[PATCH v10 0/5] arm64 userspace counter support

[PATCH v10 0/5] arm64 userspace counter support

[Rob Herring]

Another version of arm64 userspace counter access support.

The arm64 support departs from the x86 implementation by requiring the user
to explicitly request user access (via attr.config1) and only enables access
for task bound events. Since usage is explicitly requested, access is
enabled at perf_event_open() rather than on mmap() as that greatly
simplifies the implementation. Rather than trying to lock down the access
as the x86 implementation has been doing, we can start with only a limited
use case enabled and later expand it if needed.

I've run this version thru Vince's perf tests[12] with arm64 support added.
I wish I'd found these tests sooner...

This originally resurrected Raphael's series[1] to enable userspace counter
access on arm64. My previous versions are here[2][3][4][5][6][7][8][9][10].
A git branch is here[11].

Changes in v10:
 - Drop changing event_mapped/event_unmapped ops to run on the event's
   current CPU. This won't work for x86 where any thread sharing an mm
   context will have rdpmc enabled. Instead, simply track user access
   events when added to a perf context and use that count.
 - Documentation for the sysctl disable

Changes in v9:
 - Reworked x86 and perf core to handle user access tracking and call
   .event_mapped() and .event_unmapped() on the CPU with the event like
   other changes to events.
 - Use sysctl instead of sysfs to disable user access.

Changes in v8:
 - Restrict user access to thread bound events which simplifies the
   implementation. A couple of perf core changes (patches 1 and 2) are
   needed to do this.
 - Always require the user to request userspace access.

Changes in v7:
 - Handling of dirty counter leakage and reworking of context switch and
   user access enabling. The .sched_task hook and undef instruction handler
   are now utilized. (Patch 3)
 - Add a userspace disable switch like x86. (Patch 5)

Changes in v6:
 - Reworking of the handling of 64-bit counters and user access. There's
   a new config1 flag to request user access. This takes priority over
   the 64-bit flag and the user will get the maximum size the h/w
   supports without chaining.
 - The libperf evsel mmap struct is stored in its own xyarray
 - New tests for user 64-bit and 32-bit counters
 - Rebase to v5.12-rc2

Changes in v5:
 - Limit enabling/disabling access to CPUs associated with the PMU
   (supported_cpus) and with the mm_struct matching current->active_mm.
   The x86 method of using mm_cpumask doesn't work for arm64 as it is not
   updated.
 - Only set cap_user_rdpmc if event is on current cpu. See patch 2.
 - Create an mmap for every event in an evsel. This results in some changes
   to the libperf mmap API from the last version.
 - Rebase to v5.11-rc2

Changes in v4:
 - Dropped 'arm64: pmu: Add hook to handle pmu-related undefined instructions'.
   The onus is on userspace to pin itself to a homogeneous subset of CPUs
   and avoid any aborts on heterogeneous systems, so the hook is not needed.
 - Make perf_evsel__mmap() take pages rather than bytes for size
 - Fix building arm64 heterogeneous test.

Changes in v3:
 - Dropped removing x86 rdpmc test until libperf tests can run via 'perf test'
 - Added verbose prints for tests
 - Split adding perf_evsel__mmap() to separate patch

The following changes to the arm64 support have been made compared to
Raphael's last version:

The major change is support for heterogeneous systems with some
restrictions. Specifically, userspace must pin itself to like CPUs, open
a specific PMU by type, and use h/w specific events. The tests have been
reworked to demonstrate this.

Chained events are not supported. The problem with supporting chained
events was there's no way to distinguish between a chained event and a
native 64-bit counter. We could add some flag, but do self monitoring
processes really need that? Native 64-bit counters are supported if the
PMU h/w has support. As there's already an explicit ABI to request 64-bit
counters, userspace can request 64-bit counters and if user
access is not enabled, then it must retry with 32-bit counters.

Prior versions broke the build on arm32 (surprisingly never caught by
0-day). As a result, event_mapped and event_unmapped implementations have
been moved into the arm64 code.

There was a bug in that pmc_width was not set in the user page. The tests
now check for this.

The documentation has been converted to rST. I've added sections on
chained events and heterogeneous.

Rob


[1] https://lore.kernel.org/r/20190822144220.27860-1-raphael.gault@arm.com/
[2] https://lore.kernel.org/r/20200707205333.624938-1-robh@kernel.org/
[3] https://lore.kernel.org/r/20200828205614.3391252-1-robh@kernel.org/
[4] https://lore.kernel.org/r/20200911215118.2887710-1-robh@kernel.org/
[5] https://lore.kernel.org/r/20201001140116.651970-1-robh@kernel.org/
[6] https://lore.kernel.org/r/20210114020605.3943992-1-robh@kernel.org/
[7] https://lore.kernel.org/r/20210311000837.3630499-1-robh@kernel.org/
[8] https://lore.kernel.org/r/20210420031511.2348977-1-robh@kernel.org/
[9] https://lore.kernel.org/r/20210517195405.3079458-1-robh@kernel.org/
[10] https://lore.kernel.org/all/20210806225123.1958497-1-robh@kernel.org/
[11] git://git.kernel.org/pub/scm/linux/kernel/git/robh/linux.git arm64-user-perf-event-v10
[12] https://github.com/deater/perf_event_tests

Raphael Gault (1):
  Documentation: arm64: Document PMU counters access from userspace

Rob Herring (4):
  x86: perf: Move RDPMC event flag to a common definition
  perf: Add a counter for number of user access events in context
  arm64: perf: Add userspace counter access disable switch
  arm64: perf: Enable PMU counter userspace access for perf event

 Documentation/admin-guide/sysctl/kernel.rst |  11 ++
 Documentation/arm64/perf.rst                |  73 +++++++++++-
 arch/arm64/kernel/perf_event.c              | 126 ++++++++++++++++++--
 arch/x86/events/core.c                      |  10 +-
 arch/x86/events/perf_event.h                |   2 +-
 include/linux/perf_event.h                  |   3 +
 kernel/events/core.c                        |   4 +
 7 files changed, 215 insertions(+), 14 deletions(-)

--
2.30.2

[PATCH v10 1/5] x86: perf: Move RDPMC event flag to a common definition

[Rob Herring]

In preparation to enable user counter access on arm64 and to move some
of the user access handling to perf core, create a common event flag for
user counter access and convert x86 to use it.

Since the architecture specific flags start at the LSB, starting at the
MSB for common flags.

Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Borislav Petkov <bp@alien8.de>
Cc: x86@kernel.org
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: linux-perf-users@vger.kernel.org
Signed-off-by: Rob Herring <robh@kernel.org>
---
 arch/x86/events/core.c       | 10 +++++-----
 arch/x86/events/perf_event.h |  2 +-
 include/linux/perf_event.h   |  2 ++
 3 files changed, 8 insertions(+), 6 deletions(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 2a57dbed4894..2bd50fc061e1 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -2469,7 +2469,7 @@ static int x86_pmu_event_init(struct perf_event *event)
 
 	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
 	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
-		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
+		event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
 
 	return err;
 }
@@ -2503,7 +2503,7 @@ void perf_clear_dirty_counters(void)
 
 static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
 {
-	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
+	if (!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT))
 		return;
 
 	/*
@@ -2524,7 +2524,7 @@ static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
 
 static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
 {
-	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
+	if (!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT))
 		return;
 
 	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
@@ -2535,7 +2535,7 @@ static int x86_pmu_event_idx(struct perf_event *event)
 {
 	struct hw_perf_event *hwc = &event->hw;
 
-	if (!(hwc->flags & PERF_X86_EVENT_RDPMC_ALLOWED))
+	if (!(hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
 		return 0;
 
 	if (is_metric_idx(hwc->idx))
@@ -2718,7 +2718,7 @@ void arch_perf_update_userpage(struct perf_event *event,
 	userpg->cap_user_time = 0;
 	userpg->cap_user_time_zero = 0;
 	userpg->cap_user_rdpmc =
-		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
+		!!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT);
 	userpg->pmc_width = x86_pmu.cntval_bits;
 
 	if (!using_native_sched_clock() || !sched_clock_stable())
diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
index e3ac05c97b5e..49f68b15745f 100644
--- a/arch/x86/events/perf_event.h
+++ b/arch/x86/events/perf_event.h
@@ -73,7 +73,7 @@ static inline bool constraint_match(struct event_constraint *c, u64 ecode)
 #define PERF_X86_EVENT_PEBS_NA_HSW	0x0010 /* haswell style datala, unknown */
 #define PERF_X86_EVENT_EXCL		0x0020 /* HT exclusivity on counter */
 #define PERF_X86_EVENT_DYNAMIC		0x0040 /* dynamic alloc'd constraint */
-#define PERF_X86_EVENT_RDPMC_ALLOWED	0x0080 /* grant rdpmc permission */
+
 #define PERF_X86_EVENT_EXCL_ACCT	0x0100 /* accounted EXCL event */
 #define PERF_X86_EVENT_AUTO_RELOAD	0x0200 /* use PEBS auto-reload */
 #define PERF_X86_EVENT_LARGE_PEBS	0x0400 /* use large PEBS */
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index fe156a8170aa..12debf008d39 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -142,6 +142,8 @@ struct hw_perf_event {
 			int		event_base_rdpmc;
 			int		idx;
 			int		last_cpu;
+
+#define PERF_EVENT_FLAG_USER_READ_CNT	0x80000000
 			int		flags;
 
 			struct hw_perf_event_extra extra_reg;
-- 
2.30.2

[PATCH v10 2/5] perf: Add a counter for number of user access events in context

[Rob Herring]

For controlling user space counter access, we need to know if any event
in a context (currently scheduled or not) is using user space counters.
Walking the context's list of events would be slow, so add a counter
to track this.

