2 * Copyright 2017 Benjamin Herrenschmidt, IBM Corporation.
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License, version 2, as
6 * published by the Free Software Foundation.
9 #define pr_fmt(fmt) "xive-kvm: " fmt
11 #include <linux/kernel.h>
12 #include <linux/kvm_host.h>
13 #include <linux/err.h>
14 #include <linux/gfp.h>
15 #include <linux/spinlock.h>
16 #include <linux/delay.h>
17 #include <linux/percpu.h>
18 #include <linux/cpumask.h>
19 #include <asm/uaccess.h>
20 #include <asm/kvm_book3s.h>
21 #include <asm/kvm_ppc.h>
22 #include <asm/hvcall.h>
25 #include <asm/xive-regs.h>
26 #include <asm/debug.h>
27 #include <asm/debugfs.h>
31 #include <linux/debugfs.h>
32 #include <linux/seq_file.h>
34 #include "book3s_xive.h"
38 * Virtual mode variants of the hcalls for use on radix/radix
39 * with AIL. They require the VCPU's VP to be "pushed"
41 * We still instanciate them here because we use some of the
42 * generated utility functions as well in this file.
44 #define XIVE_RUNTIME_CHECKS
45 #define X_PFX xive_vm_
46 #define X_STATIC static
47 #define X_STAT_PFX stat_vm_
48 #define __x_tima xive_tima
49 #define __x_eoi_page(xd) ((void __iomem *)((xd)->eoi_mmio))
50 #define __x_trig_page(xd) ((void __iomem *)((xd)->trig_mmio))
51 #define __x_readb __raw_readb
52 #define __x_writeb __raw_writeb
53 #define __x_readw __raw_readw
54 #define __x_readq __raw_readq
55 #define __x_writeq __raw_writeq
57 #include "book3s_xive_template.c"
60 * We leave a gap of a couple of interrupts in the queue to
61 * account for the IPI and additional safety guard.
66 * This is a simple trigger for a generic XIVE IRQ. This must
67 * only be called for interrupts that support a trigger page
69 static bool xive_irq_trigger(struct xive_irq_data *xd)
71 /* This should be only for MSIs */
72 if (WARN_ON(xd->flags & XIVE_IRQ_FLAG_LSI))
75 /* Those interrupts should always have a trigger page */
76 if (WARN_ON(!xd->trig_mmio))
79 out_be64(xd->trig_mmio, 0);
84 static irqreturn_t xive_esc_irq(int irq, void *data)
86 struct kvm_vcpu *vcpu = data;
88 /* We use the existing H_PROD mechanism to wake up the target */
89 vcpu->arch.prodded = 1;
92 kvmppc_fast_vcpu_kick(vcpu);
97 static int xive_attach_escalation(struct kvm_vcpu *vcpu, u8 prio)
99 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
100 struct xive_q *q = &xc->queues[prio];
104 /* Already there ? */
105 if (xc->esc_virq[prio])
108 /* Hook up the escalation interrupt */
109 xc->esc_virq[prio] = irq_create_mapping(NULL, q->esc_irq);
110 if (!xc->esc_virq[prio]) {
111 pr_err("Failed to map escalation interrupt for queue %d of VCPU %d\n",
112 prio, xc->server_num);
117 * Future improvement: start with them disabled
118 * and handle DD2 and later scheme of merged escalation
121 name = kasprintf(GFP_KERNEL, "kvm-%d-%d-%d",
122 vcpu->kvm->arch.lpid, xc->server_num, prio);
124 pr_err("Failed to allocate escalation irq name for queue %d of VCPU %d\n",
125 prio, xc->server_num);
129 rc = request_irq(xc->esc_virq[prio], xive_esc_irq,
130 IRQF_NO_THREAD, name, vcpu);
132 pr_err("Failed to request escalation interrupt for queue %d of VCPU %d\n",
133 prio, xc->server_num);
136 xc->esc_virq_names[prio] = name;
139 irq_dispose_mapping(xc->esc_virq[prio]);
140 xc->esc_virq[prio] = 0;
145 static int xive_provision_queue(struct kvm_vcpu *vcpu, u8 prio)
147 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
148 struct kvmppc_xive *xive = xc->xive;
149 struct xive_q *q = &xc->queues[prio];
153 if (WARN_ON(q->qpage))
156 /* Allocate the queue and retrieve infos on current node for now */
157 qpage = (__be32 *)__get_free_pages(GFP_KERNEL, xive->q_page_order);
159 pr_err("Failed to allocate queue %d for VCPU %d\n",
160 prio, xc->server_num);
163 memset(qpage, 0, 1 << xive->q_order);
166 * Reconfigure the queue. This will set q->qpage only once the
167 * queue is fully configured. This is a requirement for prio 0
168 * as we will stop doing EOIs for every IPI as soon as we observe
169 * qpage being non-NULL, and instead will only EOI when we receive
170 * corresponding queue 0 entries
172 rc = xive_native_configure_queue(xc->vp_id, q, prio, qpage,
173 xive->q_order, true);
175 pr_err("Failed to configure queue %d for VCPU %d\n",
176 prio, xc->server_num);
180 /* Called with kvm_lock held */
181 static int xive_check_provisioning(struct kvm *kvm, u8 prio)
183 struct kvmppc_xive *xive = kvm->arch.xive;
184 struct kvm_vcpu *vcpu;
187 lockdep_assert_held(&kvm->lock);
189 /* Already provisioned ? */
190 if (xive->qmap & (1 << prio))
193 pr_devel("Provisioning prio... %d\n", prio);
195 /* Provision each VCPU and enable escalations */
196 kvm_for_each_vcpu(i, vcpu, kvm) {
197 if (!vcpu->arch.xive_vcpu)
199 rc = xive_provision_queue(vcpu, prio);
201 xive_attach_escalation(vcpu, prio);
206 /* Order previous stores and mark it as provisioned */
208 xive->qmap |= (1 << prio);
212 static void xive_inc_q_pending(struct kvm *kvm, u32 server, u8 prio)
214 struct kvm_vcpu *vcpu;
215 struct kvmppc_xive_vcpu *xc;
218 /* Locate target server */
219 vcpu = kvmppc_xive_find_server(kvm, server);
221 pr_warn("%s: Can't find server %d\n", __func__, server);
224 xc = vcpu->arch.xive_vcpu;
228 q = &xc->queues[prio];
229 atomic_inc(&q->pending_count);
232 static int xive_try_pick_queue(struct kvm_vcpu *vcpu, u8 prio)
234 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
243 q = &xc->queues[prio];
244 if (WARN_ON(!q->qpage))
247 /* Calculate max number of interrupts in that queue. */
248 max = (q->msk + 1) - XIVE_Q_GAP;
249 return atomic_add_unless(&q->count, 1, max) ? 0 : -EBUSY;
252 static int xive_select_target(struct kvm *kvm, u32 *server, u8 prio)
254 struct kvm_vcpu *vcpu;
257 /* Locate target server */
258 vcpu = kvmppc_xive_find_server(kvm, *server);
260 pr_devel("Can't find server %d\n", *server);
264 pr_devel("Finding irq target on 0x%x/%d...\n", *server, prio);
267 rc = xive_try_pick_queue(vcpu, prio);
271 pr_devel(" .. failed, looking up candidate...\n");
273 /* Failed, pick another VCPU */
274 kvm_for_each_vcpu(i, vcpu, kvm) {
275 if (!vcpu->arch.xive_vcpu)
277 rc = xive_try_pick_queue(vcpu, prio);
279 *server = vcpu->arch.xive_vcpu->server_num;
280 pr_devel(" found on 0x%x/%d\n", *server, prio);
284 pr_devel(" no available target !\n");
286 /* No available target ! */
290 static u8 xive_lock_and_mask(struct kvmppc_xive *xive,
291 struct kvmppc_xive_src_block *sb,
292 struct kvmppc_xive_irq_state *state)
294 struct xive_irq_data *xd;
300 * Take the lock, set masked, try again if racing
304 arch_spin_lock(&sb->lock);
305 old_prio = state->guest_priority;
306 state->guest_priority = MASKED;
310 state->guest_priority = old_prio;
311 arch_spin_unlock(&sb->lock);
314 /* No change ? Bail */
315 if (old_prio == MASKED)
318 /* Get the right irq */
319 kvmppc_xive_select_irq(state, &hw_num, &xd);
322 * If the interrupt is marked as needing masking via
323 * firmware, we do it here. Firmware masking however
324 * is "lossy", it won't return the old p and q bits
325 * and won't set the interrupt to a state where it will
326 * record queued ones. If this is an issue we should do
327 * lazy masking instead.
