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 all use the same logical irq number */
  67static u32 xive_ipi_irq;
  68#endif
  69
  70/* Xive state for each CPU */
  71static DEFINE_PER_CPU(struct xive_cpu *, xive_cpu);
  72
  73/* An invalid CPU target */
  74#define XIVE_INVALID_TARGET	(-1)
  75
  76/*
  77 * Read the next entry in a queue, return its content if it's valid
  78 * or 0 if there is no new entry.
  79 *
  80 * The queue pointer is moved forward unless "just_peek" is set
  81 */
  82static u32 xive_read_eq(struct xive_q *q, bool just_peek)
  83{
  84	u32 cur;
  85
  86	if (!q->qpage)
  87		return 0;
  88	cur = be32_to_cpup(q->qpage + q->idx);
  89
  90	/* Check valid bit (31) vs current toggle polarity */
  91	if ((cur >> 31) == q->toggle)
  92		return 0;
  93
  94	/* If consuming from the queue ... */
  95	if (!just_peek) {
  96		/* Next entry */
  97		q->idx = (q->idx + 1) & q->msk;
  98
  99		/* Wrap around: flip valid toggle */
 100		if (q->idx == 0)
 101			q->toggle ^= 1;
 102	}
 103	/* Mask out the valid bit (31) */
 104	return cur & 0x7fffffff;
 105}
 106
 107/*
 108 * Scans all the queue that may have interrupts in them
 109 * (based on "pending_prio") in priority order until an
 110 * interrupt is found or all the queues are empty.
 111 *
 112 * Then updates the CPPR (Current Processor Priority
 113 * Register) based on the most favored interrupt found
 114 * (0xff if none) and return what was found (0 if none).
 115 *
 116 * If just_peek is set, return the most favored pending
 117 * interrupt if any but don't update the queue pointers.
 118 *
 119 * Note: This function can operate generically on any number
 120 * of queues (up to 8). The current implementation of the XIVE
 121 * driver only uses a single queue however.
 122 *
 123 * Note2: This will also "flush" "the pending_count" of a queue
 124 * into the "count" when that queue is observed to be empty.
 125 * This is used to keep track of the amount of interrupts
 126 * targetting a queue. When an interrupt is moved away from
 127 * a queue, we only decrement that queue count once the queue
 128 * has been observed empty to avoid races.
 129 */
 130static u32 xive_scan_interrupts(struct xive_cpu *xc, bool just_peek)
 131{
 132	u32 irq = 0;
 133	u8 prio = 0;
 134
 135	/* Find highest pending priority */
 136	while (xc->pending_prio != 0) {
 137		struct xive_q *q;
 138
 139		prio = ffs(xc->pending_prio) - 1;
 140		DBG_VERBOSE("scan_irq: trying prio %d\n", prio);
 141
 142		/* Try to fetch */
 143		irq = xive_read_eq(&xc->queue[prio], just_peek);
 144
 145		/* Found something ? That's it */
 146		if (irq) {
 147			if (just_peek || irq_to_desc(irq))
 148				break;
 149			/*
 150			 * We should never get here; if we do then we must
 151			 * have failed to synchronize the interrupt properly
 152			 * when shutting it down.
 153			 */
 154			pr_crit("xive: got interrupt %d without descriptor, dropping\n",
 155				irq);
 156			WARN_ON(1);
 157			continue;
 158		}
 159
 160		/* Clear pending bits */
 161		xc->pending_prio &= ~(1 << prio);
 162
 163		/*
 164		 * Check if the queue count needs adjusting due to
 165		 * interrupts being moved away. See description of
 166		 * xive_dec_target_count()
 167		 */
 168		q = &xc->queue[prio];
 169		if (atomic_read(&q->pending_count)) {
 170			int p = atomic_xchg(&q->pending_count, 0);
 171			if (p) {
 172				WARN_ON(p > atomic_read(&q->count));
 173				atomic_sub(p, &q->count);
 174			}
 175		}
 176	}
 177
 178	/* If nothing was found, set CPPR to 0xff */
 179	if (irq == 0)
 180		prio = 0xff;
 181
 182	/* Update HW CPPR to match if necessary */
 183	if (prio != xc->cppr) {
 184		DBG_VERBOSE("scan_irq: adjusting CPPR to %d\n", prio);
 185		xc->cppr = prio;
 186		out_8(xive_tima + xive_tima_offset + TM_CPPR, prio);
 187	}
 188
 189	return irq;
 190}
 191
 192/*
 193 * This is used to perform the magic loads from an ESB
 194 * described in xive-regs.h
 195 */
 196static notrace u8 xive_esb_read(struct xive_irq_data *xd, u32 offset)
 197{
 198	u64 val;
 199
 200	if (offset == XIVE_ESB_SET_PQ_10 && xd->flags & XIVE_IRQ_FLAG_STORE_EOI)
 201		offset |= XIVE_ESB_LD_ST_MO;
 202
 203	/* Handle HW errata */
 204	if (xd->flags & XIVE_IRQ_FLAG_SHIFT_BUG)
 205		offset |= offset << 4;
 206
 207	if ((xd->flags & XIVE_IRQ_FLAG_H_INT_ESB) && xive_ops->esb_rw)
 208		val = xive_ops->esb_rw(xd->hw_irq, offset, 0, 0);
 209	else
 210		val = in_be64(xd->eoi_mmio + offset);
 211
 212	return (u8)val;
 213}
 214
 215static void xive_esb_write(struct xive_irq_data *xd, u32 offset, u64 data)
 216{
 217	/* Handle HW errata */
 218	if (xd->flags & XIVE_IRQ_FLAG_SHIFT_BUG)
 219		offset |= offset << 4;
 220
 221	if ((xd->flags & XIVE_IRQ_FLAG_H_INT_ESB) && xive_ops->esb_rw)
 222		xive_ops->esb_rw(xd->hw_irq, offset, data, 1);
 223	else
 224		out_be64(xd->eoi_mmio + offset, data);
 225}
 226
 227#ifdef CONFIG_XMON
 228static notrace void xive_dump_eq(const char *name, struct xive_q *q)
 229{
 230	u32 i0, i1, idx;
 231
 232	if (!q->qpage)
 233		return;
 234	idx = q->idx;
 235	i0 = be32_to_cpup(q->qpage + idx);
 236	idx = (idx + 1) & q->msk;
 237	i1 = be32_to_cpup(q->qpage + idx);
 238	xmon_printf("%s idx=%d T=%d %08x %08x ...", name,
 239		     q->idx, q->toggle, i0, i1);
 240}
 241
 242notrace void xmon_xive_do_dump(int cpu)
 243{
 244	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
 245
 246	xmon_printf("CPU %d:", cpu);
 247	if (xc) {
 248		xmon_printf("pp=%02x CPPR=%02x ", xc->pending_prio, xc->cppr);
 249
 250#ifdef CONFIG_SMP
 251		{
 252			u64 val = xive_esb_read(&xc->ipi_data, XIVE_ESB_GET);
 253
 254			xmon_printf("IPI=0x%08x PQ=%c%c ", xc->hw_ipi,
 255				    val & XIVE_ESB_VAL_P ? 'P' : '-',
 256				    val & XIVE_ESB_VAL_Q ? 'Q' : '-');
 257		}
 258#endif
 259		xive_dump_eq("EQ", &xc->queue[xive_irq_priority]);
 260	}
 261	xmon_printf("\n");
 262}
 263
 264int xmon_xive_get_irq_config(u32 hw_irq, struct irq_data *d)
 265{
 266	struct irq_chip *chip = irq_data_get_irq_chip(d);
 267	int rc;
 268	u32 target;
 269	u8 prio;
 270	u32 lirq;
 271
 272	if (!is_xive_irq(chip))
 273		return -EINVAL;
 274
 275	rc = xive_ops->get_irq_config(hw_irq, &target, &prio, &lirq);
 276	if (rc) {
 277		xmon_printf("IRQ 0x%08x : no config rc=%d\n", hw_irq, rc);
 278		return rc;
 279	}
 280
 281	xmon_printf("IRQ 0x%08x : target=0x%x prio=%02x lirq=0x%x ",
 282		    hw_irq, target, prio, lirq);
 283
 284	if (d) {
 285		struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 286		u64 val = xive_esb_read(xd, XIVE_ESB_GET);
 287
 288		xmon_printf("flags=%c%c%c PQ=%c%c",
 289			    xd->flags & XIVE_IRQ_FLAG_STORE_EOI ? 'S' : ' ',
 290			    xd->flags & XIVE_IRQ_FLAG_LSI ? 'L' : ' ',
 291			    xd->flags & XIVE_IRQ_FLAG_H_INT_ESB ? 'H' : ' ',
 292			    val & XIVE_ESB_VAL_P ? 'P' : '-',
 293			    val & XIVE_ESB_VAL_Q ? 'Q' : '-');
 294	}
 295
 296	xmon_printf("\n");
 297	return 0;
 298}
 299
 300#endif /* CONFIG_XMON */
 301
 302static unsigned int xive_get_irq(void)
 303{
 304	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
 305	u32 irq;
 306
 307	/*
 308	 * This can be called either as a result of a HW interrupt or
 309	 * as a "replay" because EOI decided there was still something
 310	 * in one of the queues.
