All of lore.kernel.org
 help / color / mirror / Atom feed
From: Jim Mattson <jmattson@google.com>
To: Manali Shukla <manali.shukla@amd.com>
Cc: Jim Mattson <jmattson@google.com>,
	seanjc@google.com, pbonzini@redhat.com,  mingo@redhat.com,
	bp@alien8.de, kvm@vger.kernel.org, x86@kernel.org,
	 santosh.shukla@amd.com, nikunj.dadhania@amd.com,
	Naveen.Rao@amd.com,  dapeng1.mi@linux.intel.com,
	ravi.bangoria@amd.com, peterz@infradead.org,
	 Sandipan.Das@amd.com, Yosry Ahmed <yosry@kernel.org>,
	linux-perf-users@vger.kernel.org
Subject: Re: [PATCH v3 1/9] perf/amd/ibs: Fix race condition in IBS
Date: Thu,  8 Oct 2026 10:19:18 -0700	[thread overview]
Message-ID: <20261008171921.668587-1-jmattson@google.com> (raw)
In-Reply-To: <20260310060022.15120-2-manali.shukla@amd.com>

On Tue, Mar 10, 2026 at 06:00:13AM +0000, Manali Shukla wrote:
> Consider the following scenario,
>
> While scheduling out an IBS event from perf's core scheduling path,
> event_sched_out() disables the IBS event by clearing the IBS enable
> bit in perf_ibs_disable_event(). However, if a delayed IBS NMI is
> delivered after the IBS enable bit is cleared, the IBS NMI handler
> may still observe the valid bit set and incorrectly treat the sample
> as valid.

The sample is valid. It was collected while the event was active, and
it is correct to record it. The bug is only that the handler re-arms
the hardware after perf_ibs_stop() disables the event.

> As a result, it re-enables IBS by setting the enable bit,
> even though the event has already been scheduled out.
>
> This leads to a situation where IBS is re-enabled after being
> explicitly disabled, which is incorrect. Although this race does not
> have visible side effects, it violates the expected behavior of the
> perf subsystem.

This race does have visible side effects:

  1. When the delayed NMI arrives before perf_ibs_stop() clears
     IBS_STARTED, the handler takes the normal path (not the fail:
     path), leaves IBS_STOPPED set, and re-arms the hardware for one
     more period. When that extra period overflows after
     perf_ibs_stop() clears IBS_STARTED, a second NMI arrives. If an
     unrelated NMI arrives first, the IBS handler takes the fail:
     path, clears IBS_STOPPED, and claims that NMI. The second IBS
     NMI is then unhandled ("Uhhuh. NMI received for unknown
     reason").
  2. With VIBS enabled, on hardware without IBS_CAPS_DIS, if this
     race happens when perf schedules out a host IBS event before
     VMRUN, IbsFetchEn or IbsOpEn is 1 at VMRUN. APM vol. 2, section
     15.38, says that these bits must be 0 at VMRUN of an SEV-ES or
     SEV-SNP guest with IBS virtualization enabled.

> The race is particularly noticeable when userspace repeatedly disables
> and re-enables IBS using PERF_EVENT_IOC_DISABLE and
> PERF_EVENT_IOC_ENABLE ioctls in a loop.
>
> Fix this by checking the IBS_STOPPING bit in the IBS NMI handler before
> re-enabling the IBS event. If the IBS_STOPPING bit is set, it indicates
> that the event is either disabled or in the process of being disabled,
> and the NMI handler should not re-enable it.
>
> Signed-off-by: Manali Shukla <manali.shukla@amd.com>

I think this warrants a Fixes tag:

Fixes: 85dc600263c2 ("perf/x86/amd/ibs: Fix pmu::stop() nesting")

This fix does not depend on VIBS. It is probably better to send it
separately through tip/perf:core, so that it can go in before the
rest of this series.

> ---
>  arch/x86/events/amd/ibs.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
> index eeb607b84dda..09b56bab510a 100644
> --- a/arch/x86/events/amd/ibs.c
> +++ b/arch/x86/events/amd/ibs.c
> @@ -1582,7 +1582,8 @@ static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
>  		}
>  		new_config |= period >> 4;
>
> -		perf_ibs_enable_event(perf_ibs, hwc, new_config);
> +		if (!test_bit(IBS_STOPPING, pcpu->state))
> +			perf_ibs_enable_event(perf_ibs, hwc, new_config);

This stops the late re-arm. perf_ibs_stop() sets IBS_STOPPING first,
with test_and_set_bit(). An NMI before that point can re-arm the
hardware, but perf_ibs_stop() then disables the hardware. An NMI
after that point does not re-arm. Both sides run on the same CPU, so
a plain test_bit() in NMI context is sufficient.

However, the skipped re-arm causes three problems when an NMI arrives
after perf_ibs_stop() sets IBS_STOPPING and before it clears
IBS_STARTED.

First, the event count can increase twice for the same sample:

  1. The handler calls perf_ibs_event_update() for the sample and
     adds a full period. perf_ibs_set_period() sets prev_count to 0.
  2. Because of this patch, the handler does not re-arm. CTL (and
     perf_ibs_stop()'s local config copy, if already read) still
     holds the old sample with Val=1, and the handler does not set
     PERF_HES_UPTODATE.
  3. perf_ibs_stop() clears Val in its config copy and calls
     perf_ibs_event_update() again, because PERF_HES_UPTODATE is
     clear. Because prev_count is now 0, the delta is the whole
     count field: CurCnt (op) or FetchCnt (fetch) from the old
     sample, as if it were progress in a new period.