329 * For now, we work around this in unmask by forcing
330 * an interrupt whenever we unmask a non-LSI via FW
333 if (xd->flags & OPAL_XIVE_IRQ_MASK_VIA_FW) {
334 xive_native_configure_irq(hw_num,
335 xive->vp_base + state->act_server,
336 MASKED, state->number);
337 /* set old_p so we can track if an H_EOI was done */
339 state->old_q = false;
341 /* Set PQ to 10, return old P and old Q and remember them */
342 val = xive_vm_esb_load(xd, XIVE_ESB_SET_PQ_10);
343 state->old_p = !!(val & 2);
344 state->old_q = !!(val & 1);
347 * Synchronize hardware to sensure the queues are updated
350 xive_native_sync_source(hw_num);
356 static void xive_lock_for_unmask(struct kvmppc_xive_src_block *sb,
357 struct kvmppc_xive_irq_state *state)
360 * Take the lock try again if racing with H_EOI
363 arch_spin_lock(&sb->lock);
366 arch_spin_unlock(&sb->lock);
370 static void xive_finish_unmask(struct kvmppc_xive *xive,
371 struct kvmppc_xive_src_block *sb,
372 struct kvmppc_xive_irq_state *state,
375 struct xive_irq_data *xd;
378 /* If we aren't changing a thing, move on */
379 if (state->guest_priority != MASKED)
382 /* Get the right irq */
383 kvmppc_xive_select_irq(state, &hw_num, &xd);
386 * See command in xive_lock_and_mask() concerning masking
389 if (xd->flags & OPAL_XIVE_IRQ_MASK_VIA_FW) {
390 xive_native_configure_irq(hw_num,
391 xive->vp_base + state->act_server,
392 state->act_priority, state->number);
393 /* If an EOI is needed, do it here */
395 xive_vm_source_eoi(hw_num, xd);
396 /* If this is not an LSI, force a trigger */
397 if (!(xd->flags & OPAL_XIVE_IRQ_LSI))
398 xive_irq_trigger(xd);
402 /* Old Q set, set PQ to 11 */
404 xive_vm_esb_load(xd, XIVE_ESB_SET_PQ_11);
407 * If not old P, then perform an "effective" EOI,
408 * on the source. This will handle the cases where
412 xive_vm_source_eoi(hw_num, xd);
414 /* Synchronize ordering and mark unmasked */
417 state->guest_priority = prio;
421 * Target an interrupt to a given server/prio, this will fallback
422 * to another server if necessary and perform the HW targetting
425 * NOTE: Must be called with the state lock held
427 static int xive_target_interrupt(struct kvm *kvm,
428 struct kvmppc_xive_irq_state *state,
431 struct kvmppc_xive *xive = kvm->arch.xive;
436 * This will return a tentative server and actual
437 * priority. The count for that new target will have
438 * already been incremented.
440 rc = xive_select_target(kvm, &server, prio);
443 * We failed to find a target ? Not much we can do
444 * at least until we support the GIQ.
450 * Increment the old queue pending count if there
451 * was one so that the old queue count gets adjusted later
452 * when observed to be empty.
454 if (state->act_priority != MASKED)
455 xive_inc_q_pending(kvm,
457 state->act_priority);
459 * Update state and HW
461 state->act_priority = prio;
462 state->act_server = server;
464 /* Get the right irq */
465 kvmppc_xive_select_irq(state, &hw_num, NULL);
467 return xive_native_configure_irq(hw_num,
468 xive->vp_base + server,
469 prio, state->number);
473 * Targetting rules: In order to avoid losing track of
474 * pending interrupts accross mask and unmask, which would
475 * allow queue overflows, we implement the following rules:
477 * - Unless it was never enabled (or we run out of capacity)
478 * an interrupt is always targetted at a valid server/queue
479 * pair even when "masked" by the guest. This pair tends to
480 * be the last one used but it can be changed under some
481 * circumstances. That allows us to separate targetting
482 * from masking, we only handle accounting during (re)targetting,
483 * this also allows us to let an interrupt drain into its target
484 * queue after masking, avoiding complex schemes to remove
485 * interrupts out of remote processor queues.
487 * - When masking, we set PQ to 10 and save the previous value
490 * - When unmasking, if saved Q was set, we set PQ to 11
491 * otherwise we leave PQ to the HW state which will be either
492 * 10 if nothing happened or 11 if the interrupt fired while
493 * masked. Effectively we are OR'ing the previous Q into the
496 * Then if saved P is clear, we do an effective EOI (Q->P->Trigger)
497 * which will unmask the interrupt and shoot a new one if Q was
500 * Otherwise (saved P is set) we leave PQ unchanged (so 10 or 11,
501 * effectively meaning an H_EOI from the guest is still expected
502 * for that interrupt).
504 * - If H_EOI occurs while masked, we clear the saved P.
506 * - When changing target, we account on the new target and
507 * increment a separate "pending" counter on the old one.
508 * This pending counter will be used to decrement the old
509 * target's count when its queue has been observed empty.
