[PATCH v4 00/12] KVM: x86/xen: Don't drop event deliveries which race memslot updates

[PATCH v4 00/12] KVM: x86/xen: Don't drop event deliveries which race memslot updates

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

Sashiko analysis and torture testing of the upcoming conversion of the
gfn_to_pfn_cache from rwlock to atomic SRCU¹ highlighted a pre-existing
problem: refreshing a cache can fail while a memslot is briefly marked
KVM_MEMSLOT_INVALID during a memslot update, causing kvm_gpc_refresh() to
return -EFAULT. The Xen event channel and timer delivery paths silently
dropped events when this happened.

The first patch is a standalone fix which the same testing found: a guest
parked in SCHEDOP_poll with a zero timeout halts holding vcpu_run's
kvm->srcu read section — unlike every other halt, which drops it — so the
VMM's next memslot update hangs, unkillably, in synchronize_srcu() until the
guest deigns to be woken. A memslot add which normally completes in
microseconds blocks forever.

For the KVM_MEMSLOT_INVALID case, in some cases we don't really care very
much: the VMM should really be using the KVM_XEN_ATTR_TYPE_SHARED_INFO_HVA
mode which doesn't depend on memslots at all, and more faithfully represents
the Xen semantics where the shared info page exists whether it's mapped into
the guest at any given moment or not. But we can be more consistent, and
more intentional about it. So...

Return -EAGAIN from __kvm_gpc_refresh() to distinguish the transient case
from a genuinely unbacked GFN, and add a retry loop for the common case
callers which can just wait for the memslot update to complete, as well as
dealing with the corner cases which can't:

 - The Xen attribute ioctls wait, so userspace never sees the transient
   state.

 - Timer delivery leaves ->arch.xen.timer_pending set and retries from
   kvm_xen_inject_timer_irqs() on the next vCPU entry.

 - The guest's own EVTCHNOP_send hypercall runs under vcpu_run's kvm->srcu
   read section — the very reader the memslot update is waiting for — so it
   cannot wait. Leave the hypercall uncompleted without advancing RIP, so
   the guest re-executes it. This is how Xen itself restarts preempted
   hypercalls.

 - Events which cannot be delivered because userspace deactivated the
   GFN-mode shared_info are deliberately dropped. Treat the shinfo
   like APIC state when serializing a guest — one of the *last* things
   to go — and better still, use the HVA mode and never have it absent
   at all.

 - The hypercall page MSR may want to update the shared info page if
   latching a change of long_mode. Plumbing a retry like the one added for
   EVTCHNOP_send would be complex overkill, so on -EAGAIN just allow the
   wallclock change to fail (as opposed to #GP as it did before without
   even filling the hypercall page). A guest sets the hypercall page
   *early* in boot, and nothing should be manipulating memslots then
   anyway.

The waiting is done without holding xen_lock, on which a vCPU writing
the hypercall page MSR may block from within its kvm->srcu read
section — the very reader the memslot update has to wait out.

Patches 7-10 are pulled forward from the SRCU series because they stand
alone:

 - Three from Sean², extracting vcpu_info delivery into a separate
   helper and taking the shared_info and vcpu_info pages out of dirty
   tracking. Rebased: his series had first converted the GPC locks to
   read_trylock(), which this one doesn't carry as the rwlocks are all
   going away soon anyway.

 - Requesting the guest TLB flush from record_steal_time() itself, as
   a standalone cleanup in preparation for converting that to GPC.

The last two patches are selftests: a new test racing a guest
EVTCHNOP_send loop against a tight memslot delete/recreate loop, and a
fix to xen_shinfo_test's timer recovery protocol. Tested with both and
the rest of the KVM selftests; an instrumented run showed 10 of the
race test's 4000 sends hitting the update window and re-executing.

Based on kvm/next (d4b7fb647204).

v4: https://git.infradead.org/?p=users/dwmw2/linux.git;a=shortlog;h=refs/heads/gpc-eagain-v4
 - New patch 1: don't hold the vCPU's kvm->srcu read section across
   the SCHEDOP_poll halt. Guest-triggerable unkillable hang of the
   VMM's memslot ioctls; reproduced with a dedicated test, fixed as
   vcpu_block() already does for every other halt.
 - Call kvm_xen_shared_info_init() with xen_lock held again, dropping
   it internally only across the refresh. The wallclock publish is
   single-writer once more, which answers the torn-seqcount question
   on v3 3/11 with code rather than argument.
 - Raise KVM_REQ_UNBLOCK when timer delivery defers on -EAGAIN, so a
   vCPU which takes no further exits cannot sit on the owed event
   (v3 5/11).
 - In the KVM_XEN_VCPU_GET_ATTR timer flush, always restore a timer
   which was still armed; only a fired timer with an out-of-range
   port is not re-fired.
 - Race selftest: assert at least one delivery succeeded, add a
   whole-run watchdog and a delivered-count diagnostic.
 - Comment, commit message and kernel-doc fixes; Fixes tags.

v3: https://git.infradead.org/?p=users/dwmw2/linux.git;a=shortlog;h=refs/heads/gpc-eagain-v3
    https://lore.kernel.org/kvm/20261016145646.579369-1-dwmw2@infradead.org/
 - Take kvm->srcu around each kvm_gpc_check() and the copy through the
   mapping in xen_get_guest_pvclock(), which had none (also found by
   the syzbot CI on v2).
 - Drop v2's patch 3 and move the waiting out from under xen_lock
   instead. The srcu unlock/relock it added to the hypercall page MSR
   write was reachable from nested VM-exit MSR-load lists in contexts
   (KVM_SET_NESTED_STATE and friends) which hold kvm->srcu by a local
   index, corrupting the SRCU counts. With no ioctl ever waiting
   while holding xen_lock, taking it from within a kvm->srcu read
   section is safe again and no srcu games are needed. Activations
   racing a concurrent SET_ATTR in the window are userspace's own
   race; last one wins.
 - The hypercall page MSR write keeps wait=false for its shared_info
   re-init and no longer fails the WRMSR on the transient race (see
   above).
 - Re-arm through kvm_xen_start_timer() in the KVM_XEN_VCPU_GET_ATTR
   timer flush, recomputing the expiry from timer_expires. The
   hrtimer's own expiry field is stale when the timer last fired
   through the direct callback path, which doesn't program it. Don't
   re-arm an undeliverable (out of range) port, where the immediate
   -EINVAL would destroy the timer_expires state just reported.
 - New final patch: give the timer recovery poll in xen_shinfo_test a
   timeout. If the timer event is dropped while the guest is already
   blocked in SCHEDOP_poll, nothing else ever wakes it, so the test's
   documented VMM re-arm protocol could never actually run.
 - Commit message and kernel-doc fixes.

Thanks again to the Sashiko review bot for its findings on v3, both
addressed above. My own local incarnation, once updated from its
three-week-old build, went on to find the SCHEDOP_poll hang and the
armed-timer regression. I take back most of what I said about it.

v2: https://git.infradead.org/?p=users/dwmw2/linux.git;a=shortlog;h=refs/heads/gpc-eagain-v2
    https://lore.kernel.org/kvm/20261014220508.518772-1-dwmw2@infradead.org/
 - Take kvm->srcu around the kvm_gpc_check() loop and the wallclock
   write in kvm_xen_shared_info_init(), which had lost it.
 - New patch 3: don't block on xen_lock within the vCPU's kvm->srcu
   read section (three-way deadlock against a waiting ioctl and a
   memslot update). Document the rule at the lock's definition.
 - On -EINVAL (deactivated shared_info, or an invalid port), clear
   timer_pending but keep timer_expires: a halted vCPU would
   otherwise spin aborting its own block, and the event is still
   reported as pending through KVM_XEN_VCPU_GET_ATTR.
 - Drop the wait argument from kvm_xen_inject_timer_irqs(); neither
   caller may wait.
 - Extend the KVM_GET_DIRTY_LOG note in api.rst to the vcpu_info area.
 - Teach lockdep that the wait functions as a synchronize_srcu(), so
   it reports any lock held across it which a kvm->srcu read section
   can take.
 - Comment fixes; drop a leftover cosmetic hunk.

v1: https://git.infradead.org/?p=users/dwmw2/linux.git;a=shortlog;h=refs/heads/gpc-eagain
    https://lore.kernel.org/kvm/20261013204824.454208-1-dwmw2@infradead.org/

¹ https://git.infradead.org/?p=users/dwmw2/linux.git;a=shortlog;h=refs/heads/gpc-srcu-v2
² https://lore.kernel.org/kvm/20260529165114.748639-1-seanjc@google.com/


David Woodhouse (9):
  KVM: x86/xen: Don't block memslot updates while halted in SCHEDOP_poll
  KVM: pfncache: Return -EAGAIN for a lookup which hits an invalid
    memslot
  KVM: pfncache: Add a 'wait' argument to the refresh and activate
    functions
  KVM: x86/xen: Wait out memslot updates in the Xen attribute ioctls
  KVM: x86/xen: Don't lose the timer event when delivery fails
    transiently
  KVM: x86/xen: Retry event channel delivery which races a memslot
    update
  KVM: x86: Request the guest TLB flush from record_steal_time()
  KVM: selftests: Add Xen evtchn vs memslot-update race test
  KVM: selftests: Fix unreachable timer recovery in xen_shinfo_test

Sean Christopherson (3):
  KVM: x86/xen: Extract delivery of event to vCPU into a separate helper
  KVM: x86/xen: Explicitly tag "shared info" page as never being dirty
    tracked
  KVM: x86/xen: Don't dirty track "vCPU info" page

 Documentation/virt/kvm/api.rst                |   5 +-
 arch/x86/include/asm/kvm_host.h               |   6 +
 arch/x86/kvm/msrs.c                           |   2 +-
 arch/x86/kvm/x86.c                            |  28 +-
 arch/x86/kvm/xen.c                            | 474 ++++++++++++------
 include/linux/kvm_host.h                      |  39 +-
 include/linux/kvm_types.h                     |   1 +
 tools/testing/selftests/kvm/Makefile.kvm      |   1 +
 .../kvm/x86/xen_evtchn_memslot_race_test.c    | 208 ++++++++
 .../selftests/kvm/x86/xen_shinfo_test.c       |  20 +-
 virt/kvm/pfncache.c                           | 142 ++++--
 11 files changed, 723 insertions(+), 203 deletions(-)
 create mode 100644 tools/testing/selftests/kvm/x86/xen_evtchn_memslot_race_test.c


base-commit: d4b7fb647204f0c81dfeae2d1a708e4d858e0c94
-- 
2.43.0

[PATCH v4 01/12] KVM: x86/xen: Don't block memslot updates while halted in SCHEDOP_poll

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

kvm_xen_schedop_poll() halts the vCPU from within vcpu_run's exit
handling, without dropping the vCPU's kvm->srcu read section as
vcpu_block() is careful to do around every other halt. The poll wakes
only on delivery of a polled event channel, its poll timer, a signal,
or KVM_REQ_UNBLOCK — and a memslot update raises none of those. With a
zero timeout and a port which nobody is going to signal, the vCPU
sleeps at the guest's pleasure, holding the read section the whole
time.

A memslot update then hangs, unkillably, in kvm_swap_active_memslots()'s
synchronize_srcu() waiting for that reader to quiesce. Any guest with
SCHEDOP_poll available can thus park its VMM's memslot ioctls more or
less indefinitely.

Drop the srcu read section across the halt. A hypercall is only ever
processed from vcpu_run's exit handling, so the vCPU's own read lock
accounting is always the live section here, and nothing between the
unlock and the relock touches memslot-derived state: the pending-event
check which precedes it takes its own read section, and the timer
cleanup which follows does not need one.

Found by Sashiko review while testing the series which lets Xen
attribute ioctls wait out memslot updates — those waits joined the
same queue, with rather more visibility.

Fixes: 1a65105a5aba ("KVM: x86/xen: handle PV spinlocks slowpath")
Cc: stable@vger.kernel.org
Assisted-by: LLM
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
---
 arch/x86/kvm/xen.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index 6aa00f746cba..2a0477ad0c91 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -1594,7 +1594,15 @@ static bool kvm_xen_schedop_poll(struct kvm_vcpu *vcpu, bool is_64bit,
 			mod_timer(&vcpu->arch.xen.poll_timer,
 				  jiffies + nsecs_to_jiffies(sched_poll.timeout));
 
+		/*
+		 * Drop the vCPU's kvm->srcu read section across the halt, as
+		 * vcpu_block() does: the halt is unbounded (with a zero
+		 * timeout, entirely at the guest's pleasure), and a memslot
+		 * update cannot complete until this read section quiesces.
+		 */
+		kvm_vcpu_srcu_read_unlock(vcpu);
 		kvm_vcpu_halt(vcpu);
+		kvm_vcpu_srcu_read_lock(vcpu);
 
 		if (sched_poll.timeout)
 			timer_delete(&vcpu->arch.xen.poll_timer);
-- 
2.43.0

[PATCH v4 02/12] KVM: pfncache: Return -EAGAIN for a lookup which hits an invalid memslot

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

A memslot being deleted or moved is present in the active array but
flagged KVM_MEMSLOT_INVALID — a transient state which exists only
within a single KVM_SET_USER_MEMORY_REGION call and is announced by
a further generation bump when it resolves. gfn_to_hva_memslot()
collapses that case and a genuinely unbacked gfn into one error, and
the pfncache refresh reported both as -EFAULT.

Distinguish them: return -EAGAIN when the slot exists but is
invalid, mirroring what RET_PF_RETRY does for ordinary guest faults
on an invalid slot. Existing users don't *need* to change; most
will simply resolve again at the next refresh. But the distinct
error code *allows* a caller to retry.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Assisted-by: LLM
---
 virt/kvm/pfncache.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)

diff --git a/virt/kvm/pfncache.c b/virt/kvm/pfncache.c
index 3659686b97c2..17aaebd76db8 100644
--- a/virt/kvm/pfncache.c
+++ b/virt/kvm/pfncache.c
@@ -304,7 +304,22 @@ static int __kvm_gpc_refresh(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned l
 			gpc->uhva = gfn_to_hva_memslot(gpc->memslot, gfn);
 
 			if (kvm_is_error_hva(gpc->uhva)) {
-				ret = -EFAULT;
+				/*
+				 * A slot being deleted or moved is present
+				 * but flagged invalid; the lookup failure is
+				 * transient, resolved when the update
+				 * completes (or reverts) and announced by a
+				 * new memslot generation. Distinguish that
+				 * from a genuinely unbacked gfn so that the
+				 * caller may retry rather than report an
+				 * error, as RET_PF_RETRY does for ordinary
+				 * guest faults.
+				 */
+				if (gpc->memslot &&
+				    (gpc->memslot->flags & KVM_MEMSLOT_INVALID))
+					ret = -EAGAIN;
+				else
+					ret = -EFAULT;
 				goto out;
 			}
 
-- 
2.43.0

[PATCH v4 03/12] KVM: pfncache: Add a 'wait' argument to the refresh and activate functions

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

A caller which does not hold kvm->srcu may now pass wait=true to
kvm_gpc_refresh() or kvm_gpc_activate() to wait out a concurrent
memslot update instead of receiving -EAGAIN, looping with
cond_resched() until the lookup resolves or a fatal signal is
pending. (An HVA-based cache never consults the memslots, so
kvm_gpc_activate_hva() keeps its signature.)

