[Intel-gfx] [PATCH 2/4] drm/i915/guc: Move GuC workqueue allocations outside of the mutex

Michał Winiarski michal.winiarski at intel.com
Mon Dec 11 15:12:54 UTC 2017


This gets rid of the following lockdep splat:

======================================================
WARNING: possible circular locking dependency detected
4.15.0-rc2-CI-Patchwork_7428+ #1 Not tainted
------------------------------------------------------
debugfs_test/1351 is trying to acquire lock:
 (&dev->struct_mutex){+.+.}, at: [<000000009d90d1a3>] i915_mutex_lock_interruptible+0x47/0x130 [i915]

but task is already holding lock:
 (&mm->mmap_sem){++++}, at: [<000000005df01c1e>] __do_page_fault+0x106/0x560

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #6 (&mm->mmap_sem){++++}:
       __might_fault+0x63/0x90
       _copy_to_user+0x1e/0x70
       filldir+0x8c/0xf0
       dcache_readdir+0xeb/0x160
       iterate_dir+0xe6/0x150
       SyS_getdents+0xa0/0x130
       entry_SYSCALL_64_fastpath+0x1c/0x89

-> #5 (&sb->s_type->i_mutex_key#5){++++}:
       lockref_get+0x9/0x20

-> #4 ((completion)&req.done){+.+.}:
       wait_for_common+0x54/0x210
       devtmpfs_create_node+0x130/0x150
       device_add+0x5ad/0x5e0
       device_create_groups_vargs+0xd4/0xe0
       device_create+0x35/0x40
       msr_device_create+0x22/0x40
       cpuhp_invoke_callback+0xc5/0xbf0
       cpuhp_thread_fun+0x167/0x210
       smpboot_thread_fn+0x17f/0x270
       kthread+0x173/0x1b0
       ret_from_fork+0x24/0x30

-> #3 (cpuhp_state-up){+.+.}:
       cpuhp_issue_call+0x132/0x1c0
       __cpuhp_setup_state_cpuslocked+0x12f/0x2a0
       __cpuhp_setup_state+0x3a/0x50
       page_writeback_init+0x3a/0x5c
       start_kernel+0x393/0x3e2
       secondary_startup_64+0xa5/0xb0

-> #2 (cpuhp_state_mutex){+.+.}:
       __mutex_lock+0x81/0x9b0
       __cpuhp_setup_state_cpuslocked+0x4b/0x2a0
       __cpuhp_setup_state+0x3a/0x50
       page_alloc_init+0x1f/0x26
       start_kernel+0x139/0x3e2
       secondary_startup_64+0xa5/0xb0

-> #1 (cpu_hotplug_lock.rw_sem){++++}:
       cpus_read_lock+0x34/0xa0
       apply_workqueue_attrs+0xd/0x40
       __alloc_workqueue_key+0x2c7/0x4e1
       intel_guc_submission_init+0x10c/0x650 [i915]
       intel_uc_init_hw+0x29e/0x460 [i915]
       i915_gem_init_hw+0xca/0x290 [i915]
       i915_gem_init+0x115/0x3a0 [i915]
       i915_driver_load+0x9a8/0x16c0 [i915]
       i915_pci_probe+0x2e/0x90 [i915]
       pci_device_probe+0x9c/0x120
       driver_probe_device+0x2a3/0x480
       __driver_attach+0xd9/0xe0
       bus_for_each_dev+0x57/0x90
       bus_add_driver+0x168/0x260
       driver_register+0x52/0xc0
       do_one_initcall+0x39/0x150
       do_init_module+0x56/0x1ef
       load_module+0x231c/0x2d70
       SyS_finit_module+0xa5/0xe0
       entry_SYSCALL_64_fastpath+0x1c/0x89

-> #0 (&dev->struct_mutex){+.+.}:
       lock_acquire+0xaf/0x200
       __mutex_lock+0x81/0x9b0
       i915_mutex_lock_interruptible+0x47/0x130 [i915]
       i915_gem_fault+0x201/0x760 [i915]
       __do_fault+0x15/0x70
       __handle_mm_fault+0x85b/0xe40
       handle_mm_fault+0x14f/0x2f0
       __do_page_fault+0x2d1/0x560
       page_fault+0x22/0x30

other info that might help us debug this:

Chain exists of:
  &dev->struct_mutex --> &sb->s_type->i_mutex_key#5 --> &mm->mmap_sem

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&mm->mmap_sem);
                               lock(&sb->s_type->i_mutex_key#5);
                               lock(&mm->mmap_sem);
  lock(&dev->struct_mutex);