512 int kvmppc_xive_set_xive(struct kvm *kvm, u32 irq, u32 server,
515 struct kvmppc_xive *xive = kvm->arch.xive;
516 struct kvmppc_xive_src_block *sb;
517 struct kvmppc_xive_irq_state *state;
525 pr_devel("set_xive ! irq 0x%x server 0x%x prio %d\n",
526 irq, server, priority);
528 /* First, check provisioning of queues */
529 if (priority != MASKED)
530 rc = xive_check_provisioning(xive->kvm,
531 xive_prio_from_guest(priority));
533 pr_devel(" provisioning failure %d !\n", rc);
537 sb = kvmppc_xive_find_source(xive, irq, &idx);
540 state = &sb->irq_state[idx];
543 * We first handle masking/unmasking since the locking
544 * might need to be retried due to EOIs, we'll handle
545 * targetting changes later. These functions will return
546 * with the SB lock held.
548 * xive_lock_and_mask() will also set state->guest_priority
549 * but won't otherwise change other fields of the state.
551 * xive_lock_for_unmask will not actually unmask, this will
552 * be done later by xive_finish_unmask() once the targetting
553 * has been done, so we don't try to unmask an interrupt
554 * that hasn't yet been targetted.
556 if (priority == MASKED)
557 xive_lock_and_mask(xive, sb, state);
559 xive_lock_for_unmask(sb, state);
563 * Then we handle targetting.
565 * First calculate a new "actual priority"
567 new_act_prio = state->act_priority;
568 if (priority != MASKED)
569 new_act_prio = xive_prio_from_guest(priority);
571 pr_devel(" new_act_prio=%x act_server=%x act_prio=%x\n",
572 new_act_prio, state->act_server, state->act_priority);
575 * Then check if we actually need to change anything,
577 * The condition for re-targetting the interrupt is that
578 * we have a valid new priority (new_act_prio is not 0xff)
579 * and either the server or the priority changed.
581 * Note: If act_priority was ff and the new priority is
582 * also ff, we don't do anything and leave the interrupt
583 * untargetted. An attempt of doing an int_on on an
584 * untargetted interrupt will fail. If that is a problem
585 * we could initialize interrupts with valid default
588 if (new_act_prio != MASKED &&
589 (state->act_server != server ||
590 state->act_priority != new_act_prio))
591 rc = xive_target_interrupt(kvm, state, server, new_act_prio);
594 * Perform the final unmasking of the interrupt source
597 if (priority != MASKED)
598 xive_finish_unmask(xive, sb, state, priority);
601 * Finally Update saved_priority to match. Only int_on/off
602 * set this field to a different value.
604 state->saved_priority = priority;
606 arch_spin_unlock(&sb->lock);
610 int kvmppc_xive_get_xive(struct kvm *kvm, u32 irq, u32 *server,
613 struct kvmppc_xive *xive = kvm->arch.xive;
614 struct kvmppc_xive_src_block *sb;
615 struct kvmppc_xive_irq_state *state;
621 sb = kvmppc_xive_find_source(xive, irq, &idx);
624 state = &sb->irq_state[idx];
625 arch_spin_lock(&sb->lock);
626 *server = state->guest_server;
627 *priority = state->guest_priority;
628 arch_spin_unlock(&sb->lock);
633 int kvmppc_xive_int_on(struct kvm *kvm, u32 irq)
635 struct kvmppc_xive *xive = kvm->arch.xive;
636 struct kvmppc_xive_src_block *sb;
637 struct kvmppc_xive_irq_state *state;
643 sb = kvmppc_xive_find_source(xive, irq, &idx);
646 state = &sb->irq_state[idx];
648 pr_devel("int_on(irq=0x%x)\n", irq);
651 * Check if interrupt was not targetted
653 if (state->act_priority == MASKED) {
654 pr_devel("int_on on untargetted interrupt\n");
658 /* If saved_priority is 0xff, do nothing */
659 if (state->saved_priority == MASKED)
663 * Lock and unmask it.
665 xive_lock_for_unmask(sb, state);
666 xive_finish_unmask(xive, sb, state, state->saved_priority);
667 arch_spin_unlock(&sb->lock);
672 int kvmppc_xive_int_off(struct kvm *kvm, u32 irq)
674 struct kvmppc_xive *xive = kvm->arch.xive;
675 struct kvmppc_xive_src_block *sb;
676 struct kvmppc_xive_irq_state *state;
682 sb = kvmppc_xive_find_source(xive, irq, &idx);
685 state = &sb->irq_state[idx];
687 pr_devel("int_off(irq=0x%x)\n", irq);
692 state->saved_priority = xive_lock_and_mask(xive, sb, state);
693 arch_spin_unlock(&sb->lock);
698 static bool xive_restore_pending_irq(struct kvmppc_xive *xive, u32 irq)
700 struct kvmppc_xive_src_block *sb;
701 struct kvmppc_xive_irq_state *state;
704 sb = kvmppc_xive_find_source(xive, irq, &idx);
707 state = &sb->irq_state[idx];
712 * Trigger the IPI. This assumes we never restore a pass-through
713 * interrupt which should be safe enough
715 xive_irq_trigger(&state->ipi_data);
720 u64 kvmppc_xive_get_icp(struct kvm_vcpu *vcpu)
722 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
727 /* Return the per-cpu state for state saving/migration */
728 return (u64)xc->cppr << KVM_REG_PPC_ICP_CPPR_SHIFT |
729 (u64)xc->mfrr << KVM_REG_PPC_ICP_MFRR_SHIFT;
732 int kvmppc_xive_set_icp(struct kvm_vcpu *vcpu, u64 icpval)
734 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
735 struct kvmppc_xive *xive = vcpu->kvm->arch.xive;
742 /* Grab individual state fields. We don't use pending_pri */
743 cppr = icpval >> KVM_REG_PPC_ICP_CPPR_SHIFT;
744 xisr = (icpval >> KVM_REG_PPC_ICP_XISR_SHIFT) &
745 KVM_REG_PPC_ICP_XISR_MASK;
746 mfrr = icpval >> KVM_REG_PPC_ICP_MFRR_SHIFT;
748 pr_devel("set_icp vcpu %d cppr=0x%x mfrr=0x%x xisr=0x%x\n",
749 xc->server_num, cppr, mfrr, xisr);
752 * We can't update the state of a "pushed" VCPU, but that
755 if (WARN_ON(vcpu->arch.xive_pushed))
758 /* Update VCPU HW saved state */
759 vcpu->arch.xive_saved_state.cppr = cppr;
760 xc->hw_cppr = xc->cppr = cppr;
763 * Update MFRR state. If it's not 0xff, we mark the VCPU as
764 * having a pending MFRR change, which will re-evaluate the
765 * target. The VCPU will thus potentially get a spurious
766 * interrupt but that's not a big deal.
