1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright 2016,2017 IBM Corporation.
6 #define pr_fmt(fmt) "xive: " fmt
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>
19 #include <linux/slab.h>
20 #include <linux/spinlock.h>
21 #include <linux/msi.h>
26 #include <asm/machdep.h>
28 #include <asm/errno.h>
30 #include <asm/xive-regs.h>
33 #include "xive-internal.h"
39 #define DBG_VERBOSE(fmt, ...) pr_devel("cpu %d - " fmt, \
40 smp_processor_id(), ## __VA_ARGS__)
42 #define DBG_VERBOSE(fmt...) do { } while(0)
46 EXPORT_SYMBOL_GPL(__xive_enabled);
47 bool xive_cmdline_disabled;
49 /* We use only one priority for now */
50 static u8 xive_irq_priority;
52 /* TIMA exported to KVM */
53 void __iomem *xive_tima;
54 EXPORT_SYMBOL_GPL(xive_tima);
58 static const struct xive_ops *xive_ops;
60 /* Our global interrupt domain */
61 static struct irq_domain *xive_irq_domain;
64 /* The IPIs all use the same logical irq number */
65 static u32 xive_ipi_irq;
68 /* Xive state for each CPU */
69 static DEFINE_PER_CPU(struct xive_cpu *, xive_cpu);
72 * A "disabled" interrupt should never fire, to catch problems
73 * we set its logical number to this
75 #define XIVE_BAD_IRQ 0x7fffffff
76 #define XIVE_MAX_IRQ (XIVE_BAD_IRQ - 1)
78 /* An invalid CPU target */
79 #define XIVE_INVALID_TARGET (-1)
82 * Read the next entry in a queue, return its content if it's valid
83 * or 0 if there is no new entry.
85 * The queue pointer is moved forward unless "just_peek" is set
87 static u32 xive_read_eq(struct xive_q *q, bool just_peek)
93 cur = be32_to_cpup(q->qpage + q->idx);
95 /* Check valid bit (31) vs current toggle polarity */
96 if ((cur >> 31) == q->toggle)
99 /* If consuming from the queue ... */
102 q->idx = (q->idx + 1) & q->msk;
104 /* Wrap around: flip valid toggle */
108 /* Mask out the valid bit (31) */
109 return cur & 0x7fffffff;
113 * Scans all the queue that may have interrupts in them
114 * (based on "pending_prio") in priority order until an
115 * interrupt is found or all the queues are empty.
117 * Then updates the CPPR (Current Processor Priority
118 * Register) based on the most favored interrupt found
119 * (0xff if none) and return what was found (0 if none).
121 * If just_peek is set, return the most favored pending
122 * interrupt if any but don't update the queue pointers.
124 * Note: This function can operate generically on any number
125 * of queues (up to 8). The current implementation of the XIVE
126 * driver only uses a single queue however.
128 * Note2: This will also "flush" "the pending_count" of a queue
129 * into the "count" when that queue is observed to be empty.
130 * This is used to keep track of the amount of interrupts
131 * targetting a queue. When an interrupt is moved away from
132 * a queue, we only decrement that queue count once the queue
133 * has been observed empty to avoid races.
135 static u32 xive_scan_interrupts(struct xive_cpu *xc, bool just_peek)
140 /* Find highest pending priority */
141 while (xc->pending_prio != 0) {
144 prio = ffs(xc->pending_prio) - 1;
145 DBG_VERBOSE("scan_irq: trying prio %d\n", prio);
148 irq = xive_read_eq(&xc->queue[prio], just_peek);
150 /* Found something ? That's it */
154 /* Clear pending bits */
155 xc->pending_prio &= ~(1 << prio);
158 * Check if the queue count needs adjusting due to
159 * interrupts being moved away. See description of
160 * xive_dec_target_count()
162 q = &xc->queue[prio];
163 if (atomic_read(&q->pending_count)) {
164 int p = atomic_xchg(&q->pending_count, 0);
166 WARN_ON(p > atomic_read(&q->count));
167 atomic_sub(p, &q->count);
172 /* If nothing was found, set CPPR to 0xff */
176 /* Update HW CPPR to match if necessary */
177 if (prio != xc->cppr) {
178 DBG_VERBOSE("scan_irq: adjusting CPPR to %d\n", prio);
180 out_8(xive_tima + xive_tima_offset + TM_CPPR, prio);
187 * This is used to perform the magic loads from an ESB
188 * described in xive.h
190 static notrace u8 xive_esb_read(struct xive_irq_data *xd, u32 offset)
194 /* Handle HW errata */
195 if (xd->flags & XIVE_IRQ_FLAG_SHIFT_BUG)
196 offset |= offset << 4;
198 if ((xd->flags & XIVE_IRQ_FLAG_H_INT_ESB) && xive_ops->esb_rw)
199 val = xive_ops->esb_rw(xd->hw_irq, offset, 0, 0);
201 val = in_be64(xd->eoi_mmio + offset);
206 static void xive_esb_write(struct xive_irq_data *xd, u32 offset, u64 data)
208 /* Handle HW errata */
209 if (xd->flags & XIVE_IRQ_FLAG_SHIFT_BUG)
210 offset |= offset << 4;
212 if ((xd->flags & XIVE_IRQ_FLAG_H_INT_ESB) && xive_ops->esb_rw)
213 xive_ops->esb_rw(xd->hw_irq, offset, data, 1);
215 out_be64(xd->eoi_mmio + offset, data);
219 static notrace void xive_dump_eq(const char *name, struct xive_q *q)
