Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.15.
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * Copyright 2016,2017 IBM Corporation.
   4 */
   5
   6#define pr_fmt(fmt) "xive: " fmt
   7
   8#include <linux/types.h>
   9#include <linux/threads.h>
  10#include <linux/kernel.h>
  11#include <linux/irq.h>
  12#include <linux/debugfs.h>
  13#include <linux/smp.h>
  14#include <linux/interrupt.h>
  15#include <linux/seq_file.h>
  16#include <linux/init.h>
  17#include <linux/cpu.h>
  18#include <linux/of.h>
  19#include <linux/slab.h>
  20#include <linux/spinlock.h>
  21#include <linux/msi.h>
  22#include <linux/vmalloc.h>
  23
  24#include <asm/debugfs.h>
  25#include <asm/prom.h>
  26#include <asm/io.h>
  27#include <asm/smp.h>
  28#include <asm/machdep.h>
  29#include <asm/irq.h>
  30#include <asm/errno.h>
  31#include <asm/xive.h>
  32#include <asm/xive-regs.h>
  33#include <asm/xmon.h>
  34
  35#include "xive-internal.h"
  36
  37#undef DEBUG_FLUSH
  38#undef DEBUG_ALL
  39
  40#ifdef DEBUG_ALL
  41#define DBG_VERBOSE(fmt, ...)	pr_devel("cpu %d - " fmt, \
  42					 smp_processor_id(), ## __VA_ARGS__)
  43#else
  44#define DBG_VERBOSE(fmt...)	do { } while(0)
  45#endif
  46
  47bool __xive_enabled;
  48EXPORT_SYMBOL_GPL(__xive_enabled);
  49bool xive_cmdline_disabled;
  50
  51/* We use only one priority for now */
  52static u8 xive_irq_priority;
  53
  54/* TIMA exported to KVM */
  55void __iomem *xive_tima;
  56EXPORT_SYMBOL_GPL(xive_tima);
  57u32 xive_tima_offset;
  58
  59/* Backend ops */
  60static const struct xive_ops *xive_ops;
  61
  62/* Our global interrupt domain */
  63static struct irq_domain *xive_irq_domain;
  64
  65#ifdef CONFIG_SMP
  66/* The IPIs use the same logical irq number when on the same chip */
  67static struct xive_ipi_desc {
  68	unsigned int irq;
  69	char name[16];
  70	atomic_t started;
  71} *xive_ipis;
  72
  73/*
  74 * Use early_cpu_to_node() for hot-plugged CPUs
  75 */
  76static unsigned int xive_ipi_cpu_to_irq(unsigned int cpu)
  77{
  78	return xive_ipis[early_cpu_to_node(cpu)].irq;
  79}
  80#endif
  81
  82/* Xive state for each CPU */
  83static DEFINE_PER_CPU(struct xive_cpu *, xive_cpu);
  84
  85/* An invalid CPU target */
  86#define XIVE_INVALID_TARGET	(-1)
  87
  88/*
  89 * Read the next entry in a queue, return its content if it's valid
  90 * or 0 if there is no new entry.
  91 *
  92 * The queue pointer is moved forward unless "just_peek" is set
  93 */
  94static u32 xive_read_eq(struct xive_q *q, bool just_peek)
  95{
  96	u32 cur;
  97
  98	if (!q->qpage)
  99		return 0;
 100	cur = be32_to_cpup(q->qpage + q->idx);
 101
 102	/* Check valid bit (31) vs current toggle polarity */
 103	if ((cur >> 31) == q->toggle)
 104		return 0;
 105
 106	/* If consuming from the queue ... */
 107	if (!just_peek) {
 108		/* Next entry */
 109		q->idx = (q->idx + 1) & q->msk;
 110
 111		/* Wrap around: flip valid toggle */
 112		if (q->idx == 0)
 113			q->toggle ^= 1;
 114	}
 115	/* Mask out the valid bit (31) */
 116	return cur & 0x7fffffff;
 117}
 118
 119/*
 120 * Scans all the queue that may have interrupts in them
 121 * (based on "pending_prio") in priority order until an
 122 * interrupt is found or all the queues are empty.
 123 *
 124 * Then updates the CPPR (Current Processor Priority
 125 * Register) based on the most favored interrupt found
 126 * (0xff if none) and return what was found (0 if none).
 127 *
 128 * If just_peek is set, return the most favored pending
 129 * interrupt if any but don't update the queue pointers.
 130 *
 131 * Note: This function can operate generically on any number
 132 * of queues (up to 8). The current implementation of the XIVE
 133 * driver only uses a single queue however.
 134 *
 135 * Note2: This will also "flush" "the pending_count" of a queue
 136 * into the "count" when that queue is observed to be empty.
 137 * This is used to keep track of the amount of interrupts
 138 * targetting a queue. When an interrupt is moved away from
 139 * a queue, we only decrement that queue count once the queue
 140 * has been observed empty to avoid races.
 141 */
 142static u32 xive_scan_interrupts(struct xive_cpu *xc, bool just_peek)
 143{
 144	u32 irq = 0;
 145	u8 prio = 0;
 146
 147	/* Find highest pending priority */
 148	while (xc->pending_prio != 0) {
 149		struct xive_q *q;
 150
 151		prio = ffs(xc->pending_prio) - 1;
 152		DBG_VERBOSE("scan_irq: trying prio %d\n", prio);
 153
 154		/* Try to fetch */
 155		irq = xive_read_eq(&xc->queue[prio], just_peek);
 156
 157		/* Found something ? That's it */
 158		if (irq) {
 159			if (just_peek || irq_to_desc(irq))
 160				break;
 161			/*
 162			 * We should never get here; if we do then we must
 163			 * have failed to synchronize the interrupt properly
 164			 * when shutting it down.