770 xive_irq_trigger(&xc->vp_ipi_data);
773 * Now saved XIRR is "interesting". It means there's something in
774 * the legacy "1 element" queue... for an IPI we simply ignore it,
775 * as the MFRR restore will handle that. For anything else we need
776 * to force a resend of the source.
777 * However the source may not have been setup yet. If that's the
778 * case, we keep that info and increment a counter in the xive to
779 * tell subsequent xive_set_source() to go look.
781 if (xisr > XICS_IPI && !xive_restore_pending_irq(xive, xisr)) {
782 xc->delayed_irq = xisr;
783 xive->delayed_irqs++;
784 pr_devel(" xisr restore delayed\n");
790 int kvmppc_xive_set_mapped(struct kvm *kvm, unsigned long guest_irq,
791 struct irq_desc *host_desc)
793 struct kvmppc_xive *xive = kvm->arch.xive;
794 struct kvmppc_xive_src_block *sb;
795 struct kvmppc_xive_irq_state *state;
796 struct irq_data *host_data = irq_desc_get_irq_data(host_desc);
797 unsigned int host_irq = irq_desc_get_irq(host_desc);
798 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(host_data);
806 pr_devel("set_mapped girq 0x%lx host HW irq 0x%x...\n",guest_irq, hw_irq);
808 sb = kvmppc_xive_find_source(xive, guest_irq, &idx);
811 state = &sb->irq_state[idx];
814 * Mark the passed-through interrupt as going to a VCPU,
815 * this will prevent further EOIs and similar operations
816 * from the XIVE code. It will also mask the interrupt
817 * to either PQ=10 or 11 state, the latter if the interrupt
818 * is pending. This will allow us to unmask or retrigger it
819 * after routing it to the guest with a simple EOI.
821 * The "state" argument is a "token", all it needs is to be
822 * non-NULL to switch to passed-through or NULL for the
823 * other way around. We may not yet have an actual VCPU
824 * target here and we don't really care.
826 rc = irq_set_vcpu_affinity(host_irq, state);
828 pr_err("Failed to set VCPU affinity for irq %d\n", host_irq);
833 * Mask and read state of IPI. We need to know if its P bit
834 * is set as that means it's potentially already using a
835 * queue entry in the target
837 prio = xive_lock_and_mask(xive, sb, state);
838 pr_devel(" old IPI prio %02x P:%d Q:%d\n", prio,
839 state->old_p, state->old_q);
841 /* Turn the IPI hard off */
842 xive_vm_esb_load(&state->ipi_data, XIVE_ESB_SET_PQ_01);
844 /* Grab info about irq */
845 state->pt_number = hw_irq;
846 state->pt_data = irq_data_get_irq_handler_data(host_data);
849 * Configure the IRQ to match the existing configuration of
850 * the IPI if it was already targetted. Otherwise this will
851 * mask the interrupt in a lossy way (act_priority is 0xff)
852 * which is fine for a never started interrupt.
854 xive_native_configure_irq(hw_irq,
855 xive->vp_base + state->act_server,
856 state->act_priority, state->number);
859 * We do an EOI to enable the interrupt (and retrigger if needed)
860 * if the guest has the interrupt unmasked and the P bit was *not*
861 * set in the IPI. If it was set, we know a slot may still be in
862 * use in the target queue thus we have to wait for a guest
865 if (prio != MASKED && !state->old_p)
866 xive_vm_source_eoi(hw_irq, state->pt_data);
868 /* Clear old_p/old_q as they are no longer relevant */
869 state->old_p = state->old_q = false;
871 /* Restore guest prio (unlocks EOI) */
873 state->guest_priority = prio;
874 arch_spin_unlock(&sb->lock);
878 EXPORT_SYMBOL_GPL(kvmppc_xive_set_mapped);
880 int kvmppc_xive_clr_mapped(struct kvm *kvm, unsigned long guest_irq,
881 struct irq_desc *host_desc)
883 struct kvmppc_xive *xive = kvm->arch.xive;
884 struct kvmppc_xive_src_block *sb;
885 struct kvmppc_xive_irq_state *state;
886 unsigned int host_irq = irq_desc_get_irq(host_desc);
894 pr_devel("clr_mapped girq 0x%lx...\n", guest_irq);
896 sb = kvmppc_xive_find_source(xive, guest_irq, &idx);
899 state = &sb->irq_state[idx];
902 * Mask and read state of IRQ. We need to know if its P bit
903 * is set as that means it's potentially already using a
904 * queue entry in the target
906 prio = xive_lock_and_mask(xive, sb, state);
907 pr_devel(" old IRQ prio %02x P:%d Q:%d\n", prio,
908 state->old_p, state->old_q);
911 * If old_p is set, the interrupt is pending, we switch it to
912 * PQ=11. This will force a resend in the host so the interrupt
913 * isn't lost to whatver host driver may pick it up
916 xive_vm_esb_load(state->pt_data, XIVE_ESB_SET_PQ_11);
918 /* Release the passed-through interrupt to the host */
919 rc = irq_set_vcpu_affinity(host_irq, NULL);
921 pr_err("Failed to clr VCPU affinity for irq %d\n", host_irq);
925 /* Forget about the IRQ */
926 state->pt_number = 0;
927 state->pt_data = NULL;
929 /* Reconfigure the IPI */
930 xive_native_configure_irq(state->ipi_number,
931 xive->vp_base + state->act_server,
932 state->act_priority, state->number);
935 * If old_p is set (we have a queue entry potentially
936 * occupied) or the interrupt is masked, we set the IPI
937 * to PQ=10 state. Otherwise we just re-enable it (PQ=00).