Signed-off-by: Rob Herring <robh@kernel.org>
---
v10:
 - Re-added.
 - Maintain the count in the perf core
v9:
 - Dropped
v8:
 - new patch
---
 include/linux/perf_event.h | 1 +
 kernel/events/core.c       | 4 ++++
 2 files changed, 5 insertions(+)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 12debf008d39..4f82a4d47139 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -821,6 +821,7 @@ struct perf_event_context {
 
 	int				nr_events;
 	int				nr_active;
+	int				nr_user;
 	int				is_active;
 	int				nr_stat;
 	int				nr_freq;
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 744e8726c5b2..01290d150da3 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -1808,6 +1808,8 @@ list_add_event(struct perf_event *event, struct perf_event_context *ctx)
 
 	list_add_rcu(&event->event_entry, &ctx->event_list);
 	ctx->nr_events++;
+	if (event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT)
+		ctx->nr_user++;
 	if (event->attr.inherit_stat)
 		ctx->nr_stat++;
 
@@ -1999,6 +2001,8 @@ list_del_event(struct perf_event *event, struct perf_event_context *ctx)
 	event->attach_state &= ~PERF_ATTACH_CONTEXT;
 
 	ctx->nr_events--;
+	if (event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT)
+		ctx->nr_user--;
 	if (event->attr.inherit_stat)
 		ctx->nr_stat--;
 
-- 
2.30.2

[PATCH v10 3/5] arm64: perf: Add userspace counter access disable switch

[Rob Herring]

Like x86, some users may want to disable userspace PMU counter
altogether. Add a sysctl 'perf_user_access' file to control userspace
counter access. The default is '0' which is disabled. Writing '1'
enables access.

Note that x86 also supports writing '2' to globally enable user access.
As there's not existing userspace support to worry about, this shouldn't
be necessary for Arm. It could be added later if the need arises.

Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: linux-perf-users@vger.kernel.org
Acked-by: Will Deacon <will@kernel.org>
Signed-off-by: Rob Herring <robh@kernel.org>
---
v10:
 - Add documentation
 - Use a custom handler (needed on the next patch)
v9:
 - Use sysctl instead of sysfs attr
 - Default to disabled
v8:
 - New patch
---
 Documentation/admin-guide/sysctl/kernel.rst | 11 +++++++++
 arch/arm64/kernel/perf_event.c              | 27 +++++++++++++++++++++
 2 files changed, 38 insertions(+)

diff --git a/Documentation/admin-guide/sysctl/kernel.rst b/Documentation/admin-guide/sysctl/kernel.rst
index 426162009ce9..346a0dba5703 100644
--- a/Documentation/admin-guide/sysctl/kernel.rst
+++ b/Documentation/admin-guide/sysctl/kernel.rst
@@ -905,6 +905,17 @@ enabled, otherwise writing to this file will return ``-EBUSY``.
 The default value is 8.
 
 
+perf_user_access (arm64 only)
+=================================
+
+Controls user space access for reading perf event counters. When set to 1,
+user space can read performance monitor counter registers directly.
+
+The default value is 0 (access disabled).
+
+See Documentation/arm64/perf.rst for more information.
+
+
 pid_max
 =======
 
diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
index b4044469527e..a8f8dd741aeb 100644
--- a/arch/arm64/kernel/perf_event.c
+++ b/arch/arm64/kernel/perf_event.c
@@ -286,6 +286,8 @@ static const struct attribute_group armv8_pmuv3_events_attr_group = {
 PMU_FORMAT_ATTR(event, "config:0-15");
 PMU_FORMAT_ATTR(long, "config1:0");
 
+static int sysctl_perf_user_access __read_mostly;
+
 static inline bool armv8pmu_event_is_64bit(struct perf_event *event)
 {
 	return event->attr.config1 & 0x1;
@@ -1104,6 +1106,29 @@ static int armv8pmu_probe_pmu(struct arm_pmu *cpu_pmu)
 	return probe.present ? 0 : -ENODEV;
 }
 
+int armv8pmu_proc_user_access_handler(struct ctl_table *table, int write,
+                void *buffer, size_t *lenp, loff_t *ppos)
+{
+	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
+	if (ret || !write || sysctl_perf_user_access)
+		return ret;
+
+	return 0;
+}
+
+static struct ctl_table armv8_pmu_sysctl_table[] = {
+	{
+		.procname       = "perf_user_access",
+		.data		= &sysctl_perf_user_access,
+		.maxlen		= sizeof(unsigned int),
+		.mode           = 0644,
+		.proc_handler	= armv8pmu_proc_user_access_handler,
+		.extra1		= SYSCTL_ZERO,
+		.extra2		= SYSCTL_ONE,
+	},
+	{ }
+};
+
 static int armv8_pmu_init(struct arm_pmu *cpu_pmu, char *name,
 			  int (*map_event)(struct perf_event *event),
 			  const struct attribute_group *events,
@@ -1136,6 +1161,8 @@ static int armv8_pmu_init(struct arm_pmu *cpu_pmu, char *name,
 	cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_CAPS] = caps ?
 			caps : &armv8_pmuv3_caps_attr_group;
 
+	register_sysctl("kernel", armv8_pmu_sysctl_table);
+
 	return 0;
 }
 
-- 
2.30.2

[PATCH v10 4/5] arm64: perf: Enable PMU counter userspace access for perf event

[Rob Herring]

Arm PMUs can support direct userspace access of counters which allows for
low overhead (i.e. no syscall) self-monitoring of tasks. The same feature
exists on x86 called 'rdpmc'. Unlike x86, userspace access will only be
enabled for thread bound events. This could be extended if needed, but
simplifies the implementation and reduces the chances for any
information leaks (which the x86 implementation suffers from).

PMU EL0 access will be enabled when an event with userspace access is
part of the thread's context. This includes when the event is not
scheduled on the PMU. There's some additional overhead clearing
dirty counters when access is enabled in order to prevent leaking
disabled counter data from other tasks.

Unlike x86, enabling of userspace access must be requested with a new
attr bit: config1:1. If the user requests userspace access and 64-bit
counters, then chaining will be disabled and the user will get the
maximum size counter the underlying h/w can support. The modes for
config1 are as follows:

config1 = 0 : user access disabled and always 32-bit
config1 = 1 : user access disabled and always 64-bit (using chaining if needed)
config1 = 2 : user access enabled and always 32-bit
config1 = 3 : user access enabled and counter size matches underlying counter.

Based on work by Raphael Gault <raphael.gault@arm.com>, but has been
completely re-written.

Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: linux-arm-kernel@lists.infradead.org
Cc: linux-perf-users@vger.kernel.org
Signed-off-by: Rob Herring <robh@kernel.org>

---
v10:
 - Don't control enabling user access based on mmap(). Changing the
   event_(un)mapped to run on the event's cpu doesn't work for x86.
   Triggering on mmap() doesn't limit access in any way and complicates
   the implementation.
 - Drop dirty counter tracking and just clear all unused counters.
 - Make the sysctl immediately disable access via IPI.
 - Merge armv8pmu_event_is_chained() and armv8pmu_event_can_chain()

v9:
 - Enabling/disabling of user access is now controlled in .start() and
   mmap hooks which are now called on CPUs that the event is on.
   Depends on rework of perf core and x86 RDPMC code posted here:
   https://lore.kernel.org/lkml/20210728230230.1911468-1-robh@kernel.org/

v8:
 - Rework user access tracking and enabling to be done on task
   context changes using sched_task() hook. This avoids the need for any
   IPIs, mm_switch hooks or undef instr handler.
 - Only support user access when explicitly requested on open and
   only for a thread bound events. This avoids some of the information
   leaks x86 has and simplifies the implementation.

v7:
 - Clear disabled counters when user access is enabled for a task to
   avoid leaking other tasks counter data.
 - Rework context switch handling utilizing sched_task callback
 - Add armv8pmu_event_can_chain() helper
 - Rework config1 flags handling structure
 - Use ARMV8_IDX_CYCLE_COUNTER_USER define for remapped user cycle
   counter index

v6:
 - Add new attr.config1 rdpmc bit for userspace to hint it wants
   userspace access when also requesting 64-bit counters.

v5:
 - Only set cap_user_rdpmc if event is on current cpu
 - Limit enabling/disabling access to CPUs associated with the PMU
   (supported_cpus) and with the mm_struct matching current->active_mm.

v2:
 - Move mapped/unmapped into arm64 code. Fixes arm32.
 - Rebase on cap_user_time_short changes

Changes from Raphael's v4:
  - Drop homogeneous check
  - Disable access for chained counters
  - Set pmc_width in user page
---
 arch/arm64/kernel/perf_event.c | 99 +++++++++++++++++++++++++++++++---
 1 file changed, 92 insertions(+), 7 deletions(-)

diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
index a8f8dd741aeb..91af3d1c254b 100644
--- a/arch/arm64/kernel/perf_event.c
+++ b/arch/arm64/kernel/perf_event.c
@@ -285,6 +285,7 @@ static const struct attribute_group armv8_pmuv3_events_attr_group = {
 
 PMU_FORMAT_ATTR(event, "config:0-15");
 PMU_FORMAT_ATTR(long, "config1:0");
+PMU_FORMAT_ATTR(rdpmc, "config1:1");
 
 static int sysctl_perf_user_access __read_mostly;
 
@@ -293,9 +294,15 @@ static inline bool armv8pmu_event_is_64bit(struct perf_event *event)
 	return event->attr.config1 & 0x1;
 }
 
+static inline bool armv8pmu_event_want_user_access(struct perf_event *event)
+{
+	return event->attr.config1 & 0x2;
+}
+
 static struct attribute *armv8_pmuv3_format_attrs[] = {
 	&format_attr_event.attr,
 	&format_attr_long.attr,
+	&format_attr_rdpmc.attr,
 	NULL,
 };
 
