Function paicrypt_event_init() initializes the PMU device driver
specific details for an event. It is called once per event creation.
The function paicrypt_event_init() is not necessarily executed on
that CPU the event will be used for.
When an event is activated, function paicrypt_start() is used to
start the event on that CPU.
The per CPU data structure struct paicrypt_map has a pointer to
the event which is active for a particular CPU. This pointer is
set in function paicrypt_start() to point to the currently installed
event. There is no need to also set this pointer in function
paicrypt_event_init() where is might be assigned to the wrong CPU.
Therefore remove this assignment in paicrypt_event_init().
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Event count value is initialized and set to zero in function
paicrypt_start(). This function is called once per CPU when an
event is started on that CPU. This leads to event count value
being set to zero as many times as there are online CPUs.
This is not necessary. The event count value is bound to the event
and it is sufficient to initialize the event counter once at
event creation time. This is done when the event structure
is dynamicly allocated with __GFP_ZERO flag. This sets
member count to zero.
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Use define PERF_ATTACH_TASK bit in event->attach_state to determine
system wide invocation or per task invocation in event initialization.
This bit is set in common code and before calling PMU device driver
specific event initialization.
It is set once and never changes. It is save to use and also in
sync with other PMU specific code.
No functional change.
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Struct paicrypt_map is a data structure and is statically defined
for each possible CPU. Rework this and replace it by dynamically
allocated data structures created when a perf_event_open() system call
is invoked.
It is replaced by an array of pointers to all possible CPUs and
reference counting. The array of pointers is allocated when the first
event is created. For each online CPU an event is installed on, a struct
paicrypt_map is allocated and a pointer to struct cpu_cf_events is
stored in the array:
CPU 0 1 2 3 ... N
+---+---+---+---+---+---+
paicrypt_root::mapptr--> | * | | | |...| |
+-|-+---+---+---+---+---+
|
|
\|/
+--------------+
| paicrypt_map |
+--------------+
With this approach the large data structure is only allocated when
an event is actually installed and used.
Also implement proper reference counting for allocation and removal.
PAI crypto counter events can not be created when a CPU hot plug
add is processed. This means a CPU hot plug add does not get
the necessary PAI event to record PAI cryptography counter increments
on the newly added CPU. There is no possibility to notify user space
of a new CPU and the necessary event infrastructure assoiciated with
the file descriptor returned by perf_event_open() system call.
However system call perf_event_open() can use the newly added CPU
when issued after the CPU hot plug add.
Kernel CPU hot plug remove deletes the CPU and stops the PAI counters on
that CPU. When the process closes the file descriptor associated
with that event, the event's destroy() function removes any
allocated data structures and adjusts the reference counts.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Acked-by: Heiko Carstens <hca@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Use control register bit defines instead of plain numbers where
possible.
Reviewed-by: Alexander Gordeev <agordeev@linux.ibm.com>
Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Add local and system prefix to some functions to clarify they change
control register contents on either the local CPU or the on all CPUs.
This results in the following API:
Two defines which load and save multiple control registers.
The defines correlate with the following C prototypes:
void __local_ctl_load(unsigned long *, unsigned int cr_low, unsigned int cr_high);
void __local_ctl_store(unsigned long *, unsigned int cr_low, unsigned int cr_high);
Two functions which locally set or clear one bit for a specified
control register:
void local_ctl_set_bit(unsigned int cr, unsigned int bit);
void local_ctl_clear_bit(unsigned int cr, unsigned int bit);
Two functions which set or clear one bit for a specified control
register on all CPUs:
void system_ctl_set_bit(unsigned int cr, unsigned int bit);
void system_ctl_clear_bit(unsigend int cr, unsigned int bit);
Reviewed-by: Alexander Gordeev <agordeev@linux.ibm.com>
Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Rename ctl_reg.h to ctlreg.h so it matches not only ctlreg.c but also
other control register related function, union, and structure names,
which all come with a ctlreg prefix.
