMD_INV);
 340	its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
 341	its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
 342
 343	its_fixup_cmd(cmd);
 344
 345	return col;
 346}
 347
 348static struct its_collection *its_build_invall_cmd(struct its_cmd_block *cmd,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 349						   struct its_cmd_desc *desc)
 350{
 351	its_encode_cmd(cmd, GITS_CMD_INVALL);
 352	its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
 353
 354	its_fixup_cmd(cmd);
 355
 356	return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 357}
 358
 359static u64 its_cmd_ptr_to_offset(struct its_node *its,
 360				 struct its_cmd_block *ptr)
 361{
 362	return (ptr - its->cmd_base) * sizeof(*ptr);
 363}
 364
 365static int its_queue_full(struct its_node *its)
 366{
 367	int widx;
 368	int ridx;
 369
 370	widx = its->cmd_write - its->cmd_base;
 371	ridx = readl_relaxed(its->base + GITS_CREADR) / sizeof(struct its_cmd_block);
 372
 373	/* This is incredibly unlikely to happen, unless the ITS locks up. */
 374	if (((widx + 1) % ITS_CMD_QUEUE_NR_ENTRIES) == ridx)
 375		return 1;
 376
 377	return 0;
 378}
 379
 380static struct its_cmd_block *its_allocate_entry(struct its_node *its)
 381{
 382	struct its_cmd_block *cmd;
 383	u32 count = 1000000;	/* 1s! */
 384
 385	while (its_queue_full(its)) {
 386		count--;
 387		if (!count) {
 388			pr_err_ratelimited("ITS queue not draining\n");
 389			return NULL;
 390		}
 391		cpu_relax();
 392		udelay(1);
 393	}
 394
 395	cmd = its->cmd_write++;
 396
 397	/* Handle queue wrapping */
 398	if (its->cmd_write == (its->cmd_base + ITS_CMD_QUEUE_NR_ENTRIES))
 399		its->cmd_write = its->cmd_base;
 400
 
 
 
 
 
 
 401	return cmd;
 402}
 403
 404static struct its_cmd_block *its_post_commands(struct its_node *its)
 405{
 406	u64 wr = its_cmd_ptr_to_offset(its, its->cmd_write);
 407
 408	writel_relaxed(wr, its->base + GITS_CWRITER);
 409
 410	return its->cmd_write;
 411}
 412
 413static void its_flush_cmd(struct its_node *its, struct its_cmd_block *cmd)
 414{
 415	/*
 416	 * Make sure the commands written to memory are observable by
 417	 * the ITS.
 418	 */
 419	if (its->flags & ITS_FLAGS_CMDQ_NEEDS_FLUSHING)
 420		__flush_dcache_area(cmd, sizeof(*cmd));
 421	else
 422		dsb(ishst);
 423}
 424
 425static void its_wait_for_range_completion(struct its_node *its,
 426					  struct its_cmd_block *from,
 427					  struct its_cmd_block *to)
 428{
 429	u64 rd_idx, from_idx, to_idx;
 430	u32 count = 1000000;	/* 1s! */
 431
 432	from_idx = its_cmd_ptr_to_offset(its, from);
 433	to_idx = its_cmd_ptr_to_offset(its, to);
 
 
 
 
 434
 435	while (1) {
 
 
 436		rd_idx = readl_relaxed(its->base + GITS_CREADR);
 437		if (rd_idx >= to_idx || rd_idx < from_idx)
 
 
 
 
 
 
 
 
 
 
 438			break;
 439
 440		count--;
 441		if (!count) {
 442			pr_err_ratelimited("ITS queue timeout\n");
 443			return;
 
 444		}
 
 445		cpu_relax();
 446		udelay(1);
 447	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 448}
 449
 450static void its_send_single_command(struct its_node *its,
 451				    its_cmd_builder_t builder,
 452				    struct its_cmd_desc *desc)
 
 453{
 454	struct its_cmd_block *cmd, *sync_cmd, *next_cmd;
 455	struct its_collection *sync_col;
 456	unsigned long flags;
 457
 458	raw_spin_lock_irqsave(&its->lock, flags);
 
 459
 460	cmd = its_allocate_entry(its);
 461	if (!cmd) {		/* We're soooooo screewed... */
 462		pr_err_ratelimited("ITS can't allocate, dropping command\n");
 463		raw_spin_unlock_irqrestore(&its->lock, flags);
 464		return;
 465	}
 466	sync_col = builder(cmd, desc);
 467	its_flush_cmd(its, cmd);
 468
 469	if (sync_col) {
 470		sync_cmd = its_allocate_entry(its);
 471		if (!sync_cmd) {
 472			pr_err_ratelimited("ITS can't SYNC, skipping\n");
 473			goto post;
 474		}
 475		its_encode_cmd(sync_cmd, GITS_CMD_SYNC);
 476		its_encode_target(sync_cmd, sync_col->target_address);
 477		its_fixup_cmd(sync_cmd);
 478		its_flush_cmd(its, sync_cmd);
 479	}
 480
 481post:
 482	next_cmd = its_post_commands(its);
 483	raw_spin_unlock_irqrestore(&its->lock, flags);
 484
 485	its_wait_for_range_completion(its, cmd, next_cmd);
 486}
 487
 488static void its_send_inv(struct its_device *dev, u32 event_id)
 489{
 490	struct its_cmd_desc desc;
 491
 492	desc.its_inv_cmd.dev = dev;
 493	desc.its_inv_cmd.event_id = event_id;
 494
 495	its_send_single_command(dev->its, its_build_inv_cmd, &desc);
 496}
 497
 498static void its_send_mapd(struct its_device *dev, int valid)
 499{
 500	struct its_cmd_desc desc;
 501
 502	desc.its_mapd_cmd.dev = dev;
 503	desc.its_mapd_cmd.valid = !!valid;
 504
 505	its_send_single_command(dev->its, its_build_mapd_cmd, &desc);
 506}
 507
 508static void its_send_mapc(struct its_node *its, struct its_collection *col,
 509			  int valid)
 510{
 511	struct its_cmd_desc desc;
 512
 513	desc.its_mapc_cmd.col = col;
 514	desc.its_mapc_cmd.valid = !!valid;
 515
 516	its_send_single_command(its, its_build_mapc_cmd, &desc);
 517}
 518
 519static void its_send_mapvi(struct its_device *dev, u32 irq_id, u32 id)
 520{
 521	struct its_cmd_desc desc;
 522
 523	desc.its_mapvi_cmd.dev = dev;
 524	desc.its_mapvi_cmd.phys_id = irq_id;
 525	desc.its_mapvi_cmd.event_id = id;
 526
 527	its_send_single_command(dev->its, its_build_mapvi_cmd, &desc);
 528}
 529
 530static void its_send_movi(struct its_device *dev,
 531			  struct its_collection *col, u32 id)
 532{
 533	struct its_cmd_desc desc;
 534
 535	desc.its_movi_cmd.dev = dev;
 536	desc.its_movi_cmd.col = col;
 537	desc.its_movi_cmd.event_id = id;
 538
 539	its_send_single_command(dev->its, its_build_movi_cmd, &desc);
 540}
 541
 542static void its_send_discard(struct its_device *dev, u32 id)
 543{
 544	struct its_cmd_desc desc;
 545
 546	desc.its_discard_cmd.dev = dev;
 547	desc.its_discard_cmd.event_id = id;
 548
 549	its_send_single_command(dev->its, its_build_discard_cmd, &desc);
 550}
 551
 552static void its_send_invall(struct its_node *its, struct its_collection *col)
 553{
 554	struct its_cmd_desc desc;
 555
 556	desc.its_invall_cmd.col = col;
 557
 558	its_send_single_command(its, its_build_invall_cmd, &desc);
 559}
 560
 561/*
 562 * irqchip functions - assumes MSI, mostly.
 563 */
 
 564
 565static inline u32 its_get_event_id(struct irq_data *d)
 
 
 
 
 
 
 
 
 
 566{
 567	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 568	return d->hwirq - its_dev->event_map.lpi_base;
 
 
 
 
 
 
 
 569}
 570
 571static void lpi_set_config(struct irq_data *d, bool enable)
 
 572{
 573	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 574	irq_hw_number_t hwirq = d->hwirq;
 575	u32 id = its_get_event_id(d);
 576	u8 *cfg = page_address(gic_rdists->prop_page) + hwirq - 8192;
 577
 578	if (enable)
 579		*cfg |= LPI_PROP_ENABLED;
 580	else
 581		*cfg &= ~LPI_PROP_ENABLED;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 582
 583	/*
 584	 * Make the above write visible to the redistributors.
 585	 * And yes, we're flushing exactly: One. Single. Byte.
 586	 * Humpf...
 587	 */
 588	if (gic_rdists->flags & RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING)
 589		__flush_dcache_area(cfg, sizeof(*cfg));
 590	else
 591		dsb(ishst);
 592	its_send_inv(its_dev, id);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 593}
 594
 595static void its_mask_irq(struct irq_data *d)
 596{
 597	lpi_set_config(d, false);
 
 
 
 598}
 599
 600static void its_unmask_irq(struct irq_data *d)
 601{
 602	lpi_set_config(d, true);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 603}
 604
 605static int its_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
 606			    bool force)
 607{
 608	unsigned int cpu = cpumask_any_and(mask_val, cpu_online_mask);
 609	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 610	struct its_collection *target_col;
 611	u32 id = its_get_event_id(d);
 
 612
 613	if (cpu >= nr_cpu_ids)
 
 614		return -EINVAL;
 615
 616	target_col = &its_dev->its->collections[cpu];
 617	its_send_movi(its_dev, target_col, id);
 618	its_dev->event_map.col_map[id] = cpu;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 619
 620	return IRQ_SET_MASK_OK_DONE;
 
 
 
 
 
 
 
 
 
 
 
 621}
 622
 623static void its_irq_compose_msi_msg(struct irq_data *d, struct msi_msg *msg)
 624{
 625	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 626	struct its_node *its;
 627	u64 addr;
 628
 629	its = its_dev->its;
 630	addr = its->phys_base + GITS_TRANSLATER;
 631
 632	msg->address_lo		= addr & ((1UL << 32) - 1);
 633	msg->address_hi		= addr >> 32;
 634	msg->data		= its_get_event_id(d);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 635}
 636
 637static struct irq_chip its_irq_chip = {
 638	.name			= "ITS",
 639	.irq_mask		= its_mask_irq,
 640	.irq_unmask		= its_unmask_irq,
 641	.irq_eoi		= irq_chip_eoi_parent,
 642	.irq_set_affinity	= its_set_affinity,
 643	.irq_compose_msi_msg	= its_irq_compose_msi_msg,
 
 
 
 644};
 645
 
 646/*
 647 * How we allocate LPIs:
 648 *
 649 * The GIC has id_bits bits for interrupt identifiers. From there, we
 650 * must subtract 8192 which are reserved for SGIs/PPIs/SPIs. Then, as
 651 * we allocate LPIs by chunks of 32, we can shift the whole thing by 5
 652 * bits to the right.
 
 
 
 
 653 *
 654 * This gives us (((1UL << id_bits) - 8192) >> 5) possible allocations.
 
 655 */
 656#define IRQS_PER_CHUNK_SHIFT	5
 657#define IRQS_PER_CHUNK		(1 << IRQS_PER_CHUNK_SHIFT)
 658
 659static unsigned long *lpi_bitmap;
 660static u32 lpi_chunks;
 661static DEFINE_SPINLOCK(lpi_lock);
 662
 663static int its_lpi_to_chunk(int lpi)
 
 
 
 
 
 
 664{
 665	return (lpi - 8192) >> IRQS_PER_CHUNK_SHIFT;
 
 
 
 
 
 
 
 
 666}
 667
 668static int its_chunk_to_lpi(int chunk)
 669{
 670	return (chunk << IRQS_PER_CHUNK_SHIFT) + 8192;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 671}
 672
 673static int __init its_lpi_init(u32 id_bits)
 
 
 
 
 
 
 
 
 
 
 
 
 674{
 675	lpi_chunks = its_lpi_to_chunk(1UL << id_bits);
 676
 677	lpi_bitmap = kzalloc(BITS_TO_LONGS(lpi_chunks) * sizeof(long),
 678			     GFP_KERNEL);
 679	if (!lpi_bitmap) {
 680		lpi_chunks = 0;
 681		return -ENOMEM;
 
 
 
 
 
 
 682	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 683
 684	pr_info("ITS: Allocated %d chunks for LPIs\n", (int)lpi_chunks);
 685	return 0;
 686}
 687
 688static unsigned long *its_lpi_alloc_chunks(int nr_irqs, int *base, int *nr_ids)
 689{
 690	unsigned long *bitmap = NULL;
 691	int chunk_id;
 692	int nr_chunks;
 693	int i;
 694
 695	nr_chunks = DIV_ROUND_UP(nr_irqs, IRQS_PER_CHUNK);
 
 
 
 
 
 
 696
 697	spin_lock(&lpi_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 698
 699	do {
 700		chunk_id = bitmap_find_next_zero_area(lpi_bitmap, lpi_chunks,
 701						      0, nr_chunks, 0);
 702		if (chunk_id < lpi_chunks)
 703			break;
 704
 705		nr_chunks--;
 706	} while (nr_chunks > 0);
 707
 708	if (!nr_chunks)
 
 
 
 709		goto out;
 710
 711	bitmap = kzalloc(BITS_TO_LONGS(nr_chunks * IRQS_PER_CHUNK) * sizeof (long),
 712			 GFP_ATOMIC);
 713	if (!bitmap)
 714		goto out;
 715
 716	for (i = 0; i < nr_chunks; i++)
 717		set_bit(chunk_id + i, lpi_bitmap);
 718
 719	*base = its_chunk_to_lpi(chunk_id);
 720	*nr_ids = nr_chunks * IRQS_PER_CHUNK;
 721
 722out:
 723	spin_unlock(&lpi_lock);
 724
 725	if (!bitmap)
 726		*base = *nr_ids = 0;
 727
 728	return bitmap;
 729}
 730
 731static void its_lpi_free(struct event_lpi_map *map)
 732{
 733	int base = map->lpi_base;
 734	int nr_ids = map->nr_lpis;
 735	int lpi;
 736
 737	spin_lock(&lpi_lock);
 738
 739	for (lpi = base; lpi < (base + nr_ids); lpi += IRQS_PER_CHUNK) {
 740		int chunk = its_lpi_to_chunk(lpi);
 741		BUG_ON(chunk > lpi_chunks);
 742		if (test_bit(chunk, lpi_bitmap)) {
 743			clear_bit(chunk, lpi_bitmap);
 744		} else {
 745			pr_err("Bad LPI chunk %d\n", chunk);
 746		}
 747	}
 748
 749	spin_unlock(&lpi_lock);
 
 
 
 750
 751	kfree(map->lpi_map);
 752	kfree(map->col_map);
 753}
 754
 755/*
 756 * We allocate 64kB for PROPBASE. That gives us at most 64K LPIs to
 757 * deal with (one configuration byte per interrupt). PENDBASE has to
 758 * be 64kB aligned (one bit per LPI, plus 8192 bits for SPI/PPI/SGI).
 759 */
 760#define LPI_PROPBASE_SZ		SZ_64K
 761#define LPI_PENDBASE_SZ		(LPI_PROPBASE_SZ / 8 + SZ_1K)
 762
 763/*
 764 * This is how many bits of ID we need, including the useless ones.
 765 */
 766#define LPI_NRBITS		ilog2(LPI_PROPBASE_SZ + SZ_8K)
 
 767
 768#define LPI_PROP_DEFAULT_PRIO	0xa0
 
 769
 770static int __init its_alloc_lpi_tables(void)
 771{
 772	phys_addr_t paddr;
 
 
 773
 774	gic_rdists->prop_page = alloc_pages(GFP_NOWAIT,
 775					   get_order(LPI_PROPBASE_SZ));
 776	if (!gic_rdists->prop_page) {
 777		pr_err("Failed to allocate PROPBASE\n");
 778		return -ENOMEM;
 779	}
 780
 781	paddr = page_to_phys(gic_rdists->prop_page);
 782	pr_info("GIC: using LPI property table @%pa\n", &paddr);
 
 
 
 
 
 783
 784	/* Priority 0xa0, Group-1, disabled */
 785	memset(page_address(gic_rdists->prop_page),
 786	       LPI_PROP_DEFAULT_PRIO | LPI_PROP_GROUP1,
 787	       LPI_PROPBASE_SZ);
 788
 789	/* Make sure the GIC will observe the written configuration */
 790	__flush_dcache_area(page_address(gic_rdists->prop_page), LPI_PROPBASE_SZ);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 791
 792	return 0;
 793}
 794
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 795static const char *its_base_type_string[] = {
 796	[GITS_BASER_TYPE_DEVICE]	= "Devices",
 797	[GITS_BASER_TYPE_VCPU]		= "Virtual CPUs",
 798	[GITS_BASER_TYPE_CPU]		= "Physical CPUs",
 799	[GITS_BASER_TYPE_COLLECTION]	= "Interrupt Collections",
 800	[GITS_BASER_TYPE_RESERVED5] 	= "Reserved (5)",
 801	[GITS_BASER_TYPE_RESERVED6] 	= "Reserved (6)",
 802	[GITS_BASER_TYPE_RESERVED7] 	= "Reserved (7)",
 803};
 804
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 805static void its_free_tables(struct its_node *its)
 806{
 807	int i;
 808
 809	for (i = 0; i < GITS_BASER_NR_REGS; i++) {
 810		if (its->tables[i].base) {
 811			free_pages((unsigned long)its->tables[i].base,
 812				   its->tables[i].order);
 813			its->tables[i].base = NULL;
 814		}
 815	}
 816}
 817
 818static int its_alloc_tables(const char *node_name, struct its_node *its)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 819{
 820	int err;
 821	int i;
 822	int psz = SZ_64K;
 823	u64 shr = GITS_BASER_InnerShareable;
 824	u64 cache;
 825	u64 typer;
 826	u32 ids;
 827
 828	if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_22375) {
 829		/*
 830		 * erratum 22375: only alloc 8MB table size
 831		 * erratum 24313: ignore memory access type
 832		 */
 833		cache	= 0;
 834		ids	= 0x14;			/* 20 bits, 8MB */
 835	} else {
 836		cache	= GITS_BASER_WaWb;
 837		typer	= readq_relaxed(its->base + GITS_TYPER);
 838		ids	= GITS_TYPER_DEVBITS(typer);
 839	}
 840
 841	for (i = 0; i < GITS_BASER_NR_REGS; i++) {
 842		u64 val = readq_relaxed(its->base + GITS_BASER + i * 8);
 
 843		u64 type = GITS_BASER_TYPE(val);
 844		u64 entry_size = GITS_BASER_ENTRY_SIZE(val);
 845		int order = get_order(psz);
 846		int alloc_pages;
 847		u64 tmp;
 848		void *base;
 849
 850		if (type == GITS_BASER_TYPE_NONE)
 851			continue;
 852
 853		/*
 854		 * Allocate as many entries as required to fit the
 855		 * range of device IDs that the ITS can grok... The ID
 856		 * space being incredibly sparse, this results in a
 857		 * massive waste of memory.
 858		 *
 859		 * For other tables, only allocate a single page.
 860		 */
 861		if (type == GITS_BASER_TYPE_DEVICE) {
 862			/*
 863			 * 'order' was initialized earlier to the default page
 864			 * granule of the the ITS.  We can't have an allocation
 865			 * smaller than that.  If the requested allocation
 866			 * is smaller, round up to the default page granule.
 867			 */
 868			order = max(get_order((1UL << ids) * entry_size),
 869				    order);
 870			if (order >= MAX_ORDER) {
 871				order = MAX_ORDER - 1;
 872				pr_warn("%s: Device Table too large, reduce its page order to %u\n",
 873					node_name, order);
 
 
 874			}
 
 
 
 
 875		}
 876
 877retry_alloc_baser:
 878		alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz);
 879		if (alloc_pages > GITS_BASER_PAGES_MAX) {
 880			alloc_pages = GITS_BASER_PAGES_MAX;
 881			order = get_order(GITS_BASER_PAGES_MAX * psz);
 882			pr_warn("%s: Device Table too large, reduce its page order to %u (%u pages)\n",
 883				node_name, order, alloc_pages);
 884		}
 885
 886		base = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
 887		if (!base) {
 888			err = -ENOMEM;
 889			goto out_free;
 890		}
 891
 892		its->tables[i].base = base;
 893		its->tables[i].order = order;
 
 
 894
 895retry_baser:
 896		val = (virt_to_phys(base) 				 |
 897		       (type << GITS_BASER_TYPE_SHIFT)			 |
 898		       ((entry_size - 1) << GITS_BASER_ENTRY_SIZE_SHIFT) |
 899		       cache						 |
 900		       shr						 |
 901		       GITS_BASER_VALID);
 902
 903		switch (psz) {
 904		case SZ_4K:
 905			val |= GITS_BASER_PAGE_SIZE_4K;
 906			break;
 907		case SZ_16K:
 908			val |= GITS_BASER_PAGE_SIZE_16K;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 909			break;
 910		case SZ_64K:
 911			val |= GITS_BASER_PAGE_SIZE_64K;
 912			break;
 913		}
 
 
 
 
 
 
 
 
 914
 915		val |= alloc_pages - 1;
 
 916
 917		writeq_relaxed(val, its->base + GITS_BASER + i * 8);
 918		tmp = readq_relaxed(its->base + GITS_BASER + i * 8);
 919
 920		if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
 921			/*
 922			 * Shareability didn't stick. Just use
 923			 * whatever the read reported, which is likely
 924			 * to be the only thing this redistributor
 925			 * supports. If that's zero, make it
 926			 * non-cacheable as well.
 927			 */
 928			shr = tmp & GITS_BASER_SHAREABILITY_MASK;
 929			if (!shr) {
 930				cache = GITS_BASER_nC;
 931				__flush_dcache_area(base, PAGE_ORDER_TO_SIZE(order));
 932			}
 933			goto retry_baser;
 934		}
 935
 936		if ((val ^ tmp) & GITS_BASER_PAGE_SIZE_MASK) {
 937			/*
 938			 * Page size didn't stick. Let's try a smaller
 939			 * size and retry. If we reach 4K, then
 940			 * something is horribly wrong...
 941			 */
 942			free_pages((unsigned long)base, order);
 943			its->tables[i].base = NULL;
 944
 945			switch (psz) {
 946			case SZ_16K:
 947				psz = SZ_4K;
 948				goto retry_alloc_baser;
 949			case SZ_64K:
 950				psz = SZ_16K;
 951				goto retry_alloc_baser;
 952			}
 953		}
 954
 955		if (val != tmp) {
 956			pr_err("ITS: %s: GITS_BASER%d doesn't stick: %lx %lx\n",
 957			       node_name, i,
 958			       (unsigned long) val, (unsigned long) tmp);
 959			err = -ENXIO;
 960			goto out_free;
 961		}
 
 
 
 
 
 
 
 
 962
 963		pr_info("ITS: allocated %d %s @%lx (psz %dK, shr %d)\n",
 964			(int)(PAGE_ORDER_TO_SIZE(order) / entry_size),
 965			its_base_type_string[type],
 966			(unsigned long)virt_to_phys(base),
 967			psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
 968	}
 969
 970	return 0;
 
 971
 972out_free:
 973	its_free_tables(its);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 974
 975	return err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 976}
 977
 978static int its_alloc_collections(struct its_node *its)
 979{
 980	its->collections = kzalloc(nr_cpu_ids * sizeof(*its->collections),
 
 
 981				   GFP_KERNEL);
 982	if (!its->collections)
 983		return -ENOMEM;
 984
 
 
 
 985	return 0;
 986}
 987
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 988static void its_cpu_init_lpis(void)
 989{
 990	void __iomem *rbase = gic_data_rdist_rd_base();
 991	struct page *pend_page;
 
 992	u64 val, tmp;
 993
 994	/* If we didn't allocate the pending table yet, do it now */
 995	pend_page = gic_data_rdist()->pend_page;
 996	if (!pend_page) {
 997		phys_addr_t paddr;
 
 
 998		/*
 999		 * The pending pages have to be at least 64kB aligned,
1000		 * hence the 'max(LPI_PENDBASE_SZ, SZ_64K)' below.
1001		 */
1002		pend_page = alloc_pages(GFP_NOWAIT | __GFP_ZERO,
1003					get_order(max(LPI_PENDBASE_SZ, SZ_64K)));
1004		if (!pend_page) {
1005			pr_err("Failed to allocate PENDBASE for CPU%d\n",
1006			       smp_processor_id());
1007			return;
1008		}
1009
1010		/* Make sure the GIC will observe the zero-ed page */
1011		__flush_dcache_area(page_address(pend_page), LPI_PENDBASE_SZ);
1012
1013		paddr = page_to_phys(pend_page);
1014		pr_info("CPU%d: using LPI pending table @%pa\n",
1015			smp_processor_id(), &paddr);
1016		gic_data_rdist()->pend_page = pend_page;
1017	}
1018
1019	/* Disable LPIs */
1020	val = readl_relaxed(rbase + GICR_CTLR);
1021	val &= ~GICR_CTLR_ENABLE_LPIS;
1022	writel_relaxed(val, rbase + GICR_CTLR);
1023
1024	/*
1025	 * Make sure any change to the table is observable by the GIC.
1026	 */
1027	dsb(sy);
1028
1029	/* set PROPBASE */
1030	val = (page_to_phys(gic_rdists->prop_page) |
1031	       GICR_PROPBASER_InnerShareable |
1032	       GICR_PROPBASER_WaWb |
1033	       ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
1034
1035	writeq_relaxed(val, rbase + GICR_PROPBASER);
1036	tmp = readq_relaxed(rbase + GICR_PROPBASER);
 
 
 
1037
1038	if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
1039		if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
1040			/*
1041			 * The HW reports non-shareable, we must
1042			 * remove the cacheability attributes as
1043			 * well.
1044			 */
1045			val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
1046				 GICR_PROPBASER_CACHEABILITY_MASK);
1047			val |= GICR_PROPBASER_nC;
1048			writeq_relaxed(val, rbase + GICR_PROPBASER);
1049		}
1050		pr_info_once("GIC: using cache flushing for LPI property table\n");
1051		gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
1052	}
1053
1054	/* set PENDBASE */
1055	val = (page_to_phys(pend_page) |
1056	       GICR_PENDBASER_InnerShareable |
1057	       GICR_PENDBASER_WaWb);
1058
1059	writeq_relaxed(val, rbase + GICR_PENDBASER);
1060	tmp = readq_relaxed(rbase + GICR_PENDBASER);
 
 
 
1061
1062	if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
1063		/*
1064		 * The HW reports non-shareable, we must remove the
1065		 * cacheability attributes as well.
1066		 */
1067		val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
1068			 GICR_PENDBASER_CACHEABILITY_MASK);
1069		val |= GICR_PENDBASER_nC;
1070		writeq_relaxed(val, rbase + GICR_PENDBASER);
1071	}
1072
1073	/* Enable LPIs */
1074	val = readl_relaxed(rbase + GICR_CTLR);
1075	val |= GICR_CTLR_ENABLE_LPIS;
1076	writel_relaxed(val, rbase + GICR_CTLR);
1077
1078	/* Make sure the GIC has seen the above */
1079	dsb(sy);
1080}
1081
1082static void its_cpu_init_collection(void)
1083{
1084	struct its_node *its;
1085	int cpu;
 
 
 
 
 
 
 
1086
1087	spin_lock(&its_lock);
1088	cpu = smp_processor_id();
 
 
 
 
 
1089
1090	list_for_each_entry(its, &its_nodes, entry) {
1091		u64 target;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1092
 
 
 
 
 
1093		/*
1094		 * We now have to bind each collection to its target
1095		 * redistributor.
1096		 */
1097		if (readq_relaxed(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
1098			/*
1099			 * This ITS wants the physical address of the
1100			 * redistributor.
1101			 */
1102			target = gic_data_rdist()->phys_base;
1103		} else {
1104			/*
1105			 * This ITS wants a linear CPU number.
1106			 */
1107			target = readq_relaxed(gic_data_rdist_rd_base() + GICR_TYPER);
1108			target = GICR_TYPER_CPU_NUMBER(target) << 16;
1109		}
1110
1111		/* Perform collection mapping */
1112		its->collections[cpu].target_address = target;
1113		its->collections[cpu].col_id = cpu;
1114
1115		its_send_mapc(its, &its->collections[cpu], 1);
1116		its_send_invall(its, &its->collections[cpu]);
1117	}
 
 
 
 
1118
1119	spin_unlock(&its_lock);
 
 
 
 
 
1120}
1121
1122static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
1123{
1124	struct its_device *its_dev = NULL, *tmp;
1125	unsigned long flags;
1126
1127	raw_spin_lock_irqsave(&its->lock, flags);
1128
1129	list_for_each_entry(tmp, &its->its_device_list, entry) {
1130		if (tmp->device_id == dev_id) {
1131			its_dev = tmp;
1132			break;
1133		}
1134	}
1135
1136	raw_spin_unlock_irqrestore(&its->lock, flags);
1137
1138	return its_dev;
1139}
1140
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1141static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
1142					    int nvecs)
1143{
1144	struct its_device *dev;
1145	unsigned long *lpi_map;
1146	unsigned long flags;
1147	u16 *col_map = NULL;
1148	void *itt;
1149	int lpi_base;
1150	int nr_lpis;
1151	int nr_ites;
1152	int sz;
1153
 
 
 
 
 
 
1154	dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1155	/*
1156	 * At least one bit of EventID is being used, hence a minimum
1157	 * of two entries. No, the architecture doesn't let you
1158	 * express an ITT with a single entry.
1159	 */
1160	nr_ites = max(2UL, roundup_pow_of_two(nvecs));
1161	sz = nr_ites * its->ite_size;
1162	sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
1163	itt = kzalloc(sz, GFP_KERNEL);
1164	lpi_map = its_lpi_alloc_chunks(nvecs, &lpi_base, &nr_lpis);
1165	if (lpi_map)
1166		col_map = kzalloc(sizeof(*col_map) * nr_lpis, GFP_KERNEL);
 
 
 
 
 
 
 
1167
1168	if (!dev || !itt || !lpi_map || !col_map) {
1169		kfree(dev);
1170		kfree(itt);
1171		kfree(lpi_map);
1172		kfree(col_map);
1173		return NULL;
1174	}
1175
1176	__flush_dcache_area(itt, sz);
1177
1178	dev->its = its;
1179	dev->itt = itt;
1180	dev->nr_ites = nr_ites;
1181	dev->event_map.lpi_map = lpi_map;
1182	dev->event_map.col_map = col_map;
1183	dev->event_map.lpi_base = lpi_base;
1184	dev->event_map.nr_lpis = nr_lpis;
 
1185	dev->device_id = dev_id;
1186	INIT_LIST_HEAD(&dev->entry);
1187
1188	raw_spin_lock_irqsave(&its->lock, flags);
1189	list_add(&dev->entry, &its->its_device_list);
1190	raw_spin_unlock_irqrestore(&its->lock, flags);
1191
1192	/* Map device to its ITT */
1193	its_send_mapd(dev, 1);
1194
1195	return dev;
1196}
1197
1198static void its_free_device(struct its_device *its_dev)
1199{
1200	unsigned long flags;
1201
1202	raw_spin_lock_irqsave(&its_dev->its->lock, flags);
1203	list_del(&its_dev->entry);
1204	raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
 
1205	kfree(its_dev->itt);
1206	kfree(its_dev);
1207}
1208
1209static int its_alloc_device_irq(struct its_device *dev, irq_hw_number_t *hwirq)
1210{
1211	int idx;
1212
1213	idx = find_first_zero_bit(dev->event_map.lpi_map,
1214				  dev->event_map.nr_lpis);
1215	if (idx == dev->event_map.nr_lpis)
 
