The resent conversion of the cpu hotplug support in the uncore driver
introduced a regression due to the way the callbacks are invoked at
initialization time.
The old code called the prepare/starting/online function on each online cpu
as a block. The new code registers the hotplug callbacks in the core for
each state. The core invokes the callbacks at each registration on all
online cpus.
The code implicitely relied on the prepare/starting/online callbacks being
called as combo on a particular cpu, which was not obvious and completely
undocumented.
The resulting subtle wreckage happens due to the way how the uncore code
manages shared data structures for cpus which share an uncore resource in
hardware. The sharing is determined in the cpu starting callback, but the
prepare callback allocates per cpu data for the upcoming cpu because
potential sharing is unknown at this point. If the starting callback finds
a online cpu which shares the hardware resource it takes a refcount on the
percpu data of that cpu and puts the own data structure into a
'free_at_online' pointer of that shared data structure. The online callback
frees that.
With the old model this worked because in a starting callback only one non
unused structure (the one of the starting cpu) was available. The new code
allocates the data structures for all cpus when the prepare callback is
registered.
Now the starting function iterates through all online cpus and looks for a
data structure (skipping its own) which has a matching hardware id. The id
member of the data structure is initialized to 0, but the hardware id can
be 0 as well. The resulting wreckage is:
CPU0 finds a matching id on CPU1, takes a refcount on CPU1 data and puts
its own data structure into CPU1s data structure to be freed.
CPU1 skips CPU0 because the data structure is its allegedly unsued own.
It finds a matching id on CPU2, takes a refcount on CPU1 data and puts
its own data structure into CPU2s data structure to be freed.
....
Now the online callbacks are invoked.
CPU0 has a pointer to CPU1s data and frees the original CPU0 data. So
far so good.
CPU1 has a pointer to CPU2s data and frees the original CPU1 data, which
is still referenced by CPU0 ---> Booom
So there are two issues to be solved here:
1) The id field must be initialized at allocation time to a value which
cannot be a valid hardware id, i.e. -1
This prevents the above scenario, but now CPU1 and CPU2 both stick their
own data structure into the free_at_online pointer of CPU0. So we leak
CPU1s data structure.
2) Fix the memory leak described in #1
Instead of having a single pointer, use a hlist to enqueue the
superflous data structures which are then freed by the first cpu
invoking the online callback.
Ideally we should know the sharing _before_ invoking the prepare callback,
but that's way beyond the scope of this bug fix.
[ tglx: Rewrote changelog ]
Fixes: 96b2bd3866 ("perf/x86/amd/uncore: Convert to hotplug state machine")
Reported-and-tested-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Borislav Petkov <bp@suse.de>
Link: http://lkml.kernel.org/r/20160909160822.lowgmkdwms2dheyv@linutronix.de
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Yanqiu Zhang reported kernel panic when using mbm event
on system where CQM is detected but without mbm event
support, like with perf:
# perf stat -e 'intel_cqm/event=3/' -a
BUG: unable to handle kernel NULL pointer dereference at 0000000000000020
IP: [<ffffffff8100d64c>] update_sample+0xbc/0xe0
...
<IRQ>
[<ffffffff8100d688>] __intel_mbm_event_init+0x18/0x20
[<ffffffff81113d6b>] flush_smp_call_function_queue+0x7b/0x160
[<ffffffff81114853>] generic_smp_call_function_single_interrupt+0x13/0x60
[<ffffffff81052017>] smp_call_function_interrupt+0x27/0x40
[<ffffffff816fb06c>] call_function_interrupt+0x8c/0xa0
...
The reason is that we currently allow to init mbm event
even if mbm support is not detected. Adding checks for
both cqm and mbm events and support into cqm's event_init.
