[PATCH] perf/amd/ibs: Clear stale IBS_{FETCH|OP}_CTL in perf_ibs_start() with CTL2[Dis]

[Jim Mattson <jmattson@google.com>]

Commit 1b044ff3c17e ("perf/amd/ibs: Avoid race between event add and
NMI") makes perf_ibs_start() reset IBS_{FETCH|OP}_CTL before it sets
IBS_STARTED. Thus, an NMI from another source that arrives before the
event is enabled finds Val=0 and does not process a stale sample or
enable the event too early.

The reset is a call to perf_ibs_disable_event() with config=0. A later
commit changed perf_ibs_disable_event():

  commit efa5700ec0da ("perf/amd/ibs: Support IBS_{FETCH|OP}_CTL2[Dis] to eliminate RMW race")

When IBS_CAPS_DIS is set, the function now writes only
IBS_{FETCH|OP}_CTL2 and returns. It does not write IBS_{FETCH|OP}_CTL.
On hardware with IBS_CAPS_DIS, the reset does nothing, and the race is
possible again: CTL still holds the previous sample, with Val=1, when
IBS_STARTED is set.

Clear IBS_{FETCH|OP}_CTL explicitly when IBS_CAPS_DIS is set, as
perf_ibs_disable_event() did before CTL2 support. Everything in CTL is
stale at this point, so it does not matter if this write discards a
Val that the hardware sets late.

Fixes: efa5700ec0da ("perf/amd/ibs: Support IBS_{FETCH|OP}_CTL2[Dis] to eliminate RMW race")
Assisted-by: LLM
Signed-off-by: Jim Mattson <jmattson@google.com>
---
I reproduced the race on hardware with IBS_CAPS_DIS. To make the window
wider, I added a debug-only udelay() between set_bit(IBS_STARTED) and
perf_ibs_enable_event() in perf_ibs_start(). A throttled ibs_op event
gave about 1000 perf_ibs_start() calls per second, and a cycles event
gave NMIs from the core PMU. In 60 seconds:

  - Without this patch, IBS_OP_CTL[Val] was 1 at every
    perf_ibs_start() call (60003 of 60003). The IBS op NMI handler took
    the stale sample and re-enabled the event during the window 17
    times.
  - With this patch, the handler never took a sample during the window.

Without the udelay(), the window is only a few instructions long.

A related question for AMD: with IBS_CAPS_DIS, perf_ibs_stop() now
leaves IBS_{FETCH|OP}_CTL[En]=1. APM vol. 2, section 15.38, says that
IbsFetchEn and IbsOpEn must be 0 at VMRUN of an SEV-ES or SEV-SNP
guest with IBS virtualization enabled, and should be 0 for other
guests. Can the hardware set CTL[Val] or change CTL in any other way
after CTL2[Dis] is set to 1? If not, perf_ibs_stop() can safely clear
CTL[En] with a read-modify-write after it sets CTL2[Dis].

 arch/x86/events/amd/ibs.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
index 3531f9c23b8c..3a5475006b59 100644
--- a/arch/x86/events/amd/ibs.c
+++ b/arch/x86/events/amd/ibs.c
@@ -580,8 +580,15 @@ static void perf_ibs_start(struct perf_event *event, int flags)
 	 * Doing so prevents a race condition in which an NMI due to other
 	 * source might accidentally activate the event before we enable
 	 * it ourselves.
+	 *
+	 * With IBS_CAPS_DIS, perf_ibs_disable_event() only sets CTL2[Dis]
+	 * and leaves the previous sample in CTL, so clear CTL explicitly.
+	 * Anything in CTL is stale, so it does not matter if this write
+	 * discards a late Val.
 	 */
 	perf_ibs_disable_event(perf_ibs, hwc, 0);
+	if (ibs_caps & IBS_CAPS_DIS)
+		wrmsrq(hwc->config_base, 0);
 
 	/*
 	 * Set STARTED before enabling the hardware, such that a subsequent NMI
-- 
2.56.0.385.gd3acb90ef8-goog