Pull kvm fixes from Paolo Bonzini:
"This is a bit on the large side, mostly due to two changes:
- Changes to disable some broken PMU virtualization (see below for
details under "x86 PMU")
- Clean up SVM's enter/exit assembly code so that it can be compiled
without OBJECT_FILES_NON_STANDARD. This fixes a warning "Unpatched
return thunk in use. This should not happen!" when running KVM
selftests.
Everything else is small bugfixes and selftest changes:
- Fix a mostly benign bug in the gfn_to_pfn_cache infrastructure
where KVM would allow userspace to refresh the cache with a bogus
GPA. The bug has existed for quite some time, but was exposed by a
new sanity check added in 6.9 (to ensure a cache is either
GPA-based or HVA-based).
- Drop an unused param from gfn_to_pfn_cache_invalidate_start() that
got left behind during a 6.9 cleanup.
- Fix a math goof in x86's hugepage logic for
KVM_SET_MEMORY_ATTRIBUTES that results in an array overflow
(detected by KASAN).
- Fix a bug where KVM incorrectly clears root_role.direct when
userspace sets guest CPUID.
- Fix a dirty logging bug in the where KVM fails to write-protect
SPTEs used by a nested guest, if KVM is using Page-Modification
Logging and the nested hypervisor is NOT using EPT.
x86 PMU:
- Drop support for virtualizing adaptive PEBS, as KVM's
implementation is architecturally broken without an obvious/easy
path forward, and because exposing adaptive PEBS can leak host LBRs
to the guest, i.e. can leak host kernel addresses to the guest.
- Set the enable bits for general purpose counters in
PERF_GLOBAL_CTRL at RESET time, as done by both Intel and AMD
processors.
- Disable LBR virtualization on CPUs that don't support LBR
callstacks, as KVM unconditionally uses
PERF_SAMPLE_BRANCH_CALL_STACK when creating the perf event, and
would fail on such CPUs.
Tests:
- Fix a flaw in the max_guest_memory selftest that results in it
exhausting the supply of ucall structures when run with more than
256 vCPUs.
- Mark KVM_MEM_READONLY as supported for RISC-V in
set_memory_region_test"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (30 commits)
KVM: Drop unused @may_block param from gfn_to_pfn_cache_invalidate_start()
KVM: selftests: Add coverage of EPT-disabled to vmx_dirty_log_test
KVM: x86/mmu: Fix and clarify comments about clearing D-bit vs. write-protecting
KVM: x86/mmu: Remove function comments above clear_dirty_{gfn_range,pt_masked}()
KVM: x86/mmu: Write-protect L2 SPTEs in TDP MMU when clearing dirty status
KVM: x86/mmu: Precisely invalidate MMU root_role during CPUID update
KVM: VMX: Disable LBR virtualization if the CPU doesn't support LBR callstacks
perf/x86/intel: Expose existence of callback support to KVM
KVM: VMX: Snapshot LBR capabilities during module initialization
KVM: x86/pmu: Do not mask LVTPC when handling a PMI on AMD platforms
KVM: x86: Snapshot if a vCPU's vendor model is AMD vs. Intel compatible
KVM: x86: Stop compiling vmenter.S with OBJECT_FILES_NON_STANDARD
KVM: SVM: Create a stack frame in __svm_sev_es_vcpu_run()
KVM: SVM: Save/restore args across SEV-ES VMRUN via host save area
KVM: SVM: Save/restore non-volatile GPRs in SEV-ES VMRUN via host save area
KVM: SVM: Clobber RAX instead of RBX when discarding spec_ctrl_intercepted
KVM: SVM: Drop 32-bit "support" from __svm_sev_es_vcpu_run()
KVM: SVM: Wrap __svm_sev_es_vcpu_run() with #ifdef CONFIG_KVM_AMD_SEV
KVM: SVM: Create a stack frame in __svm_vcpu_run() for unwinding
KVM: SVM: Remove a useless zeroing of allocated memory
...
Add a "has_callstack" field to the x86_pmu_lbr structure used to pass
information to KVM, and set it accordingly in x86_perf_get_lbr(). KVM
will use has_callstack to avoid trying to create perf LBR events with
PERF_SAMPLE_BRANCH_CALL_STACK on CPUs that don't support callstacks.
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Link: https://lore.kernel.org/r/20240307011344.835640-3-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
On x86 each struct cpu_hw_events maintains a table for counter assignment but
it missed to update one for the deleted event in x86_pmu_del(). This
can make perf_clear_dirty_counters() reset used counter if it's called
before event scheduling or enabling. Then it would return out of range
data which doesn't make sense.
The following code can reproduce the problem.