Reviewed-by: Alexander Gordeev <agordeev@linux.ibm.com>
Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
The s390 PMU of PAI crypto counters uses atomic_t for reference
counting. Replace this with the proper data type refcount_t.
No functional change.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
When we save the raw_data to the perf sample data, we need to update
the sample flags and the dynamic size. To make sure this is done
consistently, add the perf_sample_save_raw_data() helper and convert
all call sites.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20230118060559.615653-4-namhyung@kernel.org
- Thoroughly rewrite the data structures that implement perf task context handling,
with the goal of fixing various quirks and unfeatures both in already merged,
and in upcoming proposed code.
The old data structure is the per task and per cpu perf_event_contexts:
task_struct::perf_events_ctxp[] <-> perf_event_context <-> perf_cpu_context
^ | ^ | ^
`---------------------------------' | `--> pmu ---'
v ^
perf_event ------'
In this new design this is replaced with a single task context and
a single CPU context, plus intermediate data-structures:
task_struct::perf_event_ctxp -> perf_event_context <- perf_cpu_context
^ | ^ ^
`---------------------------' | |
| | perf_cpu_pmu_context <--.
| `----. ^ |
| | | |
| v v |
| ,--> perf_event_pmu_context |
| | |
| | |
v v |
perf_event ---> pmu ----------------'
[ See commit bd27568117 for more details. ]
This rewrite was developed by Peter Zijlstra and Ravi Bangoria.
- Optimize perf_tp_event()
- Update the Intel uncore PMU driver, extending it with UPI topology discovery
on various hardware models.
- Misc fixes & cleanups
Signed-off-by: Ingo Molnar <mingo@kernel.org>
-----BEGIN PGP SIGNATURE-----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=KNaI
-----END PGP SIGNATURE-----
Merge tag 'perf-core-2022-12-12' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull perf events updates from Ingo Molnar:
- Thoroughly rewrite the data structures that implement perf task
context handling, with the goal of fixing various quirks and
unfeatures both in already merged, and in upcoming proposed code.
The old data structure is the per task and per cpu
perf_event_contexts:
task_struct::perf_events_ctxp[] <-> perf_event_context <-> perf_cpu_context
^ | ^ | ^
`---------------------------------' | `--> pmu ---'
v ^
perf_event ------'
In this new design this is replaced with a single task context and a
single CPU context, plus intermediate data-structures:
task_struct::perf_event_ctxp -> perf_event_context <- perf_cpu_context
^ | ^ ^
`---------------------------' | |
| | perf_cpu_pmu_context <--.
| `----. ^ |
| | | |
| v v |
| ,--> perf_event_pmu_context |
| | |
| | |
v v |
perf_event ---> pmu ----------------'
[ See commit bd27568117 for more details. ]
This rewrite was developed by Peter Zijlstra and Ravi Bangoria.
- Optimize perf_tp_event()
- Update the Intel uncore PMU driver, extending it with UPI topology
discovery on various hardware models.
- Misc fixes & cleanups
* tag 'perf-core-2022-12-12' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (25 commits)
perf/x86/intel/uncore: Fix reference count leak in __uncore_imc_init_box()
perf/x86/intel/uncore: Fix reference count leak in snr_uncore_mmio_map()
perf/x86/intel/uncore: Fix reference count leak in hswep_has_limit_sbox()
perf/x86/intel/uncore: Fix reference count leak in sad_cfg_iio_topology()
perf/x86/intel/uncore: Make set_mapping() procedure void
perf/x86/intel/uncore: Update sysfs-devices-mapping file
perf/x86/intel/uncore: Enable UPI topology discovery for Sapphire Rapids
perf/x86/intel/uncore: Enable UPI topology discovery for Icelake Server
perf/x86/intel/uncore: Get UPI NodeID and GroupID
perf/x86/intel/uncore: Enable UPI topology discovery for Skylake Server
perf/x86/intel/uncore: Generalize get_topology() for SKX PMUs
perf/x86/intel/uncore: Disable I/O stacks to PMU mapping on ICX-D
perf/x86/intel/uncore: Clear attr_update properly
perf/x86/intel/uncore: Introduce UPI topology type
perf/x86/intel/uncore: Generalize IIO topology support
perf/core: Don't allow grouping events from different hw pmus
perf/amd/ibs: Make IBS a core pmu
perf: Fix function pointer case
perf/x86/amd: Remove the repeated declaration
perf: Fix possible memleak in pmu_dev_alloc()
...
