[PATCH v1] perf_event_open.2: clarify and expand memory barrier requirements

[PATCH v1] perf_event_open.2: clarify and expand memory barrier requirements

[Avi Kivity]

perf_event_open(2) instructs the user to issue an rmb() after reading
data_head to ensure that user-space sees all writes to the memory
it reads. rmb() is a kernel-internal term that might not mean much
to the reader; and further it is too strict. It's enough to require
the weaker load-acquire fence. This is an industry standard term
that does not require the user to understand kernel terminology.

In addition, require a store-release fence before writing data_tail.
This prevents the user's reads from being reordered with the kernel's
writes to the just-freed space. The documentation in <linux/perf_event.h>
also suggests doing this.

Signed-off-by: Avi Kivity <avi@scylladb.com>
---
 man2/perf_event_open.2 | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/man2/perf_event_open.2 b/man2/perf_event_open.2
index 81c1b10f2..db5ce746b 100644
--- a/man2/perf_event_open.2
+++ b/man2/perf_event_open.2
@@ -1837,18 +1837,19 @@ The value needs to be manually wrapped by the size of the mmap buffer
 before accessing the samples.
 .IP
 On SMP-capable platforms, after reading the
 .I data_head
 value,
-user space should issue an rmb().
+user space should issue a load-acquire fence.
 .TP
 .I data_tail
 When the mapping is
 .BR PROT_WRITE ,
 the
 .I data_tail
 value should be written by user space to reflect the last read data.
+Before writing, issue a store-release fence.
 In this case, the kernel will not overwrite unread data.
 .TP
 .IR data_offset " (since Linux 4.1)"
 .\" commit e8c6deac69629c0cb97c3d3272f8631ef17f8f0f
 Contains the offset of the location in the mmap buffer
-- 
2.31.1

Re: [PATCH v1] perf_event_open.2: clarify and expand memory barrier requirements

[Avi Kivity]

Bump

On 19/09/2021 20.36, Avi Kivity wrote:
> perf_event_open(2) instructs the user to issue an rmb() after reading
> data_head to ensure that user-space sees all writes to the memory
> it reads. rmb() is a kernel-internal term that might not mean much
> to the reader; and further it is too strict. It's enough to require
> the weaker load-acquire fence. This is an industry standard term
> that does not require the user to understand kernel terminology.
>
> In addition, require a store-release fence before writing data_tail.
> This prevents the user's reads from being reordered with the kernel's
> writes to the just-freed space. The documentation in <linux/perf_event.h>
> also suggests doing this.
>
> Signed-off-by: Avi Kivity <avi@scylladb.com>
> ---
>   man2/perf_event_open.2 | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/man2/perf_event_open.2 b/man2/perf_event_open.2
> index 81c1b10f2..db5ce746b 100644
> --- a/man2/perf_event_open.2
> +++ b/man2/perf_event_open.2
> @@ -1837,18 +1837,19 @@ The value needs to be manually wrapped by the size of the mmap buffer
>   before accessing the samples.
>   .IP
>   On SMP-capable platforms, after reading the
>   .I data_head
>   value,
> -user space should issue an rmb().
> +user space should issue a load-acquire fence.
>   .TP
>   .I data_tail
>   When the mapping is
>   .BR PROT_WRITE ,
>   the
>   .I data_tail
>   value should be written by user space to reflect the last read data.
> +Before writing, issue a store-release fence.
>   In this case, the kernel will not overwrite unread data.
>   .TP
>   .IR data_offset " (since Linux 4.1)"
>   .\" commit e8c6deac69629c0cb97c3d3272f8631ef17f8f0f
>   Contains the offset of the location in the mmap buffer