@@ -364,7 +371,7 @@ static const struct attribute_group armv8_pmuv3_caps_attr_group = {
  */
 #define	ARMV8_IDX_CYCLE_COUNTER	0
 #define	ARMV8_IDX_COUNTER0	1
-
+#define	ARMV8_IDX_CYCLE_COUNTER_USER	32
 
 /*
  * We unconditionally enable ARMv8.5-PMU long event counter support
@@ -379,15 +386,14 @@ static bool armv8pmu_has_long_event(struct arm_pmu *cpu_pmu)
 /*
  * We must chain two programmable counters for 64 bit events,
  * except when we have allocated the 64bit cycle counter (for CPU
- * cycles event). This must be called only when the event has
- * a counter allocated.
+ * cycles event) or when user space counter access is enabled.
  */
 static inline bool armv8pmu_event_is_chained(struct perf_event *event)
 {
 	int idx = event->hw.idx;
 	struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
 
-	return !WARN_ON(idx < 0) &&
+	return !(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT) &&
 	       armv8pmu_event_is_64bit(event) &&
 	       !armv8pmu_has_long_event(cpu_pmu) &&
 	       (idx != ARMV8_IDX_CYCLE_COUNTER);
@@ -720,6 +726,27 @@ static inline u32 armv8pmu_getreset_flags(void)
 	return value;
 }
 
+static void armv8pmu_disable_user_access(void)
+{
+	write_sysreg(0, pmuserenr_el0);
+}
+
+static void armv8pmu_enable_user_access(struct arm_pmu *cpu_pmu)
+{
+	int i;
+	struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
+
+	/* Clear any unused counters to avoid leaking their contents */
+	for_each_clear_bit(i, cpuc->used_mask, cpu_pmu->num_events) {
+		if (i == ARMV8_IDX_CYCLE_COUNTER)
+			write_sysreg(0, pmccntr_el0);
+		else
+			armv8pmu_write_evcntr(i, 0);
+	}
+
+	write_sysreg(ARMV8_PMU_USERENR_ER | ARMV8_PMU_USERENR_CR, pmuserenr_el0);
+}
+
 static void armv8pmu_enable_event(struct perf_event *event)
 {
 	/*
@@ -763,6 +790,14 @@ static void armv8pmu_disable_event(struct perf_event *event)
 
 static void armv8pmu_start(struct arm_pmu *cpu_pmu)
 {
+	struct perf_event_context *task_ctx =
+		this_cpu_ptr(cpu_pmu->pmu.pmu_cpu_context)->task_ctx;
+
+	if (sysctl_perf_user_access && task_ctx && task_ctx->nr_user)
+		armv8pmu_enable_user_access(cpu_pmu);
+	else
+		armv8pmu_disable_user_access();
+
 	/* Enable all counters */
 	armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMU_PMCR_E);
 }
@@ -880,13 +915,16 @@ static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc,
 	if (evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) {
 		if (!test_and_set_bit(ARMV8_IDX_CYCLE_COUNTER, cpuc->used_mask))
 			return ARMV8_IDX_CYCLE_COUNTER;
+		else if (armv8pmu_event_is_64bit(event) &&
+			   armv8pmu_event_want_user_access(event) &&
+			   !armv8pmu_has_long_event(cpu_pmu))
+				return -EAGAIN;
 	}
 
 	/*
 	 * Otherwise use events counters
 	 */
-	if (armv8pmu_event_is_64bit(event) &&
-	    !armv8pmu_has_long_event(cpu_pmu))
+	if (armv8pmu_event_is_chained(event))
 		return	armv8pmu_get_chain_idx(cpuc, cpu_pmu);
 	else
 		return armv8pmu_get_single_idx(cpuc, cpu_pmu);
@@ -902,6 +940,23 @@ static void armv8pmu_clear_event_idx(struct pmu_hw_events *cpuc,
 		clear_bit(idx - 1, cpuc->used_mask);
 }
 
+static int armv8pmu_access_event_idx(struct perf_event *event)
+{
+	if (!sysctl_perf_user_access ||
+	    !(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT))
+		return 0;
+
+	/*
+	 * We remap the cycle counter index to 32 to
+	 * match the offset applied to the rest of
+	 * the counter indices.
+	 */
+	if (event->hw.idx == ARMV8_IDX_CYCLE_COUNTER)
+		return ARMV8_IDX_CYCLE_COUNTER_USER;
+
+	return event->hw.idx;
+}
+
 /*
  * Add an event filter to a given event.
  */
@@ -995,9 +1050,23 @@ static int __armv8_pmuv3_map_event(struct perf_event *event,
 				       &armv8_pmuv3_perf_cache_map,
 				       ARMV8_PMU_EVTYPE_EVENT);
 
-	if (armv8pmu_event_is_64bit(event))
+	/*
+	 * At this point, the counter is not assigned. If a 64-bit counter is
+	 * requested, we must make sure the h/w has 64-bit counters if we set
+	 * the event size to 64-bit because chaining is not supported with
+	 * userspace access. This may still fail later on if the CPU cycle
+	 * counter is in use.
+	 */
+	if (armv8pmu_event_is_64bit(event) &&
+	    (!armv8pmu_event_want_user_access(event) ||
+	     armv8pmu_has_long_event(armpmu) || (hw_event_id == ARMV8_PMUV3_PERFCTR_CPU_CYCLES)))
 		event->hw.flags |= ARMPMU_EVT_64BIT;
 
+	/* Userspace counter access only enabled if requested and a per task event */
+	if (sysctl_perf_user_access && armv8pmu_event_want_user_access(event) &&
+	    (event->attach_state & PERF_ATTACH_TASK))
+		event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
+
 	/* Only expose micro/arch events supported by this PMU */
 	if ((hw_event_id > 0) && (hw_event_id < ARMV8_PMUV3_MAX_COMMON_EVENTS)
 	    && test_bit(hw_event_id, armpmu->pmceid_bitmap)) {
@@ -1106,6 +1175,11 @@ static int armv8pmu_probe_pmu(struct arm_pmu *cpu_pmu)
 	return probe.present ? 0 : -ENODEV;
 }
 
+static void armv8pmu_disable_user_access_ipi(void *unused)
+{
+	armv8pmu_disable_user_access();
+}
+
 int armv8pmu_proc_user_access_handler(struct ctl_table *table, int write,
                 void *buffer, size_t *lenp, loff_t *ppos)
 {
@@ -1113,6 +1187,7 @@ int armv8pmu_proc_user_access_handler(struct ctl_table *table, int write,
 	if (ret || !write || sysctl_perf_user_access)
 		return ret;
 
+	on_each_cpu(armv8pmu_disable_user_access_ipi, NULL, 1);
 	return 0;
 }
 
@@ -1152,6 +1227,8 @@ static int armv8_pmu_init(struct arm_pmu *cpu_pmu, char *name,
 	cpu_pmu->set_event_filter	= armv8pmu_set_event_filter;
 	cpu_pmu->filter_match		= armv8pmu_filter_match;
 
+	cpu_pmu->pmu.event_idx		= armv8pmu_access_event_idx;
+
 	cpu_pmu->name			= name;
 	cpu_pmu->map_event		= map_event;
 	cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_EVENTS] = events ?
@@ -1328,6 +1405,14 @@ void arch_perf_update_userpage(struct perf_event *event,
 	userpg->cap_user_time = 0;
 	userpg->cap_user_time_zero = 0;
 	userpg->cap_user_time_short = 0;
+	userpg->cap_user_rdpmc = !!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT);
+
+	if (userpg->cap_user_rdpmc) {
+		if (event->hw.flags & ARMPMU_EVT_64BIT)
+			userpg->pmc_width = 64;
+		else
+			userpg->pmc_width = 32;
+	}
 
 	do {
 		rd = sched_clock_read_begin(&seq);
-- 
2.30.2

[PATCH v10 5/5] Documentation: arm64: Document PMU counters access from userspace

[Rob Herring]

From: Raphael Gault <raphael.gault@arm.com>

Add documentation to describe the access to the pmu hardware counters from
userspace.

Signed-off-by: Raphael Gault <raphael.gault@arm.com>
Signed-off-by: Rob Herring <robh@kernel.org>
---
v10:
 - Add details on perf_user_access sysctl
v9:
 - No change
v8:
 - Reword that config1:1 must always be set to request user access
v7:
 - Merge into existing arm64 perf.rst
v6:
  - Update the chained event section with attr.config1 details
v2:
  - Update links to test examples

Changes from Raphael's v4:
  - Convert to rSt
  - Update chained event status
  - Add section for heterogeneous systems
---
 Documentation/arm64/perf.rst | 73 +++++++++++++++++++++++++++++++++++-
 1 file changed, 72 insertions(+), 1 deletion(-)

diff --git a/Documentation/arm64/perf.rst b/Documentation/arm64/perf.rst
index b567f177d385..47980c9058e3 100644
--- a/Documentation/arm64/perf.rst
+++ b/Documentation/arm64/perf.rst
@@ -2,7 +2,10 @@
 
 .. _perf_index:
 
-=====================
+====
+Perf
+====
+
 Perf Event Attributes
 =====================
 
@@ -88,3 +91,71 @@ exclude_host. However when using !exclude_hv there is a small blackout
 window at the guest entry/exit where host events are not captured.
 