The amount added in step 3 depends on what the hardware leaves in the
count fields after a sample. For op, the comment in
get_ibs_op_count() says that the lower 7 bits of CurCnt are
randomized after a rollover, so the amount is in general not zero.

Second, IBS_STOPPED can stay set in pcpu->state. perf_ibs_stop() sets
IBS_STOPPED so that a late NMI can clear it at the fail: label. When
the NMI instead arrives before perf_ibs_stop() clears IBS_STARTED,
the handler takes the normal path and does not clear IBS_STOPPED.
Before this patch, the re-armed period gave a second NMI that cleared
it. Now that the handler does not re-arm, IBS_STOPPED can stay set
and falsely claim a later unrelated NMI.

Third, the same sample can be recorded twice. After the handler skips
the re-arm, CTL still holds the old sample with Val=1, and
IBS_STARTED is still set. This is true at least until
perf_ibs_stop() calls perf_ibs_disable_event(). With IBS_CAPS_DIS,
that call writes only CTL2, so it stays true until perf_ibs_stop()
clears IBS_STARTED. If an unrelated NMI arrives in this window, the
IBS handler takes the normal path again, records the same sample a
second time, and adds another full period, because prev_count is 0.
Before this patch, the re-arm cleared Val, so this could not happen.

The throttle path (throttle != 0, so the handler does not re-arm) has
the first two problems already, when perf_event_overflow() calls
pmu::stop() from the NMI. They are not caused by this patch, but they
may be worth a look at the same time.

>  	}
>
>  	perf_event_update_userpage(event);

  reply	other threads:[~2026-10-08 17:19 UTC|newest]

Thread overview: 29+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-03-10  6:00 [PATCH v3 0/9] Implement support for IBS virtualization Manali Shukla
2026-03-10  6:00 ` [PATCH v3 1/9] perf/amd/ibs: Fix race condition in IBS Manali Shukla
2026-10-08 17:19   ` Jim Mattson [this message]
2026-03-10  6:00 ` [PATCH v3 2/9] x86/cpufeatures: Add CPUID feature bit for VIBS in SVM/SEV guests Manali Shukla
2026-10-08 17:28   ` Jim Mattson
2026-03-10  6:00 ` [PATCH v3 3/9] KVM: x86/cpuid: Add a KVM-only leaf for IBS capabilities Manali Shukla
2026-10-08 17:44   ` Jim Mattson
2026-10-22 10:16     ` Manali Shukla
2026-03-10  6:00 ` [PATCH v3 4/9] KVM: x86: Extend CPUID range to include new leaf Manali Shukla
2026-10-08 17:59   ` Jim Mattson
2026-10-22 12:00     ` Manali Shukla
2026-10-22 17:15       ` Jim Mattson
2026-03-10  6:00 ` [PATCH v3 5/9] KVM: SVM: Extend VMCB area for virtualized IBS registers Manali Shukla
2026-10-08 18:01   ` Jim Mattson
2026-03-10  6:00 ` [PATCH v3 6/9] KVM: SVM: Add support for IBS Virtualization Manali Shukla
2026-10-08 19:19   ` Jim Mattson
2026-10-08 21:28     ` Jim Mattson
2026-10-16 16:01       ` Sean Christopherson
2026-11-10 12:03         ` Manali Shukla
2026-11-10 11:30     ` Manali Shukla
2026-03-10  6:00 ` [PATCH v3 7/9] perf/x86/amd: Enable VPMU passthrough capability for IBS PMU Manali Shukla
2026-10-08 19:32   ` Jim Mattson
2026-11-18 13:39     ` Manali Shukla
2026-03-10  6:00 ` [PATCH v3 8/9] perf/x86/amd: Remove exclude_guest check from perf_ibs_init() Manali Shukla
2026-10-08 19:38   ` Jim Mattson
2026-11-18 13:38     ` Manali Shukla
2026-03-10  6:00 ` [PATCH v3 9/9] KVM: SVM: Add newly added IBS capabilities and MSRs Manali Shukla
2026-10-08 21:03   ` Jim Mattson
2026-11-11 12:11     ` Manali Shukla

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20261008171921.668587-1-jmattson@google.com \
    --to=jmattson@google.com \
    --cc=Naveen.Rao@amd.com \
    --cc=Sandipan.Das@amd.com \
    --cc=bp@alien8.de \
    --cc=dapeng1.mi@linux.intel.com \
    --cc=kvm@vger.kernel.org \
    --cc=linux-perf-users@vger.kernel.org \
    --cc=manali.shukla@amd.com \
    --cc=mingo@redhat.com \
    --cc=nikunj.dadhania@amd.com \
    --cc=pbonzini@redhat.com \
    --cc=peterz@infradead.org \
    --cc=ravi.bangoria@amd.com \
    --cc=santosh.shukla@amd.com \
    --cc=seanjc@google.com \
    --cc=x86@kernel.org \
    --cc=yosry@kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.