 311	 *
 312	 * First we perform an ACK cycle in order to update our mask
 313	 * of pending priorities. This will also have the effect of
 314	 * updating the CPPR to the most favored pending interrupts.
 315	 *
 316	 * In the future, if we have a way to differentiate a first
 317	 * entry (on HW interrupt) from a replay triggered by EOI,
 318	 * we could skip this on replays unless we soft-mask tells us
 319	 * that a new HW interrupt occurred.
 320	 */
 321	xive_ops->update_pending(xc);
 322
 323	DBG_VERBOSE("get_irq: pending=%02x\n", xc->pending_prio);
 324
 325	/* Scan our queue(s) for interrupts */
 326	irq = xive_scan_interrupts(xc, false);
 327
 328	DBG_VERBOSE("get_irq: got irq 0x%x, new pending=0x%02x\n",
 329	    irq, xc->pending_prio);
 330
 331	/* Return pending interrupt if any */
 332	if (irq == XIVE_BAD_IRQ)
 333		return 0;
 334	return irq;
 335}
 336
 337/*
 338 * After EOI'ing an interrupt, we need to re-check the queue
 339 * to see if another interrupt is pending since multiple
 340 * interrupts can coalesce into a single notification to the
 341 * CPU.
 342 *
 343 * If we find that there is indeed more in there, we call
 344 * force_external_irq_replay() to make Linux synthetize an
 345 * external interrupt on the next call to local_irq_restore().
 346 */
 347static void xive_do_queue_eoi(struct xive_cpu *xc)
 348{
 349	if (xive_scan_interrupts(xc, true) != 0) {
 350		DBG_VERBOSE("eoi: pending=0x%02x\n", xc->pending_prio);
 351		force_external_irq_replay();
 352	}
 353}
 354
 355/*
 356 * EOI an interrupt at the source. There are several methods
 357 * to do this depending on the HW version and source type
 358 */
 359static void xive_do_source_eoi(u32 hw_irq, struct xive_irq_data *xd)
 360{
 361	xd->stale_p = false;
 362	/* If the XIVE supports the new "store EOI facility, use it */
 363	if (xd->flags & XIVE_IRQ_FLAG_STORE_EOI)
 364		xive_esb_write(xd, XIVE_ESB_STORE_EOI, 0);
 365	else if (hw_irq && xd->flags & XIVE_IRQ_FLAG_EOI_FW) {
 366		/*
 367		 * The FW told us to call it. This happens for some
 368		 * interrupt sources that need additional HW whacking
 369		 * beyond the ESB manipulation. For example LPC interrupts
 370		 * on P9 DD1.0 needed a latch to be clared in the LPC bridge
 371		 * itself. The Firmware will take care of it.
 372		 */
 373		if (WARN_ON_ONCE(!xive_ops->eoi))
 374			return;
 375		xive_ops->eoi(hw_irq);
 376	} else {
 377		u8 eoi_val;
 378
 379		/*
 380		 * Otherwise for EOI, we use the special MMIO that does
 381		 * a clear of both P and Q and returns the old Q,
 382		 * except for LSIs where we use the "EOI cycle" special
 383		 * load.
 384		 *
 385		 * This allows us to then do a re-trigger if Q was set
 386		 * rather than synthesizing an interrupt in software
 387		 *
 388		 * For LSIs the HW EOI cycle is used rather than PQ bits,
 389		 * as they are automatically re-triggred in HW when still
 390		 * pending.
 391		 */
 392		if (xd->flags & XIVE_IRQ_FLAG_LSI)
 393			xive_esb_read(xd, XIVE_ESB_LOAD_EOI);
 394		else {
 395			eoi_val = xive_esb_read(xd, XIVE_ESB_SET_PQ_00);
 396			DBG_VERBOSE("eoi_val=%x\n", eoi_val);
 397
 398			/* Re-trigger if needed */
 399			if ((eoi_val & XIVE_ESB_VAL_Q) && xd->trig_mmio)
 400				out_be64(xd->trig_mmio, 0);
 401		}
 402	}
 403}
 404
 405/* irq_chip eoi callback, called with irq descriptor lock held */
 406static void xive_irq_eoi(struct irq_data *d)
 407{
 408	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 409	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
 410
 411	DBG_VERBOSE("eoi_irq: irq=%d [0x%lx] pending=%02x\n",
 412		    d->irq, irqd_to_hwirq(d), xc->pending_prio);
 413
 414	/*
 415	 * EOI the source if it hasn't been disabled and hasn't
 416	 * been passed-through to a KVM guest
 417	 */
 418	if (!irqd_irq_disabled(d) && !irqd_is_forwarded_to_vcpu(d) &&
 419	    !(xd->flags & XIVE_IRQ_NO_EOI))
 420		xive_do_source_eoi(irqd_to_hwirq(d), xd);
 421	else
 422		xd->stale_p = true;
 423
 424	/*
 425	 * Clear saved_p to indicate that it's no longer occupying
 426	 * a queue slot on the target queue
 427	 */
 428	xd->saved_p = false;
 429
 430	/* Check for more work in the queue */
 431	xive_do_queue_eoi(xc);
 432}
 433
 434/*
 435 * Helper used to mask and unmask an interrupt source. This
 436 * is only called for normal interrupts that do not require
 437 * masking/unmasking via firmware.
 438 */
 439static void xive_do_source_set_mask(struct xive_irq_data *xd,
 440				    bool mask)
 441{
 442	u64 val;
 443
 444	/*
 445	 * If the interrupt had P set, it may be in a queue.
