Add perf core PMU support for future Intel Skylake CPU cores.
The code is based on Haswell/Broadwell.
There is a new cache event list, based on the updated Haswell
event list.
Skylake has removed most counter constraints on basic
events, so the basic constraints table now only has a single
entry (plus the fixed counters).
TSX support and various other setups are all shared with Haswell.
Skylake has 32 LBR entries. Add a new LBR init function
to set this up. The filters are all the same as Haswell.
It also has a new LBR format with a separate LBR_INFO_* MSR,
but that has been already added earlier.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: eranian@google.com
Link: http://lkml.kernel.org/r/1431285767-27027-7-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Add support for the new LBRv5 format used on Intel Skylake CPUs.
The flags for mispredict, abort, in_tx etc. moved to range of separate
LBR_INFO_* MSRs. Teach the LBR code to read those. The original
LBR registers stay the same, except they have full sign
extension now.
LBR_INFO also reports a cycle count to the last branch.
Report the cycle information using the new "cycles" branch_info
output field.
In addition we have to context switch and clear the new INFO
MSRs to avoid any information leaks.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: eranian@google.com
Link: http://lkml.kernel.org/r/1431285767-27027-6-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull perf fixes from Ingo Molnar:
"These are the left over fixes from the v4.1 cycle"
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf tools: Fix build breakage if prefix= is specified
perf/x86: Honor the architectural performance monitoring version
perf/x86/intel: Fix PMI handling for Intel PT
perf/x86/intel/bts: Fix DS area sharing with x86_pmu events
perf/x86: Add more Broadwell model numbers
perf: Fix ring_buffer_attach() RCU sync, again
PEBS always had the capability to log samples to its buffers without
an interrupt. Traditionally perf has not used this but always set the
PEBS threshold to one.
For frequently occurring events (like cycles or branches or load/store)
this in term requires using a relatively high sampling period to avoid
overloading the system, by only processing PMIs. This in term increases
sampling error.
For the common cases we still need to use the PMI because the PEBS
hardware has various limitations. The biggest one is that it can not
supply a callgraph. It also requires setting a fixed period, as the
hardware does not support adaptive period. Another issue is that it
cannot supply a time stamp and some other options. To supply a TID it
requires flushing on context switch. It can however supply the IP, the
load/store address, TSX information, registers, and some other things.
So we can make PEBS work for some specific cases, basically as long as
you can do without a callgraph and can set the period you can use this
new PEBS mode.
The main benefit is the ability to support much lower sampling period
(down to -c 1000) without extensive overhead.
One use cases is for example to increase the resolution of the c2c tool.
Another is double checking when you suspect the standard sampling has
too much sampling error.
Some numbers on the overhead, using cycle soak, comparing the elapsed
time from "kernbench -M -H" between plain (threshold set to one) and
multi (large threshold).
The test command for plain:
"perf record --time -e cycles:p -c $period -- kernbench -M -H"
The test command for multi:
"perf record --no-time -e cycles:p -c $period -- kernbench -M -H"
( The only difference of test command between multi and plain is time
stamp options. Since time stamp is not supported by large PEBS
threshold, it can be used as a flag to indicate if large threshold is
enabled during the test. )
period plain(Sec) multi(Sec) Delta
10003 32.7 16.5 16.2
20003 30.2 16.2 14.0
40003 18.6 14.1 4.5
80003 16.8 14.6 2.2
100003 16.9 14.1 2.8
800003 15.4 15.7 -0.3
1000003 15.3 15.2 0.2
2000003 15.3 15.1 0.1
With periods below 100003, plain (threshold one) cause much more
overhead. With 10003 sampling period, the Elapsed Time for multi is
even 2X faster than plain.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: acme@infradead.org
Cc: eranian@google.com
Link: http://lkml.kernel.org/r/1430940834-8964-5-git-send-email-kan.liang@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
For some obscure reason intel_{start,stop}_scheduling() copy the HT
state to an intermediate array. This would make sense if we ever were
to make changes to it which we'd have to discard.
