RE: scsi-mq V2

["Elliott, Robert (Server Storage)" <Elliott@hp.com>]

> -----Original Message-----
> From: Benjamin LaHaise [mailto:bcrl@kvack.org]
> Sent: Friday, 11 July, 2014 9:55 AM
> To: Elliott, Robert (Server Storage)
> Cc: Christoph Hellwig; Jeff Moyer; Jens Axboe; dgilbert@interlog.com; James
> Bottomley; Bart Van Assche; linux-scsi@vger.kernel.org; linux-
> kernel@vger.kernel.org
> Subject: Re: scsi-mq V2
...
> Can you try the below totally untested patch instead?  It looks like
> put_reqs_available() is not irq-safe.
> 

With that addition alone, fio still runs into the same problem.

I added the same fix to get_reqs_available, which also accesses 
kcpu->reqs_available, and the test has run for 35 minutes with 
no problem.

Patch applied:

diff --git a/fs/aio.c b/fs/aio.c
index e59bba8..8e85e26 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -830,16 +830,20 @@ void exit_aio(struct mm_struct *mm)
 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
 {
 	struct kioctx_cpu *kcpu;
+	unsigned long flags;
 
 	preempt_disable();
 	kcpu = this_cpu_ptr(ctx->cpu);
 
+	local_irq_save(flags);
 	kcpu->reqs_available += nr;
+
 	while (kcpu->reqs_available >= ctx->req_batch * 2) {
 		kcpu->reqs_available -= ctx->req_batch;
 		atomic_add(ctx->req_batch, &ctx->reqs_available);
 	}
 
+	local_irq_restore(flags);
 	preempt_enable();
 }
 
@@ -847,10 +851,12 @@ static bool get_reqs_available(struct kioctx *ctx)
 {
 	struct kioctx_cpu *kcpu;
 	bool ret = false;
+	unsigned long flags;
 
 	preempt_disable();
 	kcpu = this_cpu_ptr(ctx->cpu);
 
+	local_irq_save(flags);
 	if (!kcpu->reqs_available) {
 		int old, avail = atomic_read(&ctx->reqs_available);
 
@@ -869,6 +875,7 @@ static bool get_reqs_available(struct kioctx *ctx)
 	ret = true;
 	kcpu->reqs_available--;
 out:
+	local_irq_restore(flags);
 	preempt_enable();
 	return ret;
 }

--
I will see if that solves the problem with the scsi-mq-3 tree, or 
at least some of the bisect trees leading up to it.

A few other comments:

1. Those changes boost _raw_spin_lock_irqsave into first place
in perf top:

  6.59%  [kernel]                    [k] _raw_spin_lock_irqsave
  4.37%  [kernel]                    [k] put_compound_page
  2.87%  [scsi_debug]                [k] sdebug_q_cmd_hrt_complete
  2.74%  [kernel]                    [k] _raw_spin_lock
  2.73%  [kernel]                    [k] apic_timer_interrupt
  2.41%  [kernel]                    [k] do_blockdev_direct_IO
  2.24%  [kernel]                    [k] __get_page_tail
  1.97%  [kernel]                    [k] _raw_spin_unlock_irqrestore
  1.87%  [kernel]                    [k] scsi_queue_rq
  1.76%  [scsi_debug]                [k] schedule_resp

Maybe (later) kcpu->reqs_available should converted to an atomic,
like ctx->reqs_available, to reduce that overhead?

2. After the f8567a3 patch, aio_complete has one early return that 
bypasses the call to put_reqs_available.  Is that OK, or does
that mean that sync iocbs will now eat up reqs_available?

        /*
         * Special case handling for sync iocbs:
         *  - events go directly into the iocb for fast handling
         *  - the sync task with the iocb in its stack holds the single iocb
         *    ref, no other paths have a way to get another ref
         *  - the sync task helpfully left a reference to itself in the iocb
         */
        if (is_sync_kiocb(iocb)) {
                iocb->ki_user_data = res;
                smp_wmb();
                iocb->ki_ctx = ERR_PTR(-EXDEV);
                wake_up_process(iocb->ki_obj.tsk);
                return;
        }


3. The f8567a3 patch renders this comment in aio.c out of date - it's 
no longer incremented when pulled off the ringbuffer, but is now 
incremented when aio_complete is called.

        struct {
                /*
                 * This counts the number of available slots in the ringbuffer,
                 * so we avoid overflowing it: it's decremented (if positive)
                 * when allocating a kiocb and incremented when the resulting
                 * io_event is pulled off the ringbuffer.
                 *
                 * We batch accesses to it with a percpu version.
                 */
                atomic_t        reqs_available;
        } ____cacheline_aligned_in_smp;


---
Rob Elliott    HP Server Storage