On 15/12/2021 09:57, Emanuele Giuseppe Esposito wrote:
On 10/12/2021 15:38, Hanna Reitz wrote:
On 24.11.21 07:43, Emanuele Giuseppe Esposito wrote:
Fuse logic can be classified as I/O, so there is no BQL held
during its execution. And yet, it uses blk_{get/set}_perm
functions, that are classified as BQL and clearly require
the BQL lock. Since there is no easy solution for this,
add a couple of TODOs and FIXME in the relevant sections of the
code.
Well, the problem goes deeper, because we still consider them GS
functions. So it’s fine for the permission function
raw_handle_perm_lock() to call bdrv_get_flags(), which is a GS
function, and then you still get:
qemu-storage-daemon: ../block.c:6195: bdrv_get_flags: Assertion
`qemu_in_main_thread()' failed.
Wait... Now that I read it more carefully I am confused about this. I
don't think the above has to do with fuse, right?
Can you share the command that makes qemu-storage-daemon fail?
raw_handle_perm_lock() is currently called by these three callbacks:
.bdrv_check_perm = raw_check_perm,
.bdrv_set_perm = raw_set_perm,
.bdrv_abort_perm_update = raw_abort_perm_update,
all three function pointers are defined as GS functions. So I don't
understand why is this failing.
It looks like in this case making bdrv_get_flags() not a GS function
would be feasible and might solve the problem, but I guess we should
instead think about adding something like
if (!exp->growable && !qemu_in_main_thread()) {
/* Changing permissions like below only works in the main thread */
return -EPERM;
}
to fuse_do_truncate().
Ideally, to make up for this, we should probably take the RESIZE
permission in fuse_export_create() for writable exports in iothreads.
I think I got it. I will send the RESIZE permission fix in a separate
patch.
Thank you,
Emanuele