Fixes: 33c3cc7acf ("perf/x86/mbm: Add Intel Memory B/W Monitoring enumeration and init")
Reported-by: Yanqiu Zhang <yanqzhan@redhat.com>
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Vikas Shivappa <vikas.shivappa@linux.intel.com>
Cc: Tony Luck <tony.luck@intel.com>
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/1473089407-21857-1-git-send-email-jolsa@kernel.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
There are bug reports about miscounting uncore counters on some
client machines like Sandybridge, Broadwell and Skylake. It is
very likely to be observed on idle systems.
This issue is caused by a hardware issue. PERF_GLOBAL_CTL could be
cleared after Package C7, and nothing will be count.
The related errata (HSD 158) could be found in:
www.intel.com/content/dam/www/public/us/en/documents/specification-updates/4th-gen-core-family-desktop-specification-update.pdf
This patch tries to work around this issue by re-enabling PERF_GLOBAL_CTL
in ->enable_box(). The workaround does not cover all cases. It helps for new
events after returning from C7. But it cannot prevent C7, it will still
miscount if a counter is already active.
There is no drawback in leaving it enabled, so it does not need
disable_box() here.
Signed-off-by: Kan Liang <kan.liang@intel.com>
Cc: <stable@vger.kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
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>
Link: http://lkml.kernel.org/r/1470925874-59943-1-git-send-email-kan.liang@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The lbr_context logic confused me; it appears to me to try and do the
same thing the pmu::sched_task() callback does now, but limited to
per-task events.
So rip it out. Afaict this should also improve performance, because I
think the current code can end up doing lbr_reset() twice, once from
the pmu::add() and then again from pmu::sched_task(), and MSR writes
(all 3*16 of them) are expensive!!
While thinking through the cases that need the reset it occured to me
the first install of an event in an active context needs to reset the
LBR (who knows what crap is in there), but detecting this case is
somewhat hard.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
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>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In order to allow optimizing perf_pmu_sched_task() we must ensure
perf_sched_cb_{inc,dec}() are no longer called from NMI context; this
means that pmu::{start,stop}() can no longer use them.
Prepare for this by reworking the whole large PEBS setup code.
The current code relied on the cpuc->pebs_enabled state, however since
that reflects the current active state as per pmu::{start,stop}() we
can no longer rely on this.
Introduce two counters: cpuc->n_pebs and cpuc->n_large_pebs which
count the total number of PEBS events and the number of PEBS events
that have FREERUNNING set, resp.. With this we can tell if the current
setup requires a single record interrupt threshold or can use a larger
buffer.
This also improves the code in that it re-enables the large threshold
once the PEBS event that required single record gets removed.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
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>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull x86 header cleanups from Ingo Molnar:
"This tree is a cleanup of the x86 tree reducing spurious uses of
module.h - which should improve build performance a bit"
* 'x86-headers-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86, crypto: Restore MODULE_LICENSE() to glue_helper.c so it loads
x86/apic: Remove duplicated include from probe_64.c
x86/ce4100: Remove duplicated include from ce4100.c
x86/headers: Include spinlock_types.h in x8664_ksyms_64.c for missing spinlock_t
x86/platform: Delete extraneous MODULE_* tags fromm ts5500
x86: Audit and remove any remaining unnecessary uses of module.h
x86/kvm: Audit and remove any unnecessary uses of module.h
x86/xen: Audit and remove any unnecessary uses of module.h
x86/platform: Audit and remove any unnecessary uses of module.h
x86/lib: Audit and remove any unnecessary uses of module.h
x86/kernel: Audit and remove any unnecessary uses of module.h
x86/mm: Audit and remove any unnecessary uses of module.h
x86: Don't use module.h just for AUTHOR / LICENSE tags
Pull x86 cpufeature updates from Thomas Gleixner:
- a workaround for the MONITOR instruction erratum of Goldmont CPUs
- small fixes and cleanups here and there
* 'x86-cpu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/cpu: Add workaround for MONITOR instruction erratum on Goldmont based CPUs
x86/cpu: Rename "WESTMERE2" family to "NEHALEM_G"
x86/amd_nb: Clean up init path
x86/cpufeature: Add helper macro for mask check macros
x86/cpufeature: Make sure DISABLED/REQUIRED macros are updated
x86/cpufeature: Update cpufeaure macros
Pull smp hotplug updates from Thomas Gleixner:
"This is the next part of the hotplug rework.
