drm/i915: Move execlists irq handler to a bottom half
authorTvrtko Ursulin <tvrtko.ursulin@intel.com>
Mon, 4 Apr 2016 11:11:56 +0000 (12:11 +0100)
committerTvrtko Ursulin <tvrtko.ursulin@intel.com>
Mon, 4 Apr 2016 13:08:52 +0000 (14:08 +0100)
Doing a lot of work in the interrupt handler introduces huge
latencies to the system as a whole.

Most dramatic effect can be seen by running an all engine
stress test like igt/gem_exec_nop/all where, when the kernel
config is lean enough, the whole system can be brought into
multi-second periods of complete non-interactivty. That can
look for example like this:

 NMI watchdog: BUG: soft lockup - CPU#0 stuck for 23s! [kworker/u8:3:143]
 Modules linked in: [redacted for brevity]
 CPU: 0 PID: 143 Comm: kworker/u8:3 Tainted: G     U       L  4.5.0-160321+ #183
 Hardware name: Intel Corporation Broadwell Client platform/WhiteTip Mountain 1
 Workqueue: i915 gen6_pm_rps_work [i915]
 task: ffff8800aae88000 ti: ffff8800aae90000 task.ti: ffff8800aae90000
 RIP: 0010:[<ffffffff8104a3c2>]  [<ffffffff8104a3c2>] __do_softirq+0x72/0x1d0
 RSP: 0000:ffff88014f403f38  EFLAGS: 00000206
 RAX: ffff8800aae94000 RBX: 0000000000000000 RCX: 00000000000006e0
 RDX: 0000000000000020 RSI: 0000000004208060 RDI: 0000000000215d80
 RBP: ffff88014f403f80 R08: 0000000b1b42c180 R09: 0000000000000022
 R10: 0000000000000004 R11: 00000000ffffffff R12: 000000000000a030
 R13: 0000000000000082 R14: ffff8800aa4d0080 R15: 0000000000000082
 FS:  0000000000000000(0000) GS:ffff88014f400000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 00007fa53b90c000 CR3: 0000000001a0a000 CR4: 00000000001406f0
 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
 Stack:
  042080601b33869f ffff8800aae94000 00000000fffc2678 ffff88010000000a
  0000000000000000 000000000000a030 0000000000005302 ffff8800aa4d0080
  0000000000000206 ffff88014f403f90 ffffffff8104a716 ffff88014f403fa8
 Call Trace:
  <IRQ>
  [<ffffffff8104a716>] irq_exit+0x86/0x90
  [<ffffffff81031e7d>] smp_apic_timer_interrupt+0x3d/0x50
  [<ffffffff814f3eac>] apic_timer_interrupt+0x7c/0x90
  <EOI>
  [<ffffffffa01c5b40>] ? gen8_write64+0x1a0/0x1a0 [i915]
  [<ffffffff814f2b39>] ? _raw_spin_unlock_irqrestore+0x9/0x20
  [<ffffffffa01c5c44>] gen8_write32+0x104/0x1a0 [i915]
  [<ffffffff8132c6a2>] ? n_tty_receive_buf_common+0x372/0xae0
  [<ffffffffa017cc9e>] gen6_set_rps_thresholds+0x1be/0x330 [i915]
  [<ffffffffa017eaf0>] gen6_set_rps+0x70/0x200 [i915]
  [<ffffffffa0185375>] intel_set_rps+0x25/0x30 [i915]
  [<ffffffffa01768fd>] gen6_pm_rps_work+0x10d/0x2e0 [i915]
  [<ffffffff81063852>] ? finish_task_switch+0x72/0x1c0
  [<ffffffff8105ab29>] process_one_work+0x139/0x350
  [<ffffffff8105b186>] worker_thread+0x126/0x490
  [<ffffffff8105b060>] ? rescuer_thread+0x320/0x320
  [<ffffffff8105fa64>] kthread+0xc4/0xe0
  [<ffffffff8105f9a0>] ? kthread_create_on_node+0x170/0x170
  [<ffffffff814f351f>] ret_from_fork+0x3f/0x70
  [<ffffffff8105f9a0>] ? kthread_create_on_node+0x170/0x170

I could not explain, or find a code path, which would explain
a +20 second lockup, but from some instrumentation it was
apparent the interrupts off proportion of time was between
10-25% under heavy load which is quite bad.