939 if (prio == MASKED || state->old_p)
940 xive_vm_esb_load(&state->ipi_data, XIVE_ESB_SET_PQ_10);
942 xive_vm_esb_load(&state->ipi_data, XIVE_ESB_SET_PQ_00);
944 /* Restore guest prio (unlocks EOI) */
946 state->guest_priority = prio;
947 arch_spin_unlock(&sb->lock);
951 EXPORT_SYMBOL_GPL(kvmppc_xive_clr_mapped);
953 static void kvmppc_xive_disable_vcpu_interrupts(struct kvm_vcpu *vcpu)
955 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
956 struct kvm *kvm = vcpu->kvm;
957 struct kvmppc_xive *xive = kvm->arch.xive;
960 for (i = 0; i <= xive->max_sbid; i++) {
961 struct kvmppc_xive_src_block *sb = xive->src_blocks[i];
965 for (j = 0; j < KVMPPC_XICS_IRQ_PER_ICS; j++) {
966 struct kvmppc_xive_irq_state *state = &sb->irq_state[j];
970 if (state->act_priority == MASKED)
972 if (state->act_server != xc->server_num)
976 arch_spin_lock(&sb->lock);
977 state->act_priority = MASKED;
978 xive_vm_esb_load(&state->ipi_data, XIVE_ESB_SET_PQ_01);
979 xive_native_configure_irq(state->ipi_number, 0, MASKED, 0);
980 if (state->pt_number) {
981 xive_vm_esb_load(state->pt_data, XIVE_ESB_SET_PQ_01);
982 xive_native_configure_irq(state->pt_number, 0, MASKED, 0);
984 arch_spin_unlock(&sb->lock);
989 void kvmppc_xive_cleanup_vcpu(struct kvm_vcpu *vcpu)
991 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
992 struct kvmppc_xive *xive = xc->xive;
995 pr_devel("cleanup_vcpu(cpu=%d)\n", xc->server_num);
997 /* Ensure no interrupt is still routed to that VP */
999 kvmppc_xive_disable_vcpu_interrupts(vcpu);
1001 /* Mask the VP IPI */
1002 xive_vm_esb_load(&xc->vp_ipi_data, XIVE_ESB_SET_PQ_01);
1004 /* Disable the VP */
1005 xive_native_disable_vp(xc->vp_id);
1007 /* Free the queues & associated interrupts */
1008 for (i = 0; i < KVMPPC_XIVE_Q_COUNT; i++) {
1009 struct xive_q *q = &xc->queues[i];
1011 /* Free the escalation irq */
1012 if (xc->esc_virq[i]) {
1013 free_irq(xc->esc_virq[i], vcpu);
1014 irq_dispose_mapping(xc->esc_virq[i]);
1015 kfree(xc->esc_virq_names[i]);
1017 /* Free the queue */
1018 xive_native_disable_queue(xc->vp_id, q, i);
1020 free_pages((unsigned long)q->qpage,
1021 xive->q_page_order);
1028 xive_cleanup_irq_data(&xc->vp_ipi_data);
1029 xive_native_free_irq(xc->vp_ipi);
1035 int kvmppc_xive_connect_vcpu(struct kvm_device *dev,
1036 struct kvm_vcpu *vcpu, u32 cpu)
1038 struct kvmppc_xive *xive = dev->private;
1039 struct kvmppc_xive_vcpu *xc;
1042 pr_devel("connect_vcpu(cpu=%d)\n", cpu);
1044 if (dev->ops != &kvm_xive_ops) {
1045 pr_devel("Wrong ops !\n");
1048 if (xive->kvm != vcpu->kvm)
1050 if (vcpu->arch.irq_type)
1052 if (kvmppc_xive_find_server(vcpu->kvm, cpu)) {
1053 pr_devel("Duplicate !\n");
1056 if (cpu >= KVM_MAX_VCPUS) {
1057 pr_devel("Out of bounds !\n");
1060 xc = kzalloc(sizeof(*xc), GFP_KERNEL);
1064 /* We need to synchronize with queue provisioning */
1065 mutex_lock(&vcpu->kvm->lock);
1066 vcpu->arch.xive_vcpu = xc;
1069 xc->server_num = cpu;
1070 xc->vp_id = xive->vp_base + cpu;
1074 r = xive_native_get_vp_info(xc->vp_id, &xc->vp_cam, &xc->vp_chip_id);
1078 /* Configure VCPU fields for use by assembly push/pull */
1079 vcpu->arch.xive_saved_state.w01 = cpu_to_be64(0xff000000);
1080 vcpu->arch.xive_cam_word = cpu_to_be32(xc->vp_cam | TM_QW1W2_VO);
1083 xc->vp_ipi = xive_native_alloc_irq();
1088 pr_devel(" IPI=0x%x\n", xc->vp_ipi);
1090 r = xive_native_populate_irq_data(xc->vp_ipi, &xc->vp_ipi_data);
1095 * Initialize queues. Initially we set them all for no queueing
1096 * and we enable escalation for queue 0 only which we'll use for
1097 * our mfrr change notifications. If the VCPU is hot-plugged, we
1098 * do handle provisioning however.
1100 for (i = 0; i < KVMPPC_XIVE_Q_COUNT; i++) {
1101 struct xive_q *q = &xc->queues[i];
1103 /* Is queue already enabled ? Provision it */
1104 if (xive->qmap & (1 << i)) {
1105 r = xive_provision_queue(vcpu, i);
1107 xive_attach_escalation(vcpu, i);
1111 r = xive_native_configure_queue(xc->vp_id,
1112 q, i, NULL, 0, true);
1114 pr_err("Failed to configure queue %d for VCPU %d\n",
1121 /* If not done above, attach priority 0 escalation */
1122 r = xive_attach_escalation(vcpu, 0);
1127 r = xive_native_enable_vp(xc->vp_id);
1132 r = xive_native_configure_irq(xc->vp_ipi, xc->vp_id, 0, XICS_IPI);
1134 xive_vm_esb_load(&xc->vp_ipi_data, XIVE_ESB_SET_PQ_00);
1137 mutex_unlock(&vcpu->kvm->lock);
1139 kvmppc_xive_cleanup_vcpu(vcpu);
1143 vcpu->arch.irq_type = KVMPPC_IRQ_XICS;
1148 * Scanning of queues before/after migration save
1150 static void xive_pre_save_set_queued(struct kvmppc_xive *xive, u32 irq)
1152 struct kvmppc_xive_src_block *sb;
1153 struct kvmppc_xive_irq_state *state;
1156 sb = kvmppc_xive_find_source(xive, irq, &idx);
1160 state = &sb->irq_state[idx];
1162 /* Some sanity checking */
1163 if (!state->valid) {
1164 pr_err("invalid irq 0x%x in cpu queue!\n", irq);
1169 * If the interrupt is in a queue it should have P set.
1170 * We warn so that gets reported. A backtrace isn't useful
1171 * so no need to use a WARN_ON.
1173 if (!state->saved_p)
1174 pr_err("Interrupt 0x%x is marked in a queue but P not set !\n", irq);
1177 state->in_queue = true;
1180 static void xive_pre_save_mask_irq(struct kvmppc_xive *xive,
1181 struct kvmppc_xive_src_block *sb,
1184 struct kvmppc_xive_irq_state *state = &sb->irq_state[irq];
1189 /* Mask and save state, this will also sync HW queues */
1190 state->saved_scan_prio = xive_lock_and_mask(xive, sb, state);
1192 /* Transfer P and Q */
1193 state->saved_p = state->old_p;
1194 state->saved_q = state->old_q;
1197 arch_spin_unlock(&sb->lock);
1200 static void xive_pre_save_unmask_irq(struct kvmppc_xive *xive,
1201 struct kvmppc_xive_src_block *sb,
1204 struct kvmppc_xive_irq_state *state = &sb->irq_state[irq];
1210 * Lock / exclude EOI (not technically necessary if the
1211 * guest isn't running concurrently. If this becomes a
1212 * performance issue we can probably remove the lock.