 On VHE systems there are no blackout windows.
+
+Perf Userspace PMU Hardware Counter Access
+==========================================
+
+Overview
+--------
+The perf userspace tool relies on the PMU to monitor events. It offers an
+abstraction layer over the hardware counters since the underlying
+implementation is cpu-dependent.
+Arm64 allows userspace tools to have access to the registers storing the
+hardware counters' values directly.
+
+This targets specifically self-monitoring tasks in order to reduce the overhead
+by directly accessing the registers without having to go through the kernel.
+
+How-to
+------
+The focus is set on the armv8 PMUv3 which makes sure that the access to the pmu
+registers is enabled and that the userspace has access to the relevant
+information in order to use them.
+
+In order to have access to the hardware counters, the global sysctl
+kernel/perf_user_access must first be enabled:
+
+.. code-block:: sh
+
+  echo 1 > /proc/sys/kernel/perf_user_access
+
+It is necessary to open the event using the perf tool interface with config1:1
+attr bit set: the sys_perf_event_open syscall returns a fd which can
+subsequently be used with the mmap syscall in order to retrieve a page of memory
+containing information about the event. The PMU driver uses this page to expose
+to the user the hardware counter's index and other necessary data. Using this
+index enables the user to access the PMU registers using the `mrs` instruction.
+
+The userspace access is supported in libperf using the perf_evsel__mmap()
+and perf_evsel__read() functions. See `tools/lib/perf/tests/test-evsel.c`_ for
+an example.
+
+About heterogeneous systems
+---------------------------
+On heterogeneous systems such as big.LITTLE, userspace PMU counter access can
+only be enabled when the tasks are pinned to a homogeneous subset of cores and
+the corresponding PMU instance is opened by specifying the 'type' attribute.
+The use of generic event types is not supported in this case.
+
+Have a look at `tools/perf/arch/arm64/tests/user-events.c`_ for an example. It
+can be run using the perf tool to check that the access to the registers works
+correctly from userspace:
+
+.. code-block:: sh
+
+  perf test -v user
+
+About chained events and 64-bit counters
+----------------------------------------
+Chained events are not supported in conjunction with userspace counter
+access. If a 64-bit counter is requested (attr.config1:0) with userspace
+access (attr.config1:1 set), then counter chaining will be disabled. The
+'pmc_width' in the user page will indicate the actual width of the
+counter which could be only 32-bits depending on the event and PMU
+features.
+
+.. Links
+.. _tools/perf/arch/arm64/tests/user-events.c:
+   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/arch/arm64/tests/user-events.c
+.. _tools/lib/perf/tests/test-evsel.c:
+   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/perf/tests/test-evsel.c
-- 
2.30.2

Re: [PATCH v10 0/5] arm64 userspace counter support

[Rob Herring]

On Tue, Sep 14, 2021 at 3:48 PM Rob Herring <robh@kernel.org> wrote:
>
> Another version of arm64 userspace counter access support.

Any chance someone can review this... Ideally before the end of the
cycle this time so there's time for me to respin if needed.


> The arm64 support departs from the x86 implementation by requiring the user
> to explicitly request user access (via attr.config1) and only enables access
> for task bound events. Since usage is explicitly requested, access is
> enabled at perf_event_open() rather than on mmap() as that greatly
> simplifies the implementation. Rather than trying to lock down the access
> as the x86 implementation has been doing, we can start with only a limited
> use case enabled and later expand it if needed.
>
> I've run this version thru Vince's perf tests[12] with arm64 support added.
> I wish I'd found these tests sooner...
>
> This originally resurrected Raphael's series[1] to enable userspace counter
> access on arm64. My previous versions are here[2][3][4][5][6][7][8][9][10].
> A git branch is here[11].
>
> Changes in v10:
>  - Drop changing event_mapped/event_unmapped ops to run on the event's
>    current CPU. This won't work for x86 where any thread sharing an mm
>    context will have rdpmc enabled. Instead, simply track user access
>    events when added to a perf context and use that count.
>  - Documentation for the sysctl disable
>
> Changes in v9:
>  - Reworked x86 and perf core to handle user access tracking and call
>    .event_mapped() and .event_unmapped() on the CPU with the event like
>    other changes to events.
>  - Use sysctl instead of sysfs to disable user access.
>
> Changes in v8:
>  - Restrict user access to thread bound events which simplifies the
>    implementation. A couple of perf core changes (patches 1 and 2) are
>    needed to do this.
>  - Always require the user to request userspace access.
>
> Changes in v7:
>  - Handling of dirty counter leakage and reworking of context switch and
>    user access enabling. The .sched_task hook and undef instruction handler
>    are now utilized. (Patch 3)
>  - Add a userspace disable switch like x86. (Patch 5)
>
> Changes in v6:
>  - Reworking of the handling of 64-bit counters and user access. There's
>    a new config1 flag to request user access. This takes priority over
>    the 64-bit flag and the user will get the maximum size the h/w
>    supports without chaining.
>  - The libperf evsel mmap struct is stored in its own xyarray
>  - New tests for user 64-bit and 32-bit counters
>  - Rebase to v5.12-rc2
>
> Changes in v5:
>  - Limit enabling/disabling access to CPUs associated with the PMU
>    (supported_cpus) and with the mm_struct matching current->active_mm.
>    The x86 method of using mm_cpumask doesn't work for arm64 as it is not
>    updated.
>  - Only set cap_user_rdpmc if event is on current cpu. See patch 2.
>  - Create an mmap for every event in an evsel. This results in some changes
>    to the libperf mmap API from the last version.
>  - Rebase to v5.11-rc2
>
> Changes in v4:
>  - Dropped 'arm64: pmu: Add hook to handle pmu-related undefined instructions'.
>    The onus is on userspace to pin itself to a homogeneous subset of CPUs
>    and avoid any aborts on heterogeneous systems, so the hook is not needed.
>  - Make perf_evsel__mmap() take pages rather than bytes for size
>  - Fix building arm64 heterogeneous test.
>
> Changes in v3:
>  - Dropped removing x86 rdpmc test until libperf tests can run via 'perf test'
>  - Added verbose prints for tests
>  - Split adding perf_evsel__mmap() to separate patch
>
> The following changes to the arm64 support have been made compared to
> Raphael's last version:
>
> The major change is support for heterogeneous systems with some
> restrictions. Specifically, userspace must pin itself to like CPUs, open
> a specific PMU by type, and use h/w specific events. The tests have been
> reworked to demonstrate this.
>
> Chained events are not supported. The problem with supporting chained
> events was there's no way to distinguish between a chained event and a
> native 64-bit counter. We could add some flag, but do self monitoring
> processes really need that? Native 64-bit counters are supported if the
> PMU h/w has support. As there's already an explicit ABI to request 64-bit
> counters, userspace can request 64-bit counters and if user
> access is not enabled, then it must retry with 32-bit counters.
>
> Prior versions broke the build on arm32 (surprisingly never caught by
> 0-day). As a result, event_mapped and event_unmapped implementations have
> been moved into the arm64 code.
>
> There was a bug in that pmc_width was not set in the user page. The tests
> now check for this.
>
> The documentation has been converted to rST. I've added sections on
> chained events and heterogeneous.
>
> Rob
>
>
> [1] https://lore.kernel.org/r/20190822144220.27860-1-raphael.gault@arm.com/
> [2] https://lore.kernel.org/r/20200707205333.624938-1-robh@kernel.org/
> [3] https://lore.kernel.org/r/20200828205614.3391252-1-robh@kernel.org/
> [4] https://lore.kernel.org/r/20200911215118.2887710-1-robh@kernel.org/
> [5] https://lore.kernel.org/r/20201001140116.651970-1-robh@kernel.org/
> [6] https://lore.kernel.org/r/20210114020605.3943992-1-robh@kernel.org/
> [7] https://lore.kernel.org/r/20210311000837.3630499-1-robh@kernel.org/
> [8] https://lore.kernel.org/r/20210420031511.2348977-1-robh@kernel.org/
> [9] https://lore.kernel.org/r/20210517195405.3079458-1-robh@kernel.org/
> [10] https://lore.kernel.org/all/20210806225123.1958497-1-robh@kernel.org/
> [11] git://git.kernel.org/pub/scm/linux/kernel/git/robh/linux.git arm64-user-perf-event-v10
> [12] https://github.com/deater/perf_event_tests
>
> Raphael Gault (1):
>   Documentation: arm64: Document PMU counters access from userspace
>
> Rob Herring (4):
>   x86: perf: Move RDPMC event flag to a common definition
>   perf: Add a counter for number of user access events in context
>   arm64: perf: Add userspace counter access disable switch
>   arm64: perf: Enable PMU counter userspace access for perf event
>
>  Documentation/admin-guide/sysctl/kernel.rst |  11 ++
>  Documentation/arm64/perf.rst                |  73 +++++++++++-
>  arch/arm64/kernel/perf_event.c              | 126 ++++++++++++++++++--
>  arch/x86/events/core.c                      |  10 +-
>  arch/x86/events/perf_event.h                |   2 +-
>  include/linux/perf_event.h                  |   3 +
>  kernel/events/core.c                        |   4 +
>  7 files changed, 215 insertions(+), 14 deletions(-)
>
> --
> 2.30.2

Re: [PATCH v10 1/5] x86: perf: Move RDPMC event flag to a common definition

[Mark Rutland]

Hi Rob,

On Tue, Sep 14, 2021 at 03:47:56PM -0500, Rob Herring wrote:
> In preparation to enable user counter access on arm64 and to move some
> of the user access handling to perf core, create a common event flag for
> user counter access and convert x86 to use it.
> 
> Since the architecture specific flags start at the LSB, starting at the
> MSB for common flags.

Minor comments below (definition rename, and a comment block), but with
those:

Reviewed-by: Mark Rutland <mark.rutland@arm.com>

Peter, are you happy with this from the x86 side?

> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
> Cc: Mark Rutland <mark.rutland@arm.com>
> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
> Cc: Jiri Olsa <jolsa@redhat.com>
> Cc: Namhyung Kim <namhyung@kernel.org>
> Cc: Kan Liang <kan.liang@linux.intel.com>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: Borislav Petkov <bp@alien8.de>
> Cc: x86@kernel.org
> Cc: "H. Peter Anvin" <hpa@zytor.com>
> Cc: linux-perf-users@vger.kernel.org
> Signed-off-by: Rob Herring <robh@kernel.org>
> ---
>  arch/x86/events/core.c       | 10 +++++-----
>  arch/x86/events/perf_event.h |  2 +-
>  include/linux/perf_event.h   |  2 ++
>  3 files changed, 8 insertions(+), 6 deletions(-)
> 
> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> index 2a57dbed4894..2bd50fc061e1 100644
> --- a/arch/x86/events/core.c
> +++ b/arch/x86/events/core.c
> @@ -2469,7 +2469,7 @@ static int x86_pmu_event_init(struct perf_event *event)
>  
>  	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
>  	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
> -		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
> +		event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
>  
>  	return err;
>  }
> @@ -2503,7 +2503,7 @@ void perf_clear_dirty_counters(void)
>  
>  static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
>  {
> -	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
> +	if (!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT))
>  		return;
>  
>  	/*
> @@ -2524,7 +2524,7 @@ static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
>  
>  static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
>  {
> -	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
> +	if (!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT))
>  		return;
>  
>  	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
> @@ -2535,7 +2535,7 @@ static int x86_pmu_event_idx(struct perf_event *event)
>  {
>  	struct hw_perf_event *hwc = &event->hw;
>  
> -	if (!(hwc->flags & PERF_X86_EVENT_RDPMC_ALLOWED))
> +	if (!(hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
>  		return 0;
>  
>  	if (is_metric_idx(hwc->idx))
> @@ -2718,7 +2718,7 @@ void arch_perf_update_userpage(struct perf_event *event,
>  	userpg->cap_user_time = 0;
>  	userpg->cap_user_time_zero = 0;
>  	userpg->cap_user_rdpmc =
> -		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
> +		!!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT);
>  	userpg->pmc_width = x86_pmu.cntval_bits;
>  
>  	if (!using_native_sched_clock() || !sched_clock_stable())
> diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
> index e3ac05c97b5e..49f68b15745f 100644
> --- a/arch/x86/events/perf_event.h
> +++ b/arch/x86/events/perf_event.h
> @@ -73,7 +73,7 @@ static inline bool constraint_match(struct event_constraint *c, u64 ecode)
>  #define PERF_X86_EVENT_PEBS_NA_HSW	0x0010 /* haswell style datala, unknown */
>  #define PERF_X86_EVENT_EXCL		0x0020 /* HT exclusivity on counter */
>  #define PERF_X86_EVENT_DYNAMIC		0x0040 /* dynamic alloc'd constraint */
> -#define PERF_X86_EVENT_RDPMC_ALLOWED	0x0080 /* grant rdpmc permission */
> +
>  #define PERF_X86_EVENT_EXCL_ACCT	0x0100 /* accounted EXCL event */
>  #define PERF_X86_EVENT_AUTO_RELOAD	0x0200 /* use PEBS auto-reload */
>  #define PERF_X86_EVENT_LARGE_PEBS	0x0400 /* use large PEBS */
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index fe156a8170aa..12debf008d39 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -142,6 +142,8 @@ struct hw_perf_event {
>  			int		event_base_rdpmc;
>  			int		idx;
>  			int		last_cpu;
> +
> +#define PERF_EVENT_FLAG_USER_READ_CNT	0x80000000

I realise this matches the style of PERF_HES_* and PERF_EF_*. but could
we please arrange this like the PERF_PMU_CAP_* definitions, and move
this immediately before the struct hw_perf_event defintion, with a
comment block, e.g.

/*
 * hw_perf_event::flag values
 *
 * PERF_EVENT_FLAG_ARCH bits are reserved for architecture-specific
 * usage.
 */
#define PERF_EVENT_FLAG_ARCH			0x0000ffff
#define PERF_EVENT_FLAG_USER_READ_CNT		0x80000000

Thanks,
Mark.

Re: [PATCH v10 2/5] perf: Add a counter for number of user access events in context

[Mark Rutland]

On Tue, Sep 14, 2021 at 03:47:57PM -0500, Rob Herring wrote:
> For controlling user space counter access, we need to know if any event
> in a context (currently scheduled or not) is using user space counters.
> Walking the context's list of events would be slow, so add a counter
> to track this.
> 
> Signed-off-by: Rob Herring <robh@kernel.org>

Reviewed-by: Mark Rutland <mark.rutland@arm.com>

Mark.

> ---
> v10:
>  - Re-added.
>  - Maintain the count in the perf core
> v9:
>  - Dropped
> v8:
>  - new patch
> ---
>  include/linux/perf_event.h | 1 +
>  kernel/events/core.c       | 4 ++++
>  2 files changed, 5 insertions(+)
> 
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index 12debf008d39..4f82a4d47139 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -821,6 +821,7 @@ struct perf_event_context {
>  
>  	int				nr_events;
>  	int				nr_active;
> +	int				nr_user;
>  	int				is_active;
>  	int				nr_stat;
>  	int				nr_freq;
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 744e8726c5b2..01290d150da3 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -1808,6 +1808,8 @@ list_add_event(struct perf_event *event, struct perf_event_context *ctx)
>  
>  	list_add_rcu(&event->event_entry, &ctx->event_list);
>  	ctx->nr_events++;
> +	if (event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT)
> +		ctx->nr_user++;
>  	if (event->attr.inherit_stat)
>  		ctx->nr_stat++;
>  
> @@ -1999,6 +2001,8 @@ list_del_event(struct perf_event *event, struct perf_event_context *ctx)
>  	event->attach_state &= ~PERF_ATTACH_CONTEXT;
>  
>  	ctx->nr_events--;
> +	if (event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT)
> +		ctx->nr_user--;
>  	if (event->attr.inherit_stat)
>  		ctx->nr_stat--;
>  
> -- 
> 2.30.2
> 

Re: [PATCH v10 3/5] arm64: perf: Add userspace counter access disable switch

[Mark Rutland]

On Tue, Sep 14, 2021 at 03:47:58PM -0500, Rob Herring wrote:
> Like x86, some users may want to disable userspace PMU counter
> altogether. Add a sysctl 'perf_user_access' file to control userspace
> counter access. The default is '0' which is disabled. Writing '1'
> enables access.
> 
> Note that x86 also supports writing '2' to globally enable user access.

For clarity it might be worth mentioning that on x86 this is controlled
by the PMU's `rdpmc` sysfs attribute, i.e.

  Note that x86 supports globally enabling user access by writing '2' to
  /sys/bus/event_source/devices/cpu/rdpmc

> As there's not existing userspace support to worry about, this shouldn't
> be necessary for Arm. It could be added later if the need arises.
> 
> Cc: Mark Rutland <mark.rutland@arm.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
> Cc: Jiri Olsa <jolsa@redhat.com>
> Cc: Namhyung Kim <namhyung@kernel.org>
> Cc: Catalin Marinas <catalin.marinas@arm.com>
> Cc: linux-perf-users@vger.kernel.org
> Acked-by: Will Deacon <will@kernel.org>
> Signed-off-by: Rob Herring <robh@kernel.org>
> ---
> v10:
>  - Add documentation
>  - Use a custom handler (needed on the next patch)
> v9:
>  - Use sysctl instead of sysfs attr
>  - Default to disabled
> v8:
>  - New patch
> ---
>  Documentation/admin-guide/sysctl/kernel.rst | 11 +++++++++
>  arch/arm64/kernel/perf_event.c              | 27 +++++++++++++++++++++
>  2 files changed, 38 insertions(+)
> 
> diff --git a/Documentation/admin-guide/sysctl/kernel.rst b/Documentation/admin-guide/sysctl/kernel.rst
> index 426162009ce9..346a0dba5703 100644
> --- a/Documentation/admin-guide/sysctl/kernel.rst
> +++ b/Documentation/admin-guide/sysctl/kernel.rst
> @@ -905,6 +905,17 @@ enabled, otherwise writing to this file will return ``-EBUSY``.
>  The default value is 8.
>  
>  
> +perf_user_access (arm64 only)
> +=================================
> +
> +Controls user space access for reading perf event counters. When set to 1,
> +user space can read performance monitor counter registers directly.
> +
> +The default value is 0 (access disabled).
> +
> +See Documentation/arm64/perf.rst for more information.

Looking at the existing perf sysctls:

# ls /proc/sys/kernel/perf*
/proc/sys/kernel/perf_cpu_time_max_percent
/proc/sys/kernel/perf_event_max_contexts_per_stack
/proc/sys/kernel/perf_event_max_sample_rate
/proc/sys/kernel/perf_event_max_stack
/proc/sys/kernel/perf_event_mlock_kb
/proc/sys/kernel/perf_event_paranoid

I see that other than `perf_cpu_time_max_percent`, we've used
`perf_event_` as the prefix, and I suspect we should do the same here,
but I guess it may not matter either way.

> +
> +
>  pid_max
>  =======
>  
> diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
> index b4044469527e..a8f8dd741aeb 100644
> --- a/arch/arm64/kernel/perf_event.c
> +++ b/arch/arm64/kernel/perf_event.c
> @@ -286,6 +286,8 @@ static const struct attribute_group armv8_pmuv3_events_attr_group = {
>  PMU_FORMAT_ATTR(event, "config:0-15");
>  PMU_FORMAT_ATTR(long, "config1:0");
>  
> +static int sysctl_perf_user_access __read_mostly;
> +
>  static inline bool armv8pmu_event_is_64bit(struct perf_event *event)
>  {
>  	return event->attr.config1 & 0x1;
> @@ -1104,6 +1106,29 @@ static int armv8pmu_probe_pmu(struct arm_pmu *cpu_pmu)
>  	return probe.present ? 0 : -ENODEV;
>  }
>  
> +int armv8pmu_proc_user_access_handler(struct ctl_table *table, int write,
> +                void *buffer, size_t *lenp, loff_t *ppos)
> +{
> +	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
> +	if (ret || !write || sysctl_perf_user_access)
> +		return ret;
> +
> +	return 0;
> +}

Maybe this is needed in the next patch, but the if statement is entirely
redundant on this patch and looks really odd.

