Hi,
在 2025/07/23 19:05, Nilay Shroff 写道:
On 7/23/25 12:28 PM, Yu Kuai wrote:
BTW, this is not related to this patch. Should we handle fall_back
failure like blk_mq_sysfs_register_hctxs()?
OKay I guess you meant here handle failure case by unwinding the
queue instead of looping through it from start to end. If yes, then
it could be done but again we may not want to do it the bug fix patch.
Not like that, actually I don't have any ideas for now, the hctxs is
unregistered first, and if register failed, for example, due to -ENOMEM,
I can't find a way to fallback :(
If registering new hctxs fails, we fall back to the previous value of
nr_hw_queues (prev_nr_hw_queues). When prev_nr_hw_queues is less than
the new nr_hw_queues, we do not reallocate memory for the existing hctxs—
instead, we reuse the memory that was already allocated.
Memory allocation is only attempted for the additional hctxs beyond
prev_nr_hw_queues. Therefore, if memory allocation for these new hctxs
fails, we can safely fall back to prev_nr_hw_queues because the memory
of the previously allocated hctxs remains intact.
No, like I said before, blk_mq_sysfs_unregister_hctxs() will free memory
by kobject_del() for hctx->kobj and ctx->kobj, and
__blk_mq_update_nr_hw_queues() call that helper in the beginning.
And later in the fall back code, blk_mq_sysfs_register_hctxs() can fail
by memory allocation in kobject_add(), however, the return value is not
checked.
This can be done checking the kobject state in sysfs: kobj->state_in_sysfs.
If kobj->state_in_sysfs is 1 then it implies that kobject_add() for this
kobj was successful and we can safely call kobject_del() on it otherwise
we can skip it. We already have few places in the kernel using this trick.
For instance, check sysfs_slab_unlink(). So, IMO, similar technique could be
used for hctx->kobj and ctx->kobj as well while we attempt to delete these
kobjects from unregistering queue and nr_hw_queue update.
I don't like missing sysfs attributes, however I don't have better
solution, I'm good with this.
Thanks,
Kuai
Thanks,
--Nilay
.