226 i0 = be32_to_cpup(q->qpage + idx);
227 idx = (idx + 1) & q->msk;
228 i1 = be32_to_cpup(q->qpage + idx);
229 xmon_printf(" %s Q T=%d %08x %08x ...\n", name,
233 notrace void xmon_xive_do_dump(int cpu)
235 struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
237 xmon_printf("XIVE state for CPU %d:\n", cpu);
238 xmon_printf(" pp=%02x cppr=%02x\n", xc->pending_prio, xc->cppr);
239 xive_dump_eq("IRQ", &xc->queue[xive_irq_priority]);
242 u64 val = xive_esb_read(&xc->ipi_data, XIVE_ESB_GET);
243 xmon_printf(" IPI state: %x:%c%c\n", xc->hw_ipi,
244 val & XIVE_ESB_VAL_P ? 'P' : 'p',
245 val & XIVE_ESB_VAL_Q ? 'Q' : 'q');
249 #endif /* CONFIG_XMON */
251 static unsigned int xive_get_irq(void)
253 struct xive_cpu *xc = __this_cpu_read(xive_cpu);
257 * This can be called either as a result of a HW interrupt or
258 * as a "replay" because EOI decided there was still something
259 * in one of the queues.
261 * First we perform an ACK cycle in order to update our mask
262 * of pending priorities. This will also have the effect of
263 * updating the CPPR to the most favored pending interrupts.
265 * In the future, if we have a way to differentiate a first
266 * entry (on HW interrupt) from a replay triggered by EOI,
267 * we could skip this on replays unless we soft-mask tells us
268 * that a new HW interrupt occurred.
270 xive_ops->update_pending(xc);
272 DBG_VERBOSE("get_irq: pending=%02x\n", xc->pending_prio);
274 /* Scan our queue(s) for interrupts */
275 irq = xive_scan_interrupts(xc, false);
277 DBG_VERBOSE("get_irq: got irq 0x%x, new pending=0x%02x\n",
278 irq, xc->pending_prio);
280 /* Return pending interrupt if any */
281 if (irq == XIVE_BAD_IRQ)
287 * After EOI'ing an interrupt, we need to re-check the queue
288 * to see if another interrupt is pending since multiple
289 * interrupts can coalesce into a single notification to the
292 * If we find that there is indeed more in there, we call
293 * force_external_irq_replay() to make Linux synthetize an
294 * external interrupt on the next call to local_irq_restore().
296 static void xive_do_queue_eoi(struct xive_cpu *xc)
298 if (xive_scan_interrupts(xc, true) != 0) {
299 DBG_VERBOSE("eoi: pending=0x%02x\n", xc->pending_prio);
300 force_external_irq_replay();
305 * EOI an interrupt at the source. There are several methods
306 * to do this depending on the HW version and source type
308 static void xive_do_source_eoi(u32 hw_irq, struct xive_irq_data *xd)
310 /* If the XIVE supports the new "store EOI facility, use it */
311 if (xd->flags & XIVE_IRQ_FLAG_STORE_EOI)
312 xive_esb_write(xd, XIVE_ESB_STORE_EOI, 0);
313 else if (hw_irq && xd->flags & XIVE_IRQ_FLAG_EOI_FW) {
315 * The FW told us to call it. This happens for some
316 * interrupt sources that need additional HW whacking
317 * beyond the ESB manipulation. For example LPC interrupts
318 * on P9 DD1.0 needed a latch to be clared in the LPC bridge
319 * itself. The Firmware will take care of it.
321 if (WARN_ON_ONCE(!xive_ops->eoi))
323 xive_ops->eoi(hw_irq);
328 * Otherwise for EOI, we use the special MMIO that does
329 * a clear of both P and Q and returns the old Q,
330 * except for LSIs where we use the "EOI cycle" special
333 * This allows us to then do a re-trigger if Q was set
334 * rather than synthesizing an interrupt in software
336 * For LSIs the HW EOI cycle is used rather than PQ bits,
337 * as they are automatically re-triggred in HW when still
340 if (xd->flags & XIVE_IRQ_FLAG_LSI)
341 xive_esb_read(xd, XIVE_ESB_LOAD_EOI);
343 eoi_val = xive_esb_read(xd, XIVE_ESB_SET_PQ_00);
344 DBG_VERBOSE("eoi_val=%x\n", eoi_val);
346 /* Re-trigger if needed */
347 if ((eoi_val & XIVE_ESB_VAL_Q) && xd->trig_mmio)
348 out_be64(xd->trig_mmio, 0);
353 /* irq_chip eoi callback */
354 static void xive_irq_eoi(struct irq_data *d)
356 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
357 struct xive_cpu *xc = __this_cpu_read(xive_cpu);
359 DBG_VERBOSE("eoi_irq: irq=%d [0x%lx] pending=%02x\n",
360 d->irq, irqd_to_hwirq(d), xc->pending_prio);
363 * EOI the source if it hasn't been disabled and hasn't
364 * been passed-through to a KVM guest
366 if (!irqd_irq_disabled(d) && !irqd_is_forwarded_to_vcpu(d) &&
367 !(xd->flags & XIVE_IRQ_NO_EOI))
368 xive_do_source_eoi(irqd_to_hwirq(d), xd);
371 * Clear saved_p to indicate that it's no longer occupying
372 * a queue slot on the target queue
376 /* Check for more work in the queue */
377 xive_do_queue_eoi(xc);
381 * Helper used to mask and unmask an interrupt source. This
382 * is only called for normal interrupts that do not require
383 * masking/unmasking via firmware.