- Convert all notifiers with a priority assigned
- Convert all CPU_STARTING/DYING notifiers
The final removal of the STARTING/DYING infrastructure will happen
when the merge window closes.
Another 700 hundred line of unpenetrable maze gone :)"
* 'smp-hotplug-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (70 commits)
timers/core: Correct callback order during CPU hot plug
leds/trigger/cpu: Move from CPU_STARTING to ONLINE level
powerpc/numa: Convert to hotplug state machine
arm/perf: Fix hotplug state machine conversion
irqchip/armada: Avoid unused function warnings
ARC/time: Convert to hotplug state machine
clocksource/atlas7: Convert to hotplug state machine
clocksource/armada-370-xp: Convert to hotplug state machine
clocksource/exynos_mct: Convert to hotplug state machine
clocksource/arm_global_timer: Convert to hotplug state machine
rcu: Convert rcutree to hotplug state machine
KVM/arm/arm64/vgic-new: Convert to hotplug state machine
smp/cfd: Convert core to hotplug state machine
x86/x2apic: Convert to CPU hotplug state machine
profile: Convert to hotplug state machine
timers/core: Convert to hotplug state machine
hrtimer: Convert to hotplug state machine
x86/tboot: Convert to hotplug state machine
arm64/armv8 deprecated: Convert to hotplug state machine
hwtracing/coresight-etm4x: Convert to hotplug state machine
...
Pull networking updates from David Miller:
1) Unified UDP encapsulation offload methods for drivers, from
Alexander Duyck.
2) Make DSA binding more sane, from Andrew Lunn.
3) Support QCA9888 chips in ath10k, from Anilkumar Kolli.
4) Several workqueue usage cleanups, from Bhaktipriya Shridhar.
5) Add XDP (eXpress Data Path), essentially running BPF programs on RX
packets as soon as the device sees them, with the option to mirror
the packet on TX via the same interface. From Brenden Blanco and
others.
6) Allow qdisc/class stats dumps to run lockless, from Eric Dumazet.
7) Add VLAN support to b53 and bcm_sf2, from Florian Fainelli.
8) Simplify netlink conntrack entry layout, from Florian Westphal.
9) Add ipv4 forwarding support to mlxsw spectrum driver, from Ido
Schimmel, Yotam Gigi, and Jiri Pirko.
10) Add SKB array infrastructure and convert tun and macvtap over to it.
From Michael S Tsirkin and Jason Wang.
11) Support qdisc packet injection in pktgen, from John Fastabend.
12) Add neighbour monitoring framework to TIPC, from Jon Paul Maloy.
13) Add NV congestion control support to TCP, from Lawrence Brakmo.
14) Add GSO support to SCTP, from Marcelo Ricardo Leitner.
15) Allow GRO and RPS to function on macsec devices, from Paolo Abeni.
16) Support MPLS over IPV4, from Simon Horman.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1622 commits)
xgene: Fix build warning with ACPI disabled.
be2net: perform temperature query in adapter regardless of its interface state
l2tp: Correctly return -EBADF from pppol2tp_getname.
net/mlx5_core/health: Remove deprecated create_singlethread_workqueue
net: ipmr/ip6mr: update lastuse on entry change
macsec: ensure rx_sa is set when validation is disabled
tipc: dump monitor attributes
tipc: add a function to get the bearer name
tipc: get monitor threshold for the cluster
tipc: make cluster size threshold for monitoring configurable
tipc: introduce constants for tipc address validation
net: neigh: disallow transition to NUD_STALE if lladdr is unchanged in neigh_update()
MAINTAINERS: xgene: Add driver and documentation path
Documentation: dtb: xgene: Add MDIO node
dtb: xgene: Add MDIO node
drivers: net: xgene: ethtool: Use phy_ethtool_gset and sset
drivers: net: xgene: Use exported functions
drivers: net: xgene: Enable MDIO driver
drivers: net: xgene: Add backward compatibility
drivers: net: phy: xgene: Add MDIO driver
...