 
1216		return -ENOSPC;
1217
1218	*hwirq = dev->event_map.lpi_base + idx;
1219	set_bit(idx, dev->event_map.lpi_map);
1220
1221	return 0;
1222}
1223
1224static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
1225			   int nvec, msi_alloc_info_t *info)
1226{
1227	struct its_node *its;
1228	struct its_device *its_dev;
1229	struct msi_domain_info *msi_info;
1230	u32 dev_id;
 
1231
1232	/*
1233	 * We ignore "dev" entierely, and rely on the dev_id that has
1234	 * been passed via the scratchpad. This limits this domain's
1235	 * usefulness to upper layers that definitely know that they
1236	 * are built on top of the ITS.
1237	 */
1238	dev_id = info->scratchpad[0].ul;
1239
1240	msi_info = msi_get_domain_info(domain);
1241	its = msi_info->data;
1242
 
 
 
 
 
 
 
 
 
 
 
1243	its_dev = its_find_device(its, dev_id);
1244	if (its_dev) {
1245		/*
1246		 * We already have seen this ID, probably through
1247		 * another alias (PCI bridge of some sort). No need to
1248		 * create the device.
1249		 */
 
1250		pr_debug("Reusing ITT for devID %x\n", dev_id);
1251		goto out;
1252	}
1253
1254	its_dev = its_create_device(its, dev_id, nvec);
1255	if (!its_dev)
1256		return -ENOMEM;
 
 
 
 
 
1257
1258	pr_debug("ITT %d entries, %d bits\n", nvec, ilog2(nvec));
1259out:
 
1260	info->scratchpad[0].ptr = its_dev;
1261	return 0;
1262}
1263
1264static struct msi_domain_ops its_msi_domain_ops = {
1265	.msi_prepare	= its_msi_prepare,
1266};
1267
1268static int its_irq_gic_domain_alloc(struct irq_domain *domain,
1269				    unsigned int virq,
1270				    irq_hw_number_t hwirq)
1271{
1272	struct irq_fwspec fwspec;
1273
1274	if (irq_domain_get_of_node(domain->parent)) {
1275		fwspec.fwnode = domain->parent->fwnode;
1276		fwspec.param_count = 3;
1277		fwspec.param[0] = GIC_IRQ_TYPE_LPI;
1278		fwspec.param[1] = hwirq;
1279		fwspec.param[2] = IRQ_TYPE_EDGE_RISING;
 
 
 
 
 
1280	} else {
1281		return -EINVAL;
1282	}
1283
1284	return irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
1285}
1286
1287static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
1288				unsigned int nr_irqs, void *args)
1289{
1290	msi_alloc_info_t *info = args;
1291	struct its_device *its_dev = info->scratchpad[0].ptr;
 
 
1292	irq_hw_number_t hwirq;
1293	int err;
1294	int i;
1295
1296	for (i = 0; i < nr_irqs; i++) {
1297		err = its_alloc_device_irq(its_dev, &hwirq);
1298		if (err)
1299			return err;
 
 
 
1300
1301		err = its_irq_gic_domain_alloc(domain, virq + i, hwirq);
 
1302		if (err)
1303			return err;
1304
1305		irq_domain_set_hwirq_and_chip(domain, virq + i,
1306					      hwirq, &its_irq_chip, its_dev);
 
 
 
 
1307		pr_debug("ID:%d pID:%d vID:%d\n",
1308			 (int)(hwirq - its_dev->event_map.lpi_base),
1309			 (int) hwirq, virq + i);
1310	}
1311
1312	return 0;
1313}
1314
1315static void its_irq_domain_activate(struct irq_domain *domain,
1316				    struct irq_data *d)
1317{
1318	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1319	u32 event = its_get_event_id(d);
 
 
 
 
 
1320
1321	/* Bind the LPI to the first possible CPU */
1322	its_dev->event_map.col_map[event] = cpumask_first(cpu_online_mask);
 
1323
1324	/* Map the GIC IRQ and event to the device */
1325	its_send_mapvi(its_dev, d->hwirq, event);
 
1326}
1327
1328static void its_irq_domain_deactivate(struct irq_domain *domain,
1329				      struct irq_data *d)
1330{
1331	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1332	u32 event = its_get_event_id(d);
1333
 
1334	/* Stop the delivery of interrupts */
1335	its_send_discard(its_dev, event);
1336}
1337
1338static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
1339				unsigned int nr_irqs)
1340{
1341	struct irq_data *d = irq_domain_get_irq_data(domain, virq);
1342	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 
1343	int i;
1344
 
 
 
 
1345	for (i = 0; i < nr_irqs; i++) {
1346		struct irq_data *data = irq_domain_get_irq_data(domain,
1347								virq + i);
1348		u32 event = its_get_event_id(data);
1349
1350		/* Mark interrupt index as unused */
1351		clear_bit(event, its_dev->event_map.lpi_map);
1352
1353		/* Nuke the entry in the domain */
1354		irq_domain_reset_irq_data(data);
1355	}
1356
1357	/* If all interrupts have been freed, start mopping the floor */
1358	if (bitmap_empty(its_dev->event_map.lpi_map,
 
 
 
 
 
 
1359			 its_dev->event_map.nr_lpis)) {
1360		its_lpi_free(&its_dev->event_map);
 
 
1361
1362		/* Unmap device/itt */
1363		its_send_mapd(its_dev, 0);
1364		its_free_device(its_dev);
1365	}
1366
 
 
1367	irq_domain_free_irqs_parent(domain, virq, nr_irqs);
1368}
1369
1370static const struct irq_domain_ops its_domain_ops = {
1371	.alloc			= its_irq_domain_alloc,
1372	.free			= its_irq_domain_free,
1373	.activate		= its_irq_domain_activate,
1374	.deactivate		= its_irq_domain_deactivate,
1375};
1376
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1377static int its_force_quiescent(void __iomem *base)
1378{
1379	u32 count = 1000000;	/* 1s */
1380	u32 val;
1381
1382	val = readl_relaxed(base + GITS_CTLR);
1383	if (val & GITS_CTLR_QUIESCENT)
 
 
 
 
 
1384		return 0;
1385
1386	/* Disable the generation of all interrupts to this ITS */
1387	val &= ~GITS_CTLR_ENABLE;
1388	writel_relaxed(val, base + GITS_CTLR);
1389
1390	/* Poll GITS_CTLR and wait until ITS becomes quiescent */
1391	while (1) {
1392		val = readl_relaxed(base + GITS_CTLR);
1393		if (val & GITS_CTLR_QUIESCENT)
1394			return 0;
1395
1396		count--;
1397		if (!count)
1398			return -EBUSY;
1399
1400		cpu_relax();
1401		udelay(1);
1402	}
1403}
1404
1405static void __maybe_unused its_enable_quirk_cavium_22375(void *data)
1406{
1407	struct its_node *its = data;
1408
 
 
 
1409	its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_22375;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1410}
1411
1412static const struct gic_quirk its_quirks[] = {
1413#ifdef CONFIG_CAVIUM_ERRATUM_22375
1414	{
1415		.desc	= "ITS: Cavium errata 22375, 24313",
1416		.iidr	= 0xa100034c,	/* ThunderX pass 1.x */
1417		.mask	= 0xffff0fff,
1418		.init	= its_enable_quirk_cavium_22375,
1419	},
1420#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1421	{
1422	}
1423};
1424
1425static void its_enable_quirks(struct its_node *its)
1426{
1427	u32 iidr = readl_relaxed(its->base + GITS_IIDR);
1428
1429	gic_enable_quirks(iidr, its_quirks, its);
 
 
 
 
1430}
1431
1432static int __init its_probe(struct device_node *node,
1433			    struct irq_domain *parent)
1434{
1435	struct resource res;
1436	struct its_node *its;
1437	void __iomem *its_base;
1438	struct irq_domain *inner_domain;
1439	u32 val;
1440	u64 baser, tmp;
1441	int err;
 
 
 
 
 
 
 
 
 
 
 
 
 
1442
1443	err = of_address_to_resource(node, 0, &res);
1444	if (err) {
1445		pr_warn("%s: no regs?\n", node->full_name);
1446		return -ENXIO;
 
 
 
 
1447	}
 
 
 
 
 
 
 
 
 
1448
1449	its_base = ioremap(res.start, resource_size(&res));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1450	if (!its_base) {
1451		pr_warn("%s: unable to map registers\n", node->full_name);
1452		return -ENOMEM;
 
1453	}
1454
1455	val = readl_relaxed(its_base + GITS_PIDR2) & GIC_PIDR2_ARCH_MASK;
1456	if (val != 0x30 && val != 0x40) {
1457		pr_warn("%s: no ITS detected, giving up\n", node->full_name);
1458		err = -ENODEV;
1459		goto out_unmap;
1460	}
1461
1462	err = its_force_quiescent(its_base);
1463	if (err) {
1464		pr_warn("%s: failed to quiesce, giving up\n",
1465			node->full_name);
1466		goto out_unmap;
1467	}
1468
1469	pr_info("ITS: %s\n", node->full_name);
1470
1471	its = kzalloc(sizeof(*its), GFP_KERNEL);
1472	if (!its) {
1473		err = -ENOMEM;
1474		goto out_unmap;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1475	}
1476
1477	raw_spin_lock_init(&its->lock);
1478	INIT_LIST_HEAD(&its->entry);
1479	INIT_LIST_HEAD(&its->its_device_list);
1480	its->base = its_base;
1481	its->phys_base = res.start;
1482	its->ite_size = ((readl_relaxed(its_base + GITS_TYPER) >> 4) & 0xf) + 1;
1483
1484	its->cmd_base = kzalloc(ITS_CMD_QUEUE_SZ, GFP_KERNEL);
1485	if (!its->cmd_base) {
1486		err = -ENOMEM;
1487		goto out_free_its;
 
 
 
 
 
 
 
 
1488	}
1489	its->cmd_write = its->cmd_base;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1490
1491	its_enable_quirks(its);
1492
1493	err = its_alloc_tables(node->full_name, its);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1494	if (err)
1495		goto out_free_cmd;
1496
1497	err = its_alloc_collections(its);
1498	if (err)
1499		goto out_free_tables;
1500
1501	baser = (virt_to_phys(its->cmd_base)	|
1502		 GITS_CBASER_WaWb		|
1503		 GITS_CBASER_InnerShareable	|
1504		 (ITS_CMD_QUEUE_SZ / SZ_4K - 1)	|
1505		 GITS_CBASER_VALID);
1506
1507	writeq_relaxed(baser, its->base + GITS_CBASER);
1508	tmp = readq_relaxed(its->base + GITS_CBASER);
 
 
 
1509
1510	if ((tmp ^ baser) & GITS_CBASER_SHAREABILITY_MASK) {
1511		if (!(tmp & GITS_CBASER_SHAREABILITY_MASK)) {
1512			/*
1513			 * The HW reports non-shareable, we must
1514			 * remove the cacheability attributes as
1515			 * well.
1516			 */
1517			baser &= ~(GITS_CBASER_SHAREABILITY_MASK |
1518				   GITS_CBASER_CACHEABILITY_MASK);
1519			baser |= GITS_CBASER_nC;
1520			writeq_relaxed(baser, its->base + GITS_CBASER);
1521		}
1522		pr_info("ITS: using cache flushing for cmd queue\n");
1523		its->flags |= ITS_FLAGS_CMDQ_NEEDS_FLUSHING;
1524	}
1525
1526	writeq_relaxed(0, its->base + GITS_CWRITER);
1527	writel_relaxed(GITS_CTLR_ENABLE, its->base + GITS_CTLR);
1528
1529	if (of_property_read_bool(node, "msi-controller")) {
1530		struct msi_domain_info *info;
1531
1532		info = kzalloc(sizeof(*info), GFP_KERNEL);
1533		if (!info) {
1534			err = -ENOMEM;
1535			goto out_free_tables;
1536		}
1537
1538		inner_domain = irq_domain_add_tree(node, &its_domain_ops, its);
1539		if (!inner_domain) {
1540			err = -ENOMEM;
1541			kfree(info);
1542			goto out_free_tables;
1543		}
1544
1545		inner_domain->parent = parent;
1546		inner_domain->bus_token = DOMAIN_BUS_NEXUS;
1547		info->ops = &its_msi_domain_ops;
1548		info->data = its;
1549		inner_domain->host_data = info;
1550	}
1551
1552	spin_lock(&its_lock);
1553	list_add(&its->entry, &its_nodes);
1554	spin_unlock(&its_lock);
1555
1556	return 0;
1557
1558out_free_tables:
1559	its_free_tables(its);
1560out_free_cmd:
1561	kfree(its->cmd_base);
1562out_free_its:
1563	kfree(its);
1564out_unmap:
1565	iounmap(its_base);
1566	pr_err("ITS: failed probing %s (%d)\n", node->full_name, err);
1567	return err;
1568}
1569
1570static bool gic_rdists_supports_plpis(void)
1571{
1572	return !!(readl_relaxed(gic_data_rdist_rd_base() + GICR_TYPER) & GICR_TYPER_PLPIS);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1573}
1574
1575int its_cpu_init(void)
1576{
1577	if (!list_empty(&its_nodes)) {
1578		if (!gic_rdists_supports_plpis()) {
1579			pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
1580			return -ENXIO;
1581		}
 
 
1582		its_cpu_init_lpis();
1583		its_cpu_init_collection();
1584	}
1585
1586	return 0;
1587}
1588
1589static struct of_device_id its_device_id[] = {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1590	{	.compatible	= "arm,gic-v3-its",	},
1591	{},
1592};
1593
1594int __init its_init(struct device_node *node, struct rdists *rdists,
1595	     struct irq_domain *parent_domain)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1596{
1597	struct device_node *np;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1598
1599	for (np = of_find_matching_node(node, its_device_id); np;
1600	     np = of_find_matching_node(np, its_device_id)) {
1601		its_probe(np, parent_domain);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1602	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1603
1604	if (list_empty(&its_nodes)) {
1605		pr_warn("ITS: No ITS available, not enabling LPIs\n");
1606		return -ENXIO;
1607	}
1608
1609	gic_rdists = rdists;
1610	its_alloc_lpi_tables();
1611	its_lpi_init(rdists->id_bits);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1612
1613	return 0;
1614}
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2013-2017 ARM Limited, All Rights Reserved.
   4 * Author: Marc Zyngier <marc.zyngier@arm.com>
 
 
 
 
 
 
 
 
 
 
 
 
   5 */
   6
   7#include <linux/acpi.h>
   8#include <linux/acpi_iort.h>
   9#include <linux/bitfield.h>
  10#include <linux/bitmap.h>
  11#include <linux/cpu.h>
  12#include <linux/crash_dump.h>
  13#include <linux/delay.h>
  14#include <linux/efi.h>
  15#include <linux/interrupt.h>
  16#include <linux/iommu.h>
  17#include <linux/iopoll.h>
  18#include <linux/irqdomain.h>
  19#include <linux/list.h>
  20#include <linux/log2.h>
  21#include <linux/memblock.h>
  22#include <linux/mm.h>
  23#include <linux/msi.h>
  24#include <linux/of.h>
  25#include <linux/of_address.h>
  26#include <linux/of_irq.h>
  27#include <linux/of_pci.h>
  28#include <linux/of_platform.h>
  29#include <linux/percpu.h>
  30#include <linux/slab.h>
  31#include <linux/syscore_ops.h>
  32
  33#include <linux/irqchip.h>
  34#include <linux/irqchip/arm-gic-v3.h>
  35#include <linux/irqchip/arm-gic-v4.h>
  36
 
  37#include <asm/cputype.h>
  38#include <asm/exception.h>
  39
  40#include "irq-gic-common.h"
  41
  42#define ITS_FLAGS_CMDQ_NEEDS_FLUSHING		(1ULL << 0)
  43#define ITS_FLAGS_WORKAROUND_CAVIUM_22375	(1ULL << 1)
  44#define ITS_FLAGS_WORKAROUND_CAVIUM_23144	(1ULL << 2)
  45#define ITS_FLAGS_FORCE_NON_SHAREABLE		(1ULL << 3)
  46
  47#define RD_LOCAL_LPI_ENABLED                    BIT(0)
  48#define RD_LOCAL_PENDTABLE_PREALLOCATED         BIT(1)
  49#define RD_LOCAL_MEMRESERVE_DONE                BIT(2)
  50
  51static u32 lpi_id_bits;
  52
  53/*
  54 * We allocate memory for PROPBASE to cover 2 ^ lpi_id_bits LPIs to
  55 * deal with (one configuration byte per interrupt). PENDBASE has to
  56 * be 64kB aligned (one bit per LPI, plus 8192 bits for SPI/PPI/SGI).
  57 */
  58#define LPI_NRBITS		lpi_id_bits
  59#define LPI_PROPBASE_SZ		ALIGN(BIT(LPI_NRBITS), SZ_64K)
  60#define LPI_PENDBASE_SZ		ALIGN(BIT(LPI_NRBITS) / 8, SZ_64K)
  61
  62#define LPI_PROP_DEFAULT_PRIO	GICD_INT_DEF_PRI
  63
  64/*
  65 * Collection structure - just an ID, and a redistributor address to
  66 * ping. We use one per CPU as a bag of interrupts assigned to this
  67 * CPU.
  68 */
  69struct its_collection {
  70	u64			target_address;
  71	u16			col_id;
  72};
  73
  74/*
  75 * The ITS_BASER structure - contains memory information, cached
  76 * value of BASER register configuration and ITS page size.
  77 */
  78struct its_baser {
  79	void		*base;
  80	u64		val;
  81	u32		order;
  82	u32		psz;
  83};
  84
  85struct its_device;
  86
  87/*
  88 * The ITS structure - contains most of the infrastructure, with the
  89 * top-level MSI domain, the command queue, the collections, and the
  90 * list of devices writing to it.
  91 *
  92 * dev_alloc_lock has to be taken for device allocations, while the
  93 * spinlock must be taken to parse data structures such as the device
  94 * list.
  95 */
  96struct its_node {
  97	raw_spinlock_t		lock;
  98	struct mutex		dev_alloc_lock;
  99	struct list_head	entry;
 100	void __iomem		*base;
 101	void __iomem		*sgir_base;
 102	phys_addr_t		phys_base;
 103	struct its_cmd_block	*cmd_base;
 104	struct its_cmd_block	*cmd_write;
 105	struct its_baser	tables[GITS_BASER_NR_REGS];
 
 
 
 106	struct its_collection	*collections;
 107	struct fwnode_handle	*fwnode_handle;
 108	u64			(*get_msi_base)(struct its_device *its_dev);
 109	u64			typer;
 110	u64			cbaser_save;
 111	u32			ctlr_save;
 112	u32			mpidr;
 113	struct list_head	its_device_list;
 114	u64			flags;
 115	unsigned long		list_nr;
 116	int			numa_node;
 117	unsigned int		msi_domain_flags;
 118	u32			pre_its_base; /* for Socionext Synquacer */
 119	int			vlpi_redist_offset;
 120};
 121
 122#define is_v4(its)		(!!((its)->typer & GITS_TYPER_VLPIS))
 123#define is_v4_1(its)		(!!((its)->typer & GITS_TYPER_VMAPP))
 124#define device_ids(its)		(FIELD_GET(GITS_TYPER_DEVBITS, (its)->typer) + 1)
 125
 126#define ITS_ITT_ALIGN		SZ_256
 127
 128/* The maximum number of VPEID bits supported by VLPI commands */
 129#define ITS_MAX_VPEID_BITS						\
 130	({								\
 131		int nvpeid = 16;					\
 132		if (gic_rdists->has_rvpeid &&				\
 133		    gic_rdists->gicd_typer2 & GICD_TYPER2_VIL)		\
 134			nvpeid = 1 + (gic_rdists->gicd_typer2 &		\
 135				      GICD_TYPER2_VID);			\
 136									\
 137		nvpeid;							\
 138	})
 139#define ITS_MAX_VPEID		(1 << (ITS_MAX_VPEID_BITS))
 140
 141/* Convert page order to size in bytes */
 142#define PAGE_ORDER_TO_SIZE(o)	(PAGE_SIZE << (o))
 143
 144struct event_lpi_map {
 145	unsigned long		*lpi_map;
 146	u16			*col_map;
 147	irq_hw_number_t		lpi_base;
 148	int			nr_lpis;
 149	raw_spinlock_t		vlpi_lock;
 150	struct its_vm		*vm;
 151	struct its_vlpi_map	*vlpi_maps;
 152	int			nr_vlpis;
 153};
 154
 155/*
 156 * The ITS view of a device - belongs to an ITS, owns an interrupt
 157 * translation table, and a list of interrupts.  If it some of its
 158 * LPIs are injected into a guest (GICv4), the event_map.vm field
 159 * indicates which one.
 160 */
 161struct its_device {
 162	struct list_head	entry;
 163	struct its_node		*its;
 164	struct event_lpi_map	event_map;
 165	void			*itt;
 166	u32			nr_ites;
 167	u32			device_id;
 168	bool			shared;
 169};
 170
 171static struct {
 172	raw_spinlock_t		lock;
 173	struct its_device	*dev;
 174	struct its_vpe		**vpes;
 175	int			next_victim;
 176} vpe_proxy;
 177
 178struct cpu_lpi_count {
 179	atomic_t	managed;
 180	atomic_t	unmanaged;
 181};
 182
 183static DEFINE_PER_CPU(struct cpu_lpi_count, cpu_lpi_count);
 184
 185static LIST_HEAD(its_nodes);
 186static DEFINE_RAW_SPINLOCK(its_lock);
 187static struct rdists *gic_rdists;
 188static struct irq_domain *its_parent;
 189
 190static unsigned long its_list_map;
 191static u16 vmovp_seq_num;
 192static DEFINE_RAW_SPINLOCK(vmovp_lock);
 193
 194static DEFINE_IDA(its_vpeid_ida);
 195
 196#define gic_data_rdist()		(raw_cpu_ptr(gic_rdists->rdist))
 197#define gic_data_rdist_cpu(cpu)		(per_cpu_ptr(gic_rdists->rdist, cpu))
 198#define gic_data_rdist_rd_base()	(gic_data_rdist()->rd_base)
 199#define gic_data_rdist_vlpi_base()	(gic_data_rdist_rd_base() + SZ_128K)
 200
 201/*
 202 * Skip ITSs that have no vLPIs mapped, unless we're on GICv4.1, as we
 203 * always have vSGIs mapped.
 204 */
 205static bool require_its_list_vmovp(struct its_vm *vm, struct its_node *its)
 206{
 207	return (gic_rdists->has_rvpeid || vm->vlpi_count[its->list_nr]);
 208}
 209
 210static bool rdists_support_shareable(void)
 211{
 212	return !(gic_rdists->flags & RDIST_FLAGS_FORCE_NON_SHAREABLE);
 213}
 214
 215static u16 get_its_list(struct its_vm *vm)
 216{
 217	struct its_node *its;
 218	unsigned long its_list = 0;
 219
 220	list_for_each_entry(its, &its_nodes, entry) {
 221		if (!is_v4(its))
 222			continue;
 223
 224		if (require_its_list_vmovp(vm, its))
 225			__set_bit(its->list_nr, &its_list);
 226	}
 227
 228	return (u16)its_list;
 229}
 230
 231static inline u32 its_get_event_id(struct irq_data *d)
 232{
 233	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 234	return d->hwirq - its_dev->event_map.lpi_base;
 235}
 236
 237static struct its_collection *dev_event_to_col(struct its_device *its_dev,
 238					       u32 event)
 239{
 240	struct its_node *its = its_dev->its;
 241
 242	return its->collections + its_dev->event_map.col_map[event];
 243}
 244
 245static struct its_vlpi_map *dev_event_to_vlpi_map(struct its_device *its_dev,
 246					       u32 event)
 247{
 248	if (WARN_ON_ONCE(event >= its_dev->event_map.nr_lpis))
 249		return NULL;
 250
 251	return &its_dev->event_map.vlpi_maps[event];
 252}
 253
 254static struct its_vlpi_map *get_vlpi_map(struct irq_data *d)
 255{
 256	if (irqd_is_forwarded_to_vcpu(d)) {
 257		struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 258		u32 event = its_get_event_id(d);
 259
 260		return dev_event_to_vlpi_map(its_dev, event);
 261	}
 262
 263	return NULL;
 264}
 265
 266static int vpe_to_cpuid_lock(struct its_vpe *vpe, unsigned long *flags)
 267{
 268	raw_spin_lock_irqsave(&vpe->vpe_lock, *flags);
 269	return vpe->col_idx;
 270}
 271
 272static void vpe_to_cpuid_unlock(struct its_vpe *vpe, unsigned long flags)
 273{
 274	raw_spin_unlock_irqrestore(&vpe->vpe_lock, flags);
 275}
 276
 277static struct irq_chip its_vpe_irq_chip;
 278
 279static int irq_to_cpuid_lock(struct irq_data *d, unsigned long *flags)
 280{
 281	struct its_vpe *vpe = NULL;
 282	int cpu;
 283
 284	if (d->chip == &its_vpe_irq_chip) {
 285		vpe = irq_data_get_irq_chip_data(d);
 286	} else {
 287		struct its_vlpi_map *map = get_vlpi_map(d);
 288		if (map)
 289			vpe = map->vpe;
 290	}
 291
 292	if (vpe) {
 293		cpu = vpe_to_cpuid_lock(vpe, flags);
 294	} else {
 295		/* Physical LPIs are already locked via the irq_desc lock */
 296		struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 297		cpu = its_dev->event_map.col_map[its_get_event_id(d)];
 298		/* Keep GCC quiet... */
 299		*flags = 0;
 300	}
 301
 302	return cpu;
 303}
 304
 305static void irq_to_cpuid_unlock(struct irq_data *d, unsigned long flags)
 306{
 307	struct its_vpe *vpe = NULL;
 308
 309	if (d->chip == &its_vpe_irq_chip) {
 310		vpe = irq_data_get_irq_chip_data(d);
 311	} else {
 312		struct its_vlpi_map *map = get_vlpi_map(d);
 313		if (map)
 314			vpe = map->vpe;
 315	}
 316
 317	if (vpe)
 318		vpe_to_cpuid_unlock(vpe, flags);
 319}
 320
 321static struct its_collection *valid_col(struct its_collection *col)
 322{
 323	if (WARN_ON_ONCE(col->target_address & GENMASK_ULL(15, 0)))
 324		return NULL;
 325
 326	return col;
 327}
 328
 329static struct its_vpe *valid_vpe(struct its_node *its, struct its_vpe *vpe)
 330{
 331	if (valid_col(its->collections + vpe->col_idx))
 332		return vpe;
 333
 334	return NULL;
 335}
 336
 337/*
 338 * ITS command descriptors - parameters to be encoded in a command
 339 * block.
 340 */
 341struct its_cmd_desc {
 342	union {
 343		struct {
 344			struct its_device *dev;
 345			u32 event_id;
 346		} its_inv_cmd;
 347
 348		struct {
 349			struct its_device *dev;
 350			u32 event_id;
 351		} its_clear_cmd;
 352
 353		struct {
 354			struct its_device *dev;
 355			u32 event_id;
 356		} its_int_cmd;
 357
 358		struct {
 359			struct its_device *dev;
 360			int valid;
 361		} its_mapd_cmd;
 362
 363		struct {
 364			struct its_collection *col;
 365			int valid;
 366		} its_mapc_cmd;
 367
 368		struct {
 369			struct its_device *dev;
 370			u32 phys_id;
 371			u32 event_id;
 372		} its_mapti_cmd;
 373
 374		struct {
 375			struct its_device *dev;
 376			struct its_collection *col;
 377			u32 event_id;
 378		} its_movi_cmd;
 379
 380		struct {
 381			struct its_device *dev;
 382			u32 event_id;
 383		} its_discard_cmd;
 384
 385		struct {
 386			struct its_collection *col;
 387		} its_invall_cmd;
 388
 389		struct {
 390			struct its_vpe *vpe;
 391		} its_vinvall_cmd;
 392
 393		struct {
 394			struct its_vpe *vpe;
 395			struct its_collection *col;
 396			bool valid;
 397		} its_vmapp_cmd;
 398
 399		struct {
 400			struct its_vpe *vpe;
 401			struct its_device *dev;
 402			u32 virt_id;
 403			u32 event_id;
 404			bool db_enabled;
 405		} its_vmapti_cmd;
 406
 407		struct {
 408			struct its_vpe *vpe;
 409			struct its_device *dev;
 410			u32 event_id;
 411			bool db_enabled;
 412		} its_vmovi_cmd;
 413
 414		struct {
 415			struct its_vpe *vpe;
 416			struct its_collection *col;
 417			u16 seq_num;
 418			u16 its_list;
 419		} its_vmovp_cmd;
 420
 421		struct {
 422			struct its_vpe *vpe;
 423		} its_invdb_cmd;
 424
 425		struct {
 426			struct its_vpe *vpe;
 427			u8 sgi;
 428			u8 priority;
 429			bool enable;
 430			bool group;
 431			bool clear;
 432		} its_vsgi_cmd;
 433	};
 434};
 435
 436/*
 437 * The ITS command block, which is what the ITS actually parses.
 438 */
 439struct its_cmd_block {
 440	union {
 441		u64	raw_cmd[4];
 442		__le64	raw_cmd_le[4];
 443	};
 444};
 445
 446#define ITS_CMD_QUEUE_SZ		SZ_64K
 447#define ITS_CMD_QUEUE_NR_ENTRIES	(ITS_CMD_QUEUE_SZ / sizeof(struct its_cmd_block))
 448
 449typedef struct its_collection *(*its_cmd_builder_t)(struct its_node *,
 450						    struct its_cmd_block *,
 451						    struct its_cmd_desc *);
 452
 453typedef struct its_vpe *(*its_cmd_vbuilder_t)(struct its_node *,
 454					      struct its_cmd_block *,
 455					      struct its_cmd_desc *);
 456
 457static void its_mask_encode(u64 *raw_cmd, u64 val, int h, int l)
 458{
 459	u64 mask = GENMASK_ULL(h, l);
 460	*raw_cmd &= ~mask;
 461	*raw_cmd |= (val << l) & mask;
 462}
 463
 464static void its_encode_cmd(struct its_cmd_block *cmd, u8 cmd_nr)
 465{
 466	its_mask_encode(&cmd->raw_cmd[0], cmd_nr, 7, 0);
 
 467}
 468
 469static void its_encode_devid(struct its_cmd_block *cmd, u32 devid)
 470{
 471	its_mask_encode(&cmd->raw_cmd[0], devid, 63, 32);
 
 472}
 473
 474static void its_encode_event_id(struct its_cmd_block *cmd, u32 id)
 475{
 476	its_mask_encode(&cmd->raw_cmd[1], id, 31, 0);
 
 477}
 478
 479static void its_encode_phys_id(struct its_cmd_block *cmd, u32 phys_id)
 480{
 481	its_mask_encode(&cmd->raw_cmd[1], phys_id, 63, 32);
 
 482}
 483
 484static void its_encode_size(struct its_cmd_block *cmd, u8 size)
 485{
 486	its_mask_encode(&cmd->raw_cmd[1], size, 4, 0);
 
 487}
 488
 489static void its_encode_itt(struct its_cmd_block *cmd, u64 itt_addr)
 490{
 491	its_mask_encode(&cmd->raw_cmd[2], itt_addr >> 8, 51, 8);
 
 492}
 493
 494static void its_encode_valid(struct its_cmd_block *cmd, int valid)
 495{
 496	its_mask_encode(&cmd->raw_cmd[2], !!valid, 63, 63);
 
 497}
 498
 499static void its_encode_target(struct its_cmd_block *cmd, u64 target_addr)
 500{
 501	its_mask_encode(&cmd->raw_cmd[2], target_addr >> 16, 51, 16);
 