385 static void xive_do_source_set_mask(struct xive_irq_data *xd,
391 * If the interrupt had P set, it may be in a queue.
393 * We need to make sure we don't re-enable it until it
394 * has been fetched from that queue and EOId. We keep
395 * a copy of that P state and use it to restore the
396 * ESB accordingly on unmask.
399 val = xive_esb_read(xd, XIVE_ESB_SET_PQ_01);
400 xd->saved_p = !!(val & XIVE_ESB_VAL_P);
401 } else if (xd->saved_p)
402 xive_esb_read(xd, XIVE_ESB_SET_PQ_10);
404 xive_esb_read(xd, XIVE_ESB_SET_PQ_00);
408 * Try to chose "cpu" as a new interrupt target. Increments
409 * the queue accounting for that target if it's not already
412 static bool xive_try_pick_target(int cpu)
414 struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
415 struct xive_q *q = &xc->queue[xive_irq_priority];
419 * Calculate max number of interrupts in that queue.
421 * We leave a gap of 1 just in case...
423 max = (q->msk + 1) - 1;
424 return !!atomic_add_unless(&q->count, 1, max);
428 * Un-account an interrupt for a target CPU. We don't directly
429 * decrement q->count since the interrupt might still be present
432 * Instead increment a separate counter "pending_count" which
433 * will be substracted from "count" later when that CPU observes
434 * the queue to be empty.
436 static void xive_dec_target_count(int cpu)
438 struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
439 struct xive_q *q = &xc->queue[xive_irq_priority];
441 if (WARN_ON(cpu < 0 || !xc)) {
442 pr_err("%s: cpu=%d xc=%p\n", __func__, cpu, xc);
447 * We increment the "pending count" which will be used
448 * to decrement the target queue count whenever it's next
449 * processed and found empty. This ensure that we don't
450 * decrement while we still have the interrupt there
453 atomic_inc(&q->pending_count);
456 /* Find a tentative CPU target in a CPU mask */
457 static int xive_find_target_in_mask(const struct cpumask *mask,
460 int cpu, first, num, i;
462 /* Pick up a starting point CPU in the mask based on fuzz */
463 num = min_t(int, cpumask_weight(mask), nr_cpu_ids);
467 cpu = cpumask_first(mask);
468 for (i = 0; i < first && cpu < nr_cpu_ids; i++)
469 cpu = cpumask_next(cpu, mask);
472 if (WARN_ON(cpu >= nr_cpu_ids))
473 cpu = cpumask_first(cpu_online_mask);
475 /* Remember first one to handle wrap-around */
479 * Now go through the entire mask until we find a valid
484 * We re-check online as the fallback case passes us
485 * an untested affinity mask
487 if (cpu_online(cpu) && xive_try_pick_target(cpu))
489 cpu = cpumask_next(cpu, mask);
491 if (cpu >= nr_cpu_ids)
492 cpu = cpumask_first(mask);
493 } while (cpu != first);
499 * Pick a target CPU for an interrupt. This is done at
500 * startup or if the affinity is changed in a way that
501 * invalidates the current target.
503 static int xive_pick_irq_target(struct irq_data *d,
504 const struct cpumask *affinity)
506 static unsigned int fuzz;
507 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
512 * If we have chip IDs, first we try to build a mask of
513 * CPUs matching the CPU and find a target in there
515 if (xd->src_chip != XIVE_INVALID_CHIP_ID &&
516 zalloc_cpumask_var(&mask, GFP_ATOMIC)) {
517 /* Build a mask of matching chip IDs */
518 for_each_cpu_and(cpu, affinity, cpu_online_mask) {
519 struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
520 if (xc->chip_id == xd->src_chip)
521 cpumask_set_cpu(cpu, mask);
523 /* Try to find a target */
524 if (cpumask_empty(mask))
527 cpu = xive_find_target_in_mask(mask, fuzz++);
528 free_cpumask_var(mask);
534 /* No chip IDs, fallback to using the affinity mask */
535 return xive_find_target_in_mask(affinity, fuzz++);
538 static unsigned int xive_irq_startup(struct irq_data *d)
540 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
541 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
544 pr_devel("xive_irq_startup: irq %d [0x%x] data @%p\n",
547 #ifdef CONFIG_PCI_MSI
549 * The generic MSI code returns with the interrupt disabled on the
550 * card, using the MSI mask bits. Firmware doesn't appear to unmask
551 * at that level, so we do it here by hand.