Pull perf updates from Ingo Molnar:
"With over 300 commits it's been a busy cycle - with most of the work
concentrated on the tooling side (as it should).
The main kernel side enhancements were:
- Add per event callchain limit: Recently we introduced a sysctl to
tune the max-stack for all events for which callchains were
requested:
$ sysctl kernel.perf_event_max_stack
kernel.perf_event_max_stack = 127
Now this patch introduces a way to configure this per event, i.e.
this becomes possible:
$ perf record -e sched:*/max-stack=2/ -e block:*/max-stack=10/ -a
allowing finer tuning of how much buffer space callchains use.
This uses an u16 from the reserved space at the end, leaving
another u16 for future use.
There has been interest in even finer tuning, namely to control the
max stack for kernel and userspace callchains separately. Further
discussion is needed, we may for instance use the remaining u16 for
that and when it is present, assume that the sample_max_stack
introduced in this patch applies for the kernel, and the u16 left
is used for limiting the userspace callchain (Arnaldo Carvalho de
Melo)
- Optimize AUX event (hardware assisted side-band event) delivery
(Kan Liang)
- Rework Intel family name macro usage (this is partially x86 arch
work) (Dave Hansen)
- Refine and fix Intel LBR support (David Carrillo-Cisneros)
- Add support for Intel 'TopDown' events (Andi Kleen)
- Intel uncore PMU driver fixes and enhancements (Kan Liang)
- ... other misc changes.
Here's an incomplete list of the tooling enhancements (but there's
much more, see the shortlog and the git log for details):
- Support cross unwinding, i.e. collecting '--call-graph dwarf'
perf.data files in one machine and then doing analysis in another
machine of a different hardware architecture. This enables, for
instance, to do:
$ perf record -a --call-graph dwarf
on a x86-32 or aarch64 system and then do 'perf report' on it on a
x86_64 workstation (He Kuang)
- Allow reading from a backward ring buffer (one setup via
sys_perf_event_open() with perf_event_attr.write_backward = 1)
(Wang Nan)
- Finish merging initial SDT (Statically Defined Traces) support, see
cset comments for details about how it all works (Masami Hiramatsu)
- Support attaching eBPF programs to tracepoints (Wang Nan)
- Add demangling of symbols in programs written in the Rust language
(David Tolnay)
- Add support for tracepoints in the python binding, including an
example, that sets up and parses sched:sched_switch events,
tools/perf/python/tracepoint.py (Jiri Olsa)
- Introduce --stdio-color to set up the color output mode selection
in 'annotate' and 'report', allowing emit color escape sequences
when redirecting the output of these tools (Arnaldo Carvalho de
Melo)
- Add 'callindent' option to 'perf script -F', to indent the Intel PT
call stack, making this output more ftrace-like (Adrian Hunter,
Andi Kleen)
- Allow dumping the object files generated by llvm when processing
eBPF scriptlet events (Wang Nan)
- Add stackcollapse.py script to help generating flame graphs (Paolo
Bonzini)
- Add --ldlat option to 'perf mem' to specify load latency for loads
event (e.g. cpu/mem-loads/ ) (Jiri Olsa)
- Tooling support for Intel TopDown counters, recently added to the
kernel (Andi Kleen)"
* 'perf-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (303 commits)
perf tests: Add is_printable_array test
perf tools: Make is_printable_array global
perf script python: Fix string vs byte array resolving
perf probe: Warn unmatched function filter correctly
perf cpu_map: Add more helpers
perf stat: Balance opening and reading events
tools: Copy linux/{hash,poison}.h and check for drift
perf tools: Remove include/linux/list.h from perf's MANIFEST
tools: Copy the bitops files accessed from the kernel and check for drift
Remove: kernel unistd*h files from perf's MANIFEST, not used
perf tools: Remove tools/perf/util/include/linux/const.h
perf tools: Remove tools/perf/util/include/asm/byteorder.h
perf tools: Add missing linux/compiler.h include to perf-sys.h
perf jit: Remove some no-op error handling
perf jit: Add missing curly braces
objtool: Initialize variable to silence old compiler
objtool: Add -I$(srctree)/tools/arch/$(ARCH)/include/uapi
perf record: Add --tail-synthesize option
perf session: Don't warn about out of order event if write_backward is used
perf tools: Enable overwrite settings
...
