Improve block layer scheduling fairness

Recent tests revealed that on kernels with CONFIG_PREEMPT disabled and with kernel debugging options like CONFIG_PROVE_LOCKING enabled that block/SCSI/dm workloads with a high queue depth cause soft lockup complaints to appear. Avoid this by voluntarily giving up the CPU in time for such workloads. The patches in this series are:

0001-kernel-kthread.c-Avoid-CPU-lockups.patch 0002-block-Limit-work-processed-in-softirq-context.patch 0003-Avoid-that-I-O-completion-processing-triggers-lockup.patch

Avoid that complaints similar to the one below are reported against a debug kernel:

NMI watchdog: BUG: soft lockup - CPU#1 stuck for 22s! [kdmwork-25 4:2:23313]

irq event stamp: 16320042
hardirqs last  enabled at (16320041): [<ffffffff815992c7>] _raw_spin_unlock_irq+0x27/0x40
hardirqs last disabled at (16320042): [<ffff8803ffbe3cd8>] 0xffff8803ffbe3cd8
softirqs last  enabled at (16319960): [<ffffffff8105bcdb>] __do_softirq+0x1cb/0x230
softirqs last disabled at (16319715): [<ffffffff8105bec8>] irq_exit+0xa8/0xb0
CPU: 1 PID: 23313 Comm: kdmwork-254:2
RIP: 0010:[<ffffffff815992cf>]  [<ffffffff815992cf>] _raw_spin_unlock_irq+0x2f/0x40
Call Trace:
 [<ffffffff813f702f>] scsi_request_fn+0x11f/0x630
 [<ffffffff812977fe>] __blk_run_queue+0x2e/0x40
 [<ffffffff81297065>] __elv_add_request+0x75/0x1f0
 [<ffffffff8129ba91>] blk_insert_cloned_request+0x101/0x190
 [<ffffffffa0189f1a>] map_request+0x16a/0x1b0 [dm_mod]
 [<ffffffffa0189f7d>] map_tio_request+0x1d/0x40 [dm_mod]
 [<ffffffff81077812>] kthread_worker_fn+0x82/0x1a0
 [<ffffffff8107771a>] kthread+0xea/0x100
 [<ffffffff81599b72>] ret_from_fork+0x22/0x40

---
 kernel/kthread.c | 1 +
 1 file changed, 1 insertion(+)

index 9ff173d..516ca6b 100644
@@ -593,6 +593,7 @@ repeat:
 	if (work) {
 		__set_current_state(TASK_RUNNING);
 		work->func(work);
+		cond_resched_rcu_qs();
 	} else if (!freezing(current))
 		schedule();
 

2.7.3

Avoid that complaints like the one below are reported against a debug kernel:

NMI watchdog: BUG: soft lockup - CPU#1 stuck for 22s! [disk11_0:2708]

irq event stamp: 17120809
hardirqs last  enabled at (17120808): [<ffffffff81599191>] _raw_spin_unlock_irqrestore+0x31/0x50
hardirqs last disabled at (17120809): [<ffff88046f223bd0>] 0xffff88046f223bd0
softirqs last  enabled at (17120794): [<ffffffffa060aa67>] scst_check_blocked_dev+0x77/0x1c0 [scst]
softirqs last disabled at (17120795): [<ffffffff8159acbc>] do_softirq_own_stack+0x1c/0x30
RIP: 0010:[<ffffffff81599193>]  [<ffffffff81599193>] _raw_spin_unlock_irqrestore+0x33/0x50
Call Trace:
 <IRQ>
 [<ffffffff81171450>] free_debug_processing+0x270/0x3a0
 [<ffffffff8117277a>] __slab_free+0x17a/0x2c0
 [<ffffffff81172a74>] kmem_cache_free+0x1b4/0x1d0
 [<ffffffff8111a6c2>] mempool_free_slab+0x12/0x20
 [<ffffffff8111a846>] mempool_free+0x26/0x80
 [<ffffffff81294cb9>] bio_free+0x49/0x60
 [<ffffffff81294cee>] bio_put+0x1e/0x30
 [<ffffffffa0199d31>] end_clone_bio+0x21/0x70 [dm_mod]
 [<ffffffff81294d52>] bio_endio+0x52/0x60
 [<ffffffff8129aaec>] blk_update_request+0x7c/0x2a0
 [<ffffffff813f4a8e>] scsi_end_request+0x2e/0x1d0
 [<ffffffff813f7674>] scsi_io_completion+0xb4/0x610
 [<ffffffff813ee79a>] scsi_finish_command+0xca/0x120
 [<ffffffff813f6ef0>] scsi_softirq_done+0x120/0x140
 [<ffffffff812a26e6>] blk_done_softirq+0x76/0x90
 [<ffffffff8105bc1f>] __do_softirq+0x10f/0x230
 [<ffffffff8159acbc>] do_softirq_own_stack+0x1c/0x30
 <EOI>