There have been various issues and limitations with the way perf uses
(task) contexts to track events. Most notable is the single hardware
PMU task context, which has resulted in a number of yucky things (both
proposed and merged).
Notably:
- HW breakpoint PMU
- ARM big.little PMU / Intel ADL PMU
- Intel Branch Monitoring PMU
- AMD IBS PMU
- S390 cpum_cf PMU
- PowerPC trace_imc PMU
*Current design:*
Currently we have a per task and per cpu perf_event_contexts:
task_struct::perf_events_ctxp[] <-> perf_event_context <-> perf_cpu_context
^ | ^ | ^
`---------------------------------' | `--> pmu ---'
v ^
perf_event ------'
Each task has an array of pointers to a perf_event_context. Each
perf_event_context has a direct relation to a PMU and a group of
events for that PMU. The task related perf_event_context's have a
pointer back to that task.
Each PMU has a per-cpu pointer to a per-cpu perf_cpu_context, which
includes a perf_event_context, which again has a direct relation to
that PMU, and a group of events for that PMU.
The perf_cpu_context also tracks which task context is currently
associated with that CPU and includes a few other things like the
hrtimer for rotation etc.
Each perf_event is then associated with its PMU and one
perf_event_context.
*Proposed design:*
New design proposed by this patch reduce to a single task context and
a single CPU context but adds some intermediate data-structures:
task_struct::perf_event_ctxp -> perf_event_context <- perf_cpu_context
^ | ^ ^
`---------------------------' | |
| | perf_cpu_pmu_context <--.
| `----. ^ |
| | | |
| v v |
| ,--> perf_event_pmu_context |
| | |
| | |
v v |
perf_event ---> pmu ----------------'
With the new design, perf_event_context will hold all events for all
pmus in the (respective pinned/flexible) rbtrees. This can be achieved
by adding pmu to rbtree key:
{cpu, pmu, cgroup, group_index}
Each perf_event_context carries a list of perf_event_pmu_context which
is used to hold per-pmu-per-context state. For example, it keeps track
of currently active events for that pmu, a pmu specific task_ctx_data,
a flag to tell whether rotation is required or not etc.
Additionally, perf_cpu_pmu_context is used to hold per-pmu-per-cpu
state like hrtimer details to drive the event rotation, a pointer to
perf_event_pmu_context of currently running task and some other
ancillary information.
Each perf_event is associated to it's pmu, perf_event_context and
perf_event_pmu_context.
Further optimizations to current implementation are possible. For
example, ctx_resched() can be optimized to reschedule only single pmu
events.
Much thanks to Ravi for picking this up and pushing it towards
completion.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Co-developed-by: Ravi Bangoria <ravi.bangoria@amd.com>
Signed-off-by: Ravi Bangoria <ravi.bangoria@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20221008062424.313-1-ravi.bangoria@amd.com
Rename structure member users to active_events to make it consistent
with PMU pai_ext. Also use the same prefix syntax for increment and
decrement operators in both PMUs.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Rework the mapped buffer reference count in PMU pai_crypto
to match the same technique as in PMU pai_ext.
This simplifies the logic.
Do not count the individual number of counter and sampling
processes. Remember the type of access and the total number of
references to the buffer.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Use the new sample_flags to indicate whether the raw data field is
filled by the PMU driver. Although it could check with the NULL,
follow the same rule with other fields.