$ cat repro.c
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <linux/perf_event.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/syscall.h>
struct perf_event_attr attr = {
.type = PERF_TYPE_HARDWARE,
.config = PERF_COUNT_HW_CPU_CYCLES,
.disabled = 1,
};
void *worker(void *arg)
{
int cpu = (long)arg;
int fd1 = syscall(SYS_perf_event_open, &attr, -1, cpu, -1, 0);
int fd2 = syscall(SYS_perf_event_open, &attr, -1, cpu, -1, 0);
void *p;
do {
ioctl(fd1, PERF_EVENT_IOC_ENABLE, 0);
p = mmap(NULL, 4096, PROT_READ, MAP_SHARED, fd1, 0);
ioctl(fd2, PERF_EVENT_IOC_ENABLE, 0);
ioctl(fd2, PERF_EVENT_IOC_DISABLE, 0);
munmap(p, 4096);
ioctl(fd1, PERF_EVENT_IOC_DISABLE, 0);
} while (1);
return NULL;
}
int main(void)
{
int i;
int n = sysconf(_SC_NPROCESSORS_ONLN);
pthread_t *th = calloc(n, sizeof(*th));
for (i = 0; i < n; i++)
pthread_create(&th[i], NULL, worker, (void *)(long)i);
for (i = 0; i < n; i++)
pthread_join(th[i], NULL);
free(th);
return 0;
}
And you can see the out of range data using perf stat like this.
Probably it'd be easier to see on a large machine.
$ gcc -o repro repro.c -pthread
$ ./repro &
$ sudo perf stat -A -I 1000 2>&1 | awk '{ if (length($3) > 15) print }'
1.001028462 CPU6 196,719,295,683,763 cycles # 194290.996 GHz (71.54%)
1.001028462 CPU3 396,077,485,787,730 branch-misses # 15804359784.80% of all branches (71.07%)
1.001028462 CPU17 197,608,350,727,877 branch-misses # 14594186554.56% of all branches (71.22%)
2.020064073 CPU4 198,372,472,612,140 cycles # 194681.113 GHz (70.95%)
2.020064073 CPU6 199,419,277,896,696 cycles # 195720.007 GHz (70.57%)
2.020064073 CPU20 198,147,174,025,639 cycles # 194474.654 GHz (71.03%)
2.020064073 CPU20 198,421,240,580,145 stalled-cycles-frontend # 100.14% frontend cycles idle (70.93%)
3.037443155 CPU4 197,382,689,923,416 cycles # 194043.065 GHz (71.30%)
3.037443155 CPU20 196,324,797,879,414 cycles # 193003.773 GHz (71.69%)
3.037443155 CPU5 197,679,956,608,205 stalled-cycles-backend # 1315606428.66% backend cycles idle (71.19%)
3.037443155 CPU5 198,571,860,474,851 instructions # 13215422.58 insn per cycle
It should move the contents in the cpuc->assign as well.
Fixes: 5471eea5d3 ("perf/x86: Reset the dirty counter to prevent the leak for an RDPMC task")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kan Liang <kan.liang@linux.intel.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20240306061003.1894224-1-namhyung@kernel.org
The MSR_PEBS_DATA_CFG MSR register is used to configure which data groups
should be generated into a PEBS record, and it's shared among all counters.
If there are different configurations among counters, perf combines all the
configurations.
The first perf command as below requires a complete PEBS record
(including memory info, GPRs, XMMs, and LBRs). The second perf command
only requires a basic group. However, after the second perf command is
running, the MSR_PEBS_DATA_CFG register is cleared. Only a basic group is
generated in a PEBS record, which is wrong. The required information
for the first perf command is missed.
$ perf record --intr-regs=AX,SP,XMM0 -a -C 8 -b -W -d -c 100000003 -o /dev/null -e cpu/event=0xd0,umask=0x81/upp &
$ sleep 5
$ perf record --per-thread -c 1 -e cycles:pp --no-timestamp --no-tid taskset -c 8 ./noploop 1000
The first PEBS event is a system-wide PEBS event. The second PEBS event
is a per-thread event. When the thread is scheduled out, the
intel_pmu_pebs_del() function is invoked to update the PEBS state.
Since the system-wide event is still available, the cpuc->n_pebs is 1.
The cpuc->pebs_data_cfg is cleared. The data configuration for the
system-wide PEBS event is lost.
The (cpuc->n_pebs == 1) check was introduced in commit:
b6a32f023f ("perf/x86: Fix PEBS threshold initialization")
At that time, it indeed didn't hurt whether the state was updated
during the removal, because only the threshold is updated.
The calculation of the threshold takes the last PEBS event into
account.
However, since commit:
b752ea0c28 ("perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG")
we delay the threshold update, and clear the PEBS data config, which triggers
the bug.
The PEBS data config update scope should not be shrunk during removal.
[ mingo: Improved the changelog & comments. ]
Fixes: b752ea0c28 ("perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG")
Reported-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20240401133320.703971-1-kan.liang@linux.intel.com
AMD processors based on Zen 2 and later microarchitectures do not
support PMCx087 (instruction pipe stalls) which is used as the backing
event for "stalled-cycles-frontend" and "stalled-cycles-backend".
