Re: [PATCH v1 2/2] epoll: Use __user_write_access_begin() and unsafe_put_user() in epoll_put_uevent().

On Fri, 24 Oct 2025 07:05:50 -0700
Dave Hansen [off-list ref] wrote:

On 10/23/25 22:16, Kuniyuki Iwashima wrote:

quoted

quoted

This makes me nervous. The access_ok() check is quite a distance away.
I'd kinda want to see some performance numbers before doing this. Is
removing a single access_ok() even measurable?  

I noticed I made a typo in commit message, s/tcp_rr/udp_rr/.

epoll_put_uevent() can be called multiple times in a single
epoll_wait(), and we can see 1.7% more pps on UDP even when
1 thread has 1000 sockets only:

server: $ udp_rr --nolog -6 -F 1000 -T 1 -l 3600
client: $ udp_rr --nolog -6 -F 1000 -T 256 -l 3600 -c -H $SERVER
server: $ nstat > /dev/null; sleep 10; nstat | grep -i udp

Without patch (2 stac/clac):
Udp6InDatagrams                 2205209            0.0

With patch (1 stac/clac):
Udp6InDatagrams                 2242602            0.0  

I'm totally with you about removing a stac/clac:

	https://lore.kernel.org/lkml/20250228203722.CAEB63AC@davehans-spike.ostc.intel.com/ (local)

The thing I'm worried about is having the access_ok() so distant
from the unsafe_put_user(). I'm wondering if this:

-	__user_write_access_begin(uevent, sizeof(*uevent));
+	if (!user_write_access_begin(uevent, sizeof(*uevent))
+		return NULL;
	unsafe_put_user(revents, &uevent->events, efault);
	unsafe_put_user(data, &uevent->data, efault);
	user_access_end();

is measurably slower than what was in your series. If it is
not measurably slower, then the series gets simpler because it
does not need to refactor user_write_access_begin(). It also ends
up more obviously correct because the access check is closer to
the unsafe_put_user() calls.

Also, the extra access_ok() is *much* cheaper than stac/clac.

access_ok() does contain a conditional branch
- just waiting for the misprediction penalty (say 20 clocks).
OTOH you shouldn't get that more that twice for the loop.

I'm pretty sure access_ok() itself contains an lfence - needed for reads.
But that ought to be absent from user_write_access_begin().

The 'masked' version uses alu operations (on x86-64) and don't need
lfence (or anything else) and don't contain a mispredictable branch.
They should be faster than the above - unless the code has serious
register pressure and too much gets spilled to stack.

The timings may also depend on the cpu you are using.
I'm sure I remember some of the very recent ones having much faster
stac/clac and/or lfence.

	David