When a interrupt "cliff" is reached, which was >~320k irq/s on
my machine, the whole system goes into a terrible state of the
above described multi-second lockups.

By moving the GT interrupt handling to a tasklet in a most
simple way, the problem above disappears completely.

Testing the effect on sytem-wide latencies using
igt/gem_syslatency shows the following before this patch:

gem_syslatency: cycles=1532739, latency mean=416531.829us max=2499237us
gem_syslatency: cycles=1839434, latency mean=1458099.157us max=4998944us
gem_syslatency: cycles=1432570, latency mean=2688.451us max=1201185us
gem_syslatency: cycles=1533543, latency mean=416520.499us max=2498886us

This shows that the unrelated process is experiencing huge
delays in its wake-up latency. After the patch the results
look like this:

gem_syslatency: cycles=808907, latency mean=53.133us max=1640us
gem_syslatency: cycles=862154, latency mean=62.778us max=2117us
gem_syslatency: cycles=856039, latency mean=58.079us max=2123us
gem_syslatency: cycles=841683, latency mean=56.914us max=1667us

Showing a huge improvement in the unrelated process wake-up
latency. It also shows an approximate halving in the number
of total empty batches submitted during the test. This may
not be worrying since the test puts the driver under
a very unrealistic load with ncpu threads doing empty batch
submission to all GPU engines each.

Another benefit compared to the hard-irq handling is that now
work on all engines can be dispatched in parallel since we can
have up to number of CPUs active tasklets. (While previously
a single hard-irq would serially dispatch on one engine after
another.)

More interesting scenario with regards to throughput is
"gem_latency -n 100" which  shows 25% better throughput and
CPU usage, and 14% better dispatch latencies.

I did not find any gains or regressions with Synmark2 or
GLbench under light testing. More benchmarking is certainly
required.

v2:
   * execlists_lock should be taken as spin_lock_bh when
     queuing work from userspace now. (Chris Wilson)
   * uncore.lock must be taken with spin_lock_irq when
     submitting requests since that now runs from either
     softirq or process context.

v3:
   * Expanded commit message with more testing data;
   * converted missed locking sites to _bh;
   * added execlist_lock comment. (Chris Wilson)

v4:
   * Mention dispatch parallelism in commit. (Chris Wilson)
   * Do not hold uncore.lock over MMIO reads since the block
     is already serialised per-engine via the tasklet itself.
     (Chris Wilson)
   * intel_lrc_irq_handler should be static. (Chris Wilson)
   * Cancel/sync the tasklet on GPU reset. (Chris Wilson)
   * Document and WARN that tasklet cannot be active/pending
     on engine cleanup. (Chris Wilson/Imre Deak)

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Imre Deak <imre.deak@intel.com>
Testcase: igt/gem_exec_nop/all
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94350
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1459768316-6670-1-git-send-email-tvrtko.ursulin@linux.intel.com
drivers/gpu/drm/i915/i915_debugfs.c
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_irq.c
drivers/gpu/drm/i915/intel_lrc.c
drivers/gpu/drm/i915/intel_lrc.h
drivers/gpu/drm/i915/intel_ringbuffer.h

index 0b25228c202e2ff593257d9fdabc72a0a56c8040..a2e3af02c292d025e93a8ba88f5dd656da396816 100644 (file)
@@ -2100,7 +2100,6 @@ static int i915_execlists(struct seq_file *m, void *data)
        for_each_engine(engine, dev_priv) {
                struct drm_i915_gem_request *head_req = NULL;
                int count = 0;
-               unsigned long flags;
 
                seq_printf(m, "%s\n", engine->name);
 
@@ -2127,13 +2126,13 @@ static int i915_execlists(struct seq_file *m, void *data)
                                   i, status, ctx_id);
                }
 
-               spin_lock_irqsave(&engine->execlist_lock, flags);
+               spin_lock_bh(&engine->execlist_lock);
                list_for_each(cursor, &engine->execlist_queue)
                        count++;
                head_req = list_first_entry_or_null(&engine->execlist_queue,
                                                    struct drm_i915_gem_request,
                                                    execlist_link);
-               spin_unlock_irqrestore(&engine->execlist_lock, flags);
+               spin_unlock_bh(&engine->execlist_lock);
 
                seq_printf(m, "\t%d requests in queue\n", count);
                if (head_req) {
index ca96fc12cdf43322663dded885c86744247937e3..40f90c7e718a5d08f0de3f50991a83d811a78895 100644 (file)
@@ -2842,13 +2842,15 @@ static void i915_gem_reset_engine_cleanup(struct drm_i915_private *dev_priv,
         */
 
        if (i915.enable_execlists) {
-               spin_lock_irq(&engine->execlist_lock);
+               /* Ensure irq handler finishes or is cancelled. */
+               tasklet_kill(&engine->irq_tasklet);
 