1214 xive_lock_for_unmask(sb, state);
1216 /* Restore mask/prio if it wasn't masked */
1217 if (state->saved_scan_prio != MASKED)
1218 xive_finish_unmask(xive, sb, state, state->saved_scan_prio);
1221 arch_spin_unlock(&sb->lock);
1224 static void xive_pre_save_queue(struct kvmppc_xive *xive, struct xive_q *q)
1227 u32 toggle = q->toggle;
1231 irq = __xive_read_eq(q->qpage, q->msk, &idx, &toggle);
1233 xive_pre_save_set_queued(xive, irq);
1237 static void xive_pre_save_scan(struct kvmppc_xive *xive)
1239 struct kvm_vcpu *vcpu = NULL;
1243 * See comment in xive_get_source() about how this
1244 * work. Collect a stable state for all interrupts
1246 for (i = 0; i <= xive->max_sbid; i++) {
1247 struct kvmppc_xive_src_block *sb = xive->src_blocks[i];
1250 for (j = 0; j < KVMPPC_XICS_IRQ_PER_ICS; j++)
1251 xive_pre_save_mask_irq(xive, sb, j);
1254 /* Then scan the queues and update the "in_queue" flag */
1255 kvm_for_each_vcpu(i, vcpu, xive->kvm) {
1256 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
1259 for (j = 0; j < KVMPPC_XIVE_Q_COUNT; j++) {
1260 if (xc->queues[i].qpage)
1261 xive_pre_save_queue(xive, &xc->queues[i]);
1265 /* Finally restore interrupt states */
1266 for (i = 0; i <= xive->max_sbid; i++) {
1267 struct kvmppc_xive_src_block *sb = xive->src_blocks[i];
1270 for (j = 0; j < KVMPPC_XICS_IRQ_PER_ICS; j++)
1271 xive_pre_save_unmask_irq(xive, sb, j);
1275 static void xive_post_save_scan(struct kvmppc_xive *xive)
1279 /* Clear all the in_queue flags */
1280 for (i = 0; i <= xive->max_sbid; i++) {
1281 struct kvmppc_xive_src_block *sb = xive->src_blocks[i];
1284 for (j = 0; j < KVMPPC_XICS_IRQ_PER_ICS; j++)
1285 sb->irq_state[j].in_queue = false;
1288 /* Next get_source() will do a new scan */
1289 xive->saved_src_count = 0;
1293 * This returns the source configuration and state to user space.
1295 static int xive_get_source(struct kvmppc_xive *xive, long irq, u64 addr)
1297 struct kvmppc_xive_src_block *sb;
1298 struct kvmppc_xive_irq_state *state;
1299 u64 __user *ubufp = (u64 __user *) addr;
1303 sb = kvmppc_xive_find_source(xive, irq, &idx);
1307 state = &sb->irq_state[idx];
1312 pr_devel("get_source(%ld)...\n", irq);
1315 * So to properly save the state into something that looks like a
1316 * XICS migration stream we cannot treat interrupts individually.
1318 * We need, instead, mask them all (& save their previous PQ state)
1319 * to get a stable state in the HW, then sync them to ensure that
1320 * any interrupt that had already fired hits its queue, and finally
1321 * scan all the queues to collect which interrupts are still present
1322 * in the queues, so we can set the "pending" flag on them and
1323 * they can be resent on restore.
1325 * So we do it all when the "first" interrupt gets saved, all the
1326 * state is collected at that point, the rest of xive_get_source()
1327 * will merely collect and convert that state to the expected
1328 * userspace bit mask.
1330 if (xive->saved_src_count == 0)
1331 xive_pre_save_scan(xive);
1332 xive->saved_src_count++;
1334 /* Convert saved state into something compatible with xics */
1335 val = state->guest_server;
1336 prio = state->saved_scan_prio;
1338 if (prio == MASKED) {
1339 val |= KVM_XICS_MASKED;
1340 prio = state->saved_priority;
1342 val |= prio << KVM_XICS_PRIORITY_SHIFT;
1344 val |= KVM_XICS_LEVEL_SENSITIVE;
1346 val |= KVM_XICS_PENDING;
1349 val |= KVM_XICS_PRESENTED;
1352 val |= KVM_XICS_QUEUED;
1355 * We mark it pending (which will attempt a re-delivery)
1356 * if we are in a queue *or* we were masked and had
1357 * Q set which is equivalent to the XICS "masked pending"
1360 if (state->in_queue || (prio == MASKED && state->saved_q))
1361 val |= KVM_XICS_PENDING;
1365 * If that was the last interrupt saved, reset the
1368 if (xive->saved_src_count == xive->src_count)
1369 xive_post_save_scan(xive);
1371 /* Copy the result to userspace */
1372 if (put_user(val, ubufp))
1378 static struct kvmppc_xive_src_block *xive_create_src_block(struct kvmppc_xive *xive,
1381 struct kvm *kvm = xive->kvm;
1382 struct kvmppc_xive_src_block *sb;
1385 bid = irq >> KVMPPC_XICS_ICS_SHIFT;
1387 mutex_lock(&kvm->lock);
1389 /* block already exists - somebody else got here first */
1390 if (xive->src_blocks[bid])
1393 /* Create the ICS */
1394 sb = kzalloc(sizeof(*sb), GFP_KERNEL);
1400 for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
1401 sb->irq_state[i].number = (bid << KVMPPC_XICS_ICS_SHIFT) | i;
1402 sb->irq_state[i].guest_priority = MASKED;
1403 sb->irq_state[i].saved_priority = MASKED;
1404 sb->irq_state[i].act_priority = MASKED;
1407 xive->src_blocks[bid] = sb;
1409 if (bid > xive->max_sbid)
1410 xive->max_sbid = bid;
1413 mutex_unlock(&kvm->lock);
1414 return xive->src_blocks[bid];
1417 static bool xive_check_delayed_irq(struct kvmppc_xive *xive, u32 irq)
1419 struct kvm *kvm = xive->kvm;
1420 struct kvm_vcpu *vcpu = NULL;
1423 kvm_for_each_vcpu(i, vcpu, kvm) {
1424 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
1429 if (xc->delayed_irq == irq) {
1430 xc->delayed_irq = 0;
1431 xive->delayed_irqs--;
1438 static int xive_set_source(struct kvmppc_xive *xive, long irq, u64 addr)
1440 struct kvmppc_xive_src_block *sb;
1441 struct kvmppc_xive_irq_state *state;
1442 u64 __user *ubufp = (u64 __user *) addr;
1445 u8 act_prio, guest_prio;
1449 if (irq < KVMPPC_XICS_FIRST_IRQ || irq >= KVMPPC_XICS_NR_IRQS)
1452 pr_devel("set_source(irq=0x%lx)\n", irq);
1454 /* Find the source */
1455 sb = kvmppc_xive_find_source(xive, irq, &idx);
1457 pr_devel("No source, creating source block...\n");
1458 sb = xive_create_src_block(xive, irq);
1460 pr_devel("Failed to create block...\n");
1464 state = &sb->irq_state[idx];
1466 /* Read user passed data */
1467 if (get_user(val, ubufp)) {
1468 pr_devel("fault getting user info !\n");
1472 server = val & KVM_XICS_DESTINATION_MASK;
1473 guest_prio = val >> KVM_XICS_PRIORITY_SHIFT;
1475 pr_devel(" val=0x016%llx (server=0x%x, guest_prio=%d)\n",
1476 val, server, guest_prio);
1478 * If the source doesn't already have an IPI, allocate
1479 * one and get the corresponding data
1481 if (!state->ipi_number) {
1482 state->ipi_number = xive_native_alloc_irq();
1483 if (state->ipi_number == 0) {
1484 pr_devel("Failed to allocate IPI !\n");
1487 xive_native_populate_irq_data(state->ipi_number, &state->ipi_data);
1488 pr_devel(" src_ipi=0x%x\n", state->ipi_number);
1492 * We use lock_and_mask() to set us in the right masked
1493 * state. We will override that state from the saved state
1494 * further down, but this will handle the cases of interrupts
1495 * that need FW masking. We set the initial guest_priority to
1496 * 0 before calling it to ensure it actually performs the masking.
