On 04/06/2018 03:41 PM, Ming Lei wrote:
> On Fri, Apr 06, 2018 at 12:19:19PM +0200, Christian Borntraeger wrote:
>>
>>
>> On 04/06/2018 11:23 AM, Ming Lei wrote:
>>> On Fri, Apr 06, 2018 at 10:51:28AM +0200, Christian Borntraeger wrote:
>>>>
>>>>
>>>> On 04/06/2018 10:41 AM, Ming Lei wrote:
>>>>> On Thu, Apr 05, 2018 at 07:39:56PM +0200, Christian Borntraeger wrote:
>>>>>>
>>>>>>
>>>>>> On 04/05/2018 06:11 PM, Ming Lei wrote:
>>>>>>>>
>>>>>>>> Could you please apply the following patch and provide the dmesg boot log?
>>>>>>>
>>>>>>> And please post out the 'lscpu' log together from the test machine too.
>>>>>>
>>>>>> attached.
>>>>>>
>>>>>> As I said before this seems to go way with CONFIG_NR_CPUS=64 or smaller.
>>>>>> We have 282 nr_cpu_ids here (max 141CPUs on that z13 with SMT2) but only 8 Cores
>>>>>> == 16 threads.
>>>>>
>>>>> OK, thanks!
>>>>>
>>>>> The most weird thing is that hctx->next_cpu is computed as 512 since
>>>>> nr_cpu_id is 282, and hctx->next_cpu should have pointed to one of
>>>>> possible CPU.
>>>>>
>>>>> Looks like it is a s390 specific issue, since I can setup one queue
>>>>> which has same mapping with yours:
>>>>>
>>>>> - nr_cpu_id is 282
>>>>> - CPU 0~15 is online
>>>>> - 64 queues null_blk
>>>>> - still run all hw queues in .complete handler
>>>>>
>>>>> But can't reproduce this issue at all.
>>>>>
>>>>> So please test the following patch, which may tell us why hctx->next_cpu
>>>>> is computed wrong:
>>>>
>>>> I see things like
>>>>
>>>> [ 8.196907] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>> [ 8.196910] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>> [ 8.196912] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>> [ 8.196913] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>> [ 8.196914] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>> [ 8.196915] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>> [ 8.196916] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>> [ 8.196916] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>> [ 8.196917] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>> [ 8.196918] wrong next_cpu 512, blk_mq_map_swqueue, first_and
>>>>
>>>> which is exactly what happens if the find and and operation fails (returns size of bitmap).
>>>
>>> Given both 'cpu_online_mask' and 'hctx->cpumask' are shown as correct
>>> in your previous debug log, it means the following function returns
>>> totally wrong result on S390.
>>>
>>> cpumask_first_and(hctx->cpumask, cpu_online_mask);
>>>
>>> The debugfs log shows that each hctx->cpumask includes one online
>>> CPU(0~15).
>>
>> Really? the last log (with the latest patch applied shows a lot of contexts
>> that do not have CPUs in 0-15:
>>
>> e.g.
>> [ 4.049828] dump CPUs mapped to this hctx:
>> [ 4.049829] 18
>> [ 4.049829] 82
>> [ 4.049830] 146
>> [ 4.049830] 210
>> [ 4.049831] 274
>
> That won't be an issue, since no IO can be submitted from these offline
> CPUs, then these hctx shouldn't have been run at all.
>
> But hctx->next_cpu can be set as 512 for these inactive hctx in
> blk_mq_map_swqueue(), then please test the attached patch, and if
> hctx->next_cpu is still set as 512, something is still wrong.
WIth this patch I no longer see the "run queue from wrong CPU x, hctx active" messages.
your debug code still triggers, though.
wrong next_cpu 512, blk_mq_hctx_next_cpu, first_and
wrong next_cpu 512, blk_mq_hctx_next_cpu, next_and
If we would remove the debug code then dmesg would be clean it seems.