 165			 */
 166			pr_crit("xive: got interrupt %d without descriptor, dropping\n",
 167				irq);
 168			WARN_ON(1);
 169			continue;
 170		}
 171
 172		/* Clear pending bits */
 173		xc->pending_prio &= ~(1 << prio);
 174
 175		/*
 176		 * Check if the queue count needs adjusting due to
 177		 * interrupts being moved away. See description of
 178		 * xive_dec_target_count()
 179		 */
 180		q = &xc->queue[prio];
 181		if (atomic_read(&q->pending_count)) {
 182			int p = atomic_xchg(&q->pending_count, 0);
 183			if (p) {
 184				WARN_ON(p > atomic_read(&q->count));
 185				atomic_sub(p, &q->count);
 186			}
 187		}
 188	}
 189
 190	/* If nothing was found, set CPPR to 0xff */
 191	if (irq == 0)
 192		prio = 0xff;
 193
 194	/* Update HW CPPR to match if necessary */
 195	if (prio != xc->cppr) {
 196		DBG_VERBOSE("scan_irq: adjusting CPPR to %d\n", prio);
 197		xc->cppr = prio;
 198		out_8(xive_tima + xive_tima_offset + TM_CPPR, prio);
 199	}
 200
 201	return irq;
 202}
 203
 204/*
 205 * This is used to perform the magic loads from an ESB
 206 * described in xive-regs.h
 207 */
 208static notrace u8 xive_esb_read(struct xive_irq_data *xd, u32 offset)
 209{
 210	u64 val;
 211
 212	if (offset == XIVE_ESB_SET_PQ_10 && xd->flags & XIVE_IRQ_FLAG_STORE_EOI)
 213		offset |= XIVE_ESB_LD_ST_MO;
 214
 215	if ((xd->flags & XIVE_IRQ_FLAG_H_INT_ESB) && xive_ops->esb_rw)
 216		val = xive_ops->esb_rw(xd->hw_irq, offset, 0, 0);
 217	else
 218		val = in_be64(xd->eoi_mmio + offset);
 219
 220	return (u8)val;
 221}
 222
 223static void xive_esb_write(struct xive_irq_data *xd, u32 offset, u64 data)
 224{
 225	if ((xd->flags & XIVE_IRQ_FLAG_H_INT_ESB) && xive_ops->esb_rw)
 226		xive_ops->esb_rw(xd->hw_irq, offset, data, 1);
 227	else
 228		out_be64(xd->eoi_mmio + offset, data);
 229}
 230
 231#ifdef CONFIG_XMON
 232static notrace void xive_dump_eq(const char *name, struct xive_q *q)
 233{
 234	u32 i0, i1, idx;
 235
 236	if (!q->qpage)
 237		return;
 238	idx = q->idx;
 239	i0 = be32_to_cpup(q->qpage + idx);
 240	idx = (idx + 1) & q->msk;
 241	i1 = be32_to_cpup(q->qpage + idx);
 242	xmon_printf("%s idx=%d T=%d %08x %08x ...", name,
 243		     q->idx, q->toggle, i0, i1);
 244}
 245
 246notrace void xmon_xive_do_dump(int cpu)
 247{
 248	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
 249
 250	xmon_printf("CPU %d:", cpu);
 251	if (xc) {
 252		xmon_printf("pp=%02x CPPR=%02x ", xc->pending_prio, xc->cppr);
 253
 254#ifdef CONFIG_SMP
 255		{
 256			u64 val = xive_esb_read(&xc->ipi_data, XIVE_ESB_GET);
 257
 258			xmon_printf("IPI=0x%08x PQ=%c%c ", xc->hw_ipi,
 259				    val & XIVE_ESB_VAL_P ? 'P' : '-',
 260				    val & XIVE_ESB_VAL_Q ? 'Q' : '-');
 261		}
 262#endif
 263		xive_dump_eq("EQ", &xc->queue[xive_irq_priority]);
 264	}
 265	xmon_printf("\n");
 266}
 267
 268static struct irq_data *xive_get_irq_data(u32 hw_irq)
 269{
 270	unsigned int irq = irq_find_mapping(xive_irq_domain, hw_irq);
 271
 272	return irq ? irq_get_irq_data(irq) : NULL;
 273}
 274
 275int xmon_xive_get_irq_config(u32 hw_irq, struct irq_data *d)
 276{
 277	int rc;
 278	u32 target;
 279	u8 prio;
 280	u32 lirq;
 281
 282	rc = xive_ops->get_irq_config(hw_irq, &target, &prio, &lirq);
 283	if (rc) {
 284		xmon_printf("IRQ 0x%08x : no config rc=%d\n", hw_irq, rc);
 285		return rc;
 286	}
 287
 288	xmon_printf("IRQ 0x%08x : target=0x%x prio=%02x lirq=0x%x ",
 289		    hw_irq, target, prio, lirq);
 290
 291	if (!d)
 292		d = xive_get_irq_data(hw_irq);
 293
 294	if (d) {
 295		struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 296		u64 val = xive_esb_read(xd, XIVE_ESB_GET);
 297
 298		xmon_printf("flags=%c%c%c PQ=%c%c",
 299			    xd->flags & XIVE_IRQ_FLAG_STORE_EOI ? 'S' : ' ',
 300			    xd->flags & XIVE_IRQ_FLAG_LSI ? 'L' : ' ',
 301			    xd->flags & XIVE_IRQ_FLAG_H_INT_ESB ? 'H' : ' ',
 302			    val & XIVE_ESB_VAL_P ? 'P' : '-',
 303			    val & XIVE_ESB_VAL_Q ? 'Q' : '-');
 304	}
 305
 306	xmon_printf("\n");
 307	return 0;
 308}
 309
 310void xmon_xive_get_irq_all(void)
 311{
 312	unsigned int i;
 313	struct irq_desc *desc;
 314
 315	for_each_irq_desc(i, desc) {
 316		struct irq_data *d = irq_desc_get_irq_data(desc);
 317		unsigned int hwirq = (unsigned int)irqd_to_hwirq(d);
 318
 319		if (d->domain == xive_irq_domain)
 320			xmon_xive_get_irq_config(hwirq, d);
 321	}
 322}
 323
 324#endif /* CONFIG_XMON */
 325
 326static unsigned int xive_get_irq(void)
 327{
 328	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
 329	u32 irq;
 330
 331	/*
 332	 * This can be called either as a result of a HW interrupt or
 333	 * as a "replay" because EOI decided there was still something
 334	 * in one of the queues.
 335	 *
 336	 * First we perform an ACK cycle in order to update our mask
 337	 * of pending priorities. This will also have the effect of
 338	 * updating the CPPR to the most favored pending interrupts.
 339	 *
 340	 * In the future, if we have a way to differentiate a first
 341	 * entry (on HW interrupt) from a replay triggered by EOI,
 342	 * we could skip this on replays unless we soft-mask tells us
 343	 * that a new HW interrupt occurred.
 344	 */
 345	xive_ops->update_pending(xc);
 346
 347	DBG_VERBOSE("get_irq: pending=%02x\n", xc->pending_prio);
 348
 349	/* Scan our queue(s) for interrupts */
 350	irq = xive_scan_interrupts(xc, false);
 351
 352	DBG_VERBOSE("get_irq: got irq 0x%x, new pending=0x%02x\n",
 353	    irq, xc->pending_prio);
 354
 355	/* Return pending interrupt if any */
 356	if (irq == XIVE_BAD_IRQ)
 357		return 0;
 358	return irq;
 359}
 360
 361/*
 362 * After EOI'ing an interrupt, we need to re-check the queue
 363 * to see if another interrupt is pending since multiple
 364 * interrupts can coalesce into a single notification to the
 365 * CPU.
 366 *
 367 * If we find that there is indeed more in there, we call
 368 * force_external_irq_replay() to make Linux synthetize an
 369 * external interrupt on the next call to local_irq_restore().
 370 */
 371static void xive_do_queue_eoi(struct xive_cpu *xc)
 372{
 373	if (xive_scan_interrupts(xc, true) != 0) {
 374		DBG_VERBOSE("eoi: pending=0x%02x\n", xc->pending_prio);
 375		force_external_irq_replay();
 376	}
 377}
 378
 379/*
 380 * EOI an interrupt at the source. There are several methods
 381 * to do this depending on the HW version and source type
 382 */
 383static void xive_do_source_eoi(struct xive_irq_data *xd)
 384{
 385	u8 eoi_val;
 386
 387	xd->stale_p = false;
 388
 389	/* If the XIVE supports the new "store EOI facility, use it */
 390	if (xd->flags & XIVE_IRQ_FLAG_STORE_EOI) {
 391		xive_esb_write(xd, XIVE_ESB_STORE_EOI, 0);
 392		return;
 393	}
 394
 395	/*
 396	 * For LSIs, we use the "EOI cycle" special load rather than
 397	 * PQ bits, as they are automatically re-triggered in HW when
 398	 * still pending.
 399	 */
 400	if (xd->flags & XIVE_IRQ_FLAG_LSI) {
 401		xive_esb_read(xd, XIVE_ESB_LOAD_EOI);
 402		return;
 403	}
 404
 405	/*
 406	 * Otherwise, we use the special MMIO that does a clear of
 407	 * both P and Q and returns the old Q. This allows us to then
 408	 * do a re-trigger if Q was set rather than synthesizing an
 409	 * interrupt in software
 410	 */
 411	eoi_val = xive_esb_read(xd, XIVE_ESB_SET_PQ_00);
 412	DBG_VERBOSE("eoi_val=%x\n", eoi_val);
 413
 414	/* Re-trigger if needed */
 415	if ((eoi_val & XIVE_ESB_VAL_Q) && xd->trig_mmio)
 416		out_be64(xd->trig_mmio, 0);
 417}
 418
 419/* irq_chip eoi callback, called with irq descriptor lock held */
 420static void xive_irq_eoi(struct irq_data *d)
 421{
 422	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 423	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
 424
 425	DBG_VERBOSE("eoi_irq: irq=%d [0x%lx] pending=%02x\n",
 426		    d->irq, irqd_to_hwirq(d), xc->pending_prio);
 427
 428	/*
 429	 * EOI the source if it hasn't been disabled and hasn't
 430	 * been passed-through to a KVM guest
 431	 */
 432	if (!irqd_irq_disabled(d) && !irqd_is_forwarded_to_vcpu(d) &&
 433	    !(xd->flags & XIVE_IRQ_FLAG_NO_EOI))
 434		xive_do_source_eoi(xd);
 435	else
 436		xd->stale_p = true;
 437
 438	/*
 439	 * Clear saved_p to indicate that it's no longer occupying
 440	 * a queue slot on the target queue
 441	 */
 442	xd->saved_p = false;
 443
 444	/* Check for more work in the queue */
 445	xive_do_queue_eoi(xc);
 446}
 447
 448/*
 449 * Helper used to mask and unmask an interrupt source.