---
 block/blk-softirq.c | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

index 53b1737..d4d2fe4 100644
@@ -21,10 +21,22 @@ static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
 static void blk_done_softirq(struct softirq_action *h)
 {
 	struct list_head *cpu_list, local_list;
+	struct request *e;
+	bool reraise = false;
+	int i = 0;
 
 	local_irq_disable();
 	cpu_list = this_cpu_ptr(&blk_cpu_done);
+	list_for_each_entry(e, cpu_list, ipi_list) {
+		if (++i < 256)
+			continue;
+		list_cut_position(&local_list, cpu_list, &e->ipi_list);
+		reraise = true;
+		goto enable_irq;
+	}
 	list_replace_init(cpu_list, &local_list);
+
+enable_irq:
 	local_irq_enable();
 
 	while (!list_empty(&local_list)) {

@@ -34,6 +46,14 @@ static void blk_done_softirq(struct softirq_action *h)
 		list_del_init(&rq->ipi_list);
 		rq->q->softirq_done_fn(rq);
 	}
+
+	if (!reraise)
+		return;
+
+	local_irq_disable();
+	if (!list_empty(cpu_list))
+		raise_softirq_irqoff(BLOCK_SOFTIRQ);
+	local_irq_enable();
 }
 
 #ifdef CONFIG_SMP

2.7.3

Avoid that I/O completion processing triggers the following complaints if kernel debug options that slow down the kernel significantly are enabled:

NMI watchdog: BUG: soft lockup - CPU#3 stuck for 22s! [kdmwork-254:2:358]

irq event stamp: 15233868
hardirqs last  enabled at (15233867): [<ffffffff81578167>] _raw_spin_unlock_irq+0x27/0x40
hardirqs last disabled at (15233868): [<ffffffff81579664>] apic_timer_interrupt+0x84/0x90
softirqs last  enabled at (15233850): [<ffffffff8105a55b>] __do_softirq+0x1cb/0x230
softirqs last disabled at (15233743): [<ffffffff8105a748>] irq_exit+0xa8/0xb0
CPU: 3 PID: 358 Comm: kdmwork-254:2
RIP: 0010:[<ffffffff8157816f>]  [<ffffffff8157816f>] _raw_spin_unlock_irq+0x2f/0x40
Call Trace:
 [<ffffffff813dbe28>] scsi_request_fn+0x118/0x600
 [<ffffffff8128190e>] __blk_run_queue+0x2e/0x40
 [<ffffffff81281175>] __elv_add_request+0x75/0x1f0
 [<ffffffff81285b11>] blk_insert_cloned_request+0x101/0x190
 [<ffffffffa0170eee>] map_request+0x18e/0x210 [dm_mod]
 [<ffffffffa0170f8d>] map_tio_request+0x1d/0x40 [dm_mod]
 [<ffffffff81075a0d>] kthread_worker_fn+0x7d/0x1a0
 [<ffffffff8107591a>] kthread+0xea/0x100
 [<ffffffff81578c7f>] ret_from_fork+0x3f/0x70