> ---
>
> diff --git a/block/blk-mq-cpumap.c b/block/blk-mq-cpumap.c
> index 9f8cffc8a701..638ab5c11b3c 100644
> --- a/block/blk-mq-cpumap.c
> +++ b/block/blk-mq-cpumap.c
> @@ -14,13 +14,12 @@
> #include "blk.h"
> #include "blk-mq.h"
>
> +/*
> + * Given there isn't CPU hotplug handler in blk-mq, map all CPUs to
> + * queues even it isn't present yet.
> + */
> static int cpu_to_queue_index(unsigned int nr_queues, const int cpu)
> {
> - /*
> - * Non present CPU will be mapped to queue index 0.
> - */
> - if (!cpu_present(cpu))
> - return 0;
> return cpu % nr_queues;
> }
>
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 90838e998f66..1a834d96a718 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -1343,6 +1343,13 @@ static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
> hctx_unlock(hctx, srcu_idx);
> }
>
> +static void check_next_cpu(int next_cpu, const char *str1, const char *str2)
> +{
> + if (next_cpu > nr_cpu_ids)
> + printk_ratelimited("wrong next_cpu %d, %s, %s\n",
> + next_cpu, str1, str2);
> +}
> +
> /*
> * It'd be great if the workqueue API had a way to pass
> * in a mask and had some smarts for more clever placement.
> @@ -1352,26 +1359,29 @@ static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
> static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
> {
> bool tried = false;
> + int next_cpu = hctx->next_cpu;
>
> if (hctx->queue->nr_hw_queues == 1)
> return WORK_CPU_UNBOUND;
>
> if (--hctx->next_cpu_batch <= 0) {
> - int next_cpu;
> select_cpu:
> - next_cpu = cpumask_next_and(hctx->next_cpu, hctx->cpumask,
> + next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
> cpu_online_mask);
> - if (next_cpu >= nr_cpu_ids)
> + check_next_cpu(next_cpu, __func__, "next_and");
> + if (next_cpu >= nr_cpu_ids) {
> next_cpu = cpumask_first_and(hctx->cpumask,cpu_online_mask);
> + check_next_cpu(next_cpu, __func__, "first_and");
> + }
>
> /*
> * No online CPU is found, so have to make sure hctx->next_cpu
> * is set correctly for not breaking workqueue.
> */
> - if (next_cpu >= nr_cpu_ids)
> - hctx->next_cpu = cpumask_first(hctx->cpumask);
> - else
> - hctx->next_cpu = next_cpu;
> + if (next_cpu >= nr_cpu_ids) {
> + next_cpu = cpumask_first(hctx->cpumask);
> + check_next_cpu(next_cpu, __func__, "first");
> + }
> hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
> }
>
> @@ -1379,7 +1389,7 @@ static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
> * Do unbound schedule if we can't find a online CPU for this hctx,
> * and it should only happen in the path of handling CPU DEAD.
> */
> - if (!cpu_online(hctx->next_cpu)) {
> + if (!cpu_online(next_cpu)) {
> if (!tried) {
> tried = true;
> goto select_cpu;
> @@ -1392,7 +1402,9 @@ static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
> hctx->next_cpu_batch = 1;
> return WORK_CPU_UNBOUND;
> }
> - return hctx->next_cpu;
> +
> + hctx->next_cpu = next_cpu;
> + return next_cpu;
> }
>
> static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
> @@ -2408,6 +2420,8 @@ static void blk_mq_map_swqueue(struct request_queue *q)
> mutex_unlock(&q->sysfs_lock);
>
> queue_for_each_hw_ctx(q, hctx, i) {
> + int next_cpu;
> +
> /*
> * If no software queues are mapped to this hardware queue,
> * disable it and free the request entries.
> @@ -2437,8 +2451,12 @@ static void blk_mq_map_swqueue(struct request_queue *q)
> /*
> * Initialize batch roundrobin counts
> */
> - hctx->next_cpu = cpumask_first_and(hctx->cpumask,
> + next_cpu = cpumask_first_and(hctx->cpumask,
> cpu_online_mask);
> + if (next_cpu >= nr_cpu_ids)
> + next_cpu = cpumask_first(hctx->cpumask);
> + check_next_cpu(next_cpu, __func__, "first_and");
> + hctx->next_cpu = next_cpu;
> hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
> }
> }
> Thanks,
> Ming
>