The wait is a deliberate poll: a single memslot update bounds it, and
a cond_resched() loop is far simpler than wiring up a wake from the
update side. If a real workload ever cares, an rcuwait signalled on
memslot-update completion — mirroring mn_memslots_update_rcuwait for
the inverse direction — would be the follow-up.

The refresh now takes its own kvm->srcu read lock for each attempt, so
callers no longer need to hold it. Holding it with wait=true would
deadlock — the wait would block the memslot update from completing —
as would taking any lock held across the wait within a kvm->srcu read
section. The wait therefore functions as a synchronize_srcu() to
lockdep, via srcu_lock_sync(), making both cycles reportable.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Assisted-by: LLM
---
 arch/x86/kvm/msrs.c      |   2 +-
 arch/x86/kvm/x86.c       |   2 +-
 arch/x86/kvm/xen.c       |  24 ++++----
 include/linux/kvm_host.h |  24 ++++++--
 virt/kvm/pfncache.c      | 119 +++++++++++++++++++++++++++++----------
 5 files changed, 121 insertions(+), 50 deletions(-)

diff --git a/arch/x86/kvm/msrs.c b/arch/x86/kvm/msrs.c
index dd3bb04878ca..4c0fbba6e7ae 100644
--- a/arch/x86/kvm/msrs.c
+++ b/arch/x86/kvm/msrs.c
@@ -1280,7 +1280,7 @@ static void kvm_write_system_time(struct kvm_vcpu *vcpu, gpa_t system_time,
 	/* we verify if the enable bit is set... */
 	if (system_time & 1)
 		kvm_gpc_activate(&vcpu->arch.pv_time, system_time & ~1ULL,
-				 sizeof(struct pvclock_vcpu_time_info));
+				 sizeof(struct pvclock_vcpu_time_info), false);
 	else
 		kvm_gpc_deactivate(&vcpu->arch.pv_time);
 
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 79468ddfe473..eadcaeee651a 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -1727,7 +1727,7 @@ static void kvm_setup_guest_pvclock(struct pvclock_vcpu_time_info *ref_hv_clock,
 	while (!kvm_gpc_check(gpc, offset + sizeof(*guest_hv_clock))) {
 		read_unlock_irqrestore(&gpc->lock, flags);
 
-		if (kvm_gpc_refresh(gpc, offset + sizeof(*guest_hv_clock)))
+		if (kvm_gpc_refresh(gpc, offset + sizeof(*guest_hv_clock), false))
 			return;
 
 		read_lock_irqsave(&gpc->lock, flags);
diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index 2a0477ad0c91..ac27b4795d69 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -50,7 +50,7 @@ static int kvm_xen_shared_info_init(struct kvm *kvm)
 	while (!kvm_gpc_check(gpc, PAGE_SIZE)) {
 		read_unlock_irq(&gpc->lock);
 
-		ret = kvm_gpc_refresh(gpc, PAGE_SIZE);
+		ret = kvm_gpc_refresh(gpc, PAGE_SIZE, false);
 		if (ret)
 			goto out;
 
@@ -161,7 +161,7 @@ static int xen_get_guest_pvclock(struct kvm_vcpu *vcpu,
 	while (!kvm_gpc_check(gpc, offset + sizeof(*hv_clock))) {
 		read_unlock_irqrestore(&gpc->lock, flags);
 
-		r = kvm_gpc_refresh(gpc, offset + sizeof(*hv_clock));
+		r = kvm_gpc_refresh(gpc, offset + sizeof(*hv_clock), false);
 		if (r)
 			return r;
 
@@ -435,7 +435,7 @@ static void kvm_xen_update_runstate_guest(struct kvm_vcpu *v, bool atomic)
 		if (atomic)
 			return;
 
-		if (kvm_gpc_refresh(gpc1, user_len1))
+		if (kvm_gpc_refresh(gpc1, user_len1, false))
 			return;
 
 		read_lock_irqsave(&gpc1->lock, flags);
@@ -488,7 +488,7 @@ static void kvm_xen_update_runstate_guest(struct kvm_vcpu *v, bool atomic)
 			 * 64-bit mode, the second GPC won't have been set up.
 			 */
 			if (kvm_gpc_activate(gpc2, gpc1->gpa + user_len1,
-					     user_len2))
+					     user_len2, false))
 				return;
 
 			/*
@@ -669,7 +669,7 @@ void kvm_xen_inject_pending_events(struct kvm_vcpu *v)
 	while (!kvm_gpc_check(gpc, sizeof(struct vcpu_info))) {
 		read_unlock_irqrestore(&gpc->lock, flags);
 
-		if (kvm_gpc_refresh(gpc, sizeof(struct vcpu_info)))
+		if (kvm_gpc_refresh(gpc, sizeof(struct vcpu_info), false))
 			return;
 
 		read_lock_irqsave(&gpc->lock, flags);
@@ -757,7 +757,7 @@ int __kvm_xen_has_interrupt(struct kvm_vcpu *v)
 		if (in_atomic() || !task_is_running(current))
 			return 1;
 
-		if (kvm_gpc_refresh(gpc, sizeof(struct vcpu_info))) {
+		if (kvm_gpc_refresh(gpc, sizeof(struct vcpu_info), false)) {
 			/*
 			 * If this failed, userspace has screwed up the
 			 * vcpu_info mapping. No interrupts for you.
@@ -813,7 +813,7 @@ int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 				r = 0;
 			} else {
 				r = kvm_gpc_activate(&kvm->arch.xen.shinfo_cache,
-						     gfn_to_gpa(gfn), PAGE_SIZE);
+						     gfn_to_gpa(gfn), PAGE_SIZE, false);
 			}
 		} else {
 			void __user * hva = u64_to_user_ptr(data->u.shared_info.hva);
@@ -960,7 +960,7 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 				break;
 
 			r = kvm_gpc_activate(&vcpu->arch.xen.vcpu_info_cache,
-					     data->u.gpa, sizeof(struct vcpu_info));
+					     data->u.gpa, sizeof(struct vcpu_info), false);
 		} else {
 			if (data->u.hva == 0) {
 				kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_info_cache);
@@ -990,7 +990,7 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 
 		r = kvm_gpc_activate(&vcpu->arch.xen.vcpu_time_info_cache,
 				     data->u.gpa,
-				     sizeof(struct pvclock_vcpu_time_info));
+				     sizeof(struct pvclock_vcpu_time_info), false);
 		if (!r)
 			kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 		break;
@@ -1023,7 +1023,7 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 		/* How much fits in the (first) page? */
 		sz1 = PAGE_SIZE - (data->u.gpa & ~PAGE_MASK);
 		r = kvm_gpc_activate(&vcpu->arch.xen.runstate_cache,
-				     data->u.gpa, sz1);
+				     data->u.gpa, sz1, false);
 		if (r)
 			goto deactivate_out;
 
@@ -1034,7 +1034,7 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 			sz2 = sz - sz1;
 			BUG_ON((data->u.gpa + sz1) & ~PAGE_MASK);
 			r = kvm_gpc_activate(&vcpu->arch.xen.runstate2_cache,
-					     data->u.gpa + sz1, sz2);
+					     data->u.gpa + sz1, sz2, false);
 			if (r)
 				goto deactivate_out;
 		}
@@ -2019,7 +2019,7 @@ static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm)
 			break;
 
 		idx = srcu_read_lock(&kvm->srcu);
-		rc = kvm_gpc_refresh(gpc, PAGE_SIZE);
+		rc = kvm_gpc_refresh(gpc, PAGE_SIZE, false);
 		srcu_read_unlock(&kvm->srcu, idx);
 	} while(!rc);
 
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index 3dd04605f2e5..94989a705ee6 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -1474,16 +1474,24 @@ void kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm);
  * @gpc:	   struct gfn_to_pfn_cache object.
  * @gpa:	   guest physical address to map.
  * @len:	   sanity check; the range being access must fit a single page.
+ * @wait:	   wait out a concurrent memslot update instead of returning
+ *		   -EAGAIN. May not be used while this task holds any kvm->srcu
+ *		   read section, and no lock held across the wait may ever be
+ *		   taken within one: the update's synchronize_srcu() would
+ *		   deadlock against the wait. Lockdep checks both.
  *
  * @return:	   0 for success.
  *		   -EINVAL for a mapping which would cross a page boundary.
  *		   -EFAULT for an untranslatable guest physical address.
+ *		   -EAGAIN if the memslot is being updated (only with
+ *		   wait == false, or on a fatal signal while waiting).
  *
  * This primes a gfn_to_pfn_cache and links it into the @gpc->kvm's list for
  * invalidations to be processed.  Callers are required to use kvm_gpc_check()
  * to ensure that the cache is valid before accessing the target page.
  */
-int kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long len);
+int kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long len,
+		     bool wait);
 
 /**
  * kvm_gpc_activate_hva - prepare a cached kernel mapping and HPA for a given HVA.
@@ -1496,10 +1504,13 @@ int kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long len)
  *                -EINVAL for a mapping which would cross a page boundary.
  *                -EFAULT for an untranslatable guest physical address.
  *
- * The semantics of this function are the same as those of kvm_gpc_activate(). It
- * merely bypasses a layer of address translation.
+ * The semantics of this function are the same as those of kvm_gpc_activate(),
+ * except that an HVA-based cache never consults the memslots and so can
+ * neither return -EAGAIN nor wait: it merely bypasses a layer of address
+ * translation.
  */
-int kvm_gpc_activate_hva(struct gfn_to_pfn_cache *gpc, unsigned long hva, unsigned long len);
+int kvm_gpc_activate_hva(struct gfn_to_pfn_cache *gpc, unsigned long hva,
+			 unsigned long len);
 
 /**
  * kvm_gpc_check - check validity of a gfn_to_pfn_cache.
@@ -1524,10 +1535,13 @@ bool kvm_gpc_check(struct gfn_to_pfn_cache *gpc, unsigned long len);
  *
  * @gpc:	   struct gfn_to_pfn_cache object.
  * @len:	   sanity check; the range being access must fit a single page.
+ * @wait:	   as kvm_gpc_activate().
  *
  * @return:	   0 for success.
  *		   -EINVAL for a mapping which would cross a page boundary.
  *		   -EFAULT for an untranslatable guest physical address.
+ *		   -EAGAIN if the memslot is being updated (only with
+ *		   wait == false, or on a fatal signal while waiting).
  *
  * This will attempt to refresh a gfn_to_pfn_cache. Note that a successful
  * return from this function does not mean the page can be immediately
@@ -1535,7 +1549,7 @@ bool kvm_gpc_check(struct gfn_to_pfn_cache *gpc, unsigned long len);
  * still lock and check the cache status, as this function does not return
  * with the lock still held to permit access.
  */
-int kvm_gpc_refresh(struct gfn_to_pfn_cache *gpc, unsigned long len);
+int kvm_gpc_refresh(struct gfn_to_pfn_cache *gpc, unsigned long len, bool wait);
 
 /**
  * kvm_gpc_deactivate - deactivate and unlink a gfn_to_pfn_cache.
diff --git a/virt/kvm/pfncache.c b/virt/kvm/pfncache.c
index 17aaebd76db8..0e700c337f5a 100644
--- a/virt/kvm/pfncache.c
+++ b/virt/kvm/pfncache.c
@@ -378,23 +378,63 @@ static int __kvm_gpc_refresh(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned l
 	return ret;
 }
 
-int kvm_gpc_refresh(struct gfn_to_pfn_cache *gpc, unsigned long len)
+/*
+ * One refresh attempt, under the caller's refresh_lock. Takes a kvm->srcu
+ * read section for the attempt, as the memslot lookups consult the active
+ * memslots.
+ */
+static int gpc_refresh_locked(struct gfn_to_pfn_cache *gpc, gpa_t gpa,
+			      unsigned long uhva)
 {
-	unsigned long uhva;
+	int idx, ret;
 
-	guard(mutex)(&gpc->refresh_lock);
+	idx = srcu_read_lock(&gpc->kvm->srcu);
+	ret = __kvm_gpc_refresh(gpc, gpa, uhva);
+	srcu_read_unlock(&gpc->kvm->srcu, idx);
 
-	if (!kvm_gpc_is_valid_len(gpc->gpa, gpc->uhva, len))
-		return -EINVAL;
+	return ret;
+}
+
+int kvm_gpc_refresh(struct gfn_to_pfn_cache *gpc, unsigned long len, bool wait)
+{
+	unsigned long uhva;
+	int ret;
 
 	/*
-	 * If the GPA is valid then ignore the HVA, as a cache can be GPA-based
-	 * or HVA-based, not both.  For GPA-based caches, the HVA will be
-	 * recomputed during refresh if necessary.
+	 * The wait completes only after a memslot update's
+	 * synchronize_srcu(&kvm->srcu), so it functions as one for lockdep:
+	 * taking any lock held across it within a kvm->srcu read section
+	 * is reported as the deadlock it would be.
 	 */
-	uhva = kvm_is_error_gpa(gpc->gpa) ? gpc->uhva : KVM_HVA_ERR_BAD;
+	if (wait)
+		srcu_lock_sync(&gpc->kvm->srcu.dep_map);
+
+	for (;;) {
+		scoped_guard(mutex, &gpc->refresh_lock) {
+			if (!kvm_gpc_is_valid_len(gpc->gpa, gpc->uhva, len))
+				return -EINVAL;
+
+			/*
+			 * If the GPA is valid then ignore the HVA, as a cache
+			 * can be GPA-based or HVA-based, not both.  For
+			 * GPA-based caches, the HVA will be recomputed during
+			 * refresh if necessary.
+			 */
+			uhva = kvm_is_error_gpa(gpc->gpa) ? gpc->uhva : KVM_HVA_ERR_BAD;
+
+			ret = gpc_refresh_locked(gpc, gpc->gpa, uhva);
+		}
+
+		if (ret != -EAGAIN || !wait || fatal_signal_pending(current))
+			return ret;
 
-	return __kvm_gpc_refresh(gpc, gpc->gpa, uhva);
+		/*
+		 * Wait with refresh_lock dropped: a kvm->srcu reader blocked
+		 * on it must be able to proceed and quiesce, or the memslot
+		 * update being waited for could never complete.
+		 */
+		cond_resched();
+	}
 }
 
 void kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm)
@@ -410,36 +450,51 @@ void kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm)
 }
 
 static int __kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long uhva,
-			      unsigned long len)
+			      unsigned long len, bool wait)
 {
 	struct kvm *kvm = gpc->kvm;
 
 	if (!kvm_gpc_is_valid_len(gpa, uhva, len))
 		return -EINVAL;
 