 446	 *
 447	 * We need to make sure we don't re-enable it until it
 448	 * has been fetched from that queue and EOId. We keep
 449	 * a copy of that P state and use it to restore the
 450	 * ESB accordingly on unmask.
 451	 */
 452	if (mask) {
 453		val = xive_esb_read(xd, XIVE_ESB_SET_PQ_01);
 454		if (!xd->stale_p && !!(val & XIVE_ESB_VAL_P))
 455			xd->saved_p = true;
 456		xd->stale_p = false;
 457	} else if (xd->saved_p) {
 458		xive_esb_read(xd, XIVE_ESB_SET_PQ_10);
 459		xd->saved_p = false;
 460	} else {
 461		xive_esb_read(xd, XIVE_ESB_SET_PQ_00);
 462		xd->stale_p = false;
 463	}
 464}
 465
 466/*
 467 * Try to chose "cpu" as a new interrupt target. Increments
 468 * the queue accounting for that target if it's not already
 469 * full.
 470 */
 471static bool xive_try_pick_target(int cpu)
 472{
 473	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
 474	struct xive_q *q = &xc->queue[xive_irq_priority];
 475	int max;
 476
 477	/*
 478	 * Calculate max number of interrupts in that queue.
 479	 *
 480	 * We leave a gap of 1 just in case...
 481	 */
 482	max = (q->msk + 1) - 1;
 483	return !!atomic_add_unless(&q->count, 1, max);
 484}
 485
 486/*
 487 * Un-account an interrupt for a target CPU. We don't directly
 488 * decrement q->count since the interrupt might still be present
 489 * in the queue.
 490 *
 491 * Instead increment a separate counter "pending_count" which
 492 * will be substracted from "count" later when that CPU observes
 493 * the queue to be empty.
 494 */
 495static void xive_dec_target_count(int cpu)
 496{
 497	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
 498	struct xive_q *q = &xc->queue[xive_irq_priority];
 499
 500	if (WARN_ON(cpu < 0 || !xc)) {
 501		pr_err("%s: cpu=%d xc=%p\n", __func__, cpu, xc);
 502		return;
 503	}
 504
 505	/*
 506	 * We increment the "pending count" which will be used
 507	 * to decrement the target queue count whenever it's next
 508	 * processed and found empty. This ensure that we don't
 509	 * decrement while we still have the interrupt there
 510	 * occupying a slot.
 511	 */
 512	atomic_inc(&q->pending_count);
 513}
 514
 515/* Find a tentative CPU target in a CPU mask */
 516static int xive_find_target_in_mask(const struct cpumask *mask,
 517				    unsigned int fuzz)
 518{
 519	int cpu, first, num, i;
 520
 521	/* Pick up a starting point CPU in the mask based on  fuzz */
 522	num = min_t(int, cpumask_weight(mask), nr_cpu_ids);
 523	first = fuzz % num;
 524
 525	/* Locate it */
 526	cpu = cpumask_first(mask);
 527	for (i = 0; i < first && cpu < nr_cpu_ids; i++)
 528		cpu = cpumask_next(cpu, mask);
 529
 530	/* Sanity check */
 531	if (WARN_ON(cpu >= nr_cpu_ids))
 532		cpu = cpumask_first(cpu_online_mask);
 533
 534	/* Remember first one to handle wrap-around */
 535	first = cpu;
 536
 537	/*
 538	 * Now go through the entire mask until we find a valid
 539	 * target.
 540	 */
 541	do {
 542		/*
 543		 * We re-check online as the fallback case passes us
 544		 * an untested affinity mask
 545		 */
 546		if (cpu_online(cpu) && xive_try_pick_target(cpu))
 547			return cpu;
 548		cpu = cpumask_next(cpu, mask);
 549		/* Wrap around */
 550		if (cpu >= nr_cpu_ids)
 551			cpu = cpumask_first(mask);
 552	} while (cpu != first);
 553
 554	return -1;
 555}
 556
 557/*
 558 * Pick a target CPU for an interrupt. This is done at
 559 * startup or if the affinity is changed in a way that
 560 * invalidates the current target.
 561 */
 562static int xive_pick_irq_target(struct irq_data *d,
 563				const struct cpumask *affinity)
 564{
 565	static unsigned int fuzz;
 566	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 567	cpumask_var_t mask;
 568	int cpu = -1;
 569
 570	/*
 571	 * If we have chip IDs, first we try to build a mask of
 572	 * CPUs matching the CPU and find a target in there
 573	 */
 574	if (xd->src_chip != XIVE_INVALID_CHIP_ID &&
 575		zalloc_cpumask_var(&mask, GFP_ATOMIC)) {
 576		/* Build a mask of matching chip IDs */
 577		for_each_cpu_and(cpu, affinity, cpu_online_mask) {
 578			struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
 579			if (xc->chip_id == xd->src_chip)
 580				cpumask_set_cpu(cpu, mask);
 581		}
 582		/* Try to find a target */
 583		if (cpumask_empty(mask))
 584			cpu = -1;
 585		else
 586			cpu = xive_find_target_in_mask(mask, fuzz++);
 587		free_cpumask_var(mask);
 588		if (cpu >= 0)
 589			return cpu;
 590		fuzz--;
 591	}
 592
 593	/* No chip IDs, fallback to using the affinity mask */
 594	return xive_find_target_in_mask(affinity, fuzz++);
 595}
 596
 597static unsigned int xive_irq_startup(struct irq_data *d)
 598{
 599	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 600	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 601	int target, rc;
 602
 603	xd->saved_p = false;
 604	xd->stale_p = false;
 605	pr_devel("xive_irq_startup: irq %d [0x%x] data @%p\n",
 606		 d->irq, hw_irq, d);
 607
 608#ifdef CONFIG_PCI_MSI
 609	/*
 610	 * The generic MSI code returns with the interrupt disabled on the
 611	 * card, using the MSI mask bits. Firmware doesn't appear to unmask
 612	 * at that level, so we do it here by hand.
 613	 */
 614	if (irq_data_get_msi_desc(d))
 615		pci_msi_unmask_irq(d);
 616#endif
 617
 618	/* Pick a target */
 619	target = xive_pick_irq_target(d, irq_data_get_affinity_mask(d));
 620	if (target == XIVE_INVALID_TARGET) {
 621		/* Try again breaking affinity */
 622		target = xive_pick_irq_target(d, cpu_online_mask);
 623		if (target == XIVE_INVALID_TARGET)
 624			return -ENXIO;
 625		pr_warn("irq %d started with broken affinity\n", d->irq);
 626	}
 627
 628	/* Sanity check */
 629	if (WARN_ON(target == XIVE_INVALID_TARGET ||
 630		    target >= nr_cpu_ids))
 631		target = smp_processor_id();
 632
 633	xd->target = target;
 634
 635	/*
 636	 * Configure the logical number to be the Linux IRQ number
 637	 * and set the target queue
 638	 */
 639	rc = xive_ops->configure_irq(hw_irq,
 640				     get_hard_smp_processor_id(target),
 641				     xive_irq_priority, d->irq);
 642	if (rc)
 643		return rc;
 644
 645	/* Unmask the ESB */
 646	xive_do_source_set_mask(xd, false);
 647
 648	return 0;
 649}
 650
 651/* called with irq descriptor lock held */
 652static void xive_irq_shutdown(struct irq_data *d)
 653{
 654	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 655	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 656
 657	pr_devel("xive_irq_shutdown: irq %d [0x%x] data @%p\n",
 658		 d->irq, hw_irq, d);
 659
 660	if (WARN_ON(xd->target == XIVE_INVALID_TARGET))
 661		return;
 662
 663	/* Mask the interrupt at the source */
 664	xive_do_source_set_mask(xd, true);
 665
 666	/*
 667	 * Mask the interrupt in HW in the IVT/EAS and set the number
 668	 * to be the "bad" IRQ number
 669	 */
 670	xive_ops->configure_irq(hw_irq,
 671				get_hard_smp_processor_id(xd->target),
 672				0xff, XIVE_BAD_IRQ);
 673
 674	xive_dec_target_count(xd->target);
 675	xd->target = XIVE_INVALID_TARGET;
 676}
 677
 678static void xive_irq_unmask(struct irq_data *d)
 679{
 680	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 681
 682	pr_devel("xive_irq_unmask: irq %d data @%p\n", d->irq, xd);
 683
 684	/*
 685	 * This is a workaround for PCI LSI problems on P9, for
 686	 * these, we call FW to set the mask. The problems might
 687	 * be fixed by P9 DD2.0, if that is the case, firmware
 688	 * will no longer set that flag.
