The perf_prepare_sample() sets the perf_sample_data according to the
attr->sample_type before copying it to the ring buffer. But BPF also
wants to access the sample data so it needs to prepare the sample even
before the regular path.
That means the perf_prepare_sample() can be called more than once. Set
the data->sample_flags consistently so that it can indicate which fields
are set already and skip them if sets.
Also update the filtered_sample_type to have the dependent flags to
reduce the number of branches.
Suggested-by: Peter Zijlstra <[email protected]>
Signed-off-by: Namhyung Kim <[email protected]>
---
kernel/events/core.c | 85 +++++++++++++++++++++++++++++++++-----------
1 file changed, 65 insertions(+), 20 deletions(-)
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 0218b6ffaf36..7c5555051662 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -7053,6 +7053,14 @@ static void perf_aux_sample_output(struct perf_event *event,
ring_buffer_put(rb);
}
+/*
+ * A set of common sample data types saved even for non-sample records
+ * when event->attr.sample_id_all is set.
+ */
+#define PERF_SAMPLE_ID_ALL (PERF_SAMPLE_TID | PERF_SAMPLE_TIME | \
+ PERF_SAMPLE_CPU | PERF_SAMPLE_ID | \
+ PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_STREAM_ID)
+
static void __perf_event_header__init_id(struct perf_event_header *header,
struct perf_sample_data *data,
struct perf_event *event,
@@ -7561,6 +7569,11 @@ perf_callchain(struct perf_event *event, struct pt_regs *regs)
return callchain ?: &__empty_callchain;
}
+static __always_inline u64 __cond_set(u64 flags, u64 s, u64 d)
+{
+ return d * !!(flags & s);
+}
+
void perf_prepare_sample(struct perf_event_header *header,
struct perf_sample_data *data,
struct perf_event *event,
@@ -7576,14 +7589,25 @@ void perf_prepare_sample(struct perf_event_header *header,
header->misc |= perf_misc_flags(regs);
/*
- * Clear the sample flags that have already been done by the
- * PMU driver.
+ * Add the sample flags that are dependent to others. And clear the
+ * sample flags that have already been done by the PMU driver.
*/
- filtered_sample_type = sample_type & ~data->sample_flags;
+ filtered_sample_type = sample_type;
+ filtered_sample_type |= __cond_set(sample_type, PERF_SAMPLE_CODE_PAGE_SIZE,
+ PERF_SAMPLE_IP);
+ filtered_sample_type |= __cond_set(sample_type, PERF_SAMPLE_DATA_PAGE_SIZE |
+ PERF_SAMPLE_PHYS_ADDR, PERF_SAMPLE_ADDR);
+ filtered_sample_type |= __cond_set(sample_type, PERF_SAMPLE_STACK_USER,
+ PERF_SAMPLE_REGS_USER);
+ filtered_sample_type &= ~data->sample_flags;
+
__perf_event_header__init_id(header, data, event, filtered_sample_type);
+ data->sample_flags |= sample_type & PERF_SAMPLE_ID_ALL;
- if (sample_type & (PERF_SAMPLE_IP | PERF_SAMPLE_CODE_PAGE_SIZE))
+ if (filtered_sample_type & PERF_SAMPLE_IP) {
data->ip = perf_instruction_pointer(regs);
+ data->sample_flags |= PERF_SAMPLE_IP;
+ }
if (filtered_sample_type & PERF_SAMPLE_CALLCHAIN)
perf_sample_save_callchain(data, event, regs);
@@ -7600,10 +7624,15 @@ void perf_prepare_sample(struct perf_event_header *header,
data->sample_flags |= PERF_SAMPLE_BRANCH_STACK;
}
- if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
+ if (filtered_sample_type & PERF_SAMPLE_REGS_USER)
perf_sample_regs_user(&data->regs_user, regs);
- if (sample_type & PERF_SAMPLE_REGS_USER) {
+ /*
+ * It cannot use the filtered_sample_type here as REGS_USER can be set
+ * by STACK_USER (using __cond_set() above) and we don't want to update
+ * the dyn_size if it's not requested by users.