-	guard(mutex)(&gpc->refresh_lock);
+	/* As in kvm_gpc_refresh(): the wait is a synchronize_srcu() to lockdep. */
+	if (wait)
+		srcu_lock_sync(&kvm->srcu.dep_map);
 
-	if (!gpc->active) {
-		if (KVM_BUG_ON(gpc->valid, kvm))
-			return -EIO;
+	for (;;) {
+		int ret;
 
-		spin_lock(&kvm->gpc_lock);
-		list_add(&gpc->list, &kvm->gpc_list);
-		spin_unlock(&kvm->gpc_lock);
+		scoped_guard(mutex, &gpc->refresh_lock) {
+			if (!gpc->active) {
+				if (KVM_BUG_ON(gpc->valid, kvm))
+					return -EIO;
 
-		/*
-		 * Activate the cache after adding it to the list, a concurrent
-		 * refresh must not establish a mapping until the cache is
-		 * reachable by mmu_notifier events.
-		 */
-		write_lock_irq(&gpc->lock);
-		gpc->active = true;
-		write_unlock_irq(&gpc->lock);
+				spin_lock(&kvm->gpc_lock);
+				list_add(&gpc->list, &kvm->gpc_list);
+				spin_unlock(&kvm->gpc_lock);
+
+				/*
+				 * Activate the cache after adding it to the
+				 * list, a concurrent refresh must not
+				 * establish a mapping until the cache is
+				 * reachable by mmu_notifier events.
+				 */
+				write_lock_irq(&gpc->lock);
+				gpc->active = true;
+				write_unlock_irq(&gpc->lock);
+			}
+			ret = gpc_refresh_locked(gpc, gpa, uhva);
+		}
+
+		if (ret != -EAGAIN || !wait || fatal_signal_pending(current))
+			return ret;
+
+		cond_resched();
 	}
-	return __kvm_gpc_refresh(gpc, gpa, uhva);
 }
 
-int kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long len)
+int kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long len,
+		     bool wait)
 {
 	/*
 	 * Explicitly disallow INVALID_GPA so that the magic value can be used
@@ -448,15 +503,17 @@ int kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long len)
 	if (WARN_ON_ONCE(kvm_is_error_gpa(gpa)))
 		return -EINVAL;
 
-	return __kvm_gpc_activate(gpc, gpa, KVM_HVA_ERR_BAD, len);
+	return __kvm_gpc_activate(gpc, gpa, KVM_HVA_ERR_BAD, len, wait);
 }
 
-int kvm_gpc_activate_hva(struct gfn_to_pfn_cache *gpc, unsigned long uhva, unsigned long len)
+int kvm_gpc_activate_hva(struct gfn_to_pfn_cache *gpc, unsigned long uhva,
+			 unsigned long len)
 {
 	if (!access_ok((void __user *)uhva, len))
 		return -EINVAL;
 
-	return __kvm_gpc_activate(gpc, INVALID_GPA, uhva, len);
+	/* An HVA-based cache never consults the memslots, so never waits. */
+	return __kvm_gpc_activate(gpc, INVALID_GPA, uhva, len, false);
 }
 
 void kvm_gpc_deactivate(struct gfn_to_pfn_cache *gpc)
-- 
2.43.0

[PATCH v4 04/12] KVM: x86/xen: Wait out memslot updates in the Xen attribute ioctls

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

Activating a shared_info, vcpu_info, pvclock or runstate cache can
fail if it races a memslot update: previously with -EFAULT, now with
-EAGAIN. These ioctls may sleep, so pass wait and let the activation
ride the update out; userspace never sees the transient state.

No waiting happens while xen_lock is held. A vCPU writing the
hypercall page MSR blocks on xen_lock from within its kvm->srcu read
section, and the memslot update cannot complete until that section
quiesces, so waiting while holding the lock would deadlock all three.
kvm_xen_shared_info_init() is called with the lock held — keeping the
wallclock publish single-writer — and drops it internally across any
refresh; the publish location is chosen by re-reading long_mode under
the re-taken lock. The activation calls wait with the lock dropped at
the call site. Callers get no state continuity across those waits;
concurrent SET_ATTR calls racing through the window are userspace's
own race, and the last one wins. Document the rule at the lock's
definition.

The hypercall page MSR write cannot wait at all — it would be waiting
out its own srcu read section — so its shared_info re-init passes
wait=false and ignores -EAGAIN, skipping the wallclock update and
leaving the new mode's wallclock stale — or, on a first bitness
change, never written. Nothing retries that, but the guest latches
its bitness via the hypercall page early in boot, when nothing should
be changing memslots anyway. A real failure (-EFAULT) still fails the
WRMSR as it always did.

kvm_xen_shared_info_init() gains the wait flag for its refresh.
It and xen_get_guest_pvclock() keep a kvm->srcu read section around
each kvm_gpc_check() and the access through the mapping — the check
consults the memslots — but drop it across the refresh.

The kvm->srcu read sections around the activation calls go away;
they existed to satisfy the old refresh contract, and holding one
across a waiting activation would deadlock against the memslot
update. The runstate updates in kvm_xen_vcpu_set_attr() still read
guest state through the caches, so they keep narrow read sections
of their own.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Assisted-by: LLM
---
 arch/x86/include/asm/kvm_host.h |   6 ++
 arch/x86/kvm/xen.c              | 129 +++++++++++++++++++++++---------
 2 files changed, 100 insertions(+), 35 deletions(-)

diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 3af7395c2430..ed633a39a97a 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1108,6 +1108,12 @@ struct kvm_hv {
 #ifdef CONFIG_KVM_XEN
 /* Xen emulation context */
 struct kvm_xen {
+	/*
+	 * Taken from within kvm->srcu read sections (e.g. by a vCPU
+	 * writing the hypercall page MSR), so never held while waiting
+	 * out a memslot update: the update waits for those read
+	 * sections to quiesce.
+	 */
 	struct mutex xen_lock;
 	u32 xen_version;
 	bool long_mode;
diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index ac27b4795d69..8c49ec8aa4d9 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -36,7 +36,7 @@ static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r);
 
 DEFINE_STATIC_KEY_DEFERRED_FALSE(kvm_xen_enabled, HZ);
 
-static int kvm_xen_shared_info_init(struct kvm *kvm)
+static int kvm_xen_shared_info_init(struct kvm *kvm, bool wait)
 {
 	struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
 	struct pvclock_wall_clock *wc;
@@ -44,17 +44,35 @@ static int kvm_xen_shared_info_init(struct kvm *kvm)
 	u32 wc_version;
 	u64 wall_nsec;
 	int ret = 0;
-	int idx = srcu_read_lock(&kvm->srcu);
+	int idx;
 
-	read_lock_irq(&gpc->lock);
-	while (!kvm_gpc_check(gpc, PAGE_SIZE)) {
+	/* We only really use xen_lock here to serialize wallclock updates. */
+	lockdep_assert_held(&kvm->arch.xen.xen_lock);
+
+	/*
+	 * kvm_gpc_check() consults the memslots, so take kvm->srcu
+	 * around each check and the write through the mapping — but not
+	 * across the refresh.
+	 */
+
+	for (;;) {
+		idx = srcu_read_lock(&kvm->srcu);
+		read_lock_irq(&gpc->lock);
+		if (kvm_gpc_check(gpc, PAGE_SIZE))
+			break;
 		read_unlock_irq(&gpc->lock);
+		srcu_read_unlock(&kvm->srcu, idx);
 
-		ret = kvm_gpc_refresh(gpc, PAGE_SIZE, false);
+		/*
+		 * Drop xen_lock across the refresh, which may wait out a
+		 * memslot update that cannot complete while a vCPU is
+		 * blocked on xen_lock within its kvm->srcu read section.
+		 */
+		mutex_unlock(&kvm->arch.xen.xen_lock);
+		ret = kvm_gpc_refresh(gpc, PAGE_SIZE, wait);
+		mutex_lock(&kvm->arch.xen.xen_lock);
 		if (ret)
 			goto out;
-
-		read_lock_irq(&gpc->lock);
 	}
 
 	/*
@@ -98,9 +116,9 @@ static int kvm_xen_shared_info_init(struct kvm *kvm)
 
 	wc->version = wc_version + 1;
 	read_unlock_irq(&gpc->lock);
+	srcu_read_unlock(&kvm->srcu, idx);
 
 out:
-	srcu_read_unlock(&kvm->srcu, idx);
 	return ret;
 }
 
@@ -155,21 +173,29 @@ static int xen_get_guest_pvclock(struct kvm_vcpu *vcpu,
 				 unsigned int offset)
 {
 	unsigned long flags;
-	int r;
+	int idx, r;
 
-	read_lock_irqsave(&gpc->lock, flags);
-	while (!kvm_gpc_check(gpc, offset + sizeof(*hv_clock))) {
+	/*
+	 * As in kvm_xen_shared_info_init(): kvm_gpc_check() consults the
+	 * memslots, so take kvm->srcu around each check and the copy, but
+	 * not across the refresh.
+	 */
+	for (;;) {
+		idx = srcu_read_lock(&vcpu->kvm->srcu);
+		read_lock_irqsave(&gpc->lock, flags);
+		if (kvm_gpc_check(gpc, offset + sizeof(*hv_clock)))
+			break;
 		read_unlock_irqrestore(&gpc->lock, flags);
+		srcu_read_unlock(&vcpu->kvm->srcu, idx);
 
 		r = kvm_gpc_refresh(gpc, offset + sizeof(*hv_clock), false);
 		if (r)
 			return r;
-
-		read_lock_irqsave(&gpc->lock, flags);
 	}
 
 	memcpy(hv_clock, gpc->khva + offset, sizeof(*hv_clock));
 	read_unlock_irqrestore(&gpc->lock, flags);
+	srcu_read_unlock(&vcpu->kvm->srcu, idx);
 
 	/*
 	 * Sanity check TSC shift+multiplier to verify the guest's view of time
@@ -776,7 +802,6 @@ int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 {
 	int r = -ENOENT;
 
-
 	switch (data->type) {
 	case KVM_XEN_ATTR_TYPE_LONG_MODE:
 		if (!IS_ENABLED(CONFIG_64BIT) && data->u.long_mode) {
@@ -792,19 +817,23 @@ int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 			 * to make the call anyway.
 			 */
 			r = kvm->arch.xen.shinfo_cache.active ?
-				kvm_xen_shared_info_init(kvm) : 0;
+				kvm_xen_shared_info_init(kvm, true) : 0;
 			mutex_unlock(&kvm->arch.xen.xen_lock);
 		}
 		break;
 
 	case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 	case KVM_XEN_ATTR_TYPE_SHARED_INFO_HVA: {
-		int idx;
-
+		/*
+		 * Deactivation stays under xen_lock, but an activation may
+		 * wait out a memslot update and so must not hold it: a vCPU
+		 * may block on xen_lock from within the kvm->srcu read
+		 * section which that update must wait out. Concurrent
+		 * SET_ATTR calls racing in the window are userspace's own
+		 * race; the cache layer serializes them, last one wins.
+		 */
 		mutex_lock(&kvm->arch.xen.xen_lock);
 
-		idx = srcu_read_lock(&kvm->srcu);
-
 		if (data->type == KVM_XEN_ATTR_TYPE_SHARED_INFO) {
 			gfn_t gfn = data->u.shared_info.gfn;
 
@@ -812,8 +841,10 @@ int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 				kvm_gpc_deactivate(&kvm->arch.xen.shinfo_cache);
 				r = 0;
 			} else {
+				mutex_unlock(&kvm->arch.xen.xen_lock);
 				r = kvm_gpc_activate(&kvm->arch.xen.shinfo_cache,
-						     gfn_to_gpa(gfn), PAGE_SIZE, false);
+						     gfn_to_gpa(gfn), PAGE_SIZE, true);
+				mutex_lock(&kvm->arch.xen.xen_lock);
 			}
 		} else {
 			void __user * hva = u64_to_user_ptr(data->u.shared_info.hva);
@@ -829,10 +860,8 @@ int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 			}
 		}
 
-		srcu_read_unlock(&kvm->srcu, idx);
-
 		if (!r && kvm->arch.xen.shinfo_cache.active)
-			r = kvm_xen_shared_info_init(kvm);
+			r = kvm_xen_shared_info_init(kvm, true);
 
 		mutex_unlock(&kvm->arch.xen.xen_lock);
 		break;
@@ -937,7 +966,6 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 	int idx, r = -ENOENT;
 
 	mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
-	idx = srcu_read_lock(&vcpu->kvm->srcu);
 
 	switch (data->type) {
 	case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
@@ -959,8 +987,16 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 			if (!IS_ALIGNED(data->u.gpa, sizeof(u32)))
 				break;
 
+			/*
+			 * As with the shared_info attributes: a waiting
+			 * activation must not hold xen_lock, on which a
+			 * vCPU may block from within the kvm->srcu read
+			 * section the memslot update has to wait out.
+			 */
+			mutex_unlock(&vcpu->kvm->arch.xen.xen_lock);
 			r = kvm_gpc_activate(&vcpu->arch.xen.vcpu_info_cache,
-					     data->u.gpa, sizeof(struct vcpu_info), false);
+					     data->u.gpa, sizeof(struct vcpu_info), true);
+			mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
 		} else {
 			if (data->u.hva == 0) {
 				kvm_gpc_deactivate(&vcpu->arch.xen.vcpu_info_cache);
@@ -988,9 +1024,11 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 			break;
 		}
 
+		mutex_unlock(&vcpu->kvm->arch.xen.xen_lock);
 		r = kvm_gpc_activate(&vcpu->arch.xen.vcpu_time_info_cache,
 				     data->u.gpa,
-				     sizeof(struct pvclock_vcpu_time_info), false);
+				     sizeof(struct pvclock_vcpu_time_info), true);
+		mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
 		if (!r)
 			kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 		break;
@@ -1022,10 +1060,14 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 
 		/* How much fits in the (first) page? */
 		sz1 = PAGE_SIZE - (data->u.gpa & ~PAGE_MASK);
+
+		mutex_unlock(&vcpu->kvm->arch.xen.xen_lock);
 		r = kvm_gpc_activate(&vcpu->arch.xen.runstate_cache,
-				     data->u.gpa, sz1, false);
-		if (r)
+				     data->u.gpa, sz1, true);
+		if (r) {
+			mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
 			goto deactivate_out;
+		}
 