 450 */
 451static void xive_do_source_set_mask(struct xive_irq_data *xd,
 452				    bool mask)
 453{
 454	u64 val;
 455
 456	/*
 457	 * If the interrupt had P set, it may be in a queue.
 458	 *
 459	 * We need to make sure we don't re-enable it until it
 460	 * has been fetched from that queue and EOId. We keep
 461	 * a copy of that P state and use it to restore the
 462	 * ESB accordingly on unmask.
 463	 */
 464	if (mask) {
 465		val = xive_esb_read(xd, XIVE_ESB_SET_PQ_01);
 466		if (!xd->stale_p && !!(val & XIVE_ESB_VAL_P))
 467			xd->saved_p = true;
 468		xd->stale_p = false;
 469	} else if (xd->saved_p) {
 470		xive_esb_read(xd, XIVE_ESB_SET_PQ_10);
 471		xd->saved_p = false;
 472	} else {
 473		xive_esb_read(xd, XIVE_ESB_SET_PQ_00);
 474		xd->stale_p = false;
 475	}
 476}
 477
 478/*
 479 * Try to chose "cpu" as a new interrupt target. Increments
 480 * the queue accounting for that target if it's not already
 481 * full.
 482 */
 483static bool xive_try_pick_target(int cpu)
 484{
 485	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
 486	struct xive_q *q = &xc->queue[xive_irq_priority];
 487	int max;
 488
 489	/*
 490	 * Calculate max number of interrupts in that queue.
 491	 *
 492	 * We leave a gap of 1 just in case...
 493	 */
 494	max = (q->msk + 1) - 1;
 495	return !!atomic_add_unless(&q->count, 1, max);
 496}
 497
 498/*
 499 * Un-account an interrupt for a target CPU. We don't directly
 500 * decrement q->count since the interrupt might still be present
 501 * in the queue.
 502 *
 503 * Instead increment a separate counter "pending_count" which
 504 * will be substracted from "count" later when that CPU observes
 505 * the queue to be empty.
 506 */
 507static void xive_dec_target_count(int cpu)
 508{
 509	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
 510	struct xive_q *q = &xc->queue[xive_irq_priority];
 511
 512	if (WARN_ON(cpu < 0 || !xc)) {
 513		pr_err("%s: cpu=%d xc=%p\n", __func__, cpu, xc);
 514		return;
 515	}
 516
 517	/*
 518	 * We increment the "pending count" which will be used
 519	 * to decrement the target queue count whenever it's next
 520	 * processed and found empty. This ensure that we don't
 521	 * decrement while we still have the interrupt there
 522	 * occupying a slot.
 523	 */
 524	atomic_inc(&q->pending_count);
 525}
 526
 527/* Find a tentative CPU target in a CPU mask */
 528static int xive_find_target_in_mask(const struct cpumask *mask,
 529				    unsigned int fuzz)
 530{
 531	int cpu, first, num, i;
 532
 533	/* Pick up a starting point CPU in the mask based on  fuzz */
 534	num = min_t(int, cpumask_weight(mask), nr_cpu_ids);
 535	first = fuzz % num;
 536
 537	/* Locate it */
 538	cpu = cpumask_first(mask);
 539	for (i = 0; i < first && cpu < nr_cpu_ids; i++)
 540		cpu = cpumask_next(cpu, mask);
 541
 542	/* Sanity check */
 543	if (WARN_ON(cpu >= nr_cpu_ids))
 544		cpu = cpumask_first(cpu_online_mask);
 545
 546	/* Remember first one to handle wrap-around */
 547	first = cpu;
 548
 549	/*
 550	 * Now go through the entire mask until we find a valid
 551	 * target.
 552	 */
 553	do {
 554		/*
 555		 * We re-check online as the fallback case passes us
 556		 * an untested affinity mask
 557		 */
 558		if (cpu_online(cpu) && xive_try_pick_target(cpu))
 559			return cpu;
 560		cpu = cpumask_next(cpu, mask);
 561		/* Wrap around */
 562		if (cpu >= nr_cpu_ids)
 563			cpu = cpumask_first(mask);
 564	} while (cpu != first);
 565
 566	return -1;
 567}
 568
 569/*
 570 * Pick a target CPU for an interrupt. This is done at
 571 * startup or if the affinity is changed in a way that
 572 * invalidates the current target.
 573 */
 574static int xive_pick_irq_target(struct irq_data *d,
 575				const struct cpumask *affinity)
 576{
 577	static unsigned int fuzz;
 578	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 579	cpumask_var_t mask;
 580	int cpu = -1;
 581
 582	/*
 583	 * If we have chip IDs, first we try to build a mask of
 584	 * CPUs matching the CPU and find a target in there
 585	 */
 586	if (xd->src_chip != XIVE_INVALID_CHIP_ID &&
 587		zalloc_cpumask_var(&mask, GFP_ATOMIC)) {
 588		/* Build a mask of matching chip IDs */
 589		for_each_cpu_and(cpu, affinity, cpu_online_mask) {
 590			struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
 591			if (xc->chip_id == xd->src_chip)
 592				cpumask_set_cpu(cpu, mask);
 593		}
 594		/* Try to find a target */
 595		if (cpumask_empty(mask))
 596			cpu = -1;
 597		else
 598			cpu = xive_find_target_in_mask(mask, fuzz++);
 599		free_cpumask_var(mask);
 600		if (cpu >= 0)
 601			return cpu;
 602		fuzz--;
 603	}
 604
 605	/* No chip IDs, fallback to using the affinity mask */
 606	return xive_find_target_in_mask(affinity, fuzz++);
 607}
 608
 609static unsigned int xive_irq_startup(struct irq_data *d)
 610{
 611	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 612	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 613	int target, rc;
 614
 615	xd->saved_p = false;
 616	xd->stale_p = false;
 617	pr_devel("xive_irq_startup: irq %d [0x%x] data @%p\n",
 618		 d->irq, hw_irq, d);
 619
 620#ifdef CONFIG_PCI_MSI
 621	/*
 622	 * The generic MSI code returns with the interrupt disabled on the
 623	 * card, using the MSI mask bits. Firmware doesn't appear to unmask
 624	 * at that level, so we do it here by hand.