+               spin_lock_bh(&engine->execlist_lock);
                /* list_splice_tail_init checks for empty lists */
                list_splice_tail_init(&engine->execlist_queue,
                                      &engine->execlist_retired_req_list);
+               spin_unlock_bh(&engine->execlist_lock);
 
-               spin_unlock_irq(&engine->execlist_lock);
                intel_execlists_retire_requests(engine);
        }
 
@@ -2968,9 +2970,9 @@ i915_gem_retire_requests(struct drm_device *dev)
                i915_gem_retire_requests_ring(engine);
                idle &= list_empty(&engine->request_list);
                if (i915.enable_execlists) {
-                       spin_lock_irq(&engine->execlist_lock);
+                       spin_lock_bh(&engine->execlist_lock);
                        idle &= list_empty(&engine->execlist_queue);
-                       spin_unlock_irq(&engine->execlist_lock);
+                       spin_unlock_bh(&engine->execlist_lock);
 
                        intel_execlists_retire_requests(engine);
                }
index a9c18137e3f5de35ef76eaa711f8d26f3a534d69..c85eb8dec2dce01a6c23d3f8c79491a9d3b14801 100644 (file)
@@ -1323,7 +1323,7 @@ gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir, int test_shift)
        if (iir & (GT_RENDER_USER_INTERRUPT << test_shift))
                notify_ring(engine);
        if (iir & (GT_CONTEXT_SWITCH_INTERRUPT << test_shift))
-               intel_lrc_irq_handler(engine);
+               tasklet_schedule(&engine->irq_tasklet);
 }
 
 static irqreturn_t gen8_gt_irq_handler(struct drm_i915_private *dev_priv,
index 5d4ca3b11ae2b10d2a07305755bb6ff94fdb04bb..a1db6a02cf23592d72d085109c42585f362fbd45 100644 (file)
  * preemption, but just sampling the new tail pointer).
  *
  */
+#include <linux/interrupt.h>
 
 #include <drm/drmP.h>
 #include <drm/i915_drm.h>
@@ -418,20 +419,18 @@ static void execlists_submit_requests(struct drm_i915_gem_request *rq0,
 {
        struct drm_i915_private *dev_priv = rq0->i915;
 
-       /* BUG_ON(!irqs_disabled());  */
-
        execlists_update_context(rq0);
 
        if (rq1)
                execlists_update_context(rq1);
 
-       spin_lock(&dev_priv->uncore.lock);
+       spin_lock_irq(&dev_priv->uncore.lock);
        intel_uncore_forcewake_get__locked(dev_priv, FORCEWAKE_ALL);
 
        execlists_elsp_write(rq0, rq1);
 
        intel_uncore_forcewake_put__locked(dev_priv, FORCEWAKE_ALL);
-       spin_unlock(&dev_priv->uncore.lock);
+       spin_unlock_irq(&dev_priv->uncore.lock);
 }
 
 static void execlists_context_unqueue(struct intel_engine_cs *engine)
@@ -538,13 +537,14 @@ get_context_status(struct intel_engine_cs *engine, unsigned int read_pointer,
 
 /**
  * intel_lrc_irq_handler() - handle Context Switch interrupts
- * @ring: Engine Command Streamer to handle.
+ * @engine: Engine Command Streamer to handle.
  *
  * Check the unread Context Status Buffers and manage the submission of new
  * contexts to the ELSP accordingly.
  */
-void intel_lrc_irq_handler(struct intel_engine_cs *engine)
+static void intel_lrc_irq_handler(unsigned long data)
 {
+       struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
        struct drm_i915_private *dev_priv = engine->dev->dev_private;
        u32 status_pointer;
        unsigned int read_pointer, write_pointer;
@@ -552,8 +552,7 @@ void intel_lrc_irq_handler(struct intel_engine_cs *engine)
        unsigned int csb_read = 0, i;
        unsigned int submit_contexts = 0;
 