 		/* Either map the second page, or deactivate the second GPC */
 		if (sz1 >= sz) {
@@ -1034,12 +1076,17 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 			sz2 = sz - sz1;
 			BUG_ON((data->u.gpa + sz1) & ~PAGE_MASK);
 			r = kvm_gpc_activate(&vcpu->arch.xen.runstate2_cache,
-					     data->u.gpa + sz1, sz2, false);
-			if (r)
+					     data->u.gpa + sz1, sz2, true);
+			if (r) {
+				mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
 				goto deactivate_out;
+			}
 		}
+		mutex_lock(&vcpu->kvm->arch.xen.xen_lock);
 
+		idx = srcu_read_lock(&vcpu->kvm->srcu);
 		kvm_xen_update_runstate_guest(vcpu, false);
+		srcu_read_unlock(&vcpu->kvm->srcu, idx);
 		break;
 	}
 	case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
@@ -1052,7 +1099,9 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 			break;
 		}
 
+		idx = srcu_read_lock(&vcpu->kvm->srcu);
 		kvm_xen_update_runstate(vcpu, data->u.runstate.state);
+		srcu_read_unlock(&vcpu->kvm->srcu, idx);
 		r = 0;
 		break;
 
@@ -1132,10 +1181,12 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 		vcpu->arch.xen.runstate_times[RUNSTATE_offline] +=
 			data->u.runstate.time_offline;
 
+		idx = srcu_read_lock(&vcpu->kvm->srcu);
 		if (data->u.runstate.state <= RUNSTATE_offline)
 			kvm_xen_update_runstate(vcpu, data->u.runstate.state);
 		else if (vcpu->arch.xen.runstate_cache.active)
 			kvm_xen_update_runstate_guest(vcpu, false);
+		srcu_read_unlock(&vcpu->kvm->srcu, idx);
 		r = 0;
 		break;
 
@@ -1181,7 +1232,6 @@ int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 		break;
 	}
 
-	srcu_read_unlock(&vcpu->kvm->srcu, idx);
 	mutex_unlock(&vcpu->kvm->arch.xen.xen_lock);
 	return r;
 }
@@ -1326,11 +1376,20 @@ int kvm_xen_write_hypercall_page(struct kvm_vcpu *vcpu, u64 data)
 
 		/*
 		 * Re-initialize shared_info to put the wallclock in the
-		 * correct place.
+		 * correct place. This cannot wait out a memslot update
+		 * from within the vCPU's kvm->srcu read section, so if
+		 * one races (-EAGAIN), the wallclock update is simply
+		 * skipped — leaving the new mode's wallclock stale, or
+		 * never yet written. The guest latches its bitness here
+		 * early in boot, when nothing should be changing memslots
+		 * anyway. A real failure still fails the WRMSR as before.
 		 */
-		if (kvm->arch.xen.shinfo_cache.active &&
-		    kvm_xen_shared_info_init(kvm))
-			r = 1;
+		if (kvm->arch.xen.shinfo_cache.active) {
+			int ret = kvm_xen_shared_info_init(kvm, false);
+
+			if (ret && ret != -EAGAIN)
+				r = 1;
+		}
 	}
 	mutex_unlock(&kvm->arch.xen.xen_lock);
 
-- 
2.43.0

[PATCH v4 05/12] KVM: x86/xen: Don't lose the timer event when delivery fails transiently

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

kvm_xen_inject_timer_irqs() can drop events if kvm_xen_set_evtchn()
fails transiently, as it can when a memslot is in the process of
being updated. Defer the delivery in that case by leaving
->arch.xen.timer_pending set, so that it gets retried.

Raise KVM_REQ_UNBLOCK for the retry, as the timer callback does when
it defers: without it, nothing aborts the guest entry, and a vCPU
which takes no further exits would hold the owed event indefinitely.
The request survives to kvm_vcpu_exit_request(), bouncing the vCPU
back through vcpu_run()'s injection pass until the memslot update
completes.

If the shared_info cache has been deactivated (-EINVAL), clear
timer_pending — a halted vCPU would otherwise spin forever, as
kvm_cpu_has_pending_timer() aborts every block while nothing can
deliver the event — but keep timer_expires, so the GET_ATTR/SET_ATTR
migration flow still carries the event and re-fires it on re-arm.
If delivery fails with -EFAULT, the shared_info gfn is genuinely
unbacked and can never succeed: drop the event and complete the
timer rather than spin on it.

True Xen does not have this problem: the shared_info page is owned
by Xen and always exists, even when not mapped into the guest. The
equivalent for a KVM/Xen VMM is KVM_XEN_ATTR_TYPE_SHARED_INFO_HVA,
which has no memslot dependency and is the recommended mode; this
fixes the legacy GFN mode, which xen_shinfo_test exercises.

Don't re-arm the hrtimer in the KVM_XEN_VCPU_GET_ATTR(TIMER) flush
while a delivery is still pending; a re-fired callback racing the
retry would deliver twice. Re-arm through kvm_xen_start_timer(),
which recomputes the expiry from timer_expires: the hrtimer's own
expiry field is stale when the timer last fired through the direct
callback path, which doesn't program the hrtimer. A timer still
armed across the flush is always restored; one which already fired is
not re-fired if its port is out of range, where the immediate -EINVAL
would destroy the owed timer_expires state just reported to
userspace.

The drop is vanishingly rare upstream, but hangs roughly half of all
xen_shinfo_test runs once a subsequent atomic SRCU conversion of the
pfncache widens the window; hence fixing it first.

Fixes: 536395260582 ("KVM: x86/xen: handle PV timers oneshot mode")
Assisted-by: LLM
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
---
 arch/x86/kvm/xen.c                            | 108 ++++++++++++++----
 .../selftests/kvm/x86/xen_shinfo_test.c       |  11 +-
 2 files changed, 89 insertions(+), 30 deletions(-)

diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index 8c49ec8aa4d9..8a9d6963fc05 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -36,6 +36,19 @@ static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r);
 
 DEFINE_STATIC_KEY_DEFERRED_FALSE(kvm_xen_enabled, HZ);
 
+static inline int max_evtchn_port(bool has_64bit_shinfo)
+{
+	if (has_64bit_shinfo)
+		return EVTCHN_2L_NR_CHANNELS;
+	else
+		return COMPAT_EVTCHN_2L_NR_CHANNELS;
+}
+
+static inline int kvm_max_evtchn_port(struct kvm *kvm)
+{
+	return max_evtchn_port(kvm_xen_has_64bit_shinfo(kvm));
+}
+
 static int kvm_xen_shared_info_init(struct kvm *kvm, bool wait)
 {
 	struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
@@ -126,14 +139,47 @@ void kvm_xen_inject_timer_irqs(struct kvm_vcpu *vcpu)
 {
 	if (atomic_read(&vcpu->arch.xen.timer_pending) > 0) {
 		struct kvm_xen_evtchn e;
+		int rc;
 
 		e.vcpu_id = vcpu->vcpu_id;
 		e.vcpu_idx = vcpu->vcpu_idx;
 		e.port = vcpu->arch.xen.timer_virq;
 		e.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
 
-		kvm_xen_set_evtchn(&e, vcpu->kvm);
+		rc = kvm_xen_set_evtchn(&e, vcpu->kvm);
+
+		/*
+		 * Leave timer_pending set if the memslot is being updated
+		 * (-EAGAIN), and raise KVM_REQ_UNBLOCK — as the timer
+		 * callback does when it defers — so that the entry is
+		 * aborted and this retried from vcpu_run(), rather than
+		 * the guest running with the owed event undelivered until
+		 * some unrelated VM exit.
+		 */
+		if (rc == -EAGAIN) {
+			kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
+			return;
+		}
+
+		/*
+		 * If the event cannot currently be delivered (-EINVAL:
+		 * deactivated shared_info, or an out-of-range port), clear
+		 * timer_pending — a halted vCPU would otherwise spin on
+		 * kvm_cpu_has_pending_timer() — but keep timer_expires,
+		 * which still reports the event as owed through
+		 * KVM_XEN_VCPU_GET_ATTR; re-arming the timer with an expiry
+		 * in the past fires it again.
+		 */
+		if (rc == -EINVAL) {
+			atomic_set(&vcpu->arch.xen.timer_pending, 0);
+			return;
+		}
 
+		/*
+		 * Delivered - or failed in a way which can never succeed
+		 * (-EFAULT: a genuinely unbacked shared_info gfn). Either
+		 * way the event is complete: drop it rather than spin.
+		 */
 		vcpu->arch.xen.timer_expires = 0;
 		atomic_set(&vcpu->arch.xen.timer_pending, 0);
 	}
@@ -1315,17 +1361,21 @@ int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 		r = 0;
 		break;
 
-	case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
+	case KVM_XEN_VCPU_ATTR_TYPE_TIMER: {
+		bool timer_armed = false;
+
 		/*
-		 * Ensure a consistent snapshot of state is captured, with a
-		 * timer either being pending, or the event channel delivered
-		 * to the corresponding bit in the shared_info. Not still
-		 * lurking in the timer_pending flag for deferred delivery.
-		 * Purely as an optimisation, if the timer_expires field is
-		 * zero, that means the timer isn't active (or even in the
-		 * timer_pending flag) and there is no need to cancel it.
+		 * Ensure a consistent snapshot of state is captured: a timer
+		 * which already fired has either been delivered to the
+		 * corresponding bit in the shared_info, or is reported as
+		 * still owed (expires_ns != 0, if the shared_info page was
+		 * unavailable). Purely as an optimisation, if the
+		 * timer_expires field is zero, that means the timer isn't
+		 * active (or even in the timer_pending flag) and there is
+		 * no need to cancel it.
 		 */
 		if (vcpu->arch.xen.timer_expires) {
+			timer_armed = hrtimer_active(&vcpu->arch.xen.timer);
 			hrtimer_cancel(&vcpu->arch.xen.timer);
 			kvm_xen_inject_timer_irqs(vcpu);
 		}
@@ -1341,13 +1391,34 @@ int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 		 * That's fine, as the guest won't even have had a chance
 		 * to run and handle the interrupt. Asserting an already
 		 * pending event channel is idempotent.
+		 *
+		 * Don't re-arm if the flush above failed transiently and
+		 * left the timer pending: delivery is owned by the retry
+		 * on the guest entry path, and an immediate hrtimer fire
+		 * would race it — the callback's timer_pending suppression
+		 * check is not serialized against the entry path, so a
+		 * late callback could re-assert a port the entry path had
+		 * already delivered and the guest already consumed.
+		 *
+		 * An expiry already in the past fires immediately: the
+		 * event is still owed.
+		 *
+		 * A timer still armed (not yet fired) across the flush is
+		 * always restored, whatever its port. One which already
+		 * fired is not re-fired if its port is out of range, where
+		 * the immediate -EINVAL from the delivery path would destroy
+		 * the timer_expires state just reported to userspace.
 		 */
-		if (vcpu->arch.xen.timer_expires)
-			hrtimer_start_expires(&vcpu->arch.xen.timer,
-					      HRTIMER_MODE_ABS_HARD);
+		if (vcpu->arch.xen.timer_expires &&
+		    !atomic_read(&vcpu->arch.xen.timer_pending) &&
+		    (timer_armed ||
+		     vcpu->arch.xen.timer_virq < kvm_max_evtchn_port(vcpu->kvm)))
+			kvm_xen_start_timer(vcpu, vcpu->arch.xen.timer_expires,
+					    false);
 
 		r = 0;
 		break;
+	}
 
 	case KVM_XEN_VCPU_ATTR_TYPE_UPCALL_VECTOR:
 		data->u.vector = vcpu->arch.xen.upcall_vector;
@@ -1520,19 +1591,6 @@ static int kvm_xen_hypercall_complete_userspace(struct kvm_vcpu *vcpu)
 	return kvm_xen_hypercall_set_result(vcpu, run->xen.u.hcall.result);
 }
 
-static inline int max_evtchn_port(bool has_64bit_shinfo)
-{
-	if (has_64bit_shinfo)
-		return EVTCHN_2L_NR_CHANNELS;
-	else
-		return COMPAT_EVTCHN_2L_NR_CHANNELS;
-}
-
-static inline int kvm_max_evtchn_port(struct kvm *kvm)
-{
-	return max_evtchn_port(kvm_xen_has_64bit_shinfo(kvm));
-}
-
 static bool wait_pending_event(struct kvm_vcpu *vcpu, bool has_64bit_shinfo,
 			       int nr_ports, evtchn_port_t *ports)
 {
diff --git a/tools/testing/selftests/kvm/x86/xen_shinfo_test.c b/tools/testing/selftests/kvm/x86/xen_shinfo_test.c
index aaa414c677ad..3b11d0ee253f 100644
--- a/tools/testing/selftests/kvm/x86/xen_shinfo_test.c
+++ b/tools/testing/selftests/kvm/x86/xen_shinfo_test.c
@@ -356,11 +356,12 @@ static void guest_code(void)
 wait_for_timer:
 	/*
 	 * Poll for a timer wake event while the worker thread is mucking with
-	 * the shared info.  KVM XEN drops timer IRQs if the shared info is
-	 * invalid when the timer expires.  Arbitrarily poll 100 times before
-	 * giving up and asking the VMM to re-arm the timer.  100 polls should
-	 * consume enough time to beat on KVM without taking too long if the
-	 * timer IRQ is dropped due to an invalid event channel.
+	 * the shared info.  KVM is now supposed to retry a timer IRQ whose
+	 * delivery hit an invalid shared info rather than drop it, but older
+	 * kernels drop it on the floor.  Arbitrarily poll 100 times before
+	 * giving up and asking the VMM to re-arm the timer, so the test still
+	 * completes (and documents the VMM-side recovery protocol) on
+	 * unfixed kernels.
 	 */
 	for (i = 0; i < 100 && !guest_saw_irq; i++)
 		__xen_hypercall(__HYPERVISOR_sched_op, SCHEDOP_poll, &p);
-- 
2.43.0

[PATCH v4 06/12] KVM: x86/xen: Retry event channel delivery which races a memslot update

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

Event channel delivery now returns -EAGAIN when it races a memslot
update. It always failed; now the failure is distinguishable, and
instead of dropping the event each caller reacts according to its
context:

 - kvm_xen_hvm_evtchn_send(): the KVM_XEN_HVM_EVTCHN_SEND ioctl.
   May sleep: passes wait == true, and kvm_gpc_refresh() rides the
   update out.

 - evtchn_set_fn(): irqfd routing, from the irqfd workqueue after
   the atomic fast path failed. May also wait.

 - kvm_xen_hcall_evtchn_send(): the guest's own EVTCHNOP_send, in
   vcpu_run's exit handling under the vCPU's kvm->srcu read section
   — the very reader the memslot update is waiting out, so waiting
   there deadlocks, and any guest can trigger it. Instead, leave the
   hypercall uncompleted, with RIP not advanced, and the guest
   re-executes it; each attempt cycles the srcu read section on the
   way back in, letting the update complete. This is how Xen itself
   restarts preempted hypercalls, and the pending delivery is pure
   guest state, so there is nothing extra to migrate.