 502}
 503
 504static void its_encode_collection(struct its_cmd_block *cmd, u16 col)
 505{
 506	its_mask_encode(&cmd->raw_cmd[2], col, 15, 0);
 507}
 508
 509static void its_encode_vpeid(struct its_cmd_block *cmd, u16 vpeid)
 510{
 511	its_mask_encode(&cmd->raw_cmd[1], vpeid, 47, 32);
 512}
 513
 514static void its_encode_virt_id(struct its_cmd_block *cmd, u32 virt_id)
 515{
 516	its_mask_encode(&cmd->raw_cmd[2], virt_id, 31, 0);
 517}
 518
 519static void its_encode_db_phys_id(struct its_cmd_block *cmd, u32 db_phys_id)
 520{
 521	its_mask_encode(&cmd->raw_cmd[2], db_phys_id, 63, 32);
 522}
 523
 524static void its_encode_db_valid(struct its_cmd_block *cmd, bool db_valid)
 525{
 526	its_mask_encode(&cmd->raw_cmd[2], db_valid, 0, 0);
 527}
 528
 529static void its_encode_seq_num(struct its_cmd_block *cmd, u16 seq_num)
 530{
 531	its_mask_encode(&cmd->raw_cmd[0], seq_num, 47, 32);
 532}
 533
 534static void its_encode_its_list(struct its_cmd_block *cmd, u16 its_list)
 535{
 536	its_mask_encode(&cmd->raw_cmd[1], its_list, 15, 0);
 537}
 538
 539static void its_encode_vpt_addr(struct its_cmd_block *cmd, u64 vpt_pa)
 540{
 541	its_mask_encode(&cmd->raw_cmd[3], vpt_pa >> 16, 51, 16);
 542}
 543
 544static void its_encode_vpt_size(struct its_cmd_block *cmd, u8 vpt_size)
 545{
 546	its_mask_encode(&cmd->raw_cmd[3], vpt_size, 4, 0);
 547}
 548
 549static void its_encode_vconf_addr(struct its_cmd_block *cmd, u64 vconf_pa)
 550{
 551	its_mask_encode(&cmd->raw_cmd[0], vconf_pa >> 16, 51, 16);
 552}
 553
 554static void its_encode_alloc(struct its_cmd_block *cmd, bool alloc)
 555{
 556	its_mask_encode(&cmd->raw_cmd[0], alloc, 8, 8);
 557}
 558
 559static void its_encode_ptz(struct its_cmd_block *cmd, bool ptz)
 560{
 561	its_mask_encode(&cmd->raw_cmd[0], ptz, 9, 9);
 562}
 563
 564static void its_encode_vmapp_default_db(struct its_cmd_block *cmd,
 565					u32 vpe_db_lpi)
 566{
 567	its_mask_encode(&cmd->raw_cmd[1], vpe_db_lpi, 31, 0);
 568}
 569
 570static void its_encode_vmovp_default_db(struct its_cmd_block *cmd,
 571					u32 vpe_db_lpi)
 572{
 573	its_mask_encode(&cmd->raw_cmd[3], vpe_db_lpi, 31, 0);
 574}
 575
 576static void its_encode_db(struct its_cmd_block *cmd, bool db)
 577{
 578	its_mask_encode(&cmd->raw_cmd[2], db, 63, 63);
 579}
 580
 581static void its_encode_sgi_intid(struct its_cmd_block *cmd, u8 sgi)
 582{
 583	its_mask_encode(&cmd->raw_cmd[0], sgi, 35, 32);
 584}
 585
 586static void its_encode_sgi_priority(struct its_cmd_block *cmd, u8 prio)
 587{
 588	its_mask_encode(&cmd->raw_cmd[0], prio >> 4, 23, 20);
 589}
 590
 591static void its_encode_sgi_group(struct its_cmd_block *cmd, bool grp)
 592{
 593	its_mask_encode(&cmd->raw_cmd[0], grp, 10, 10);
 594}
 595
 596static void its_encode_sgi_clear(struct its_cmd_block *cmd, bool clr)
 597{
 598	its_mask_encode(&cmd->raw_cmd[0], clr, 9, 9);
 599}
 600
 601static void its_encode_sgi_enable(struct its_cmd_block *cmd, bool en)
 602{
 603	its_mask_encode(&cmd->raw_cmd[0], en, 8, 8);
 604}
 605
 606static inline void its_fixup_cmd(struct its_cmd_block *cmd)
 607{
 608	/* Let's fixup BE commands */
 609	cmd->raw_cmd_le[0] = cpu_to_le64(cmd->raw_cmd[0]);
 610	cmd->raw_cmd_le[1] = cpu_to_le64(cmd->raw_cmd[1]);
 611	cmd->raw_cmd_le[2] = cpu_to_le64(cmd->raw_cmd[2]);
 612	cmd->raw_cmd_le[3] = cpu_to_le64(cmd->raw_cmd[3]);
 613}
 614
 615static struct its_collection *its_build_mapd_cmd(struct its_node *its,
 616						 struct its_cmd_block *cmd,
 617						 struct its_cmd_desc *desc)
 618{
 619	unsigned long itt_addr;
 620	u8 size = ilog2(desc->its_mapd_cmd.dev->nr_ites);
 621
 622	itt_addr = virt_to_phys(desc->its_mapd_cmd.dev->itt);
 623	itt_addr = ALIGN(itt_addr, ITS_ITT_ALIGN);
 624
 625	its_encode_cmd(cmd, GITS_CMD_MAPD);
 626	its_encode_devid(cmd, desc->its_mapd_cmd.dev->device_id);
 627	its_encode_size(cmd, size - 1);
 628	its_encode_itt(cmd, itt_addr);
 629	its_encode_valid(cmd, desc->its_mapd_cmd.valid);
 630
 631	its_fixup_cmd(cmd);
 632
 633	return NULL;
 634}
 635
 636static struct its_collection *its_build_mapc_cmd(struct its_node *its,
 637						 struct its_cmd_block *cmd,
 638						 struct its_cmd_desc *desc)
 639{
 640	its_encode_cmd(cmd, GITS_CMD_MAPC);
 641	its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
 642	its_encode_target(cmd, desc->its_mapc_cmd.col->target_address);
 643	its_encode_valid(cmd, desc->its_mapc_cmd.valid);
 644
 645	its_fixup_cmd(cmd);
 646
 647	return desc->its_mapc_cmd.col;
 648}
 649
 650static struct its_collection *its_build_mapti_cmd(struct its_node *its,
 651						  struct its_cmd_block *cmd,
 652						  struct its_cmd_desc *desc)
 653{
 654	struct its_collection *col;
 655
 656	col = dev_event_to_col(desc->its_mapti_cmd.dev,
 657			       desc->its_mapti_cmd.event_id);
 658
 659	its_encode_cmd(cmd, GITS_CMD_MAPTI);
 660	its_encode_devid(cmd, desc->its_mapti_cmd.dev->device_id);
 661	its_encode_event_id(cmd, desc->its_mapti_cmd.event_id);
 662	its_encode_phys_id(cmd, desc->its_mapti_cmd.phys_id);
 663	its_encode_collection(cmd, col->col_id);
 664
 665	its_fixup_cmd(cmd);
 666
 667	return valid_col(col);
 668}
 669
 670static struct its_collection *its_build_movi_cmd(struct its_node *its,
 671						 struct its_cmd_block *cmd,
 672						 struct its_cmd_desc *desc)
 673{
 674	struct its_collection *col;
 675
 676	col = dev_event_to_col(desc->its_movi_cmd.dev,
 677			       desc->its_movi_cmd.event_id);
 678
 679	its_encode_cmd(cmd, GITS_CMD_MOVI);
 680	its_encode_devid(cmd, desc->its_movi_cmd.dev->device_id);
 681	its_encode_event_id(cmd, desc->its_movi_cmd.event_id);
 682	its_encode_collection(cmd, desc->its_movi_cmd.col->col_id);
 683
 684	its_fixup_cmd(cmd);
 685
 686	return valid_col(col);
 687}
 688
 689static struct its_collection *its_build_discard_cmd(struct its_node *its,
 690						    struct its_cmd_block *cmd,
 691						    struct its_cmd_desc *desc)
 692{
 693	struct its_collection *col;
 694
 695	col = dev_event_to_col(desc->its_discard_cmd.dev,
 696			       desc->its_discard_cmd.event_id);
 697
 698	its_encode_cmd(cmd, GITS_CMD_DISCARD);
 699	its_encode_devid(cmd, desc->its_discard_cmd.dev->device_id);
 700	its_encode_event_id(cmd, desc->its_discard_cmd.event_id);
 701
 702	its_fixup_cmd(cmd);
 703
 704	return valid_col(col);
 705}
 706
 707static struct its_collection *its_build_inv_cmd(struct its_node *its,
 708						struct its_cmd_block *cmd,
 709						struct its_cmd_desc *desc)
 710{
 711	struct its_collection *col;
 712
 713	col = dev_event_to_col(desc->its_inv_cmd.dev,
 714			       desc->its_inv_cmd.event_id);
 715
 716	its_encode_cmd(cmd, GITS_CMD_INV);
 717	its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
 718	its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
 719
 720	its_fixup_cmd(cmd);
 721
 722	return valid_col(col);
 723}
 724
 725static struct its_collection *its_build_int_cmd(struct its_node *its,
 726						struct its_cmd_block *cmd,
 727						struct its_cmd_desc *desc)
 728{
 729	struct its_collection *col;
 730
 731	col = dev_event_to_col(desc->its_int_cmd.dev,
 732			       desc->its_int_cmd.event_id);
 733
 734	its_encode_cmd(cmd, GITS_CMD_INT);
 735	its_encode_devid(cmd, desc->its_int_cmd.dev->device_id);
 736	its_encode_event_id(cmd, desc->its_int_cmd.event_id);
 737
 738	its_fixup_cmd(cmd);
 739
 740	return valid_col(col);
 741}
 742
 743static struct its_collection *its_build_clear_cmd(struct its_node *its,
 744						  struct its_cmd_block *cmd,
 745						  struct its_cmd_desc *desc)
 746{
 747	struct its_collection *col;
 748
 749	col = dev_event_to_col(desc->its_clear_cmd.dev,
 750			       desc->its_clear_cmd.event_id);
 751
 752	its_encode_cmd(cmd, GITS_CMD_CLEAR);
 753	its_encode_devid(cmd, desc->its_clear_cmd.dev->device_id);
 754	its_encode_event_id(cmd, desc->its_clear_cmd.event_id);
 755
 756	its_fixup_cmd(cmd);
 757
 758	return valid_col(col);
 759}
 760
 761static struct its_collection *its_build_invall_cmd(struct its_node *its,
 762						   struct its_cmd_block *cmd,
 763						   struct its_cmd_desc *desc)
 764{
 765	its_encode_cmd(cmd, GITS_CMD_INVALL);
 766	its_encode_collection(cmd, desc->its_invall_cmd.col->col_id);
 767
 768	its_fixup_cmd(cmd);
 769
 770	return desc->its_invall_cmd.col;
 771}
 772
 773static struct its_vpe *its_build_vinvall_cmd(struct its_node *its,
 774					     struct its_cmd_block *cmd,
 775					     struct its_cmd_desc *desc)
 776{
 777	its_encode_cmd(cmd, GITS_CMD_VINVALL);
 778	its_encode_vpeid(cmd, desc->its_vinvall_cmd.vpe->vpe_id);
 779
 780	its_fixup_cmd(cmd);
 781
 782	return valid_vpe(its, desc->its_vinvall_cmd.vpe);
 783}
 784
 785static struct its_vpe *its_build_vmapp_cmd(struct its_node *its,
 786					   struct its_cmd_block *cmd,
 787					   struct its_cmd_desc *desc)
 788{
 789	struct its_vpe *vpe = valid_vpe(its, desc->its_vmapp_cmd.vpe);
 790	unsigned long vpt_addr, vconf_addr;
 791	u64 target;
 792	bool alloc;
 793
 794	its_encode_cmd(cmd, GITS_CMD_VMAPP);
 795	its_encode_vpeid(cmd, desc->its_vmapp_cmd.vpe->vpe_id);
 796	its_encode_valid(cmd, desc->its_vmapp_cmd.valid);
 797
 798	if (!desc->its_vmapp_cmd.valid) {
 799		if (is_v4_1(its)) {
 800			alloc = !atomic_dec_return(&desc->its_vmapp_cmd.vpe->vmapp_count);
 801			its_encode_alloc(cmd, alloc);
 802			/*
 803			 * Unmapping a VPE is self-synchronizing on GICv4.1,
 804			 * no need to issue a VSYNC.
 805			 */
 806			vpe = NULL;
 807		}
 808
 809		goto out;
 810	}
 811
 812	vpt_addr = virt_to_phys(page_address(desc->its_vmapp_cmd.vpe->vpt_page));
 813	target = desc->its_vmapp_cmd.col->target_address + its->vlpi_redist_offset;
 814
 815	its_encode_target(cmd, target);
 816	its_encode_vpt_addr(cmd, vpt_addr);
 817	its_encode_vpt_size(cmd, LPI_NRBITS - 1);
 818
 819	if (!is_v4_1(its))
 820		goto out;
 821
 822	vconf_addr = virt_to_phys(page_address(desc->its_vmapp_cmd.vpe->its_vm->vprop_page));
 823
 824	alloc = !atomic_fetch_inc(&desc->its_vmapp_cmd.vpe->vmapp_count);
 825
 826	its_encode_alloc(cmd, alloc);
 827
 828	/*
 829	 * GICv4.1 provides a way to get the VLPI state, which needs the vPE
 830	 * to be unmapped first, and in this case, we may remap the vPE
 831	 * back while the VPT is not empty. So we can't assume that the
 832	 * VPT is empty on map. This is why we never advertise PTZ.
 833	 */
 834	its_encode_ptz(cmd, false);
 835	its_encode_vconf_addr(cmd, vconf_addr);
 836	its_encode_vmapp_default_db(cmd, desc->its_vmapp_cmd.vpe->vpe_db_lpi);
 837
 838out:
 839	its_fixup_cmd(cmd);
 840
 841	return vpe;
 842}
 843
 844static struct its_vpe *its_build_vmapti_cmd(struct its_node *its,
 845					    struct its_cmd_block *cmd,
 846					    struct its_cmd_desc *desc)
 847{
 848	u32 db;
 849
 850	if (!is_v4_1(its) && desc->its_vmapti_cmd.db_enabled)
 851		db = desc->its_vmapti_cmd.vpe->vpe_db_lpi;
 852	else
 853		db = 1023;
 854
 855	its_encode_cmd(cmd, GITS_CMD_VMAPTI);
 856	its_encode_devid(cmd, desc->its_vmapti_cmd.dev->device_id);
 857	its_encode_vpeid(cmd, desc->its_vmapti_cmd.vpe->vpe_id);
 858	its_encode_event_id(cmd, desc->its_vmapti_cmd.event_id);
 859	its_encode_db_phys_id(cmd, db);
 860	its_encode_virt_id(cmd, desc->its_vmapti_cmd.virt_id);
 861
 862	its_fixup_cmd(cmd);
 863
 864	return valid_vpe(its, desc->its_vmapti_cmd.vpe);
 865}
 866
 867static struct its_vpe *its_build_vmovi_cmd(struct its_node *its,
 868					   struct its_cmd_block *cmd,
 869					   struct its_cmd_desc *desc)
 870{
 871	u32 db;
 872
 873	if (!is_v4_1(its) && desc->its_vmovi_cmd.db_enabled)
 874		db = desc->its_vmovi_cmd.vpe->vpe_db_lpi;
 875	else
 876		db = 1023;
 877
 878	its_encode_cmd(cmd, GITS_CMD_VMOVI);
 879	its_encode_devid(cmd, desc->its_vmovi_cmd.dev->device_id);
 880	its_encode_vpeid(cmd, desc->its_vmovi_cmd.vpe->vpe_id);
 881	its_encode_event_id(cmd, desc->its_vmovi_cmd.event_id);
 882	its_encode_db_phys_id(cmd, db);
 883	its_encode_db_valid(cmd, true);
 884
 885	its_fixup_cmd(cmd);
 886
 887	return valid_vpe(its, desc->its_vmovi_cmd.vpe);
 888}
 889
 890static struct its_vpe *its_build_vmovp_cmd(struct its_node *its,
 891					   struct its_cmd_block *cmd,
 892					   struct its_cmd_desc *desc)
 893{
 894	u64 target;
 895
 896	target = desc->its_vmovp_cmd.col->target_address + its->vlpi_redist_offset;
 897	its_encode_cmd(cmd, GITS_CMD_VMOVP);
 898	its_encode_seq_num(cmd, desc->its_vmovp_cmd.seq_num);
 899	its_encode_its_list(cmd, desc->its_vmovp_cmd.its_list);
 900	its_encode_vpeid(cmd, desc->its_vmovp_cmd.vpe->vpe_id);
 901	its_encode_target(cmd, target);
 902
 903	if (is_v4_1(its)) {
 904		its_encode_db(cmd, true);
 905		its_encode_vmovp_default_db(cmd, desc->its_vmovp_cmd.vpe->vpe_db_lpi);
 906	}
 907
 908	its_fixup_cmd(cmd);
 909
 910	return valid_vpe(its, desc->its_vmovp_cmd.vpe);
 911}
 912
 913static struct its_vpe *its_build_vinv_cmd(struct its_node *its,
 914					  struct its_cmd_block *cmd,
 915					  struct its_cmd_desc *desc)
 916{
 917	struct its_vlpi_map *map;
 918
 919	map = dev_event_to_vlpi_map(desc->its_inv_cmd.dev,
 920				    desc->its_inv_cmd.event_id);
 921
 922	its_encode_cmd(cmd, GITS_CMD_INV);
 923	its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
 924	its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
 925
 926	its_fixup_cmd(cmd);
 927
 928	return valid_vpe(its, map->vpe);
 929}
 930
 931static struct its_vpe *its_build_vint_cmd(struct its_node *its,
 932					  struct its_cmd_block *cmd,
 933					  struct its_cmd_desc *desc)
 934{
 935	struct its_vlpi_map *map;
 936
 937	map = dev_event_to_vlpi_map(desc->its_int_cmd.dev,
 938				    desc->its_int_cmd.event_id);
 939
 940	its_encode_cmd(cmd, GITS_CMD_INT);
 941	its_encode_devid(cmd, desc->its_int_cmd.dev->device_id);
 942	its_encode_event_id(cmd, desc->its_int_cmd.event_id);
 943
 944	its_fixup_cmd(cmd);
 945
 946	return valid_vpe(its, map->vpe);
 947}
 948
 949static struct its_vpe *its_build_vclear_cmd(struct its_node *its,
 950					    struct its_cmd_block *cmd,
 951					    struct its_cmd_desc *desc)
 952{
 953	struct its_vlpi_map *map;
 954
 955	map = dev_event_to_vlpi_map(desc->its_clear_cmd.dev,
 956				    desc->its_clear_cmd.event_id);
 957
 958	its_encode_cmd(cmd, GITS_CMD_CLEAR);
 959	its_encode_devid(cmd, desc->its_clear_cmd.dev->device_id);
 960	its_encode_event_id(cmd, desc->its_clear_cmd.event_id);
 961
 962	its_fixup_cmd(cmd);
 963
 964	return valid_vpe(its, map->vpe);
 965}
 966
 967static struct its_vpe *its_build_invdb_cmd(struct its_node *its,
 968					   struct its_cmd_block *cmd,
 969					   struct its_cmd_desc *desc)
 970{
 971	if (WARN_ON(!is_v4_1(its)))
 972		return NULL;
 973
 974	its_encode_cmd(cmd, GITS_CMD_INVDB);
 975	its_encode_vpeid(cmd, desc->its_invdb_cmd.vpe->vpe_id);
 976
 977	its_fixup_cmd(cmd);
 978
 979	return valid_vpe(its, desc->its_invdb_cmd.vpe);
 980}
 981
 982static struct its_vpe *its_build_vsgi_cmd(struct its_node *its,
 983					  struct its_cmd_block *cmd,
 984					  struct its_cmd_desc *desc)
 985{
 986	if (WARN_ON(!is_v4_1(its)))
 987		return NULL;
 988
 989	its_encode_cmd(cmd, GITS_CMD_VSGI);
 990	its_encode_vpeid(cmd, desc->its_vsgi_cmd.vpe->vpe_id);
 991	its_encode_sgi_intid(cmd, desc->its_vsgi_cmd.sgi);
 992	its_encode_sgi_priority(cmd, desc->its_vsgi_cmd.priority);
 993	its_encode_sgi_group(cmd, desc->its_vsgi_cmd.group);
 994	its_encode_sgi_clear(cmd, desc->its_vsgi_cmd.clear);
 995	its_encode_sgi_enable(cmd, desc->its_vsgi_cmd.enable);
 996
 997	its_fixup_cmd(cmd);
 998
 999	return valid_vpe(its, desc->its_vsgi_cmd.vpe);
1000}
1001
1002static u64 its_cmd_ptr_to_offset(struct its_node *its,
1003				 struct its_cmd_block *ptr)
1004{
1005	return (ptr - its->cmd_base) * sizeof(*ptr);
1006}
1007
1008static int its_queue_full(struct its_node *its)
1009{
1010	int widx;
1011	int ridx;
1012
1013	widx = its->cmd_write - its->cmd_base;
1014	ridx = readl_relaxed(its->base + GITS_CREADR) / sizeof(struct its_cmd_block);
1015
1016	/* This is incredibly unlikely to happen, unless the ITS locks up. */
1017	if (((widx + 1) % ITS_CMD_QUEUE_NR_ENTRIES) == ridx)
1018		return 1;
1019
1020	return 0;
1021}
1022
1023static struct its_cmd_block *its_allocate_entry(struct its_node *its)
1024{
1025	struct its_cmd_block *cmd;
1026	u32 count = 1000000;	/* 1s! */
1027
1028	while (its_queue_full(its)) {
1029		count--;
1030		if (!count) {
1031			pr_err_ratelimited("ITS queue not draining\n");
1032			return NULL;
1033		}
1034		cpu_relax();
1035		udelay(1);
1036	}
1037
1038	cmd = its->cmd_write++;
1039
1040	/* Handle queue wrapping */
1041	if (its->cmd_write == (its->cmd_base + ITS_CMD_QUEUE_NR_ENTRIES))
1042		its->cmd_write = its->cmd_base;
1043
1044	/* Clear command  */
1045	cmd->raw_cmd[0] = 0;
1046	cmd->raw_cmd[1] = 0;
1047	cmd->raw_cmd[2] = 0;
1048	cmd->raw_cmd[3] = 0;
1049
1050	return cmd;
1051}
1052
1053static struct its_cmd_block *its_post_commands(struct its_node *its)
1054{
1055	u64 wr = its_cmd_ptr_to_offset(its, its->cmd_write);
1056
1057	writel_relaxed(wr, its->base + GITS_CWRITER);
1058
1059	return its->cmd_write;
1060}
1061
1062static void its_flush_cmd(struct its_node *its, struct its_cmd_block *cmd)
1063{
1064	/*
1065	 * Make sure the commands written to memory are observable by
1066	 * the ITS.
1067	 */
1068	if (its->flags & ITS_FLAGS_CMDQ_NEEDS_FLUSHING)
1069		gic_flush_dcache_to_poc(cmd, sizeof(*cmd));
1070	else
1071		dsb(ishst);
1072}
1073
1074static int its_wait_for_range_completion(struct its_node *its,
1075					 u64	prev_idx,
1076					 struct its_cmd_block *to)
1077{
1078	u64 rd_idx, to_idx, linear_idx;
1079	u32 count = 1000000;	/* 1s! */
1080
1081	/* Linearize to_idx if the command set has wrapped around */
1082	to_idx = its_cmd_ptr_to_offset(its, to);
1083	if (to_idx < prev_idx)
1084		to_idx += ITS_CMD_QUEUE_SZ;
1085
1086	linear_idx = prev_idx;
1087
1088	while (1) {
1089		s64 delta;
1090
1091		rd_idx = readl_relaxed(its->base + GITS_CREADR);
1092
1093		/*
1094		 * Compute the read pointer progress, taking the
1095		 * potential wrap-around into account.
1096		 */
1097		delta = rd_idx - prev_idx;
1098		if (rd_idx < prev_idx)
1099			delta += ITS_CMD_QUEUE_SZ;
1100
1101		linear_idx += delta;
1102		if (linear_idx >= to_idx)
1103			break;
1104
1105		count--;
1106		if (!count) {
1107			pr_err_ratelimited("ITS queue timeout (%llu %llu)\n",
1108					   to_idx, linear_idx);
1109			return -1;
1110		}
1111		prev_idx = rd_idx;
1112		cpu_relax();
1113		udelay(1);
1114	}
1115
1116	return 0;
1117}
1118
1119/* Warning, macro hell follows */
1120#define BUILD_SINGLE_CMD_FUNC(name, buildtype, synctype, buildfn)	\
1121void name(struct its_node *its,						\
1122	  buildtype builder,						\
1123	  struct its_cmd_desc *desc)					\
1124{									\
1125	struct its_cmd_block *cmd, *sync_cmd, *next_cmd;		\
1126	synctype *sync_obj;						\
1127	unsigned long flags;						\
1128	u64 rd_idx;							\
1129									\
1130	raw_spin_lock_irqsave(&its->lock, flags);			\
1131									\
1132	cmd = its_allocate_entry(its);					\
1133	if (!cmd) {		/* We're soooooo screewed... */		\
1134		raw_spin_unlock_irqrestore(&its->lock, flags);		\
1135		return;							\
1136	}								\
1137	sync_obj = builder(its, cmd, desc);				\
1138	its_flush_cmd(its, cmd);					\
1139									\
1140	if (sync_obj) {							\
1141		sync_cmd = its_allocate_entry(its);			\
1142		if (!sync_cmd)						\
1143			goto post;					\
1144									\
1145		buildfn(its, sync_cmd, sync_obj);			\
1146		its_flush_cmd(its, sync_cmd);				\
1147	}								\
1148									\
1149post:									\
1150	rd_idx = readl_relaxed(its->base + GITS_CREADR);		\
1151	next_cmd = its_post_commands(its);				\
1152	raw_spin_unlock_irqrestore(&its->lock, flags);			\
1153									\
1154	if (its_wait_for_range_completion(its, rd_idx, next_cmd))	\
1155		pr_err_ratelimited("ITS cmd %ps failed\n", builder);	\
1156}
1157
1158static void its_build_sync_cmd(struct its_node *its,
1159			       struct its_cmd_block *sync_cmd,
1160			       struct its_collection *sync_col)
1161{
1162	its_encode_cmd(sync_cmd, GITS_CMD_SYNC);
1163	its_encode_target(sync_cmd, sync_col->target_address);
1164
1165	its_fixup_cmd(sync_cmd);
1166}
1167
1168static BUILD_SINGLE_CMD_FUNC(its_send_single_command, its_cmd_builder_t,
1169			     struct its_collection, its_build_sync_cmd)
1170
1171static void its_build_vsync_cmd(struct its_node *its,
1172				struct its_cmd_block *sync_cmd,
1173				struct its_vpe *sync_vpe)
1174{
1175	its_encode_cmd(sync_cmd, GITS_CMD_VSYNC);
1176	its_encode_vpeid(sync_cmd, sync_vpe->vpe_id);
1177
1178	its_fixup_cmd(sync_cmd);
1179}
1180
1181static BUILD_SINGLE_CMD_FUNC(its_send_single_vcommand, its_cmd_vbuilder_t,
1182			     struct its_vpe, its_build_vsync_cmd)
1183
1184static void its_send_int(struct its_device *dev, u32 event_id)
1185{
1186	struct its_cmd_desc desc;
 
 
1187
1188	desc.its_int_cmd.dev = dev;
1189	desc.its_int_cmd.event_id = event_id;
1190
1191	its_send_single_command(dev->its, its_build_int_cmd, &desc);
1192}
 
 
 
 
 
 
1193
1194static void its_send_clear(struct its_device *dev, u32 event_id)
1195{
1196	struct its_cmd_desc desc;
 
 
 
 
 
 
 
 
1197
1198	desc.its_clear_cmd.dev = dev;
1199	desc.its_clear_cmd.event_id = event_id;
 
1200
1201	its_send_single_command(dev->its, its_build_clear_cmd, &desc);
1202}
1203
1204static void its_send_inv(struct its_device *dev, u32 event_id)
1205{
1206	struct its_cmd_desc desc;
1207
1208	desc.its_inv_cmd.dev = dev;
1209	desc.its_inv_cmd.event_id = event_id;
1210
1211	its_send_single_command(dev->its, its_build_inv_cmd, &desc);
1212}
1213
1214static void its_send_mapd(struct its_device *dev, int valid)
1215{
1216	struct its_cmd_desc desc;
1217
1218	desc.its_mapd_cmd.dev = dev;
1219	desc.its_mapd_cmd.valid = !!valid;
1220
1221	its_send_single_command(dev->its, its_build_mapd_cmd, &desc);
1222}
1223
1224static void its_send_mapc(struct its_node *its, struct its_collection *col,
1225			  int valid)
1226{
1227	struct its_cmd_desc desc;
1228
1229	desc.its_mapc_cmd.col = col;
1230	desc.its_mapc_cmd.valid = !!valid;
1231
1232	its_send_single_command(its, its_build_mapc_cmd, &desc);
1233}
1234
1235static void its_send_mapti(struct its_device *dev, u32 irq_id, u32 id)
1236{
1237	struct its_cmd_desc desc;
1238
1239	desc.its_mapti_cmd.dev = dev;
1240	desc.its_mapti_cmd.phys_id = irq_id;
1241	desc.its_mapti_cmd.event_id = id;
1242
1243	its_send_single_command(dev->its, its_build_mapti_cmd, &desc);
1244}
1245
1246static void its_send_movi(struct its_device *dev,
1247			  struct its_collection *col, u32 id)
1248{
1249	struct its_cmd_desc desc;
1250
1251	desc.its_movi_cmd.dev = dev;
1252	desc.its_movi_cmd.col = col;
1253	desc.its_movi_cmd.event_id = id;
1254
1255	its_send_single_command(dev->its, its_build_movi_cmd, &desc);
1256}
1257
1258static void its_send_discard(struct its_device *dev, u32 id)
1259{
1260	struct its_cmd_desc desc;
1261
1262	desc.its_discard_cmd.dev = dev;
1263	desc.its_discard_cmd.event_id = id;
1264
1265	its_send_single_command(dev->its, its_build_discard_cmd, &desc);
1266}
1267
1268static void its_send_invall(struct its_node *its, struct its_collection *col)
1269{
1270	struct its_cmd_desc desc;
1271
1272	desc.its_invall_cmd.col = col;
1273
1274	its_send_single_command(its, its_build_invall_cmd, &desc);
1275}
1276
1277static void its_send_vmapti(struct its_device *dev, u32 id)
1278{
1279	struct its_vlpi_map *map = dev_event_to_vlpi_map(dev, id);
1280	struct its_cmd_desc desc;
1281
1282	desc.its_vmapti_cmd.vpe = map->vpe;
1283	desc.its_vmapti_cmd.dev = dev;
1284	desc.its_vmapti_cmd.virt_id = map->vintid;
1285	desc.its_vmapti_cmd.event_id = id;
1286	desc.its_vmapti_cmd.db_enabled = map->db_enabled;
1287
1288	its_send_single_vcommand(dev->its, its_build_vmapti_cmd, &desc);
1289}
1290
1291static void its_send_vmovi(struct its_device *dev, u32 id)
1292{
1293	struct its_vlpi_map *map = dev_event_to_vlpi_map(dev, id);
1294	struct its_cmd_desc desc;
1295
1296	desc.its_vmovi_cmd.vpe = map->vpe;
1297	desc.its_vmovi_cmd.dev = dev;
1298	desc.its_vmovi_cmd.event_id = id;
1299	desc.its_vmovi_cmd.db_enabled = map->db_enabled;
1300
1301	its_send_single_vcommand(dev->its, its_build_vmovi_cmd, &desc);
1302}
1303
1304static void its_send_vmapp(struct its_node *its,
1305			   struct its_vpe *vpe, bool valid)
1306{
1307	struct its_cmd_desc desc;
 