553 if (irq_data_get_msi_desc(d))
554 pci_msi_unmask_irq(d);
558 target = xive_pick_irq_target(d, irq_data_get_affinity_mask(d));
559 if (target == XIVE_INVALID_TARGET) {
560 /* Try again breaking affinity */
561 target = xive_pick_irq_target(d, cpu_online_mask);
562 if (target == XIVE_INVALID_TARGET)
564 pr_warn("irq %d started with broken affinity\n", d->irq);
568 if (WARN_ON(target == XIVE_INVALID_TARGET ||
569 target >= nr_cpu_ids))
570 target = smp_processor_id();
575 * Configure the logical number to be the Linux IRQ number
576 * and set the target queue
578 rc = xive_ops->configure_irq(hw_irq,
579 get_hard_smp_processor_id(target),
580 xive_irq_priority, d->irq);
585 xive_do_source_set_mask(xd, false);
590 static void xive_irq_shutdown(struct irq_data *d)
592 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
593 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
595 pr_devel("xive_irq_shutdown: irq %d [0x%x] data @%p\n",
598 if (WARN_ON(xd->target == XIVE_INVALID_TARGET))
601 /* Mask the interrupt at the source */
602 xive_do_source_set_mask(xd, true);
605 * The above may have set saved_p. We clear it otherwise it
606 * will prevent re-enabling later on. It is ok to forget the
607 * fact that the interrupt might be in a queue because we are
608 * accounting that already in xive_dec_target_count() and will
609 * be re-routing it to a new queue with proper accounting when
610 * it's started up again
615 * Mask the interrupt in HW in the IVT/EAS and set the number
616 * to be the "bad" IRQ number
618 xive_ops->configure_irq(hw_irq,
619 get_hard_smp_processor_id(xd->target),
622 xive_dec_target_count(xd->target);
623 xd->target = XIVE_INVALID_TARGET;
626 static void xive_irq_unmask(struct irq_data *d)
628 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
630 pr_devel("xive_irq_unmask: irq %d data @%p\n", d->irq, xd);
633 * This is a workaround for PCI LSI problems on P9, for
634 * these, we call FW to set the mask. The problems might
635 * be fixed by P9 DD2.0, if that is the case, firmware
636 * will no longer set that flag.
638 if (xd->flags & XIVE_IRQ_FLAG_MASK_FW) {
639 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
640 xive_ops->configure_irq(hw_irq,
641 get_hard_smp_processor_id(xd->target),
642 xive_irq_priority, d->irq);
646 xive_do_source_set_mask(xd, false);
649 static void xive_irq_mask(struct irq_data *d)
651 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
653 pr_devel("xive_irq_mask: irq %d data @%p\n", d->irq, xd);
656 * This is a workaround for PCI LSI problems on P9, for
657 * these, we call OPAL to set the mask. The problems might
658 * be fixed by P9 DD2.0, if that is the case, firmware
659 * will no longer set that flag.
661 if (xd->flags & XIVE_IRQ_FLAG_MASK_FW) {
662 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
663 xive_ops->configure_irq(hw_irq,
664 get_hard_smp_processor_id(xd->target),
669 xive_do_source_set_mask(xd, true);
672 static int xive_irq_set_affinity(struct irq_data *d,
673 const struct cpumask *cpumask,
676 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
677 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
678 u32 target, old_target;
681 pr_devel("xive_irq_set_affinity: irq %d\n", d->irq);
683 /* Is this valid ? */
684 if (cpumask_any_and(cpumask, cpu_online_mask) >= nr_cpu_ids)
687 /* Don't do anything if the interrupt isn't started */
688 if (!irqd_is_started(d))
689 return IRQ_SET_MASK_OK;
692 * If existing target is already in the new mask, and is
693 * online then do nothing.
695 if (xd->target != XIVE_INVALID_TARGET &&
696 cpu_online(xd->target) &&
697 cpumask_test_cpu(xd->target, cpumask))
698 return IRQ_SET_MASK_OK;
700 /* Pick a new target */
701 target = xive_pick_irq_target(d, cpumask);
703 /* No target found */
704 if (target == XIVE_INVALID_TARGET)
708 if (WARN_ON(target >= nr_cpu_ids))
709 target = smp_processor_id();
711 old_target = xd->target;
714 * Only configure the irq if it's not currently passed-through to
717 if (!irqd_is_forwarded_to_vcpu(d))
718 rc = xive_ops->configure_irq(hw_irq,
719 get_hard_smp_processor_id(target),
720 xive_irq_priority, d->irq);
722 pr_err("Error %d reconfiguring irq %d\n", rc, d->irq);
726 pr_devel(" target: 0x%x\n", target);
729 /* Give up previous target */
730 if (old_target != XIVE_INVALID_TARGET)
731 xive_dec_target_count(old_target);
733 return IRQ_SET_MASK_OK;
736 static int xive_irq_set_type(struct irq_data *d, unsigned int flow_type)
738 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
741 * We only support these. This has really no effect other than setting
742 * the corresponding descriptor bits mind you but those will in turn
743 * affect the resend function when re-enabling an edge interrupt.