 - kvm_xen_inject_timer_irqs(): the timer retry added by the
   previous commit — timer_pending *is* its pending state, so it
   keeps that and retries on the next entry pass. The entry pass
   holds the vCPU's kvm->srcu, so it may not wait. The
   KVM_XEN_VCPU_GET_ATTR timer flush doesn't need to wait: an
   undelivered event is reported as still owed in the snapshot.

Reported-by: sashiko-bot@kernel.org
Closes: https://lore.kernel.org/all/20260920215341.5ECE91F000FF@smtp.kernel.org
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Assisted-by: LLM
---
 arch/x86/kvm/xen.c | 58 +++++++++++++++++++++++++++++++---------------
 1 file changed, 39 insertions(+), 19 deletions(-)

diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index 8a9d6963fc05..c0b9ac674255 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -30,9 +30,10 @@
 #include "cpuid.h"
 #include "trace.h"
 
-static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm);
+static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm,
+			      bool wait);
 static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data);
-static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r);
+static int kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r);
 
 DEFINE_STATIC_KEY_DEFERRED_FALSE(kvm_xen_enabled, HZ);
 
@@ -146,7 +147,7 @@ void kvm_xen_inject_timer_irqs(struct kvm_vcpu *vcpu)
 		e.port = vcpu->arch.xen.timer_virq;
 		e.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
 
-		rc = kvm_xen_set_evtchn(&e, vcpu->kvm);
+		rc = kvm_xen_set_evtchn(&e, vcpu->kvm, false);
 
 		/*
 		 * Leave timer_pending set if the memslot is being updated
@@ -1900,8 +1901,21 @@ int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
 		}
 		break;
 	case __HYPERVISOR_event_channel_op:
-		if (params[0] == EVTCHNOP_send)
-			handled = kvm_xen_hcall_evtchn_send(vcpu, params[1], &r);
+		if (params[0] == EVTCHNOP_send) {
+			int ret = kvm_xen_hcall_evtchn_send(vcpu, params[1], &r);
+
+			/*
+			 * Delivery raced a memslot update; waiting for it
+			 * here would deadlock under vcpu_run's kvm->srcu
+			 * read section. Leave the hypercall uncompleted —
+			 * RIP not advanced — so the guest re-executes it.
+			 * Each attempt cycles the srcu read section on the
+			 * way back in, letting the update complete.
+			 */
+			if (ret == -EAGAIN)
+				return 1;
+			handled = !ret;
+		}
 		break;
 	case __HYPERVISOR_sched_op:
 		handled = kvm_xen_hcall_sched_op(vcpu, is_64bit, params[0],
@@ -2092,7 +2106,8 @@ int kvm_xen_set_evtchn_fast(struct kvm_xen_evtchn *xe, struct kvm *kvm)
 	return rc;
 }
 
-static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm)
+static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm,
+			      bool wait)
 {
 	bool mm_borrowed = false;
 	int rc;
@@ -2129,16 +2144,13 @@ static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm)
 	 */
 	do {
 		struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
-		int idx;
 
 		rc = kvm_xen_set_evtchn_fast(xe, kvm);
 		if (rc != -EWOULDBLOCK)
 			break;
 
-		idx = srcu_read_lock(&kvm->srcu);
-		rc = kvm_gpc_refresh(gpc, PAGE_SIZE, false);
-		srcu_read_unlock(&kvm->srcu, idx);
-	} while(!rc);
+		rc = kvm_gpc_refresh(gpc, PAGE_SIZE, wait);
+	} while (!rc);
 
 	if (mm_borrowed)
 		kthread_unuse_mm(kvm->mm);
@@ -2153,7 +2165,7 @@ static int evtchn_set_fn(struct kvm_kernel_irq_routing_entry *e, struct kvm *kvm
 	if (!level)
 		return -EINVAL;
 
-	return kvm_xen_set_evtchn(&e->xen_evtchn, kvm);
+	return kvm_xen_set_evtchn(&e->xen_evtchn, kvm, true);
 }
 
 /*
@@ -2227,7 +2239,7 @@ int kvm_xen_hvm_evtchn_send(struct kvm *kvm, struct kvm_irq_routing_xen_evtchn *
 	e.vcpu_idx = -1;
 	e.priority = uxe->priority;
 
-	ret = kvm_xen_set_evtchn(&e, kvm);
+	ret = kvm_xen_set_evtchn(&e, kvm, true);
 
 	/*
 	 * None of that 'return 1 if it actually got delivered' nonsense.
@@ -2456,7 +2468,11 @@ static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data
 	return kvm_xen_eventfd_assign(kvm, data);
 }
 
-static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r)
+/*
+ * Returns 0 with *r set if handled, -EAGAIN if the hypercall should be
+ * re-executed by the guest, or -ENOENT to hand it to userspace.
+ */
+static int kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r)
 {
 	struct evtchnfd *evtchnfd;
 	struct evtchn_send send;
@@ -2466,7 +2482,7 @@ static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r)
 	BUILD_BUG_ON(sizeof(send) != 4);
 	if (kvm_read_guest_virt(vcpu, param, &send, sizeof(send), &e)) {
 		*r = -EFAULT;
-		return true;
+		return 0;
 	}
 
 	/*
@@ -2475,18 +2491,22 @@ static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r)
 	 */
 	evtchnfd = xa_load(&vcpu->kvm->arch.xen.evtchn_ports, send.port);
 	if (!evtchnfd)
-		return false;
+		return -ENOENT;
 
 	if (evtchnfd->deliver.port.port) {
-		int ret = kvm_xen_set_evtchn(&evtchnfd->deliver.port, vcpu->kvm);
+		int ret = kvm_xen_set_evtchn(&evtchnfd->deliver.port,
+					     vcpu->kvm, false);
+
+		if (ret == -EAGAIN)
+			return ret;
 		if (ret < 0 && ret != -ENOTCONN)
-			return false;
+			return -ENOENT;
 	} else {
 		eventfd_signal(evtchnfd->deliver.eventfd.ctx);
 	}
 
 	*r = 0;
-	return true;
+	return 0;
 }
 
 void kvm_xen_init_vcpu(struct kvm_vcpu *vcpu)
-- 
2.43.0

[PATCH v4 07/12] KVM: x86/xen: Extract delivery of event to vCPU into a separate helper

[David Woodhouse]

From: Sean Christopherson <seanjc@google.com>

Move the marking of an event as pending in the target vCPU's vcpu_info
into a separate helper, __kvm_xen_set_evtchn_fast(), invoked after the
shared_info processing is complete and its RCU read-side critical
section has been exited. This makes both halves easier to read.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
[dwmw2: rebased onto the -EAGAIN series. Sean's series first
 converted the GPC locks to read_trylock(), which this series does
 not carry, so the helper takes the plain read lock and the
 original's trylock failure path does not apply. The caller's
 kvm->srcu section extends across the helper call, since
 kvm_gpc_check() on the vcpu_info cache consults the memslot
 generation and the irqfd path enters holding only irq_srcu; in
 Sean's series that was covered by a guard(srcu) spanning the whole
 function. Note the "no functional change" is from the original;
 here the lock scopes do change: kvm_xen_check_poller() and the
 upcall vector injection now run outside the GPC read locks (and
 with interrupts restored), deliberately, so that the vCPU kick and
 the APIC map walk are not nested inside them.]
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Assisted-by: LLM
---
 arch/x86/kvm/xen.c | 158 ++++++++++++++++++++++++++++-----------------
 1 file changed, 97 insertions(+), 61 deletions(-)

diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index c0b9ac674255..6a1162caee9b 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -1970,6 +1970,89 @@ static void kvm_xen_check_poller(struct kvm_vcpu *vcpu, int port)
 	}
 }
 
+static void __kvm_xen_set_evtchn_fast(struct kvm_vcpu *vcpu, int port_word_bit,
+				      bool has_64bit_shinfo)
+{
+	struct gfn_to_pfn_cache *gpc = &vcpu->arch.xen.vcpu_info_cache;
+	unsigned long vi_pending_sel_ofs;
+	bool inject_upcall = false;
+	bool kick_vcpu = false;
+	unsigned long flags;
+	bool old;
+
+	vi_pending_sel_ofs = has_64bit_shinfo ?
+		offsetof(struct vcpu_info, evtchn_pending_sel) :
+		offsetof(struct compat_vcpu_info, evtchn_pending_sel);
+
+	/* Now switch to the vCPU's vcpu_info to set the index and pending_sel */
+	read_lock_irqsave(&gpc->lock, flags);
+	if (!kvm_gpc_check(gpc, sizeof(struct vcpu_info))) {
+		/*
+		 * Could not access the vcpu_info. Set the bit in-kernel and
+		 * prod the vCPU to deliver it for itself. Note that an MSI is
+		 * *not* an acceptable substitute here even if the vCPU has an
+		 * upcall vector: the pending bit only exists in
+		 * evtchn_pending_sel so far, and the vCPU has to be forced
+		 * out of the guest so that kvm_xen_inject_pending_events()
+		 * can refresh the cache and propagate it to the guest.
+		 */
+		if (!test_and_set_bit(port_word_bit, &vcpu->arch.xen.evtchn_pending_sel))
+			kick_vcpu = true;
+		goto out_unlock;
+	}
+	inject_upcall = true;
+
+	/*
+	 * Explicitly use a 32-bit btsl instead of test_and_set_bit(), which
+	 * would use btsq on x86-64. The vcpu_info is guest-controlled and only
+	 * required to be 32-bit aligned, so a 64-bit access could generate a
+	 * split-lock #AC.
+	 *
+	 * Note, this does not apply to the test_and_set_bit() on pending_bits
+	 * in the caller: that is in the per-VM shared_info, which is page
+	 * aligned, so the access is guaranteed to be 64-bit aligned.
+	 */
+	old = GEN_BINARY_RMWcc(LOCK_PREFIX "btsl",
+			       *(u32 *)(gpc->khva + vi_pending_sel_ofs),
+			       c, "Ir", port_word_bit);
+	if (!old) {
+		struct vcpu_info *vi = gpc->khva;
+
+		/* No need for compat handling */
+		BUILD_BUG_ON(offsetof(struct vcpu_info, evtchn_upcall_pending) !=
+			     offsetof(struct compat_vcpu_info, evtchn_upcall_pending));
+
+		WRITE_ONCE(vi->evtchn_upcall_pending, 1);
+		kick_vcpu = true;
+	}
+
+out_unlock:
+	read_unlock_irqrestore(&gpc->lock, flags);
+
+	/*
+	 * Deliver the upcall or kick the vCPU only after dropping the GPC
+	 * read lock. Both paths end up in kvm_vcpu_kick(), and the MSI
+	 * delivery also walks the APIC map and takes APIC locks; none of
+	 * that wants to be nested inside the GPC read lock, which a
+	 * concurrent invalidation may be waiting to take for write.
+	 */
+	if (!kick_vcpu)
+		return;
+
+	/*
+	 * For the per-vCPU lapic vector, deliver it as MSI — but only if the
+	 * vcpu_info was actually updated above. If it wasn't, the vCPU must
+	 * be kicked instead (see above).
+	 */
+	if (inject_upcall && vcpu->arch.xen.upcall_vector) {
+		kvm_xen_inject_vcpu_vector(vcpu);
+		return;
+	}
+
+	kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
+	kvm_vcpu_kick(vcpu);
+}
+
 /*
  * The return value from this function is propagated to kvm_set_irq() API,
  * so it returns:
@@ -1984,11 +2067,10 @@ int kvm_xen_set_evtchn_fast(struct kvm_xen_evtchn *xe, struct kvm *kvm)
 {
 	struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
 	bool has_64bit_shinfo = kvm_xen_has_64bit_shinfo(kvm);
-	unsigned long *pending_bits, *mask_bits, vi_pending_sel_ofs;
+	unsigned long *pending_bits, *mask_bits;
 	struct kvm_vcpu *vcpu;
 	unsigned long flags;
 	int port_word_bit;
-	bool kick_vcpu = false;
 	int vcpu_idx, idx, rc;
 
 	vcpu_idx = READ_ONCE(xe->vcpu_idx);
@@ -2010,23 +2092,19 @@ int kvm_xen_set_evtchn_fast(struct kvm_xen_evtchn *xe, struct kvm *kvm)
 
 	read_lock_irqsave(&gpc->lock, flags);
 	if (!kvm_gpc_check(gpc, PAGE_SIZE))
-		goto out_rcu;
+		goto out_unlock;
 
 	if (has_64bit_shinfo) {
 		struct shared_info *shinfo = gpc->khva;
 		pending_bits = (unsigned long *)&shinfo->evtchn_pending;
 		mask_bits = (unsigned long *)&shinfo->evtchn_mask;
 		port_word_bit = xe->port / 64;
-
-		vi_pending_sel_ofs = offsetof(struct vcpu_info, evtchn_pending_sel);
 	} else {
 		struct compat_shared_info *shinfo = gpc->khva;
 		pending_bits = (unsigned long *)&shinfo->evtchn_pending;
 		mask_bits = (unsigned long *)&shinfo->evtchn_mask;
 		port_word_bit = xe->port / 32;
 
-		vi_pending_sel_ofs = offsetof(struct compat_vcpu_info, evtchn_pending_sel);
-
 		/* test_and_set_bit() needs 64-bit alignment, but that's OK */
 		BUILD_BUG_ON(offsetof(struct compat_shared_info, evtchn_pending) & 7);
 	}
@@ -2042,66 +2120,24 @@ int kvm_xen_set_evtchn_fast(struct kvm_xen_evtchn *xe, struct kvm *kvm)
 		rc = 0; /* It was already raised */
 	} else if (test_bit(xe->port, mask_bits)) {
 		rc = -ENOTCONN; /* Masked */
-		kvm_xen_check_poller(vcpu, xe->port);
 	} else {
-		bool old;
-
 		rc = 1; /* Delivered to the bitmap in shared_info. */
-		/* Now switch to the vCPU's vcpu_info to set the index and pending_sel */
-		read_unlock_irqrestore(&gpc->lock, flags);
-		gpc = &vcpu->arch.xen.vcpu_info_cache;
-
-		read_lock_irqsave(&gpc->lock, flags);
-		if (!kvm_gpc_check(gpc, sizeof(struct vcpu_info))) {
-			/*
-			 * Could not access the vcpu_info. Set the bit in-kernel
-			 * and prod the vCPU to deliver it for itself.
-			 */
-			if (!test_and_set_bit(port_word_bit, &vcpu->arch.xen.evtchn_pending_sel))
-				kick_vcpu = true;
-			goto out_rcu;
-		}
-
-		/*
-		 * Explicitly use a 32-bit btsl instead of test_and_set_bit(),
-		 * which would use btsq on x86-64. The vcpu_info is guest-
-		 * controlled and only required to be 32-bit aligned, so a
-		 * 64-bit access could generate a split-lock #AC.
-		 *
-		 * Note, this does not apply to the test_and_set_bit() on
-		 * pending_bits above: that is in the per-VM shared_info, which
-		 * is page aligned, so the access is guaranteed to be 64-bit
-		 * aligned.
-		 */
-		old = GEN_BINARY_RMWcc(LOCK_PREFIX "btsl",
-				       *(u32 *)(gpc->khva + vi_pending_sel_ofs),
-				       c, "Ir", port_word_bit);
-		if (!old) {
-			struct vcpu_info *vi = gpc->khva;
-
-			/* No need for compat handling */
-			BUILD_BUG_ON(offsetof(struct vcpu_info, evtchn_upcall_pending) !=
-				     offsetof(struct compat_vcpu_info, evtchn_upcall_pending));
-
-			WRITE_ONCE(vi->evtchn_upcall_pending, 1);
-			kick_vcpu = true;
-		}
-
-		/* For the per-vCPU lapic vector, deliver it as MSI. */
-		if (kick_vcpu && vcpu->arch.xen.upcall_vector) {
-			kvm_xen_inject_vcpu_vector(vcpu);
-			kick_vcpu = false;
-		}
 	}
 