 
 
1308
1309	desc.its_vmapp_cmd.vpe = vpe;
1310	desc.its_vmapp_cmd.valid = valid;
1311	desc.its_vmapp_cmd.col = &its->collections[vpe->col_idx];
1312
1313	its_send_single_vcommand(its, its_build_vmapp_cmd, &desc);
1314}
1315
1316static void its_send_vmovp(struct its_vpe *vpe)
1317{
1318	struct its_cmd_desc desc = {};
1319	struct its_node *its;
1320	unsigned long flags;
1321	int col_id = vpe->col_idx;
1322
1323	desc.its_vmovp_cmd.vpe = vpe;
1324
1325	if (!its_list_map) {
1326		its = list_first_entry(&its_nodes, struct its_node, entry);
1327		desc.its_vmovp_cmd.col = &its->collections[col_id];
1328		its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
1329		return;
1330	}
1331
1332	/*
1333	 * Yet another marvel of the architecture. If using the
1334	 * its_list "feature", we need to make sure that all ITSs
1335	 * receive all VMOVP commands in the same order. The only way
1336	 * to guarantee this is to make vmovp a serialization point.
1337	 *
1338	 * Wall <-- Head.
1339	 */
1340	raw_spin_lock_irqsave(&vmovp_lock, flags);
1341
1342	desc.its_vmovp_cmd.seq_num = vmovp_seq_num++;
1343	desc.its_vmovp_cmd.its_list = get_its_list(vpe->its_vm);
1344
1345	/* Emit VMOVPs */
1346	list_for_each_entry(its, &its_nodes, entry) {
1347		if (!is_v4(its))
1348			continue;
1349
1350		if (!require_its_list_vmovp(vpe->its_vm, its))
1351			continue;
1352
1353		desc.its_vmovp_cmd.col = &its->collections[col_id];
1354		its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
1355	}
1356
1357	raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1358}
1359
1360static void its_send_vinvall(struct its_node *its, struct its_vpe *vpe)
1361{
1362	struct its_cmd_desc desc;
1363
1364	desc.its_vinvall_cmd.vpe = vpe;
1365	its_send_single_vcommand(its, its_build_vinvall_cmd, &desc);
1366}
1367
1368static void its_send_vinv(struct its_device *dev, u32 event_id)
1369{
1370	struct its_cmd_desc desc;
1371
1372	/*
1373	 * There is no real VINV command. This is just a normal INV,
1374	 * with a VSYNC instead of a SYNC.
1375	 */
1376	desc.its_inv_cmd.dev = dev;
1377	desc.its_inv_cmd.event_id = event_id;
1378
1379	its_send_single_vcommand(dev->its, its_build_vinv_cmd, &desc);
1380}
1381
1382static void its_send_vint(struct its_device *dev, u32 event_id)
1383{
1384	struct its_cmd_desc desc;
1385
1386	/*
1387	 * There is no real VINT command. This is just a normal INT,
1388	 * with a VSYNC instead of a SYNC.
1389	 */
1390	desc.its_int_cmd.dev = dev;
1391	desc.its_int_cmd.event_id = event_id;
1392
1393	its_send_single_vcommand(dev->its, its_build_vint_cmd, &desc);
1394}
1395
1396static void its_send_vclear(struct its_device *dev, u32 event_id)
1397{
1398	struct its_cmd_desc desc;
1399
1400	/*
1401	 * There is no real VCLEAR command. This is just a normal CLEAR,
1402	 * with a VSYNC instead of a SYNC.
1403	 */
1404	desc.its_clear_cmd.dev = dev;
1405	desc.its_clear_cmd.event_id = event_id;
1406
1407	its_send_single_vcommand(dev->its, its_build_vclear_cmd, &desc);
1408}
1409
1410static void its_send_invdb(struct its_node *its, struct its_vpe *vpe)
1411{
1412	struct its_cmd_desc desc;
1413
1414	desc.its_invdb_cmd.vpe = vpe;
1415	its_send_single_vcommand(its, its_build_invdb_cmd, &desc);
1416}
1417
1418/*
1419 * irqchip functions - assumes MSI, mostly.
1420 */
1421static void lpi_write_config(struct irq_data *d, u8 clr, u8 set)
1422{
1423	struct its_vlpi_map *map = get_vlpi_map(d);
1424	irq_hw_number_t hwirq;
1425	void *va;
1426	u8 *cfg;
1427
1428	if (map) {
1429		va = page_address(map->vm->vprop_page);
1430		hwirq = map->vintid;
1431
1432		/* Remember the updated property */
1433		map->properties &= ~clr;
1434		map->properties |= set | LPI_PROP_GROUP1;
1435	} else {
1436		va = gic_rdists->prop_table_va;
1437		hwirq = d->hwirq;
1438	}
1439
1440	cfg = va + hwirq - 8192;
1441	*cfg &= ~clr;
1442	*cfg |= set | LPI_PROP_GROUP1;
1443
1444	/*
1445	 * Make the above write visible to the redistributors.
1446	 * And yes, we're flushing exactly: One. Single. Byte.
1447	 * Humpf...
1448	 */
1449	if (gic_rdists->flags & RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING)
1450		gic_flush_dcache_to_poc(cfg, sizeof(*cfg));
1451	else
1452		dsb(ishst);
1453}
1454
1455static void wait_for_syncr(void __iomem *rdbase)
1456{
1457	while (readl_relaxed(rdbase + GICR_SYNCR) & 1)
1458		cpu_relax();
1459}
1460
1461static void __direct_lpi_inv(struct irq_data *d, u64 val)
1462{
1463	void __iomem *rdbase;
1464	unsigned long flags;
1465	int cpu;
1466
1467	/* Target the redistributor this LPI is currently routed to */
1468	cpu = irq_to_cpuid_lock(d, &flags);
1469	raw_spin_lock(&gic_data_rdist_cpu(cpu)->rd_lock);
1470
1471	rdbase = per_cpu_ptr(gic_rdists->rdist, cpu)->rd_base;
1472	gic_write_lpir(val, rdbase + GICR_INVLPIR);
1473	wait_for_syncr(rdbase);
1474
1475	raw_spin_unlock(&gic_data_rdist_cpu(cpu)->rd_lock);
1476	irq_to_cpuid_unlock(d, flags);
1477}
1478
1479static void direct_lpi_inv(struct irq_data *d)
1480{
1481	struct its_vlpi_map *map = get_vlpi_map(d);
1482	u64 val;
1483
1484	if (map) {
1485		struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1486
1487		WARN_ON(!is_v4_1(its_dev->its));
1488
1489		val  = GICR_INVLPIR_V;
1490		val |= FIELD_PREP(GICR_INVLPIR_VPEID, map->vpe->vpe_id);
1491		val |= FIELD_PREP(GICR_INVLPIR_INTID, map->vintid);
1492	} else {
1493		val = d->hwirq;
1494	}
1495
1496	__direct_lpi_inv(d, val);
1497}
1498
1499static void lpi_update_config(struct irq_data *d, u8 clr, u8 set)
1500{
1501	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1502
1503	lpi_write_config(d, clr, set);
1504	if (gic_rdists->has_direct_lpi &&
1505	    (is_v4_1(its_dev->its) || !irqd_is_forwarded_to_vcpu(d)))
1506		direct_lpi_inv(d);
1507	else if (!irqd_is_forwarded_to_vcpu(d))
1508		its_send_inv(its_dev, its_get_event_id(d));
1509	else
1510		its_send_vinv(its_dev, its_get_event_id(d));
1511}
1512
1513static void its_vlpi_set_doorbell(struct irq_data *d, bool enable)
1514{
1515	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1516	u32 event = its_get_event_id(d);
1517	struct its_vlpi_map *map;
1518
1519	/*
1520	 * GICv4.1 does away with the per-LPI nonsense, nothing to do
1521	 * here.
1522	 */
1523	if (is_v4_1(its_dev->its))
1524		return;
1525
1526	map = dev_event_to_vlpi_map(its_dev, event);
1527
1528	if (map->db_enabled == enable)
1529		return;
1530
1531	map->db_enabled = enable;
1532
1533	/*
1534	 * More fun with the architecture:
1535	 *
1536	 * Ideally, we'd issue a VMAPTI to set the doorbell to its LPI
1537	 * value or to 1023, depending on the enable bit. But that
1538	 * would be issuing a mapping for an /existing/ DevID+EventID
1539	 * pair, which is UNPREDICTABLE. Instead, let's issue a VMOVI
1540	 * to the /same/ vPE, using this opportunity to adjust the
1541	 * doorbell. Mouahahahaha. We loves it, Precious.
1542	 */
1543	its_send_vmovi(its_dev, event);
1544}
1545
1546static void its_mask_irq(struct irq_data *d)
1547{
1548	if (irqd_is_forwarded_to_vcpu(d))
1549		its_vlpi_set_doorbell(d, false);
1550
1551	lpi_update_config(d, LPI_PROP_ENABLED, 0);
1552}
1553
1554static void its_unmask_irq(struct irq_data *d)
1555{
1556	if (irqd_is_forwarded_to_vcpu(d))
1557		its_vlpi_set_doorbell(d, true);
1558
1559	lpi_update_config(d, 0, LPI_PROP_ENABLED);
1560}
1561
1562static __maybe_unused u32 its_read_lpi_count(struct irq_data *d, int cpu)
1563{
1564	if (irqd_affinity_is_managed(d))
1565		return atomic_read(&per_cpu_ptr(&cpu_lpi_count, cpu)->managed);
1566
1567	return atomic_read(&per_cpu_ptr(&cpu_lpi_count, cpu)->unmanaged);
1568}
1569
1570static void its_inc_lpi_count(struct irq_data *d, int cpu)
1571{
1572	if (irqd_affinity_is_managed(d))
1573		atomic_inc(&per_cpu_ptr(&cpu_lpi_count, cpu)->managed);
1574	else
1575		atomic_inc(&per_cpu_ptr(&cpu_lpi_count, cpu)->unmanaged);
1576}
1577
1578static void its_dec_lpi_count(struct irq_data *d, int cpu)
1579{
1580	if (irqd_affinity_is_managed(d))
1581		atomic_dec(&per_cpu_ptr(&cpu_lpi_count, cpu)->managed);
1582	else
1583		atomic_dec(&per_cpu_ptr(&cpu_lpi_count, cpu)->unmanaged);
1584}
1585
1586static unsigned int cpumask_pick_least_loaded(struct irq_data *d,
1587					      const struct cpumask *cpu_mask)
1588{
1589	unsigned int cpu = nr_cpu_ids, tmp;
1590	int count = S32_MAX;
1591
1592	for_each_cpu(tmp, cpu_mask) {
1593		int this_count = its_read_lpi_count(d, tmp);
1594		if (this_count < count) {
1595			cpu = tmp;
1596		        count = this_count;
1597		}
1598	}
1599
1600	return cpu;
1601}
1602
1603/*
1604 * As suggested by Thomas Gleixner in:
1605 * https://lore.kernel.org/r/87h80q2aoc.fsf@nanos.tec.linutronix.de
1606 */
1607static int its_select_cpu(struct irq_data *d,
1608			  const struct cpumask *aff_mask)
1609{
1610	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1611	static DEFINE_RAW_SPINLOCK(tmpmask_lock);
1612	static struct cpumask __tmpmask;
1613	struct cpumask *tmpmask;
1614	unsigned long flags;
1615	int cpu, node;
1616	node = its_dev->its->numa_node;
1617	tmpmask = &__tmpmask;
1618
1619	raw_spin_lock_irqsave(&tmpmask_lock, flags);
1620
1621	if (!irqd_affinity_is_managed(d)) {
1622		/* First try the NUMA node */
1623		if (node != NUMA_NO_NODE) {
1624			/*
1625			 * Try the intersection of the affinity mask and the
1626			 * node mask (and the online mask, just to be safe).
1627			 */
1628			cpumask_and(tmpmask, cpumask_of_node(node), aff_mask);
1629			cpumask_and(tmpmask, tmpmask, cpu_online_mask);
1630
1631			/*
1632			 * Ideally, we would check if the mask is empty, and
1633			 * try again on the full node here.
1634			 *
1635			 * But it turns out that the way ACPI describes the
1636			 * affinity for ITSs only deals about memory, and
1637			 * not target CPUs, so it cannot describe a single
1638			 * ITS placed next to two NUMA nodes.
1639			 *
1640			 * Instead, just fallback on the online mask. This
1641			 * diverges from Thomas' suggestion above.
1642			 */
1643			cpu = cpumask_pick_least_loaded(d, tmpmask);
1644			if (cpu < nr_cpu_ids)
1645				goto out;
1646
1647			/* If we can't cross sockets, give up */
1648			if ((its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144))
1649				goto out;
1650
1651			/* If the above failed, expand the search */
1652		}
1653
1654		/* Try the intersection of the affinity and online masks */
1655		cpumask_and(tmpmask, aff_mask, cpu_online_mask);
1656
1657		/* If that doesn't fly, the online mask is the last resort */
1658		if (cpumask_empty(tmpmask))
1659			cpumask_copy(tmpmask, cpu_online_mask);
1660
1661		cpu = cpumask_pick_least_loaded(d, tmpmask);
1662	} else {
1663		cpumask_copy(tmpmask, aff_mask);
1664
1665		/* If we cannot cross sockets, limit the search to that node */
1666		if ((its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) &&
1667		    node != NUMA_NO_NODE)
1668			cpumask_and(tmpmask, tmpmask, cpumask_of_node(node));
1669
1670		cpu = cpumask_pick_least_loaded(d, tmpmask);
1671	}
1672out:
1673	raw_spin_unlock_irqrestore(&tmpmask_lock, flags);
1674
1675	pr_debug("IRQ%d -> %*pbl CPU%d\n", d->irq, cpumask_pr_args(aff_mask), cpu);
1676	return cpu;
1677}
1678
1679static int its_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
1680			    bool force)
1681{
 
1682	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1683	struct its_collection *target_col;
1684	u32 id = its_get_event_id(d);
1685	int cpu, prev_cpu;
1686
1687	/* A forwarded interrupt should use irq_set_vcpu_affinity */
1688	if (irqd_is_forwarded_to_vcpu(d))
1689		return -EINVAL;
1690
1691	prev_cpu = its_dev->event_map.col_map[id];
1692	its_dec_lpi_count(d, prev_cpu);
1693
1694	if (!force)
1695		cpu = its_select_cpu(d, mask_val);
1696	else
1697		cpu = cpumask_pick_least_loaded(d, mask_val);
1698
1699	if (cpu < 0 || cpu >= nr_cpu_ids)
1700		goto err;
1701
1702	/* don't set the affinity when the target cpu is same as current one */
1703	if (cpu != prev_cpu) {
1704		target_col = &its_dev->its->collections[cpu];
1705		its_send_movi(its_dev, target_col, id);
1706		its_dev->event_map.col_map[id] = cpu;
1707		irq_data_update_effective_affinity(d, cpumask_of(cpu));
1708	}
1709
1710	its_inc_lpi_count(d, cpu);
1711
1712	return IRQ_SET_MASK_OK_DONE;
1713
1714err:
1715	its_inc_lpi_count(d, prev_cpu);
1716	return -EINVAL;
1717}
1718
1719static u64 its_irq_get_msi_base(struct its_device *its_dev)
1720{
1721	struct its_node *its = its_dev->its;
1722
1723	return its->phys_base + GITS_TRANSLATER;
1724}
1725
1726static void its_irq_compose_msi_msg(struct irq_data *d, struct msi_msg *msg)
1727{
1728	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1729	struct its_node *its;
1730	u64 addr;
1731
1732	its = its_dev->its;
1733	addr = its->get_msi_base(its_dev);
1734
1735	msg->address_lo		= lower_32_bits(addr);
1736	msg->address_hi		= upper_32_bits(addr);
1737	msg->data		= its_get_event_id(d);
1738
1739	iommu_dma_compose_msi_msg(irq_data_get_msi_desc(d), msg);
1740}
1741
1742static int its_irq_set_irqchip_state(struct irq_data *d,
1743				     enum irqchip_irq_state which,
1744				     bool state)
1745{
1746	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1747	u32 event = its_get_event_id(d);
1748
1749	if (which != IRQCHIP_STATE_PENDING)
1750		return -EINVAL;
1751
1752	if (irqd_is_forwarded_to_vcpu(d)) {
1753		if (state)
1754			its_send_vint(its_dev, event);
1755		else
1756			its_send_vclear(its_dev, event);
1757	} else {
1758		if (state)
1759			its_send_int(its_dev, event);
1760		else
1761			its_send_clear(its_dev, event);
1762	}
1763
1764	return 0;
1765}
1766
1767static int its_irq_retrigger(struct irq_data *d)
1768{
1769	return !its_irq_set_irqchip_state(d, IRQCHIP_STATE_PENDING, true);
1770}
1771
1772/*
1773 * Two favourable cases:
1774 *
1775 * (a) Either we have a GICv4.1, and all vPEs have to be mapped at all times
1776 *     for vSGI delivery
1777 *
1778 * (b) Or the ITSs do not use a list map, meaning that VMOVP is cheap enough
1779 *     and we're better off mapping all VPEs always
1780 *
1781 * If neither (a) nor (b) is true, then we map vPEs on demand.
1782 *
1783 */
1784static bool gic_requires_eager_mapping(void)
1785{
1786	if (!its_list_map || gic_rdists->has_rvpeid)
1787		return true;
1788
1789	return false;
1790}
1791
1792static void its_map_vm(struct its_node *its, struct its_vm *vm)
1793{
1794	unsigned long flags;
1795
1796	if (gic_requires_eager_mapping())
1797		return;
1798
1799	raw_spin_lock_irqsave(&vmovp_lock, flags);
1800
1801	/*
1802	 * If the VM wasn't mapped yet, iterate over the vpes and get
1803	 * them mapped now.
1804	 */
1805	vm->vlpi_count[its->list_nr]++;
1806
1807	if (vm->vlpi_count[its->list_nr] == 1) {
1808		int i;
1809
1810		for (i = 0; i < vm->nr_vpes; i++) {
1811			struct its_vpe *vpe = vm->vpes[i];
1812			struct irq_data *d = irq_get_irq_data(vpe->irq);
1813
1814			/* Map the VPE to the first possible CPU */
1815			vpe->col_idx = cpumask_first(cpu_online_mask);
1816			its_send_vmapp(its, vpe, true);
1817			its_send_vinvall(its, vpe);
1818			irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
1819		}
1820	}
1821
1822	raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1823}
1824
1825static void its_unmap_vm(struct its_node *its, struct its_vm *vm)
1826{
1827	unsigned long flags;
1828
1829	/* Not using the ITS list? Everything is always mapped. */
1830	if (gic_requires_eager_mapping())
1831		return;
1832
1833	raw_spin_lock_irqsave(&vmovp_lock, flags);
1834
1835	if (!--vm->vlpi_count[its->list_nr]) {
1836		int i;
1837
1838		for (i = 0; i < vm->nr_vpes; i++)
1839			its_send_vmapp(its, vm->vpes[i], false);
1840	}
1841
1842	raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1843}
1844
1845static int its_vlpi_map(struct irq_data *d, struct its_cmd_info *info)
1846{
1847	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1848	u32 event = its_get_event_id(d);
1849	int ret = 0;
1850
1851	if (!info->map)
1852		return -EINVAL;
1853
1854	raw_spin_lock(&its_dev->event_map.vlpi_lock);
1855
1856	if (!its_dev->event_map.vm) {
1857		struct its_vlpi_map *maps;
1858
1859		maps = kcalloc(its_dev->event_map.nr_lpis, sizeof(*maps),
1860			       GFP_ATOMIC);
1861		if (!maps) {
1862			ret = -ENOMEM;
1863			goto out;
1864		}
1865
1866		its_dev->event_map.vm = info->map->vm;
1867		its_dev->event_map.vlpi_maps = maps;
1868	} else if (its_dev->event_map.vm != info->map->vm) {
1869		ret = -EINVAL;
1870		goto out;
1871	}
1872
1873	/* Get our private copy of the mapping information */
1874	its_dev->event_map.vlpi_maps[event] = *info->map;
1875
1876	if (irqd_is_forwarded_to_vcpu(d)) {
1877		/* Already mapped, move it around */
1878		its_send_vmovi(its_dev, event);
1879	} else {
1880		/* Ensure all the VPEs are mapped on this ITS */
1881		its_map_vm(its_dev->its, info->map->vm);
1882
1883		/*
1884		 * Flag the interrupt as forwarded so that we can
1885		 * start poking the virtual property table.
1886		 */
1887		irqd_set_forwarded_to_vcpu(d);
1888
1889		/* Write out the property to the prop table */
1890		lpi_write_config(d, 0xff, info->map->properties);
1891
1892		/* Drop the physical mapping */
1893		its_send_discard(its_dev, event);
1894
1895		/* and install the virtual one */
1896		its_send_vmapti(its_dev, event);
1897
1898		/* Increment the number of VLPIs */
1899		its_dev->event_map.nr_vlpis++;
1900	}
1901
1902out:
1903	raw_spin_unlock(&its_dev->event_map.vlpi_lock);
1904	return ret;
1905}
1906
1907static int its_vlpi_get(struct irq_data *d, struct its_cmd_info *info)
1908{
1909	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1910	struct its_vlpi_map *map;
1911	int ret = 0;
1912
1913	raw_spin_lock(&its_dev->event_map.vlpi_lock);
1914
1915	map = get_vlpi_map(d);
1916
1917	if (!its_dev->event_map.vm || !map) {
1918		ret = -EINVAL;
1919		goto out;
1920	}
1921
1922	/* Copy our mapping information to the incoming request */
1923	*info->map = *map;
1924
1925out:
1926	raw_spin_unlock(&its_dev->event_map.vlpi_lock);
1927	return ret;
1928}
1929
1930static int its_vlpi_unmap(struct irq_data *d)
1931{
1932	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1933	u32 event = its_get_event_id(d);
1934	int ret = 0;
1935
1936	raw_spin_lock(&its_dev->event_map.vlpi_lock);
1937
1938	if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d)) {
1939		ret = -EINVAL;
1940		goto out;
1941	}
1942
1943	/* Drop the virtual mapping */
1944	its_send_discard(its_dev, event);
1945
1946	/* and restore the physical one */
1947	irqd_clr_forwarded_to_vcpu(d);
1948	its_send_mapti(its_dev, d->hwirq, event);
1949	lpi_update_config(d, 0xff, (LPI_PROP_DEFAULT_PRIO |
1950				    LPI_PROP_ENABLED |
1951				    LPI_PROP_GROUP1));
1952
1953	/* Potentially unmap the VM from this ITS */
1954	its_unmap_vm(its_dev->its, its_dev->event_map.vm);
1955
1956	/*
1957	 * Drop the refcount and make the device available again if
1958	 * this was the last VLPI.
1959	 */
1960	if (!--its_dev->event_map.nr_vlpis) {
1961		its_dev->event_map.vm = NULL;
1962		kfree(its_dev->event_map.vlpi_maps);
1963	}
1964
1965out:
1966	raw_spin_unlock(&its_dev->event_map.vlpi_lock);
1967	return ret;
1968}
1969
1970static int its_vlpi_prop_update(struct irq_data *d, struct its_cmd_info *info)
1971{
1972	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1973
1974	if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d))
1975		return -EINVAL;
1976
1977	if (info->cmd_type == PROP_UPDATE_AND_INV_VLPI)
1978		lpi_update_config(d, 0xff, info->config);
1979	else
1980		lpi_write_config(d, 0xff, info->config);
1981	its_vlpi_set_doorbell(d, !!(info->config & LPI_PROP_ENABLED));
1982
1983	return 0;
1984}
1985
1986static int its_irq_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
1987{
1988	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1989	struct its_cmd_info *info = vcpu_info;
1990
1991	/* Need a v4 ITS */
1992	if (!is_v4(its_dev->its))
1993		return -EINVAL;
1994
1995	/* Unmap request? */
1996	if (!info)
1997		return its_vlpi_unmap(d);
1998
1999	switch (info->cmd_type) {
2000	case MAP_VLPI:
2001		return its_vlpi_map(d, info);
2002
2003	case GET_VLPI:
2004		return its_vlpi_get(d, info);
2005
2006	case PROP_UPDATE_VLPI:
2007	case PROP_UPDATE_AND_INV_VLPI:
2008		return its_vlpi_prop_update(d, info);
2009
2010	default:
2011		return -EINVAL;
2012	}
2013}
2014
2015static struct irq_chip its_irq_chip = {
2016	.name			= "ITS",
2017	.irq_mask		= its_mask_irq,
2018	.irq_unmask		= its_unmask_irq,
2019	.irq_eoi		= irq_chip_eoi_parent,
2020	.irq_set_affinity	= its_set_affinity,
2021	.irq_compose_msi_msg	= its_irq_compose_msi_msg,
2022	.irq_set_irqchip_state	= its_irq_set_irqchip_state,
2023	.irq_retrigger		= its_irq_retrigger,
2024	.irq_set_vcpu_affinity	= its_irq_set_vcpu_affinity,
2025};
2026
2027
2028/*
2029 * How we allocate LPIs:
2030 *
2031 * lpi_range_list contains ranges of LPIs that are to available to
2032 * allocate from. To allocate LPIs, just pick the first range that
2033 * fits the required allocation, and reduce it by the required
2034 * amount. Once empty, remove the range from the list.
2035 *
2036 * To free a range of LPIs, add a free range to the list, sort it and
2037 * merge the result if the new range happens to be adjacent to an
2038 * already free block.
2039 *
2040 * The consequence of the above is that allocation is cost is low, but
2041 * freeing is expensive. We assumes that freeing rarely occurs.
2042 */
2043#define ITS_MAX_LPI_NRBITS	16 /* 64K LPIs */
 
2044
2045static DEFINE_MUTEX(lpi_range_lock);
2046static LIST_HEAD(lpi_range_list);
 
2047
2048struct lpi_range {
2049	struct list_head	entry;
2050	u32			base_id;
2051	u32			span;
2052};
2053
2054static struct lpi_range *mk_lpi_range(u32 base, u32 span)
2055{
2056	struct lpi_range *range;
2057
2058	range = kmalloc(sizeof(*range), GFP_KERNEL);
2059	if (range) {
2060		range->base_id = base;
2061		range->span = span;
2062	}
2063
2064	return range;
2065}
2066
2067static int alloc_lpi_range(u32 nr_lpis, u32 *base)
2068{
2069	struct lpi_range *range, *tmp;
2070	int err = -ENOSPC;
2071
2072	mutex_lock(&lpi_range_lock);
2073
2074	list_for_each_entry_safe(range, tmp, &lpi_range_list, entry) {
2075		if (range->span >= nr_lpis) {
2076			*base = range->base_id;
2077			range->base_id += nr_lpis;
2078			range->span -= nr_lpis;
2079
2080			if (range->span == 0) {
2081				list_del(&range->entry);
2082				kfree(range);
2083			}
2084
2085			err = 0;
2086			break;
2087		}
2088	}
2089
2090	mutex_unlock(&lpi_range_lock);
2091
2092	pr_debug("ITS: alloc %u:%u\n", *base, nr_lpis);
2093	return err;
2094}
2095
2096static void merge_lpi_ranges(struct lpi_range *a, struct lpi_range *b)
2097{
2098	if (&a->entry == &lpi_range_list || &b->entry == &lpi_range_list)
2099		return;
2100	if (a->base_id + a->span != b->base_id)
2101		return;
2102	b->base_id = a->base_id;
2103	b->span += a->span;
2104	list_del(&a->entry);
2105	kfree(a);
2106}
2107
2108static int free_lpi_range(u32 base, u32 nr_lpis)
2109{
2110	struct lpi_range *new, *old;
2111
2112	new = mk_lpi_range(base, nr_lpis);
2113	if (!new)
 
 
2114		return -ENOMEM;
2115
2116	mutex_lock(&lpi_range_lock);
2117
2118	list_for_each_entry_reverse(old, &lpi_range_list, entry) {
2119		if (old->base_id < base)
2120			break;
2121	}
2122	/*
2123	 * old is the last element with ->base_id smaller than base,
2124	 * so new goes right after it. If there are no elements with
2125	 * ->base_id smaller than base, &old->entry ends up pointing
2126	 * at the head of the list, and inserting new it the start of
2127	 * the list is the right thing to do in that case as well.
2128	 */
2129	list_add(&new->entry, &old->entry);
2130	/*
2131	 * Now check if we can merge with the preceding and/or
2132	 * following ranges.
2133	 */
2134	merge_lpi_ranges(old, new);
2135	merge_lpi_ranges(new, list_next_entry(new, entry));
2136
2137	mutex_unlock(&lpi_range_lock);
2138	return 0;
2139}
2140
2141static int __init its_lpi_init(u32 id_bits)
2142{
2143	u32 lpis = (1UL << id_bits) - 8192;
2144	u32 numlpis;
2145	int err;
 
2146
2147	numlpis = 1UL << GICD_TYPER_NUM_LPIS(gic_rdists->gicd_typer);
2148
2149	if (numlpis > 2 && !WARN_ON(numlpis > lpis)) {
2150		lpis = numlpis;
2151		pr_info("ITS: Using hypervisor restricted LPI range [%u]\n",
2152			lpis);
2153	}
2154
2155	/*
2156	 * Initializing the allocator is just the same as freeing the
2157	 * full range of LPIs.
2158	 */
2159	err = free_lpi_range(8192, lpis);
2160	pr_debug("ITS: Allocator initialized for %u LPIs\n", lpis);
2161	return err;
2162}
2163
2164static unsigned long *its_lpi_alloc(int nr_irqs, u32 *base, int *nr_ids)
2165{
2166	unsigned long *bitmap = NULL;
2167	int err = 0;
2168
2169	do {
2170		err = alloc_lpi_range(nr_irqs, base);
2171		if (!err)
 
2172			break;
2173
2174		nr_irqs /= 2;
2175	} while (nr_irqs > 0);
2176
2177	if (!nr_irqs)
2178		err = -ENOSPC;
2179
2180	if (err)
2181		goto out;
2182
2183	bitmap = bitmap_zalloc(nr_irqs, GFP_ATOMIC);
 
2184	if (!bitmap)
2185		goto out;
2186
2187	*nr_ids = nr_irqs;
 
 
 
 
2188
2189out:
 
 
2190	if (!bitmap)
2191		*base = *nr_ids = 0;
2192
2193	return bitmap;
2194}
2195
2196static void its_lpi_free(unsigned long *bitmap, u32 base, u32 nr_ids)
2197{
2198	WARN_ON(free_lpi_range(base, nr_ids));
2199	bitmap_free(bitmap);
2200}
 
 
 
 
 
 
 
 
 
 
 
 
2201
2202static void gic_reset_prop_table(void *va)
2203{
2204	/* Priority 0xa0, Group-1, disabled */
2205	memset(va, LPI_PROP_DEFAULT_PRIO | LPI_PROP_GROUP1, LPI_PROPBASE_SZ);
2206
2207	/* Make sure the GIC will observe the written configuration */
2208	gic_flush_dcache_to_poc(va, LPI_PROPBASE_SZ);
2209}
2210
2211static struct page *its_allocate_prop_table(gfp_t gfp_flags)
2212{
2213	struct page *prop_page;
 
 
 
 
2214
2215	prop_page = alloc_pages(gfp_flags, get_order(LPI_PROPBASE_SZ));
2216	if (!prop_page)
2217		return NULL;
2218
2219	gic_reset_prop_table(page_address(prop_page));
2220
2221	return prop_page;
2222}
2223
2224static void its_free_prop_table(struct page *prop_page)
2225{
2226	free_pages((unsigned long)page_address(prop_page),
2227		   get_order(LPI_PROPBASE_SZ));
2228}
2229
2230static bool gic_check_reserved_range(phys_addr_t addr, unsigned long size)
2231{
2232	phys_addr_t start, end, addr_end;
2233	u64 i;
 