Use PMCx0A9 (cycles where micro-op queue is empty) instead to count
frontend stalls and remove the entry for backend stalls since there
is no direct replacement.
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Ian Rogers <irogers@google.com>
Fixes: 3fe3331bb2 ("perf/x86/amd: Add event map for AMD Family 17h")
Link: https://lore.kernel.org/r/03d7fc8fa2a28f9be732116009025bdec1b3ec97.1711352180.git.sandipan.das@amd.com
Currently, the LBR code assumes that LBR Freeze is supported on all processors
when X86_FEATURE_AMD_LBR_V2 is available i.e. CPUID leaf 0x80000022[EAX]
bit 1 is set. This is incorrect as the availability of the feature is
additionally dependent on CPUID leaf 0x80000022[EAX] bit 2 being set,
which may not be set for all Zen 4 processors.
Define a new feature bit for LBR and PMC freeze and set the freeze enable bit
(FLBRI) in DebugCtl (MSR 0x1d9) conditionally.
It should still be possible to use LBR without freeze for profile-guided
optimization of user programs by using an user-only branch filter during
profiling. When the user-only filter is enabled, branches are no longer
recorded after the transition to CPL 0 upon PMI arrival. When branch
entries are read in the PMI handler, the branch stack does not change.
E.g.
$ perf record -j any,u -e ex_ret_brn_tkn ./workload
Since the feature bit is visible under flags in /proc/cpuinfo, it can be
used to determine the feasibility of use-cases which require LBR Freeze
to be supported by the hardware such as profile-guided optimization of
kernels.
Fixes: ca5b7c0d96 ("perf/x86/amd/lbr: Add LbrExtV2 branch record support")
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/69a453c97cfd11c6f2584b19f937fe6df741510f.1711091584.git.sandipan.das@amd.com
When bringing a CPU online, some of the PMC and LBR related registers
are reset. The same is done when a CPU is taken offline although that
is unnecessary. This currently happens in the "cpu_dead" callback which
is also incorrect as the callback runs on a control CPU instead of the
one that is being taken offline. This also affects hibernation and
suspend to RAM on some platforms as reported in the link below.
Fixes: 21d59e3e2c ("perf/x86/amd/core: Detect PerfMonV2 support")
Reported-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/550a026764342cf7e5812680e3e2b91fe662b5ac.1706526029.git.sandipan.das@amd.com
Pull x86 cleanups from Ingo Molnar:
"Misc cleanups, including a large series from Thomas Gleixner to cure
sparse warnings"
* tag 'x86-cleanups-2024-03-11' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/nmi: Drop unused declaration of proc_nmi_enabled()
x86/callthunks: Use EXPORT_PER_CPU_SYMBOL_GPL() for per CPU variables
x86/cpu: Provide a declaration for itlb_multihit_kvm_mitigation
x86/cpu: Use EXPORT_PER_CPU_SYMBOL_GPL() for x86_spec_ctrl_current
x86/uaccess: Add missing __force to casts in __access_ok() and valid_user_address()
x86/percpu: Cure per CPU madness on UP
smp: Consolidate smp_prepare_boot_cpu()
x86/msr: Add missing __percpu annotations
x86/msr: Prepare for including <linux/percpu.h> into <asm/msr.h>
perf/x86/amd/uncore: Fix __percpu annotation
x86/nmi: Remove an unnecessary IS_ENABLED(CONFIG_SMP)
x86/apm_32: Remove dead function apm_get_battery_status()
x86/insn-eval: Fix function param name in get_eff_addr_sib()
To clean up the per CPU insanity of UP which causes sparse to be rightfully
unhappy and prevents the usage of the generic per CPU accessors on cpu_info
it is necessary to include <linux/percpu.h> into <asm/msr.h>.
Including <linux/percpu.h> into <asm/msr.h> is impossible because it ends
up in header dependency hell. The problem is that <asm/processor.h>
includes <asm/msr.h>. The inclusion of <linux/percpu.h> results in a
compile fail where the compiler cannot longer handle an include in
<asm/cpufeature.h> which references boot_cpu_data which is
defined in <asm/processor.h>.
The only reason why <asm/msr.h> is included in <asm/processor.h> are the
set/get_debugctlmsr() inlines. They are defined there because <asm/processor.h>
is such a nice dump ground for everything. In fact they belong obviously
into <asm/debugreg.h>.
Move them to <asm/debugreg.h> and fix up the resulting damage which is just
exposing the reliance on random include chains.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20240304005104.454678686@linutronix.de
The __percpu annotation in struct amd_uncore is confusing Sparse:
uncore.c:649:10: sparse: warning: incorrect type in initializer (different address spaces)
uncore.c:649:10: sparse: expected void const [noderef] __percpu *__vpp_verify
uncore.c:649:10: sparse: got union amd_uncore_info *
The reason is that the __percpu annotation sits between the '*'
dereferencing operator and the member name.
