On 9/23/21 11:55 AM, Jens Axboe wrote:
> On 9/23/21 11:51 AM, Bart Van Assche wrote:
>> On 9/23/21 9:39 AM, Jens Axboe wrote:
>>> On 9/23/21 10:22 AM, Bart Van Assche wrote:
>>>> On 9/23/21 9:04 AM, Jens Axboe wrote:
>>>>> What options are you loading null_blk with?
>>>>
>>>> The issues I reported are the result of running test blk/010 from the
>>>> blktests suite. That test loads the null_blk kernel module twice:
>>>>
>>>> _init_null_blk queue_mode=2 submit_queues=16 nr_devices=32
>>>> [ ... ]
>>>> _exit_null_blk
>>>> _init_null_blk queue_mode=2 submit_queues=16 nr_devices=32 shared_tags=1
>>>> [ ... ]
>>>> _exit_null_blk
>>>>
>>>> Please let me know if you need more information.
>>>
>>> Tried both that and running block/010, didn't trigger anything for me.
>>> Odd...
>>
>> Hi Jens,
>>
>> I took another look at the kernel logs from yesterday of the VM that I use
>> for testing. In that kernel log I found the following:
>> * Without any changes on top of the for-next branch of
>> git://git.kernel.dk/linux-block (commit 4129031563d0 ("Merge branch
>> 'for-5.16/io_uring' into for-next"), test block/010 triggers the oops
>> reported at the start of this email thread.
>> * With the patch at the start of this email thread applied, the first test
>> that triggers a kernel oops is block/020 (blk_mq_free_rqs+0x1f4).
>>
>> This morning I rebuilt the block for-next kernel with and without my
>> null_blk patch applied. I was able to reproduce what I observed yesterday.
>> Test block/020 passes with if commit 5f7acddf706c ("null_blk: poll queue
>> support") is reverted. This is why I wrote that my patch does not seem to
>> be sufficient to fix commit 5f7acddf706c.
>
> Ah ok, so it's block/020, not block/010 for the later one. I'll take a
> look.
I think it's the assumption that poll_queueus == submit_queues. Does this
work for you?
diff --git a/drivers/block/null_blk/main.c b/drivers/block/null_blk/main.c
index eb5cfe189e90..dac88c5daff9 100644
--- a/drivers/block/null_blk/main.c
+++ b/drivers/block/null_blk/main.c
@@ -1748,7 +1757,7 @@ static int setup_queues(struct nullb *nullb)
int nqueues = nr_cpu_ids;
if (g_poll_queues)
- nqueues *= 2;
+ nqueues += g_poll_queues;
nullb->queues = kcalloc(nqueues, sizeof(struct nullb_queue),
GFP_KERNEL);
@@ -1814,7 +1823,7 @@ static int null_init_tag_set(struct nullb *nullb, struct blk_mq_tag_set *set)
g_submit_queues;
poll_queues = nullb ? nullb->dev->poll_queues : g_poll_queues;
if (poll_queues)
- set->nr_hw_queues *= 2;
+ set->nr_hw_queues += poll_queues;
set->queue_depth = nullb ? nullb->dev->hw_queue_depth :
g_hw_queue_depth;
set->numa_node = nullb ? nullb->dev->home_node : g_home_node;
--
Jens Axboe
