linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: "tip-bot2 for Suravee Suthikulpanit" <tip-bot2@linutronix.de>
To: linux-tip-commits@vger.kernel.org
Cc: Alexander Monakov <amonakov@ispras.ru>,
	Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>,
	"Peter Zijlstra (Intel)" <peterz@infradead.org>,
	x86@kernel.org, linux-kernel@vger.kernel.org
Subject: [tip: perf/urgent] x86/events/amd/iommu: Fix invalid Perf result due to IOMMU PMC power-gating
Date: Thu, 06 May 2021 13:48:18 -0000	[thread overview]
Message-ID: <162030889887.29796.12426716742809018792.tip-bot2@tip-bot2> (raw)
In-Reply-To: <20210504065236.4415-1-suravee.suthikulpanit@amd.com>

The following commit has been merged into the perf/urgent branch of tip:

Commit-ID:     e10de314287c2c14b0e6f0e3e961975ce2f4a83d
Gitweb:        https://git.kernel.org/tip/e10de314287c2c14b0e6f0e3e961975ce2f4a83d
Author:        Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
AuthorDate:    Tue, 04 May 2021 01:52:36 -05:00
Committer:     Peter Zijlstra <peterz@infradead.org>
CommitterDate: Thu, 06 May 2021 15:33:37 +02:00

x86/events/amd/iommu: Fix invalid Perf result due to IOMMU PMC power-gating

On certain AMD platforms, when the IOMMU performance counter source
(csource) field is zero, power-gating for the counter is enabled, which
prevents write access and returns zero for read access.

This can cause invalid perf result especially when event multiplexing
is needed (i.e. more number of events than available counters) since
the current logic keeps track of the previously read counter value,
and subsequently re-program the counter to continue counting the event.
With power-gating enabled, we cannot gurantee successful re-programming
of the counter.

Workaround this issue by :

1. Modifying the ordering of setting/reading counters and enabing/
   disabling csources to only access the counter when the csource
   is set to non-zero.

2. Since AMD IOMMU PMU does not support interrupt mode, the logic
   can be simplified to always start counting with value zero,
   and accumulate the counter value when stopping without the need
   to keep track and reprogram the counter with the previously read
   counter value.

This has been tested on systems with and without power-gating.

Fixes: 994d6608efe4 ("iommu/amd: Remove performance counter pre-initialization test")
Suggested-by: Alexander Monakov <amonakov@ispras.ru>
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20210504065236.4415-1-suravee.suthikulpanit@amd.com
---
 arch/x86/events/amd/iommu.c | 47 +++++++++++++++++++-----------------
 1 file changed, 26 insertions(+), 21 deletions(-)

diff --git a/arch/x86/events/amd/iommu.c b/arch/x86/events/amd/iommu.c
index 6a98a76..2da6139 100644
--- a/arch/x86/events/amd/iommu.c
+++ b/arch/x86/events/amd/iommu.c
@@ -18,8 +18,6 @@
 #include "../perf_event.h"
 #include "iommu.h"
 
-#define COUNTER_SHIFT		16
-
 /* iommu pmu conf masks */
 #define GET_CSOURCE(x)     ((x)->conf & 0xFFULL)
 #define GET_DEVID(x)       (((x)->conf >> 8)  & 0xFFFFULL)
@@ -285,22 +283,31 @@ static void perf_iommu_start(struct perf_event *event, int flags)
 	WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
 	hwc->state = 0;
 
+	/*
+	 * To account for power-gating, which prevents write to
+	 * the counter, we need to enable the counter
+	 * before setting up counter register.
+	 */
+	perf_iommu_enable_event(event);
+
 	if (flags & PERF_EF_RELOAD) {
-		u64 prev_raw_count = local64_read(&hwc->prev_count);
+		u64 count = 0;
 		struct amd_iommu *iommu = perf_event_2_iommu(event);
 
+		/*
+		 * Since the IOMMU PMU only support counting mode,
+		 * the counter always start with value zero.
+		 */
 		amd_iommu_pc_set_reg(iommu, hwc->iommu_bank, hwc->iommu_cntr,
-				     IOMMU_PC_COUNTER_REG, &prev_raw_count);
+				     IOMMU_PC_COUNTER_REG, &count);
 	}
 
-	perf_iommu_enable_event(event);
 	perf_event_update_userpage(event);
-
 }
 
 static void perf_iommu_read(struct perf_event *event)
 {
-	u64 count, prev, delta;
+	u64 count;
 	struct hw_perf_event *hwc = &event->hw;
 	struct amd_iommu *iommu = perf_event_2_iommu(event);
 
@@ -311,14 +318,11 @@ static void perf_iommu_read(struct perf_event *event)
 	/* IOMMU pc counter register is only 48 bits */
 	count &= GENMASK_ULL(47, 0);
 
-	prev = local64_read(&hwc->prev_count);
-	if (local64_cmpxchg(&hwc->prev_count, prev, count) != prev)
-		return;
-
-	/* Handle 48-bit counter overflow */
-	delta = (count << COUNTER_SHIFT) - (prev << COUNTER_SHIFT);
-	delta >>= COUNTER_SHIFT;
-	local64_add(delta, &event->count);
+	/*
+	 * Since the counter always start with value zero,
+	 * simply just accumulate the count for the event.
+	 */
+	local64_add(count, &event->count);
 }
 
 static void perf_iommu_stop(struct perf_event *event, int flags)
@@ -328,15 +332,16 @@ static void perf_iommu_stop(struct perf_event *event, int flags)
 	if (hwc->state & PERF_HES_UPTODATE)
 		return;
 
+	/*
+	 * To account for power-gating, in which reading the counter would
+	 * return zero, we need to read the register before disabling.
+	 */
+	perf_iommu_read(event);
+	hwc->state |= PERF_HES_UPTODATE;
+
 	perf_iommu_disable_event(event);
 	WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
 	hwc->state |= PERF_HES_STOPPED;
-
-	if (hwc->state & PERF_HES_UPTODATE)
-		return;
-
-	perf_iommu_read(event);
-	hwc->state |= PERF_HES_UPTODATE;
 }
 
 static int perf_iommu_add(struct perf_event *event, int flags)

  parent reply	other threads:[~2021-05-06 13:48 UTC|newest]

Thread overview: 11+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-05-04  6:52 [PATCH] x86/events/amd/iommu: Fix invalid Perf result due to IOMMU PMC power-gating Suravee Suthikulpanit
2021-05-04  9:39 ` Peter Zijlstra
2021-05-04 11:58   ` Suthikulpanit, Suravee
2021-05-04 12:13     ` Peter Zijlstra
2021-05-05 12:39       ` Suthikulpanit, Suravee
2021-05-05 13:05         ` Peter Zijlstra
2021-05-10  2:08           ` Suthikulpanit, Suravee
2021-05-04 17:04 ` David Coe
2021-05-05 10:24 ` David Coe
2021-05-06 13:48 ` tip-bot2 for Suravee Suthikulpanit [this message]
2021-05-14 10:48 ` David Coe

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=162030889887.29796.12426716742809018792.tip-bot2@tip-bot2 \
    --to=tip-bot2@linutronix.de \
    --cc=amonakov@ispras.ru \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-tip-commits@vger.kernel.org \
    --cc=peterz@infradead.org \
    --cc=suravee.suthikulpanit@amd.com \
    --cc=x86@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 a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).