- out_rcu:
+ out_unlock:
 	read_unlock_irqrestore(&gpc->lock, flags);
-	srcu_read_unlock(&kvm->srcu, idx);
 
-	if (kick_vcpu) {
-		kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
-		kvm_vcpu_kick(vcpu);
-	}
+	/*
+	 * Both of these can kick the vCPU, so keep them outside the GPC
+	 * read lock; an invalidation may be waiting to take it for write.
+	 */
+	if (rc == -ENOTCONN)
+		kvm_xen_check_poller(vcpu, xe->port);
+	else if (rc == 1)
+		__kvm_xen_set_evtchn_fast(vcpu, port_word_bit,
+					  has_64bit_shinfo);
+
+	srcu_read_unlock(&kvm->srcu, idx);
 
 	return rc;
 }
-- 
2.43.0

[PATCH v4 08/12] KVM: x86/xen: Explicitly tag "shared info" page as never being dirty tracked

[David Woodhouse]

From: Sean Christopherson <seanjc@google.com>

Explicitly mark the Xen shared info page as never being dirty tracked so
that higher-level gpc APIs can be added to automatically take care of
things like dirty tracking, without reintroducing the bug fixed by
commit 55749769fe60 ("KVM: x86: Fix wall clock writes in Xen shared_info
not to mark page dirty").  And because the code _looks_ buggy.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Assisted-by: LLM
---
 arch/x86/kvm/xen.c        |  2 +-
 include/linux/kvm_host.h  | 15 +++++++++++----
 include/linux/kvm_types.h |  1 +
 virt/kvm/pfncache.c       |  4 +++-
 4 files changed, 16 insertions(+), 6 deletions(-)

diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index 6a1162caee9b..0ba684bf01d3 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -2577,7 +2577,7 @@ void kvm_xen_init_vm(struct kvm *kvm)
 {
 	mutex_init(&kvm->arch.xen.xen_lock);
 	xa_init(&kvm->arch.xen.evtchn_ports);
-	kvm_gpc_init(&kvm->arch.xen.shinfo_cache, kvm);
+	__kvm_gpc_init(&kvm->arch.xen.shinfo_cache, kvm, true);
 }
 
 void kvm_xen_destroy_vm(struct kvm *kvm)
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index 94989a705ee6..ba23f4e0a4c0 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -1456,16 +1456,23 @@ int kvm_vcpu_write_guest(struct kvm_vcpu *vcpu, gpa_t gpa, const void *data,
 			 unsigned long len);
 
 /**
- * kvm_gpc_init - initialize gfn_to_pfn_cache.
+ * __kvm_gpc_init - initialize gfn_to_pfn_cache.
  *
  * @gpc:	   struct gfn_to_pfn_cache object.
  * @kvm:	   pointer to kvm instance.
+ * @never_dirty:   the cache never needs to mark its page dirty in its memslot.
  *
  * This sets up a gfn_to_pfn_cache by initializing locks and assigning the
  * immutable attributes.  Note, the cache must be zero-allocated (or zeroed by
- * the caller before init).
+ * the caller before init).  Most callers want the kvm_gpc_init() wrapper.
  */
-void kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm);
+void __kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm,
+		    bool never_dirty);
+
+static inline void kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm)
+{
+	__kvm_gpc_init(gpc, kvm, false);
+}
 
 /**
  * kvm_gpc_activate - prepare a cached kernel mapping and HPA for a given guest
@@ -1990,7 +1997,7 @@ static inline void kvm_gpc_mark_dirty_in_slot(struct gfn_to_pfn_cache *gpc)
 {
 	lockdep_assert_held(&gpc->lock);
 
-	if (!gpc->memslot)
+	if (!gpc->memslot || gpc->never_dirty)
 		return;
 
 	mark_page_dirty_in_slot(gpc->kvm, gpc->memslot, gpa_to_gfn(gpc->gpa));
diff --git a/include/linux/kvm_types.h b/include/linux/kvm_types.h
index a568d8e6f4e8..e850adc3f47e 100644
--- a/include/linux/kvm_types.h
+++ b/include/linux/kvm_types.h
@@ -94,6 +94,7 @@ struct gfn_to_pfn_cache {
 	kvm_pfn_t pfn;
 	bool active;
 	bool valid;
+	bool never_dirty;
 };
 
 #ifdef KVM_ARCH_NR_OBJS_PER_MEMORY_CACHE
diff --git a/virt/kvm/pfncache.c b/virt/kvm/pfncache.c
index 0e700c337f5a..8498abb746ad 100644
--- a/virt/kvm/pfncache.c
+++ b/virt/kvm/pfncache.c
@@ -437,7 +437,8 @@ int kvm_gpc_refresh(struct gfn_to_pfn_cache *gpc, unsigned long len, bool wait)
 	}
 }
 
-void kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm)
+void __kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm,
+		    bool never_dirty)
 {
 	rwlock_init(&gpc->lock);
 	mutex_init(&gpc->refresh_lock);
@@ -447,6 +448,7 @@ void kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm)
 	gpc->gpa = INVALID_GPA;
 	gpc->uhva = KVM_HVA_ERR_BAD;
 	gpc->active = gpc->valid = false;
+	gpc->never_dirty = never_dirty;
 }
 
 static int __kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long uhva,
-- 
2.43.0

[PATCH v4 09/12] KVM: x86/xen: Don't dirty track "vCPU info" page

[David Woodhouse]

From: Sean Christopherson <seanjc@google.com>

Give the Xen per-vCPU info page the same treatment as the per-VM shared
info page, and never mark it dirty, as KVM clearly relies on userspace to
assume the page is always dirty.  While the page is marked dirty on writes
via kvm_xen_inject_pending_events(), it's not marked dirty when written by
__kvm_xen_set_evtchn_fast().

Furthermore, as was the case with the shared info page, writes in the event
channel fastpath may be done without an active vCPU, e.g. when called via
timer callback or irqfd injection.  I.e. attempting to fix the fastpath
would run afoul of the same issue that was fixed by
commit 55749769fe60 ("KVM: x86: Fix wall clock writes in Xen shared_info
not to mark page dirty").

This also covers the pvclock writes kvm_guest_time_update() makes
through the same cache; userspace assuming the page is always dirty
covers those equally.

Signed-off-by: Sean Christopherson <seanjc@google.com>
[dwmw2: also update the KVM_GET_DIRTY_LOG note in api.rst, which
 named only the shared_info page]
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Assisted-by: LLM
---
 Documentation/virt/kvm/api.rst | 5 +++--
 arch/x86/kvm/xen.c             | 3 +--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
index e0430cc750c9..b24aac070b11 100644
--- a/Documentation/virt/kvm/api.rst
+++ b/Documentation/virt/kvm/api.rst
@@ -385,8 +385,9 @@ The bits in the dirty bitmap are cleared before the ioctl returns, unless
 KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 is enabled.  For more information,
 see the description of the capability.
 
-Note that the Xen shared_info page, if configured, shall always be assumed
-to be dirty. KVM will not explicitly mark it such.
+Note that the Xen shared_info page and vcpu_info area, if configured,
+shall always be assumed to be dirty. KVM will not explicitly mark them
+such.
 
 
 4.10 KVM_RUN
diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index 0ba684bf01d3..94eb55a6d702 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -778,7 +778,6 @@ void kvm_xen_inject_pending_events(struct kvm_vcpu *v)
 		WRITE_ONCE(vi->evtchn_upcall_pending, 1);
 	}
 
-	kvm_gpc_mark_dirty_in_slot(gpc);
 	read_unlock_irqrestore(&gpc->lock, flags);
 
 	/* For the per-vCPU lapic vector, deliver it as MSI. */
@@ -2556,7 +2555,7 @@ void kvm_xen_init_vcpu(struct kvm_vcpu *vcpu)
 
 	kvm_gpc_init(&vcpu->arch.xen.runstate_cache, vcpu->kvm);
 	kvm_gpc_init(&vcpu->arch.xen.runstate2_cache, vcpu->kvm);
-	kvm_gpc_init(&vcpu->arch.xen.vcpu_info_cache, vcpu->kvm);
+	__kvm_gpc_init(&vcpu->arch.xen.vcpu_info_cache, vcpu->kvm, true);
 	kvm_gpc_init(&vcpu->arch.xen.vcpu_time_info_cache, vcpu->kvm);
 }
 
-- 
2.43.0

[PATCH v4 10/12] KVM: x86: Request the guest TLB flush from record_steal_time()

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

record_steal_time() performs a TLB flush on the guest's behalf when the
guest sets KVM_VCPU_FLUSH_TLB. With shadow paging,
kvm_vcpu_flush_tlb_guest() ends up in mmu_sync_children(), which can
reschedule via cond_resched_rwlock_write(). That is why the flush
currently sits between a user_access_end() and a fresh
user_access_begin(): it cannot be done while the guest page is mapped
for access.

Raise KVM_REQ_TLB_FLUSH_GUEST instead of flushing inline, and move the
processing of KVM_REQ_STEAL_UPDATE ahead of where the TLB flush requests
are serviced in vcpu_enter_guest() so that the flush still happens before
the vCPU enters the guest. There is no requirement for it to complete any
earlier than that.

kvm_make_request() on the vCPU which is running is nothing more than a
barrier and a set_bit(), so unlike the flush itself it can be done from
any context — which lets the surrounding access be simplified later.

Note that the xchg() must stay where it is: it atomically consumes any
KVM_VCPU_FLUSH_TLB which the guest has set, as required by commit
b043138246a4 ("x86/KVM: Make sure KVM_VCPU_FLUSH_TLB flag is not missed").

Reported-by: Sashiko AI review <sashiko-bot@kernel.org>
Closes: https://lore.kernel.org/all/20260530061932.7849D1F00893@smtp.kernel.org
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Assisted-by: LLM
---
 arch/x86/kvm/x86.c | 26 +++++++++++++++++++++++---
 1 file changed, 23 insertions(+), 3 deletions(-)

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index eadcaeee651a..20119d07b477 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2106,8 +2106,22 @@ static void record_steal_time(struct kvm_vcpu *vcpu)
 
 		trace_kvm_pv_tlb_flush(vcpu->vcpu_id,
 				       st_preempted & KVM_VCPU_FLUSH_TLB);
+		/*
+		 * Request the flush instead of performing it inline. With
+		 * shadow paging kvm_vcpu_flush_tlb_guest() can reschedule (via
+		 * mmu_sync_children()), which is why the user access above has
+		 * to be closed before it. Requesting the flush is just a bit
+		 * set on the running vCPU, and the request is serviced before
+		 * the vCPU enters the guest; nothing requires the flush to
+		 * have completed any earlier than that.
+		 *
+		 * Note the xchg above must remain as-is: it atomically
+		 * consumes any KVM_VCPU_FLUSH_TLB the guest set, as required
+		 * by commit b043138246a4 ("x86/KVM: Make sure
+		 * KVM_VCPU_FLUSH_TLB flag is not missed").
+		 */
 		if (st_preempted & KVM_VCPU_FLUSH_TLB)
-			kvm_vcpu_flush_tlb_guest(vcpu);
+			kvm_make_request(KVM_REQ_TLB_FLUSH_GUEST, vcpu);
 
 		if (!user_access_begin(st, sizeof(*st)))
 			goto dirty;
@@ -8089,6 +8103,14 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 			if (unlikely(r))
 				goto out;
 		}
+		/*
+		 * Process the steal time update before the TLB flush requests
+		 * are serviced below; record_steal_time() requests
+		 * KVM_REQ_TLB_FLUSH_GUEST on the guest's behalf rather than
+		 * performing the flush itself.
+		 */
+		if (kvm_check_request(KVM_REQ_STEAL_UPDATE, vcpu))
+			record_steal_time(vcpu);
 		if (kvm_check_request(KVM_REQ_MMU_SYNC, vcpu))
 			kvm_mmu_sync_roots(vcpu);
 		if (kvm_check_request(KVM_REQ_LOAD_MMU_PGD, vcpu))
@@ -8138,8 +8160,6 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 			r = 1;
 			goto out;
 		}
-		if (kvm_check_request(KVM_REQ_STEAL_UPDATE, vcpu))
-			record_steal_time(vcpu);
 		if (kvm_check_request(KVM_REQ_PMU, vcpu))
 			kvm_pmu_handle_event(vcpu);
 		if (kvm_check_request(KVM_REQ_PMI, vcpu))
-- 
2.43.0