Move it before the dereferencing operator to cure this.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20240304005104.394845326@linutronix.de
AMD (ab)uses topology_die_id() to store the Node ID information and
topology_max_dies_per_pkg to store the number of nodes per package.
This collides with the proper processor die level enumeration which is
coming on AMD with CPUID 8000_0026, unless there is a correlation between
the two. There is zero documentation about that.
So provide new storage and new accessors which for now still access die_id
and topology_max_die_per_pkg(). Will be mopped up after AMD and HYGON are
converted over.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Juergen Gross <jgross@suse.com>
Tested-by: Sohil Mehta <sohil.mehta@intel.com>
Tested-by: Michael Kelley <mhklinux@outlook.com>
Tested-by: Zhang Rui <rui.zhang@intel.com>
Tested-by: Wang Wendy <wendy.wang@intel.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://lore.kernel.org/r/20240212153624.956116738@linutronix.de
Pull performance events updates from Ingo Molnar:
- Add branch stack counters ABI extension to better capture the growing
amount of information the PMU exposes via branch stack sampling.
There's matching tooling support.
- Fix race when creating the nr_addr_filters sysfs file
- Add Intel Sierra Forest and Grand Ridge intel/cstate PMU support
- Add Intel Granite Rapids, Sierra Forest and Grand Ridge uncore PMU
support
- Misc cleanups & fixes
* tag 'perf-core-2024-01-08' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf/x86/intel/uncore: Factor out topology_gidnid_map()
perf/x86/intel/uncore: Fix NULL pointer dereference issue in upi_fill_topology()
perf/x86/amd: Reject branch stack for IBS events
perf/x86/intel/uncore: Support Sierra Forest and Grand Ridge
perf/x86/intel/uncore: Support IIO free-running counters on GNR
perf/x86/intel/uncore: Support Granite Rapids
perf/x86/uncore: Use u64 to replace unsigned for the uncore offsets array
perf/x86/intel/uncore: Generic uncore_get_uncores and MMIO format of SPR
perf: Fix the nr_addr_filters fix
perf/x86/intel/cstate: Add Grand Ridge support
perf/x86/intel/cstate: Add Sierra Forest support
x86/smp: Export symbol cpu_clustergroup_mask()
perf/x86/intel/cstate: Cleanup duplicate attr_groups
perf/core: Fix narrow startup race when creating the perf nr_addr_filters sysfs file
perf/x86/intel: Support branch counters logging
perf/x86/intel: Reorganize attrs and is_visible
perf: Add branch_sample_call_stack
perf/x86: Add PERF_X86_EVENT_NEEDS_BRANCH_STACK flag
perf: Add branch stack counters
Pull x86 cleanups from Ingo Molnar:
- Change global variables to local
- Add missing kernel-doc function parameter descriptions
- Remove unused parameter from a macro
- Remove obsolete Kconfig entry
- Fix comments
- Fix typos, mostly scripted, manually reviewed
and a micro-optimization got misplaced as a cleanup:
- Micro-optimize the asm code in secondary_startup_64_no_verify()
* tag 'x86-cleanups-2024-01-08' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
arch/x86: Fix typos
x86/head_64: Use TESTB instead of TESTL in secondary_startup_64_no_verify()
x86/docs: Remove reference to syscall trampoline in PTI
x86/Kconfig: Remove obsolete config X86_32_SMP
x86/io: Remove the unused 'bw' parameter from the BUILDIO() macro
x86/mtrr: Document missing function parameters in kernel-doc
x86/setup: Make relocated_ramdisk a local variable of relocate_initrd()
When commit c59a1f106f ("KVM: x86/pmu: Add IA32_PEBS_ENABLE
MSR emulation for extended PEBS") switched the initialization of
cpuc->guest_switch_msrs to use compound literals, it screwed up
the boolean logic:
+ u64 pebs_mask = cpuc->pebs_enabled & x86_pmu.pebs_capable;
...
- arr[0].guest = intel_ctrl & ~cpuc->intel_ctrl_host_mask;
- arr[0].guest &= ~(cpuc->pebs_enabled & x86_pmu.pebs_capable);
+ .guest = intel_ctrl & (~cpuc->intel_ctrl_host_mask | ~pebs_mask),
Before the patch, the value of arr[0].guest would have been intel_ctrl &
~cpuc->intel_ctrl_host_mask & ~pebs_mask. The intent is to always treat
PEBS events as host-only because, while the guest runs, there is no way
to tell the processor about the virtual address where to put PEBS records
intended for the host.
Unfortunately, the new expression can be expanded to
(intel_ctrl & ~cpuc->intel_ctrl_host_mask) | (intel_ctrl & ~pebs_mask)
which makes no sense; it includes any bit that isn't *both* marked as
exclude_guest and using PEBS. So, reinstate the old logic. Another
way to write it could be "intel_ctrl & ~(cpuc->intel_ctrl_host_mask |
pebs_mask)", presumably the intention of the author of the faulty.