This patch adds support for non-linear data on raw records. It
extends raw records to have one or multiple fragments that will
be written linearly into the ring slot, where each fragment can
optionally have a custom callback handler to walk and extract
complex, possibly non-linear data.
If a callback handler is provided for a fragment, then the new
__output_custom() will be used instead of __output_copy() for
the perf_output_sample() part. perf_prepare_sample() does all
the size calculation only once, so perf_output_sample() doesn't
need to redo the same work anymore, meaning real_size and padding
will be cached in the raw record. The raw record becomes 32 bytes
in size without holes; to not increase it further and to avoid
doing unnecessary recalculations in fast-path, we can reuse
next pointer of the last fragment, idea here is borrowed from
ZERO_OR_NULL_PTR(), which should keep the perf_output_sample()
path for PERF_SAMPLE_RAW minimal.
This facility is needed for BPF's event output helper as a first
user that will, in a follow-up, add an additional perf_raw_frag
to its perf_raw_record in order to be able to more efficiently
dump skb context after a linear head meta data related to it.
skbs can be non-linear and thus need a custom output function to
dump buffers. Currently, the skb data needs to be copied twice;
with the help of __output_custom() this work only needs to be
done once. Future users could be things like XDP/BPF programs
that work on different context though and would thus also have
a different callback function.
The few users of raw records are adapted to initialize their frag
data from the raw record itself, no change in behavior for them.
The code is based upon a PoC diff provided by Peter Zijlstra [1].
[1] http://thread.gmane.org/gmane.linux.network/421294
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Historically a lot of these existed because we did not have
a distinction between what was modular code and what was providing
support to modules via EXPORT_SYMBOL and friends. That changed
when we forked out support for the latter into the export.h file.
This means we should be able to reduce the usage of module.h
in code that is obj-y Makefile or bool Kconfig. In the case of
some of these which are modular, we can extend that to also include
files that are building basic support functionality but not related
to loading or registering the final module; such files also have
no need whatsoever for module.h
The advantage in removing such instances is that module.h itself
sources about 15 other headers; adding significantly to what we feed
cpp, and it can obscure what headers we are effectively using.
Since module.h was the source for init.h (for __init) and for
export.h (for EXPORT_SYMBOL) we consider each instance for the
presence of either and replace as needed.
In the case of crypto/glue_helper.c we delete a redundant instance
of MODULE_LICENSE in order to delete module.h -- the license info
is already present at the top of the file.
The uncore change warrants a mention too; it is uncore.c that uses
module.h and not uncore.h; hence the relocation done there.
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20160714001901.31603-9-paul.gortmaker@windriver.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Add quirk for context switch to save/restore the value of
MSR_LAST_BRANCH_FROM_x when LBR is enabled and there is potential for
kernel addresses to be in the lbr_from register.