745 * Set set the default to edge as explained in map().
747 if (flow_type == IRQ_TYPE_DEFAULT || flow_type == IRQ_TYPE_NONE)
748 flow_type = IRQ_TYPE_EDGE_RISING;
750 if (flow_type != IRQ_TYPE_EDGE_RISING &&
751 flow_type != IRQ_TYPE_LEVEL_LOW)
754 irqd_set_trigger_type(d, flow_type);
757 * Double check it matches what the FW thinks
759 * NOTE: We don't know yet if the PAPR interface will provide
760 * the LSI vs MSI information apart from the device-tree so
761 * this check might have to move into an optional backend call
762 * that is specific to the native backend
764 if ((flow_type == IRQ_TYPE_LEVEL_LOW) !=
765 !!(xd->flags & XIVE_IRQ_FLAG_LSI)) {
766 pr_warn("Interrupt %d (HW 0x%x) type mismatch, Linux says %s, FW says %s\n",
767 d->irq, (u32)irqd_to_hwirq(d),
768 (flow_type == IRQ_TYPE_LEVEL_LOW) ? "Level" : "Edge",
769 (xd->flags & XIVE_IRQ_FLAG_LSI) ? "Level" : "Edge");
772 return IRQ_SET_MASK_OK_NOCOPY;
775 static int xive_irq_retrigger(struct irq_data *d)
777 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
779 /* This should be only for MSIs */
780 if (WARN_ON(xd->flags & XIVE_IRQ_FLAG_LSI))
784 * To perform a retrigger, we first set the PQ bits to
785 * 11, then perform an EOI.
787 xive_esb_read(xd, XIVE_ESB_SET_PQ_11);
790 * Note: We pass "0" to the hw_irq argument in order to
791 * avoid calling into the backend EOI code which we don't
792 * want to do in the case of a re-trigger. Backends typically
793 * only do EOI for LSIs anyway.
795 xive_do_source_eoi(0, xd);
800 static int xive_irq_set_vcpu_affinity(struct irq_data *d, void *state)
802 struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
803 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
808 * We only support this on interrupts that do not require
809 * firmware calls for masking and unmasking
811 if (xd->flags & XIVE_IRQ_FLAG_MASK_FW)
815 * This is called by KVM with state non-NULL for enabling
816 * pass-through or NULL for disabling it
819 irqd_set_forwarded_to_vcpu(d);
821 /* Set it to PQ=10 state to prevent further sends */
822 pq = xive_esb_read(xd, XIVE_ESB_SET_PQ_10);
824 /* No target ? nothing to do */
825 if (xd->target == XIVE_INVALID_TARGET) {
827 * An untargetted interrupt should have been
828 * also masked at the source
836 * If P was set, adjust state to PQ=11 to indicate
837 * that a resend is needed for the interrupt to reach
838 * the guest. Also remember the value of P.
840 * This also tells us that it's in flight to a host queue
841 * or has already been fetched but hasn't been EOIed yet
842 * by the host. This it's potentially using up a host
843 * queue slot. This is important to know because as long
844 * as this is the case, we must not hard-unmask it when
845 * "returning" that interrupt to the host.
847 * This saved_p is cleared by the host EOI, when we know
848 * for sure the queue slot is no longer in use.
851 pq = xive_esb_read(xd, XIVE_ESB_SET_PQ_11);
855 * Sync the XIVE source HW to ensure the interrupt
856 * has gone through the EAS before we change its
857 * target to the guest. That should guarantee us
858 * that we *will* eventually get an EOI for it on
859 * the host. Otherwise there would be a small window
860 * for P to be seen here but the interrupt going
861 * to the guest queue.
863 if (xive_ops->sync_source)
864 xive_ops->sync_source(hw_irq);
868 irqd_clr_forwarded_to_vcpu(d);
870 /* No host target ? hard mask and return */
871 if (xd->target == XIVE_INVALID_TARGET) {
872 xive_do_source_set_mask(xd, true);
877 * Sync the XIVE source HW to ensure the interrupt
878 * has gone through the EAS before we change its
879 * target to the host.
881 if (xive_ops->sync_source)
882 xive_ops->sync_source(hw_irq);
885 * By convention we are called with the interrupt in
886 * a PQ=10 or PQ=11 state, ie, it won't fire and will
887 * have latched in Q whether there's a pending HW
890 * First reconfigure the target.
892 rc = xive_ops->configure_irq(hw_irq,
893 get_hard_smp_processor_id(xd->target),
894 xive_irq_priority, d->irq);
899 * Then if saved_p is not set, effectively re-enable the
900 * interrupt with an EOI. If it is set, we know there is
901 * still a message in a host queue somewhere that will be
904 * Note: We don't check irqd_irq_disabled(). Effectively,
905 * we *will* let the irq get through even if masked if the
906 * HW is still firing it in order to deal with the whole
907 * saved_p business properly. If the interrupt triggers
908 * while masked, the generic code will re-mask it anyway.