 
2234
2235	/*
2236	 * We don't bother checking for a kdump kernel as by
2237	 * construction, the LPI tables are out of this kernel's
2238	 * memory map.
2239	 */
2240	if (is_kdump_kernel())
2241		return true;
2242
2243	addr_end = addr + size - 1;
 
 
 
2244
2245	for_each_reserved_mem_range(i, &start, &end) {
2246		if (addr >= start && addr_end <= end)
2247			return true;
2248	}
2249
2250	/* Not found, not a good sign... */
2251	pr_warn("GICv3: Expected reserved range [%pa:%pa], not found\n",
2252		&addr, &addr_end);
2253	add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
2254	return false;
2255}
2256
2257static int gic_reserve_range(phys_addr_t addr, unsigned long size)
2258{
2259	if (efi_enabled(EFI_CONFIG_TABLES))
2260		return efi_mem_reserve_persistent(addr, size);
2261
2262	return 0;
2263}
2264
2265static int __init its_setup_lpi_prop_table(void)
2266{
2267	if (gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED) {
2268		u64 val;
2269
2270		val = gicr_read_propbaser(gic_data_rdist_rd_base() + GICR_PROPBASER);
2271		lpi_id_bits = (val & GICR_PROPBASER_IDBITS_MASK) + 1;
2272
2273		gic_rdists->prop_table_pa = val & GENMASK_ULL(51, 12);
2274		gic_rdists->prop_table_va = memremap(gic_rdists->prop_table_pa,
2275						     LPI_PROPBASE_SZ,
2276						     MEMREMAP_WB);
2277		gic_reset_prop_table(gic_rdists->prop_table_va);
2278	} else {
2279		struct page *page;
2280
2281		lpi_id_bits = min_t(u32,
2282				    GICD_TYPER_ID_BITS(gic_rdists->gicd_typer),
2283				    ITS_MAX_LPI_NRBITS);
2284		page = its_allocate_prop_table(GFP_NOWAIT);
2285		if (!page) {
2286			pr_err("Failed to allocate PROPBASE\n");
2287			return -ENOMEM;
2288		}
2289
2290		gic_rdists->prop_table_pa = page_to_phys(page);
2291		gic_rdists->prop_table_va = page_address(page);
2292		WARN_ON(gic_reserve_range(gic_rdists->prop_table_pa,
2293					  LPI_PROPBASE_SZ));
2294	}
2295
2296	pr_info("GICv3: using LPI property table @%pa\n",
2297		&gic_rdists->prop_table_pa);
2298
2299	return its_lpi_init(lpi_id_bits);
2300}
2301
2302static const char *its_base_type_string[] = {
2303	[GITS_BASER_TYPE_DEVICE]	= "Devices",
2304	[GITS_BASER_TYPE_VCPU]		= "Virtual CPUs",
2305	[GITS_BASER_TYPE_RESERVED3]	= "Reserved (3)",
2306	[GITS_BASER_TYPE_COLLECTION]	= "Interrupt Collections",
2307	[GITS_BASER_TYPE_RESERVED5] 	= "Reserved (5)",
2308	[GITS_BASER_TYPE_RESERVED6] 	= "Reserved (6)",
2309	[GITS_BASER_TYPE_RESERVED7] 	= "Reserved (7)",
2310};
2311
2312static u64 its_read_baser(struct its_node *its, struct its_baser *baser)
2313{
2314	u32 idx = baser - its->tables;
2315
2316	return gits_read_baser(its->base + GITS_BASER + (idx << 3));
2317}
2318
2319static void its_write_baser(struct its_node *its, struct its_baser *baser,
2320			    u64 val)
2321{
2322	u32 idx = baser - its->tables;
2323
2324	gits_write_baser(val, its->base + GITS_BASER + (idx << 3));
2325	baser->val = its_read_baser(its, baser);
2326}
2327
2328static int its_setup_baser(struct its_node *its, struct its_baser *baser,
2329			   u64 cache, u64 shr, u32 order, bool indirect)
2330{
2331	u64 val = its_read_baser(its, baser);
2332	u64 esz = GITS_BASER_ENTRY_SIZE(val);
2333	u64 type = GITS_BASER_TYPE(val);
2334	u64 baser_phys, tmp;
2335	u32 alloc_pages, psz;
2336	struct page *page;
2337	void *base;
2338
2339	psz = baser->psz;
2340	alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz);
2341	if (alloc_pages > GITS_BASER_PAGES_MAX) {
2342		pr_warn("ITS@%pa: %s too large, reduce ITS pages %u->%u\n",
2343			&its->phys_base, its_base_type_string[type],
2344			alloc_pages, GITS_BASER_PAGES_MAX);
2345		alloc_pages = GITS_BASER_PAGES_MAX;
2346		order = get_order(GITS_BASER_PAGES_MAX * psz);
2347	}
2348
2349	page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO, order);
2350	if (!page)
2351		return -ENOMEM;
2352
2353	base = (void *)page_address(page);
2354	baser_phys = virt_to_phys(base);
2355
2356	/* Check if the physical address of the memory is above 48bits */
2357	if (IS_ENABLED(CONFIG_ARM64_64K_PAGES) && (baser_phys >> 48)) {
2358
2359		/* 52bit PA is supported only when PageSize=64K */
2360		if (psz != SZ_64K) {
2361			pr_err("ITS: no 52bit PA support when psz=%d\n", psz);
2362			free_pages((unsigned long)base, order);
2363			return -ENXIO;
2364		}
2365
2366		/* Convert 52bit PA to 48bit field */
2367		baser_phys = GITS_BASER_PHYS_52_to_48(baser_phys);
2368	}
2369
2370retry_baser:
2371	val = (baser_phys					 |
2372		(type << GITS_BASER_TYPE_SHIFT)			 |
2373		((esz - 1) << GITS_BASER_ENTRY_SIZE_SHIFT)	 |
2374		((alloc_pages - 1) << GITS_BASER_PAGES_SHIFT)	 |
2375		cache						 |
2376		shr						 |
2377		GITS_BASER_VALID);
2378
2379	val |=	indirect ? GITS_BASER_INDIRECT : 0x0;
2380
2381	switch (psz) {
2382	case SZ_4K:
2383		val |= GITS_BASER_PAGE_SIZE_4K;
2384		break;
2385	case SZ_16K:
2386		val |= GITS_BASER_PAGE_SIZE_16K;
2387		break;
2388	case SZ_64K:
2389		val |= GITS_BASER_PAGE_SIZE_64K;
2390		break;
2391	}
2392
2393	if (!shr)
2394		gic_flush_dcache_to_poc(base, PAGE_ORDER_TO_SIZE(order));
2395
2396	its_write_baser(its, baser, val);
2397	tmp = baser->val;
2398
2399	if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
2400		/*
2401		 * Shareability didn't stick. Just use
2402		 * whatever the read reported, which is likely
2403		 * to be the only thing this redistributor
2404		 * supports. If that's zero, make it
2405		 * non-cacheable as well.
2406		 */
2407		shr = tmp & GITS_BASER_SHAREABILITY_MASK;
2408		if (!shr)
2409			cache = GITS_BASER_nC;
2410
2411		goto retry_baser;
2412	}
2413
2414	if (val != tmp) {
2415		pr_err("ITS@%pa: %s doesn't stick: %llx %llx\n",
2416		       &its->phys_base, its_base_type_string[type],
2417		       val, tmp);
2418		free_pages((unsigned long)base, order);
2419		return -ENXIO;
2420	}
2421
2422	baser->order = order;
2423	baser->base = base;
2424	baser->psz = psz;
2425	tmp = indirect ? GITS_LVL1_ENTRY_SIZE : esz;
2426
2427	pr_info("ITS@%pa: allocated %d %s @%lx (%s, esz %d, psz %dK, shr %d)\n",
2428		&its->phys_base, (int)(PAGE_ORDER_TO_SIZE(order) / (int)tmp),
2429		its_base_type_string[type],
2430		(unsigned long)virt_to_phys(base),
2431		indirect ? "indirect" : "flat", (int)esz,
2432		psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
2433
2434	return 0;
2435}
2436
2437static bool its_parse_indirect_baser(struct its_node *its,
2438				     struct its_baser *baser,
2439				     u32 *order, u32 ids)
2440{
2441	u64 tmp = its_read_baser(its, baser);
2442	u64 type = GITS_BASER_TYPE(tmp);
2443	u64 esz = GITS_BASER_ENTRY_SIZE(tmp);
2444	u64 val = GITS_BASER_InnerShareable | GITS_BASER_RaWaWb;
2445	u32 new_order = *order;
2446	u32 psz = baser->psz;
2447	bool indirect = false;
2448
2449	/* No need to enable Indirection if memory requirement < (psz*2)bytes */
2450	if ((esz << ids) > (psz * 2)) {
2451		/*
2452		 * Find out whether hw supports a single or two-level table by
2453		 * table by reading bit at offset '62' after writing '1' to it.
2454		 */
2455		its_write_baser(its, baser, val | GITS_BASER_INDIRECT);
2456		indirect = !!(baser->val & GITS_BASER_INDIRECT);
2457
2458		if (indirect) {
2459			/*
2460			 * The size of the lvl2 table is equal to ITS page size
2461			 * which is 'psz'. For computing lvl1 table size,
2462			 * subtract ID bits that sparse lvl2 table from 'ids'
2463			 * which is reported by ITS hardware times lvl1 table
2464			 * entry size.
2465			 */
2466			ids -= ilog2(psz / (int)esz);
2467			esz = GITS_LVL1_ENTRY_SIZE;
2468		}
2469	}
2470
2471	/*
2472	 * Allocate as many entries as required to fit the
2473	 * range of device IDs that the ITS can grok... The ID
2474	 * space being incredibly sparse, this results in a
2475	 * massive waste of memory if two-level device table
2476	 * feature is not supported by hardware.
2477	 */
2478	new_order = max_t(u32, get_order(esz << ids), new_order);
2479	if (new_order > MAX_PAGE_ORDER) {
2480		new_order = MAX_PAGE_ORDER;
2481		ids = ilog2(PAGE_ORDER_TO_SIZE(new_order) / (int)esz);
2482		pr_warn("ITS@%pa: %s Table too large, reduce ids %llu->%u\n",
2483			&its->phys_base, its_base_type_string[type],
2484			device_ids(its), ids);
2485	}
2486
2487	*order = new_order;
2488
2489	return indirect;
2490}
2491
2492static u32 compute_common_aff(u64 val)
2493{
2494	u32 aff, clpiaff;
2495
2496	aff = FIELD_GET(GICR_TYPER_AFFINITY, val);
2497	clpiaff = FIELD_GET(GICR_TYPER_COMMON_LPI_AFF, val);
2498
2499	return aff & ~(GENMASK(31, 0) >> (clpiaff * 8));
2500}
2501
2502static u32 compute_its_aff(struct its_node *its)
2503{
2504	u64 val;
2505	u32 svpet;
2506
2507	/*
2508	 * Reencode the ITS SVPET and MPIDR as a GICR_TYPER, and compute
2509	 * the resulting affinity. We then use that to see if this match
2510	 * our own affinity.
2511	 */
2512	svpet = FIELD_GET(GITS_TYPER_SVPET, its->typer);
2513	val  = FIELD_PREP(GICR_TYPER_COMMON_LPI_AFF, svpet);
2514	val |= FIELD_PREP(GICR_TYPER_AFFINITY, its->mpidr);
2515	return compute_common_aff(val);
2516}
2517
2518static struct its_node *find_sibling_its(struct its_node *cur_its)
2519{
2520	struct its_node *its;
2521	u32 aff;
2522
2523	if (!FIELD_GET(GITS_TYPER_SVPET, cur_its->typer))
2524		return NULL;
2525
2526	aff = compute_its_aff(cur_its);
2527
2528	list_for_each_entry(its, &its_nodes, entry) {
2529		u64 baser;
2530
2531		if (!is_v4_1(its) || its == cur_its)
2532			continue;
2533
2534		if (!FIELD_GET(GITS_TYPER_SVPET, its->typer))
2535			continue;
2536
2537		if (aff != compute_its_aff(its))
2538			continue;
2539
2540		/* GICv4.1 guarantees that the vPE table is GITS_BASER2 */
2541		baser = its->tables[2].val;
2542		if (!(baser & GITS_BASER_VALID))
2543			continue;
2544
2545		return its;
2546	}
2547
2548	return NULL;
2549}
2550
2551static void its_free_tables(struct its_node *its)
2552{
2553	int i;
2554
2555	for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2556		if (its->tables[i].base) {
2557			free_pages((unsigned long)its->tables[i].base,
2558				   its->tables[i].order);
2559			its->tables[i].base = NULL;
2560		}
2561	}
2562}
2563
2564static int its_probe_baser_psz(struct its_node *its, struct its_baser *baser)
2565{
2566	u64 psz = SZ_64K;
2567
2568	while (psz) {
2569		u64 val, gpsz;
2570
2571		val = its_read_baser(its, baser);
2572		val &= ~GITS_BASER_PAGE_SIZE_MASK;
2573
2574		switch (psz) {
2575		case SZ_64K:
2576			gpsz = GITS_BASER_PAGE_SIZE_64K;
2577			break;
2578		case SZ_16K:
2579			gpsz = GITS_BASER_PAGE_SIZE_16K;
2580			break;
2581		case SZ_4K:
2582		default:
2583			gpsz = GITS_BASER_PAGE_SIZE_4K;
2584			break;
2585		}
2586
2587		gpsz >>= GITS_BASER_PAGE_SIZE_SHIFT;
2588
2589		val |= FIELD_PREP(GITS_BASER_PAGE_SIZE_MASK, gpsz);
2590		its_write_baser(its, baser, val);
2591
2592		if (FIELD_GET(GITS_BASER_PAGE_SIZE_MASK, baser->val) == gpsz)
2593			break;
2594
2595		switch (psz) {
2596		case SZ_64K:
2597			psz = SZ_16K;
2598			break;
2599		case SZ_16K:
2600			psz = SZ_4K;
2601			break;
2602		case SZ_4K:
2603		default:
2604			return -1;
2605		}
2606	}
2607
2608	baser->psz = psz;
2609	return 0;
2610}
2611
2612static int its_alloc_tables(struct its_node *its)
2613{
 
 
 
2614	u64 shr = GITS_BASER_InnerShareable;
2615	u64 cache = GITS_BASER_RaWaWb;
2616	int err, i;
 
2617
2618	if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_22375)
2619		/* erratum 24313: ignore memory access type */
2620		cache = GITS_BASER_nCnB;
2621
2622	if (its->flags & ITS_FLAGS_FORCE_NON_SHAREABLE) {
2623		cache = GITS_BASER_nC;
2624		shr = 0;
 
 
 
 
2625	}
2626
2627	for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2628		struct its_baser *baser = its->tables + i;
2629		u64 val = its_read_baser(its, baser);
2630		u64 type = GITS_BASER_TYPE(val);
2631		bool indirect = false;
2632		u32 order;
 
 
 
2633
2634		if (type == GITS_BASER_TYPE_NONE)
2635			continue;
2636
2637		if (its_probe_baser_psz(its, baser)) {
2638			its_free_tables(its);
2639			return -ENXIO;
2640		}
2641
2642		order = get_order(baser->psz);
2643
2644		switch (type) {
2645		case GITS_BASER_TYPE_DEVICE:
2646			indirect = its_parse_indirect_baser(its, baser, &order,
2647							    device_ids(its));
2648			break;
2649
2650		case GITS_BASER_TYPE_VCPU:
2651			if (is_v4_1(its)) {
2652				struct its_node *sibling;
2653
2654				WARN_ON(i != 2);
2655				if ((sibling = find_sibling_its(its))) {
2656					*baser = sibling->tables[2];
2657					its_write_baser(its, baser, baser->val);
2658					continue;
2659				}
2660			}
2661
2662			indirect = its_parse_indirect_baser(its, baser, &order,
2663							    ITS_MAX_VPEID_BITS);
2664			break;
2665		}
2666
2667		err = its_setup_baser(its, baser, cache, shr, order, indirect);
2668		if (err < 0) {
2669			its_free_tables(its);
2670			return err;
 
 
 
 
 
 
 
 
 
2671		}
2672
2673		/* Update settings which will be used for next BASERn */
2674		cache = baser->val & GITS_BASER_CACHEABILITY_MASK;
2675		shr = baser->val & GITS_BASER_SHAREABILITY_MASK;
2676	}
2677
2678	return 0;
2679}
 
 
 
 
 
2680
2681static u64 inherit_vpe_l1_table_from_its(void)
2682{
2683	struct its_node *its;
2684	u64 val;
2685	u32 aff;
2686
2687	val = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
2688	aff = compute_common_aff(val);
2689
2690	list_for_each_entry(its, &its_nodes, entry) {
2691		u64 baser, addr;
2692
2693		if (!is_v4_1(its))
2694			continue;
2695
2696		if (!FIELD_GET(GITS_TYPER_SVPET, its->typer))
2697			continue;
2698
2699		if (aff != compute_its_aff(its))
2700			continue;
2701
2702		/* GICv4.1 guarantees that the vPE table is GITS_BASER2 */
2703		baser = its->tables[2].val;
2704		if (!(baser & GITS_BASER_VALID))
2705			continue;
2706
2707		/* We have a winner! */
2708		gic_data_rdist()->vpe_l1_base = its->tables[2].base;
2709
2710		val  = GICR_VPROPBASER_4_1_VALID;
2711		if (baser & GITS_BASER_INDIRECT)
2712			val |= GICR_VPROPBASER_4_1_INDIRECT;
2713		val |= FIELD_PREP(GICR_VPROPBASER_4_1_PAGE_SIZE,
2714				  FIELD_GET(GITS_BASER_PAGE_SIZE_MASK, baser));
2715		switch (FIELD_GET(GITS_BASER_PAGE_SIZE_MASK, baser)) {
2716		case GIC_PAGE_SIZE_64K:
2717			addr = GITS_BASER_ADDR_48_to_52(baser);
2718			break;
2719		default:
2720			addr = baser & GENMASK_ULL(47, 12);
2721			break;
2722		}
2723		val |= FIELD_PREP(GICR_VPROPBASER_4_1_ADDR, addr >> 12);
2724		if (rdists_support_shareable()) {
2725			val |= FIELD_PREP(GICR_VPROPBASER_SHAREABILITY_MASK,
2726					  FIELD_GET(GITS_BASER_SHAREABILITY_MASK, baser));
2727			val |= FIELD_PREP(GICR_VPROPBASER_INNER_CACHEABILITY_MASK,
2728					  FIELD_GET(GITS_BASER_INNER_CACHEABILITY_MASK, baser));
2729		}
2730		val |= FIELD_PREP(GICR_VPROPBASER_4_1_SIZE, GITS_BASER_NR_PAGES(baser) - 1);
2731
2732		return val;
2733	}
2734
2735	return 0;
2736}
2737
2738static u64 inherit_vpe_l1_table_from_rd(cpumask_t **mask)
2739{
2740	u32 aff;
2741	u64 val;
2742	int cpu;
 
 
 
 
 
 
 
 
 
 
2743
2744	val = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
2745	aff = compute_common_aff(val);
 
 
 
 
 
 
2746
2747	for_each_possible_cpu(cpu) {
2748		void __iomem *base = gic_data_rdist_cpu(cpu)->rd_base;
 
 
 
 
 
 
 
2749
2750		if (!base || cpu == smp_processor_id())
2751			continue;
2752
2753		val = gic_read_typer(base + GICR_TYPER);
2754		if (aff != compute_common_aff(val))
2755			continue;
2756
2757		/*
2758		 * At this point, we have a victim. This particular CPU
2759		 * has already booted, and has an affinity that matches
2760		 * ours wrt CommonLPIAff. Let's use its own VPROPBASER.
2761		 * Make sure we don't write the Z bit in that case.
2762		 */
2763		val = gicr_read_vpropbaser(base + SZ_128K + GICR_VPROPBASER);
2764		val &= ~GICR_VPROPBASER_4_1_Z;
2765
2766		gic_data_rdist()->vpe_l1_base = gic_data_rdist_cpu(cpu)->vpe_l1_base;
2767		*mask = gic_data_rdist_cpu(cpu)->vpe_table_mask;
2768
2769		return val;
 
2770	}
2771
2772	return 0;
2773}
2774
2775static bool allocate_vpe_l2_table(int cpu, u32 id)
2776{
2777	void __iomem *base = gic_data_rdist_cpu(cpu)->rd_base;
2778	unsigned int psz, esz, idx, npg, gpsz;
2779	u64 val;
2780	struct page *page;
2781	__le64 *table;
2782
2783	if (!gic_rdists->has_rvpeid)
2784		return true;
2785
2786	/* Skip non-present CPUs */
2787	if (!base)
2788		return true;
2789
2790	val  = gicr_read_vpropbaser(base + SZ_128K + GICR_VPROPBASER);
2791
2792	esz  = FIELD_GET(GICR_VPROPBASER_4_1_ENTRY_SIZE, val) + 1;
2793	gpsz = FIELD_GET(GICR_VPROPBASER_4_1_PAGE_SIZE, val);
2794	npg  = FIELD_GET(GICR_VPROPBASER_4_1_SIZE, val) + 1;
2795
2796	switch (gpsz) {
2797	default:
2798		WARN_ON(1);
2799		fallthrough;
2800	case GIC_PAGE_SIZE_4K:
2801		psz = SZ_4K;
2802		break;
2803	case GIC_PAGE_SIZE_16K:
2804		psz = SZ_16K;
2805		break;
2806	case GIC_PAGE_SIZE_64K:
2807		psz = SZ_64K;
2808		break;
2809	}
2810
2811	/* Don't allow vpe_id that exceeds single, flat table limit */
2812	if (!(val & GICR_VPROPBASER_4_1_INDIRECT))
2813		return (id < (npg * psz / (esz * SZ_8)));
2814
2815	/* Compute 1st level table index & check if that exceeds table limit */
2816	idx = id >> ilog2(psz / (esz * SZ_8));
2817	if (idx >= (npg * psz / GITS_LVL1_ENTRY_SIZE))
2818		return false;
2819
2820	table = gic_data_rdist_cpu(cpu)->vpe_l1_base;
2821
2822	/* Allocate memory for 2nd level table */
2823	if (!table[idx]) {
2824		page = alloc_pages(GFP_KERNEL | __GFP_ZERO, get_order(psz));
2825		if (!page)
2826			return false;
2827
2828		/* Flush Lvl2 table to PoC if hw doesn't support coherency */
2829		if (!(val & GICR_VPROPBASER_SHAREABILITY_MASK))
2830			gic_flush_dcache_to_poc(page_address(page), psz);
2831
2832		table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
2833
2834		/* Flush Lvl1 entry to PoC if hw doesn't support coherency */
2835		if (!(val & GICR_VPROPBASER_SHAREABILITY_MASK))
2836			gic_flush_dcache_to_poc(table + idx, GITS_LVL1_ENTRY_SIZE);
2837
2838		/* Ensure updated table contents are visible to RD hardware */
2839		dsb(sy);
2840	}
2841
2842	return true;
2843}
2844
2845static int allocate_vpe_l1_table(void)
2846{
2847	void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2848	u64 val, gpsz, npg, pa;
2849	unsigned int psz = SZ_64K;
2850	unsigned int np, epp, esz;
2851	struct page *page;
2852
2853	if (!gic_rdists->has_rvpeid)
2854		return 0;
2855
2856	/*
2857	 * if VPENDBASER.Valid is set, disable any previously programmed
2858	 * VPE by setting PendingLast while clearing Valid. This has the
2859	 * effect of making sure no doorbell will be generated and we can
2860	 * then safely clear VPROPBASER.Valid.
2861	 */
2862	if (gicr_read_vpendbaser(vlpi_base + GICR_VPENDBASER) & GICR_VPENDBASER_Valid)
2863		gicr_write_vpendbaser(GICR_VPENDBASER_PendingLast,
2864				      vlpi_base + GICR_VPENDBASER);
2865
2866	/*
2867	 * If we can inherit the configuration from another RD, let's do
2868	 * so. Otherwise, we have to go through the allocation process. We
2869	 * assume that all RDs have the exact same requirements, as
2870	 * nothing will work otherwise.
2871	 */
2872	val = inherit_vpe_l1_table_from_rd(&gic_data_rdist()->vpe_table_mask);
2873	if (val & GICR_VPROPBASER_4_1_VALID)
2874		goto out;
2875
2876	gic_data_rdist()->vpe_table_mask = kzalloc(sizeof(cpumask_t), GFP_ATOMIC);
2877	if (!gic_data_rdist()->vpe_table_mask)
2878		return -ENOMEM;
2879
2880	val = inherit_vpe_l1_table_from_its();
2881	if (val & GICR_VPROPBASER_4_1_VALID)
2882		goto out;
2883
2884	/* First probe the page size */
2885	val = FIELD_PREP(GICR_VPROPBASER_4_1_PAGE_SIZE, GIC_PAGE_SIZE_64K);
2886	gicr_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2887	val = gicr_read_vpropbaser(vlpi_base + GICR_VPROPBASER);
2888	gpsz = FIELD_GET(GICR_VPROPBASER_4_1_PAGE_SIZE, val);
2889	esz = FIELD_GET(GICR_VPROPBASER_4_1_ENTRY_SIZE, val);
2890
2891	switch (gpsz) {
2892	default:
2893		gpsz = GIC_PAGE_SIZE_4K;
2894		fallthrough;
2895	case GIC_PAGE_SIZE_4K:
2896		psz = SZ_4K;
2897		break;
2898	case GIC_PAGE_SIZE_16K:
2899		psz = SZ_16K;
2900		break;
2901	case GIC_PAGE_SIZE_64K:
2902		psz = SZ_64K;
2903		break;
2904	}
2905
2906	/*
2907	 * Start populating the register from scratch, including RO fields
2908	 * (which we want to print in debug cases...)
2909	 */
2910	val = 0;
2911	val |= FIELD_PREP(GICR_VPROPBASER_4_1_PAGE_SIZE, gpsz);
2912	val |= FIELD_PREP(GICR_VPROPBASER_4_1_ENTRY_SIZE, esz);
2913
2914	/* How many entries per GIC page? */
2915	esz++;
2916	epp = psz / (esz * SZ_8);
2917
2918	/*
2919	 * If we need more than just a single L1 page, flag the table
2920	 * as indirect and compute the number of required L1 pages.
2921	 */
2922	if (epp < ITS_MAX_VPEID) {
2923		int nl2;
2924
2925		val |= GICR_VPROPBASER_4_1_INDIRECT;
2926
2927		/* Number of L2 pages required to cover the VPEID space */
2928		nl2 = DIV_ROUND_UP(ITS_MAX_VPEID, epp);
2929
2930		/* Number of L1 pages to point to the L2 pages */
2931		npg = DIV_ROUND_UP(nl2 * SZ_8, psz);
2932	} else {
2933		npg = 1;
2934	}
2935
2936	val |= FIELD_PREP(GICR_VPROPBASER_4_1_SIZE, npg - 1);
2937
2938	/* Right, that's the number of CPU pages we need for L1 */
2939	np = DIV_ROUND_UP(npg * psz, PAGE_SIZE);
2940
2941	pr_debug("np = %d, npg = %lld, psz = %d, epp = %d, esz = %d\n",
2942		 np, npg, psz, epp, esz);
2943	page = alloc_pages(GFP_ATOMIC | __GFP_ZERO, get_order(np * PAGE_SIZE));
2944	if (!page)
2945		return -ENOMEM;
2946
2947	gic_data_rdist()->vpe_l1_base = page_address(page);
2948	pa = virt_to_phys(page_address(page));
2949	WARN_ON(!IS_ALIGNED(pa, psz));
2950
2951	val |= FIELD_PREP(GICR_VPROPBASER_4_1_ADDR, pa >> 12);
2952	if (rdists_support_shareable()) {
2953		val |= GICR_VPROPBASER_RaWb;
2954		val |= GICR_VPROPBASER_InnerShareable;
2955	}
2956	val |= GICR_VPROPBASER_4_1_Z;
2957	val |= GICR_VPROPBASER_4_1_VALID;
2958
2959out:
2960	gicr_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2961	cpumask_set_cpu(smp_processor_id(), gic_data_rdist()->vpe_table_mask);
2962
2963	pr_debug("CPU%d: VPROPBASER = %llx %*pbl\n",
2964		 smp_processor_id(), val,
2965		 cpumask_pr_args(gic_data_rdist()->vpe_table_mask));
2966
2967	return 0;
2968}
2969
2970static int its_alloc_collections(struct its_node *its)
2971{
2972	int i;
2973
2974	its->collections = kcalloc(nr_cpu_ids, sizeof(*its->collections),
2975				   GFP_KERNEL);
2976	if (!its->collections)
2977		return -ENOMEM;
2978
2979	for (i = 0; i < nr_cpu_ids; i++)
2980		its->collections[i].target_address = ~0ULL;
2981
2982	return 0;
2983}
2984
2985static struct page *its_allocate_pending_table(gfp_t gfp_flags)
2986{
2987	struct page *pend_page;
2988
2989	pend_page = alloc_pages(gfp_flags | __GFP_ZERO,
2990				get_order(LPI_PENDBASE_SZ));
2991	if (!pend_page)
2992		return NULL;
2993
2994	/* Make sure the GIC will observe the zero-ed page */
2995	gic_flush_dcache_to_poc(page_address(pend_page), LPI_PENDBASE_SZ);
2996
2997	return pend_page;
2998}
2999
3000static void its_free_pending_table(struct page *pt)
3001{
3002	free_pages((unsigned long)page_address(pt), get_order(LPI_PENDBASE_SZ));
3003}
3004
3005/*
3006 * Booting with kdump and LPIs enabled is generally fine. Any other
3007 * case is wrong in the absence of firmware/EFI support.
3008 */
3009static bool enabled_lpis_allowed(void)
3010{
3011	phys_addr_t addr;
3012	u64 val;
3013
3014	/* Check whether the property table is in a reserved region */
3015	val = gicr_read_propbaser(gic_data_rdist_rd_base() + GICR_PROPBASER);
3016	addr = val & GENMASK_ULL(51, 12);
3017
3018	return gic_check_reserved_range(addr, LPI_PROPBASE_SZ);
3019}
3020
3021static int __init allocate_lpi_tables(void)
3022{
3023	u64 val;
3024	int err, cpu;
3025
3026	/*
3027	 * If LPIs are enabled while we run this from the boot CPU,
3028	 * flag the RD tables as pre-allocated if the stars do align.
3029	 */
3030	val = readl_relaxed(gic_data_rdist_rd_base() + GICR_CTLR);
3031	if ((val & GICR_CTLR_ENABLE_LPIS) && enabled_lpis_allowed()) {
3032		gic_rdists->flags |= (RDIST_FLAGS_RD_TABLES_PREALLOCATED |
3033				      RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING);
3034		pr_info("GICv3: Using preallocated redistributor tables\n");
3035	}
3036
3037	err = its_setup_lpi_prop_table();
3038	if (err)
3039		return err;
3040
3041	/*
3042	 * We allocate all the pending tables anyway, as we may have a
3043	 * mix of RDs that have had LPIs enabled, and some that
3044	 * don't. We'll free the unused ones as each CPU comes online.
3045	 */
3046	for_each_possible_cpu(cpu) {
3047		struct page *pend_page;
3048
3049		pend_page = its_allocate_pending_table(GFP_NOWAIT);
3050		if (!pend_page) {
3051			pr_err("Failed to allocate PENDBASE for CPU%d\n", cpu);
3052			return -ENOMEM;
3053		}
3054
3055		gic_data_rdist_cpu(cpu)->pend_page = pend_page;
3056	}
3057
3058	return 0;
3059}
3060
3061static u64 read_vpend_dirty_clear(void __iomem *vlpi_base)
3062{
3063	u32 count = 1000000;	/* 1s! */
3064	bool clean;
3065	u64 val;
3066
3067	do {
3068		val = gicr_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
3069		clean = !(val & GICR_VPENDBASER_Dirty);
3070		if (!clean) {
3071			count--;
3072			cpu_relax();
3073			udelay(1);
3074		}
3075	} while (!clean && count);
3076
3077	if (unlikely(!clean))
3078		pr_err_ratelimited("ITS virtual pending table not cleaning\n");
3079
3080	return val;
3081}
3082
3083static u64 its_clear_vpend_valid(void __iomem *vlpi_base, u64 clr, u64 set)
3084{
3085	u64 val;
3086
3087	/* Make sure we wait until the RD is done with the initial scan */
3088	val = read_vpend_dirty_clear(vlpi_base);
3089	val &= ~GICR_VPENDBASER_Valid;
3090	val &= ~clr;
3091	val |= set;
3092	gicr_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
3093
3094	val = read_vpend_dirty_clear(vlpi_base);
3095	if (unlikely(val & GICR_VPENDBASER_Dirty))
3096		val |= GICR_VPENDBASER_PendingLast;
3097
3098	return val;
3099}
3100
3101static void its_cpu_init_lpis(void)
3102{
3103	void __iomem *rbase = gic_data_rdist_rd_base();
3104	struct page *pend_page;
3105	phys_addr_t paddr;
3106	u64 val, tmp;
3107
3108	if (gic_data_rdist()->flags & RD_LOCAL_LPI_ENABLED)
3109		return;
3110
3111	val = readl_relaxed(rbase + GICR_CTLR);
3112	if ((gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED) &&
3113	    (val & GICR_CTLR_ENABLE_LPIS)) {
3114		/*
3115		 * Check that we get the same property table on all
3116		 * RDs. If we don't, this is hopeless.
3117		 */
3118		paddr = gicr_read_propbaser(rbase + GICR_PROPBASER);
3119		paddr &= GENMASK_ULL(51, 12);
3120		if (WARN_ON(gic_rdists->prop_table_pa != paddr))
3121			add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
 