Except we don't. By the time we call intel_commit_scheduling() we're;
as the name implies; committed to them. We'll never back out.
A further hint its pointless is that stop_scheduling() unconditionally
publishes the state.
So the intermediate array is pointless, modify the state in place and
kill the extra array.
And remove the pointless array initialization: INTEL_EXCL_UNUSED == 0.
Note; all is serialized by intel_excl_cntr::lock.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The (SNB/IVB/HSW) HT bug only affects events that can be programmed
onto GP counters, therefore we should only limit the number of GP
counters that can be used per cpu -- iow we should not constrain the
FP counters.
Furthermore, we should only enfore such a limit when there are in fact
exclusive events being scheduled on either sibling.
Reported-by: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
[ Fixed build fail for the !CONFIG_CPU_SUP_INTEL case. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Commit 43b4578071 ("perf/x86: Reduce stack usage of
x86_schedule_events()") violated the rule that 'fake' scheduling; as
used for event/group validation; should not change the event state.
This went mostly un-noticed because repeated calls of
x86_pmu::get_event_constraints() would give the same result. And
x86_pmu::put_event_constraints() would mostly not do anything.
Commit e979121b1b ("perf/x86/intel: Implement cross-HT corruption
bug workaround") made the situation much worse by actually setting the
event->hw.constraint value to NULL, so when validation and actual
scheduling interact we get NULL ptr derefs.
Fix it by removing the constraint pointer from the event and move it
back to an array, this time in cpuc instead of on the stack.
validate_group()
x86_schedule_events()
event->hw.constraint = c; # store
<context switch>
perf_task_event_sched_in()
...
x86_schedule_events();
event->hw.constraint = c2; # store
...
put_event_constraints(event); # assume failure to schedule
intel_put_event_constraints()
event->hw.constraint = NULL;
<context switch end>
c = event->hw.constraint; # read -> NULL
if (!test_bit(hwc->idx, c->idxmsk)) # <- *BOOM* NULL deref
This in particular is possible when the event in question is a
cpu-wide event and group-leader, where the validate_group() tries to
add an event to the group.
Reported-by: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Hunter <ahh@google.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Fixes: 43b4578071 ("perf/x86: Reduce stack usage of x86_schedule_events()")
Fixes: e979121b1b ("perf/x86/intel: Implement cross-HT corruption bug workaround")
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The perf PMI currently does unnecessary MSR accesses when
LBRs are enabled. We use LBR freezing, or when in callstack
mode force the LBRs to only filter on ring 3.
So there is no need to disable the LBRs explicitely in the
PMI handler.
Also we always unnecessarily rewrite LBR_SELECT in the LBR
handler, even though it can never change.
5) | /* write_msr: MSR_LBR_SELECT(1c8), value 0 */
5) | /* read_msr: MSR_IA32_DEBUGCTLMSR(1d9), value 1801 */
5) | /* write_msr: MSR_IA32_DEBUGCTLMSR(1d9), value 1801 */
5) | /* write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value 70000000f */
5) | /* write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value 0 */
5) | /* write_msr: MSR_LBR_SELECT(1c8), value 0 */
5) | /* read_msr: MSR_IA32_DEBUGCTLMSR(1d9), value 1801 */
5) | /* write_msr: MSR_IA32_DEBUGCTLMSR(1d9), value 1801 */
This patch:
- Avoids disabling already frozen LBRs unnecessarily in the PMI
- Avoids changing LBR_SELECT in the PMI
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: eranian@google.com
Link: http://lkml.kernel.org/r/1426871484-21285-1-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch disables the PMU HT bug when Hyperthreading (HT)
is disabled. We cannot do this test immediately when perf_events
is initialized. We need to wait until the topology information
is setup properly. As such, we register a later initcall, check
the topology and potentially disable the workaround. To do this,
we need to ensure there is no user of the PMU. At this point of
the boot, the only user is the NMI watchdog, thus we disable
it during the switch and re-enable it right after.