 689	 */
 690	if (xd->flags & XIVE_IRQ_FLAG_MASK_FW) {
 691		unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 692		xive_ops->configure_irq(hw_irq,
 693					get_hard_smp_processor_id(xd->target),
 694					xive_irq_priority, d->irq);
 695		return;
 696	}
 697
 698	xive_do_source_set_mask(xd, false);
 699}
 700
 701static void xive_irq_mask(struct irq_data *d)
 702{
 703	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 704
 705	pr_devel("xive_irq_mask: irq %d data @%p\n", d->irq, xd);
 706
 707	/*
 708	 * This is a workaround for PCI LSI problems on P9, for
 709	 * these, we call OPAL to set the mask. The problems might
 710	 * be fixed by P9 DD2.0, if that is the case, firmware
 711	 * will no longer set that flag.
 712	 */
 713	if (xd->flags & XIVE_IRQ_FLAG_MASK_FW) {
 714		unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 715		xive_ops->configure_irq(hw_irq,
 716					get_hard_smp_processor_id(xd->target),
 717					0xff, d->irq);
 718		return;
 719	}
 720
 721	xive_do_source_set_mask(xd, true);
 722}
 723
 724static int xive_irq_set_affinity(struct irq_data *d,
 725				 const struct cpumask *cpumask,
 726				 bool force)
 727{
 728	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 729	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 730	u32 target, old_target;
 731	int rc = 0;
 732
 733	pr_devel("xive_irq_set_affinity: irq %d\n", d->irq);
 734
 735	/* Is this valid ? */
 736	if (cpumask_any_and(cpumask, cpu_online_mask) >= nr_cpu_ids)
 737		return -EINVAL;
 738
 739	/* Don't do anything if the interrupt isn't started */
 740	if (!irqd_is_started(d))
 741		return IRQ_SET_MASK_OK;
 742
 743	/*
 744	 * If existing target is already in the new mask, and is
 745	 * online then do nothing.
 746	 */
 747	if (xd->target != XIVE_INVALID_TARGET &&
 748	    cpu_online(xd->target) &&
 749	    cpumask_test_cpu(xd->target, cpumask))
 750		return IRQ_SET_MASK_OK;
 751
 752	/* Pick a new target */
 753	target = xive_pick_irq_target(d, cpumask);
 754
 755	/* No target found */
 756	if (target == XIVE_INVALID_TARGET)
 757		return -ENXIO;
 758
 759	/* Sanity check */
 760	if (WARN_ON(target >= nr_cpu_ids))
 761		target = smp_processor_id();
 762
 763	old_target = xd->target;
 764
 765	/*
 766	 * Only configure the irq if it's not currently passed-through to
 767	 * a KVM guest
 768	 */
 769	if (!irqd_is_forwarded_to_vcpu(d))
 770		rc = xive_ops->configure_irq(hw_irq,
 771					     get_hard_smp_processor_id(target),
 772					     xive_irq_priority, d->irq);
 773	if (rc < 0) {
 774		pr_err("Error %d reconfiguring irq %d\n", rc, d->irq);
 775		return rc;
 776	}
 777
 778	pr_devel("  target: 0x%x\n", target);
 779	xd->target = target;
 780
 781	/* Give up previous target */
 782	if (old_target != XIVE_INVALID_TARGET)
 783	    xive_dec_target_count(old_target);
 784
 785	return IRQ_SET_MASK_OK;
 786}
 787
 788static int xive_irq_set_type(struct irq_data *d, unsigned int flow_type)
 789{
 790	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 791
 792	/*
 793	 * We only support these. This has really no effect other than setting
 794	 * the corresponding descriptor bits mind you but those will in turn
 795	 * affect the resend function when re-enabling an edge interrupt.
 796	 *
 797	 * Set set the default to edge as explained in map().
 798	 */
 799	if (flow_type == IRQ_TYPE_DEFAULT || flow_type == IRQ_TYPE_NONE)
 800		flow_type = IRQ_TYPE_EDGE_RISING;
 801
 802	if (flow_type != IRQ_TYPE_EDGE_RISING &&
 803	    flow_type != IRQ_TYPE_LEVEL_LOW)
 804		return -EINVAL;
 805
 806	irqd_set_trigger_type(d, flow_type);
 807
 808	/*
 809	 * Double check it matches what the FW thinks
 810	 *
 811	 * NOTE: We don't know yet if the PAPR interface will provide
 812	 * the LSI vs MSI information apart from the device-tree so
 813	 * this check might have to move into an optional backend call
 814	 * that is specific to the native backend
 815	 */
 816	if ((flow_type == IRQ_TYPE_LEVEL_LOW) !=
 817	    !!(xd->flags & XIVE_IRQ_FLAG_LSI)) {
 818		pr_warn("Interrupt %d (HW 0x%x) type mismatch, Linux says %s, FW says %s\n",
 819			d->irq, (u32)irqd_to_hwirq(d),
 820			(flow_type == IRQ_TYPE_LEVEL_LOW) ? "Level" : "Edge",
 821			(xd->flags & XIVE_IRQ_FLAG_LSI) ? "Level" : "Edge");
 822	}
 823
 824	return IRQ_SET_MASK_OK_NOCOPY;
 825}
 826
 827static int xive_irq_retrigger(struct irq_data *d)
 828{
 829	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 830
 831	/* This should be only for MSIs */
 832	if (WARN_ON(xd->flags & XIVE_IRQ_FLAG_LSI))
 833		return 0;
 834
 835	/*
 836	 * To perform a retrigger, we first set the PQ bits to
 837	 * 11, then perform an EOI.
 838	 */
 839	xive_esb_read(xd, XIVE_ESB_SET_PQ_11);
 840
 841	/*
 842	 * Note: We pass "0" to the hw_irq argument in order to
 843	 * avoid calling into the backend EOI code which we don't
 844	 * want to do in the case of a re-trigger. Backends typically
 845	 * only do EOI for LSIs anyway.
 846	 */
 847	xive_do_source_eoi(0, xd);
 848
 849	return 1;
 850}
 851
 852/*
 853 * Caller holds the irq descriptor lock, so this won't be called
 854 * concurrently with xive_get_irqchip_state on the same interrupt.