 
 
3122
3123		paddr = gicr_read_pendbaser(rbase + GICR_PENDBASER);
3124		paddr &= GENMASK_ULL(51, 16);
3125
3126		WARN_ON(!gic_check_reserved_range(paddr, LPI_PENDBASE_SZ));
3127		gic_data_rdist()->flags |= RD_LOCAL_PENDTABLE_PREALLOCATED;
 
 
 
3128
3129		goto out;
3130	}
 
 
3131
3132	pend_page = gic_data_rdist()->pend_page;
3133	paddr = page_to_phys(pend_page);
 
 
3134
3135	/* set PROPBASE */
3136	val = (gic_rdists->prop_table_pa |
3137	       GICR_PROPBASER_InnerShareable |
3138	       GICR_PROPBASER_RaWaWb |
3139	       ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
3140
3141	gicr_write_propbaser(val, rbase + GICR_PROPBASER);
3142	tmp = gicr_read_propbaser(rbase + GICR_PROPBASER);
3143
3144	if (!rdists_support_shareable())
3145		tmp &= ~GICR_PROPBASER_SHAREABILITY_MASK;
3146
3147	if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
3148		if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
3149			/*
3150			 * The HW reports non-shareable, we must
3151			 * remove the cacheability attributes as
3152			 * well.
3153			 */
3154			val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
3155				 GICR_PROPBASER_CACHEABILITY_MASK);
3156			val |= GICR_PROPBASER_nC;
3157			gicr_write_propbaser(val, rbase + GICR_PROPBASER);
3158		}
3159		pr_info_once("GIC: using cache flushing for LPI property table\n");
3160		gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
3161	}
3162
3163	/* set PENDBASE */
3164	val = (page_to_phys(pend_page) |
3165	       GICR_PENDBASER_InnerShareable |
3166	       GICR_PENDBASER_RaWaWb);
3167
3168	gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
3169	tmp = gicr_read_pendbaser(rbase + GICR_PENDBASER);
3170
3171	if (!rdists_support_shareable())
3172		tmp &= ~GICR_PENDBASER_SHAREABILITY_MASK;
3173
3174	if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
3175		/*
3176		 * The HW reports non-shareable, we must remove the
3177		 * cacheability attributes as well.
3178		 */
3179		val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
3180			 GICR_PENDBASER_CACHEABILITY_MASK);
3181		val |= GICR_PENDBASER_nC;
3182		gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
3183	}
3184
3185	/* Enable LPIs */
3186	val = readl_relaxed(rbase + GICR_CTLR);
3187	val |= GICR_CTLR_ENABLE_LPIS;
3188	writel_relaxed(val, rbase + GICR_CTLR);
3189
3190out:
3191	if (gic_rdists->has_vlpis && !gic_rdists->has_rvpeid) {
3192		void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
3193
3194		/*
3195		 * It's possible for CPU to receive VLPIs before it is
3196		 * scheduled as a vPE, especially for the first CPU, and the
3197		 * VLPI with INTID larger than 2^(IDbits+1) will be considered
3198		 * as out of range and dropped by GIC.
3199		 * So we initialize IDbits to known value to avoid VLPI drop.
3200		 */
3201		val = (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
3202		pr_debug("GICv4: CPU%d: Init IDbits to 0x%llx for GICR_VPROPBASER\n",
3203			smp_processor_id(), val);
3204		gicr_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
3205
3206		/*
3207		 * Also clear Valid bit of GICR_VPENDBASER, in case some
3208		 * ancient programming gets left in and has possibility of
3209		 * corrupting memory.
3210		 */
3211		val = its_clear_vpend_valid(vlpi_base, 0, 0);
3212	}
3213
3214	if (allocate_vpe_l1_table()) {
3215		/*
3216		 * If the allocation has failed, we're in massive trouble.
3217		 * Disable direct injection, and pray that no VM was
3218		 * already running...
3219		 */
3220		gic_rdists->has_rvpeid = false;
3221		gic_rdists->has_vlpis = false;
3222	}
3223
3224	/* Make sure the GIC has seen the above */
3225	dsb(sy);
3226	gic_data_rdist()->flags |= RD_LOCAL_LPI_ENABLED;
3227	pr_info("GICv3: CPU%d: using %s LPI pending table @%pa\n",
3228		smp_processor_id(),
3229		gic_data_rdist()->flags & RD_LOCAL_PENDTABLE_PREALLOCATED ?
3230		"reserved" : "allocated",
3231		&paddr);
3232}
3233
3234static void its_cpu_init_collection(struct its_node *its)
3235{
3236	int cpu = smp_processor_id();
3237	u64 target;
3238
3239	/* avoid cross node collections and its mapping */
3240	if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) {
3241		struct device_node *cpu_node;
3242
3243		cpu_node = of_get_cpu_node(cpu, NULL);
3244		if (its->numa_node != NUMA_NO_NODE &&
3245			its->numa_node != of_node_to_nid(cpu_node))
3246			return;
3247	}
3248
3249	/*
3250	 * We now have to bind each collection to its target
3251	 * redistributor.
3252	 */
3253	if (gic_read_typer(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
3254		/*
3255		 * This ITS wants the physical address of the
3256		 * redistributor.
3257		 */
3258		target = gic_data_rdist()->phys_base;
3259	} else {
3260		/* This ITS wants a linear CPU number. */
3261		target = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
3262		target = GICR_TYPER_CPU_NUMBER(target) << 16;
3263	}
 
 
 
 
 
 
 
3264
3265	/* Perform collection mapping */
3266	its->collections[cpu].target_address = target;
3267	its->collections[cpu].col_id = cpu;
3268
3269	its_send_mapc(its, &its->collections[cpu], 1);
3270	its_send_invall(its, &its->collections[cpu]);
3271}
3272
3273static void its_cpu_init_collections(void)
3274{
3275	struct its_node *its;
3276
3277	raw_spin_lock(&its_lock);
3278
3279	list_for_each_entry(its, &its_nodes, entry)
3280		its_cpu_init_collection(its);
3281
3282	raw_spin_unlock(&its_lock);
3283}
3284
3285static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
3286{
3287	struct its_device *its_dev = NULL, *tmp;
3288	unsigned long flags;
3289
3290	raw_spin_lock_irqsave(&its->lock, flags);
3291
3292	list_for_each_entry(tmp, &its->its_device_list, entry) {
3293		if (tmp->device_id == dev_id) {
3294			its_dev = tmp;
3295			break;
3296		}
3297	}
3298
3299	raw_spin_unlock_irqrestore(&its->lock, flags);
3300
3301	return its_dev;
3302}
3303
3304static struct its_baser *its_get_baser(struct its_node *its, u32 type)
3305{
3306	int i;
3307
3308	for (i = 0; i < GITS_BASER_NR_REGS; i++) {
3309		if (GITS_BASER_TYPE(its->tables[i].val) == type)
3310			return &its->tables[i];
3311	}
3312
3313	return NULL;
3314}
3315
3316static bool its_alloc_table_entry(struct its_node *its,
3317				  struct its_baser *baser, u32 id)
3318{
3319	struct page *page;
3320	u32 esz, idx;
3321	__le64 *table;
3322
3323	/* Don't allow device id that exceeds single, flat table limit */
3324	esz = GITS_BASER_ENTRY_SIZE(baser->val);
3325	if (!(baser->val & GITS_BASER_INDIRECT))
3326		return (id < (PAGE_ORDER_TO_SIZE(baser->order) / esz));
3327
3328	/* Compute 1st level table index & check if that exceeds table limit */
3329	idx = id >> ilog2(baser->psz / esz);
3330	if (idx >= (PAGE_ORDER_TO_SIZE(baser->order) / GITS_LVL1_ENTRY_SIZE))
3331		return false;
3332
3333	table = baser->base;
3334
3335	/* Allocate memory for 2nd level table */
3336	if (!table[idx]) {
3337		page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO,
3338					get_order(baser->psz));
3339		if (!page)
3340			return false;
3341
3342		/* Flush Lvl2 table to PoC if hw doesn't support coherency */
3343		if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
3344			gic_flush_dcache_to_poc(page_address(page), baser->psz);
3345
3346		table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
3347
3348		/* Flush Lvl1 entry to PoC if hw doesn't support coherency */
3349		if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
3350			gic_flush_dcache_to_poc(table + idx, GITS_LVL1_ENTRY_SIZE);
3351
3352		/* Ensure updated table contents are visible to ITS hardware */
3353		dsb(sy);
3354	}
3355
3356	return true;
3357}
3358
3359static bool its_alloc_device_table(struct its_node *its, u32 dev_id)
3360{
3361	struct its_baser *baser;
3362
3363	baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
3364
3365	/* Don't allow device id that exceeds ITS hardware limit */
3366	if (!baser)
3367		return (ilog2(dev_id) < device_ids(its));
3368
3369	return its_alloc_table_entry(its, baser, dev_id);
3370}
3371
3372static bool its_alloc_vpe_table(u32 vpe_id)
3373{
3374	struct its_node *its;
3375	int cpu;
3376
3377	/*
3378	 * Make sure the L2 tables are allocated on *all* v4 ITSs. We
3379	 * could try and only do it on ITSs corresponding to devices
3380	 * that have interrupts targeted at this VPE, but the
3381	 * complexity becomes crazy (and you have tons of memory
3382	 * anyway, right?).
3383	 */
3384	list_for_each_entry(its, &its_nodes, entry) {
3385		struct its_baser *baser;
3386
3387		if (!is_v4(its))
3388			continue;
3389
3390		baser = its_get_baser(its, GITS_BASER_TYPE_VCPU);
3391		if (!baser)
3392			return false;
3393
3394		if (!its_alloc_table_entry(its, baser, vpe_id))
3395			return false;
3396	}
3397
3398	/* Non v4.1? No need to iterate RDs and go back early. */
3399	if (!gic_rdists->has_rvpeid)
3400		return true;
3401
3402	/*
3403	 * Make sure the L2 tables are allocated for all copies of
3404	 * the L1 table on *all* v4.1 RDs.
3405	 */
3406	for_each_possible_cpu(cpu) {
3407		if (!allocate_vpe_l2_table(cpu, vpe_id))
3408			return false;
3409	}
3410
3411	return true;
3412}
3413
3414static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
3415					    int nvecs, bool alloc_lpis)
3416{
3417	struct its_device *dev;
3418	unsigned long *lpi_map = NULL;
3419	unsigned long flags;
3420	u16 *col_map = NULL;
3421	void *itt;
3422	int lpi_base;
3423	int nr_lpis;
3424	int nr_ites;
3425	int sz;
3426
3427	if (!its_alloc_device_table(its, dev_id))
3428		return NULL;
3429
3430	if (WARN_ON(!is_power_of_2(nvecs)))
3431		nvecs = roundup_pow_of_two(nvecs);
3432
3433	dev = kzalloc(sizeof(*dev), GFP_KERNEL);
3434	/*
3435	 * Even if the device wants a single LPI, the ITT must be
3436	 * sized as a power of two (and you need at least one bit...).
 
3437	 */
3438	nr_ites = max(2, nvecs);
3439	sz = nr_ites * (FIELD_GET(GITS_TYPER_ITT_ENTRY_SIZE, its->typer) + 1);
3440	sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
3441	itt = kzalloc_node(sz, GFP_KERNEL, its->numa_node);
3442	if (alloc_lpis) {
3443		lpi_map = its_lpi_alloc(nvecs, &lpi_base, &nr_lpis);
3444		if (lpi_map)
3445			col_map = kcalloc(nr_lpis, sizeof(*col_map),
3446					  GFP_KERNEL);
3447	} else {
3448		col_map = kcalloc(nr_ites, sizeof(*col_map), GFP_KERNEL);
3449		nr_lpis = 0;
3450		lpi_base = 0;
3451	}
3452
3453	if (!dev || !itt ||  !col_map || (!lpi_map && alloc_lpis)) {
3454		kfree(dev);
3455		kfree(itt);
3456		bitmap_free(lpi_map);
3457		kfree(col_map);
3458		return NULL;
3459	}
3460
3461	gic_flush_dcache_to_poc(itt, sz);
3462
3463	dev->its = its;
3464	dev->itt = itt;
3465	dev->nr_ites = nr_ites;
3466	dev->event_map.lpi_map = lpi_map;
3467	dev->event_map.col_map = col_map;
3468	dev->event_map.lpi_base = lpi_base;
3469	dev->event_map.nr_lpis = nr_lpis;
3470	raw_spin_lock_init(&dev->event_map.vlpi_lock);
3471	dev->device_id = dev_id;
3472	INIT_LIST_HEAD(&dev->entry);
3473
3474	raw_spin_lock_irqsave(&its->lock, flags);
3475	list_add(&dev->entry, &its->its_device_list);
3476	raw_spin_unlock_irqrestore(&its->lock, flags);
3477
3478	/* Map device to its ITT */
3479	its_send_mapd(dev, 1);
3480
3481	return dev;
3482}
3483
3484static void its_free_device(struct its_device *its_dev)
3485{
3486	unsigned long flags;
3487
3488	raw_spin_lock_irqsave(&its_dev->its->lock, flags);
3489	list_del(&its_dev->entry);
3490	raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
3491	kfree(its_dev->event_map.col_map);
3492	kfree(its_dev->itt);
3493	kfree(its_dev);
3494}
3495
3496static int its_alloc_device_irq(struct its_device *dev, int nvecs, irq_hw_number_t *hwirq)
3497{
3498	int idx;
3499
3500	/* Find a free LPI region in lpi_map and allocate them. */
3501	idx = bitmap_find_free_region(dev->event_map.lpi_map,
3502				      dev->event_map.nr_lpis,
3503				      get_count_order(nvecs));
3504	if (idx < 0)
3505		return -ENOSPC;
3506
3507	*hwirq = dev->event_map.lpi_base + idx;
 
3508
3509	return 0;
3510}
3511
3512static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
3513			   int nvec, msi_alloc_info_t *info)
3514{
3515	struct its_node *its;
3516	struct its_device *its_dev;
3517	struct msi_domain_info *msi_info;
3518	u32 dev_id;
3519	int err = 0;
3520
3521	/*
3522	 * We ignore "dev" entirely, and rely on the dev_id that has
3523	 * been passed via the scratchpad. This limits this domain's
3524	 * usefulness to upper layers that definitely know that they
3525	 * are built on top of the ITS.
3526	 */
3527	dev_id = info->scratchpad[0].ul;
3528
3529	msi_info = msi_get_domain_info(domain);
3530	its = msi_info->data;
3531
3532	if (!gic_rdists->has_direct_lpi &&
3533	    vpe_proxy.dev &&
3534	    vpe_proxy.dev->its == its &&
3535	    dev_id == vpe_proxy.dev->device_id) {
3536		/* Bad luck. Get yourself a better implementation */
3537		WARN_ONCE(1, "DevId %x clashes with GICv4 VPE proxy device\n",
3538			  dev_id);
3539		return -EINVAL;
3540	}
3541
3542	mutex_lock(&its->dev_alloc_lock);
3543	its_dev = its_find_device(its, dev_id);
3544	if (its_dev) {
3545		/*
3546		 * We already have seen this ID, probably through
3547		 * another alias (PCI bridge of some sort). No need to
3548		 * create the device.
3549		 */
3550		its_dev->shared = true;
3551		pr_debug("Reusing ITT for devID %x\n", dev_id);
3552		goto out;
3553	}
3554
3555	its_dev = its_create_device(its, dev_id, nvec, true);
3556	if (!its_dev) {
3557		err = -ENOMEM;
3558		goto out;
3559	}
3560
3561	if (info->flags & MSI_ALLOC_FLAGS_PROXY_DEVICE)
3562		its_dev->shared = true;
3563
3564	pr_debug("ITT %d entries, %d bits\n", nvec, ilog2(nvec));
3565out:
3566	mutex_unlock(&its->dev_alloc_lock);
3567	info->scratchpad[0].ptr = its_dev;
3568	return err;
3569}
3570
3571static struct msi_domain_ops its_msi_domain_ops = {
3572	.msi_prepare	= its_msi_prepare,
3573};
3574
3575static int its_irq_gic_domain_alloc(struct irq_domain *domain,
3576				    unsigned int virq,
3577				    irq_hw_number_t hwirq)
3578{
3579	struct irq_fwspec fwspec;
3580
3581	if (irq_domain_get_of_node(domain->parent)) {
3582		fwspec.fwnode = domain->parent->fwnode;
3583		fwspec.param_count = 3;
3584		fwspec.param[0] = GIC_IRQ_TYPE_LPI;
3585		fwspec.param[1] = hwirq;
3586		fwspec.param[2] = IRQ_TYPE_EDGE_RISING;
3587	} else if (is_fwnode_irqchip(domain->parent->fwnode)) {
3588		fwspec.fwnode = domain->parent->fwnode;
3589		fwspec.param_count = 2;
3590		fwspec.param[0] = hwirq;
3591		fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
3592	} else {
3593		return -EINVAL;
3594	}
3595
3596	return irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
3597}
3598
3599static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
3600				unsigned int nr_irqs, void *args)
3601{
3602	msi_alloc_info_t *info = args;
3603	struct its_device *its_dev = info->scratchpad[0].ptr;
3604	struct its_node *its = its_dev->its;
3605	struct irq_data *irqd;
3606	irq_hw_number_t hwirq;
3607	int err;
3608	int i;
3609
3610	err = its_alloc_device_irq(its_dev, nr_irqs, &hwirq);
3611	if (err)
3612		return err;
3613
3614	err = iommu_dma_prepare_msi(info->desc, its->get_msi_base(its_dev));
3615	if (err)
3616		return err;
3617
3618	for (i = 0; i < nr_irqs; i++) {
3619		err = its_irq_gic_domain_alloc(domain, virq + i, hwirq + i);
3620		if (err)
3621			return err;
3622
3623		irq_domain_set_hwirq_and_chip(domain, virq + i,
3624					      hwirq + i, &its_irq_chip, its_dev);
3625		irqd = irq_get_irq_data(virq + i);
3626		irqd_set_single_target(irqd);
3627		irqd_set_affinity_on_activate(irqd);
3628		irqd_set_resend_when_in_progress(irqd);
3629		pr_debug("ID:%d pID:%d vID:%d\n",
3630			 (int)(hwirq + i - its_dev->event_map.lpi_base),
3631			 (int)(hwirq + i), virq + i);
3632	}
3633
3634	return 0;
3635}
3636
3637static int its_irq_domain_activate(struct irq_domain *domain,
3638				   struct irq_data *d, bool reserve)
3639{
3640	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
3641	u32 event = its_get_event_id(d);
3642	int cpu;
3643
3644	cpu = its_select_cpu(d, cpu_online_mask);
3645	if (cpu < 0 || cpu >= nr_cpu_ids)
3646		return -EINVAL;
3647
3648	its_inc_lpi_count(d, cpu);
3649	its_dev->event_map.col_map[event] = cpu;
3650	irq_data_update_effective_affinity(d, cpumask_of(cpu));
3651
3652	/* Map the GIC IRQ and event to the device */
3653	its_send_mapti(its_dev, d->hwirq, event);
3654	return 0;
3655}
3656
3657static void its_irq_domain_deactivate(struct irq_domain *domain,
3658				      struct irq_data *d)
3659{
3660	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
3661	u32 event = its_get_event_id(d);
3662
3663	its_dec_lpi_count(d, its_dev->event_map.col_map[event]);
3664	/* Stop the delivery of interrupts */
3665	its_send_discard(its_dev, event);
3666}
3667
3668static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
3669				unsigned int nr_irqs)
3670{
3671	struct irq_data *d = irq_domain_get_irq_data(domain, virq);
3672	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
3673	struct its_node *its = its_dev->its;
3674	int i;
3675
3676	bitmap_release_region(its_dev->event_map.lpi_map,
3677			      its_get_event_id(irq_domain_get_irq_data(domain, virq)),
3678			      get_count_order(nr_irqs));
3679
3680	for (i = 0; i < nr_irqs; i++) {
3681		struct irq_data *data = irq_domain_get_irq_data(domain,
3682								virq + i);
 
 
 
 
 
