Re: block: sleeping in atomic warnings

On Tue, Feb 07, 2023 at 08:15:04AM -0800, Linus Torvalds wrote:

On Tue, Feb 7, 2023 at 6:06 AM Dan Carpenter [off-list ref] wrote:

quoted

block/blk-crypto-profile.c:382 __blk_crypto_evict_key() warn: sleeping in atomic context
block/blk-crypto-profile.c:390 __blk_crypto_evict_key() warn: sleeping in atomic context

Yeah, that looks very real, but doesn't really seem to be a block bug.

__put_super() has a big comment that it's called under the sb_lock
spinlock, so it's all in atomic context, but then:

quoted

-> __put_super()
   -> fscrypt_destroy_keyring()
      -> fscrypt_put_master_key_activeref()
         -> fscrypt_destroy_prepared_key()
            -> fscrypt_destroy_inline_crypt_key()
               -> blk_crypto_evict_key()

and we have a comment in __blk_crypto_evict_key() that it must be
called in "process context".

However, the *normal* unmount sequence does all the cleanup *before*
it gets sb_lock, and calls fscrypt_destroy_keyring() in process
context, which is probably why it never triggers in practice, because
the "last put" is normally there, not in __put_super.

Eric? Al?

It smells like __put_super() may need to do some parts delayed, not
under sb_lock.

It's a false positive.  See the comment above fscrypt_destroy_keyring(), which
is meant to explain this, though I can update the comment to be clearer.  If the
filesystem has been mounted, then fscrypt_destroy_keyring() is called from
generic_shutdown_super(), which can sleep, and the call from __put_super() is a
no-op.  If the filesystem has not been mounted, then the call from __put_super()
is needed, but blk_crypto_evict_key() can never be executed in that case.

- Eric