911 xive_do_source_eoi(hw_irq, xd);
917 static struct irq_chip xive_irq_chip = {
919 .irq_startup = xive_irq_startup,
920 .irq_shutdown = xive_irq_shutdown,
921 .irq_eoi = xive_irq_eoi,
922 .irq_mask = xive_irq_mask,
923 .irq_unmask = xive_irq_unmask,
924 .irq_set_affinity = xive_irq_set_affinity,
925 .irq_set_type = xive_irq_set_type,
926 .irq_retrigger = xive_irq_retrigger,
927 .irq_set_vcpu_affinity = xive_irq_set_vcpu_affinity,
930 bool is_xive_irq(struct irq_chip *chip)
932 return chip == &xive_irq_chip;
934 EXPORT_SYMBOL_GPL(is_xive_irq);
936 void xive_cleanup_irq_data(struct xive_irq_data *xd)
939 iounmap(xd->eoi_mmio);
940 if (xd->eoi_mmio == xd->trig_mmio)
941 xd->trig_mmio = NULL;
945 iounmap(xd->trig_mmio);
946 xd->trig_mmio = NULL;
949 EXPORT_SYMBOL_GPL(xive_cleanup_irq_data);
951 static int xive_irq_alloc_data(unsigned int virq, irq_hw_number_t hw)
953 struct xive_irq_data *xd;
956 xd = kzalloc(sizeof(struct xive_irq_data), GFP_KERNEL);
959 rc = xive_ops->populate_irq_data(hw, xd);
964 xd->target = XIVE_INVALID_TARGET;
965 irq_set_handler_data(virq, xd);
970 static void xive_irq_free_data(unsigned int virq)
972 struct xive_irq_data *xd = irq_get_handler_data(virq);
976 irq_set_handler_data(virq, NULL);
977 xive_cleanup_irq_data(xd);
983 static void xive_cause_ipi(int cpu)
986 struct xive_irq_data *xd;
988 xc = per_cpu(xive_cpu, cpu);
990 DBG_VERBOSE("IPI CPU %d -> %d (HW IRQ 0x%x)\n",
991 smp_processor_id(), cpu, xc->hw_ipi);
994 if (WARN_ON(!xd->trig_mmio))
996 out_be64(xd->trig_mmio, 0);
999 static irqreturn_t xive_muxed_ipi_action(int irq, void *dev_id)
1001 return smp_ipi_demux();
1004 static void xive_ipi_eoi(struct irq_data *d)
1006 struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1008 /* Handle possible race with unplug and drop stale IPIs */
1012 DBG_VERBOSE("IPI eoi: irq=%d [0x%lx] (HW IRQ 0x%x) pending=%02x\n",
1013 d->irq, irqd_to_hwirq(d), xc->hw_ipi, xc->pending_prio);
1015 xive_do_source_eoi(xc->hw_ipi, &xc->ipi_data);
1016 xive_do_queue_eoi(xc);
1019 static void xive_ipi_do_nothing(struct irq_data *d)
1022 * Nothing to do, we never mask/unmask IPIs, but the callback
1023 * has to exist for the struct irq_chip.
1027 static struct irq_chip xive_ipi_chip = {
1029 .irq_eoi = xive_ipi_eoi,
1030 .irq_mask = xive_ipi_do_nothing,
1031 .irq_unmask = xive_ipi_do_nothing,
1034 static void __init xive_request_ipi(void)
1039 * Initialization failed, move on, we might manage to
1040 * reach the point where we display our errors before
1041 * the system falls appart
1043 if (!xive_irq_domain)
1047 virq = irq_create_mapping(xive_irq_domain, 0);
1048 xive_ipi_irq = virq;
1050 WARN_ON(request_irq(virq, xive_muxed_ipi_action,
1051 IRQF_PERCPU | IRQF_NO_THREAD, "IPI", NULL));
1054 static int xive_setup_cpu_ipi(unsigned int cpu)
1056 struct xive_cpu *xc;
1059 pr_debug("Setting up IPI for CPU %d\n", cpu);
1061 xc = per_cpu(xive_cpu, cpu);
1063 /* Check if we are already setup */
1064 if (xc->hw_ipi != 0)
1067 /* Grab an IPI from the backend, this will populate xc->hw_ipi */
1068 if (xive_ops->get_ipi(cpu, xc))
1072 * Populate the IRQ data in the xive_cpu structure and
1073 * configure the HW / enable the IPIs.
