From: Peter Zijlstra <peterz@infradead.org>
To: Don Zickus <dzickus@redhat.com>
Cc: dave.hansen@linux.intel.com, eranian@google.com,
ak@linux.intel.com, jmario@redhat.com,
linux-kernel@vger.kernel.org, acme@infradead.org,
mingo@kernel.org
Subject: [PATCH] perf, x86: Optimize intel_pmu_pebs_fixup_ip()
Date: Wed, 16 Oct 2013 12:57:55 +0200 [thread overview]
Message-ID: <20131016105755.GX10651@twins.programming.kicks-ass.net> (raw)
In-Reply-To: <20131015154104.GA227855@redhat.com>
A prettier patch below. The main difference is on-demand allocation of
the scratch buffer.
---
Subject: perf, x86: Optimize intel_pmu_pebs_fixup_ip()
From: Peter Zijlstra <peterz@infradead.org>
Date: Tue, 15 Oct 2013 12:14:04 +0200
On Mon, Oct 14, 2013 at 04:35:49PM -0400, Don Zickus wrote:
> While there are a few places that are causing latencies, for now I focused on
> the longest one first. It seems to be 'copy_user_from_nmi'
>
> intel_pmu_handle_irq ->
> intel_pmu_drain_pebs_nhm ->
> __intel_pmu_drain_pebs_nhm ->
> __intel_pmu_pebs_event ->
> intel_pmu_pebs_fixup_ip ->
> copy_from_user_nmi
>
> In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of
> all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles
> (there are some cases where only 10 iterations are needed to go that high
> too, but in generall over 50 or so). At this point copy_user_from_nmi
> seems to account for over 90% of the nmi latency.
So avoid having to call copy_from_user_nmi() for every instruction.
Since we already limit the max basic block size, we can easily
pre-allocate a piece of memory to copy the entire thing into in one
go.
Don reports (for a previous version):
> Your patch made a huge difference in improvement. The
> copy_from_user_nmi() no longer hits the million of cycles. I still
> have a batch of 100,000-300,000 cycles. My longest NMI paths used
> to be dominated by copy_from_user_nmi, now it is not (I have to dig
> up the new hot path).
Cc: eranian@google.com
Cc: ak@linux.intel.com
Cc: jmario@redhat.com
Cc: acme@infradead.org
Cc: dave.hansen@linux.intel.com
Reported-by: Don Zickus <dzickus@redhat.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
---
arch/x86/kernel/cpu/perf_event_intel_ds.c | 48 +++++++++++++++++++++---------
1 file changed, 34 insertions(+), 14 deletions(-)
--- a/arch/x86/kernel/cpu/perf_event_intel_ds.c
+++ b/arch/x86/kernel/cpu/perf_event_intel_ds.c
@@ -12,6 +12,7 @@
#define BTS_BUFFER_SIZE (PAGE_SIZE << 4)
#define PEBS_BUFFER_SIZE PAGE_SIZE
+#define PEBS_FIXUP_SIZE PAGE_SIZE
/*
* pebs_record_32 for p4 and core not supported
@@ -228,12 +229,14 @@ void fini_debug_store_on_cpu(int cpu)
wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
}
+static DEFINE_PER_CPU(void *, insn_buffer);
+
static int alloc_pebs_buffer(int cpu)
{
struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
int node = cpu_to_node(cpu);
int max, thresh = 1; /* always use a single PEBS record */
- void *buffer;
+ void *buffer, *ibuffer;
if (!x86_pmu.pebs)
return 0;
@@ -242,6 +245,15 @@ static int alloc_pebs_buffer(int cpu)
if (unlikely(!buffer))
return -ENOMEM;
+ if (x86_pmu.intel_cap.pebs_format < 2) {
+ ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
+ if (!ibuffer) {
+ kfree(buffer);
+ return -ENOMEM;
+ }
+ per_cpu(insn_buffer, cpu) = ibuffer;
+ }
+
max = PEBS_BUFFER_SIZE / x86_pmu.pebs_record_size;
ds->pebs_buffer_base = (u64)(unsigned long)buffer;
@@ -262,6 +274,9 @@ static void release_pebs_buffer(int cpu)
if (!ds || !x86_pmu.pebs)
return;
+ kfree(per_cpu(insn_buffer, cpu));
+ per_cpu(insn_buffer, cpu) = NULL;
+
kfree((void *)(unsigned long)ds->pebs_buffer_base);
ds->pebs_buffer_base = 0;
}
@@ -729,6 +744,7 @@ static int intel_pmu_pebs_fixup_ip(struc
unsigned long old_to, to = cpuc->lbr_entries[0].to;
unsigned long ip = regs->ip;