INFO: rcu_sched self-detected stall on CPU
 3-...: (6497 ticks this GP) idle=fb9/140000000000002/0 softirq=2044956/2045037 fqs=5414
  (t=6500 jiffies g=219289 c=219288 q=7233211)
Task dump for CPU 3:
kdmwork-254:2   R  running task        0   358      2 0x00000008
Call Trace:
 <IRQ>  [<ffffffff8108195f>] sched_show_task+0xbf/0x150
 [<ffffffff81084742>] dump_cpu_task+0x32/0x40
 [<ffffffff810b5e79>] rcu_dump_cpu_stacks+0x89/0xe0
 [<ffffffff810b9999>] rcu_check_callbacks+0x439/0x730
 [<ffffffff810bc1c4>] update_process_times+0x34/0x60
 [<ffffffff810caaa0>] tick_sched_handle.isra.18+0x20/0x50
 [<ffffffff810cb148>] tick_sched_timer+0x38/0x70
 [<ffffffff810bc865>] __hrtimer_run_queues+0xa5/0x1c0
 [<ffffffff810bcef6>] hrtimer_interrupt+0xa6/0x1b0
 [<ffffffff81038ba3>] smp_trace_apic_timer_interrupt+0x63/0x90
 [<ffffffff81038bd9>] smp_apic_timer_interrupt+0x9/0x10
 [<ffffffff81579669>] apic_timer_interrupt+0x89/0x90
 [<ffffffff811640d6>] __slab_free+0xc6/0x270
 [<ffffffff811643d9>] kmem_cache_free+0x159/0x160
 [<ffffffff811d0a12>] kiocb_free+0x32/0x40
 [<ffffffff811d1b75>] aio_complete+0x1e5/0x3c0
 [<ffffffff811bef45>] dio_complete+0x75/0x1d0
 [<ffffffff811bf11a>] dio_bio_end_aio+0x7a/0x130
 [<ffffffff812800da>] bio_endio+0x3a/0x60
 [<ffffffff81284b7c>] blk_update_request+0x7c/0x2a0
 [<ffffffffa016fcc1>] end_clone_bio+0x41/0x70 [dm_mod]
 [<ffffffff812800da>] bio_endio+0x3a/0x60
 [<ffffffff81284b7c>] blk_update_request+0x7c/0x2a0
 [<ffffffff813da01e>] scsi_end_request+0x2e/0x1d0
 [<ffffffff813dc444>] scsi_io_completion+0xb4/0x610
 [<ffffffff813d3e0a>] scsi_finish_command+0xca/0x120
 [<ffffffff813dbcf0>] scsi_softirq_done+0x120/0x140
 [<ffffffff8128c742>] blk_done_softirq+0x72/0x90
 [<ffffffff8105a49f>] __do_softirq+0x10f/0x230
 [<ffffffff8105a748>] irq_exit+0xa8/0xb0
 [<ffffffff810065a5>] do_IRQ+0x65/0x110
 [<ffffffff815793c9>] common_interrupt+0x89/0x89
 <EOI>
 [<ffffffffa010f815>] __multipath_map.isra.16+0x145/0x260 [dm_multipath]
 [<ffffffffa010f962>] multipath_map+0x12/0x20 [dm_multipath]
 [<ffffffffa0170da3>] map_request+0x43/0x210 [dm_mod]
 [<ffffffffa0170f8d>] map_tio_request+0x1d/0x40 [dm_mod]
 [<ffffffff81075a0d>] kthread_worker_fn+0x7d/0x1a0
 [<ffffffff8107591a>] kthread+0xea/0x100
 [<ffffffff81578c7f>] ret_from_fork+0x3f/0x70

---
 drivers/scsi/scsi_lib.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

index 8106515..3bf6c33 100644
@@ -1770,13 +1770,14 @@ static void scsi_request_fn(struct request_queue *q)
 	struct Scsi_Host *shost;
 	struct scsi_cmnd *cmd;
 	struct request *req;
+	int i;
 
 	/*
-	 * To start with, we keep looping until the queue is empty, or until
-	 * the host is no longer able to accept any more requests.
+	 * Loop until the queue is empty, until the host is no longer able to
+	 * accept any more requests or until 256 requests have been processed.
 	 */
 	shost = sdev->host;
-	for (;;) {
+	for (i = 256; i > 0; i--) {
 		int rtn;
 		/*
 		 * get next queueable request.  We do this early to make sure

@@ -1861,6 +1862,9 @@ static void scsi_request_fn(struct request_queue *q)
 		spin_lock_irq(q->queue_lock);
 	}
 
+	if (unlikely(i == 0))
+		blk_delay_queue(q, 0);
+
 	return;
 
  host_not_ready:

2.7.3