1498 state->guest_priority = 0;
1499 xive_lock_and_mask(xive, sb, state);
1502 * Now, we select a target if we have one. If we don't we
1503 * leave the interrupt untargetted. It means that an interrupt
1504 * can become "untargetted" accross migration if it was masked
1505 * by set_xive() but there is little we can do about it.
1508 /* First convert prio and mark interrupt as untargetted */
1509 act_prio = xive_prio_from_guest(guest_prio);
1510 state->act_priority = MASKED;
1511 state->guest_server = server;
1514 * We need to drop the lock due to the mutex below. Hopefully
1515 * nothing is touching that interrupt yet since it hasn't been
1516 * advertized to a running guest yet
1518 arch_spin_unlock(&sb->lock);
1520 /* If we have a priority target the interrupt */
1521 if (act_prio != MASKED) {
1522 /* First, check provisioning of queues */
1523 mutex_lock(&xive->kvm->lock);
1524 rc = xive_check_provisioning(xive->kvm, act_prio);
1525 mutex_unlock(&xive->kvm->lock);
1527 /* Target interrupt */
1529 rc = xive_target_interrupt(xive->kvm, state,
1532 * If provisioning or targetting failed, leave it
1533 * alone and masked. It will remain disabled until
1534 * the guest re-targets it.
1539 * Find out if this was a delayed irq stashed in an ICP,
1540 * in which case, treat it as pending
1542 if (xive->delayed_irqs && xive_check_delayed_irq(xive, irq)) {
1543 val |= KVM_XICS_PENDING;
1544 pr_devel(" Found delayed ! forcing PENDING !\n");
1547 /* Cleanup the SW state */
1548 state->old_p = false;
1549 state->old_q = false;
1551 state->asserted = false;
1553 /* Restore LSI state */
1554 if (val & KVM_XICS_LEVEL_SENSITIVE) {
1556 if (val & KVM_XICS_PENDING)
1557 state->asserted = true;
1558 pr_devel(" LSI ! Asserted=%d\n", state->asserted);
1562 * Restore P and Q. If the interrupt was pending, we
1563 * force both P and Q, which will trigger a resend.
1565 * That means that a guest that had both an interrupt
1566 * pending (queued) and Q set will restore with only
1567 * one instance of that interrupt instead of 2, but that
1568 * is perfectly fine as coalescing interrupts that haven't
1569 * been presented yet is always allowed.
1571 if (val & KVM_XICS_PRESENTED || val & KVM_XICS_PENDING)
1572 state->old_p = true;
1573 if (val & KVM_XICS_QUEUED || val & KVM_XICS_PENDING)
1574 state->old_q = true;
1576 pr_devel(" P=%d, Q=%d\n", state->old_p, state->old_q);
1579 * If the interrupt was unmasked, update guest priority and
1580 * perform the appropriate state transition and do a
1581 * re-trigger if necessary.
1583 if (val & KVM_XICS_MASKED) {
1584 pr_devel(" masked, saving prio\n");
1585 state->guest_priority = MASKED;
1586 state->saved_priority = guest_prio;
1588 pr_devel(" unmasked, restoring to prio %d\n", guest_prio);
1589 xive_finish_unmask(xive, sb, state, guest_prio);
1590 state->saved_priority = guest_prio;
1593 /* Increment the number of valid sources and mark this one valid */
1596 state->valid = true;
1601 int kvmppc_xive_set_irq(struct kvm *kvm, int irq_source_id, u32 irq, int level,
1604 struct kvmppc_xive *xive = kvm->arch.xive;
1605 struct kvmppc_xive_src_block *sb;
1606 struct kvmppc_xive_irq_state *state;
1612 sb = kvmppc_xive_find_source(xive, irq, &idx);
1616 /* Perform locklessly .... (we need to do some RCUisms here...) */
1617 state = &sb->irq_state[idx];
1621 /* We don't allow a trigger on a passed-through interrupt */
1622 if (state->pt_number)
1625 if ((level == 1 && state->lsi) || level == KVM_INTERRUPT_SET_LEVEL)
1626 state->asserted = 1;
1627 else if (level == 0 || level == KVM_INTERRUPT_UNSET) {
1628 state->asserted = 0;
1632 /* Trigger the IPI */
1633 xive_irq_trigger(&state->ipi_data);
1638 static int xive_set_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
1640 struct kvmppc_xive *xive = dev->private;
1642 /* We honor the existing XICS ioctl */
1643 switch (attr->group) {
1644 case KVM_DEV_XICS_GRP_SOURCES:
1645 return xive_set_source(xive, attr->attr, attr->addr);
1650 static int xive_get_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
1652 struct kvmppc_xive *xive = dev->private;
1654 /* We honor the existing XICS ioctl */
1655 switch (attr->group) {
1656 case KVM_DEV_XICS_GRP_SOURCES:
1657 return xive_get_source(xive, attr->attr, attr->addr);
1662 static int xive_has_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
1664 /* We honor the same limits as XICS, at least for now */
1665 switch (attr->group) {
1666 case KVM_DEV_XICS_GRP_SOURCES:
1667 if (attr->attr >= KVMPPC_XICS_FIRST_IRQ &&
1668 attr->attr < KVMPPC_XICS_NR_IRQS)
1675 static void kvmppc_xive_cleanup_irq(u32 hw_num, struct xive_irq_data *xd)
1677 xive_vm_esb_load(xd, XIVE_ESB_SET_PQ_01);
1678 xive_native_configure_irq(hw_num, 0, MASKED, 0);