 855 */
 856static int xive_irq_set_vcpu_affinity(struct irq_data *d, void *state)
 857{
 858	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 859	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
 860	int rc;
 861	u8 pq;
 862
 863	/*
 864	 * We only support this on interrupts that do not require
 865	 * firmware calls for masking and unmasking
 866	 */
 867	if (xd->flags & XIVE_IRQ_FLAG_MASK_FW)
 868		return -EIO;
 869
 870	/*
 871	 * This is called by KVM with state non-NULL for enabling
 872	 * pass-through or NULL for disabling it
 873	 */
 874	if (state) {
 875		irqd_set_forwarded_to_vcpu(d);
 876
 877		/* Set it to PQ=10 state to prevent further sends */
 878		pq = xive_esb_read(xd, XIVE_ESB_SET_PQ_10);
 879		if (!xd->stale_p) {
 880			xd->saved_p = !!(pq & XIVE_ESB_VAL_P);
 881			xd->stale_p = !xd->saved_p;
 882		}
 883
 884		/* No target ? nothing to do */
 885		if (xd->target == XIVE_INVALID_TARGET) {
 886			/*
 887			 * An untargetted interrupt should have been
 888			 * also masked at the source
 889			 */
 890			WARN_ON(xd->saved_p);
 891
 892			return 0;
 893		}
 894
 895		/*
 896		 * If P was set, adjust state to PQ=11 to indicate
 897		 * that a resend is needed for the interrupt to reach
 898		 * the guest. Also remember the value of P.
 899		 *
 900		 * This also tells us that it's in flight to a host queue
 901		 * or has already been fetched but hasn't been EOIed yet
 902		 * by the host. This it's potentially using up a host
 903		 * queue slot. This is important to know because as long
 904		 * as this is the case, we must not hard-unmask it when
 905		 * "returning" that interrupt to the host.
 906		 *
 907		 * This saved_p is cleared by the host EOI, when we know
 908		 * for sure the queue slot is no longer in use.
 909		 */
 910		if (xd->saved_p) {
 911			xive_esb_read(xd, XIVE_ESB_SET_PQ_11);
 912
 913			/*
 914			 * Sync the XIVE source HW to ensure the interrupt
 915			 * has gone through the EAS before we change its
 916			 * target to the guest. That should guarantee us
 917			 * that we *will* eventually get an EOI for it on
 918			 * the host. Otherwise there would be a small window
 919			 * for P to be seen here but the interrupt going
 920			 * to the guest queue.
 921			 */
 922			if (xive_ops->sync_source)
 923				xive_ops->sync_source(hw_irq);
 924		}
 925	} else {
 926		irqd_clr_forwarded_to_vcpu(d);
 927
 928		/* No host target ? hard mask and return */
 929		if (xd->target == XIVE_INVALID_TARGET) {
 930			xive_do_source_set_mask(xd, true);
 931			return 0;
 932		}
 933
 934		/*
 935		 * Sync the XIVE source HW to ensure the interrupt
 936		 * has gone through the EAS before we change its
 937		 * target to the host.
 938		 */
 939		if (xive_ops->sync_source)
 940			xive_ops->sync_source(hw_irq);
 941
 942		/*
 943		 * By convention we are called with the interrupt in
 944		 * a PQ=10 or PQ=11 state, ie, it won't fire and will
 945		 * have latched in Q whether there's a pending HW
 946		 * interrupt or not.
 947		 *
 948		 * First reconfigure the target.
 949		 */
 950		rc = xive_ops->configure_irq(hw_irq,
 951					     get_hard_smp_processor_id(xd->target),
 952					     xive_irq_priority, d->irq);
 953		if (rc)
 954			return rc;
 955
 956		/*
 957		 * Then if saved_p is not set, effectively re-enable the
 958		 * interrupt with an EOI. If it is set, we know there is
 959		 * still a message in a host queue somewhere that will be
 960		 * EOId eventually.
 961		 *
 962		 * Note: We don't check irqd_irq_disabled(). Effectively,
 963		 * we *will* let the irq get through even if masked if the
 964		 * HW is still firing it in order to deal with the whole
 965		 * saved_p business properly. If the interrupt triggers
 966		 * while masked, the generic code will re-mask it anyway.
 967		 */
 968		if (!xd->saved_p)
 969			xive_do_source_eoi(hw_irq, xd);
 970
 971	}
 972	return 0;
 973}
 974
 975/* Called with irq descriptor lock held. */
 976static int xive_get_irqchip_state(struct irq_data *data,
 977				  enum irqchip_irq_state which, bool *state)
 978{
 979	struct xive_irq_data *xd = irq_data_get_irq_handler_data(data);
 980	u8 pq;
 981
 982	switch (which) {
 983	case IRQCHIP_STATE_ACTIVE:
 984		pq = xive_esb_read(xd, XIVE_ESB_GET);
 985
 986		/*
 987		 * The esb value being all 1's means we couldn't get
 988		 * the PQ state of the interrupt through mmio. It may
 989		 * happen, for example when querying a PHB interrupt
 990		 * while the PHB is in an error state. We consider the
 991		 * interrupt to be inactive in that case.
 992		 */
 993		*state = (pq != XIVE_ESB_INVALID) && !xd->stale_p &&
 994			(xd->saved_p || !!(pq & XIVE_ESB_VAL_P));
 995		return 0;
 996	default:
 997		return -EINVAL;
 998	}
 999}
1000
1001static struct irq_chip xive_irq_chip = {
1002	.name = "XIVE-IRQ",
1003	.irq_startup = xive_irq_startup,
1004	.irq_shutdown = xive_irq_shutdown,
1005	.irq_eoi = xive_irq_eoi,
1006	.irq_mask = xive_irq_mask,
1007	.irq_unmask = xive_irq_unmask,
1008	.irq_set_affinity = xive_irq_set_affinity,
1009	.irq_set_type = xive_irq_set_type,
1010	.irq_retrigger = xive_irq_retrigger,
1011	.irq_set_vcpu_affinity = xive_irq_set_vcpu_affinity,
1012	.irq_get_irqchip_state = xive_get_irqchip_state,
1013};
1014
1015bool is_xive_irq(struct irq_chip *chip)
1016{
1017	return chip == &xive_irq_chip;
1018}
1019EXPORT_SYMBOL_GPL(is_xive_irq);
1020
1021void xive_cleanup_irq_data(struct xive_irq_data *xd)
1022{
1023	if (xd->eoi_mmio) {
1024		unmap_kernel_range((unsigned long)xd->eoi_mmio,
1025				   1u << xd->esb_shift);
1026		iounmap(xd->eoi_mmio);
1027		if (xd->eoi_mmio == xd->trig_mmio)
1028			xd->trig_mmio = NULL;
1029		xd->eoi_mmio = NULL;
1030	}
1031	if (xd->trig_mmio) {
1032		unmap_kernel_range((unsigned long)xd->trig_mmio,
1033				   1u << xd->esb_shift);
1034		iounmap(xd->trig_mmio);
1035		xd->trig_mmio = NULL;
1036	}
1037}
1038EXPORT_SYMBOL_GPL(xive_cleanup_irq_data);
1039
1040static int xive_irq_alloc_data(unsigned int virq, irq_hw_number_t hw)
1041{
1042	struct xive_irq_data *xd;
1043	int rc;
1044
1045	xd = kzalloc(sizeof(struct xive_irq_data), GFP_KERNEL);
1046	if (!xd)
1047		return -ENOMEM;
1048	rc = xive_ops->populate_irq_data(hw, xd);
1049	if (rc) {
1050		kfree(xd);
1051		return rc;
1052	}
1053	xd->target = XIVE_INVALID_TARGET;
1054	irq_set_handler_data(virq, xd);
1055
1056	/*
1057	 * Turn OFF by default the interrupt being mapped. A side
1058	 * effect of this check is the mapping the ESB page of the
1059	 * interrupt in the Linux address space. This prevents page
1060	 * fault issues in the crash handler which masks all
1061	 * interrupts.