 625	 */
 626	if (irq_data_get_msi_desc(d))
 627		pci_msi_unmask_irq(d);
 628#endif
 629
 630	/* Pick a target */
 631	target = xive_pick_irq_target(d, irq_data_get_affinity_mask(d));
 632	if (target == XIVE_INVALID_TARGET) {
 633		/* Try again breaking affinity */
 634		target = xive_pick_irq_target(d, cpu_online_mask);
 635		if (target == XIVE_INVALID_TARGET)
 636			return -ENXIO;
 637		pr_warn("irq %d started with broken affinity\n", d->irq);
 638	}
 639
 640	/* Sanity check */
 641	if (WARN_ON(target == XIVE_INVALID_TARGET ||
 642		    target >= nr_cpu_ids))
 643		target = smp_processor_id();
 644
 645	xd->target = target;
 646
 647	/*
 648	 * Configure the logical number to be the Linux IRQ number
 649	 * and set the target queue
 650	 */
 651	rc = xive_ops->configure_irq(hw_irq,
 652				     get_hard_smp_processor_id(target),
 653				     xive_irq_priority, d->irq);
 654	if (rc)
 655		return rc;
 656
 657	/* Unmask the ESB */
 658	xive_do_source_set_mask(xd, false);
 659
 660	return 0;
 661}
 662
 663/* called with irq descriptor lock held */
 664static void xive_irq_shutdown(struct irq_data *d)
 665{
 666	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 667	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 668
 669	pr_devel("xive_irq_shutdown: irq %d [0x%x] data @%p\n",
 670		 d->irq, hw_irq, d);
 671
 672	if (WARN_ON(xd->target == XIVE_INVALID_TARGET))
 673		return;
 674
 675	/* Mask the interrupt at the source */
 676	xive_do_source_set_mask(xd, true);
 677
 678	/*
 679	 * Mask the interrupt in HW in the IVT/EAS and set the number
 680	 * to be the "bad" IRQ number
 681	 */
 682	xive_ops->configure_irq(hw_irq,
 683				get_hard_smp_processor_id(xd->target),
 684				0xff, XIVE_BAD_IRQ);
 685
 686	xive_dec_target_count(xd->target);
 687	xd->target = XIVE_INVALID_TARGET;
 688}
 689
 690static void xive_irq_unmask(struct irq_data *d)
 691{
 692	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 693
 694	pr_devel("xive_irq_unmask: irq %d data @%p\n", d->irq, xd);
 695
 696	xive_do_source_set_mask(xd, false);
 697}
 698
 699static void xive_irq_mask(struct irq_data *d)
 700{
 701	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 702
 703	pr_devel("xive_irq_mask: irq %d data @%p\n", d->irq, xd);
 704
 705	xive_do_source_set_mask(xd, true);
 706}
 707
 708static int xive_irq_set_affinity(struct irq_data *d,
 709				 const struct cpumask *cpumask,
 710				 bool force)
 711{
 712	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 713	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 714	u32 target, old_target;
 715	int rc = 0;
 716
 717	pr_devel("xive_irq_set_affinity: irq %d\n", d->irq);
 718
 719	/* Is this valid ? */
 720	if (cpumask_any_and(cpumask, cpu_online_mask) >= nr_cpu_ids)
 721		return -EINVAL;
 722
 723	/* Don't do anything if the interrupt isn't started */
 724	if (!irqd_is_started(d))
 725		return IRQ_SET_MASK_OK;
 726
 727	/*
 728	 * If existing target is already in the new mask, and is
 729	 * online then do nothing.
 730	 */
 731	if (xd->target != XIVE_INVALID_TARGET &&
 732	    cpu_online(xd->target) &&
 733	    cpumask_test_cpu(xd->target, cpumask))
 734		return IRQ_SET_MASK_OK;
 735
 736	/* Pick a new target */
 737	target = xive_pick_irq_target(d, cpumask);
 738
 739	/* No target found */
 740	if (target == XIVE_INVALID_TARGET)
 741		return -ENXIO;
 742
 743	/* Sanity check */
 744	if (WARN_ON(target >= nr_cpu_ids))
 745		target = smp_processor_id();
 746
 747	old_target = xd->target;
 748
 749	/*
 750	 * Only configure the irq if it's not currently passed-through to
 751	 * a KVM guest
 752	 */
 753	if (!irqd_is_forwarded_to_vcpu(d))
 754		rc = xive_ops->configure_irq(hw_irq,
 755					     get_hard_smp_processor_id(target),
 756					     xive_irq_priority, d->irq);
 757	if (rc < 0) {
 758		pr_err("Error %d reconfiguring irq %d\n", rc, d->irq);
 759		return rc;
 760	}
 761
 762	pr_devel("  target: 0x%x\n", target);
 763	xd->target = target;
 764
 765	/* Give up previous target */
 766	if (old_target != XIVE_INVALID_TARGET)
 767	    xive_dec_target_count(old_target);
 768
 769	return IRQ_SET_MASK_OK;
 770}
 771
 772static int xive_irq_set_type(struct irq_data *d, unsigned int flow_type)
 773{
 774	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 775
 776	/*
 777	 * We only support these. This has really no effect other than setting
 778	 * the corresponding descriptor bits mind you but those will in turn
 779	 * affect the resend function when re-enabling an edge interrupt.
 780	 *
 781	 * Set set the default to edge as explained in map().
 782	 */
 783	if (flow_type == IRQ_TYPE_DEFAULT || flow_type == IRQ_TYPE_NONE)
 784		flow_type = IRQ_TYPE_EDGE_RISING;
 785
 786	if (flow_type != IRQ_TYPE_EDGE_RISING &&
 787	    flow_type != IRQ_TYPE_LEVEL_LOW)
 788		return -EINVAL;
 789
 790	irqd_set_trigger_type(d, flow_type);
 791
 792	/*
 793	 * Double check it matches what the FW thinks
 794	 *
 795	 * NOTE: We don't know yet if the PAPR interface will provide
 796	 * the LSI vs MSI information apart from the device-tree so
 797	 * this check might have to move into an optional backend call
 798	 * that is specific to the native backend
 799	 */
 800	if ((flow_type == IRQ_TYPE_LEVEL_LOW) !=
 801	    !!(xd->flags & XIVE_IRQ_FLAG_LSI)) {
 802		pr_warn("Interrupt %d (HW 0x%x) type mismatch, Linux says %s, FW says %s\n",
 803			d->irq, (u32)irqd_to_hwirq(d),
 804			(flow_type == IRQ_TYPE_LEVEL_LOW) ? "Level" : "Edge",
 805			(xd->flags & XIVE_IRQ_FLAG_LSI) ? "Level" : "Edge");
 806	}
 807
 808	return IRQ_SET_MASK_OK_NOCOPY;
 809}
 810
 811static int xive_irq_retrigger(struct irq_data *d)
 812{
 813	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 814
 815	/* This should be only for MSIs */
 816	if (WARN_ON(xd->flags & XIVE_IRQ_FLAG_LSI))
 817		return 0;
 818
 819	/*
 820	 * To perform a retrigger, we first set the PQ bits to
 821	 * 11, then perform an EOI.
 822	 */
 823	xive_esb_read(xd, XIVE_ESB_SET_PQ_11);
 824	xive_do_source_eoi(xd);
 825
 826	return 1;
 827}
 828
 829/*
 830 * Caller holds the irq descriptor lock, so this won't be called
 831 * concurrently with xive_get_irqchip_state on the same interrupt.
 832 */
 833static int xive_irq_set_vcpu_affinity(struct irq_data *d, void *state)
 834{
 835	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 836	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 837	int rc;
 838	u8 pq;
 839
 840	/*
 841	 * This is called by KVM with state non-NULL for enabling
 842	 * pass-through or NULL for disabling it
 843	 */
 844	if (state) {
 845		irqd_set_forwarded_to_vcpu(d);
 846
 847		/* Set it to PQ=10 state to prevent further sends */
 848		pq = xive_esb_read(xd, XIVE_ESB_SET_PQ_10);
 849		if (!xd->stale_p) {
 850			xd->saved_p = !!(pq & XIVE_ESB_VAL_P);
 851			xd->stale_p = !xd->saved_p;
 852		}
 853
 854		/* No target ? nothing to do */
 855		if (xd->target == XIVE_INVALID_TARGET) {
 856			/*
 857			 * An untargetted interrupt should have been
 858			 * also masked at the source
 859			 */
 860			WARN_ON(xd->saved_p);
 861
 862			return 0;
 863		}
 864
 865		/*
 866		 * If P was set, adjust state to PQ=11 to indicate
 867		 * that a resend is needed for the interrupt to reach
 868		 * the guest. Also remember the value of P.
 869		 *
 870		 * This also tells us that it's in flight to a host queue
 871		 * or has already been fetched but hasn't been EOIed yet
 872		 * by the host. This it's potentially using up a host
 873		 * queue slot. This is important to know because as long
 874		 * as this is the case, we must not hard-unmask it when
 875		 * "returning" that interrupt to the host.
 876		 *
 877		 * This saved_p is cleared by the host EOI, when we know
 878		 * for sure the queue slot is no longer in use.