To test this patch, use a perf tool and kernel with the patch
next in this series. That patch removes the work around that masked
the hw bug:
$ ./lbr_perf record --call-graph lbr -e cycles:k sleep 1
where lbr_perf is the patched perf tool, that allows to specify :k
on lbr mode. The above command will trigger a #GPF :
WARNING: CPU: 28 PID: 14096 at arch/x86/mm/extable.c:65 ex_handler_wrmsr_unsafe+0x70/0x80
unchecked MSR access error: WRMSR to 0x681 (tried to write 0x1fffffff81010794)
...
Call Trace:
[<ffffffff8167af49>] dump_stack+0x4d/0x63
[<ffffffff810b9b15>] __warn+0xe5/0x100
[<ffffffff810b9be9>] warn_slowpath_fmt+0x49/0x50
[<ffffffff810abb40>] ex_handler_wrmsr_unsafe+0x70/0x80
[<ffffffff810abc42>] fixup_exception+0x42/0x50
[<ffffffff81079d1a>] do_general_protection+0x8a/0x160
[<ffffffff81684ec2>] general_protection+0x22/0x30
[<ffffffff810101b9>] ? intel_pmu_lbr_sched_task+0xc9/0x380
[<ffffffff81009d7c>] intel_pmu_sched_task+0x3c/0x60
[<ffffffff81003a2b>] x86_pmu_sched_task+0x1b/0x20
[<ffffffff81192a5b>] perf_pmu_sched_task+0x6b/0xb0
[<ffffffff8119746d>] __perf_event_task_sched_in+0x7d/0x150
[<ffffffff810dd9dc>] finish_task_switch+0x15c/0x200
[<ffffffff8167f894>] __schedule+0x274/0x6cc
[<ffffffff8167fdd9>] schedule+0x39/0x90
[<ffffffff81675398>] exit_to_usermode_loop+0x39/0x89
[<ffffffff810028ce>] prepare_exit_to_usermode+0x2e/0x30
[<ffffffff81683c1b>] retint_user+0x8/0x10
Signed-off-by: David Carrillo-Cisneros <davidcc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Stephane Eranian <eranian@google.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Link: http://lkml.kernel.org/r/1466533874-52003-5-git-send-email-davidcc@google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Intel's SDM states that bits 61:62 in MSR_LAST_BRANCH_FROM_x are the
TSX flags for formats with LBR_TSX flags (i.e. LBR_FORMAT_EIP_EFLAGS2).
However, when the CPU has TSX support deactivated, bits 61:62 actually
behave as follows:
- For wrmsr(), bits 61:62 are considered part of the sign extension.
- When capturing branches, the LBR hw will always clear bits 61:62.
regardless of the sign extension.
Therefore, if:
1) LBR has TSX format.
2) CPU has no TSX support enabled.
... then any value passed to wrmsr() must be sign extended to 63 bits
and any value from rdmsr() must be converted to have a sign extension
of 61 bits, ignoring the values at TSX flags.
This bug was masked by the work-around to the Intel's CPU bug:
BJ94. "LBR May Contain Incorrect Information When Using FREEZE_LBRS_ON_PMI"
in Document Number: 324643-037US.
The aforementioned work-around uses hw flags to filter out all kernel
branches, limiting LBR callstack to user level execution only.
Since user addresses are not sign extended, they do not trigger the wrmsr()
bug in MSR_LAST_BRANCH_FROM_x when saved/restored at context switch.
To verify the hw bug:
$ perf record -b -e cycles sleep 1
$ rdmsr -p 0 0x680
0x1fffffffb0b9b0cc
$ wrmsr -p 0 0x680 0x1fffffffb0b9b0cc
write(): Input/output error
The quirk for LBR_FROM_ MSRs is required before calls to wrmsrl() and
after rdmsrl().
This patch introduces it for wrmsrl()'s done for testing LBR support.
Future patch in series adds the quirk for context switch, that would
be required if LBR callstack is to be enabled for ring 0.
Signed-off-by: David Carrillo-Cisneros <davidcc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Stephane Eranian <eranian@google.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Link: http://lkml.kernel.org/r/1466533874-52003-3-git-send-email-davidcc@google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