int is_64bit = 0;
+ void *kaddr;
/*
* We don't need to fixup if the PEBS assist is fault like
@@ -752,7 +768,7 @@ static int intel_pmu_pebs_fixup_ip(struc
* unsigned math, either ip is before the start (impossible) or
* the basic block is larger than 1 page (sanity)
*/
- if ((ip - to) > PAGE_SIZE)
+ if ((ip - to) > PEBS_FIXUP_SIZE)
return 0;
/*
@@ -763,29 +779,33 @@ static int intel_pmu_pebs_fixup_ip(struc
return 1;
}
+ if (!kernel_ip(ip)) {
+ int size, bytes;
+ u8 *buf = this_cpu_ptr(insn_buffer);
+
+ size = ip - to; /* Must fit our buffer, see above */
+ bytes = copy_from_user_nmi(buf, (void __user *)to, size);
+ if (bytes != size)
+ return 0;
+
+ kaddr = buf;
+ } else {
+ kaddr = (void *)to;
+ }
+
do {
struct insn insn;
- u8 buf[MAX_INSN_SIZE];
- void *kaddr;
old_to = to;
- if (!kernel_ip(ip)) {
- int bytes, size = MAX_INSN_SIZE;
-
- bytes = copy_from_user_nmi(buf, (void __user *)to, size);
- if (bytes != size)
- return 0;
-
- kaddr = buf;
- } else
- kaddr = (void *)to;
#ifdef CONFIG_X86_64
is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
#endif
insn_init(&insn, kaddr, is_64bit);
insn_get_length(&insn);
+
to += insn.length;
+ kaddr += insn.length;
} while (to < ip);
if (to == ip) {
next prev parent reply other threads:[~2013-10-16 10:58 UTC|newest]
Thread overview: 47+ messages / expand[flat|nested] mbox.gz Atom feed top
2013-10-14 20:35 x86, perf: throttling issues with long nmi latencies Don Zickus
2013-10-14 21:28 ` Andi Kleen
2013-10-15 10:14 ` Peter Zijlstra
2013-10-15 13:02 ` Peter Zijlstra
2013-10-15 14:32 ` Peter Zijlstra
2013-10-15 15:07 ` Peter Zijlstra
2013-10-15 15:41 ` Don Zickus
2013-10-16 10:57 ` Peter Zijlstra [this message]
2013-10-16 12:46 ` [PATCH] perf, x86: Optimize intel_pmu_pebs_fixup_ip() Don Zickus
2013-10-16 13:31 ` Peter Zijlstra
2013-10-16 13:54 ` Don Zickus
2013-10-17 11:21 ` Peter Zijlstra
2013-10-17 13:33 ` Peter Zijlstra
2013-10-29 14:07 ` [tip:perf/urgent] perf/x86: Fix NMI measurements tip-bot for Peter Zijlstra
2013-10-16 20:52 ` [PATCH] perf, x86: Optimize intel_pmu_pebs_fixup_ip() Andi Kleen
2013-10-16 21:03 ` Peter Zijlstra
2013-10-16 23:07 ` Peter Zijlstra
2013-10-17 9:41 ` Peter Zijlstra
2013-10-17 16:00 ` Don Zickus
2013-10-17 16:04 ` Don Zickus
2013-10-17 16:30 ` Peter Zijlstra
2013-10-17 18:26 ` Linus Torvalds
2013-10-17 21:08 ` Peter Zijlstra
2013-10-17 21:11 ` Peter Zijlstra
2013-10-17 22:01 ` Peter Zijlstra
2013-10-17 22:27 ` Linus Torvalds
2013-10-22 21:12 ` Peter Zijlstra
2013-10-23 7:09 ` Linus Torvalds
2013-10-23 20:48 ` Peter Zijlstra
2013-10-24 10:52 ` Peter Zijlstra
2013-10-24 13:47 ` Don Zickus
2013-10-24 14:06 ` Peter Zijlstra
2013-10-25 16:33 ` Don Zickus
2013-10-25 17:03 ` Peter Zijlstra
2013-10-26 10:36 ` Ingo Molnar
2013-10-28 13:19 ` Don Zickus
2013-10-29 14:08 ` [tip:perf/core] perf/x86: Further optimize copy_from_user_nmi() tip-bot for Peter Zijlstra
2013-10-23 7:44 ` [PATCH] perf, x86: Optimize intel_pmu_pebs_fixup_ip() Ingo Molnar
2013-10-17 14:49 ` Don Zickus
2013-10-17 14:51 ` Peter Zijlstra
2013-10-17 15:03 ` Don Zickus
2013-10-17 15:09 ` Peter Zijlstra
2013-10-17 15:11 ` Peter Zijlstra
2013-10-17 16:50 ` [tip:perf/core] perf/x86: " tip-bot for Peter Zijlstra
2013-10-15 16:22 ` x86, perf: throttling issues with long nmi latencies Don Zickus
2013-10-15 14:36 ` Don Zickus
2013-10-15 14:39 ` Peter Zijlstra
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=20131016105755.GX10651@twins.programming.kicks-ass.net \
--to=peterz@infradead.org \
--cc=acme@infradead.org \
--cc=ak@linux.intel.com \
--cc=dave.hansen@linux.intel.com \
--cc=dzickus@redhat.com \
--cc=eranian@google.com \
--cc=jmario@redhat.com \
--cc=linux-kernel@vger.kernel.org \
--cc=mingo@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).