1075 rc = xive_ops->populate_irq_data(xc->hw_ipi, &xc->ipi_data);
1077 pr_err("Failed to populate IPI data on CPU %d\n", cpu);
1080 rc = xive_ops->configure_irq(xc->hw_ipi,
1081 get_hard_smp_processor_id(cpu),
1082 xive_irq_priority, xive_ipi_irq);
1084 pr_err("Failed to map IPI CPU %d\n", cpu);
1087 pr_devel("CPU %d HW IPI %x, virq %d, trig_mmio=%p\n", cpu,
1088 xc->hw_ipi, xive_ipi_irq, xc->ipi_data.trig_mmio);
1091 xive_do_source_set_mask(&xc->ipi_data, false);
1096 static void xive_cleanup_cpu_ipi(unsigned int cpu, struct xive_cpu *xc)
1098 /* Disable the IPI and free the IRQ data */
1100 /* Already cleaned up ? */
1101 if (xc->hw_ipi == 0)
1105 xive_do_source_set_mask(&xc->ipi_data, true);
1108 * Note: We don't call xive_cleanup_irq_data() to free
1109 * the mappings as this is called from an IPI on kexec
1110 * which is not a safe environment to call iounmap()
1113 /* Deconfigure/mask in the backend */
1114 xive_ops->configure_irq(xc->hw_ipi, hard_smp_processor_id(),
1115 0xff, xive_ipi_irq);
1117 /* Free the IPIs in the backend */
1118 xive_ops->put_ipi(cpu, xc);
1121 void __init xive_smp_probe(void)
1123 smp_ops->cause_ipi = xive_cause_ipi;
1125 /* Register the IPI */
1128 /* Allocate and setup IPI for the boot CPU */
1129 xive_setup_cpu_ipi(smp_processor_id());
1132 #endif /* CONFIG_SMP */
1134 static int xive_irq_domain_map(struct irq_domain *h, unsigned int virq,
1140 * Mark interrupts as edge sensitive by default so that resend
1141 * actually works. Will fix that up below if needed.
1143 irq_clear_status_flags(virq, IRQ_LEVEL);
1146 /* IPIs are special and come up with HW number 0 */
1149 * IPIs are marked per-cpu. We use separate HW interrupts under
1150 * the hood but associated with the same "linux" interrupt
1152 irq_set_chip_and_handler(virq, &xive_ipi_chip,
1158 rc = xive_irq_alloc_data(virq, hw);
1162 irq_set_chip_and_handler(virq, &xive_irq_chip, handle_fasteoi_irq);
1167 static void xive_irq_domain_unmap(struct irq_domain *d, unsigned int virq)
1169 struct irq_data *data = irq_get_irq_data(virq);
1170 unsigned int hw_irq;
1172 /* XXX Assign BAD number */
1175 hw_irq = (unsigned int)irqd_to_hwirq(data);
1177 xive_irq_free_data(virq);
1180 static int xive_irq_domain_xlate(struct irq_domain *h, struct device_node *ct,
1181 const u32 *intspec, unsigned int intsize,
1182 irq_hw_number_t *out_hwirq, unsigned int *out_flags)
1185 *out_hwirq = intspec[0];
1188 * If intsize is at least 2, we look for the type in the second cell,
1189 * we assume the LSB indicates a level interrupt.
1193 *out_flags = IRQ_TYPE_LEVEL_LOW;
1195 *out_flags = IRQ_TYPE_EDGE_RISING;
1197 *out_flags = IRQ_TYPE_LEVEL_LOW;
1202 static int xive_irq_domain_match(struct irq_domain *h, struct device_node *node,
1203 enum irq_domain_bus_token bus_token)
1205 return xive_ops->match(node);
1208 static const struct irq_domain_ops xive_irq_domain_ops = {
1209 .match = xive_irq_domain_match,
1210 .map = xive_irq_domain_map,
1211 .unmap = xive_irq_domain_unmap,
1212 .xlate = xive_irq_domain_xlate,
1215 static void __init xive_init_host(void)
1217 xive_irq_domain = irq_domain_add_nomap(NULL, XIVE_MAX_IRQ,
1218 &xive_irq_domain_ops, NULL);
1219 if (WARN_ON(xive_irq_domain == NULL))
1221 irq_set_default_host(xive_irq_domain);
1224 static void xive_cleanup_cpu_queues(unsigned int cpu, struct xive_cpu *xc)
1226 if (xc->queue[xive_irq_priority].qpage)
1227 xive_ops->cleanup_queue(cpu, xc, xive_irq_priority);
1230 static int xive_setup_cpu_queues(unsigned int cpu, struct xive_cpu *xc)
1234 /* We setup 1 queues for now with a 64k page */
1235 if (!xc->queue[xive_irq_priority].qpage)
1236 rc = xive_ops->setup_queue(cpu, xc, xive_irq_priority);
1241 static int xive_prepare_cpu(unsigned int cpu)
1243 struct xive_cpu *xc;
1245 xc = per_cpu(xive_cpu, cpu);
1247 struct device_node *np;
1249 xc = kzalloc_node(sizeof(struct xive_cpu),
1250 GFP_KERNEL, cpu_to_node(cpu));
1253 np = of_get_cpu_node(cpu, NULL);
1255 xc->chip_id = of_get_ibm_chip_id(np);
1258 per_cpu(xive_cpu, cpu) = xc;
1261 /* Setup EQs if not already */
1262 return xive_setup_cpu_queues(cpu, xc);
1265 static void xive_setup_cpu(void)
1267 struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1269 /* The backend might have additional things to do */
1270 if (xive_ops->setup_cpu)
1271 xive_ops->setup_cpu(smp_processor_id(), xc);
1273 /* Set CPPR to 0xff to enable flow of interrupts */
1275 out_8(xive_tima + xive_tima_offset + TM_CPPR, 0xff);
1279 void xive_smp_setup_cpu(void)
1281 pr_devel("SMP setup CPU %d\n", smp_processor_id());
1283 /* This will have already been done on the boot CPU */
1284 if (smp_processor_id() != boot_cpuid)
1289 int xive_smp_prepare_cpu(unsigned int cpu)
1293 /* Allocate per-CPU data and queues */
1294 rc = xive_prepare_cpu(cpu);
1298 /* Allocate and setup IPI for the new CPU */
1299 return xive_setup_cpu_ipi(cpu);
1302 #ifdef CONFIG_HOTPLUG_CPU
1303 static void xive_flush_cpu_queue(unsigned int cpu, struct xive_cpu *xc)
1307 /* We assume local irqs are disabled */
1308 WARN_ON(!irqs_disabled());
1310 /* Check what's already in the CPU queue */
1311 while ((irq = xive_scan_interrupts(xc, false)) != 0) {
1313 * We need to re-route that interrupt to its new destination.