Can we please either:

1) Use proc_dointvec_minmax() directly in this patch (which is what Will
Acked in v9) and add the wrapper in the next patch when we need it.

2) make this:

| int armv8pmu_proc_user_access_handler(struct ctl_table *table, int write,
|                 void *buffer, size_t *lenp, loff_t *ppos)
| {
| 	return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
| }

... and flesh it out in the next patch.

With either of those two options (and regardless of whether the
attribute is renamed):

Reviewed-by: Mark Rutland <mark.rutland@arm.com>

Thanks,
Mark.

> +
> +	return 0;
> +}
> +
> +static struct ctl_table armv8_pmu_sysctl_table[] = {
> +	{
> +		.procname       = "perf_user_access",
> +		.data		= &sysctl_perf_user_access,
> +		.maxlen		= sizeof(unsigned int),
> +		.mode           = 0644,
> +		.proc_handler	= armv8pmu_proc_user_access_handler,
> +		.extra1		= SYSCTL_ZERO,
> +		.extra2		= SYSCTL_ONE,
> +	},
> +	{ }
> +};
> +
>  static int armv8_pmu_init(struct arm_pmu *cpu_pmu, char *name,
>  			  int (*map_event)(struct perf_event *event),
>  			  const struct attribute_group *events,
> @@ -1136,6 +1161,8 @@ static int armv8_pmu_init(struct arm_pmu *cpu_pmu, char *name,
>  	cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_CAPS] = caps ?
>  			caps : &armv8_pmuv3_caps_attr_group;
>  
> +	register_sysctl("kernel", armv8_pmu_sysctl_table);
> +
>  	return 0;
>  }
>  
> -- 
> 2.30.2
> 

Re: [PATCH v10 0/5] arm64 userspace counter support

[Mark Rutland]

Hi Rob,

On Mon, Sep 27, 2021 at 03:46:15PM -0500, Rob Herring wrote:
> On Tue, Sep 14, 2021 at 3:48 PM Rob Herring <robh@kernel.org> wrote:
> >
> > Another version of arm64 userspace counter access support.
> 
> Any chance someone can review this... Ideally before the end of the
> cycle this time so there's time for me to respin if needed.

Sorry for the long delay on this.

I've given some comments on patches 1-3 (which are trivial, and I think
we can fix up when applying if necessary), and I've started
reviewing/testing patch 4; I'll try to follow up with more comments
tomorrow.

Thanks,
Mark.

Re: [PATCH v10 4/5] arm64: perf: Enable PMU counter userspace access for perf event

[Mark Rutland]

Hi Rob,

This looks pretty good!

I have one largish query below, and otherwise only trivialities that I'm
happy to fix up.

On Tue, Sep 14, 2021 at 03:47:59PM -0500, Rob Herring wrote:
> Arm PMUs can support direct userspace access of counters which allows for
> low overhead (i.e. no syscall) self-monitoring of tasks. The same feature
> exists on x86 called 'rdpmc'. Unlike x86, userspace access will only be
> enabled for thread bound events. This could be extended if needed, but
> simplifies the implementation and reduces the chances for any
> information leaks (which the x86 implementation suffers from).
> 
> PMU EL0 access will be enabled when an event with userspace access is
> part of the thread's context. This includes when the event is not
> scheduled on the PMU. There's some additional overhead clearing
> dirty counters when access is enabled in order to prevent leaking
> disabled counter data from other tasks.
> 
> Unlike x86, enabling of userspace access must be requested with a new
> attr bit: config1:1. If the user requests userspace access and 64-bit
> counters, then chaining will be disabled and the user will get the
> maximum size counter the underlying h/w can support. The modes for
> config1 are as follows:
> 
> config1 = 0 : user access disabled and always 32-bit
> config1 = 1 : user access disabled and always 64-bit (using chaining if needed)
> config1 = 2 : user access enabled and always 32-bit
> config1 = 3 : user access enabled and counter size matches underlying counter.

We probably need to note somewhere (i.e. in the next patch) that we mean
*logically* 32-bit, and this could be a biased 64-bit counter, so
userspace needs to treat the upper 32-bits of counters as UNKNOWN.

For the `config1 = 3` case (potentially) overriding the usual long
semantic, I'm struggling to understand why we need that rather than
forcing the use of a 64-bit counter, because in that case:

* For a CPU_CYCLES event:
  __armv8_pmuv3_map_event() will always pick 64-bits
  get_event_idx() may fail to allocate a 64-bit counter.

* For other events:
  __armv8_pmuv3_map_event() will pick 32/64 based on long counter
  support
  get_event_idx() will only fail if there are no counters free.

Whereas if __armv8_pmuv3_map_event() returned an error for the latter
when long counter support is not implemented, we'd have consistent
`long` semantics, and the CPU_CYCLES behaviour would be identical.

What's the rationale for `3` leaving the choice to the kernel?

If the problem is discoverability, I'd be happy to add something to
sysfs to describe whether the PMU has long event support.

> Based on work by Raphael Gault <raphael.gault@arm.com>, but has been
> completely re-written.
> 
> Cc: Will Deacon <will@kernel.org>
> Cc: Mark Rutland <mark.rutland@arm.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
> Cc: Jiri Olsa <jolsa@redhat.com>
> Cc: Namhyung Kim <namhyung@kernel.org>
> Cc: Catalin Marinas <catalin.marinas@arm.com>
> Cc: linux-arm-kernel@lists.infradead.org
> Cc: linux-perf-users@vger.kernel.org
> Signed-off-by: Rob Herring <robh@kernel.org>
> 
> ---
> v10:
>  - Don't control enabling user access based on mmap(). Changing the
>    event_(un)mapped to run on the event's cpu doesn't work for x86.
>    Triggering on mmap() doesn't limit access in any way and complicates
>    the implementation.
>  - Drop dirty counter tracking and just clear all unused counters.
>  - Make the sysctl immediately disable access via IPI.
>  - Merge armv8pmu_event_is_chained() and armv8pmu_event_can_chain()
> 
> v9:
>  - Enabling/disabling of user access is now controlled in .start() and
>    mmap hooks which are now called on CPUs that the event is on.
>    Depends on rework of perf core and x86 RDPMC code posted here:
>    https://lore.kernel.org/lkml/20210728230230.1911468-1-robh@kernel.org/
> 
> v8:
>  - Rework user access tracking and enabling to be done on task
>    context changes using sched_task() hook. This avoids the need for any
>    IPIs, mm_switch hooks or undef instr handler.
>  - Only support user access when explicitly requested on open and
>    only for a thread bound events. This avoids some of the information
>    leaks x86 has and simplifies the implementation.
> 
> v7:
>  - Clear disabled counters when user access is enabled for a task to
>    avoid leaking other tasks counter data.
>  - Rework context switch handling utilizing sched_task callback
>  - Add armv8pmu_event_can_chain() helper
>  - Rework config1 flags handling structure
>  - Use ARMV8_IDX_CYCLE_COUNTER_USER define for remapped user cycle
>    counter index
> 
> v6:
>  - Add new attr.config1 rdpmc bit for userspace to hint it wants
>    userspace access when also requesting 64-bit counters.
> 
> v5:
>  - Only set cap_user_rdpmc if event is on current cpu
>  - Limit enabling/disabling access to CPUs associated with the PMU
>    (supported_cpus) and with the mm_struct matching current->active_mm.
> 
> v2:
>  - Move mapped/unmapped into arm64 code. Fixes arm32.
>  - Rebase on cap_user_time_short changes
> 
> Changes from Raphael's v4:
>   - Drop homogeneous check
>   - Disable access for chained counters
>   - Set pmc_width in user page
> ---
>  arch/arm64/kernel/perf_event.c | 99 +++++++++++++++++++++++++++++++---
>  1 file changed, 92 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
> index a8f8dd741aeb..91af3d1c254b 100644
> --- a/arch/arm64/kernel/perf_event.c
> +++ b/arch/arm64/kernel/perf_event.c
> @@ -285,6 +285,7 @@ static const struct attribute_group armv8_pmuv3_events_attr_group = {
>  
>  PMU_FORMAT_ATTR(event, "config:0-15");
>  PMU_FORMAT_ATTR(long, "config1:0");
> +PMU_FORMAT_ATTR(rdpmc, "config1:1");
>  
>  static int sysctl_perf_user_access __read_mostly;
>  
> @@ -293,9 +294,15 @@ static inline bool armv8pmu_event_is_64bit(struct perf_event *event)
>  	return event->attr.config1 & 0x1;
>  }
>  
> +static inline bool armv8pmu_event_want_user_access(struct perf_event *event)
> +{
> +	return event->attr.config1 & 0x2;
> +}
> +
>  static struct attribute *armv8_pmuv3_format_attrs[] = {
>  	&format_attr_event.attr,
>  	&format_attr_long.attr,
> +	&format_attr_rdpmc.attr,
>  	NULL,
>  };
>  
> @@ -364,7 +371,7 @@ static const struct attribute_group armv8_pmuv3_caps_attr_group = {
>   */
>  #define	ARMV8_IDX_CYCLE_COUNTER	0
>  #define	ARMV8_IDX_COUNTER0	1
> -
> +#define	ARMV8_IDX_CYCLE_COUNTER_USER	32
>  
>  /*
>   * We unconditionally enable ARMv8.5-PMU long event counter support
> @@ -379,15 +386,14 @@ static bool armv8pmu_has_long_event(struct arm_pmu *cpu_pmu)
>  /*
>   * We must chain two programmable counters for 64 bit events,
>   * except when we have allocated the 64bit cycle counter (for CPU
> - * cycles event). This must be called only when the event has
> - * a counter allocated.
> + * cycles event) or when user space counter access is enabled.
>   */
>  static inline bool armv8pmu_event_is_chained(struct perf_event *event)
>  {
>  	int idx = event->hw.idx;
>  	struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
>  
> -	return !WARN_ON(idx < 0) &&
> +	return !(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT) &&
>  	       armv8pmu_event_is_64bit(event) &&
>  	       !armv8pmu_has_long_event(cpu_pmu) &&
>  	       (idx != ARMV8_IDX_CYCLE_COUNTER);
> @@ -720,6 +726,27 @@ static inline u32 armv8pmu_getreset_flags(void)
>  	return value;

Above this, could we please add:

| static inline bool armv8pmu_event_has_user_read(struct perf_event *event)
| {
| 	return event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT;
| }

... and use that where we look at PERF_EVENT_FLAG_USER_READ_CNT?