However, I personally find the repeated application of A AND NOT B to
be a bit more readable.
This shows up as guest failures when running concurrent long-running
perf workloads on the host, and was reported to happen with rcutorture.
All guests on a given host would die simultaneously with something like an
instruction fault or a segmentation violation.
Reported-by: Paul E. McKenney <paulmck@kernel.org>
Analyzed-by: Sean Christopherson <seanjc@google.com>
Tested-by: Paul E. McKenney <paulmck@kernel.org>
Cc: stable@vger.kernel.org
Fixes: c59a1f106f ("KVM: x86/pmu: Add IA32_PEBS_ENABLE MSR emulation for extended PEBS")
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The same as Sapphire Rapids, Granite Rapids also supports the discovery
table feature. All the basic uncore PMON information can be retrieved
from the discovery table which resides in the BIOS.
There are 4 new units are added on Granite Rapids, b2cmi, b2cxl, ubox,
and mdf_sbo. The layout of the counters is exactly the same as the
generic uncore counters. Only add a name for the new units. All the
details can be retrieved from the discovery table.
The description of the new units can be found at
https://www.intel.com/content/www/us/en/secure/content-details/772943/content-details.html
The other units, e.g., cha, iio, irp, pcu, and imc, are the same as
Sapphire Rapids.
Ignore the upi and b2upi units in the discovery table, which are broken
for now.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Ammy Yi <ammy.yi@intel.com>
Link: https://lore.kernel.org/r/20231117163939.2468007-3-kan.liang@linux.intel.com
The current perf doesn't save the complete address of an uncore unit.
The complete address of each unit is calculated by the base address +
offset. The type of the base address is u64, while the type of offset is
unsigned.
In the old platforms (without the discovery table method), the base
address and offset are hard coded in the driver. Perf can always use the
lowest address as the base address. Everything works well.
In the new platforms (starting from SPR), the discovery table provides
a complete address for all uncore units. To follow the current
framework/codes, when parsing the discovery table, the complete address
of the first box is stored as a base address. The offset of the
following units is calculated by the complete address of the unit minus
the base address (the address of the first unit). On GNR, the latter
units may have a lower address compared to the first unit. So the offset
is a negative value. The upper 32 bits are lost when casting a negative
u64 to an unsigned type.
Use u64 to replace unsigned for the uncore offsets array to correct the
above case. There is no functional change.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Ammy Yi <ammy.yi@intel.com>
Link: https://lore.kernel.org/r/20231117163939.2468007-2-kan.liang@linux.intel.com
When running perf-stat command on Intel hybrid platform, perf-stat
reports the following errors:
sudo taskset -c 7 ./perf stat -vvvv -e cpu_atom/instructions/ sleep 1
Opening: cpu/cycles/:HG
------------------------------------------------------------
perf_event_attr:
type 0 (PERF_TYPE_HARDWARE)
config 0xa00000000
disabled 1
------------------------------------------------------------
sys_perf_event_open: pid 0 cpu -1 group_fd -1 flags 0x8
sys_perf_event_open failed, error -16
Performance counter stats for 'sleep 1':
<not counted> cpu_atom/instructions/
It looks the cpu_atom/instructions/ event can't be enabled on atom PMU
even when the process is pinned on atom core. Investigation shows that
exclusive_event_init() helper always returns -EBUSY error in the perf
event creation. That's strange since the atom PMU should not be an
exclusive PMU.
Further investigation shows the issue was introduced by commit:
97588df87b ("perf/x86/intel: Add common intel_pmu_init_hybrid()")
The commit originally intents to clear the bit PERF_PMU_CAP_AUX_OUTPUT
from PMU capabilities if intel_cap.pebs_output_pt_available is not set,
but it incorrectly uses 'or' operation and leads to all PMU capabilities
bits are set to 1 except bit PERF_PMU_CAP_AUX_OUTPUT.
Testing this fix on Intel hybrid platforms, the observed issues
disappear.
Fixes: 97588df87b ("perf/x86/intel: Add common intel_pmu_init_hybrid()")
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20231121014628.729989-1-dapeng1.mi@linux.intel.com
A new module C6 Residency Counter is introduced in the Sierra Forest.
The scope of the new counter is module (A cluster of cores shared L2
cache). Create a brand new cstate_module PMU to profile the new counter.
The only differences between the new cstate_module PMU and the existing
cstate PMU are the scope and events.
Regarding the choice of the new cstate_module PMU name, the current
naming rule of a cstate PMU is "cstate_" + the scope of the PMU. The
scope of the PMU is the cores shared L2. On SRF, Intel calls it
"module", while the internal Linux sched code calls it "cluster". The
"cstate_module" is used as the new PMU name, because
- The Cstate PMU driver is a Intel specific driver. It doesn't impact
other ARCHs. The name makes it consistent with the documentation.
- The "cluster" mainly be used by the scheduler developer, while the
user of cstate PMU is more likely a researcher reading HW docs and
optimizing power.