+ */
+ if ((sample_type & ~data->sample_flags) & PERF_SAMPLE_REGS_USER) {
/* regs dump ABI info */
int size = sizeof(u64);
@@ -7613,9 +7642,10 @@ void perf_prepare_sample(struct perf_event_header *header,
}
data->dyn_size += size;
+ data->sample_flags |= PERF_SAMPLE_REGS_USER;
}
- if (sample_type & PERF_SAMPLE_STACK_USER) {
+ if (filtered_sample_type & PERF_SAMPLE_STACK_USER) {
/*
* Either we need PERF_SAMPLE_STACK_USER bit to be always
* processed as the last one or have additional check added
@@ -7638,23 +7668,30 @@ void perf_prepare_sample(struct perf_event_header *header,
data->stack_user_size = stack_size;
data->dyn_size += size;
+ data->sample_flags |= PERF_SAMPLE_STACK_USER;
}
- if (filtered_sample_type & PERF_SAMPLE_WEIGHT_TYPE)
+ if (filtered_sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
data->weight.full = 0;
+ data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
+ }
- if (filtered_sample_type & PERF_SAMPLE_DATA_SRC)
+ if (filtered_sample_type & PERF_SAMPLE_DATA_SRC) {
data->data_src.val = PERF_MEM_NA;
+ data->sample_flags |= PERF_SAMPLE_DATA_SRC;
+ }
- if (filtered_sample_type & PERF_SAMPLE_TRANSACTION)
+ if (filtered_sample_type & PERF_SAMPLE_TRANSACTION) {
data->txn = 0;
+ data->sample_flags |= PERF_SAMPLE_TRANSACTION;
+ }
- if (sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR | PERF_SAMPLE_DATA_PAGE_SIZE)) {
- if (filtered_sample_type & PERF_SAMPLE_ADDR)
- data->addr = 0;
+ if (filtered_sample_type & PERF_SAMPLE_ADDR) {
+ data->addr = 0;
+ data->sample_flags |= PERF_SAMPLE_ADDR;
}
- if (sample_type & PERF_SAMPLE_REGS_INTR) {
+ if (filtered_sample_type & PERF_SAMPLE_REGS_INTR) {
/* regs dump ABI info */
int size = sizeof(u64);
@@ -7667,19 +7704,22 @@ void perf_prepare_sample(struct perf_event_header *header,
}
data->dyn_size += size;
+ data->sample_flags |= PERF_SAMPLE_REGS_INTR;
}
- if (sample_type & PERF_SAMPLE_PHYS_ADDR &&
- filtered_sample_type & PERF_SAMPLE_PHYS_ADDR)
+ if (filtered_sample_type & PERF_SAMPLE_PHYS_ADDR) {
data->phys_addr = perf_virt_to_phys(data->addr);
+ data->sample_flags |= PERF_SAMPLE_PHYS_ADDR;
+ }
#ifdef CONFIG_CGROUP_PERF
- if (sample_type & PERF_SAMPLE_CGROUP) {
+ if (filtered_sample_type & PERF_SAMPLE_CGROUP) {
struct cgroup *cgrp;
/* protected by RCU */
cgrp = task_css_check(current, perf_event_cgrp_id, 1)->cgroup;
data->cgroup = cgroup_id(cgrp);
+ data->sample_flags |= PERF_SAMPLE_CGROUP;
}
#endif
@@ -7688,13 +7728,17 @@ void perf_prepare_sample(struct perf_event_header *header,
* require PERF_SAMPLE_ADDR, kernel implicitly retrieve the data->addr,
* but the value will not dump to the userspace.
*/
- if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)
+ if (filtered_sample_type & PERF_SAMPLE_DATA_PAGE_SIZE) {
data->data_page_size = perf_get_page_size(data->addr);
+ data->sample_flags |= PERF_SAMPLE_DATA_PAGE_SIZE;
+ }
- if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)
+ if (filtered_sample_type & PERF_SAMPLE_CODE_PAGE_SIZE) {
data->code_page_size = perf_get_page_size(data->ip);
+ data->sample_flags |= PERF_SAMPLE_CODE_PAGE_SIZE;
+ }
- if (sample_type & PERF_SAMPLE_AUX) {
+ if (filtered_sample_type & PERF_SAMPLE_AUX) {
u64 size;
header->size += sizeof(u64); /* size */
@@ -7712,6 +7756,7 @@ void perf_prepare_sample(struct perf_event_header *header,
WARN_ON_ONCE(size + header->size > U16_MAX);
data->dyn_size += size + sizeof(u64); /* size above */
+ data->sample_flags |= PERF_SAMPLE_AUX;
}
header->size += data->dyn_size;
--
2.39.0.314.g84b9a713c41-goog
On Thu, Jan 12, 2023 at 01:40:12PM -0800, Namhyung Kim wrote:
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -7053,6 +7053,14 @@ static void perf_aux_sample_output(struct perf_event *event,
> ring_buffer_put(rb);
> }
>
> +/*
> + * A set of common sample data types saved even for non-sample records
> + * when event->attr.sample_id_all is set.
> + */
> +#define PERF_SAMPLE_ID_ALL (PERF_SAMPLE_TID | PERF_SAMPLE_TIME | \
> + PERF_SAMPLE_CPU | PERF_SAMPLE_ID | \
> + PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_STREAM_ID)
> +
> __perf_event_header__init_id(header, data, event, filtered_sample_type);
> + data->sample_flags |= sample_type & PERF_SAMPLE_ID_ALL;
Why not update __perf_event_header__init_id() ?
On Fri, Jan 13, 2023 at 3:20 AM Peter Zijlstra <[email protected]> wrote:
>
> On Thu, Jan 12, 2023 at 01:40:12PM -0800, Namhyung Kim wrote:
>
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -7053,6 +7053,14 @@ static void perf_aux_sample_output(struct perf_event *event,
> > ring_buffer_put(rb);
> > }
> >
> > +/*
> > + * A set of common sample data types saved even for non-sample records
> > + * when event->attr.sample_id_all is set.
> > + */
> > +#define PERF_SAMPLE_ID_ALL (PERF_SAMPLE_TID | PERF_SAMPLE_TIME | \
> > + PERF_SAMPLE_CPU | PERF_SAMPLE_ID | \
> > + PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_STREAM_ID)
> > +
>
>
> > __perf_event_header__init_id(header, data, event, filtered_sample_type);
> > + data->sample_flags |= sample_type & PERF_SAMPLE_ID_ALL;
>
> Why not update __perf_event_header__init_id() ?
My concern was the non-sample records which don't check the sample
flags at all. So I'd like to skip the write for them.
But that's not a big deal. I can move it if you want.
Thanks,
Namhyung