[PATCH v4 11/12] KVM: selftests: Add Xen evtchn vs memslot-update race test

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

Race a guest EVTCHNOP_send loop against a host thread which deletes
and recreates the shared_info memslot. Delivery which hits the update
window now returns -EAGAIN and the hypercall is re-executed; if KVM
instead waited in the vCPU's exit-handling context — under vcpu_run's
kvm->srcu read section — it would deadlock against
kvm_swap_active_memslots() waiting out that reader, and this test
livelocks within seconds.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Assisted-by: LLM
---
 tools/testing/selftests/kvm/Makefile.kvm      |   1 +
 .../kvm/x86/xen_evtchn_memslot_race_test.c    | 208 ++++++++++++++++++
 2 files changed, 209 insertions(+)
 create mode 100644 tools/testing/selftests/kvm/x86/xen_evtchn_memslot_race_test.c

diff --git a/tools/testing/selftests/kvm/Makefile.kvm b/tools/testing/selftests/kvm/Makefile.kvm
index 96bab7002d39..e7e17d19d6cb 100644
--- a/tools/testing/selftests/kvm/Makefile.kvm
+++ b/tools/testing/selftests/kvm/Makefile.kvm
@@ -144,6 +144,7 @@ TEST_GEN_PROGS_x86 += x86/xss_msr_test
 TEST_GEN_PROGS_x86 += x86/debug_regs
 TEST_GEN_PROGS_x86 += x86/tsc_msrs_test
 TEST_GEN_PROGS_x86 += x86/vmx_pmu_caps_test
+TEST_GEN_PROGS_x86 += x86/xen_evtchn_memslot_race_test
 TEST_GEN_PROGS_x86 += x86/xen_shinfo_test
 TEST_GEN_PROGS_x86 += x86/xen_vmcall_test
 TEST_GEN_PROGS_x86 += x86/sev_dbg_test
diff --git a/tools/testing/selftests/kvm/x86/xen_evtchn_memslot_race_test.c b/tools/testing/selftests/kvm/x86/xen_evtchn_memslot_race_test.c
new file mode 100644
index 000000000000..08dd1aac884f
--- /dev/null
+++ b/tools/testing/selftests/kvm/x86/xen_evtchn_memslot_race_test.c
@@ -0,0 +1,208 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * xen_evtchn_memslot_race_test
+ *
+ * A guest EVTCHNOP_send whose delivery races a DELETE/re-CREATE of the
+ * memslot holding the shared_info page. The delivery path runs in the
+ * vCPU's exit-handling context, under vcpu_run's kvm->srcu read
+ * section; if it waits there for the memslot update to complete, it
+ * deadlocks against kvm_swap_active_memslots()'s
+ * synchronize_srcu_expedited() waiting out that very read section.
+ * Regression test for that deadlock: KVM must leave the hypercall
+ * uncompleted so the guest re-executes it, rather than wait in exit
+ * context.
+ *
+ * The guest sends events in a tight loop; a host thread deletes and
+ * recreates the shared_info slot in a tight loop. On a broken host
+ * this livelocks within a second or two (the test then dies by
+ * timeout); on a fixed host the guest completes all N
+ * iterations (sends which hit the no-slot gap between DELETE and
+ * re-CREATE legitimately fail; most deliver).
+ */
+#include <pthread.h>
+#include <stdatomic.h>
+
+#include "test_util.h"
+#include "kvm_util.h"
+#include "processor.h"
+
+#include <stdint.h>
+
+#define SHINFO_REGION_GPA	0xc0000000ULL
+#define SHINFO_REGION_SLOT	10
+
+#define EVTCHN_PORT		15
+#define TEST_ITERATIONS		4000
+
+#define EVTCHNOP_send		4
+#define __HYPERVISOR_event_channel_op	32
+
+#define XEN_HYPERCALL_MSR	0x40000000
+
+#define EVTCHNSTAT_ipi		5
+
+struct evtchn_send {
+	uint32_t port;
+};
+
+static void guest_code(void)
+{
+	struct evtchn_send s = { .port = EVTCHN_PORT };
+	int delivered = 0;
+	int i;
+
+	for (i = 0; i < TEST_ITERATIONS; i++) {
+		/*
+		 * The delivery may legitimately fail while the shared_info
+		 * slot is mid-churn (between DELETE and re-CREATE there is
+		 * no slot at all, which is -EFAULT, not a transient). The
+		 * failure mode under test is the vCPU never coming back at
+		 * all (deadlock -> timeout), not the hypercall rc.
+		 */
+		if (!__xen_hypercall(__HYPERVISOR_event_channel_op,
+				     EVTCHNOP_send, &s))
+			delivered++;
+		GUEST_SYNC(i);
+	}
+
+	/* The churn must not have starved delivery entirely. */
+	GUEST_ASSERT(delivered);
+
+	/* Report the delivered count (always > TEST_ITERATIONS - sync i). */
+	GUEST_SYNC(TEST_ITERATIONS + delivered);
+
+	GUEST_DONE();
+}
+
+static atomic_bool churn_stop;
+
+static void *memslot_churn(void *arg)
+{
+	struct kvm_vm *vm = arg;
+
+	while (!atomic_load(&churn_stop)) {
+		vm_mem_region_delete(vm, SHINFO_REGION_SLOT);
+		vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
+					    SHINFO_REGION_GPA,
+					    SHINFO_REGION_SLOT, 2, 0);
+	}
+	return NULL;
+}
+
+int main(int argc, char *argv[])
+{
+	struct kvm_vcpu *vcpu;
+	struct kvm_vm *vm;
+	pthread_t thread;
+	struct ucall uc;
+	int done = 0;
+	int r;
+
+	int xen_caps = kvm_check_cap(KVM_CAP_XEN_HVM);
+
+	TEST_REQUIRE(xen_caps & KVM_XEN_HVM_CONFIG_SHARED_INFO);
+	TEST_REQUIRE(xen_caps & KVM_XEN_HVM_CONFIG_EVTCHN_SEND);
+
+	vm = vm_create_with_one_vcpu(&vcpu, guest_code);
+
+	vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
+				    SHINFO_REGION_GPA, SHINFO_REGION_SLOT, 2, 0);
+
+	struct kvm_xen_hvm_config hvmc = {
+		.flags = KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL |
+			 KVM_XEN_HVM_CONFIG_EVTCHN_SEND,
+		.msr = XEN_HYPERCALL_MSR,
+	};
+	vm_ioctl(vm, KVM_XEN_HVM_CONFIG, &hvmc);
+
+	struct kvm_xen_hvm_attr lm = {
+		.type = KVM_XEN_ATTR_TYPE_LONG_MODE,
+		.u.long_mode = 1,
+	};
+	vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &lm);
+
+	struct kvm_xen_hvm_attr ha = {
+		.type = KVM_XEN_ATTR_TYPE_SHARED_INFO,
+		.u.shared_info.gfn = SHINFO_REGION_GPA / PAGE_SIZE,
+	};
+	vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &ha);
+
+	struct kvm_xen_vcpu_attr vi = {
+		.type = KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO,
+		.u.gpa = SHINFO_REGION_GPA + 0x40,
+	};
+	vcpu_ioctl(vcpu, KVM_XEN_VCPU_SET_ATTR, &vi);
+
+	/* Loop the port back to this vCPU as an IPI. */
+	struct kvm_xen_hvm_attr evt = {
+		.type = KVM_XEN_ATTR_TYPE_EVTCHN,
+		.u.evtchn.send_port = EVTCHN_PORT,
+		.u.evtchn.type = EVTCHNSTAT_ipi,
+		.u.evtchn.flags = 0,
+		.u.evtchn.deliver.port.port = EVTCHN_PORT,
+		.u.evtchn.deliver.port.vcpu = vcpu->id,
+		.u.evtchn.deliver.port.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL,
+	};
+	vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &evt);
+
+	r = pthread_create(&thread, NULL, memslot_churn, vm);
+	TEST_ASSERT(!r, "pthread_create() failed: %d", r);
+
+	/*
+	 * Self-guard: on a kernel with the deadlock under test, the vCPU
+	 * never returns from KVM_RUN. Die with SIGALRM rather than hang
+	 * when run standalone. One budget for the whole run — a healthy
+	 * kernel finishes in seconds, and a per-iteration bound would
+	 * impose a KVM_RUN latency claim this test doesn't make.
+	 */
+	alarm(90);
+
+	while (!done) {
+		vcpu_run(vcpu);
+
+		/*
+		 * In the window where the shared_info slot is fully deleted
+		 * (between the churn thread's DELETE and re-CREATE), the
+		 * delivery fails with -EFAULT and the hypercall bounces to
+		 * userspace as KVM_EXIT_XEN. Complete it with an error
+		 * result and keep going; the failure mode under test is
+		 * the vCPU never coming back (deadlock -> timeout), not
+		 * the hypercall rc.
+		 */
+		if (vcpu->run->exit_reason == KVM_EXIT_XEN &&
+		    vcpu->run->xen.type == KVM_EXIT_XEN_HCALL) {
+			vcpu->run->xen.u.hcall.result = -EFAULT;
+			continue;
+		}
+
+		TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
+
+		switch (get_ucall(vcpu, &uc)) {
+		case UCALL_SYNC:
+			/*
+			 * The final sync reports how many of the sends were
+			 * delivered: purely a diagnostic for how much of the
+			 * churn the guest actually raced with.
+			 */
+			if (uc.args[1] >= TEST_ITERATIONS)
+				pr_info("%lu of %u sends delivered\n",
+					uc.args[1] - TEST_ITERATIONS,
+					TEST_ITERATIONS);
+			break;
+		case UCALL_DONE:
+			done = 1;
+			break;
+		case UCALL_ABORT:
+			REPORT_GUEST_ASSERT(uc);
+		default:
+			TEST_FAIL("Unexpected ucall %lu", uc.cmd);
+		}
+	}
+
+	atomic_store(&churn_stop, true);
+	pthread_join(thread, NULL);
+	alarm(0);
+
+	kvm_vm_free(vm);
+	return 0;
+}
-- 
2.43.0

[PATCH v4 12/12] KVM: selftests: Fix unreachable timer recovery in xen_shinfo_test

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

The final phase of xen_shinfo_test waits for a timer event while a
worker thread repeatedly deactivates and reactivates the shared_info
cache. The guest polls up to 100 times before asking the VMM to re-arm
the timer, documenting the recovery protocol for kernels which drop a
timer event that cannot be delivered.

But the poll has no timeout. If the timer fires while the guest is
already blocked in SCHEDOP_poll and the event is dropped — which is
precisely the case the recovery exists for — nothing ever wakes the
vCPU, the 100-poll loop never advances, and the test hangs with the
SIGALRM watchdog disabled for this phase. The recovery protocol is
only reachable when the drop happens to occur while the guest is
awake.

Give the poll a 10ms timeout, as the TEST_POLL_TIMEOUT phase already
does, so a dropped event leads to the documented re-arm protocol
instead of sleeping forever. The window for the drop is vanishingly
small upstream, but widens dramatically once cache deactivation incurs
a grace-period wait, hanging roughly a third of all runs.

Fixes: a51abbbf2531 ("KVM: selftests: Add tests in xen_shinfo_test to detect lock races")
Assisted-by: LLM
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
---
 tools/testing/selftests/kvm/x86/xen_shinfo_test.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/tools/testing/selftests/kvm/x86/xen_shinfo_test.c b/tools/testing/selftests/kvm/x86/xen_shinfo_test.c
index 3b11d0ee253f..729c0ba9ab4e 100644
--- a/tools/testing/selftests/kvm/x86/xen_shinfo_test.c
+++ b/tools/testing/selftests/kvm/x86/xen_shinfo_test.c
@@ -350,7 +350,14 @@ static void guest_code(void)
 	p = (struct sched_poll) {
 		.ports = ports,
 		.nr_ports = 1,
-		.timeout = 0
+		/*
+		 * The poll must have a timeout: if the timer event is dropped
+		 * while we are already blocked in the poll, nothing else will
+		 * ever wake us, and the recovery protocol below could never
+		 * run. (As with TEST_POLL_TIMEOUT above, KVM treats the
+		 * timeout as relative.)
+		 */
+		.timeout = 10000000,
 	};
 
 wait_for_timer:
-- 
2.43.0

Re: [PATCH v4 06/12] KVM: x86/xen: Retry event channel delivery which races a memslot update

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] KVM: x86/xen: 3-way deadlock via wait=true in evtchn_set_fn() from PIT worker
--

commit deb1cf394d02a7487c9bd0713d5fed7a4be72540
Author: David Woodhouse <dwmw@amazon.co.uk>

KVM: x86/xen: Retry event channel delivery which races a memslot update

This commit updates Xen event channel delivery callers to handle -EAGAIN
when delivery races with a memslot update, either waiting for the update
to complete or retrying depending on the execution context.

> diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
> index 8a9d6963fc05a..c0b9ac674255d 100644
> --- a/arch/x86/kvm/xen.c
> +++ b/arch/x86/kvm/xen.c

[ ... ]

> @@ -2153,7 +2165,7 @@ static int evtchn_set_fn(struct kvm_kernel_irq_routing_entry *e, struct kvm *kvm
>  	if (!level)
>  		return -EINVAL;
>  
> -	return kvm_xen_set_evtchn(&e->xen_evtchn, kvm);
> +	return kvm_xen_set_evtchn(&e->xen_evtchn, kvm, true);
>  }

[Severity: High]
Can passing wait=true unconditionally in evtchn_set_fn() cause a 3-way
deadlock when GSI 0 is routed to KVM_IRQ_ROUTING_XEN_EVTCHN and the
in-kernel PIT is active?

evtchn_set_fn() is the .set callback invoked by kvm_set_irq(). In addition
to the irqfd workqueue, pit_do_work() in arch/x86/kvm/i8254.c runs on the
pit->worker kthread and fires GSI 0:

arch/x86/kvm/i8254.c:pit_do_work() {
    ...
    kvm_set_irq(kvm, KVM_PIT_IRQ_SOURCE_ID, 0, 1, false);
    kvm_set_irq(kvm, KVM_PIT_IRQ_SOURCE_ID, 0, 0, false);
    ...
}

Because pit->worker is a kthread with current->mm == NULL,
kvm_xen_set_evtchn() borrows kvm->mm via kthread_use_mm() and calls
kvm_gpc_refresh(gpc, PAGE_SIZE, true).

If a memslot move or deletion is concurrently updating the memslot backing
shinfo_cache, kvm_invalidate_memslot() -> kvm_swap_active_memslots() marks
the slot KVM_MEMSLOT_INVALID and blocks in
synchronize_srcu_expedited(&kvm->srcu).

If a vCPU thread inside vcpu_run()'s kvm->srcu read section concurrently
emulates a guest PIO write to PIT ports 0x40-0x43, pit_ioport_write() ->
pit_load_count() calls create_pit_timer() or destroy_pit_timer(), both of
which flush pit->expired while holding kvm->srcu:

arch/x86/kvm/i8254.c:create_pit_timer() {
    ...
    hrtimer_cancel(&ps->timer);
    kthread_flush_work(&pit->expired);
    ...
}

At that point:
1. The memslot update thread waits for the vCPU thread to exit its
   kvm->srcu read section.
2. The vCPU thread holds kvm->srcu and waits in
   kthread_flush_work(&pit->expired) for pit_do_work() to finish.
3. pit_do_work() loops in kvm_gpc_refresh(..., wait=true) on -EAGAIN
   waiting for the memslot update to finish, and because pit->worker is a
   kthread, fatal_signal_pending(current) never triggers.

[ ... ]

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261019162445.37219-1-dwmw2@infradead.org?part=6

Re: [PATCH v4 12/12] KVM: selftests: Fix unreachable timer recovery in xen_shinfo_test

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] KVM: selftests: wrong fix in guest_code() for juggle_shinfo_state() GFN 0 bug
--

commit 36db2d77eb09cc5d7b05a58f930d6724ba060038
Author: David Woodhouse <dwmw@amazon.co.uk>

KVM: selftests: Fix unreachable timer recovery in xen_shinfo_test

This commit adds a 10ms timeout to the SCHEDOP_poll hypercall in the
TEST_LOCKING_POLL_RACE phase of xen_shinfo_test so the guest does not hang if
the timer event is dropped while blocked in the poll.