1062	 */
1063	xive_esb_read(xd, XIVE_ESB_SET_PQ_01);
1064
1065	return 0;
1066}
1067
1068static void xive_irq_free_data(unsigned int virq)
1069{
1070	struct xive_irq_data *xd = irq_get_handler_data(virq);
1071
1072	if (!xd)
1073		return;
1074	irq_set_handler_data(virq, NULL);
1075	xive_cleanup_irq_data(xd);
1076	kfree(xd);
1077}
1078
1079#ifdef CONFIG_SMP
1080
1081static void xive_cause_ipi(int cpu)
1082{
1083	struct xive_cpu *xc;
1084	struct xive_irq_data *xd;
1085
1086	xc = per_cpu(xive_cpu, cpu);
1087
1088	DBG_VERBOSE("IPI CPU %d -> %d (HW IRQ 0x%x)\n",
1089		    smp_processor_id(), cpu, xc->hw_ipi);
1090
1091	xd = &xc->ipi_data;
1092	if (WARN_ON(!xd->trig_mmio))
1093		return;
1094	out_be64(xd->trig_mmio, 0);
1095}
1096
1097static irqreturn_t xive_muxed_ipi_action(int irq, void *dev_id)
1098{
1099	return smp_ipi_demux();
1100}
1101
1102static void xive_ipi_eoi(struct irq_data *d)
1103{
1104	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1105
1106	/* Handle possible race with unplug and drop stale IPIs */
1107	if (!xc)
1108		return;
1109
1110	DBG_VERBOSE("IPI eoi: irq=%d [0x%lx] (HW IRQ 0x%x) pending=%02x\n",
1111		    d->irq, irqd_to_hwirq(d), xc->hw_ipi, xc->pending_prio);
1112
1113	xive_do_source_eoi(xc->hw_ipi, &xc->ipi_data);
1114	xive_do_queue_eoi(xc);
1115}
1116
1117static void xive_ipi_do_nothing(struct irq_data *d)
1118{
1119	/*
1120	 * Nothing to do, we never mask/unmask IPIs, but the callback
1121	 * has to exist for the struct irq_chip.
1122	 */
1123}
1124
1125static struct irq_chip xive_ipi_chip = {
1126	.name = "XIVE-IPI",
1127	.irq_eoi = xive_ipi_eoi,
1128	.irq_mask = xive_ipi_do_nothing,
1129	.irq_unmask = xive_ipi_do_nothing,
1130};
1131
1132static void __init xive_request_ipi(void)
1133{
1134	unsigned int virq;
1135
1136	/*
1137	 * Initialization failed, move on, we might manage to
1138	 * reach the point where we display our errors before
1139	 * the system falls appart
1140	 */
1141	if (!xive_irq_domain)
1142		return;
1143
1144	/* Initialize it */
1145	virq = irq_create_mapping(xive_irq_domain, 0);
1146	xive_ipi_irq = virq;
1147
1148	WARN_ON(request_irq(virq, xive_muxed_ipi_action,
1149			    IRQF_PERCPU | IRQF_NO_THREAD, "IPI", NULL));
1150}
1151
1152static int xive_setup_cpu_ipi(unsigned int cpu)
1153{
1154	struct xive_cpu *xc;
1155	int rc;
1156
1157	pr_debug("Setting up IPI for CPU %d\n", cpu);
1158
1159	xc = per_cpu(xive_cpu, cpu);
1160
1161	/* Check if we are already setup */
1162	if (xc->hw_ipi != XIVE_BAD_IRQ)
1163		return 0;
1164
1165	/* Grab an IPI from the backend, this will populate xc->hw_ipi */
1166	if (xive_ops->get_ipi(cpu, xc))
1167		return -EIO;
1168
1169	/*
1170	 * Populate the IRQ data in the xive_cpu structure and
1171	 * configure the HW / enable the IPIs.
1172	 */
1173	rc = xive_ops->populate_irq_data(xc->hw_ipi, &xc->ipi_data);
1174	if (rc) {
1175		pr_err("Failed to populate IPI data on CPU %d\n", cpu);
1176		return -EIO;
1177	}
1178	rc = xive_ops->configure_irq(xc->hw_ipi,
1179				     get_hard_smp_processor_id(cpu),
1180				     xive_irq_priority, xive_ipi_irq);
1181	if (rc) {
1182		pr_err("Failed to map IPI CPU %d\n", cpu);
1183		return -EIO;
1184	}
1185	pr_devel("CPU %d HW IPI %x, virq %d, trig_mmio=%p\n", cpu,
1186	    xc->hw_ipi, xive_ipi_irq, xc->ipi_data.trig_mmio);
1187
1188	/* Unmask it */
1189	xive_do_source_set_mask(&xc->ipi_data, false);
1190
1191	return 0;
1192}
1193
1194static void xive_cleanup_cpu_ipi(unsigned int cpu, struct xive_cpu *xc)
1195{
1196	/* Disable the IPI and free the IRQ data */
1197
1198	/* Already cleaned up ? */
1199	if (xc->hw_ipi == XIVE_BAD_IRQ)
1200		return;
1201
1202	/* Mask the IPI */
1203	xive_do_source_set_mask(&xc->ipi_data, true);
1204
1205	/*
1206	 * Note: We don't call xive_cleanup_irq_data() to free
1207	 * the mappings as this is called from an IPI on kexec
1208	 * which is not a safe environment to call iounmap()
1209	 */
1210
1211	/* Deconfigure/mask in the backend */
1212	xive_ops->configure_irq(xc->hw_ipi, hard_smp_processor_id(),
1213				0xff, xive_ipi_irq);
1214
1215	/* Free the IPIs in the backend */
1216	xive_ops->put_ipi(cpu, xc);
1217}
1218
1219void __init xive_smp_probe(void)
1220{
1221	smp_ops->cause_ipi = xive_cause_ipi;
1222
1223	/* Register the IPI */
1224	xive_request_ipi();
1225
1226	/* Allocate and setup IPI for the boot CPU */
1227	xive_setup_cpu_ipi(smp_processor_id());
1228}
1229
1230#endif /* CONFIG_SMP */
1231
1232static int xive_irq_domain_map(struct irq_domain *h, unsigned int virq,
1233			       irq_hw_number_t hw)
1234{
1235	int rc;
1236
1237	/*
1238	 * Mark interrupts as edge sensitive by default so that resend
1239	 * actually works. Will fix that up below if needed.