1314 * First get and lock the descriptor
1316 struct irq_desc *desc = irq_to_desc(irq);
1317 struct irq_data *d = irq_desc_get_irq_data(desc);
1318 struct xive_irq_data *xd;
1319 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
1322 * Ignore anything that isn't a XIVE irq and ignore
1323 * IPIs, so can just be dropped.
1325 if (d->domain != xive_irq_domain || hw_irq == 0)
1329 * The IRQ should have already been re-routed, it's just a
1330 * stale in the old queue, so re-trigger it in order to make
1331 * it reach is new destination.
1334 pr_info("CPU %d: Got irq %d while offline, re-sending...\n",
1337 raw_spin_lock(&desc->lock);
1338 xd = irq_desc_get_handler_data(desc);
1341 * For LSIs, we EOI, this will cause a resend if it's
1342 * still asserted. Otherwise do an MSI retrigger.
1344 if (xd->flags & XIVE_IRQ_FLAG_LSI)
1345 xive_do_source_eoi(irqd_to_hwirq(d), xd);
1347 xive_irq_retrigger(d);
1349 raw_spin_unlock(&desc->lock);
1353 void xive_smp_disable_cpu(void)
1355 struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1356 unsigned int cpu = smp_processor_id();
1358 /* Migrate interrupts away from the CPU */
1359 irq_migrate_all_off_this_cpu();
1361 /* Set CPPR to 0 to disable flow of interrupts */
1363 out_8(xive_tima + xive_tima_offset + TM_CPPR, 0);
1365 /* Flush everything still in the queue */
1366 xive_flush_cpu_queue(cpu, xc);
1368 /* Re-enable CPPR */
1370 out_8(xive_tima + xive_tima_offset + TM_CPPR, 0xff);
1373 void xive_flush_interrupt(void)
1375 struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1376 unsigned int cpu = smp_processor_id();
1378 /* Called if an interrupt occurs while the CPU is hot unplugged */
1379 xive_flush_cpu_queue(cpu, xc);
1382 #endif /* CONFIG_HOTPLUG_CPU */
1384 #endif /* CONFIG_SMP */
1386 void xive_teardown_cpu(void)
1388 struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1389 unsigned int cpu = smp_processor_id();
1391 /* Set CPPR to 0 to disable flow of interrupts */
1393 out_8(xive_tima + xive_tima_offset + TM_CPPR, 0);
1395 if (xive_ops->teardown_cpu)
1396 xive_ops->teardown_cpu(cpu, xc);
1399 /* Get rid of IPI */
1400 xive_cleanup_cpu_ipi(cpu, xc);
1403 /* Disable and free the queues */
1404 xive_cleanup_cpu_queues(cpu, xc);
1407 void xive_shutdown(void)
1409 xive_ops->shutdown();
1412 bool __init xive_core_init(const struct xive_ops *ops, void __iomem *area, u32 offset,
1416 xive_tima_offset = offset;
1418 xive_irq_priority = max_prio;
1420 ppc_md.get_irq = xive_get_irq;
1421 __xive_enabled = true;
1423 pr_devel("Initializing host..\n");
1426 pr_devel("Initializing boot CPU..\n");
1428 /* Allocate per-CPU data and queues */
1429 xive_prepare_cpu(smp_processor_id());
1431 /* Get ready for interrupts */
1434 pr_info("Interrupt handling initialized with %s backend\n",
1436 pr_info("Using priority %d for all interrupts\n", max_prio);
1441 __be32 *xive_queue_page_alloc(unsigned int cpu, u32 queue_shift)
1443 unsigned int alloc_order;
1447 alloc_order = xive_alloc_order(queue_shift);
1448 pages = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, alloc_order);
1450 return ERR_PTR(-ENOMEM);
1451 qpage = (__be32 *)page_address(pages);
1452 memset(qpage, 0, 1 << queue_shift);
1457 static int __init xive_off(char *arg)
1459 xive_cmdline_disabled = true;
1462 __setup("xive=off", xive_off);