 879		 */
 880		if (xd->saved_p) {
 881			xive_esb_read(xd, XIVE_ESB_SET_PQ_11);
 882
 883			/*
 884			 * Sync the XIVE source HW to ensure the interrupt
 885			 * has gone through the EAS before we change its
 886			 * target to the guest. That should guarantee us
 887			 * that we *will* eventually get an EOI for it on
 888			 * the host. Otherwise there would be a small window
 889			 * for P to be seen here but the interrupt going
 890			 * to the guest queue.
 891			 */
 892			if (xive_ops->sync_source)
 893				xive_ops->sync_source(hw_irq);
 894		}
 895	} else {
 896		irqd_clr_forwarded_to_vcpu(d);
 897
 898		/* No host target ? hard mask and return */
 899		if (xd->target == XIVE_INVALID_TARGET) {
 900			xive_do_source_set_mask(xd, true);
 901			return 0;
 902		}
 903
 904		/*
 905		 * Sync the XIVE source HW to ensure the interrupt
 906		 * has gone through the EAS before we change its
 907		 * target to the host.
 908		 */
 909		if (xive_ops->sync_source)
 910			xive_ops->sync_source(hw_irq);
 911
 912		/*
 913		 * By convention we are called with the interrupt in
 914		 * a PQ=10 or PQ=11 state, ie, it won't fire and will
 915		 * have latched in Q whether there's a pending HW
 916		 * interrupt or not.
 917		 *
 918		 * First reconfigure the target.
 919		 */
 920		rc = xive_ops->configure_irq(hw_irq,
 921					     get_hard_smp_processor_id(xd->target),
 922					     xive_irq_priority, d->irq);
 923		if (rc)
 924			return rc;
 925
 926		/*
 927		 * Then if saved_p is not set, effectively re-enable the
 928		 * interrupt with an EOI. If it is set, we know there is
 929		 * still a message in a host queue somewhere that will be
 930		 * EOId eventually.
 931		 *
 932		 * Note: We don't check irqd_irq_disabled(). Effectively,
 933		 * we *will* let the irq get through even if masked if the
 934		 * HW is still firing it in order to deal with the whole
 935		 * saved_p business properly. If the interrupt triggers
 936		 * while masked, the generic code will re-mask it anyway.
 937		 */
 938		if (!xd->saved_p)
 939			xive_do_source_eoi(xd);
 940
 941	}
 942	return 0;
 943}
 944
 945/* Called with irq descriptor lock held. */
 946static int xive_get_irqchip_state(struct irq_data *data,
 947				  enum irqchip_irq_state which, bool *state)
 948{
 949	struct xive_irq_data *xd = irq_data_get_irq_handler_data(data);
 950	u8 pq;
 951
 952	switch (which) {
 953	case IRQCHIP_STATE_ACTIVE:
 954		pq = xive_esb_read(xd, XIVE_ESB_GET);
 955
 956		/*
 957		 * The esb value being all 1's means we couldn't get
 958		 * the PQ state of the interrupt through mmio. It may
 959		 * happen, for example when querying a PHB interrupt
 960		 * while the PHB is in an error state. We consider the
 961		 * interrupt to be inactive in that case.
 962		 */
 963		*state = (pq != XIVE_ESB_INVALID) && !xd->stale_p &&
 964			(xd->saved_p || (!!(pq & XIVE_ESB_VAL_P) &&
 965			 !irqd_irq_disabled(data)));
 966		return 0;
 967	default:
 968		return -EINVAL;
 969	}
 970}
 971
 972static struct irq_chip xive_irq_chip = {
 973	.name = "XIVE-IRQ",
 974	.irq_startup = xive_irq_startup,
 975	.irq_shutdown = xive_irq_shutdown,
 976	.irq_eoi = xive_irq_eoi,
 977	.irq_mask = xive_irq_mask,
 978	.irq_unmask = xive_irq_unmask,
 979	.irq_set_affinity = xive_irq_set_affinity,
 980	.irq_set_type = xive_irq_set_type,
 981	.irq_retrigger = xive_irq_retrigger,
 982	.irq_set_vcpu_affinity = xive_irq_set_vcpu_affinity,
 983	.irq_get_irqchip_state = xive_get_irqchip_state,
 984};
 985
 986bool is_xive_irq(struct irq_chip *chip)
 987{
 988	return chip == &xive_irq_chip;
 989}
 990EXPORT_SYMBOL_GPL(is_xive_irq);
 991
 992void xive_cleanup_irq_data(struct xive_irq_data *xd)
 993{
 994	if (xd->eoi_mmio) {
 995		iounmap(xd->eoi_mmio);
 996		if (xd->eoi_mmio == xd->trig_mmio)
 997			xd->trig_mmio = NULL;
 998		xd->eoi_mmio = NULL;
 999	}
1000	if (xd->trig_mmio) {
1001		iounmap(xd->trig_mmio);
1002		xd->trig_mmio = NULL;
1003	}
1004}
1005EXPORT_SYMBOL_GPL(xive_cleanup_irq_data);
1006
1007static int xive_irq_alloc_data(unsigned int virq, irq_hw_number_t hw)
1008{
1009	struct xive_irq_data *xd;
1010	int rc;
1011
1012	xd = kzalloc(sizeof(struct xive_irq_data), GFP_KERNEL);
1013	if (!xd)
1014		return -ENOMEM;
1015	rc = xive_ops->populate_irq_data(hw, xd);
1016	if (rc) {
1017		kfree(xd);
1018		return rc;
1019	}
1020	xd->target = XIVE_INVALID_TARGET;
1021	irq_set_handler_data(virq, xd);
1022
1023	/*
1024	 * Turn OFF by default the interrupt being mapped. A side
1025	 * effect of this check is the mapping the ESB page of the
1026	 * interrupt in the Linux address space. This prevents page
1027	 * fault issues in the crash handler which masks all
1028	 * interrupts.
1029	 */
1030	xive_esb_read(xd, XIVE_ESB_SET_PQ_01);
1031
1032	return 0;
1033}
1034
1035static void xive_irq_free_data(unsigned int virq)
1036{
1037	struct xive_irq_data *xd = irq_get_handler_data(virq);
1038
1039	if (!xd)
1040		return;
1041	irq_set_handler_data(virq, NULL);
1042	xive_cleanup_irq_data(xd);
1043	kfree(xd);
1044}
1045
1046#ifdef CONFIG_SMP
1047
1048static void xive_cause_ipi(int cpu)
1049{
1050	struct xive_cpu *xc;
1051	struct xive_irq_data *xd;
1052
1053	xc = per_cpu(xive_cpu, cpu);
1054
1055	DBG_VERBOSE("IPI CPU %d -> %d (HW IRQ 0x%x)\n",
1056		    smp_processor_id(), cpu, xc->hw_ipi);
1057
1058	xd = &xc->ipi_data;
1059	if (WARN_ON(!xd->trig_mmio))
1060		return;
1061	out_be64(xd->trig_mmio, 0);
1062}
1063
1064static irqreturn_t xive_muxed_ipi_action(int irq, void *dev_id)
1065{
1066	return smp_ipi_demux();
1067}
1068
1069static void xive_ipi_eoi(struct irq_data *d)
1070{
1071	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1072
1073	/* Handle possible race with unplug and drop stale IPIs */
1074	if (!xc)
1075		return;
1076
1077	DBG_VERBOSE("IPI eoi: irq=%d [0x%lx] (HW IRQ 0x%x) pending=%02x\n",
1078		    d->irq, irqd_to_hwirq(d), xc->hw_ipi, xc->pending_prio);
1079
1080	xive_do_source_eoi(&xc->ipi_data);
1081	xive_do_queue_eoi(xc);
1082}
1083
1084static void xive_ipi_do_nothing(struct irq_data *d)
1085{
1086	/*
1087	 * Nothing to do, we never mask/unmask IPIs, but the callback
1088	 * has to exist for the struct irq_chip.
