From: Linus Torvalds torvalds@linux-foundation.org
[ Upstream commit ef0ba05538299f1391cbe097de36895bb36ecfe6 ]
The kernel test robot reported a -5.8% performance regression on the "poll2" test of will-it-scale, and bisected it to commit d55564cfc222 ("x86: Make __put_user() generate an out-of-line call").
I didn't expect an out-of-line __put_user() to matter, because no normal core code should use that non-checking legacy version of user access any more. But I had overlooked the very odd poll() usage, which does a __put_user() to update the 'revents' values of the poll array.
Now, Al Viro correctly points out that instead of updating just the 'revents' field, it would be much simpler to just copy the _whole_ pollfd entry, and then we could just use "copy_to_user()" on the whole array of entries, the same way we use "copy_from_user()" a few lines earlier to get the original values.
But that is not what we've traditionally done, and I worry that threaded applications might be concurrently modifying the other fields of the pollfd array. So while Al's suggestion is simpler - and perhaps worth trying in the future - this instead keeps the "just update revents" model.
To fix the performance regression, use the modern "unsafe_put_user()" instead of __put_user(), with the proper "user_write_access_begin()" guarding in place. This improves code generation enormously.
Link: https://lore.kernel.org/lkml/20210107134723.GA28532@xsang-OptiPlex-9020/ Reported-by: kernel test robot oliver.sang@intel.com Tested-by: Oliver Sang oliver.sang@intel.com Cc: Al Viro viro@zeniv.linux.org.uk Cc: David Laight David.Laight@aculab.com Cc: Peter Zijlstra peterz@infradead.org Signed-off-by: Linus Torvalds torvalds@linux-foundation.org Signed-off-by: Sasha Levin sashal@kernel.org --- fs/select.c | 14 +++++++++++--- 1 file changed, 11 insertions(+), 3 deletions(-)
diff --git a/fs/select.c b/fs/select.c index ebfebdfe5c69a..37aaa8317f3ae 100644 --- a/fs/select.c +++ b/fs/select.c @@ -1011,14 +1011,17 @@ static int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, fdcount = do_poll(head, &table, end_time); poll_freewait(&table);
+ if (!user_write_access_begin(ufds, nfds * sizeof(*ufds))) + goto out_fds; + for (walk = head; walk; walk = walk->next) { struct pollfd *fds = walk->entries; int j;
- for (j = 0; j < walk->len; j++, ufds++) - if (__put_user(fds[j].revents, &ufds->revents)) - goto out_fds; + for (j = walk->len; j; fds++, ufds++, j--) + unsafe_put_user(fds->revents, &ufds->revents, Efault); } + user_write_access_end();
err = fdcount; out_fds: @@ -1030,6 +1033,11 @@ static int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, }
return err; + +Efault: + user_write_access_end(); + err = -EFAULT; + goto out_fds; }
static long do_restart_poll(struct restart_block *restart_block)