shmem: move spinlock into shmem_recalc_inode() to fix quota support

Commit "shmem: fix quota lock nesting in huge hole handling" was not so
good: Smatch caught shmem_recalc_inode()'s shmem_inode_unacct_blocks()
descending into quota_send_warning(): where blocking GFP_NOFS is used,
yet shmem_recalc_inode() is called holding the shmem inode's info->lock.

Yes, both __dquot_alloc_space() and __dquot_free_space() are commented
"This operation can block, but only after everything is updated" - when
calling flush_warnings() at the end - both its print_warning() and its
quota_send_warning() may block.

Rework shmem_recalc_inode() to take the shmem inode's info->lock inside,
and drop it before calling shmem_inode_unacct_blocks().

And why were the spin_locks disabling interrupts?  That was just a relic
from when shmem_charge() and shmem_uncharge() were called while holding
i_pages xa_lock: stop disabling interrupts for info->lock now.

To help stop me from making the same mistake again, add a might_sleep()
into shmem_inode_acct_block() and shmem_inode_unacct_blocks(); and those
functions have grown, so let the compiler decide whether to inline them.

Reported-by: Dan Carpenter <[email protected]>
Closes: https://lore.kernel.org/linux-fsdevel/[email protected]/
Signed-off-by: Hugh Dickins <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Message-Id: <[email protected]>
Signed-off-by: Christian Brauner <[email protected]>

0 files changed, 0 insertions, 0 deletions