-       spin_lock(&dev_priv->uncore.lock);
-       intel_uncore_forcewake_get__locked(dev_priv, FORCEWAKE_ALL);
+       intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
 
        status_pointer = I915_READ_FW(RING_CONTEXT_STATUS_PTR(engine));
 
@@ -578,8 +577,7 @@ void intel_lrc_irq_handler(struct intel_engine_cs *engine)
                      _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK,
                                    engine->next_context_status_buffer << 8));
 
-       intel_uncore_forcewake_put__locked(dev_priv, FORCEWAKE_ALL);
-       spin_unlock(&dev_priv->uncore.lock);
+       intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
 
        spin_lock(&engine->execlist_lock);
 
@@ -621,7 +619,7 @@ static void execlists_context_queue(struct drm_i915_gem_request *request)
 
        i915_gem_request_reference(request);
 
-       spin_lock_irq(&engine->execlist_lock);
+       spin_lock_bh(&engine->execlist_lock);
 
        list_for_each_entry(cursor, &engine->execlist_queue, execlist_link)
                if (++num_elements > 2)
@@ -646,7 +644,7 @@ static void execlists_context_queue(struct drm_i915_gem_request *request)
        if (num_elements == 0)
                execlists_context_unqueue(engine);
 
-       spin_unlock_irq(&engine->execlist_lock);
+       spin_unlock_bh(&engine->execlist_lock);
 }
 
 static int logical_ring_invalidate_all_caches(struct drm_i915_gem_request *req)
@@ -1033,9 +1031,9 @@ void intel_execlists_retire_requests(struct intel_engine_cs *engine)
                return;
 
        INIT_LIST_HEAD(&retired_list);
-       spin_lock_irq(&engine->execlist_lock);
+       spin_lock_bh(&engine->execlist_lock);
        list_replace_init(&engine->execlist_retired_req_list, &retired_list);
-       spin_unlock_irq(&engine->execlist_lock);
+       spin_unlock_bh(&engine->execlist_lock);
 
        list_for_each_entry_safe(req, tmp, &retired_list, execlist_link) {
                struct intel_context *ctx = req->ctx;
@@ -2016,6 +2014,13 @@ void intel_logical_ring_cleanup(struct intel_engine_cs *engine)
        if (!intel_engine_initialized(engine))
                return;
 
+       /*
+        * Tasklet cannot be active at this point due intel_mark_active/idle
+        * so this is just for documentation.
+        */
+       if (WARN_ON(test_bit(TASKLET_STATE_SCHED, &engine->irq_tasklet.state)))
+               tasklet_kill(&engine->irq_tasklet);
+
        dev_priv = engine->dev->dev_private;
 
        if (engine->buffer) {
@@ -2089,6 +2094,9 @@ logical_ring_init(struct drm_device *dev, struct intel_engine_cs *engine)
        INIT_LIST_HEAD(&engine->execlist_retired_req_list);
        spin_lock_init(&engine->execlist_lock);
 
+       tasklet_init(&engine->irq_tasklet,
+                    intel_lrc_irq_handler, (unsigned long)engine);
+
        logical_ring_init_platform_invariants(engine);
 
        ret = i915_cmd_parser_init_ring(engine);
index a17cb12221bad0af70ff1474701d8e99e0001b46..0b0853eee91e1325d2f86167a17b4ab63217a09f 100644 (file)
@@ -118,7 +118,6 @@ int intel_execlists_submission(struct i915_execbuffer_params *params,
                               struct drm_i915_gem_execbuffer2 *args,
                               struct list_head *vmas);
 
-void intel_lrc_irq_handler(struct intel_engine_cs *engine);
 void intel_execlists_retire_requests(struct intel_engine_cs *engine);
 
 #endif /* _INTEL_LRC_H_ */
index 221a94627aab5d09b42a8a990d016f08d42d8e3b..18074ab55f61556f72443b4c517fbc45236e2d1b 100644 (file)
@@ -266,7 +266,8 @@ struct  intel_engine_cs {
        } semaphore;
 
        /* Execlists */
-       spinlock_t execlist_lock;
+       struct tasklet_struct irq_tasklet;
+       spinlock_t execlist_lock; /* used inside tasklet, use spin_lock_bh */
        struct list_head execlist_queue;
        struct list_head execlist_retired_req_list;
        unsigned int next_context_status_buffer;