Having the workaround disabled when it is not needed provides
some benefits by limiting the overhead is time and space.
The workaround still ensures correct scheduling of the corrupting
memory events (0xd0, 0xd1, 0xd2) when HT is off. Those events
can only be measured on counters 0-3. Something else the current
kernel did not handle correctly.
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: bp@alien8.de
Cc: jolsa@redhat.com
Cc: kan.liang@intel.com
Cc: maria.n.dimakopoulou@gmail.com
Link: http://lkml.kernel.org/r/1416251225-17721-13-git-send-email-eranian@google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch limits the number of counters available to each CPU when
the HT bug workaround is enabled.
This is necessary to avoid situation of counter starvation. Such can
arise from configuration where one HT thread, HT0, is using all 4 counters
with corrupting events which require exclusion the the sibling HT, HT1.
In such case, HT1 would not be able to schedule any event until HT0
is done. To mitigate this problem, this patch artificially limits
the number of counters to 2.
That way, we can gurantee that at least 2 counters are not in exclusive
mode and therefore allow the sibling thread to schedule events of the
same type (system vs. per-thread). The 2 counters are not determined
in advance. We simply set the limit to two events per HT.
This helps mitigate starvation in case of events with specific counter
constraints such a PREC_DIST.
Note that this does not elimintate the starvation is all cases. But
it is better than not having it.
(Solution suggested by Peter Zjilstra.)
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: bp@alien8.de
Cc: jolsa@redhat.com
Cc: kan.liang@intel.com
Cc: maria.n.dimakopoulou@gmail.com
Link: http://lkml.kernel.org/r/1416251225-17721-11-git-send-email-eranian@google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch implements a software workaround for a HW erratum
on Intel SandyBridge, IvyBridge and Haswell processors
with Hyperthreading enabled. The errata are documented for
each processor in their respective specification update
documents:
- SandyBridge: BJ122
- IvyBridge: BV98
- Haswell: HSD29
The bug causes silent counter corruption across hyperthreads only
when measuring certain memory events (0xd0, 0xd1, 0xd2, 0xd3).
Counters measuring those events may leak counts to the sibling
counter. For instance, counter 0, thread 0 measuring event 0xd0,
may leak to counter 0, thread 1, regardless of the event measured
there. The size of the leak is not predictible. It all depends on
the workload and the state of each sibling hyper-thread. The
corrupting events do undercount as a consequence of the leak. The
leak is compensated automatically only when the sibling counter measures
the exact same corrupting event AND the workload is on the two threads
is the same. Given, there is no way to guarantee this, a work-around
is necessary. Furthermore, there is a serious problem if the leaked count
is added to a low-occurrence event. In that case the corruption on
the low occurrence event can be very large, e.g., orders of magnitude.
There is no HW or FW workaround for this problem.
The bug is very easy to reproduce on a loaded system.
Here is an example on a Haswell client, where CPU0, CPU4
are siblings. We load the CPUs with a simple triad app
streaming large floating-point vector. We use 0x81d0
corrupting event (MEM_UOPS_RETIRED:ALL_LOADS) and
0x20cc (ROB_MISC_EVENTS:LBR_INSERTS). Given we are not
using the LBR, the 0x20cc event should be zero.
$ taskset -c 0 triad &
$ taskset -c 4 triad &
$ perf stat -a -C 0 -e r81d0 sleep 100 &
$ perf stat -a -C 4 -r20cc sleep 10
Performance counter stats for 'system wide':
139 277 291 r20cc
10,000969126 seconds time elapsed
In this example, 0x81d0 and r20cc ar eusing sinling counters
on CPU0 and CPU4. 0x81d0 leaks into 0x20cc and corrupts it
from 0 to 139 millions occurrences.
This patch provides a software workaround to this problem by modifying the
way events are scheduled onto counters by the kernel. The patch forces
cross-thread mutual exclusion between counters in case a corrupting event
is measured by one of the hyper-threads. If thread 0, counter 0 is measuring
event 0xd0, then nothing can be measured on counter 0, thread 1. If no corrupting
event is measured on any hyper-thread, event scheduling proceeds as before.