- In the Intel's SDM, the "cluster" has a different meaning/scope for
topology. Using it will mislead the end users.
Besides the module C6, the core C1/C6 and pkg C6 residency counters are
supported in the Sierra Forest as well.
Suggested-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20231116142245.1233485-3-kan.liang@linux.intel.com
Pull x86 core updates from Thomas Gleixner:
- Limit the hardcoded topology quirk for Hygon CPUs to those which have
a model ID less than 4.
The newer models have the topology CPUID leaf 0xB correctly
implemented and are not affected.
- Make SMT control more robust against enumeration failures
SMT control was added to allow controlling SMT at boottime or
runtime. The primary purpose was to provide a simple mechanism to
disable SMT in the light of speculation attack vectors.
It turned out that the code is sensible to enumeration failures and
worked only by chance for XEN/PV. XEN/PV has no real APIC enumeration
which means the primary thread mask is not set up correctly. By
chance a XEN/PV boot ends up with smp_num_siblings == 2, which makes
the hotplug control stay at its default value "enabled". So the mask
is never evaluated.
The ongoing rework of the topology evaluation caused XEN/PV to end up
with smp_num_siblings == 1, which sets the SMT control to "not
supported" and the empty primary thread mask causes the hotplug core
to deny the bringup of the APS.
Make the decision logic more robust and take 'not supported' and 'not
implemented' into account for the decision whether a CPU should be
booted or not.
- Fake primary thread mask for XEN/PV
Pretend that all XEN/PV vCPUs are primary threads, which makes the
usage of the primary thread mask valid on XEN/PV. That is consistent
with because all of the topology information on XEN/PV is fake or
even non-existent.
- Encapsulate topology information in cpuinfo_x86
Move the randomly scattered topology data into a separate data
structure for readability and as a preparatory step for the topology
evaluation overhaul.
- Consolidate APIC ID data type to u32
It's fixed width hardware data and not randomly u16, int, unsigned
long or whatever developers decided to use.
- Cure the abuse of cpuinfo for persisting logical IDs.
Per CPU cpuinfo is used to persist the logical package and die IDs.
That's really not the right place simply because cpuinfo is subject
to be reinitialized when a CPU goes through an offline/online cycle.
Use separate per CPU data for the persisting to enable the further
topology management rework. It will be removed once the new topology
management is in place.
- Provide a debug interface for inspecting topology information
Useful in general and extremly helpful for validating the topology
management rework in terms of correctness or "bug" compatibility.
* tag 'x86-core-2023-10-29-v2' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (23 commits)
x86/apic, x86/hyperv: Use u32 in hv_snp_boot_ap() too
x86/cpu: Provide debug interface
x86/cpu/topology: Cure the abuse of cpuinfo for persisting logical ids
x86/apic: Use u32 for wakeup_secondary_cpu[_64]()
x86/apic: Use u32 for [gs]et_apic_id()
x86/apic: Use u32 for phys_pkg_id()
x86/apic: Use u32 for cpu_present_to_apicid()
x86/apic: Use u32 for check_apicid_used()
x86/apic: Use u32 for APIC IDs in global data
x86/apic: Use BAD_APICID consistently
x86/cpu: Move cpu_l[l2]c_id into topology info
x86/cpu: Move logical package and die IDs into topology info
x86/cpu: Remove pointless evaluation of x86_coreid_bits
x86/cpu: Move cu_id into topology info
x86/cpu: Move cpu_core_id into topology info
hwmon: (fam15h_power) Use topology_core_id()
scsi: lpfc: Use topology_core_id()
x86/cpu: Move cpu_die_id into topology info
x86/cpu: Move phys_proc_id into topology info
x86/cpu: Encapsulate topology information in cpuinfo_x86
...
Pull performance event updates from Ingo Molnar:
- Add AMD Unified Memory Controller (UMC) events introduced with Zen 4
- Simplify & clean up the uncore management code
- Fall back from RDPMC to RDMSR on certain uncore PMUs
- Improve per-package and cstate event reading
- Extend the Intel ref-cycles event to GP counters
- Fix Intel MTL event constraints
- Improve the Intel hybrid CPU handling code
- Micro-optimize the RAPL code
- Optimize perf_cgroup_switch()
- Improve large AUX area error handling
- Misc fixes and cleanups
* tag 'perf-core-2023-10-28' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (26 commits)
perf/x86/amd/uncore: Pass through error code for initialization failures, instead of -ENODEV
perf/x86/amd/uncore: Fix uninitialized return value in amd_uncore_init()
x86/cpu: Fix the AMD Fam 17h, Fam 19h, Zen2 and Zen4 MSR enumerations
perf: Optimize perf_cgroup_switch()
perf/x86/amd/uncore: Add memory controller support
perf/x86/amd/uncore: Add group exclusivity
perf/x86/amd/uncore: Use rdmsr if rdpmc is unavailable
perf/x86/amd/uncore: Move discovery and registration
perf/x86/amd/uncore: Refactor uncore management
perf/core: Allow reading package events from perf_event_read_local
perf/x86/cstate: Allow reading the package statistics from local CPU
perf/x86/intel/pt: Fix kernel-doc comments
perf/x86/rapl: Annotate 'struct rapl_pmus' with __counted_by
perf/core: Rename perf_proc_update_handler() -> perf_event_max_sample_rate_handler(), for readability
perf/x86/rapl: Fix "Using plain integer as NULL pointer" Sparse warning
perf/x86/rapl: Use local64_try_cmpxchg in rapl_event_update()
perf/x86/rapl: Stop doing cpu_relax() in the local64_cmpxchg() loop in rapl_event_update()
perf/core: Bail out early if the request AUX area is out of bound
perf/x86/intel: Extend the ref-cycles event to GP counters
perf/x86/intel: Fix broken fixed event constraints extension
...