1089	 */
1090}
1091
1092static struct irq_chip xive_ipi_chip = {
1093	.name = "XIVE-IPI",
1094	.irq_eoi = xive_ipi_eoi,
1095	.irq_mask = xive_ipi_do_nothing,
1096	.irq_unmask = xive_ipi_do_nothing,
1097};
1098
1099/*
1100 * IPIs are marked per-cpu. We use separate HW interrupts under the
1101 * hood but associated with the same "linux" interrupt
1102 */
1103struct xive_ipi_alloc_info {
1104	irq_hw_number_t hwirq;
1105};
1106
1107static int xive_ipi_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
1108				     unsigned int nr_irqs, void *arg)
1109{
1110	struct xive_ipi_alloc_info *info = arg;
1111	int i;
1112
1113	for (i = 0; i < nr_irqs; i++) {
1114		irq_domain_set_info(domain, virq + i, info->hwirq + i, &xive_ipi_chip,
1115				    domain->host_data, handle_percpu_irq,
1116				    NULL, NULL);
1117	}
1118	return 0;
1119}
1120
1121static const struct irq_domain_ops xive_ipi_irq_domain_ops = {
1122	.alloc  = xive_ipi_irq_domain_alloc,
1123};
1124
1125static int __init xive_init_ipis(void)
1126{
1127	struct fwnode_handle *fwnode;
1128	struct irq_domain *ipi_domain;
1129	unsigned int node;
1130	int ret = -ENOMEM;
1131
1132	fwnode = irq_domain_alloc_named_fwnode("XIVE-IPI");
1133	if (!fwnode)
1134		goto out;
1135
1136	ipi_domain = irq_domain_create_linear(fwnode, nr_node_ids,
1137					      &xive_ipi_irq_domain_ops, NULL);
1138	if (!ipi_domain)
1139		goto out_free_fwnode;
1140
1141	xive_ipis = kcalloc(nr_node_ids, sizeof(*xive_ipis), GFP_KERNEL | __GFP_NOFAIL);
1142	if (!xive_ipis)
1143		goto out_free_domain;
1144
1145	for_each_node(node) {
1146		struct xive_ipi_desc *xid = &xive_ipis[node];
1147		struct xive_ipi_alloc_info info = { node };
1148
1149		/*
1150		 * Map one IPI interrupt per node for all cpus of that node.
1151		 * Since the HW interrupt number doesn't have any meaning,
1152		 * simply use the node number.
1153		 */
1154		ret = irq_domain_alloc_irqs(ipi_domain, 1, node, &info);
1155		if (ret < 0)
1156			goto out_free_xive_ipis;
1157		xid->irq = ret;
1158
1159		snprintf(xid->name, sizeof(xid->name), "IPI-%d", node);
1160	}
1161
1162	return ret;
1163
1164out_free_xive_ipis:
1165	kfree(xive_ipis);
1166out_free_domain:
1167	irq_domain_remove(ipi_domain);
1168out_free_fwnode:
1169	irq_domain_free_fwnode(fwnode);
1170out:
1171	return ret;
1172}
1173
1174static int xive_request_ipi(unsigned int cpu)
1175{
1176	struct xive_ipi_desc *xid = &xive_ipis[early_cpu_to_node(cpu)];
1177	int ret;
1178
1179	if (atomic_inc_return(&xid->started) > 1)
1180		return 0;
1181
1182	ret = request_irq(xid->irq, xive_muxed_ipi_action,
1183			  IRQF_PERCPU | IRQF_NO_THREAD,
1184			  xid->name, NULL);
1185
1186	WARN(ret < 0, "Failed to request IPI %d: %d\n", xid->irq, ret);
1187	return ret;
1188}
1189
1190static int xive_setup_cpu_ipi(unsigned int cpu)
1191{
1192	unsigned int xive_ipi_irq = xive_ipi_cpu_to_irq(cpu);
1193	struct xive_cpu *xc;
1194	int rc;
1195
1196	pr_debug("Setting up IPI for CPU %d\n", cpu);
1197
1198	xc = per_cpu(xive_cpu, cpu);
1199
1200	/* Check if we are already setup */
1201	if (xc->hw_ipi != XIVE_BAD_IRQ)
1202		return 0;
1203
1204	/* Register the IPI */
1205	xive_request_ipi(cpu);
1206
1207	/* Grab an IPI from the backend, this will populate xc->hw_ipi */
1208	if (xive_ops->get_ipi(cpu, xc))
1209		return -EIO;
1210
1211	/*
1212	 * Populate the IRQ data in the xive_cpu structure and
1213	 * configure the HW / enable the IPIs.
1214	 */
1215	rc = xive_ops->populate_irq_data(xc->hw_ipi, &xc->ipi_data);
1216	if (rc) {
1217		pr_err("Failed to populate IPI data on CPU %d\n", cpu);
1218		return -EIO;
1219	}
1220	rc = xive_ops->configure_irq(xc->hw_ipi,
1221				     get_hard_smp_processor_id(cpu),
1222				     xive_irq_priority, xive_ipi_irq);
1223	if (rc) {
1224		pr_err("Failed to map IPI CPU %d\n", cpu);
1225		return -EIO;
1226	}
1227	pr_devel("CPU %d HW IPI %x, virq %d, trig_mmio=%p\n", cpu,
1228	    xc->hw_ipi, xive_ipi_irq, xc->ipi_data.trig_mmio);
1229
1230	/* Unmask it */
1231	xive_do_source_set_mask(&xc->ipi_data, false);
1232
1233	return 0;
1234}
1235
1236static void xive_cleanup_cpu_ipi(unsigned int cpu, struct xive_cpu *xc)
1237{
1238	unsigned int xive_ipi_irq = xive_ipi_cpu_to_irq(cpu);
1239
1240	/* Disable the IPI and free the IRQ data */
1241
1242	/* Already cleaned up ? */
1243	if (xc->hw_ipi == XIVE_BAD_IRQ)
1244		return;
1245
1246	/* TODO: clear IPI mapping */
1247
1248	/* Mask the IPI */
1249	xive_do_source_set_mask(&xc->ipi_data, true);
1250
1251	/*
1252	 * Note: We don't call xive_cleanup_irq_data() to free
1253	 * the mappings as this is called from an IPI on kexec
1254	 * which is not a safe environment to call iounmap()
1255	 */
1256
1257	/* Deconfigure/mask in the backend */
1258	xive_ops->configure_irq(xc->hw_ipi, hard_smp_processor_id(),
1259				0xff, xive_ipi_irq);
1260
1261	/* Free the IPIs in the backend */
1262	xive_ops->put_ipi(cpu, xc);
1263}
1264
1265void __init xive_smp_probe(void)
1266{
1267	smp_ops->cause_ipi = xive_cause_ipi;
1268
1269	/* Register the IPI */
1270	xive_init_ipis();
1271
1272	/* Allocate and setup IPI for the boot CPU */
1273	xive_setup_cpu_ipi(smp_processor_id());
1274}
1275
1276#endif /* CONFIG_SMP */
1277
1278static int xive_irq_domain_map(struct irq_domain *h, unsigned int virq,
1279			       irq_hw_number_t hw)
1280{
1281	int rc;
1282
1283	/*
1284	 * Mark interrupts as edge sensitive by default so that resend
1285	 * actually works. Will fix that up below if needed.
1286	 */
1287	irq_clear_status_flags(virq, IRQ_LEVEL);
1288
1289	rc = xive_irq_alloc_data(virq, hw);
1290	if (rc)
1291		return rc;
1292
1293	irq_set_chip_and_handler(virq, &xive_irq_chip, handle_fasteoi_irq);
1294
1295	return 0;
1296}
1297
1298static void xive_irq_domain_unmap(struct irq_domain *d, unsigned int virq)
1299{
1300	xive_irq_free_data(virq);
1301}
1302
1303static int xive_irq_domain_xlate(struct irq_domain *h, struct device_node *ct,
1304				 const u32 *intspec, unsigned int intsize,
1305				 irq_hw_number_t *out_hwirq, unsigned int *out_flags)
1306
1307{
1308	*out_hwirq = intspec[0];
1309
1310	/*
1311	 * If intsize is at least 2, we look for the type in the second cell,
1312	 * we assume the LSB indicates a level interrupt.