The same example run with the workaround enabled, yield the correct answer:
$ taskset -c 0 triad &
$ taskset -c 4 triad &
$ perf stat -a -C 0 -e r81d0 sleep 100 &
$ perf stat -a -C 4 -r20cc sleep 10
Performance counter stats for 'system wide':
0 r20cc
10,000969126 seconds time elapsed
The patch does provide correctness for all non-corrupting events. It does not
"repatriate" the leaked counts back to the leaking counter. This is planned
for a second patch series. This patch series makes this repatriation more
easy by guaranteeing the sibling counter is not measuring any useful event.
The patch introduces dynamic constraints for events. That means that events which
did not have constraints, i.e., could be measured on any counters, may now be
constrained to a subset of the counters depending on what is going on the sibling
thread. The algorithm is similar to a cache coherency protocol. We call it XSU
in reference to Exclusive, Shared, Unused, the 3 possible states of a PMU
counter.
As a consequence of the workaround, users may see an increased amount of event
multiplexing, even in situtations where there are fewer events than counters
measured on a CPU.
Patch has been tested on all three impacted processors. Note that when
HT is off, there is no corruption. However, the workaround is still enabled,
yet not costing too much. Adding a dynamic detection of HT on turned out to
be complex are requiring too much to code to be justified.
This patch addresses the issue when PEBS is not used. A subsequent patch
fixes the problem when PEBS is used.
Signed-off-by: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com>
[spinlock_t -> raw_spinlock_t]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Stephane Eranian <eranian@google.com>
Cc: bp@alien8.de
Cc: jolsa@redhat.com
Cc: kan.liang@intel.com
Link: http://lkml.kernel.org/r/1416251225-17721-7-git-send-email-eranian@google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
On Broadwell INST_RETIRED.ALL cannot be used with any period
that doesn't have the lowest 6 bits cleared. And the period
should not be smaller than 128.
This is erratum BDM11 and BDM55:
http://www.intel.com/content/dam/www/public/us/en/documents/specification-updates/5th-gen-core-family-spec-update.pdf
BDM11: When using a period < 100; we may get incorrect PEBS/PMI
interrupts and/or an invalid counter state.
BDM55: When bit0-5 of the period are !0 we may get redundant PEBS
records on overflow.
Add a new callback to enforce this, and set it for Broadwell.
How does this handle the case when an app requests a specific
period with some of the bottom bits set?
Short answer:
Any useful instruction sampling period needs to be 4-6 orders
of magnitude larger than 128, as an PMI every 128 instructions
would instantly overwhelm the system and be throttled.
So the +-64 error from this is really small compared to the
period, much smaller than normal system jitter.
Long answer (by Peterz):
IFF we guarantee perf_event_attr::sample_period >= 128.
Suppose we start out with sample_period=192; then we'll set period_left
to 192, we'll end up with left = 128 (we truncate the lower bits). We
get an interrupt, find that period_left = 64 (>0 so we return 0 and
don't get an overflow handler), up that to 128. Then we trigger again,
at n=256. Then we find period_left = -64 (<=0 so we return 1 and do get
an overflow). We increment with sample_period so we get left = 128. We
fire again, at n=384, period_left = 0 (<=0 so we return 1 and get an
overflow). And on and on.
So while the individual interrupts are 'wrong' we get then with
interval=256,128 in exactly the right ratio to average out at 192. And
this works for everything >=128.
So the num_samples*fixed_period thing is still entirely correct +- 127,
which is good enough I'd say, as you already have that error anyhow.