>  
> +static void armv8pmu_disable_user_access(void)
> +{
> +	write_sysreg(0, pmuserenr_el0);
> +}
> +
> +static void armv8pmu_enable_user_access(struct arm_pmu *cpu_pmu)
> +{
> +	int i;
> +	struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
> +
> +	/* Clear any unused counters to avoid leaking their contents */
> +	for_each_clear_bit(i, cpuc->used_mask, cpu_pmu->num_events) {
> +		if (i == ARMV8_IDX_CYCLE_COUNTER)
> +			write_sysreg(0, pmccntr_el0);
> +		else
> +			armv8pmu_write_evcntr(i, 0);
> +	}
> +
> +	write_sysreg(ARMV8_PMU_USERENR_ER | ARMV8_PMU_USERENR_CR, pmuserenr_el0);
> +}
> +
>  static void armv8pmu_enable_event(struct perf_event *event)
>  {
>  	/*
> @@ -763,6 +790,14 @@ static void armv8pmu_disable_event(struct perf_event *event)
>  
>  static void armv8pmu_start(struct arm_pmu *cpu_pmu)
>  {
> +	struct perf_event_context *task_ctx =
> +		this_cpu_ptr(cpu_pmu->pmu.pmu_cpu_context)->task_ctx;
> +
> +	if (sysctl_perf_user_access && task_ctx && task_ctx->nr_user)
> +		armv8pmu_enable_user_access(cpu_pmu);
> +	else
> +		armv8pmu_disable_user_access();
> +
>  	/* Enable all counters */
>  	armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMU_PMCR_E);
>  }
> @@ -880,13 +915,16 @@ static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc,
>  	if (evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) {
>  		if (!test_and_set_bit(ARMV8_IDX_CYCLE_COUNTER, cpuc->used_mask))
>  			return ARMV8_IDX_CYCLE_COUNTER;
> +		else if (armv8pmu_event_is_64bit(event) &&
> +			   armv8pmu_event_want_user_access(event) &&
> +			   !armv8pmu_has_long_event(cpu_pmu))
> +				return -EAGAIN;
>  	}
>  
>  	/*
>  	 * Otherwise use events counters
>  	 */
> -	if (armv8pmu_event_is_64bit(event) &&
> -	    !armv8pmu_has_long_event(cpu_pmu))
> +	if (armv8pmu_event_is_chained(event))
>  		return	armv8pmu_get_chain_idx(cpuc, cpu_pmu);
>  	else
>  		return armv8pmu_get_single_idx(cpuc, cpu_pmu);
> @@ -902,6 +940,23 @@ static void armv8pmu_clear_event_idx(struct pmu_hw_events *cpuc,
>  		clear_bit(idx - 1, cpuc->used_mask);
>  }
>  
> +static int armv8pmu_access_event_idx(struct perf_event *event)

Can we please s/access/user/ here?

> +{
> +	if (!sysctl_perf_user_access ||
> +	    !(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT))
> +		return 0;
> +
> +	/*
> +	 * We remap the cycle counter index to 32 to
> +	 * match the offset applied to the rest of
> +	 * the counter indices.
> +	 */
> +	if (event->hw.idx == ARMV8_IDX_CYCLE_COUNTER)
> +		return ARMV8_IDX_CYCLE_COUNTER_USER;
> +
> +	return event->hw.idx;
> +}
> +
>  /*
>   * Add an event filter to a given event.
>   */
> @@ -995,9 +1050,23 @@ static int __armv8_pmuv3_map_event(struct perf_event *event,
>  				       &armv8_pmuv3_perf_cache_map,
>  				       ARMV8_PMU_EVTYPE_EVENT);
>  
> -	if (armv8pmu_event_is_64bit(event))
> +	/*
> +	 * At this point, the counter is not assigned. If a 64-bit counter is
> +	 * requested, we must make sure the h/w has 64-bit counters if we set
> +	 * the event size to 64-bit because chaining is not supported with
> +	 * userspace access. This may still fail later on if the CPU cycle
> +	 * counter is in use.
> +	 */
> +	if (armv8pmu_event_is_64bit(event) &&
> +	    (!armv8pmu_event_want_user_access(event) ||
> +	     armv8pmu_has_long_event(armpmu) || (hw_event_id == ARMV8_PMUV3_PERFCTR_CPU_CYCLES)))
>  		event->hw.flags |= ARMPMU_EVT_64BIT;

If we can follow my suggestion in reply to the cover text, we can make
this:

	if (armv8pmu_event_is_64bit(event))
		event->hw.flags |= ARMPMU_EVT_64BIT;

	/*
	 * User events must be allocated into a single counter, and so
	 * must not be chained.
	 *
	 * Most 64-bit events require long counter support, but 64-bit
	 * CPU_CYCLES events can be placed into the dedicated cycle
	 * counter when this is free.
	 *
	if (armv8pmu_event_want_user_access()) {
		if (armv8pmu_event_is_64bit(event) &&
		    (hw_event_id != ARMV8_PMUV3_PERFCTR_CPU_CYCLES) &&
		    !armv8pmu_has_long_event(armpmu))
			return -EINVAL;
	}
	
> +	/* Userspace counter access only enabled if requested and a per task event */
> +	if (sysctl_perf_user_access && armv8pmu_event_want_user_access(event) &&
> +	    (event->attach_state & PERF_ATTACH_TASK))
> +		event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;

Can we please explicitly reject !PERF_ATTACH_TASK case?

If the user requested something we don't intend to support, I'd rather
return -EINVAL here, rather than continue on.

Thanks,
Mark.

> +
>  	/* Only expose micro/arch events supported by this PMU */
>  	if ((hw_event_id > 0) && (hw_event_id < ARMV8_PMUV3_MAX_COMMON_EVENTS)
>  	    && test_bit(hw_event_id, armpmu->pmceid_bitmap)) {
> @@ -1106,6 +1175,11 @@ static int armv8pmu_probe_pmu(struct arm_pmu *cpu_pmu)
>  	return probe.present ? 0 : -ENODEV;
>  }
>  
> +static void armv8pmu_disable_user_access_ipi(void *unused)
> +{
> +	armv8pmu_disable_user_access();
> +}
> +
>  int armv8pmu_proc_user_access_handler(struct ctl_table *table, int write,
>                  void *buffer, size_t *lenp, loff_t *ppos)
>  {
> @@ -1113,6 +1187,7 @@ int armv8pmu_proc_user_access_handler(struct ctl_table *table, int write,
>  	if (ret || !write || sysctl_perf_user_access)
>  		return ret;
>  
> +	on_each_cpu(armv8pmu_disable_user_access_ipi, NULL, 1);
>  	return 0;
>  }
>  
> @@ -1152,6 +1227,8 @@ static int armv8_pmu_init(struct arm_pmu *cpu_pmu, char *name,
>  	cpu_pmu->set_event_filter	= armv8pmu_set_event_filter;
>  	cpu_pmu->filter_match		= armv8pmu_filter_match;
>  
> +	cpu_pmu->pmu.event_idx		= armv8pmu_access_event_idx;
> +
>  	cpu_pmu->name			= name;
>  	cpu_pmu->map_event		= map_event;
>  	cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_EVENTS] = events ?
> @@ -1328,6 +1405,14 @@ void arch_perf_update_userpage(struct perf_event *event,
>  	userpg->cap_user_time = 0;
>  	userpg->cap_user_time_zero = 0;
>  	userpg->cap_user_time_short = 0;
> +	userpg->cap_user_rdpmc = !!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT);
> +
> +	if (userpg->cap_user_rdpmc) {
> +		if (event->hw.flags & ARMPMU_EVT_64BIT)
> +			userpg->pmc_width = 64;
> +		else
> +			userpg->pmc_width = 32;
> +	}
>  
>  	do {
>  		rd = sched_clock_read_begin(&seq);
> -- 
> 2.30.2
> 

Re: [PATCH v10 4/5] arm64: perf: Enable PMU counter userspace access for perf event

[Rob Herring]

On Thu, Oct 14, 2021 at 11:58 AM Mark Rutland <mark.rutland@arm.com> wrote:
>
> Hi Rob,
>
> This looks pretty good!
>
> I have one largish query below, and otherwise only trivialities that I'm
> happy to fix up.
>
> On Tue, Sep 14, 2021 at 03:47:59PM -0500, Rob Herring wrote:
> > Arm PMUs can support direct userspace access of counters which allows for
> > low overhead (i.e. no syscall) self-monitoring of tasks. The same feature
> > exists on x86 called 'rdpmc'. Unlike x86, userspace access will only be
> > enabled for thread bound events. This could be extended if needed, but
> > simplifies the implementation and reduces the chances for any
> > information leaks (which the x86 implementation suffers from).
> >
> > PMU EL0 access will be enabled when an event with userspace access is
> > part of the thread's context. This includes when the event is not
> > scheduled on the PMU. There's some additional overhead clearing
> > dirty counters when access is enabled in order to prevent leaking
> > disabled counter data from other tasks.
> >
> > Unlike x86, enabling of userspace access must be requested with a new
> > attr bit: config1:1. If the user requests userspace access and 64-bit
> > counters, then chaining will be disabled and the user will get the
> > maximum size counter the underlying h/w can support. The modes for
> > config1 are as follows:
> >
> > config1 = 0 : user access disabled and always 32-bit
> > config1 = 1 : user access disabled and always 64-bit (using chaining if needed)
> > config1 = 2 : user access enabled and always 32-bit
> > config1 = 3 : user access enabled and counter size matches underlying counter.
>
> We probably need to note somewhere (i.e. in the next patch) that we mean
> *logically* 32-bit, and this could be a biased 64-bit counter, so
> userspace needs to treat the upper 32-bits of counters as UNKNOWN.