1240	 */
1241	irq_clear_status_flags(virq, IRQ_LEVEL);
1242
1243#ifdef CONFIG_SMP
1244	/* IPIs are special and come up with HW number 0 */
1245	if (hw == 0) {
1246		/*
1247		 * IPIs are marked per-cpu. We use separate HW interrupts under
1248		 * the hood but associated with the same "linux" interrupt
1249		 */
1250		irq_set_chip_and_handler(virq, &xive_ipi_chip,
1251					 handle_percpu_irq);
1252		return 0;
1253	}
1254#endif
1255
1256	rc = xive_irq_alloc_data(virq, hw);
1257	if (rc)
1258		return rc;
1259
1260	irq_set_chip_and_handler(virq, &xive_irq_chip, handle_fasteoi_irq);
1261
1262	return 0;
1263}
1264
1265static void xive_irq_domain_unmap(struct irq_domain *d, unsigned int virq)
1266{
1267	struct irq_data *data = irq_get_irq_data(virq);
1268	unsigned int hw_irq;
1269
1270	/* XXX Assign BAD number */
1271	if (!data)
1272		return;
1273	hw_irq = (unsigned int)irqd_to_hwirq(data);
1274	if (hw_irq)
1275		xive_irq_free_data(virq);
1276}
1277
1278static int xive_irq_domain_xlate(struct irq_domain *h, struct device_node *ct,
1279				 const u32 *intspec, unsigned int intsize,
1280				 irq_hw_number_t *out_hwirq, unsigned int *out_flags)
1281
1282{
1283	*out_hwirq = intspec[0];
1284
1285	/*
1286	 * If intsize is at least 2, we look for the type in the second cell,
1287	 * we assume the LSB indicates a level interrupt.
1288	 */
1289	if (intsize > 1) {
1290		if (intspec[1] & 1)
1291			*out_flags = IRQ_TYPE_LEVEL_LOW;
1292		else
1293			*out_flags = IRQ_TYPE_EDGE_RISING;
1294	} else
1295		*out_flags = IRQ_TYPE_LEVEL_LOW;
1296
1297	return 0;
1298}
1299
1300static int xive_irq_domain_match(struct irq_domain *h, struct device_node *node,
1301				 enum irq_domain_bus_token bus_token)
1302{
1303	return xive_ops->match(node);
1304}
1305
1306static const struct irq_domain_ops xive_irq_domain_ops = {
1307	.match = xive_irq_domain_match,
1308	.map = xive_irq_domain_map,
1309	.unmap = xive_irq_domain_unmap,
1310	.xlate = xive_irq_domain_xlate,
1311};
1312
1313static void __init xive_init_host(void)
1314{
1315	xive_irq_domain = irq_domain_add_nomap(NULL, XIVE_MAX_IRQ,
1316					       &xive_irq_domain_ops, NULL);
1317	if (WARN_ON(xive_irq_domain == NULL))
1318		return;
1319	irq_set_default_host(xive_irq_domain);
1320}
1321
1322static void xive_cleanup_cpu_queues(unsigned int cpu, struct xive_cpu *xc)
1323{
1324	if (xc->queue[xive_irq_priority].qpage)
1325		xive_ops->cleanup_queue(cpu, xc, xive_irq_priority);
1326}
1327
1328static int xive_setup_cpu_queues(unsigned int cpu, struct xive_cpu *xc)
1329{
1330	int rc = 0;
1331
1332	/* We setup 1 queues for now with a 64k page */
1333	if (!xc->queue[xive_irq_priority].qpage)
1334		rc = xive_ops->setup_queue(cpu, xc, xive_irq_priority);
1335
1336	return rc;
1337}
1338
1339static int xive_prepare_cpu(unsigned int cpu)
1340{
1341	struct xive_cpu *xc;
1342
1343	xc = per_cpu(xive_cpu, cpu);
1344	if (!xc) {
1345		struct device_node *np;
1346
1347		xc = kzalloc_node(sizeof(struct xive_cpu),
1348				  GFP_KERNEL, cpu_to_node(cpu));
1349		if (!xc)
1350			return -ENOMEM;
1351		np = of_get_cpu_node(cpu, NULL);
1352		if (np)
1353			xc->chip_id = of_get_ibm_chip_id(np);
1354		of_node_put(np);
1355		xc->hw_ipi = XIVE_BAD_IRQ;
1356
1357		per_cpu(xive_cpu, cpu) = xc;
1358	}
1359
1360	/* Setup EQs if not already */
1361	return xive_setup_cpu_queues(cpu, xc);
1362}
1363
1364static void xive_setup_cpu(void)
1365{
1366	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1367
1368	/* The backend might have additional things to do */
1369	if (xive_ops->setup_cpu)
1370		xive_ops->setup_cpu(smp_processor_id(), xc);
1371
1372	/* Set CPPR to 0xff to enable flow of interrupts */
1373	xc->cppr = 0xff;
1374	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0xff);
1375}
1376
1377#ifdef CONFIG_SMP
1378void xive_smp_setup_cpu(void)
1379{
1380	pr_devel("SMP setup CPU %d\n", smp_processor_id());
1381
1382	/* This will have already been done on the boot CPU */
1383	if (smp_processor_id() != boot_cpuid)
1384		xive_setup_cpu();
1385
1386}
1387
1388int xive_smp_prepare_cpu(unsigned int cpu)
1389{
1390	int rc;
1391
1392	/* Allocate per-CPU data and queues */
1393	rc = xive_prepare_cpu(cpu);
1394	if (rc)
1395		return rc;
1396
1397	/* Allocate and setup IPI for the new CPU */
1398	return xive_setup_cpu_ipi(cpu);
1399}
1400
1401#ifdef CONFIG_HOTPLUG_CPU
1402static void xive_flush_cpu_queue(unsigned int cpu, struct xive_cpu *xc)
1403{
1404	u32 irq;
1405
1406	/* We assume local irqs are disabled */
1407	WARN_ON(!irqs_disabled());
1408
1409	/* Check what's already in the CPU queue */
1410	while ((irq = xive_scan_interrupts(xc, false)) != 0) {
1411		/*
1412		 * We need to re-route that interrupt to its new destination.
1413		 * First get and lock the descriptor
1414		 */
1415		struct irq_desc *desc = irq_to_desc(irq);
1416		struct irq_data *d = irq_desc_get_irq_data(desc);
1417		struct xive_irq_data *xd;
1418		unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
1419
1420		/*
1421		 * Ignore anything that isn't a XIVE irq and ignore
1422		 * IPIs, so can just be dropped.
1423		 */
1424		if (d->domain != xive_irq_domain || hw_irq == 0)
1425			continue;
1426
1427		/*
1428		 * The IRQ should have already been re-routed, it's just a
1429		 * stale in the old queue, so re-trigger it in order to make
1430		 * it reach is new destination.
1431		 */
1432#ifdef DEBUG_FLUSH
1433		pr_info("CPU %d: Got irq %d while offline, re-sending...\n",
1434			cpu, irq);
1435#endif
1436		raw_spin_lock(&desc->lock);
1437		xd = irq_desc_get_handler_data(desc);
1438
1439		/*
1440		 * Clear saved_p to indicate that it's no longer pending
1441		 */
1442		xd->saved_p = false;
1443
1444		/*
1445		 * For LSIs, we EOI, this will cause a resend if it's
1446		 * still asserted. Otherwise do an MSI retrigger.
