Message ID | 20220409205134.13070-1-brgl@bgdev.pl |
---|---|
State | New |
Headers | show |
Series | [GIT,PULL] gpio: fixes for v5.18-rc2 | expand |
On Sat, Apr 9, 2022 at 10:51 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote: > > Here's a single fix for a race condition between the GPIO core and consumers of > GPIO IRQ chips. I've pulled this, but it's horribly broken. You can't just use a compiler barrier to make sure the compiler orders the data at initialization time. That doesn't take care of CPU re-ordering, but it also doesn't take care of re-ordering reads on the other side of the equation. Every write barrier needs to pair with a read barrier. And "barrier()" is only a barrier on that CPU, since it is only a barrier for code generation, not for data. There are multiple ways to do proper hand-off of data, but the best one is likely - on the initialization side, do .. initialize all the data, then do .. smp_store_release(&initialized, 1); - on the reading side, do if (!smp_load_acquire(&initialized)) return -EAGAIN; .. you can now rely on all the data having been initialized .. But honestly, the fact that you got this race condition so wrong makes me suggest you use proper locks. Because the above gives you proper ordering between the two sequences, but the sequences in question still have to have a *lot* of guarantees about the accesses actually then being valid in a lock-free environment (the only obviously safe situation is a "initialize things once, everything afterwards is only a read" - otherwise y ou need to make sure all the *updates* are safely done too). With locking, all these issues go away. The lock will take care of ordering, but also data consistency at updates. Without locking, you need to do the above kinds of careful things for _all_ the accesses that can race, not just that "initialized" flag. Linus
The pull request you sent on Sat, 9 Apr 2022 22:51:34 +0200:
> git://git.kernel.org/pub/scm/linux/kernel/git/brgl/linux.git tags/gpio-fixes-for-v5.18-rc2
has been merged into torvalds/linux.git:
https://git.kernel.org/torvalds/c/fa3b895da8e06d6e3dcf3e6941a3fd428343e3d7
Thank you!
On Sun, Apr 10, 2022 at 6:27 AM Linus Torvalds <torvalds@linux-foundation.org> wrote: > > On Sat, Apr 9, 2022 at 10:51 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote: > > > > Here's a single fix for a race condition between the GPIO core and consumers of > > GPIO IRQ chips. > > I've pulled this, but it's horribly broken. > > You can't just use a compiler barrier to make sure the compiler orders > the data at initialization time. > > That doesn't take care of CPU re-ordering, but it also doesn't take > care of re-ordering reads on the other side of the equation. > > Every write barrier needs to pair with a read barrier. > > And "barrier()" is only a barrier on that CPU, since it is only a > barrier for code generation, not for data. > > There are multiple ways to do proper hand-off of data, but the best > one is likely > > - on the initialization side, do > > .. initialize all the data, then do .. > smp_store_release(&initialized, 1); > > - on the reading side, do > > if (!smp_load_acquire(&initialized)) > return -EAGAIN; > > .. you can now rely on all the data having been initialized .. > > But honestly, the fact that you got this race condition so wrong makes > me suggest you use proper locks. Because the above gives you proper > ordering between the two sequences, but the sequences in question > still have to have a *lot* of guarantees about the accesses actually > then being valid in a lock-free environment (the only obviously safe > situation is a "initialize things once, everything afterwards is only > a read" - otherwise y ou need to make sure all the *updates* are > safely done too). > > With locking, all these issues go away. The lock will take care of > ordering, but also data consistency at updates. > > Without locking, you need to do the above kinds of careful things for > _all_ the accesses that can race, not just that "initialized" flag. > > Linus Cc'ing Shreeya Thanks, we'll see about a follow-up with a better solution. Bart