On Wed, Jul 10, 2024 at 04:19:53PM +0206, John Ogness wrote:
On 2024-07-10, Daniel Vetter daniel.vetter@ffwll.ch wrote:
console_lock is the outermost subsystem lock for a lot of subsystems, which means get/put_user must nest within. Which means it cannot be acquired somewhere deeply nested in other locks, and most definitely not while holding fs locks potentially needed to resolve faults.
console_trylock is the best we can do here. But John pointed out on a previous version that this is futile:
"Using the console lock here at all is wrong. The console lock does not prevent other CPUs from calling printk() and inserting lines in between.
"There is no way to guarantee a contiguous ringbuffer block using multiple printk() calls.
"The console_lock usage should be removed."
https://lore.kernel.org/lkml/87frsh33xp.fsf@jogness.linutronix.de/
Do that.
Note that there is more of this incorrect usage of console lock in:
fs/bcachefs/debug.c:bch2_btree_verify_replica()
fs/bcachefs/bset.c:bch2_dump_btree_node()
from commit 1c6fdbd8f246("bcachefs: Initial commit")
... and its parent bcache:
drivers/md/bcache/debug.c:bch_btree_verify()
drivers/md/bcache/bset.c:bch_dump_bucket()
from commit cafe56359144("bcache: A block layer cache")
These should also be removed. Although Kent should verify that the console lock is not providing some sort of necessary side-effect synchronization.
I'll take a look, at least some of them seem doable to audit without deep bcachefs understanding. Thanks for pointing them out, I should have looked a bit more at git grep ... -Sima