Okay, though this detail doesn't matter if the user uses the correct
read loop (now in libperf).

> For the `config1 = 3` case (potentially) overriding the usual long
> semantic, I'm struggling to understand why we need that rather than
> forcing the use of a 64-bit counter, because in that case:
>
> * For a CPU_CYCLES event:
>   __armv8_pmuv3_map_event() will always pick 64-bits
>   get_event_idx() may fail to allocate a 64-bit counter.
>
> * For other events:
>   __armv8_pmuv3_map_event() will pick 32/64 based on long counter
>   support
>   get_event_idx() will only fail if there are no counters free.
>
> Whereas if __armv8_pmuv3_map_event() returned an error for the latter
> when long counter support is not implemented, we'd have consistent
> `long` semantics, and the CPU_CYCLES behaviour would be identical.
>
> What's the rationale for `3` leaving the choice to the kernel?

It's the give me the maximum sized counter the h/w can support choice.
That's easier for userspace to implement. Bit 1 is more of a hint that
the user wants userspace access rather than a requirement.

> If the problem is discoverability, I'd be happy to add something to
> sysfs to describe whether the PMU has long event support.

Checking sysfs or a try for 64-bit support then fall back to 32-bit
support isn't much difference.

Keep in mind that x86 always succeeds here. Every userspace user will
have to add whatever dance we create here. For example, each libperf
test with user access (there's only 2 in my tree, but there's a series
adding more) has to have an '#ifdef __aarch64__' for whatever we do
here. I was seeking to minimize that. Right now, that's just a set
config1 to 0x3. Also, note that libperf will opportunistically use a
userspace read instead of read(). The user just has to mmap the event
and libperf will use a userspace read when enabled which ultimately
depends on what the mmapped page says.

[...]

> > @@ -995,9 +1050,23 @@ static int __armv8_pmuv3_map_event(struct perf_event *event,
> >                                      &armv8_pmuv3_perf_cache_map,
> >                                      ARMV8_PMU_EVTYPE_EVENT);
> >
> > -     if (armv8pmu_event_is_64bit(event))
> > +     /*
> > +      * At this point, the counter is not assigned. If a 64-bit counter is
> > +      * requested, we must make sure the h/w has 64-bit counters if we set
> > +      * the event size to 64-bit because chaining is not supported with
> > +      * userspace access. This may still fail later on if the CPU cycle
> > +      * counter is in use.
> > +      */
> > +     if (armv8pmu_event_is_64bit(event) &&
> > +         (!armv8pmu_event_want_user_access(event) ||
> > +          armv8pmu_has_long_event(armpmu) || (hw_event_id == ARMV8_PMUV3_PERFCTR_CPU_CYCLES)))
> >               event->hw.flags |= ARMPMU_EVT_64BIT;
>
> If we can follow my suggestion in reply to the cover text, we can make
> this:
>
>         if (armv8pmu_event_is_64bit(event))
>                 event->hw.flags |= ARMPMU_EVT_64BIT;
>
>         /*
>          * User events must be allocated into a single counter, and so
>          * must not be chained.
>          *
>          * Most 64-bit events require long counter support, but 64-bit
>          * CPU_CYCLES events can be placed into the dedicated cycle
>          * counter when this is free.
>          *
>         if (armv8pmu_event_want_user_access()) {
>                 if (armv8pmu_event_is_64bit(event) &&
>                     (hw_event_id != ARMV8_PMUV3_PERFCTR_CPU_CYCLES) &&
>                     !armv8pmu_has_long_event(armpmu))
>                         return -EINVAL;
>         }
>
> > +     /* Userspace counter access only enabled if requested and a per task event */
> > +     if (sysctl_perf_user_access && armv8pmu_event_want_user_access(event) &&
> > +         (event->attach_state & PERF_ATTACH_TASK))
> > +             event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
>
> Can we please explicitly reject !PERF_ATTACH_TASK case?
>
> If the user requested something we don't intend to support, I'd rather
> return -EINVAL here, rather than continue on.

This is similar to the 64-bit case though I'm somewhat less concerned
here given per cpu events aren't too useful in this case and the setup
is a bit different already.

Rob

Re: [PATCH v10 4/5] arm64: perf: Enable PMU counter userspace access for perf event

[Rob Herring]

On Thu, Oct 14, 2021 at 11:58 AM Mark Rutland <mark.rutland@arm.com> wrote:
>
> Hi Rob,
>
> This looks pretty good!
>
> I have one largish query below, and otherwise only trivialities that I'm
> happy to fix up.
>
> On Tue, Sep 14, 2021 at 03:47:59PM -0500, Rob Herring wrote:

[...]

> >  static inline bool armv8pmu_event_is_chained(struct perf_event *event)
> >  {
> >       int idx = event->hw.idx;
> >       struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
> >
> > -     return !WARN_ON(idx < 0) &&
> > +     return !(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT) &&
> >              armv8pmu_event_is_64bit(event) &&
> >              !armv8pmu_has_long_event(cpu_pmu) &&
> >              (idx != ARMV8_IDX_CYCLE_COUNTER);
> > @@ -720,6 +726,27 @@ static inline u32 armv8pmu_getreset_flags(void)
> >       return value;
>
> Above this, could we please add:
>
> | static inline bool armv8pmu_event_has_user_read(struct perf_event *event)
> | {
> |       return event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT;
> | }
>
> ... and use that where we look at PERF_EVENT_FLAG_USER_READ_CNT?

Sure, but as this is a common flag now, I should probably make that a
common function in linux/perf_event.h and have x86 code use it too.

Rob

Re: [PATCH v10 4/5] arm64: perf: Enable PMU counter userspace access for perf event

[Mark Rutland]

Hi Rob,

On Thu, Oct 14, 2021 at 02:24:46PM -0500, Rob Herring wrote:
> On Thu, Oct 14, 2021 at 11:58 AM Mark Rutland <mark.rutland@arm.com> wrote:
> > On Tue, Sep 14, 2021 at 03:47:59PM -0500, Rob Herring wrote:
> > For the `config1 = 3` case (potentially) overriding the usual long
> > semantic, I'm struggling to understand why we need that rather than
> > forcing the use of a 64-bit counter, because in that case:
> >
> > * For a CPU_CYCLES event:
> >   __armv8_pmuv3_map_event() will always pick 64-bits
> >   get_event_idx() may fail to allocate a 64-bit counter.
> >
> > * For other events:
> >   __armv8_pmuv3_map_event() will pick 32/64 based on long counter
> >   support
> >   get_event_idx() will only fail if there are no counters free.
> >
> > Whereas if __armv8_pmuv3_map_event() returned an error for the latter
> > when long counter support is not implemented, we'd have consistent
> > `long` semantics, and the CPU_CYCLES behaviour would be identical.
> >
> > What's the rationale for `3` leaving the choice to the kernel?
> 
> It's the give me the maximum sized counter the h/w can support choice.
> That's easier for userspace to implement. Bit 1 is more of a hint that
> the user wants userspace access rather than a requirement.
> 
> > If the problem is discoverability, I'd be happy to add something to
> > sysfs to describe whether the PMU has long event support.
> 
> Checking sysfs or a try for 64-bit support then fall back to 32-bit
> support isn't much difference.
> 
> Keep in mind that x86 always succeeds here. Every userspace user will
> have to add whatever dance we create here. For example, each libperf
> test with user access (there's only 2 in my tree, but there's a series
> adding more) has to have an '#ifdef __aarch64__' for whatever we do
> here. I was seeking to minimize that. Right now, that's just a set
> config1 to 0x3. Also, note that libperf will opportunistically use a
> userspace read instead of read(). The user just has to mmap the event
> and libperf will use a userspace read when enabled which ultimately
> depends on what the mmapped page says.

I think that x86 always succeeding here is more of a legacy thing that
they're stuck with rather than a design to be copied.

I'd prefer to keep the existing meaning of the `long` flag to mean "give
me 64 bits of counter, somehow", with `rdpmc` meaning "give me a single
counter I can access from userspace", even if that means the combination
of the two can sometimes be rejected. As you say, we can probe for that
as necessary by trying `long` then falling back to a plain event, and if
that ends up being a bottleneck somehow we can figure out a way of
advertising support to userspace. Regardless, we should 

Importantly, I don't think libperf should override a user's request for
`long`, since the user may want to optimize for minimal perturbation
rather than faster access.

If we want a "please give me the longest counter that's compatible with
other constraints", I think that should be a new flag e.g. `trylong`,
and shouldn't override the existing `long`. We can add that as a
follow-up if we want it.

Thanks,
Mark.