1679 xive_cleanup_irq_data(xd);
1682 static void kvmppc_xive_free_sources(struct kvmppc_xive_src_block *sb)
1686 for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
1687 struct kvmppc_xive_irq_state *state = &sb->irq_state[i];
1692 kvmppc_xive_cleanup_irq(state->ipi_number, &state->ipi_data);
1693 xive_native_free_irq(state->ipi_number);
1695 /* Pass-through, cleanup too */
1696 if (state->pt_number)
1697 kvmppc_xive_cleanup_irq(state->pt_number, state->pt_data);
1699 state->valid = false;
1703 static void kvmppc_xive_free(struct kvm_device *dev)
1705 struct kvmppc_xive *xive = dev->private;
1706 struct kvm *kvm = xive->kvm;
1709 debugfs_remove(xive->dentry);
1712 kvm->arch.xive = NULL;
1714 /* Mask and free interrupts */
1715 for (i = 0; i <= xive->max_sbid; i++) {
1716 if (xive->src_blocks[i])
1717 kvmppc_xive_free_sources(xive->src_blocks[i]);
1718 kfree(xive->src_blocks[i]);
1719 xive->src_blocks[i] = NULL;
1722 if (xive->vp_base != XIVE_INVALID_VP)
1723 xive_native_free_vp_block(xive->vp_base);
1730 static int kvmppc_xive_create(struct kvm_device *dev, u32 type)
1732 struct kvmppc_xive *xive;
1733 struct kvm *kvm = dev->kvm;
1736 pr_devel("Creating xive for partition\n");
1738 xive = kzalloc(sizeof(*xive), GFP_KERNEL);
1742 dev->private = xive;
1746 /* Already there ? */
1750 kvm->arch.xive = xive;
1752 /* We use the default queue size set by the host */
1753 xive->q_order = xive_native_default_eq_shift();
1754 if (xive->q_order < PAGE_SHIFT)
1755 xive->q_page_order = 0;
1757 xive->q_page_order = xive->q_order - PAGE_SHIFT;
1759 /* Allocate a bunch of VPs */
1760 xive->vp_base = xive_native_alloc_vp_block(KVM_MAX_VCPUS);
1761 pr_devel("VP_Base=%x\n", xive->vp_base);
1763 if (xive->vp_base == XIVE_INVALID_VP)
1775 static int xive_debug_show(struct seq_file *m, void *private)
1777 struct kvmppc_xive *xive = m->private;
1778 struct kvm *kvm = xive->kvm;
1779 struct kvm_vcpu *vcpu;
1780 u64 t_rm_h_xirr = 0;
1781 u64 t_rm_h_ipoll = 0;
1782 u64 t_rm_h_cppr = 0;
1785 u64 t_vm_h_xirr = 0;
1786 u64 t_vm_h_ipoll = 0;
1787 u64 t_vm_h_cppr = 0;
1795 seq_printf(m, "=========\nVCPU state\n=========\n");
1797 kvm_for_each_vcpu(i, vcpu, kvm) {
1798 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
1803 seq_printf(m, "cpu server %#x CPPR:%#x HWCPPR:%#x"
1804 " MFRR:%#x PEND:%#x h_xirr: R=%lld V=%lld\n",
1805 xc->server_num, xc->cppr, xc->hw_cppr,
1806 xc->mfrr, xc->pending,
1807 xc->stat_rm_h_xirr, xc->stat_vm_h_xirr);
1809 t_rm_h_xirr += xc->stat_rm_h_xirr;
1810 t_rm_h_ipoll += xc->stat_rm_h_ipoll;
1811 t_rm_h_cppr += xc->stat_rm_h_cppr;
1812 t_rm_h_eoi += xc->stat_rm_h_eoi;
1813 t_rm_h_ipi += xc->stat_rm_h_ipi;
1814 t_vm_h_xirr += xc->stat_vm_h_xirr;
1815 t_vm_h_ipoll += xc->stat_vm_h_ipoll;
1816 t_vm_h_cppr += xc->stat_vm_h_cppr;
1817 t_vm_h_eoi += xc->stat_vm_h_eoi;
1818 t_vm_h_ipi += xc->stat_vm_h_ipi;
1821 seq_printf(m, "Hcalls totals\n");
1822 seq_printf(m, " H_XIRR R=%10lld V=%10lld\n", t_rm_h_xirr, t_vm_h_xirr);
1823 seq_printf(m, " H_IPOLL R=%10lld V=%10lld\n", t_rm_h_ipoll, t_vm_h_ipoll);
1824 seq_printf(m, " H_CPPR R=%10lld V=%10lld\n", t_rm_h_cppr, t_vm_h_cppr);
1825 seq_printf(m, " H_EOI R=%10lld V=%10lld\n", t_rm_h_eoi, t_vm_h_eoi);
1826 seq_printf(m, " H_IPI R=%10lld V=%10lld\n", t_rm_h_ipi, t_vm_h_ipi);
1831 static int xive_debug_open(struct inode *inode, struct file *file)
1833 return single_open(file, xive_debug_show, inode->i_private);
1836 static const struct file_operations xive_debug_fops = {
1837 .open = xive_debug_open,
1839 .llseek = seq_lseek,
1840 .release = single_release,
1843 static void xive_debugfs_init(struct kvmppc_xive *xive)
1847 name = kasprintf(GFP_KERNEL, "kvm-xive-%p", xive);
1849 pr_err("%s: no memory for name\n", __func__);
1853 xive->dentry = debugfs_create_file(name, S_IRUGO, powerpc_debugfs_root,
1854 xive, &xive_debug_fops);
1856 pr_debug("%s: created %s\n", __func__, name);
1860 static void kvmppc_xive_init(struct kvm_device *dev)
1862 struct kvmppc_xive *xive = (struct kvmppc_xive *)dev->private;
1864 /* Register some debug interfaces */
1865 xive_debugfs_init(xive);
1868 struct kvm_device_ops kvm_xive_ops = {
1870 .create = kvmppc_xive_create,
1871 .init = kvmppc_xive_init,
1872 .destroy = kvmppc_xive_free,
1873 .set_attr = xive_set_attr,
1874 .get_attr = xive_get_attr,
1875 .has_attr = xive_has_attr,
1878 void kvmppc_xive_init_module(void)
1880 __xive_vm_h_xirr = xive_vm_h_xirr;
1881 __xive_vm_h_ipoll = xive_vm_h_ipoll;
1882 __xive_vm_h_ipi = xive_vm_h_ipi;
1883 __xive_vm_h_cppr = xive_vm_h_cppr;
1884 __xive_vm_h_eoi = xive_vm_h_eoi;
1887 void kvmppc_xive_exit_module(void)
1889 __xive_vm_h_xirr = NULL;
1890 __xive_vm_h_ipoll = NULL;
1891 __xive_vm_h_ipi = NULL;
1892 __xive_vm_h_cppr = NULL;
1893 __xive_vm_h_eoi = NULL;