1447		 */
1448		if (xd->flags & XIVE_IRQ_FLAG_LSI)
1449			xive_do_source_eoi(irqd_to_hwirq(d), xd);
1450		else
1451			xive_irq_retrigger(d);
1452
1453		raw_spin_unlock(&desc->lock);
1454	}
1455}
1456
1457void xive_smp_disable_cpu(void)
1458{
1459	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1460	unsigned int cpu = smp_processor_id();
1461
1462	/* Migrate interrupts away from the CPU */
1463	irq_migrate_all_off_this_cpu();
1464
1465	/* Set CPPR to 0 to disable flow of interrupts */
1466	xc->cppr = 0;
1467	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0);
1468
1469	/* Flush everything still in the queue */
1470	xive_flush_cpu_queue(cpu, xc);
1471
1472	/* Re-enable CPPR  */
1473	xc->cppr = 0xff;
1474	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0xff);
1475}
1476
1477void xive_flush_interrupt(void)
1478{
1479	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1480	unsigned int cpu = smp_processor_id();
1481
1482	/* Called if an interrupt occurs while the CPU is hot unplugged */
1483	xive_flush_cpu_queue(cpu, xc);
1484}
1485
1486#endif /* CONFIG_HOTPLUG_CPU */
1487
1488#endif /* CONFIG_SMP */
1489
1490void xive_teardown_cpu(void)
1491{
1492	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1493	unsigned int cpu = smp_processor_id();
1494
1495	/* Set CPPR to 0 to disable flow of interrupts */
1496	xc->cppr = 0;
1497	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0);
1498
1499	if (xive_ops->teardown_cpu)
1500		xive_ops->teardown_cpu(cpu, xc);
1501
1502#ifdef CONFIG_SMP
1503	/* Get rid of IPI */
1504	xive_cleanup_cpu_ipi(cpu, xc);
1505#endif
1506
1507	/* Disable and free the queues */
1508	xive_cleanup_cpu_queues(cpu, xc);
1509}
1510
1511void xive_shutdown(void)
1512{
1513	xive_ops->shutdown();
1514}
1515
1516bool __init xive_core_init(const struct xive_ops *ops, void __iomem *area, u32 offset,
1517			   u8 max_prio)
1518{
1519	xive_tima = area;
1520	xive_tima_offset = offset;
1521	xive_ops = ops;
1522	xive_irq_priority = max_prio;
1523
1524	ppc_md.get_irq = xive_get_irq;
1525	__xive_enabled = true;
1526
1527	pr_devel("Initializing host..\n");
1528	xive_init_host();
1529
1530	pr_devel("Initializing boot CPU..\n");
1531
1532	/* Allocate per-CPU data and queues */
1533	xive_prepare_cpu(smp_processor_id());
1534
1535	/* Get ready for interrupts */
1536	xive_setup_cpu();
1537
1538	pr_info("Interrupt handling initialized with %s backend\n",
1539		xive_ops->name);
1540	pr_info("Using priority %d for all interrupts\n", max_prio);
1541
1542	return true;
1543}
1544
1545__be32 *xive_queue_page_alloc(unsigned int cpu, u32 queue_shift)
1546{
1547	unsigned int alloc_order;
1548	struct page *pages;
1549	__be32 *qpage;
1550
1551	alloc_order = xive_alloc_order(queue_shift);
1552	pages = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, alloc_order);
1553	if (!pages)
1554		return ERR_PTR(-ENOMEM);
1555	qpage = (__be32 *)page_address(pages);
1556	memset(qpage, 0, 1 << queue_shift);
1557
1558	return qpage;
1559}
1560
1561static int __init xive_off(char *arg)
1562{
1563	xive_cmdline_disabled = true;
1564	return 0;
1565}
1566__setup("xive=off", xive_off);
1567
1568void xive_debug_show_cpu(struct seq_file *m, int cpu)
1569{
1570	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
1571
1572	seq_printf(m, "CPU %d:", cpu);
1573	if (xc) {
1574		seq_printf(m, "pp=%02x CPPR=%02x ", xc->pending_prio, xc->cppr);
1575
1576#ifdef CONFIG_SMP
1577		{
1578			u64 val = xive_esb_read(&xc->ipi_data, XIVE_ESB_GET);
1579
1580			seq_printf(m, "IPI=0x%08x PQ=%c%c ", xc->hw_ipi,
1581				   val & XIVE_ESB_VAL_P ? 'P' : '-',
1582				   val & XIVE_ESB_VAL_Q ? 'Q' : '-');
1583		}
1584#endif
1585		{
1586			struct xive_q *q = &xc->queue[xive_irq_priority];
1587			u32 i0, i1, idx;
1588
1589			if (q->qpage) {
1590				idx = q->idx;
1591				i0 = be32_to_cpup(q->qpage + idx);
1592				idx = (idx + 1) & q->msk;
1593				i1 = be32_to_cpup(q->qpage + idx);
1594				seq_printf(m, "EQ idx=%d T=%d %08x %08x ...",
1595					   q->idx, q->toggle, i0, i1);
1596			}
1597		}
1598	}
1599	seq_puts(m, "\n");
1600}
1601
1602void xive_debug_show_irq(struct seq_file *m, u32 hw_irq, struct irq_data *d)
1603{
1604	struct irq_chip *chip = irq_data_get_irq_chip(d);
1605	int rc;
1606	u32 target;
1607	u8 prio;
1608	u32 lirq;
1609
1610	if (!is_xive_irq(chip))
1611		return;
1612
1613	rc = xive_ops->get_irq_config(hw_irq, &target, &prio, &lirq);
1614	if (rc) {
1615		seq_printf(m, "IRQ 0x%08x : no config rc=%d\n", hw_irq, rc);
1616		return;
1617	}
1618
1619	seq_printf(m, "IRQ 0x%08x : target=0x%x prio=%02x lirq=0x%x ",
1620		   hw_irq, target, prio, lirq);
1621
1622	if (d) {
1623		struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
1624		u64 val = xive_esb_read(xd, XIVE_ESB_GET);
1625
1626		seq_printf(m, "flags=%c%c%c PQ=%c%c",
1627			   xd->flags & XIVE_IRQ_FLAG_STORE_EOI ? 'S' : ' ',
1628			   xd->flags & XIVE_IRQ_FLAG_LSI ? 'L' : ' ',
1629			   xd->flags & XIVE_IRQ_FLAG_H_INT_ESB ? 'H' : ' ',
1630			   val & XIVE_ESB_VAL_P ? 'P' : '-',
1631			   val & XIVE_ESB_VAL_Q ? 'Q' : '-');
1632	}
1633	seq_puts(m, "\n");
1634}
1635
1636static int xive_core_debug_show(struct seq_file *m, void *private)
1637{
1638	unsigned int i;
1639	struct irq_desc *desc;
1640	int cpu;
1641
1642	if (xive_ops->debug_show)
1643		xive_ops->debug_show(m, private);
1644
1645	for_each_possible_cpu(cpu)
1646		xive_debug_show_cpu(m, cpu);
1647
1648	for_each_irq_desc(i, desc) {
1649		struct irq_data *d = irq_desc_get_irq_data(desc);
1650		unsigned int hw_irq;
1651
1652		if (!d)
1653			continue;
1654
1655		hw_irq = (unsigned int)irqd_to_hwirq(d);
1656
1657		/* IPIs are special (HW number 0) */
1658		if (hw_irq)
1659			xive_debug_show_irq(m, hw_irq, d);
1660	}
1661	return 0;
1662}
1663DEFINE_SHOW_ATTRIBUTE(xive_core_debug);
1664
1665int xive_core_debug_init(void)
1666{
1667	if (xive_enabled())
1668		debugfs_create_file("xive", 0400, powerpc_debugfs_root,
1669				    NULL, &xive_core_debug_fops);
1670	return 0;
1671}