1313	 */
1314	if (intsize > 1) {
1315		if (intspec[1] & 1)
1316			*out_flags = IRQ_TYPE_LEVEL_LOW;
1317		else
1318			*out_flags = IRQ_TYPE_EDGE_RISING;
1319	} else
1320		*out_flags = IRQ_TYPE_LEVEL_LOW;
1321
1322	return 0;
1323}
1324
1325static int xive_irq_domain_match(struct irq_domain *h, struct device_node *node,
1326				 enum irq_domain_bus_token bus_token)
1327{
1328	return xive_ops->match(node);
1329}
1330
1331#ifdef CONFIG_GENERIC_IRQ_DEBUGFS
1332static const char * const esb_names[] = { "RESET", "OFF", "PENDING", "QUEUED" };
1333
1334static const struct {
1335	u64  mask;
1336	char *name;
1337} xive_irq_flags[] = {
1338	{ XIVE_IRQ_FLAG_STORE_EOI, "STORE_EOI" },
1339	{ XIVE_IRQ_FLAG_LSI,       "LSI"       },
1340	{ XIVE_IRQ_FLAG_H_INT_ESB, "H_INT_ESB" },
1341	{ XIVE_IRQ_FLAG_NO_EOI,    "NO_EOI"    },
1342};
1343
1344static void xive_irq_domain_debug_show(struct seq_file *m, struct irq_domain *d,
1345				       struct irq_data *irqd, int ind)
1346{
1347	struct xive_irq_data *xd;
1348	u64 val;
1349	int i;
1350
1351	/* No IRQ domain level information. To be done */
1352	if (!irqd)
1353		return;
1354
1355	if (!is_xive_irq(irq_data_get_irq_chip(irqd)))
1356		return;
1357
1358	seq_printf(m, "%*sXIVE:\n", ind, "");
1359	ind++;
1360
1361	xd = irq_data_get_irq_handler_data(irqd);
1362	if (!xd) {
1363		seq_printf(m, "%*snot assigned\n", ind, "");
1364		return;
1365	}
1366
1367	val = xive_esb_read(xd, XIVE_ESB_GET);
1368	seq_printf(m, "%*sESB:      %s\n", ind, "", esb_names[val & 0x3]);
1369	seq_printf(m, "%*sPstate:   %s %s\n", ind, "", xd->stale_p ? "stale" : "",
1370		   xd->saved_p ? "saved" : "");
1371	seq_printf(m, "%*sTarget:   %d\n", ind, "", xd->target);
1372	seq_printf(m, "%*sChip:     %d\n", ind, "", xd->src_chip);
1373	seq_printf(m, "%*sTrigger:  0x%016llx\n", ind, "", xd->trig_page);
1374	seq_printf(m, "%*sEOI:      0x%016llx\n", ind, "", xd->eoi_page);
1375	seq_printf(m, "%*sFlags:    0x%llx\n", ind, "", xd->flags);
1376	for (i = 0; i < ARRAY_SIZE(xive_irq_flags); i++) {
1377		if (xd->flags & xive_irq_flags[i].mask)
1378			seq_printf(m, "%*s%s\n", ind + 12, "", xive_irq_flags[i].name);
1379	}
1380}
1381#endif
1382
1383static const struct irq_domain_ops xive_irq_domain_ops = {
1384	.match = xive_irq_domain_match,
1385	.map = xive_irq_domain_map,
1386	.unmap = xive_irq_domain_unmap,
1387	.xlate = xive_irq_domain_xlate,
1388#ifdef CONFIG_GENERIC_IRQ_DEBUGFS
1389	.debug_show = xive_irq_domain_debug_show,
1390#endif
1391};
1392
1393static void __init xive_init_host(struct device_node *np)
1394{
1395	xive_irq_domain = irq_domain_add_nomap(np, XIVE_MAX_IRQ,
1396					       &xive_irq_domain_ops, NULL);
1397	if (WARN_ON(xive_irq_domain == NULL))
1398		return;
1399	irq_set_default_host(xive_irq_domain);
1400}
1401
1402static void xive_cleanup_cpu_queues(unsigned int cpu, struct xive_cpu *xc)
1403{
1404	if (xc->queue[xive_irq_priority].qpage)
1405		xive_ops->cleanup_queue(cpu, xc, xive_irq_priority);
1406}
1407
1408static int xive_setup_cpu_queues(unsigned int cpu, struct xive_cpu *xc)
1409{
1410	int rc = 0;
1411
1412	/* We setup 1 queues for now with a 64k page */
1413	if (!xc->queue[xive_irq_priority].qpage)
1414		rc = xive_ops->setup_queue(cpu, xc, xive_irq_priority);
1415
1416	return rc;
1417}
1418
1419static int xive_prepare_cpu(unsigned int cpu)
1420{
1421	struct xive_cpu *xc;
1422
1423	xc = per_cpu(xive_cpu, cpu);
1424	if (!xc) {
1425		xc = kzalloc_node(sizeof(struct xive_cpu),
1426				  GFP_KERNEL, cpu_to_node(cpu));
1427		if (!xc)
1428			return -ENOMEM;
1429		xc->hw_ipi = XIVE_BAD_IRQ;
1430		xc->chip_id = XIVE_INVALID_CHIP_ID;
1431		if (xive_ops->prepare_cpu)
1432			xive_ops->prepare_cpu(cpu, xc);
1433
1434		per_cpu(xive_cpu, cpu) = xc;
1435	}
1436
1437	/* Setup EQs if not already */
1438	return xive_setup_cpu_queues(cpu, xc);
1439}
1440
1441static void xive_setup_cpu(void)
1442{
1443	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1444
1445	/* The backend might have additional things to do */
1446	if (xive_ops->setup_cpu)
1447		xive_ops->setup_cpu(smp_processor_id(), xc);
1448
1449	/* Set CPPR to 0xff to enable flow of interrupts */
1450	xc->cppr = 0xff;
1451	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0xff);
1452}
1453
1454#ifdef CONFIG_SMP
1455void xive_smp_setup_cpu(void)
1456{
1457	pr_devel("SMP setup CPU %d\n", smp_processor_id());
1458
1459	/* This will have already been done on the boot CPU */
1460	if (smp_processor_id() != boot_cpuid)
1461		xive_setup_cpu();
1462
1463}
1464
1465int xive_smp_prepare_cpu(unsigned int cpu)
1466{
1467	int rc;
1468
1469	/* Allocate per-CPU data and queues */
1470	rc = xive_prepare_cpu(cpu);
1471	if (rc)
1472		return rc;
1473
1474	/* Allocate and setup IPI for the new CPU */
1475	return xive_setup_cpu_ipi(cpu);
1476}
1477
1478#ifdef CONFIG_HOTPLUG_CPU
1479static void xive_flush_cpu_queue(unsigned int cpu, struct xive_cpu *xc)
1480{
1481	u32 irq;
1482
1483	/* We assume local irqs are disabled */
1484	WARN_ON(!irqs_disabled());
1485
1486	/* Check what's already in the CPU queue */
1487	while ((irq = xive_scan_interrupts(xc, false)) != 0) {
1488		/*
1489		 * We need to re-route that interrupt to its new destination.
1490		 * First get and lock the descriptor
1491		 */
1492		struct irq_desc *desc = irq_to_desc(irq);
1493		struct irq_data *d = irq_desc_get_irq_data(desc);
1494		struct xive_irq_data *xd;
1495
1496		/*
1497		 * Ignore anything that isn't a XIVE irq and ignore
1498		 * IPIs, so can just be dropped.
1499		 */
1500		if (d->domain != xive_irq_domain)
1501			continue;
1502
1503		/*
1504		 * The IRQ should have already been re-routed, it's just a
1505		 * stale in the old queue, so re-trigger it in order to make
1506		 * it reach is new destination.