So no need to 'fix' the tools, al we need to do is refuse to create
INST_RETIRED:ALL events with sample_period < 128.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
[ Updated comments and changelog a bit. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1424225886-18652-3-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Haswell has a new feature that utilizes the existing LBR facility to
record call chains. To enable this feature, bits (JCC, NEAR_IND_JMP,
NEAR_REL_JMP, FAR_BRANCH, EN_CALLSTACK) in LBR_SELECT must be set to 1,
bits (NEAR_REL_CALL, NEAR-IND_CALL, NEAR_RET) must be cleared. Due to
a hardware bug of Haswell, this feature doesn't work well with
FREEZE_LBRS_ON_PMI.
When the call stack feature is enabled, the LBR stack will capture
unfiltered call data normally, but as return instructions are executed,
the last captured branch record is flushed from the on-chip registers
in a last-in first-out (LIFO) manner. Thus, branch information relative
to leaf functions will not be captured, while preserving the call stack
information of the main line execution path.
This patch defines a separate lbr_sel map for Haswell. The map contains
a new entry for the call stack feature.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: eranian@google.com
Cc: jolsa@redhat.com
Link: http://lkml.kernel.org/r/1415156173-10035-5-git-send-email-kan.liang@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We currently allow any process to use rdpmc. This significantly
weakens the protection offered by PR_TSC_DISABLED, and it could be
helpful to users attempting to exploit timing attacks.
Since we can't enable access to individual counters, use a very
coarse heuristic to limit access to rdpmc: allow access only when
a perf_event is mmapped. This protects seccomp sandboxes.
There is plenty of room to further tighen these restrictions. For
example, this allows rdpmc for any x86_pmu event, but it's only
useful for self-monitoring tasks.
As a side effect, cap_user_rdpmc will now be false for AMD uncore
events. This isn't a real regression, since .event_idx is disabled
for these events anyway for the time being. Whenever that gets
re-added, the cap_user_rdpmc code can be adjusted or refactored
accordingly.
Signed-off-by: Andy Lutomirski <luto@amacapital.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Kees Cook <keescook@chromium.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Vince Weaver <vince@deater.net>
Cc: "hillf.zj" <hillf.zj@alibaba-inc.com>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/a2bdb3cf3a1d70c26980d7c6dddfbaa69f3182bf.1414190806.git.luto@amacapital.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
These patches:
86a349a28b ("perf/x86/intel: Add Broadwell core support")
c46e665f03 ("perf/x86: Add INST_RETIRED.ALL workarounds")
fdda3c4aac ("perf/x86/intel: Use Broadwell cache event list for Haswell")
introduced magic constants and unexplained changes:
https://lkml.org/lkml/2014/10/28/1128https://lkml.org/lkml/2014/10/27/325https://lkml.org/lkml/2014/8/27/546https://lkml.org/lkml/2014/10/28/546
Peter Zijlstra has attempted to help out, to clean up the mess:
https://lkml.org/lkml/2014/10/28/543
But has not received helpful and constructive replies which makes
me doubt wether it can all be finished in time until v3.18 is
released.
Despite various review feedback the author (Andi Kleen) has answered
only few of the review questions and has generally been uncooperative,
only giving replies when prompted repeatedly, and only giving minimal
answers instead of constructively explaining and helping along the effort.
That kind of behavior is not acceptable.
There's also a boot crash on Intel E5-1630 v3 CPUs reported for another
commit from Andi Kleen:
e735b9db12 ("perf/x86/intel/uncore: Add Haswell-EP uncore support")
https://lkml.org/lkml/2014/10/22/730
Which is not yet resolved. The uncore driver is independent in theory,
but the crash makes me worry about how well all these patches were
tested and makes me uneasy about the level of interminging that the
Broadwell and Haswell code has received by the commits above.
As a first step to resolve the mess revert the Broadwell client commits
back to the v3.17 version, before we run out of time and problematic
code hits a stable upstream kernel.
( If the Haswell-EP crash is not resolved via a simple fix then we'll have
to revert the Haswell-EP uncore driver as well. )
The Broadwell client series has to be submitted in a clean fashion, with
single, well documented changes per patch. If they are submitted in time
and are accepted during review then they can possibly go into v3.19 but
will need additional scrutiny due to the rocky history of this patch set.