3683		/* Nuke the entry in the domain */
3684		irq_domain_reset_irq_data(data);
3685	}
3686
3687	mutex_lock(&its->dev_alloc_lock);
3688
3689	/*
3690	 * If all interrupts have been freed, start mopping the
3691	 * floor. This is conditioned on the device not being shared.
3692	 */
3693	if (!its_dev->shared &&
3694	    bitmap_empty(its_dev->event_map.lpi_map,
3695			 its_dev->event_map.nr_lpis)) {
3696		its_lpi_free(its_dev->event_map.lpi_map,
3697			     its_dev->event_map.lpi_base,
3698			     its_dev->event_map.nr_lpis);
3699
3700		/* Unmap device/itt */
3701		its_send_mapd(its_dev, 0);
3702		its_free_device(its_dev);
3703	}
3704
3705	mutex_unlock(&its->dev_alloc_lock);
3706
3707	irq_domain_free_irqs_parent(domain, virq, nr_irqs);
3708}
3709
3710static const struct irq_domain_ops its_domain_ops = {
3711	.alloc			= its_irq_domain_alloc,
3712	.free			= its_irq_domain_free,
3713	.activate		= its_irq_domain_activate,
3714	.deactivate		= its_irq_domain_deactivate,
3715};
3716
3717/*
3718 * This is insane.
3719 *
3720 * If a GICv4.0 doesn't implement Direct LPIs (which is extremely
3721 * likely), the only way to perform an invalidate is to use a fake
3722 * device to issue an INV command, implying that the LPI has first
3723 * been mapped to some event on that device. Since this is not exactly
3724 * cheap, we try to keep that mapping around as long as possible, and
3725 * only issue an UNMAP if we're short on available slots.
3726 *
3727 * Broken by design(tm).
3728 *
3729 * GICv4.1, on the other hand, mandates that we're able to invalidate
3730 * by writing to a MMIO register. It doesn't implement the whole of
3731 * DirectLPI, but that's good enough. And most of the time, we don't
3732 * even have to invalidate anything, as the redistributor can be told
3733 * whether to generate a doorbell or not (we thus leave it enabled,
3734 * always).
3735 */
3736static void its_vpe_db_proxy_unmap_locked(struct its_vpe *vpe)
3737{
3738	/* GICv4.1 doesn't use a proxy, so nothing to do here */
3739	if (gic_rdists->has_rvpeid)
3740		return;
3741
3742	/* Already unmapped? */
3743	if (vpe->vpe_proxy_event == -1)
3744		return;
3745
3746	its_send_discard(vpe_proxy.dev, vpe->vpe_proxy_event);
3747	vpe_proxy.vpes[vpe->vpe_proxy_event] = NULL;
3748
3749	/*
3750	 * We don't track empty slots at all, so let's move the
3751	 * next_victim pointer if we can quickly reuse that slot
3752	 * instead of nuking an existing entry. Not clear that this is
3753	 * always a win though, and this might just generate a ripple
3754	 * effect... Let's just hope VPEs don't migrate too often.
3755	 */
3756	if (vpe_proxy.vpes[vpe_proxy.next_victim])
3757		vpe_proxy.next_victim = vpe->vpe_proxy_event;
3758
3759	vpe->vpe_proxy_event = -1;
3760}
3761
3762static void its_vpe_db_proxy_unmap(struct its_vpe *vpe)
3763{
3764	/* GICv4.1 doesn't use a proxy, so nothing to do here */
3765	if (gic_rdists->has_rvpeid)
3766		return;
3767
3768	if (!gic_rdists->has_direct_lpi) {
3769		unsigned long flags;
3770
3771		raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
3772		its_vpe_db_proxy_unmap_locked(vpe);
3773		raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
3774	}
3775}
3776
3777static void its_vpe_db_proxy_map_locked(struct its_vpe *vpe)
3778{
3779	/* GICv4.1 doesn't use a proxy, so nothing to do here */
3780	if (gic_rdists->has_rvpeid)
3781		return;
3782
3783	/* Already mapped? */
3784	if (vpe->vpe_proxy_event != -1)
3785		return;
3786
3787	/* This slot was already allocated. Kick the other VPE out. */
3788	if (vpe_proxy.vpes[vpe_proxy.next_victim])
3789		its_vpe_db_proxy_unmap_locked(vpe_proxy.vpes[vpe_proxy.next_victim]);
3790
3791	/* Map the new VPE instead */
3792	vpe_proxy.vpes[vpe_proxy.next_victim] = vpe;
3793	vpe->vpe_proxy_event = vpe_proxy.next_victim;
3794	vpe_proxy.next_victim = (vpe_proxy.next_victim + 1) % vpe_proxy.dev->nr_ites;
3795
3796	vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = vpe->col_idx;
3797	its_send_mapti(vpe_proxy.dev, vpe->vpe_db_lpi, vpe->vpe_proxy_event);
3798}
3799
3800static void its_vpe_db_proxy_move(struct its_vpe *vpe, int from, int to)
3801{
3802	unsigned long flags;
3803	struct its_collection *target_col;
3804
3805	/* GICv4.1 doesn't use a proxy, so nothing to do here */
3806	if (gic_rdists->has_rvpeid)
3807		return;
3808
3809	if (gic_rdists->has_direct_lpi) {
3810		void __iomem *rdbase;
3811
3812		rdbase = per_cpu_ptr(gic_rdists->rdist, from)->rd_base;
3813		gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
3814		wait_for_syncr(rdbase);
3815
3816		return;
3817	}
3818
3819	raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
3820
3821	its_vpe_db_proxy_map_locked(vpe);
3822
3823	target_col = &vpe_proxy.dev->its->collections[to];
3824	its_send_movi(vpe_proxy.dev, target_col, vpe->vpe_proxy_event);
3825	vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = to;
3826
3827	raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
3828}
3829
3830static int its_vpe_set_affinity(struct irq_data *d,
3831				const struct cpumask *mask_val,
3832				bool force)
3833{
3834	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3835	struct cpumask common, *table_mask;
3836	unsigned long flags;
3837	int from, cpu;
3838
3839	/*
3840	 * Changing affinity is mega expensive, so let's be as lazy as
3841	 * we can and only do it if we really have to. Also, if mapped
3842	 * into the proxy device, we need to move the doorbell
3843	 * interrupt to its new location.
3844	 *
3845	 * Another thing is that changing the affinity of a vPE affects
3846	 * *other interrupts* such as all the vLPIs that are routed to
3847	 * this vPE. This means that the irq_desc lock is not enough to
3848	 * protect us, and that we must ensure nobody samples vpe->col_idx
3849	 * during the update, hence the lock below which must also be
3850	 * taken on any vLPI handling path that evaluates vpe->col_idx.
3851	 */
3852	from = vpe_to_cpuid_lock(vpe, &flags);
3853	table_mask = gic_data_rdist_cpu(from)->vpe_table_mask;
3854
3855	/*
3856	 * If we are offered another CPU in the same GICv4.1 ITS
3857	 * affinity, pick this one. Otherwise, any CPU will do.
3858	 */
3859	if (table_mask && cpumask_and(&common, mask_val, table_mask))
3860		cpu = cpumask_test_cpu(from, &common) ? from : cpumask_first(&common);
3861	else
3862		cpu = cpumask_first(mask_val);
3863
3864	if (from == cpu)
3865		goto out;
3866
3867	vpe->col_idx = cpu;
3868
3869	its_send_vmovp(vpe);
3870	its_vpe_db_proxy_move(vpe, from, cpu);
3871
3872out:
3873	irq_data_update_effective_affinity(d, cpumask_of(cpu));
3874	vpe_to_cpuid_unlock(vpe, flags);
3875
3876	return IRQ_SET_MASK_OK_DONE;
3877}
3878
3879static void its_wait_vpt_parse_complete(void)
3880{
3881	void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
3882	u64 val;
3883
3884	if (!gic_rdists->has_vpend_valid_dirty)
3885		return;
3886
3887	WARN_ON_ONCE(readq_relaxed_poll_timeout_atomic(vlpi_base + GICR_VPENDBASER,
3888						       val,
3889						       !(val & GICR_VPENDBASER_Dirty),
3890						       1, 500));
3891}
3892
3893static void its_vpe_schedule(struct its_vpe *vpe)
3894{
3895	void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
3896	u64 val;
3897
3898	/* Schedule the VPE */
3899	val  = virt_to_phys(page_address(vpe->its_vm->vprop_page)) &
3900		GENMASK_ULL(51, 12);
3901	val |= (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
3902	if (rdists_support_shareable()) {
3903		val |= GICR_VPROPBASER_RaWb;
3904		val |= GICR_VPROPBASER_InnerShareable;
3905	}
3906	gicr_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
3907
3908	val  = virt_to_phys(page_address(vpe->vpt_page)) &
3909		GENMASK_ULL(51, 16);
3910	if (rdists_support_shareable()) {
3911		val |= GICR_VPENDBASER_RaWaWb;
3912		val |= GICR_VPENDBASER_InnerShareable;
3913	}
3914	/*
3915	 * There is no good way of finding out if the pending table is
3916	 * empty as we can race against the doorbell interrupt very
3917	 * easily. So in the end, vpe->pending_last is only an
3918	 * indication that the vcpu has something pending, not one
3919	 * that the pending table is empty. A good implementation
3920	 * would be able to read its coarse map pretty quickly anyway,
3921	 * making this a tolerable issue.
3922	 */
3923	val |= GICR_VPENDBASER_PendingLast;
3924	val |= vpe->idai ? GICR_VPENDBASER_IDAI : 0;
3925	val |= GICR_VPENDBASER_Valid;
3926	gicr_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
3927}
3928
3929static void its_vpe_deschedule(struct its_vpe *vpe)
3930{
3931	void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
3932	u64 val;
3933
3934	val = its_clear_vpend_valid(vlpi_base, 0, 0);
3935
3936	vpe->idai = !!(val & GICR_VPENDBASER_IDAI);
3937	vpe->pending_last = !!(val & GICR_VPENDBASER_PendingLast);
3938}
3939
3940static void its_vpe_invall(struct its_vpe *vpe)
3941{
3942	struct its_node *its;
3943
3944	list_for_each_entry(its, &its_nodes, entry) {
3945		if (!is_v4(its))
3946			continue;
3947
3948		if (its_list_map && !vpe->its_vm->vlpi_count[its->list_nr])
3949			continue;
3950
3951		/*
3952		 * Sending a VINVALL to a single ITS is enough, as all
3953		 * we need is to reach the redistributors.
3954		 */
3955		its_send_vinvall(its, vpe);
3956		return;
3957	}
3958}
3959
3960static int its_vpe_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
3961{
3962	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3963	struct its_cmd_info *info = vcpu_info;
3964
3965	switch (info->cmd_type) {
3966	case SCHEDULE_VPE:
3967		its_vpe_schedule(vpe);
3968		return 0;
3969
3970	case DESCHEDULE_VPE:
3971		its_vpe_deschedule(vpe);
3972		return 0;
3973
3974	case COMMIT_VPE:
3975		its_wait_vpt_parse_complete();
3976		return 0;
3977
3978	case INVALL_VPE:
3979		its_vpe_invall(vpe);
3980		return 0;
3981
3982	default:
3983		return -EINVAL;
3984	}
3985}
3986
3987static void its_vpe_send_cmd(struct its_vpe *vpe,
3988			     void (*cmd)(struct its_device *, u32))
3989{
3990	unsigned long flags;
3991
3992	raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
3993
3994	its_vpe_db_proxy_map_locked(vpe);
3995	cmd(vpe_proxy.dev, vpe->vpe_proxy_event);
3996
3997	raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
3998}
3999
4000static void its_vpe_send_inv(struct irq_data *d)
4001{
4002	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4003
4004	if (gic_rdists->has_direct_lpi)
4005		__direct_lpi_inv(d, d->parent_data->hwirq);
4006	else
4007		its_vpe_send_cmd(vpe, its_send_inv);
4008}
4009
4010static void its_vpe_mask_irq(struct irq_data *d)
4011{
4012	/*
4013	 * We need to unmask the LPI, which is described by the parent
4014	 * irq_data. Instead of calling into the parent (which won't
4015	 * exactly do the right thing, let's simply use the
4016	 * parent_data pointer. Yes, I'm naughty.
4017	 */
4018	lpi_write_config(d->parent_data, LPI_PROP_ENABLED, 0);
4019	its_vpe_send_inv(d);
4020}
4021
4022static void its_vpe_unmask_irq(struct irq_data *d)
4023{
4024	/* Same hack as above... */
4025	lpi_write_config(d->parent_data, 0, LPI_PROP_ENABLED);
4026	its_vpe_send_inv(d);
4027}
4028
4029static int its_vpe_set_irqchip_state(struct irq_data *d,
4030				     enum irqchip_irq_state which,
4031				     bool state)
4032{
4033	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4034
4035	if (which != IRQCHIP_STATE_PENDING)
4036		return -EINVAL;
4037
4038	if (gic_rdists->has_direct_lpi) {
4039		void __iomem *rdbase;
4040
4041		rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
4042		if (state) {
4043			gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_SETLPIR);
4044		} else {
4045			gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
4046			wait_for_syncr(rdbase);
4047		}
4048	} else {
4049		if (state)
4050			its_vpe_send_cmd(vpe, its_send_int);
4051		else
4052			its_vpe_send_cmd(vpe, its_send_clear);
4053	}
4054
4055	return 0;
4056}
4057
4058static int its_vpe_retrigger(struct irq_data *d)
4059{
4060	return !its_vpe_set_irqchip_state(d, IRQCHIP_STATE_PENDING, true);
4061}
4062
4063static struct irq_chip its_vpe_irq_chip = {
4064	.name			= "GICv4-vpe",
4065	.irq_mask		= its_vpe_mask_irq,
4066	.irq_unmask		= its_vpe_unmask_irq,
4067	.irq_eoi		= irq_chip_eoi_parent,
4068	.irq_set_affinity	= its_vpe_set_affinity,
4069	.irq_retrigger		= its_vpe_retrigger,
4070	.irq_set_irqchip_state	= its_vpe_set_irqchip_state,
4071	.irq_set_vcpu_affinity	= its_vpe_set_vcpu_affinity,
4072};
4073
4074static struct its_node *find_4_1_its(void)
4075{
4076	static struct its_node *its = NULL;
4077
4078	if (!its) {
4079		list_for_each_entry(its, &its_nodes, entry) {
4080			if (is_v4_1(its))
4081				return its;
4082		}
4083
4084		/* Oops? */
4085		its = NULL;
4086	}
4087
4088	return its;
4089}
4090
4091static void its_vpe_4_1_send_inv(struct irq_data *d)
4092{
4093	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4094	struct its_node *its;
4095
4096	/*
4097	 * GICv4.1 wants doorbells to be invalidated using the
4098	 * INVDB command in order to be broadcast to all RDs. Send
4099	 * it to the first valid ITS, and let the HW do its magic.
4100	 */
4101	its = find_4_1_its();
4102	if (its)
4103		its_send_invdb(its, vpe);
4104}
4105
4106static void its_vpe_4_1_mask_irq(struct irq_data *d)
4107{
4108	lpi_write_config(d->parent_data, LPI_PROP_ENABLED, 0);
4109	its_vpe_4_1_send_inv(d);
4110}
4111
4112static void its_vpe_4_1_unmask_irq(struct irq_data *d)
4113{
4114	lpi_write_config(d->parent_data, 0, LPI_PROP_ENABLED);
4115	its_vpe_4_1_send_inv(d);
4116}
4117
4118static void its_vpe_4_1_schedule(struct its_vpe *vpe,
4119				 struct its_cmd_info *info)
4120{
4121	void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
4122	u64 val = 0;
4123
4124	/* Schedule the VPE */
4125	val |= GICR_VPENDBASER_Valid;
4126	val |= info->g0en ? GICR_VPENDBASER_4_1_VGRP0EN : 0;
4127	val |= info->g1en ? GICR_VPENDBASER_4_1_VGRP1EN : 0;
4128	val |= FIELD_PREP(GICR_VPENDBASER_4_1_VPEID, vpe->vpe_id);
4129
4130	gicr_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
4131}
4132
4133static void its_vpe_4_1_deschedule(struct its_vpe *vpe,
4134				   struct its_cmd_info *info)
4135{
4136	void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
4137	u64 val;
4138
4139	if (info->req_db) {
4140		unsigned long flags;
4141
4142		/*
4143		 * vPE is going to block: make the vPE non-resident with
4144		 * PendingLast clear and DB set. The GIC guarantees that if
4145		 * we read-back PendingLast clear, then a doorbell will be
4146		 * delivered when an interrupt comes.
4147		 *
4148		 * Note the locking to deal with the concurrent update of
4149		 * pending_last from the doorbell interrupt handler that can
4150		 * run concurrently.
4151		 */
4152		raw_spin_lock_irqsave(&vpe->vpe_lock, flags);
4153		val = its_clear_vpend_valid(vlpi_base,
4154					    GICR_VPENDBASER_PendingLast,
4155					    GICR_VPENDBASER_4_1_DB);
4156		vpe->pending_last = !!(val & GICR_VPENDBASER_PendingLast);
4157		raw_spin_unlock_irqrestore(&vpe->vpe_lock, flags);
4158	} else {
4159		/*
4160		 * We're not blocking, so just make the vPE non-resident
4161		 * with PendingLast set, indicating that we'll be back.
4162		 */
4163		val = its_clear_vpend_valid(vlpi_base,
4164					    0,
4165					    GICR_VPENDBASER_PendingLast);
4166		vpe->pending_last = true;
4167	}
4168}
4169
4170static void its_vpe_4_1_invall(struct its_vpe *vpe)
4171{
4172	void __iomem *rdbase;
4173	unsigned long flags;
4174	u64 val;
4175	int cpu;
4176
4177	val  = GICR_INVALLR_V;
4178	val |= FIELD_PREP(GICR_INVALLR_VPEID, vpe->vpe_id);
4179
4180	/* Target the redistributor this vPE is currently known on */
4181	cpu = vpe_to_cpuid_lock(vpe, &flags);
4182	raw_spin_lock(&gic_data_rdist_cpu(cpu)->rd_lock);
4183	rdbase = per_cpu_ptr(gic_rdists->rdist, cpu)->rd_base;
4184	gic_write_lpir(val, rdbase + GICR_INVALLR);
4185
4186	wait_for_syncr(rdbase);
4187	raw_spin_unlock(&gic_data_rdist_cpu(cpu)->rd_lock);
4188	vpe_to_cpuid_unlock(vpe, flags);
4189}
4190
4191static int its_vpe_4_1_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
4192{
4193	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4194	struct its_cmd_info *info = vcpu_info;
4195
4196	switch (info->cmd_type) {
4197	case SCHEDULE_VPE:
4198		its_vpe_4_1_schedule(vpe, info);
4199		return 0;
4200
4201	case DESCHEDULE_VPE:
4202		its_vpe_4_1_deschedule(vpe, info);
4203		return 0;
4204
4205	case COMMIT_VPE:
4206		its_wait_vpt_parse_complete();
4207		return 0;
4208
4209	case INVALL_VPE:
4210		its_vpe_4_1_invall(vpe);
4211		return 0;
4212
4213	default:
4214		return -EINVAL;
4215	}
4216}
4217
4218static struct irq_chip its_vpe_4_1_irq_chip = {
4219	.name			= "GICv4.1-vpe",
4220	.irq_mask		= its_vpe_4_1_mask_irq,
4221	.irq_unmask		= its_vpe_4_1_unmask_irq,
4222	.irq_eoi		= irq_chip_eoi_parent,
4223	.irq_set_affinity	= its_vpe_set_affinity,
4224	.irq_set_vcpu_affinity	= its_vpe_4_1_set_vcpu_affinity,
4225};
4226
4227static void its_configure_sgi(struct irq_data *d, bool clear)
4228{
4229	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4230	struct its_cmd_desc desc;
4231
4232	desc.its_vsgi_cmd.vpe = vpe;
4233	desc.its_vsgi_cmd.sgi = d->hwirq;
4234	desc.its_vsgi_cmd.priority = vpe->sgi_config[d->hwirq].priority;
4235	desc.its_vsgi_cmd.enable = vpe->sgi_config[d->hwirq].enabled;
4236	desc.its_vsgi_cmd.group = vpe->sgi_config[d->hwirq].group;
4237	desc.its_vsgi_cmd.clear = clear;
4238
4239	/*
4240	 * GICv4.1 allows us to send VSGI commands to any ITS as long as the
4241	 * destination VPE is mapped there. Since we map them eagerly at
4242	 * activation time, we're pretty sure the first GICv4.1 ITS will do.
4243	 */
4244	its_send_single_vcommand(find_4_1_its(), its_build_vsgi_cmd, &desc);
4245}
4246
4247static void its_sgi_mask_irq(struct irq_data *d)
4248{
4249	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4250
4251	vpe->sgi_config[d->hwirq].enabled = false;
4252	its_configure_sgi(d, false);
4253}
4254
4255static void its_sgi_unmask_irq(struct irq_data *d)
4256{
4257	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4258
4259	vpe->sgi_config[d->hwirq].enabled = true;
4260	its_configure_sgi(d, false);
4261}
4262
4263static int its_sgi_set_affinity(struct irq_data *d,
4264				const struct cpumask *mask_val,
4265				bool force)
4266{
4267	/*
4268	 * There is no notion of affinity for virtual SGIs, at least
4269	 * not on the host (since they can only be targeting a vPE).
4270	 * Tell the kernel we've done whatever it asked for.
4271	 */
4272	irq_data_update_effective_affinity(d, mask_val);
4273	return IRQ_SET_MASK_OK;
4274}
4275
4276static int its_sgi_set_irqchip_state(struct irq_data *d,
4277				     enum irqchip_irq_state which,
4278				     bool state)
4279{
4280	if (which != IRQCHIP_STATE_PENDING)
4281		return -EINVAL;
4282
4283	if (state) {
4284		struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4285		struct its_node *its = find_4_1_its();
4286		u64 val;
4287
4288		val  = FIELD_PREP(GITS_SGIR_VPEID, vpe->vpe_id);
4289		val |= FIELD_PREP(GITS_SGIR_VINTID, d->hwirq);
4290		writeq_relaxed(val, its->sgir_base + GITS_SGIR - SZ_128K);
4291	} else {
4292		its_configure_sgi(d, true);
4293	}
4294
4295	return 0;
4296}
4297
4298static int its_sgi_get_irqchip_state(struct irq_data *d,
4299				     enum irqchip_irq_state which, bool *val)
4300{
4301	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4302	void __iomem *base;
4303	unsigned long flags;
4304	u32 count = 1000000;	/* 1s! */
4305	u32 status;
4306	int cpu;
4307
4308	if (which != IRQCHIP_STATE_PENDING)
4309		return -EINVAL;
4310
4311	/*
4312	 * Locking galore! We can race against two different events:
4313	 *
4314	 * - Concurrent vPE affinity change: we must make sure it cannot
4315	 *   happen, or we'll talk to the wrong redistributor. This is
4316	 *   identical to what happens with vLPIs.
4317	 *
4318	 * - Concurrent VSGIPENDR access: As it involves accessing two
4319	 *   MMIO registers, this must be made atomic one way or another.
4320	 */
4321	cpu = vpe_to_cpuid_lock(vpe, &flags);
4322	raw_spin_lock(&gic_data_rdist_cpu(cpu)->rd_lock);
4323	base = gic_data_rdist_cpu(cpu)->rd_base + SZ_128K;
4324	writel_relaxed(vpe->vpe_id, base + GICR_VSGIR);
4325	do {
4326		status = readl_relaxed(base + GICR_VSGIPENDR);
4327		if (!(status & GICR_VSGIPENDR_BUSY))
4328			goto out;
4329
4330		count--;
4331		if (!count) {
4332			pr_err_ratelimited("Unable to get SGI status\n");
4333			goto out;
4334		}
4335		cpu_relax();
4336		udelay(1);
4337	} while (count);
4338
4339out:
4340	raw_spin_unlock(&gic_data_rdist_cpu(cpu)->rd_lock);
4341	vpe_to_cpuid_unlock(vpe, flags);
4342
4343	if (!count)
4344		return -ENXIO;
4345
4346	*val = !!(status & (1 << d->hwirq));
4347
4348	return 0;
4349}
4350
4351static int its_sgi_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
4352{
4353	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4354	struct its_cmd_info *info = vcpu_info;
4355
4356	switch (info->cmd_type) {
4357	case PROP_UPDATE_VSGI:
4358		vpe->sgi_config[d->hwirq].priority = info->priority;
4359		vpe->sgi_config[d->hwirq].group = info->group;
4360		its_configure_sgi(d, false);
4361		return 0;
4362
4363	default:
4364		return -EINVAL;
4365	}
4366}
4367
4368static struct irq_chip its_sgi_irq_chip = {
4369	.name			= "GICv4.1-sgi",
4370	.irq_mask		= its_sgi_mask_irq,
4371	.irq_unmask		= its_sgi_unmask_irq,
4372	.irq_set_affinity	= its_sgi_set_affinity,
4373	.irq_set_irqchip_state	= its_sgi_set_irqchip_state,
4374	.irq_get_irqchip_state	= its_sgi_get_irqchip_state,
4375	.irq_set_vcpu_affinity	= its_sgi_set_vcpu_affinity,
4376};
4377
4378static int its_sgi_irq_domain_alloc(struct irq_domain *domain,
4379				    unsigned int virq, unsigned int nr_irqs,
4380				    void *args)
4381{
4382	struct its_vpe *vpe = args;
4383	int i;
4384
4385	/* Yes, we do want 16 SGIs */
4386	WARN_ON(nr_irqs != 16);
4387
4388	for (i = 0; i < 16; i++) {
4389		vpe->sgi_config[i].priority = 0;
4390		vpe->sgi_config[i].enabled = false;
4391		vpe->sgi_config[i].group = false;
4392
4393		irq_domain_set_hwirq_and_chip(domain, virq + i, i,
4394					      &its_sgi_irq_chip, vpe);
4395		irq_set_status_flags(virq + i, IRQ_DISABLE_UNLAZY);
4396	}
4397
4398	return 0;
4399}
4400
4401static void its_sgi_irq_domain_free(struct irq_domain *domain,
4402				    unsigned int virq,
4403				    unsigned int nr_irqs)
4404{
4405	/* Nothing to do */
4406}
4407
4408static int its_sgi_irq_domain_activate(struct irq_domain *domain,
4409				       struct irq_data *d, bool reserve)
4410{
4411	/* Write out the initial SGI configuration */
4412	its_configure_sgi(d, false);
4413	return 0;
4414}
4415
4416static void its_sgi_irq_domain_deactivate(struct irq_domain *domain,
4417					  struct irq_data *d)
4418{
4419	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4420
4421	/*
4422	 * The VSGI command is awkward:
4423	 *
4424	 * - To change the configuration, CLEAR must be set to false,
4425	 *   leaving the pending bit unchanged.
4426	 * - To clear the pending bit, CLEAR must be set to true, leaving
4427	 *   the configuration unchanged.
4428	 *
4429	 * You just can't do both at once, hence the two commands below.
4430	 */
4431	vpe->sgi_config[d->hwirq].enabled = false;
4432	its_configure_sgi(d, false);
4433	its_configure_sgi(d, true);
4434}
4435
4436static const struct irq_domain_ops its_sgi_domain_ops = {
4437	.alloc		= its_sgi_irq_domain_alloc,
4438	.free		= its_sgi_irq_domain_free,
4439	.activate	= its_sgi_irq_domain_activate,
4440	.deactivate	= its_sgi_irq_domain_deactivate,
4441};
4442
4443static int its_vpe_id_alloc(void)
4444{
4445	return ida_alloc_max(&its_vpeid_ida, ITS_MAX_VPEID - 1, GFP_KERNEL);
4446}
4447
4448static void its_vpe_id_free(u16 id)
4449{
4450	ida_free(&its_vpeid_ida, id);
4451}
4452
4453static int its_vpe_init(struct its_vpe *vpe)
4454{
4455	struct page *vpt_page;
4456	int vpe_id;
4457
4458	/* Allocate vpe_id */
4459	vpe_id = its_vpe_id_alloc();
4460	if (vpe_id < 0)
4461		return vpe_id;
4462
4463	/* Allocate VPT */
4464	vpt_page = its_allocate_pending_table(GFP_KERNEL);
4465	if (!vpt_page) {
4466		its_vpe_id_free(vpe_id);
4467		return -ENOMEM;
4468	}
4469
4470	if (!its_alloc_vpe_table(vpe_id)) {
4471		its_vpe_id_free(vpe_id);
4472		its_free_pending_table(vpt_page);
4473		return -ENOMEM;
4474	}
4475
4476	raw_spin_lock_init(&vpe->vpe_lock);
4477	vpe->vpe_id = vpe_id;
4478	vpe->vpt_page = vpt_page;
4479	if (gic_rdists->has_rvpeid)
4480		atomic_set(&vpe->vmapp_count, 0);
4481	else
4482		vpe->vpe_proxy_event = -1;
4483
4484	return 0;
4485}
4486
4487static void its_vpe_teardown(struct its_vpe *vpe)
4488{
4489	its_vpe_db_proxy_unmap(vpe);
4490	its_vpe_id_free(vpe->vpe_id);
4491	its_free_pending_table(vpe->vpt_page);
4492}
4493
4494static void its_vpe_irq_domain_free(struct irq_domain *domain,
4495				    unsigned int virq,
4496				    unsigned int nr_irqs)
4497{
4498	struct its_vm *vm = domain->host_data;
4499	int i;
4500
4501	irq_domain_free_irqs_parent(domain, virq, nr_irqs);
4502
4503	for (i = 0; i < nr_irqs; i++) {
4504		struct irq_data *data = irq_domain_get_irq_data(domain,
4505								virq + i);
4506		struct its_vpe *vpe = irq_data_get_irq_chip_data(data);
4507
4508		BUG_ON(vm != vpe->its_vm);
4509
4510		clear_bit(data->hwirq, vm->db_bitmap);
4511		its_vpe_teardown(vpe);
4512		irq_domain_reset_irq_data(data);
4513	}
4514
4515	if (bitmap_empty(vm->db_bitmap, vm->nr_db_lpis)) {
4516		its_lpi_free(vm->db_bitmap, vm->db_lpi_base, vm->nr_db_lpis);
4517		its_free_prop_table(vm->vprop_page);
4518	}
4519}
4520
4521static int its_vpe_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
4522				    unsigned int nr_irqs, void *args)
4523{
4524	struct irq_chip *irqchip = &its_vpe_irq_chip;
4525	struct its_vm *vm = args;
4526	unsigned long *bitmap;
4527	struct page *vprop_page;
4528	int base, nr_ids, i, err = 0;
4529
4530	BUG_ON(!vm);
4531
4532	bitmap = its_lpi_alloc(roundup_pow_of_two(nr_irqs), &base, &nr_ids);
4533	if (!bitmap)
4534		return -ENOMEM;
4535
4536	if (nr_ids < nr_irqs) {
4537		its_lpi_free(bitmap, base, nr_ids);
4538		return -ENOMEM;
4539	}
4540
4541	vprop_page = its_allocate_prop_table(GFP_KERNEL);
4542	if (!vprop_page) {
4543		its_lpi_free(bitmap, base, nr_ids);
4544		return -ENOMEM;
4545	}
4546
4547	vm->db_bitmap = bitmap;
4548	vm->db_lpi_base = base;
4549	vm->nr_db_lpis = nr_ids;
4550	vm->vprop_page = vprop_page;
4551
4552	if (gic_rdists->has_rvpeid)
4553		irqchip = &its_vpe_4_1_irq_chip;
4554
4555	for (i = 0; i < nr_irqs; i++) {
4556		vm->vpes[i]->vpe_db_lpi = base + i;
4557		err = its_vpe_init(vm->vpes[i]);
4558		if (err)
4559			break;
4560		err = its_irq_gic_domain_alloc(domain, virq + i,
4561					       vm->vpes[i]->vpe_db_lpi);
4562		if (err)
4563			break;
4564		irq_domain_set_hwirq_and_chip(domain, virq + i, i,
4565					      irqchip, vm->vpes[i]);
4566		set_bit(i, bitmap);
4567		irqd_set_resend_when_in_progress(irq_get_irq_data(virq + i));
4568	}
4569
4570	if (err)
4571		its_vpe_irq_domain_free(domain, virq, i);
4572
4573	return err;
4574}
4575
4576static int its_vpe_irq_domain_activate(struct irq_domain *domain,
4577				       struct irq_data *d, bool reserve)
4578{
4579	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4580	struct its_node *its;
4581
4582	/*
4583	 * If we use the list map, we issue VMAPP on demand... Unless
4584	 * we're on a GICv4.1 and we eagerly map the VPE on all ITSs
4585	 * so that VSGIs can work.
4586	 */
4587	if (!gic_requires_eager_mapping())
4588		return 0;
4589
4590	/* Map the VPE to the first possible CPU */
4591	vpe->col_idx = cpumask_first(cpu_online_mask);
4592
4593	list_for_each_entry(its, &its_nodes, entry) {
4594		if (!is_v4(its))
4595			continue;
4596
4597		its_send_vmapp(its, vpe, true);
4598		its_send_vinvall(its, vpe);
4599	}
4600
4601	irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
4602
4603	return 0;
4604}
4605
4606static void its_vpe_irq_domain_deactivate(struct irq_domain *domain,
4607					  struct irq_data *d)
4608{
4609	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4610	struct its_node *its;
4611
4612	/*
4613	 * If we use the list map on GICv4.0, we unmap the VPE once no
4614	 * VLPIs are associated with the VM.
4615	 */
4616	if (!gic_requires_eager_mapping())
4617		return;
4618
4619	list_for_each_entry(its, &its_nodes, entry) {
4620		if (!is_v4(its))
4621			continue;
4622
4623		its_send_vmapp(its, vpe, false);
4624	}
4625
4626	/*
4627	 * There may be a direct read to the VPT after unmapping the
4628	 * vPE, to guarantee the validity of this, we make the VPT
4629	 * memory coherent with the CPU caches here.
4630	 */
4631	if (find_4_1_its() && !atomic_read(&vpe->vmapp_count))
4632		gic_flush_dcache_to_poc(page_address(vpe->vpt_page),
4633					LPI_PENDBASE_SZ);
4634}
4635
4636static const struct irq_domain_ops its_vpe_domain_ops = {
4637	.alloc			= its_vpe_irq_domain_alloc,
4638	.free			= its_vpe_irq_domain_free,
4639	.activate		= its_vpe_irq_domain_activate,
4640	.deactivate		= its_vpe_irq_domain_deactivate,
4641};
4642
4643static int its_force_quiescent(void __iomem *base)
4644{
4645	u32 count = 1000000;	/* 1s */
4646	u32 val;
4647
4648	val = readl_relaxed(base + GITS_CTLR);
4649	/*
4650	 * GIC architecture specification requires the ITS to be both
4651	 * disabled and quiescent for writes to GITS_BASER<n> or
4652	 * GITS_CBASER to not have UNPREDICTABLE results.
4653	 */
4654	if ((val & GITS_CTLR_QUIESCENT) && !(val & GITS_CTLR_ENABLE))
4655		return 0;
4656
4657	/* Disable the generation of all interrupts to this ITS */
4658	val &= ~(GITS_CTLR_ENABLE | GITS_CTLR_ImDe);
4659	writel_relaxed(val, base + GITS_CTLR);
4660
4661	/* Poll GITS_CTLR and wait until ITS becomes quiescent */
4662	while (1) {
4663		val = readl_relaxed(base + GITS_CTLR);
4664		if (val & GITS_CTLR_QUIESCENT)
4665			return 0;
4666
4667		count--;
4668		if (!count)
4669			return -EBUSY;
4670
4671		cpu_relax();
4672		udelay(1);
4673	}
4674}
4675
4676static bool __maybe_unused its_enable_quirk_cavium_22375(void *data)
4677{
4678	struct its_node *its = data;
4679
4680	/* erratum 22375: only alloc 8MB table size (20 bits) */
4681	its->typer &= ~GITS_TYPER_DEVBITS;
4682	its->typer |= FIELD_PREP(GITS_TYPER_DEVBITS, 20 - 1);
4683	its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_22375;
4684
4685	return true;
4686}
4687
4688static bool __maybe_unused its_enable_quirk_cavium_23144(void *data)
4689{
4690	struct its_node *its = data;
4691
4692	its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_23144;
4693
4694	return true;
4695}
4696
4697static bool __maybe_unused its_enable_quirk_qdf2400_e0065(void *data)
4698{
4699	struct its_node *its = data;
4700
4701	/* On QDF2400, the size of the ITE is 16Bytes */
4702	its->typer &= ~GITS_TYPER_ITT_ENTRY_SIZE;
4703	its->typer |= FIELD_PREP(GITS_TYPER_ITT_ENTRY_SIZE, 16 - 1);
4704
4705	return true;
4706}
4707
4708static u64 its_irq_get_msi_base_pre_its(struct its_device *its_dev)
4709{
4710	struct its_node *its = its_dev->its;
4711
4712	/*
4713	 * The Socionext Synquacer SoC has a so-called 'pre-ITS',
4714	 * which maps 32-bit writes targeted at a separate window of
4715	 * size '4 << device_id_bits' onto writes to GITS_TRANSLATER
4716	 * with device ID taken from bits [device_id_bits + 1:2] of
4717	 * the window offset.
4718	 */
4719	return its->pre_its_base + (its_dev->device_id << 2);
4720}
4721
4722static bool __maybe_unused its_enable_quirk_socionext_synquacer(void *data)
4723{
4724	struct its_node *its = data;
4725	u32 pre_its_window[2];
4726	u32 ids;
4727
4728	if (!fwnode_property_read_u32_array(its->fwnode_handle,
4729					   "socionext,synquacer-pre-its",
4730					   pre_its_window,
4731					   ARRAY_SIZE(pre_its_window))) {
4732
4733		its->pre_its_base = pre_its_window[0];
4734		its->get_msi_base = its_irq_get_msi_base_pre_its;
4735
4736		ids = ilog2(pre_its_window[1]) - 2;
4737		if (device_ids(its) > ids) {
4738			its->typer &= ~GITS_TYPER_DEVBITS;
4739			its->typer |= FIELD_PREP(GITS_TYPER_DEVBITS, ids - 1);
4740		}
4741
4742		/* the pre-ITS breaks isolation, so disable MSI remapping */
4743		its->msi_domain_flags &= ~IRQ_DOMAIN_FLAG_ISOLATED_MSI;
4744		return true;
4745	}
4746	return false;
4747}
4748
4749static bool __maybe_unused its_enable_quirk_hip07_161600802(void *data)
4750{
4751	struct its_node *its = data;
4752
4753	/*
4754	 * Hip07 insists on using the wrong address for the VLPI
4755	 * page. Trick it into doing the right thing...
4756	 */
4757	its->vlpi_redist_offset = SZ_128K;
4758	return true;
4759}
4760
4761static bool __maybe_unused its_enable_rk3588001(void *data)
4762{
4763	struct its_node *its = data;
4764
4765	if (!of_machine_is_compatible("rockchip,rk3588") &&
4766	    !of_machine_is_compatible("rockchip,rk3588s"))
4767		return false;
4768
4769	its->flags |= ITS_FLAGS_FORCE_NON_SHAREABLE;
4770	gic_rdists->flags |= RDIST_FLAGS_FORCE_NON_SHAREABLE;
4771
4772	return true;
4773}
4774
4775static bool its_set_non_coherent(void *data)
4776{
4777	struct its_node *its = data;
4778
4779	its->flags |= ITS_FLAGS_FORCE_NON_SHAREABLE;
4780	return true;
4781}
4782
4783static const struct gic_quirk its_quirks[] = {
4784#ifdef CONFIG_CAVIUM_ERRATUM_22375
4785	{
4786		.desc	= "ITS: Cavium errata 22375, 24313",
4787		.iidr	= 0xa100034c,	/* ThunderX pass 1.x */
4788		.mask	= 0xffff0fff,
4789		.init	= its_enable_quirk_cavium_22375,
4790	},
4791#endif
4792#ifdef CONFIG_CAVIUM_ERRATUM_23144
4793	{
4794		.desc	= "ITS: Cavium erratum 23144",
4795		.iidr	= 0xa100034c,	/* ThunderX pass 1.x */
4796		.mask	= 0xffff0fff,
4797		.init	= its_enable_quirk_cavium_23144,
4798	},
4799#endif
4800#ifdef CONFIG_QCOM_QDF2400_ERRATUM_0065
4801	{
4802		.desc	= "ITS: QDF2400 erratum 0065",
4803		.iidr	= 0x00001070, /* QDF2400 ITS rev 1.x */
4804		.mask	= 0xffffffff,
4805		.init	= its_enable_quirk_qdf2400_e0065,
4806	},
4807#endif
4808#ifdef CONFIG_SOCIONEXT_SYNQUACER_PREITS
4809	{
4810		/*
4811		 * The Socionext Synquacer SoC incorporates ARM's own GIC-500
4812		 * implementation, but with a 'pre-ITS' added that requires
4813		 * special handling in software.
4814		 */
4815		.desc	= "ITS: Socionext Synquacer pre-ITS",
4816		.iidr	= 0x0001143b,
4817		.mask	= 0xffffffff,
4818		.init	= its_enable_quirk_socionext_synquacer,
4819	},
4820#endif
4821#ifdef CONFIG_HISILICON_ERRATUM_161600802
4822	{
4823		.desc	= "ITS: Hip07 erratum 161600802",
4824		.iidr	= 0x00000004,
4825		.mask	= 0xffffffff,
4826		.init	= its_enable_quirk_hip07_161600802,
4827	},
4828#endif
4829#ifdef CONFIG_ROCKCHIP_ERRATUM_3588001
4830	{
4831		.desc   = "ITS: Rockchip erratum RK3588001",
4832		.iidr   = 0x0201743b,
4833		.mask   = 0xffffffff,
4834		.init   = its_enable_rk3588001,
4835	},
4836#endif
4837	{
4838		.desc   = "ITS: non-coherent attribute",
4839		.property = "dma-noncoherent",
4840		.init   = its_set_non_coherent,
4841	},
4842	{
4843	}
4844};
4845
4846static void its_enable_quirks(struct its_node *its)
4847{
4848	u32 iidr = readl_relaxed(its->base + GITS_IIDR);
4849
4850	gic_enable_quirks(iidr, its_quirks, its);
4851
4852	if (is_of_node(its->fwnode_handle))
4853		gic_enable_of_quirks(to_of_node(its->fwnode_handle),
4854				     its_quirks, its);
4855}
4856
4857static int its_save_disable(void)
 