Remove the raw field from the perf_sample_data_init() to minimize
the number of cache lines touched.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20220921220032.2858517-2-namhyung@kernel.org
Two different events such as pai_crypto/KM_AES_128/ and
pai_crypto/KM_AES_192/ can be installed multiple times on the same CPU
and the events are executed concurrently:
# perf stat -e pai_crypto/KM_AES_128/ -C0 -a -- sleep 5 &
# sleep 2
# perf stat -e pai_crypto/KM_AES_192/ -C0 -a -- true
This results in the first event being installed two times with two seconds
delay. The kernel does install the second event after the first
event has been deleted and re-added, as can be seen in the traces:
13:48:47.600350 paicrypt_start event 0x1007 (event KM_AES_128)
13:48:49.599359 paicrypt_stop event 0x1007 (event KM_AES_128)
13:48:49.599198 paicrypt_start event 0x1007
13:48:49.599199 paicrypt_start event 0x1008
13:48:49.599921 paicrypt_event_destroy event 0x1008
13:48:52.601507 paicrypt_event_destroy event 0x1007
This is caused by functions event_sched_in() and event_sched_out() which
call the PMU's add() and start() functions on schedule_in and the PMU's
stop() and del() functions on schedule_out. This is correct for events
attached to processes. The pai_crypto events are system-wide events
and not attached to processes.
Since the kernel common code can not be changed easily, fix this issue
and do not reset the event count value to zero each time the event is
added and started. Instead use a flag and zero the event count value
only when called immediately after the event has been initialized.
Therefore only the first invocation of the the event's add() function
initializes the event count value to zero. The following invocations
of the event's add() function leave the current event count value
untouched.
Fixes: 39d62336f5 ("s390/pai: add support for cryptography counters")
Reported-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
The pai_crypto PMU has to check the event number. It has to be in
the supported range. This is not the case, the lower limit is not
checked. Fix this and obey the lower limit.
Fixes: 39d62336f5 ("s390/pai: add support for cryptography counters")
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Suggested-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Reviewed-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
PMU device driver perf_pai_crypto supports Processor Activity
Instrumentation (PAI), available with IBM z16:
- maps a full page to lowcore address 0x1500.
- uses CR0 bit 13 to turn PAI crypto counting on and off.
- creates a sample with raw data on each context switch out when
at context switch some mapped counters have a value of nonzero.
This device driver only supports CPU wide context, no task context
is allowed.
Support for counting:
- one or more counters can be specified using
perf stat -e pai_crypto/xxx/
where xxx stands for the counter event name. Multiple invocation
of this command is possible. The counter names are listed in
/sys/devices/pai_crypto/events directory.
- one special counters can be specified using
perf stat -e pai_crypto/CRYPTO_ALL/
which returns the sum of all incremented crypto counters.
- one event pai_crypto/CRYPTO_ALL/ is reserved for sampling.
No multiple invocations are possible. The event collects data at
context switch out and saves them in the ring buffer.
Add qpaci assembly instruction to query supported memory mapped crypto
counters. It returns the number of counters (no holes allowed in that
range).
The PAI crypto counter events are system wide and can not be executed
in parallel. Therefore some restrictions documented in function
paicrypt_busy apply.
In particular event CRYPTO_ALL for sampling must run exclusive.
Only counting events can run in parallel.
PAI crypto counter events can not be created when a CPU hot plug
add is processed. This means a CPU hot plug add does not get
the necessary PAI event to record PAI cryptography counter increments
on the newly added CPU. CPU hot plug remove removes the event and
terminates the counting of PAI counters immediately.
Co-developed-by: Sven Schnelle <svens@linux.ibm.com>
Signed-off-by: Sven Schnelle <svens@linux.ibm.com>
Reviewed-by: Juergen Christ <jchrist@linux.ibm.com>
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Link: https://lore.kernel.org/r/20220504062351.2954280-3-tmricht@linux.ibm.com
Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