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: eranian@google.com
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/1409683455-29168-3-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
On Broadwell INST_RETIRED.ALL cannot be used with any period
that doesn't have the lowest 6 bits cleared. And the period
should not be smaller than 128.
Add a new callback to enforce this, and set it for Broadwell.
This is erratum BDM57 and BDM11.
How does this handle the case when an app requests a specific
period with some of the bottom bits set
The apps thinks it is sampling at X occurences per sample, when it is
in fact at X - 63 (worst case).
Short answer:
Any useful instruction sampling period needs to be 4-6 orders
of magnitude larger than 128, as an PMI every 128 instructions
would instantly overwhelm the system and be throttled.
So the +-64 error from this is really small compared to the
period, much smaller than normal system jitter.
Long answer:
<write up by Peter:>
IFF we guarantee perf_event_attr::sample_period >= 128.
Suppose we start out with sample_period=192; then we'll set period_left
to 192, we'll end up with left = 128 (we truncate the lower bits). We
get an interrupt, find that period_left = 64 (>0 so we return 0 and
don't get an overflow handler), up that to 128. Then we trigger again,
at n=256. Then we find period_left = -64 (<=0 so we return 1 and do get
an overflow). We increment with sample_period so we get left = 128. We
fire again, at n=384, period_left = 0 (<=0 so we return 1 and get an
overflow). And on and on.
So while the individual interrupts are 'wrong' we get then with
interval=256,128 in exactly the right ratio to average out at 192. And
this works for everything >=128.
So the num_samples*fixed_period thing is still entirely correct +- 127,
which is good enough I'd say, as you already have that error anyhow.
So no need to 'fix' the tools, al we need to do is refuse to create
INST_RETIRED:ALL events with sample_period < 128.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com>
Cc: Mark Davies <junk@eslaf.co.uk>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1409683455-29168-4-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The basic idea is that it does not make sense to list all PEBS
events individually. The list is very long, sometimes outdated
and the hardware doesn't need it. If an event does not support
PEBS it will just not count, there is no security issue.
We need to only list events that something special, like
supporting load or store addresses.
This vastly simplifies the PEBS event selection. It also
speeds up the scheduling because the scheduler doesn't
have to walk as many constraints.
Bugs fixed:
- We do not allow setting forbidden flags with PEBS anymore
(SDM 18.9.4), except for the special cycle event.
This is done using a new constraint macro that also
matches on the event flags.
- Correct DataLA and load/store/na flags reporting on Haswell
[Requires a followon patch]
- We did not allow all PEBS events on Haswell:
We were missing some valid subevents in d1-d2 (MEM_LOAD_UOPS_RETIRED.*,
MEM_LOAD_UOPS_RETIRED_L3_HIT_RETIRED.*)
This includes the changes proposed by Stephane earlier and obsoletes
his patchkit (except for some changes on pre Sandy Bridge/Silvermont
CPUs)
I only did Sandy Bridge and Silvermont and later so far, mostly because these
are the parts I could directly confirm the hardware behavior with hardware
architects. Also I do not believe the older CPUs have any
missing events in their PEBS list, so there's no pressing
need to change them.
I did not implement the flag proposed by Peter to allow
setting forbidden flags. If really needed this could
be implemented on to of this patch.
v2: Fix broken store events on SNB/IVB (Stephane Eranian)
v3: More fixes. Rename some arguments (Stephane Eranian)
v4: List most Haswell events individually again to report
memory operation type correctly.
Add new flags to describe load/store/na for datala.
Update description.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Reviewed-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1407785233-32193-2-git-send-email-eranian@google.com
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com>
Cc: Mark Davies <junk@eslaf.co.uk>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Yan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
With -cpu host, KVM reports LBR and extra_regs support, if the host has
support.
When the guest perf driver tries to access LBR or extra_regs MSR,
it #GPs all MSR accesses,since KVM doesn't handle LBR and extra_regs support.