The branch counters logging (A.K.A LBR event logging) introduces a
per-counter indication of precise event occurrences in LBRs. It can
provide a means to attribute exposed retirement latency to combinations
of events across a block of instructions. It also provides a means of
attributing Timed LBR latencies to events.
The feature is first introduced on SRF/GRR. It is an enhancement of the
ARCH LBR. It adds new fields in the LBR_INFO MSRs to log the occurrences
of events on the GP counters. The information is displayed by the order
of counters.
The design proposed in this patch requires that the events which are
logged must be in a group with the event that has LBR. If there are
more than one LBR group, the counters logging information only from the
current group (overflowed) are stored for the perf tool, otherwise the
perf tool cannot know which and when other groups are scheduled
especially when multiplexing is triggered. The user can ensure it uses
the maximum number of counters that support LBR info (4 by now) by
making the group large enough.
The HW only logs events by the order of counters. The order may be
different from the order of enabling which the perf tool can understand.
When parsing the information of each branch entry, convert the counter
order to the enabled order, and store the enabled order in the extension
space.
Unconditionally reset LBRs for an LBR event group when it's deleted. The
logged counter information is only valid for the current LBR group. If
another LBR group is scheduled later, the information from the stale
LBRs would be otherwise wrongly interpreted.
Add a sanity check in intel_pmu_hw_config(). Disable the feature if other
counter filters (inv, cmask, edge, in_tx) are set or LBR call stack mode
is enabled. (For the LBR call stack mode, we cannot simply flush the
LBR, since it will break the call stack. Also, there is no obvious usage
with the call stack mode for now.)
Only applying the PERF_SAMPLE_BRANCH_COUNTERS doesn't require any branch
stack setup.
Expose the maximum number of supported counters and the width of the
counters into the sysfs. The perf tool can use the information to parse
the logged counters in each branch.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20231025201626.3000228-5-kan.liang@linux.intel.com
Currently, branch_sample_type !=0 is used to check whether a branch
stack setup is required. But it doesn't check the sample type,
unnecessary branch stack setup may be done for a counting event. E.g.,
perf record -e "{branch-instructions,branch-misses}:S" -j any
Also, the event only with the new PERF_SAMPLE_BRANCH_COUNTERS branch
sample type may not require a branch stack setup either.
Add a new flag NEEDS_BRANCH_STACK to indicate whether the event requires
a branch stack setup. Replace the needs_branch_stack() by checking the
new flag.
The counting event check is implemented here. The later patch will take
the new PERF_SAMPLE_BRANCH_COUNTERS into account.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20231025201626.3000228-2-kan.liang@linux.intel.com
Currently, the additional information of a branch entry is stored in a
u64 space. With more and more information added, the space is running
out. For example, the information of occurrences of events will be added
for each branch.
Two places were suggested to append the counters.
https://lore.kernel.org/lkml/20230802215814.GH231007@hirez.programming.kicks-ass.net/
One place is right after the flags of each branch entry. It changes the
existing struct perf_branch_entry. The later ARCH specific
implementation has to be really careful to consistently pick
the right struct.
The other place is right after the entire struct perf_branch_stack.
The disadvantage is that the pointer of the extra space has to be
recorded. The common interface perf_sample_save_brstack() has to be
updated.
The latter is much straightforward, and should be easily understood and
maintained. It is implemented in the patch.
Add a new branch sample type, PERF_SAMPLE_BRANCH_COUNTERS, to indicate
the event which is recorded in the branch info.
The "u64 counters" may store the occurrences of several events. The
information regarding the number of events/counters and the width of
each counter should be exposed via sysfs as a reference for the perf
tool. Define the branch_counter_nr and branch_counter_width ABI here.
The support will be implemented later in the Intel-specific patch.
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20231025201626.3000228-1-kan.liang@linux.intel.com
Unified Memory Controller (UMC) events were introduced with Zen 4 as a
part of the Performance Monitoring Version 2 (PerfMonV2) enhancements.
An event is specified using the EventSelect bits and the RdWrMask bits
can be used for additional filtering of read and write requests.