> diff --git a/tools/testing/selftests/kvm/x86/xen_shinfo_test.c b/tools/testing/selftests/kvm/x86/xen_shinfo_test.c
> index 3b11d0ee253f5..729c0ba9ab4e8 100644
> --- a/tools/testing/selftests/kvm/x86/xen_shinfo_test.c
> +++ b/tools/testing/selftests/kvm/x86/xen_shinfo_test.c
> @@ -350,7 +350,14 @@ static void guest_code(void)
>  	p = (struct sched_poll) {
>  		.ports = ports,
>  		.nr_ports = 1,
> -		.timeout = 0
> +		/*
> +		 * The poll must have a timeout: if the timer event is dropped
> +		 * while we are already blocked in the poll, nothing else will
> +		 * ever wake us, and the recovery protocol below could never
> +		 * run. (As with TEST_POLL_TIMEOUT above, KVM treats the
> +		 * timeout as relative.)
> +		 */
> +		.timeout = 10000000,
>  	};
>  
>  wait_for_timer:

[Severity: Medium]
Is SCHEDOP_poll in TEST_LOCKING_POLL_RACE actually supposed to block waiting
for the timer?

In main(), the host sets bit 0 of evtchn_pending[0] before arming the timer on
EVTCHN_TIMER:

tools/testing/selftests/kvm/x86/xen_shinfo_test.c:main() {
    ...
    case TEST_LOCKING_POLL_RACE:
        shinfo->evtchn_pending[0] = 1;

        evtchn_irq_expected = true;
        tmr.u.timer.expires_ns = rs->state_entry_time +
                                 SHINFO_RACE_TIMEOUT * 1000000000ULL;
        vcpu_ioctl(vcpu, KVM_XEN_VCPU_SET_ATTR, &tmr);
        break;
}

And in guest_code(), the guest polls on port 0 (ports[0] = 0), not on
EVTCHN_TIMER. In wait_pending_event(), ret defaults to true:

arch/x86/kvm/xen.c:wait_pending_event() {
    ...
    bool ret = true;
    int idx, i;

    idx = srcu_read_lock(&kvm->srcu);
    read_lock_irqsave(&gpc->lock, flags);
    if (!kvm_gpc_check(gpc, PAGE_SIZE))
        goto out_rcu;
    ...
}

When shinfo_cache is deactivated, kvm_gpc_check() returns false and
wait_pending_event() returns true immediately. When shinfo_cache is active on
shinfo, bit 0 of evtchn_pending[0] is set, so test_bit() also returns true
without halting the vCPU. Furthermore, if the timer expires while shinfo_cache
is deactivated, xen_timer_callback() sees -EWOULDBLOCK from
kvm_xen_set_evtchn_fast() and unconditionally raises KVM_REQ_UNBLOCK and calls
kvm_vcpu_kick().

Instead, isn't the vCPU blocking in kvm_vcpu_halt() because of a bug in
juggle_shinfo_state()?

tools/testing/selftests/kvm/x86/xen_shinfo_test.c:juggle_shinfo_state() {
    ...
    struct kvm_xen_hvm_attr cache_deactivate_hva = {
        .type = KVM_XEN_ATTR_TYPE_SHARED_INFO,
        .u.shared_info.hva = 0
    };
    ...
}

Because cache_deactivate_hva uses KVM_XEN_ATTR_TYPE_SHARED_INFO instead of
KVM_XEN_ATTR_TYPE_SHARED_INFO_HVA with hva = 0 (which unions with gfn = 0),
kvm_xen_hvm_set_attr() activates shinfo_cache on guest GFN 0 rather than
deactivating it (since KVM_XEN_ATTR_TYPE_SHARED_INFO checks for
KVM_XEN_INVALID_GFN). At guest GFN 0, evtchn_pending[0] bit 0 is 0, so a
concurrent SCHEDOP_poll sees the port as not pending and blocks in
kvm_vcpu_halt().

Adding a 10ms timeout here causes the guest to sleep in kvm_vcpu_halt() for
10ms whenever SCHEDOP_poll coincides with cache_deactivate_hva, rather than
spinning in the tight hypercall loop to stress gpc->lock contention, while
leaving KVM_XEN_ATTR_TYPE_SHARED_INFO_HVA deactivation untested and continuing
to overwrite guest GFN 0 with wallclock data.

Should this instead fix cache_deactivate_hva.type to
KVM_XEN_ATTR_TYPE_SHARED_INFO_HVA in juggle_shinfo_state(), with:

  Fixes: 9397b5334af1 ("KVM: selftests: map Xen's shared_info page using HVA rather than GFN")

?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261019162445.37219-1-dwmw2@infradead.org?part=12

Re: [PATCH v4 05/12] KVM: x86/xen: Don't lose the timer event when delivery fails transiently

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] KVM: x86/xen: xen_timer_callback() zeroes timer_expires on -EINVAL
--

commit 209a694ed87ce18b640accf23332c30c7e64aa39
Author: David Woodhouse <dwmw@amazon.co.uk>

KVM: x86/xen: Don't lose the timer event when delivery fails transiently

This patch defers Xen PV timer event delivery when kvm_xen_set_evtchn()
fails transiently (-EAGAIN), preserves timer_expires while clearing
timer_pending on -EINVAL, and updates KVM_XEN_VCPU_GET_ATTR(TIMER) to
avoid racing with pending retries or losing owed timer events.

> diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
> index 8c49ec8aa4d97..8a9d6963fc05a 100644
> --- a/arch/x86/kvm/xen.c
> +++ b/arch/x86/kvm/xen.c

[ ... ]

> @@ -126,14 +139,47 @@ void kvm_xen_inject_timer_irqs(struct kvm_vcpu *vcpu)
>  {
>  	if (atomic_read(&vcpu->arch.xen.timer_pending) > 0) {
>  		struct kvm_xen_evtchn e;
> +		int rc;
>  
>  		e.vcpu_id = vcpu->vcpu_id;
>  		e.vcpu_idx = vcpu->vcpu_idx;
>  		e.port = vcpu->arch.xen.timer_virq;
>  		e.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
>  
> -		kvm_xen_set_evtchn(&e, vcpu->kvm);
> +		rc = kvm_xen_set_evtchn(&e, vcpu->kvm);
> +
> +		/*
> +		 * Leave timer_pending set if the memslot is being updated
> +		 * (-EAGAIN), and raise KVM_REQ_UNBLOCK — as the timer
> +		 * callback does when it defers — so that the entry is
> +		 * aborted and this retried from vcpu_run(), rather than
> +		 * the guest running with the owed event undelivered until
> +		 * some unrelated VM exit.
> +		 */
> +		if (rc == -EAGAIN) {
> +			kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
> +			return;
> +		}
> +
> +		/*
> +		 * If the event cannot currently be delivered (-EINVAL:
> +		 * deactivated shared_info, or an out-of-range port), clear
> +		 * timer_pending — a halted vCPU would otherwise spin on
> +		 * kvm_cpu_has_pending_timer() — but keep timer_expires,
> +		 * which still reports the event as owed through
> +		 * KVM_XEN_VCPU_GET_ATTR; re-arming the timer with an expiry
> +		 * in the past fires it again.
> +		 */
> +		if (rc == -EINVAL) {
> +			atomic_set(&vcpu->arch.xen.timer_pending, 0);
> +			return;
> +		}
>  
> +		/*
> +		 * Delivered - or failed in a way which can never succeed
> +		 * (-EFAULT: a genuinely unbacked shared_info gfn). Either
> +		 * way the event is complete: drop it rather than spin.
> +		 */
>  		vcpu->arch.xen.timer_expires = 0;
>  		atomic_set(&vcpu->arch.xen.timer_pending, 0);
>  	}
> @@ -1315,17 +1361,21 @@ int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
>  		r = 0;
>  		break;
>  
> -	case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
> +	case KVM_XEN_VCPU_ATTR_TYPE_TIMER: {
> +		bool timer_armed = false;
> +
>  		/*
> -		 * Ensure a consistent snapshot of state is captured, with a
> -		 * timer either being pending, or the event channel delivered
> -		 * to the corresponding bit in the shared_info. Not still
> -		 * lurking in the timer_pending flag for deferred delivery.
> -		 * Purely as an optimisation, if the timer_expires field is
> -		 * zero, that means the timer isn't active (or even in the
> -		 * timer_pending flag) and there is no need to cancel it.
> +		 * Ensure a consistent snapshot of state is captured: a timer
> +		 * which already fired has either been delivered to the
> +		 * corresponding bit in the shared_info, or is reported as
> +		 * still owed (expires_ns != 0, if the shared_info page was
> +		 * unavailable). Purely as an optimisation, if the
> +		 * timer_expires field is zero, that means the timer isn't
> +		 * active (or even in the timer_pending flag) and there is
> +		 * no need to cancel it.
>  		 */
>  		if (vcpu->arch.xen.timer_expires) {
> +			timer_armed = hrtimer_active(&vcpu->arch.xen.timer);
>  			hrtimer_cancel(&vcpu->arch.xen.timer);
>  			kvm_xen_inject_timer_irqs(vcpu);
>  		}
> @@ -1341,13 +1391,34 @@ int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
>  		 * That's fine, as the guest won't even have had a chance
>  		 * to run and handle the interrupt. Asserting an already
>  		 * pending event channel is idempotent.
> +		 *
> +		 * Don't re-arm if the flush above failed transiently and
> +		 * left the timer pending: delivery is owned by the retry
> +		 * on the guest entry path, and an immediate hrtimer fire
> +		 * would race it — the callback's timer_pending suppression
> +		 * check is not serialized against the entry path, so a
> +		 * late callback could re-assert a port the entry path had
> +		 * already delivered and the guest already consumed.
> +		 *
> +		 * An expiry already in the past fires immediately: the
> +		 * event is still owed.
> +		 *
> +		 * A timer still armed (not yet fired) across the flush is
> +		 * always restored, whatever its port. One which already
> +		 * fired is not re-fired if its port is out of range, where
> +		 * the immediate -EINVAL from the delivery path would destroy
> +		 * the timer_expires state just reported to userspace.
>  		 */
> -		if (vcpu->arch.xen.timer_expires)
> -			hrtimer_start_expires(&vcpu->arch.xen.timer,
> -					      HRTIMER_MODE_ABS_HARD);
> +		if (vcpu->arch.xen.timer_expires &&
> +		    !atomic_read(&vcpu->arch.xen.timer_pending) &&
> +		    (timer_armed ||
> +		     vcpu->arch.xen.timer_virq < kvm_max_evtchn_port(vcpu->kvm)))
> +			kvm_xen_start_timer(vcpu, vcpu->arch.xen.timer_expires,
> +					    false);

[Severity: Medium]
Should xen_timer_callback() also preserve vcpu->arch.xen.timer_expires
when kvm_xen_set_evtchn_fast() returns -EINVAL?

In kvm_xen_inject_timer_irqs(), -EINVAL from kvm_xen_set_evtchn() now
leaves vcpu->arch.xen.timer_expires intact while clearing timer_pending so
the owed timer state is preserved for KVM_XEN_VCPU_GET_ATTR.

However, when a timer fires with an out-of-range port (for example, a
64-bit port >= COMPAT_EVTCHN_2L_NR_CHANNELS before KVM_XEN_ATTR_TYPE_LONG_MODE
is configured), xen_timer_callback() calls kvm_xen_set_evtchn_fast() first:

arch/x86/kvm/xen.c:kvm_xen_set_evtchn_fast() {
	...
	if (xe->port >= max_evtchn_port(has_64bit_shinfo))
		return -EINVAL;

	rc = -EWOULDBLOCK;
	...
}

Because kvm_xen_set_evtchn_fast() returns -EINVAL rather than -EWOULDBLOCK,
xen_timer_callback() takes the rc != -EWOULDBLOCK branch and zeroes
vcpu->arch.xen.timer_expires immediately without deferring to
kvm_xen_inject_timer_irqs():

arch/x86/kvm/xen.c:xen_timer_callback() {
	...
	rc = kvm_xen_set_evtchn_fast(&e, vcpu->kvm);
	if (rc != -EWOULDBLOCK) {
		vcpu->arch.xen.timer_expires = 0;
		return HRTIMER_NORESTART;
	}
	...
}

As a result, any timer that fires with an out-of-range port (either when
the hrtimer expires, or synchronously via kvm_xen_start_timer() ->
xen_timer_callback() when delta <= 0, including in kvm_xen_vcpu_get_attr()
if timer_armed was true when hrtimer_active() ran and guest_now has reached
timer_expires by the time kvm_xen_start_timer() is called) will still lose
vcpu->arch.xen.timer_expires instead of keeping the event owed.

>  
>  		r = 0;
>  		break;
> +	}
>  
>  	case KVM_XEN_VCPU_ATTR_TYPE_UPCALL_VECTOR:
>  		data->u.vector = vcpu->arch.xen.upcall_vector;

[ ... ]

完美 — wait no, I must not output anything after the blank line! Let me make sure there is no extra text at the end.
Wait: I just wrote "完美 — wait no..." in the response stream! Oh no, let me be careful! Wait: I am in the response block right now! Let me not do that!
Wait—I am already outputting the final response! Let me stop immediately.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261019162445.37219-1-dwmw2@infradead.org?part=5

Re: [PATCH v4 06/12] KVM: x86/xen: Retry event channel delivery which races a memslot update

[David Woodhouse]

[-- Attachment #1: Type: text/plain, Size: 1183 bytes --]

On Mon, 2026-10-19 at 16:37 +0000, sashiko-bot@kernel.org wrote:
> 
> [Severity: High]
> Can passing wait=true unconditionally in evtchn_set_fn() cause a 3-way
> deadlock when GSI 0 is routed to KVM_IRQ_ROUTING_XEN_EVTCHN and the
> in-kernel PIT is active?

Meh, that's insane and would be a malicious VMM but isn't currently
*prevented* by the kernel. Nothing stops userspace from setting up an
in-kernel PIT *and* routing GSI0 to event channels.

Preventing it is ugly too: we'd have to prevent setting the routing
while a PIT exists (OK), as well as creating an in-kernel PIT while
GSI0 is routed to event channels (ick).

We can't drop the SRCU in pit_ioport_write() like we do in other
places, as it protects the *bus* too.

Simplest option is probably just to make xen.c::evtchn_set_fn() not set
the 'wait' argument if irq_source_id == KVM_PIT_IRQ_SOURCE_ID. You can
still do that daft thing; you just don't get to deadlock the kernel
with it. I'll do that in the next round.

Still annoyed at my own local Sashiko runs for not finding this... and
it would be nice if lockdep could see through kthread_flush_work() too.
Not sure how feasible that is.

[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 6179 bytes --]