So check the related MSRs access right once at initialization time to avoid
the error access at runtime.
For reproducing the issue, please build the kernel with CONFIG_KVM_INTEL = y
(for host kernel).
And CONFIG_PARAVIRT = n and CONFIG_KVM_GUEST = n (for guest kernel).
Start the guest with -cpu host.
Run perf record with --branch-any or --branch-filter in guest to trigger LBR
Run perf stat offcore events (E.g. LLC-loads/LLC-load-misses ...) in guest to
trigger offcore_rsp #GP
Signed-off-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com>
Cc: Mark Davies <junk@eslaf.co.uk>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Yan, Zheng <zheng.z.yan@intel.com>
Link: http://lkml.kernel.org/r/1405365957-20202-1-git-send-email-kan.liang@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The EVENT_CONSTRAINT_END() macro defines the end marker as
a constraint with a weight of zero. This was all fine
until we blacklisted the corrupting memory events on
Intel IvyBridge. These events are blacklisted by using
a counter bitmask of zero. Thus, they also get a constraint
weight of zero.
The iteration macro: for_each_constraint tests the weight==0.
Therefore, it was stopping at the first blacklisted event, i.e.,
0xd0. The corrupting events were therefore considered as
unconstrained and were scheduled on any of the generic counters.
This patch fixes the end marker to have a weight of -1. With
this, the blacklisted events get an empty constraint and cannot
be scheduled which is what we want for now.
Signed-off-by: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com>
Reviewed-by: Stephane Eranian <eranian@google.com>
Cc: peterz@infradead.org
Cc: ak@linux.intel.com
Cc: jolsa@redhat.com
Cc: zheng.z.yan@intel.com
Link: http://lkml.kernel.org/r/20131204232437.GA10689@starlight
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch fixes a problem with the shared registers mutual
exclusion code and incremental event scheduling by the
generic perf_event code.
There was a bug whereby the mutual exclusion on the shared
registers was not enforced because of incremental scheduling
abort due to event constraints. As an example on Intel
Nehalem, consider the following events:
group1= L1D_CACHE_LD:E_STATE,OFFCORE_RESPONSE_0:PF_RFO,L1D_CACHE_LD:I_STATE
group2= L1D_CACHE_LD:I_STATE
The L1D_CACHE_LD event can only be measured by 2 counters. Yet, there
are 3 instances here. The first group can be scheduled and is committed.
Then, the generic code tries to schedule group2 and this fails (because
there is no more counter to support the 3rd instance of L1D_CACHE_LD).
But in x86_schedule_events() error path, put_event_contraints() is invoked
on ALL the events and not just the ones that just failed. That causes the
"lock" on the shared offcore_response MSR to be released. Yet the first group
is actually scheduled and is exposed to reprogramming of that shared msr by
the sibling HT thread. In other words, there is no guarantee on what is
measured.
This patch fixes the problem by tagging committed events with the
PERF_X86_EVENT_COMMITTED tag. In the error path of x86_schedule_events(),
only the events NOT tagged have their constraint released. The tag
is eventually removed when the event in descheduled.
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20130620164254.GA3556@quad
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Recent Intel CPUs like Haswell and IvyBridge have a new
alternative MSR range for perfctrs that allows writing the full
counter width. Enable this range if the hardware reports it
using a new capability bit.
Currently the perf code queries CPUID to get the counter width,
and sign extends the counter values as needed. The traditional
PERFCTR MSRs always limit to 32bit, even though the counter
internally is larger (usually 48 bits on recent CPUs)
When the new capability is set use the alternative range which
do not have these restrictions.
This lowers the overhead of perf stat slightly because it has to
do less interrupts to accumulate the counter value. On Haswell
it also avoids some problems with TSX aborting when the end of
the counter range is reached.
( See the patch "perf/x86/intel: Avoid checkpointed counters
causing excessive TSX aborts" for more details. )
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Reviewed-by: Stephane Eranian <eranian@google.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1372173153-20215-1-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
