On 4/17/18 10:42 AM, Kees Cook wrote:
> On Mon, Apr 16, 2018 at 8:12 PM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>> With a hardware watchpoint, I've isolated the corruption to here:
>>
>> bfq_dispatch_request+0x2be/0x1610:
>> __bfq_dispatch_request at block/bfq-iosched.c:3902
>> 3900 if (rq) {
>> 3901 inc_in_driver_start_rq:
>> 3902 bfqd->rq_in_driver++;
>> 3903 start_rq:
>> 3904 rq->rq_flags |= RQF_STARTED;
>> 3905 }
>
> FWIW, the stacktrace here (removing the ? lines) is:
>
> [ 34.311980] RIP: 0010:bfq_dispatch_request+0x2be/0x1610
> [ 34.452491] blk_mq_do_dispatch_sched+0x1d9/0x260
> [ 34.454561] blk_mq_sched_dispatch_requests+0x3da/0x4b0
> [ 34.458789] __blk_mq_run_hw_queue+0xae/0x130
> [ 34.460001] __blk_mq_delay_run_hw_queue+0x192/0x280
> [ 34.460823] blk_mq_run_hw_queue+0x10b/0x1b0
> [ 34.463240] blk_mq_sched_insert_request+0x3bd/0x4d0
> [ 34.467342] blk_execute_rq+0xcf/0x140
> [ 34.468483] sg_io+0x2f7/0x730
>
> Can anyone tell me more about the memory allocation layout of the
> various variables here? It looks like struct request is a header in
> front of struct scsi_request? How do struct elevator_queue, struct
> blk_mq_ctx, and struct blk_mq_hw_ctx overlap these?
The scsi_request is a payload item for the block request, it's
located right after the request in memory. These are persistent
allocations, we don't allocate/free them per IO.
blk_mq_ctx are the blk-mq software queues, they are percpu and
allocated when the queue is setup.
blk_mq_hw_ctx is the hardware queue. You probably have just one,
it's allocated when the queue is setup.
struct elevator_queue is allocated when the scheduler is attached
to the queue. This can get freed and allocated if you switch
the scheduler on a queue, otherwise it persists until the queue
is torn down (and the scheduler data is freed).
> Regardless, I'll check for elevator data changing too...
It should not change unless you switch IO schedulers. If you're
using BFQ and not switching, then it won't change.
--
Jens Axboe