1507		 */
1508#ifdef DEBUG_FLUSH
1509		pr_info("CPU %d: Got irq %d while offline, re-sending...\n",
1510			cpu, irq);
1511#endif
1512		raw_spin_lock(&desc->lock);
1513		xd = irq_desc_get_handler_data(desc);
1514
1515		/*
1516		 * Clear saved_p to indicate that it's no longer pending
1517		 */
1518		xd->saved_p = false;
1519
1520		/*
1521		 * For LSIs, we EOI, this will cause a resend if it's
1522		 * still asserted. Otherwise do an MSI retrigger.
1523		 */
1524		if (xd->flags & XIVE_IRQ_FLAG_LSI)
1525			xive_do_source_eoi(xd);
1526		else
1527			xive_irq_retrigger(d);
1528
1529		raw_spin_unlock(&desc->lock);
1530	}
1531}
1532
1533void xive_smp_disable_cpu(void)
1534{
1535	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1536	unsigned int cpu = smp_processor_id();
1537
1538	/* Migrate interrupts away from the CPU */
1539	irq_migrate_all_off_this_cpu();
1540
1541	/* Set CPPR to 0 to disable flow of interrupts */
1542	xc->cppr = 0;
1543	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0);
1544
1545	/* Flush everything still in the queue */
1546	xive_flush_cpu_queue(cpu, xc);
1547
1548	/* Re-enable CPPR  */
1549	xc->cppr = 0xff;
1550	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0xff);
1551}
1552
1553void xive_flush_interrupt(void)
1554{
1555	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1556	unsigned int cpu = smp_processor_id();
1557
1558	/* Called if an interrupt occurs while the CPU is hot unplugged */
1559	xive_flush_cpu_queue(cpu, xc);
1560}
1561
1562#endif /* CONFIG_HOTPLUG_CPU */
1563
1564#endif /* CONFIG_SMP */
1565
1566void xive_teardown_cpu(void)
1567{
1568	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1569	unsigned int cpu = smp_processor_id();
1570
1571	/* Set CPPR to 0 to disable flow of interrupts */
1572	xc->cppr = 0;
1573	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0);
1574
1575	if (xive_ops->teardown_cpu)
1576		xive_ops->teardown_cpu(cpu, xc);
1577
1578#ifdef CONFIG_SMP
1579	/* Get rid of IPI */
1580	xive_cleanup_cpu_ipi(cpu, xc);
1581#endif
1582
1583	/* Disable and free the queues */
1584	xive_cleanup_cpu_queues(cpu, xc);
1585}
1586
1587void xive_shutdown(void)
1588{
1589	xive_ops->shutdown();
1590}
1591
1592bool __init xive_core_init(struct device_node *np, const struct xive_ops *ops,
1593			   void __iomem *area, u32 offset, u8 max_prio)
1594{
1595	xive_tima = area;
1596	xive_tima_offset = offset;
1597	xive_ops = ops;
1598	xive_irq_priority = max_prio;
1599
1600	ppc_md.get_irq = xive_get_irq;
1601	__xive_enabled = true;
1602
1603	pr_devel("Initializing host..\n");
1604	xive_init_host(np);
1605
1606	pr_devel("Initializing boot CPU..\n");
1607
1608	/* Allocate per-CPU data and queues */
1609	xive_prepare_cpu(smp_processor_id());
1610
1611	/* Get ready for interrupts */
1612	xive_setup_cpu();
1613
1614	pr_info("Interrupt handling initialized with %s backend\n",
1615		xive_ops->name);
1616	pr_info("Using priority %d for all interrupts\n", max_prio);
1617
1618	return true;
1619}
1620
1621__be32 *xive_queue_page_alloc(unsigned int cpu, u32 queue_shift)
1622{
1623	unsigned int alloc_order;
1624	struct page *pages;
1625	__be32 *qpage;
1626
1627	alloc_order = xive_alloc_order(queue_shift);
1628	pages = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, alloc_order);
1629	if (!pages)
1630		return ERR_PTR(-ENOMEM);
1631	qpage = (__be32 *)page_address(pages);
1632	memset(qpage, 0, 1 << queue_shift);
1633
1634	return qpage;
1635}
1636
1637static int __init xive_off(char *arg)
1638{
1639	xive_cmdline_disabled = true;
1640	return 0;
1641}
1642__setup("xive=off", xive_off);
1643
1644static void xive_debug_show_cpu(struct seq_file *m, int cpu)
1645{
1646	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
1647
1648	seq_printf(m, "CPU %d:", cpu);
1649	if (xc) {
1650		seq_printf(m, "pp=%02x CPPR=%02x ", xc->pending_prio, xc->cppr);
1651
1652#ifdef CONFIG_SMP
1653		{
1654			u64 val = xive_esb_read(&xc->ipi_data, XIVE_ESB_GET);
1655
1656			seq_printf(m, "IPI=0x%08x PQ=%c%c ", xc->hw_ipi,
1657				   val & XIVE_ESB_VAL_P ? 'P' : '-',
1658				   val & XIVE_ESB_VAL_Q ? 'Q' : '-');
1659		}
1660#endif
1661		{
1662			struct xive_q *q = &xc->queue[xive_irq_priority];
1663			u32 i0, i1, idx;
1664
1665			if (q->qpage) {
1666				idx = q->idx;
1667				i0 = be32_to_cpup(q->qpage + idx);
1668				idx = (idx + 1) & q->msk;
1669				i1 = be32_to_cpup(q->qpage + idx);
1670				seq_printf(m, "EQ idx=%d T=%d %08x %08x ...",
1671					   q->idx, q->toggle, i0, i1);
1672			}
1673		}
1674	}
1675	seq_puts(m, "\n");
1676}
1677
1678static void xive_debug_show_irq(struct seq_file *m, struct irq_data *d)
1679{
1680	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
1681	int rc;
1682	u32 target;
1683	u8 prio;
1684	u32 lirq;
1685	struct xive_irq_data *xd;
1686	u64 val;
1687
1688	rc = xive_ops->get_irq_config(hw_irq, &target, &prio, &lirq);
1689	if (rc) {
1690		seq_printf(m, "IRQ 0x%08x : no config rc=%d\n", hw_irq, rc);
1691		return;
1692	}
1693
1694	seq_printf(m, "IRQ 0x%08x : target=0x%x prio=%02x lirq=0x%x ",
1695		   hw_irq, target, prio, lirq);
1696
1697	xd = irq_data_get_irq_handler_data(d);
1698	val = xive_esb_read(xd, XIVE_ESB_GET);
1699	seq_printf(m, "flags=%c%c%c PQ=%c%c",
1700		   xd->flags & XIVE_IRQ_FLAG_STORE_EOI ? 'S' : ' ',
1701		   xd->flags & XIVE_IRQ_FLAG_LSI ? 'L' : ' ',
1702		   xd->flags & XIVE_IRQ_FLAG_H_INT_ESB ? 'H' : ' ',
1703		   val & XIVE_ESB_VAL_P ? 'P' : '-',
1704		   val & XIVE_ESB_VAL_Q ? 'Q' : '-');
1705	seq_puts(m, "\n");
1706}
1707
1708static int xive_core_debug_show(struct seq_file *m, void *private)
1709{
1710	unsigned int i;
1711	struct irq_desc *desc;
1712	int cpu;
1713
1714	if (xive_ops->debug_show)
1715		xive_ops->debug_show(m, private);
1716
1717	for_each_possible_cpu(cpu)
1718		xive_debug_show_cpu(m, cpu);
1719
1720	for_each_irq_desc(i, desc) {
1721		struct irq_data *d = irq_desc_get_irq_data(desc);
1722
1723		if (d->domain == xive_irq_domain)
1724			xive_debug_show_irq(m, d);
1725	}
1726	return 0;
1727}
1728DEFINE_SHOW_ATTRIBUTE(xive_core_debug);
1729
1730int xive_core_debug_init(void)
1731{
1732	if (xive_enabled())
1733		debugfs_create_file("xive", 0400, powerpc_debugfs_root,
1734				    NULL, &xive_core_debug_fops);
1735	return 0;
1736}