4858{
 
4859	struct its_node *its;
4860	int err = 0;
4861
4862	raw_spin_lock(&its_lock);
4863	list_for_each_entry(its, &its_nodes, entry) {
4864		void __iomem *base;
4865
4866		base = its->base;
4867		its->ctlr_save = readl_relaxed(base + GITS_CTLR);
4868		err = its_force_quiescent(base);
4869		if (err) {
4870			pr_err("ITS@%pa: failed to quiesce: %d\n",
4871			       &its->phys_base, err);
4872			writel_relaxed(its->ctlr_save, base + GITS_CTLR);
4873			goto err;
4874		}
4875
4876		its->cbaser_save = gits_read_cbaser(base + GITS_CBASER);
4877	}
4878
4879err:
4880	if (err) {
4881		list_for_each_entry_continue_reverse(its, &its_nodes, entry) {
4882			void __iomem *base;
4883
4884			base = its->base;
4885			writel_relaxed(its->ctlr_save, base + GITS_CTLR);
4886		}
4887	}
4888	raw_spin_unlock(&its_lock);
4889
4890	return err;
4891}
4892
4893static void its_restore_enable(void)
4894{
4895	struct its_node *its;
4896	int ret;
4897
4898	raw_spin_lock(&its_lock);
4899	list_for_each_entry(its, &its_nodes, entry) {
4900		void __iomem *base;
4901		int i;
4902
4903		base = its->base;
4904
4905		/*
4906		 * Make sure that the ITS is disabled. If it fails to quiesce,
4907		 * don't restore it since writing to CBASER or BASER<n>
4908		 * registers is undefined according to the GIC v3 ITS
4909		 * Specification.
4910		 *
4911		 * Firmware resuming with the ITS enabled is terminally broken.
4912		 */
4913		WARN_ON(readl_relaxed(base + GITS_CTLR) & GITS_CTLR_ENABLE);
4914		ret = its_force_quiescent(base);
4915		if (ret) {
4916			pr_err("ITS@%pa: failed to quiesce on resume: %d\n",
4917			       &its->phys_base, ret);
4918			continue;
4919		}
4920
4921		gits_write_cbaser(its->cbaser_save, base + GITS_CBASER);
4922
4923		/*
4924		 * Writing CBASER resets CREADR to 0, so make CWRITER and
4925		 * cmd_write line up with it.
4926		 */
4927		its->cmd_write = its->cmd_base;
4928		gits_write_cwriter(0, base + GITS_CWRITER);
4929
4930		/* Restore GITS_BASER from the value cache. */
4931		for (i = 0; i < GITS_BASER_NR_REGS; i++) {
4932			struct its_baser *baser = &its->tables[i];
4933
4934			if (!(baser->val & GITS_BASER_VALID))
4935				continue;
4936
4937			its_write_baser(its, baser, baser->val);
4938		}
4939		writel_relaxed(its->ctlr_save, base + GITS_CTLR);
4940
4941		/*
4942		 * Reinit the collection if it's stored in the ITS. This is
4943		 * indicated by the col_id being less than the HCC field.
4944		 * CID < HCC as specified in the GIC v3 Documentation.
4945		 */
4946		if (its->collections[smp_processor_id()].col_id <
4947		    GITS_TYPER_HCC(gic_read_typer(base + GITS_TYPER)))
4948			its_cpu_init_collection(its);
4949	}
4950	raw_spin_unlock(&its_lock);
4951}
4952
4953static struct syscore_ops its_syscore_ops = {
4954	.suspend = its_save_disable,
4955	.resume = its_restore_enable,
4956};
4957
4958static void __init __iomem *its_map_one(struct resource *res, int *err)
4959{
4960	void __iomem *its_base;
4961	u32 val;
4962
4963	its_base = ioremap(res->start, SZ_64K);
4964	if (!its_base) {
4965		pr_warn("ITS@%pa: Unable to map ITS registers\n", &res->start);
4966		*err = -ENOMEM;
4967		return NULL;
4968	}
4969
4970	val = readl_relaxed(its_base + GITS_PIDR2) & GIC_PIDR2_ARCH_MASK;
4971	if (val != 0x30 && val != 0x40) {
4972		pr_warn("ITS@%pa: No ITS detected, giving up\n", &res->start);
4973		*err = -ENODEV;
4974		goto out_unmap;
4975	}
4976
4977	*err = its_force_quiescent(its_base);
4978	if (*err) {
4979		pr_warn("ITS@%pa: Failed to quiesce, giving up\n", &res->start);
 
4980		goto out_unmap;
4981	}
4982
4983	return its_base;
4984
4985out_unmap:
4986	iounmap(its_base);
4987	return NULL;
4988}
4989
4990static int its_init_domain(struct its_node *its)
4991{
4992	struct irq_domain *inner_domain;
4993	struct msi_domain_info *info;
4994
4995	info = kzalloc(sizeof(*info), GFP_KERNEL);
4996	if (!info)
4997		return -ENOMEM;
4998
4999	info->ops = &its_msi_domain_ops;
5000	info->data = its;
5001
5002	inner_domain = irq_domain_create_hierarchy(its_parent,
5003						   its->msi_domain_flags, 0,
5004						   its->fwnode_handle, &its_domain_ops,
5005						   info);
5006	if (!inner_domain) {
5007		kfree(info);
5008		return -ENOMEM;
5009	}
5010
5011	irq_domain_update_bus_token(inner_domain, DOMAIN_BUS_NEXUS);
 
 
 
 
 
5012
5013	return 0;
5014}
5015
5016static int its_init_vpe_domain(void)
5017{
5018	struct its_node *its;
5019	u32 devid;
5020	int entries;
5021
5022	if (gic_rdists->has_direct_lpi) {
5023		pr_info("ITS: Using DirectLPI for VPE invalidation\n");
5024		return 0;
5025	}
5026
5027	/* Any ITS will do, even if not v4 */
5028	its = list_first_entry(&its_nodes, struct its_node, entry);
5029
5030	entries = roundup_pow_of_two(nr_cpu_ids);
5031	vpe_proxy.vpes = kcalloc(entries, sizeof(*vpe_proxy.vpes),
5032				 GFP_KERNEL);
5033	if (!vpe_proxy.vpes)
5034		return -ENOMEM;
5035
5036	/* Use the last possible DevID */
5037	devid = GENMASK(device_ids(its) - 1, 0);
5038	vpe_proxy.dev = its_create_device(its, devid, entries, false);
5039	if (!vpe_proxy.dev) {
5040		kfree(vpe_proxy.vpes);
5041		pr_err("ITS: Can't allocate GICv4 proxy device\n");
5042		return -ENOMEM;
5043	}
5044
5045	BUG_ON(entries > vpe_proxy.dev->nr_ites);
5046
5047	raw_spin_lock_init(&vpe_proxy.lock);
5048	vpe_proxy.next_victim = 0;
5049	pr_info("ITS: Allocated DevID %x as GICv4 proxy device (%d slots)\n",
5050		devid, vpe_proxy.dev->nr_ites);
5051
5052	return 0;
5053}
5054
5055static int __init its_compute_its_list_map(struct its_node *its)
5056{
5057	int its_number;
5058	u32 ctlr;
5059
5060	/*
5061	 * This is assumed to be done early enough that we're
5062	 * guaranteed to be single-threaded, hence no
5063	 * locking. Should this change, we should address
5064	 * this.
5065	 */
5066	its_number = find_first_zero_bit(&its_list_map, GICv4_ITS_LIST_MAX);
5067	if (its_number >= GICv4_ITS_LIST_MAX) {
5068		pr_err("ITS@%pa: No ITSList entry available!\n",
5069		       &its->phys_base);
5070		return -EINVAL;
5071	}
5072
5073	ctlr = readl_relaxed(its->base + GITS_CTLR);
5074	ctlr &= ~GITS_CTLR_ITS_NUMBER;
5075	ctlr |= its_number << GITS_CTLR_ITS_NUMBER_SHIFT;
5076	writel_relaxed(ctlr, its->base + GITS_CTLR);
5077	ctlr = readl_relaxed(its->base + GITS_CTLR);
5078	if ((ctlr & GITS_CTLR_ITS_NUMBER) != (its_number << GITS_CTLR_ITS_NUMBER_SHIFT)) {
5079		its_number = ctlr & GITS_CTLR_ITS_NUMBER;
5080		its_number >>= GITS_CTLR_ITS_NUMBER_SHIFT;
5081	}
5082
5083	if (test_and_set_bit(its_number, &its_list_map)) {
5084		pr_err("ITS@%pa: Duplicate ITSList entry %d\n",
5085		       &its->phys_base, its_number);
5086		return -EINVAL;
5087	}
5088
5089	return its_number;
5090}
5091
5092static int __init its_probe_one(struct its_node *its)
5093{
5094	u64 baser, tmp;
5095	struct page *page;
5096	u32 ctlr;
5097	int err;
5098
5099	its_enable_quirks(its);
5100
5101	if (is_v4(its)) {
5102		if (!(its->typer & GITS_TYPER_VMOVP)) {
5103			err = its_compute_its_list_map(its);
5104			if (err < 0)
5105				goto out;
5106
5107			its->list_nr = err;
5108
5109			pr_info("ITS@%pa: Using ITS number %d\n",
5110				&its->phys_base, err);
5111		} else {
5112			pr_info("ITS@%pa: Single VMOVP capable\n", &its->phys_base);
5113		}
5114
5115		if (is_v4_1(its)) {
5116			u32 svpet = FIELD_GET(GITS_TYPER_SVPET, its->typer);
5117
5118			its->sgir_base = ioremap(its->phys_base + SZ_128K, SZ_64K);
5119			if (!its->sgir_base) {
5120				err = -ENOMEM;
5121				goto out;
5122			}
5123
5124			its->mpidr = readl_relaxed(its->base + GITS_MPIDR);
5125
5126			pr_info("ITS@%pa: Using GICv4.1 mode %08x %08x\n",
5127				&its->phys_base, its->mpidr, svpet);
5128		}
5129	}
5130
5131	page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO,
5132				get_order(ITS_CMD_QUEUE_SZ));
5133	if (!page) {
5134		err = -ENOMEM;
5135		goto out_unmap_sgir;
5136	}
5137	its->cmd_base = (void *)page_address(page);
5138	its->cmd_write = its->cmd_base;
5139
5140	err = its_alloc_tables(its);
5141	if (err)
5142		goto out_free_cmd;
5143
5144	err = its_alloc_collections(its);
5145	if (err)
5146		goto out_free_tables;
5147
5148	baser = (virt_to_phys(its->cmd_base)	|
5149		 GITS_CBASER_RaWaWb		|
5150		 GITS_CBASER_InnerShareable	|
5151		 (ITS_CMD_QUEUE_SZ / SZ_4K - 1)	|
5152		 GITS_CBASER_VALID);
5153
5154	gits_write_cbaser(baser, its->base + GITS_CBASER);
5155	tmp = gits_read_cbaser(its->base + GITS_CBASER);
5156
5157	if (its->flags & ITS_FLAGS_FORCE_NON_SHAREABLE)
5158		tmp &= ~GITS_CBASER_SHAREABILITY_MASK;
5159
5160	if ((tmp ^ baser) & GITS_CBASER_SHAREABILITY_MASK) {
5161		if (!(tmp & GITS_CBASER_SHAREABILITY_MASK)) {
5162			/*
5163			 * The HW reports non-shareable, we must
5164			 * remove the cacheability attributes as
5165			 * well.
5166			 */
5167			baser &= ~(GITS_CBASER_SHAREABILITY_MASK |
5168				   GITS_CBASER_CACHEABILITY_MASK);
5169			baser |= GITS_CBASER_nC;
5170			gits_write_cbaser(baser, its->base + GITS_CBASER);
5171		}
5172		pr_info("ITS: using cache flushing for cmd queue\n");
5173		its->flags |= ITS_FLAGS_CMDQ_NEEDS_FLUSHING;
5174	}
5175
5176	gits_write_cwriter(0, its->base + GITS_CWRITER);
5177	ctlr = readl_relaxed(its->base + GITS_CTLR);
5178	ctlr |= GITS_CTLR_ENABLE;
5179	if (is_v4(its))
5180		ctlr |= GITS_CTLR_ImDe;
5181	writel_relaxed(ctlr, its->base + GITS_CTLR);
 
 
 
 
 
 
 
 
 
 
 
 
5182
5183	err = its_init_domain(its);
5184	if (err)
5185		goto out_free_tables;
 
 
 
5186
5187	raw_spin_lock(&its_lock);
5188	list_add(&its->entry, &its_nodes);
5189	raw_spin_unlock(&its_lock);
5190
5191	return 0;
5192
5193out_free_tables:
5194	its_free_tables(its);
5195out_free_cmd:
5196	free_pages((unsigned long)its->cmd_base, get_order(ITS_CMD_QUEUE_SZ));
5197out_unmap_sgir:
5198	if (its->sgir_base)
5199		iounmap(its->sgir_base);
5200out:
5201	pr_err("ITS@%pa: failed probing (%d)\n", &its->phys_base, err);
5202	return err;
5203}
5204
5205static bool gic_rdists_supports_plpis(void)
5206{
5207	return !!(gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER) & GICR_TYPER_PLPIS);
5208}
5209
5210static int redist_disable_lpis(void)
5211{
5212	void __iomem *rbase = gic_data_rdist_rd_base();
5213	u64 timeout = USEC_PER_SEC;
5214	u64 val;
5215
5216	if (!gic_rdists_supports_plpis()) {
5217		pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
5218		return -ENXIO;
5219	}
5220
5221	val = readl_relaxed(rbase + GICR_CTLR);
5222	if (!(val & GICR_CTLR_ENABLE_LPIS))
5223		return 0;
5224
5225	/*
5226	 * If coming via a CPU hotplug event, we don't need to disable
5227	 * LPIs before trying to re-enable them. They are already
5228	 * configured and all is well in the world.
5229	 *
5230	 * If running with preallocated tables, there is nothing to do.
5231	 */
5232	if ((gic_data_rdist()->flags & RD_LOCAL_LPI_ENABLED) ||
5233	    (gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED))
5234		return 0;
5235
5236	/*
5237	 * From that point on, we only try to do some damage control.
5238	 */
5239	pr_warn("GICv3: CPU%d: Booted with LPIs enabled, memory probably corrupted\n",
5240		smp_processor_id());
5241	add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
5242
5243	/* Disable LPIs */
5244	val &= ~GICR_CTLR_ENABLE_LPIS;
5245	writel_relaxed(val, rbase + GICR_CTLR);
5246
5247	/* Make sure any change to GICR_CTLR is observable by the GIC */
5248	dsb(sy);
5249
5250	/*
5251	 * Software must observe RWP==0 after clearing GICR_CTLR.EnableLPIs
5252	 * from 1 to 0 before programming GICR_PEND{PROP}BASER registers.
5253	 * Error out if we time out waiting for RWP to clear.
5254	 */
5255	while (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_RWP) {
5256		if (!timeout) {
5257			pr_err("CPU%d: Timeout while disabling LPIs\n",
5258			       smp_processor_id());
5259			return -ETIMEDOUT;
5260		}
5261		udelay(1);
5262		timeout--;
5263	}
5264
5265	/*
5266	 * After it has been written to 1, it is IMPLEMENTATION
5267	 * DEFINED whether GICR_CTLR.EnableLPI becomes RES1 or can be
5268	 * cleared to 0. Error out if clearing the bit failed.
5269	 */
5270	if (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_ENABLE_LPIS) {
5271		pr_err("CPU%d: Failed to disable LPIs\n", smp_processor_id());
5272		return -EBUSY;
5273	}
5274
5275	return 0;
5276}
5277
5278int its_cpu_init(void)
5279{
5280	if (!list_empty(&its_nodes)) {
5281		int ret;
5282
5283		ret = redist_disable_lpis();
5284		if (ret)
5285			return ret;
5286
5287		its_cpu_init_lpis();
5288		its_cpu_init_collections();
5289	}
5290
5291	return 0;
5292}
5293
5294static void rdist_memreserve_cpuhp_cleanup_workfn(struct work_struct *work)
5295{
5296	cpuhp_remove_state_nocalls(gic_rdists->cpuhp_memreserve_state);
5297	gic_rdists->cpuhp_memreserve_state = CPUHP_INVALID;
5298}
5299
5300static DECLARE_WORK(rdist_memreserve_cpuhp_cleanup_work,
5301		    rdist_memreserve_cpuhp_cleanup_workfn);
5302
5303static int its_cpu_memreserve_lpi(unsigned int cpu)
5304{
5305	struct page *pend_page;
5306	int ret = 0;
5307
5308	/* This gets to run exactly once per CPU */
5309	if (gic_data_rdist()->flags & RD_LOCAL_MEMRESERVE_DONE)
5310		return 0;
5311
5312	pend_page = gic_data_rdist()->pend_page;
5313	if (WARN_ON(!pend_page)) {
5314		ret = -ENOMEM;
5315		goto out;
5316	}
5317	/*
5318	 * If the pending table was pre-programmed, free the memory we
5319	 * preemptively allocated. Otherwise, reserve that memory for
5320	 * later kexecs.
5321	 */
5322	if (gic_data_rdist()->flags & RD_LOCAL_PENDTABLE_PREALLOCATED) {
5323		its_free_pending_table(pend_page);
5324		gic_data_rdist()->pend_page = NULL;
5325	} else {
5326		phys_addr_t paddr = page_to_phys(pend_page);
5327		WARN_ON(gic_reserve_range(paddr, LPI_PENDBASE_SZ));
5328	}
5329
5330out:
5331	/* Last CPU being brought up gets to issue the cleanup */
5332	if (!IS_ENABLED(CONFIG_SMP) ||
5333	    cpumask_equal(&cpus_booted_once_mask, cpu_possible_mask))
5334		schedule_work(&rdist_memreserve_cpuhp_cleanup_work);
5335
5336	gic_data_rdist()->flags |= RD_LOCAL_MEMRESERVE_DONE;
5337	return ret;
5338}
5339
5340/* Mark all the BASER registers as invalid before they get reprogrammed */
5341static int __init its_reset_one(struct resource *res)
5342{
5343	void __iomem *its_base;
5344	int err, i;
5345
5346	its_base = its_map_one(res, &err);
5347	if (!its_base)
5348		return err;
5349
5350	for (i = 0; i < GITS_BASER_NR_REGS; i++)
5351		gits_write_baser(0, its_base + GITS_BASER + (i << 3));
5352
5353	iounmap(its_base);
5354	return 0;
5355}
5356
5357static const struct of_device_id its_device_id[] = {
5358	{	.compatible	= "arm,gic-v3-its",	},
5359	{},
5360};
5361
5362static struct its_node __init *its_node_init(struct resource *res,
5363					     struct fwnode_handle *handle, int numa_node)
5364{
5365	void __iomem *its_base;
5366	struct its_node *its;
5367	int err;
5368
5369	its_base = its_map_one(res, &err);
5370	if (!its_base)
5371		return NULL;
5372
5373	pr_info("ITS %pR\n", res);
5374
5375	its = kzalloc(sizeof(*its), GFP_KERNEL);
5376	if (!its)
5377		goto out_unmap;
5378
5379	raw_spin_lock_init(&its->lock);
5380	mutex_init(&its->dev_alloc_lock);
5381	INIT_LIST_HEAD(&its->entry);
5382	INIT_LIST_HEAD(&its->its_device_list);
5383
5384	its->typer = gic_read_typer(its_base + GITS_TYPER);
5385	its->base = its_base;
5386	its->phys_base = res->start;
5387	its->get_msi_base = its_irq_get_msi_base;
5388	its->msi_domain_flags = IRQ_DOMAIN_FLAG_ISOLATED_MSI;
5389
5390	its->numa_node = numa_node;
5391	its->fwnode_handle = handle;
5392
5393	return its;
5394
5395out_unmap:
5396	iounmap(its_base);
5397	return NULL;
5398}
5399
5400static void its_node_destroy(struct its_node *its)
5401{
5402	iounmap(its->base);
5403	kfree(its);
5404}
5405
5406static int __init its_of_probe(struct device_node *node)
5407{
5408	struct device_node *np;
5409	struct resource res;
5410	int err;
5411
5412	/*
5413	 * Make sure *all* the ITS are reset before we probe any, as
5414	 * they may be sharing memory. If any of the ITS fails to
5415	 * reset, don't even try to go any further, as this could
5416	 * result in something even worse.
5417	 */
5418	for (np = of_find_matching_node(node, its_device_id); np;
5419	     np = of_find_matching_node(np, its_device_id)) {
5420		if (!of_device_is_available(np) ||
5421		    !of_property_read_bool(np, "msi-controller") ||
5422		    of_address_to_resource(np, 0, &res))
5423			continue;
5424
5425		err = its_reset_one(&res);
5426		if (err)
5427			return err;
5428	}
5429
5430	for (np = of_find_matching_node(node, its_device_id); np;
5431	     np = of_find_matching_node(np, its_device_id)) {
5432		struct its_node *its;
5433
5434		if (!of_device_is_available(np))
5435			continue;
5436		if (!of_property_read_bool(np, "msi-controller")) {
5437			pr_warn("%pOF: no msi-controller property, ITS ignored\n",
5438				np);
5439			continue;
5440		}
5441
5442		if (of_address_to_resource(np, 0, &res)) {
5443			pr_warn("%pOF: no regs?\n", np);
5444			continue;
5445		}
5446
5447
5448		its = its_node_init(&res, &np->fwnode, of_node_to_nid(np));
5449		if (!its)
5450			return -ENOMEM;
5451
5452		err = its_probe_one(its);
5453		if (err)  {
5454			its_node_destroy(its);
5455			return err;
5456		}
5457	}
5458	return 0;
5459}
5460
5461#ifdef CONFIG_ACPI
5462
5463#define ACPI_GICV3_ITS_MEM_SIZE (SZ_128K)
5464
5465#ifdef CONFIG_ACPI_NUMA
5466struct its_srat_map {
5467	/* numa node id */
5468	u32	numa_node;
5469	/* GIC ITS ID */
5470	u32	its_id;
5471};
5472
5473static struct its_srat_map *its_srat_maps __initdata;
5474static int its_in_srat __initdata;
5475
5476static int __init acpi_get_its_numa_node(u32 its_id)
5477{
5478	int i;
5479
5480	for (i = 0; i < its_in_srat; i++) {
5481		if (its_id == its_srat_maps[i].its_id)
5482			return its_srat_maps[i].numa_node;
5483	}
5484	return NUMA_NO_NODE;
5485}
5486
5487static int __init gic_acpi_match_srat_its(union acpi_subtable_headers *header,
5488					  const unsigned long end)
5489{
5490	return 0;
5491}
5492
5493static int __init gic_acpi_parse_srat_its(union acpi_subtable_headers *header,
5494			 const unsigned long end)
5495{
5496	int node;
5497	struct acpi_srat_gic_its_affinity *its_affinity;
5498
5499	its_affinity = (struct acpi_srat_gic_its_affinity *)header;
5500	if (!its_affinity)
5501		return -EINVAL;
5502
5503	if (its_affinity->header.length < sizeof(*its_affinity)) {
5504		pr_err("SRAT: Invalid header length %d in ITS affinity\n",
5505			its_affinity->header.length);
5506		return -EINVAL;
5507	}
5508
5509	/*
5510	 * Note that in theory a new proximity node could be created by this
5511	 * entry as it is an SRAT resource allocation structure.
5512	 * We do not currently support doing so.
5513	 */
5514	node = pxm_to_node(its_affinity->proximity_domain);
5515
5516	if (node == NUMA_NO_NODE || node >= MAX_NUMNODES) {
5517		pr_err("SRAT: Invalid NUMA node %d in ITS affinity\n", node);
5518		return 0;
5519	}
5520
5521	its_srat_maps[its_in_srat].numa_node = node;
5522	its_srat_maps[its_in_srat].its_id = its_affinity->its_id;
5523	its_in_srat++;
5524	pr_info("SRAT: PXM %d -> ITS %d -> Node %d\n",
5525		its_affinity->proximity_domain, its_affinity->its_id, node);
5526
5527	return 0;
5528}
5529
5530static void __init acpi_table_parse_srat_its(void)
5531{
5532	int count;
5533
5534	count = acpi_table_parse_entries(ACPI_SIG_SRAT,
5535			sizeof(struct acpi_table_srat),
5536			ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
5537			gic_acpi_match_srat_its, 0);
5538	if (count <= 0)
5539		return;
5540
5541	its_srat_maps = kmalloc_array(count, sizeof(struct its_srat_map),
5542				      GFP_KERNEL);
5543	if (!its_srat_maps)
5544		return;
5545
5546	acpi_table_parse_entries(ACPI_SIG_SRAT,
5547			sizeof(struct acpi_table_srat),
5548			ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
5549			gic_acpi_parse_srat_its, 0);
5550}
5551
5552/* free the its_srat_maps after ITS probing */
5553static void __init acpi_its_srat_maps_free(void)
5554{
5555	kfree(its_srat_maps);
5556}
5557#else
5558static void __init acpi_table_parse_srat_its(void)	{ }
5559static int __init acpi_get_its_numa_node(u32 its_id) { return NUMA_NO_NODE; }
5560static void __init acpi_its_srat_maps_free(void) { }
5561#endif
5562
5563static int __init gic_acpi_parse_madt_its(union acpi_subtable_headers *header,
5564					  const unsigned long end)
5565{
5566	struct acpi_madt_generic_translator *its_entry;
5567	struct fwnode_handle *dom_handle;
5568	struct its_node *its;
5569	struct resource res;
5570	int err;
5571
5572	its_entry = (struct acpi_madt_generic_translator *)header;
5573	memset(&res, 0, sizeof(res));
5574	res.start = its_entry->base_address;
5575	res.end = its_entry->base_address + ACPI_GICV3_ITS_MEM_SIZE - 1;
5576	res.flags = IORESOURCE_MEM;
5577
5578	dom_handle = irq_domain_alloc_fwnode(&res.start);
5579	if (!dom_handle) {
5580		pr_err("ITS@%pa: Unable to allocate GICv3 ITS domain token\n",
5581		       &res.start);
5582		return -ENOMEM;
5583	}
5584
5585	err = iort_register_domain_token(its_entry->translation_id, res.start,
5586					 dom_handle);
5587	if (err) {
5588		pr_err("ITS@%pa: Unable to register GICv3 ITS domain token (ITS ID %d) to IORT\n",
5589		       &res.start, its_entry->translation_id);
5590		goto dom_err;
5591	}
5592
5593	its = its_node_init(&res, dom_handle,
5594			    acpi_get_its_numa_node(its_entry->translation_id));
5595	if (!its) {
5596		err = -ENOMEM;
5597		goto node_err;
5598	}
5599
5600	err = its_probe_one(its);
5601	if (!err)
5602		return 0;
5603
5604node_err:
5605	iort_deregister_domain_token(its_entry->translation_id);
5606dom_err:
5607	irq_domain_free_fwnode(dom_handle);
5608	return err;
5609}
5610
5611static int __init its_acpi_reset(union acpi_subtable_headers *header,
5612				 const unsigned long end)
5613{
5614	struct acpi_madt_generic_translator *its_entry;
5615	struct resource res;
5616
5617	its_entry = (struct acpi_madt_generic_translator *)header;
5618	res = (struct resource) {
5619		.start	= its_entry->base_address,
5620		.end	= its_entry->base_address + ACPI_GICV3_ITS_MEM_SIZE - 1,
5621		.flags	= IORESOURCE_MEM,
5622	};
5623
5624	return its_reset_one(&res);
5625}
5626
5627static void __init its_acpi_probe(void)
5628{
5629	acpi_table_parse_srat_its();
5630	/*
5631	 * Make sure *all* the ITS are reset before we probe any, as
5632	 * they may be sharing memory. If any of the ITS fails to
5633	 * reset, don't even try to go any further, as this could
5634	 * result in something even worse.
5635	 */
5636	if (acpi_table_parse_madt(ACPI_MADT_TYPE_GENERIC_TRANSLATOR,
5637				  its_acpi_reset, 0) > 0)
5638		acpi_table_parse_madt(ACPI_MADT_TYPE_GENERIC_TRANSLATOR,
5639				      gic_acpi_parse_madt_its, 0);
5640	acpi_its_srat_maps_free();
5641}
5642#else
5643static void __init its_acpi_probe(void) { }
5644#endif
5645
5646int __init its_lpi_memreserve_init(void)
5647{
5648	int state;
5649
5650	if (!efi_enabled(EFI_CONFIG_TABLES))
5651		return 0;
5652
5653	if (list_empty(&its_nodes))
5654		return 0;
5655
5656	gic_rdists->cpuhp_memreserve_state = CPUHP_INVALID;
5657	state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
5658				  "irqchip/arm/gicv3/memreserve:online",
5659				  its_cpu_memreserve_lpi,
5660				  NULL);
5661	if (state < 0)
5662		return state;
5663
5664	gic_rdists->cpuhp_memreserve_state = state;
5665
5666	return 0;
5667}
5668
5669int __init its_init(struct fwnode_handle *handle, struct rdists *rdists,
5670		    struct irq_domain *parent_domain)
5671{
5672	struct device_node *of_node;
5673	struct its_node *its;
5674	bool has_v4 = false;
5675	bool has_v4_1 = false;
5676	int err;
5677
5678	gic_rdists = rdists;
5679
5680	its_parent = parent_domain;
5681	of_node = to_of_node(handle);
5682	if (of_node)
5683		its_of_probe(of_node);
5684	else
5685		its_acpi_probe();
5686
5687	if (list_empty(&its_nodes)) {
5688		pr_warn("ITS: No ITS available, not enabling LPIs\n");
5689		return -ENXIO;
5690	}
5691
5692	err = allocate_lpi_tables();
5693	if (err)
5694		return err;
5695
5696	list_for_each_entry(its, &its_nodes, entry) {
5697		has_v4 |= is_v4(its);
5698		has_v4_1 |= is_v4_1(its);
5699	}
5700
5701	/* Don't bother with inconsistent systems */
5702	if (WARN_ON(!has_v4_1 && rdists->has_rvpeid))
5703		rdists->has_rvpeid = false;
5704
5705	if (has_v4 & rdists->has_vlpis) {
5706		const struct irq_domain_ops *sgi_ops;
5707
5708		if (has_v4_1)
5709			sgi_ops = &its_sgi_domain_ops;
5710		else
5711			sgi_ops = NULL;
5712
5713		if (its_init_vpe_domain() ||
5714		    its_init_v4(parent_domain, &its_vpe_domain_ops, sgi_ops)) {
5715			rdists->has_vlpis = false;
5716			pr_err("ITS: Disabling GICv4 support\n");
5717		}
5718	}
5719
5720	register_syscore_ops(&its_syscore_ops);
5721
5722	return 0;
5723}