As of now, a maximum of 12 channels of DDR5 are available on each socket
and each channel is controlled by a dedicated UMC. Each UMC, in turn,
has its own set of performance monitoring counters.
Since the MSR address space for the UMC PERF_CTL and PERF_CTR registers
are reused across sockets, uncore groups are created on the basis of
socket IDs. Hence, group exclusivity is mandatory while opening events
so that events for an UMC can only be opened on CPUs which are on the
same socket as the corresponding memory channel.
For each socket, the total number of available UMC counters and active
memory channels are determined from CPUID leaf 0x80000022 EBX and ECX
respectively. Usually, on Zen 4, each UMC has four counters.
MSR assignments are determined on the basis of active UMCs. E.g. if
UMCs 1, 4 and 9 are active for a given socket, then
* UMC 1 gets MSRs 0xc0010800 to 0xc0010807 as PERF_CTLs and PERF_CTRs
* UMC 4 gets MSRs 0xc0010808 to 0xc001080f as PERF_CTLs and PERF_CTRs
* UMC 9 gets MSRs 0xc0010810 to 0xc0010817 as PERF_CTLs and PERF_CTRs
If there are sockets without any online CPUs when the amd_uncore driver
is loaded, UMCs for such sockets will not be discoverable since the
mechanism relies on executing the CPUID instruction on an online CPU
from the socket.
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/b25f391205c22733493abec1ed850b71784edc5f.1696425185.git.sandipan.das@amd.com
In some cases, it may be necessary to restrict opening PMU events to a
subset of CPUs. E.g. Unified Memory Controller (UMC) PMUs are specific
to each active memory channel and the MSR address space for the PERF_CTL
and PERF_CTR registers is reused on each socket. Thus, opening events
for a specific UMC PMU should be restricted to CPUs belonging to the
same socket as that of the UMC. The "cpumask" of the PMU should also
reflect this accordingly.
Uncore PMUs which require this can use the new group attribute in struct
amd_uncore_pmu to set a valid group ID during the scan() phase. Later,
during init(), an uncore context for a CPU will be unavailable if the
group ID does not match.
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/937d6d71010a48ea4e069f4904b3116a5f99ecdf.1696425185.git.sandipan.das@amd.com
Uncore PMUs have traditionally been registered in the module init path.
This is fine for the existing DF and L3 PMUs since the CPUID information
does not vary across CPUs but not for the memory controller (UMC) PMUs
since information like active memory channels can vary for each socket
depending on how the DIMMs have been physically populated.
To overcome this, the discovery of PMU information using CPUID is moved
to the startup of UNCORE_STARTING. This cannot be done in the startup of
UNCORE_PREP since the hotplug callback does not run on the CPU that is
being brought online.
Previously, the startup of UNCORE_PREP was used for allocating uncore
contexts following which, the startup of UNCORE_STARTING was used to
find and reuse an existing sibling context, if possible. Any unused
contexts were added to a list for reclaimation later during the startup
of UNCORE_ONLINE.
Since all required CPUID info is now available only after the startup of
UNCORE_STARTING has completed, context allocation has been moved to the
startup of UNCORE_ONLINE. Before allocating contexts, the first CPU that
comes online has to take up the additional responsibility of registering
the PMUs. This is a one-time process though. Since sibling discovery now
happens prior to deciding whether a new context is required, there is no
longer a need to track and free up unused contexts.
The teardown of UNCORE_ONLINE and UNCORE_PREP functionally remain the
same.
Overall, the flow of control described above is achieved using the
following handlers for managing uncore PMUs. It is mandatory to define
them for each type of uncore PMU.
* scan() runs during startup of UNCORE_STARTING and collects PMU info
using CPUID.
* init() runs during startup of UNCORE_ONLINE, registers PMUs and sets
up uncore contexts.
* move() runs during teardown of UNCORE_ONLINE and migrates uncore
contexts to a shared sibling, if possible.
* free() runs during teardown of UNCORE_PREP and frees up uncore
contexts.
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/e6c447e48872fcab8452e0dd81b1c9cb09f39eb4.1696425185.git.sandipan.das@amd.com
Since struct amd_uncore is used to manage per-cpu contexts, rename it to
amd_uncore_ctx in order to better reflect its purpose. Add a new struct
amd_uncore_pmu to encapsulate all attributes which are shared by per-cpu
contexts for a corresponding PMU. These include the number of counters,
active mask, MSR and RDPMC base addresses, etc. Since the struct pmu is
now embedded, the corresponding amd_uncore_pmu for a given event can be
found by simply using container_of().
Finally, move all PMU-specific code to separate functions. While the
original event management functions continue to provide the base
functionality, all PMU-specific quirks and customizations are applied in
separate functions.
The motivation is to simplify the management of uncore PMUs.
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/24b38c49a5dae65d8c96e5d75a2b96ae97aaa651.1696425185.git.sandipan.das@amd.com
