On Tue, Jun 27, 2023 at 08:09:46PM -0700, Suren Baghdasaryan wrote: > On Tue, Jun 27, 2023 at 6:54 PM Tejun Heo <tj@xxxxxxxxxx> wrote: > > > > Hello, > > > > On Tue, Jun 27, 2023 at 02:58:08PM -0700, Suren Baghdasaryan wrote: > > > Ok in kernfs_generic_poll() we are using kernfs_open_node.poll > > > waitqueue head for polling and kernfs_open_node is freed from inside > > > kernfs_unlink_open_file() which is called from kernfs_fop_release(). > > > So, it is destroyed only when the last fput() is done, unlike the > > > ops->release() operation which we are using for destroying PSI > > > trigger's waitqueue. So, it seems we still need an operation which > > > would indicate that the file is truly going away. > > > > If we want to stay consistent with how kernfs behaves w.r.t. severing, the > > right thing to do would be preventing any future polling at severing and > > waking up everyone currently waiting, which sounds fine from cgroup behavior > > POV too. > > That's actually what we are currently doing for PSI triggers. > ->release() is handled by cgroup_pressure_release() which signals the > waiters, waits for RCU grace period to pass (per > https://elixir.bootlin.com/linux/latest/source/include/linux/wait.h#L258) > and then releases all the trigger resources including the waitqueue > head. However as reported in > https://lore.kernel.org/all/20230613062306.101831-1-lujialin4@xxxxxxxxxx > this does not save us from the synchronous polling case: > > do_select > vfs_poll > cgroup_pressure_release > psi_trigger_destroy > wake_up_pollfree(&t->event_wait) -> unblocks vfs_poll > synchronize_rcu() > kfree(t) -> frees waitqueue head > poll_freewait() > -> uses waitqueue head > > > This happens because we release the resources associated with the file > while there are still file users (the file's refcount is non-zero). > And that happens because kernfs can call ->release() before the last > fput(). > > > > > Now, the challenge is designing an interface which is difficult to make > > mistake with. IOW, it'd be great if kernfs wraps poll call so that severing > > is implemented without kernfs users doing anything, or at least make it > > pretty obvious what the correct usage pattern is. > > > > > Christian's suggestion to rename current ops->release() operation into > > > ops->drain() (or ops->flush() per Matthew's request) and introduce a > > > "new" ops->release() which is called only when the last fput() is done > > > seems sane to me. Would everyone be happy with that approach? > > > > I'm not sure I'd go there. The contract is that once ->release() is called, > > the code backing that file can go away (e.g. rmmod'd). It really should > > behave just like the last put from kernfs users' POV. > > I 100% agree with the above statement. > > > For this specific fix, > > it's safe because we know the ops is always built into the kernel and won't > > go away but it'd be really bad if the interface says "this is a normal thing > > to do". We'd be calling into rmmod'd text pages in no time. > > > > So, I mean, even for temporary fix, we have to make it abundantly clear that > > this is not for usual usage and can only be used if the code backing the ops > > is built into the kernel and so on. > > I think the root cause of this problem is that ->release() in kernfs > does not adhere to the common rule that ->release() is called only > when the file is going away and has no users left. Am I wrong? So imho, ultimately this all comes down to rmdir() having special semantics in kernfs. On any regular filesystem an rmdir() on a directory which is still referenced by a struct file doesn't trigger an f_op->release() operation. It's just that directory is unlinked and you get some sort of errno like ENOENT when you try to create new files in there or whatever. The actual f_op->release) however is triggered on last fput(). But in essence, kernfs treats an rmdir() operation as being equivalent to a final fput() such that it somehow magically kills all file references. And that's just wrong and not supported.