 *** DEADLOCK ***

1 lock held by debugfs_test/1351:
 #0:  (&mm->mmap_sem){++++}, at: [<000000005df01c1e>] __do_page_fault+0x106/0x560

stack backtrace:
CPU: 2 PID: 1351 Comm: debugfs_test Not tainted 4.15.0-rc2-CI-Patchwork_7428+ #1
Hardware name:                  /NUC6i5SYB, BIOS SYSKLi35.86A.0057.2017.0119.1758 01/19/2017
Call Trace:
 dump_stack+0x5f/0x86
 print_circular_bug+0x230/0x3b0
 check_prev_add+0x439/0x7b0
 ? lockdep_init_map_crosslock+0x20/0x20
 ? unwind_get_return_address+0x16/0x30
 ? __lock_acquire+0x1385/0x15a0
 __lock_acquire+0x1385/0x15a0
 lock_acquire+0xaf/0x200
 ? i915_mutex_lock_interruptible+0x47/0x130 [i915]
 __mutex_lock+0x81/0x9b0
 ? i915_mutex_lock_interruptible+0x47/0x130 [i915]
 ? i915_mutex_lock_interruptible+0x47/0x130 [i915]
 ? i915_mutex_lock_interruptible+0x47/0x130 [i915]
 i915_mutex_lock_interruptible+0x47/0x130 [i915]
 ? __pm_runtime_resume+0x4f/0x80
 i915_gem_fault+0x201/0x760 [i915]
 __do_fault+0x15/0x70
 __handle_mm_fault+0x85b/0xe40
 handle_mm_fault+0x14f/0x2f0
 __do_page_fault+0x2d1/0x560
 page_fault+0x22/0x30
RIP: 0033:0x7f98d6f49116
RSP: 002b:00007ffd6ffc3278 EFLAGS: 00010283
RAX: 00007f98d39a2bc0 RBX: 0000000000000000 RCX: 0000000000001680
RDX: 0000000000001680 RSI: 00007ffd6ffc3400 RDI: 00007f98d39a2bc0
RBP: 00007ffd6ffc33a0 R08: 0000000000000000 R09: 00000000000005a0
R10: 000055e847c2a830 R11: 0000000000000002 R12: 0000000000000001
R13: 000055e847c1d040 R14: 00007ffd6ffc3400 R15: 00007f98d6752ba0

Testcase: igt/debugfs_test/read_all_entries
Signed-off-by: Michał Winiarski <michal.winiarski at intel.com>
Cc: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
Cc: Michal Wajdeczko <michal.wajdeczko at intel.com>
---
 drivers/gpu/drm/i915/i915_drv.c             |  1 +
 drivers/gpu/drm/i915/i915_gem.c             |  4 ++
 drivers/gpu/drm/i915/intel_guc.c            | 57 +++++++++++++++++++++++
 drivers/gpu/drm/i915/intel_guc.h            |  2 +
 drivers/gpu/drm/i915/intel_guc_log.c        | 23 ---------
 drivers/gpu/drm/i915/intel_guc_submission.c | 72 ++++++++---------------------
 drivers/gpu/drm/i915/intel_guc_submission.h |  2 +
 drivers/gpu/drm/i915/intel_uc.c             | 26 +++++++++++
 drivers/gpu/drm/i915/intel_uc.h             |  2 +
 9 files changed, 112 insertions(+), 77 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
index 5b1fd5f1defb..53678031ed61 100644
--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c
@@ -621,6 +621,7 @@ static void i915_gem_fini(struct drm_i915_private *dev_priv)
 	i915_gem_contexts_fini(dev_priv);
 	mutex_unlock(&dev_priv->drm.struct_mutex);
 
+	intel_uc_fini_wq(dev_priv);
 	i915_gem_cleanup_userptr(dev_priv);
 
 	i915_gem_drain_freed_objects(dev_priv);
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index fcc9b53864f0..4458a65ea580 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -5151,6 +5151,10 @@ int i915_gem_init(struct drm_i915_private *dev_priv)
 	if (ret)
 		return ret;
 
+	ret = intel_uc_init_wq(dev_priv);
+	if (ret)
+		return ret;
+
 	/* This is just a security blanket to placate dragons.
 	 * On some systems, we very sporadically observe that the first TLBs
 	 * used by the CS may be stale, despite us poking the TLB reset. If
diff --git a/drivers/gpu/drm/i915/intel_guc.c b/drivers/gpu/drm/i915/intel_guc.c
index 92ed22f38fc4..3c6bf5a34c3c 100644
--- a/drivers/gpu/drm/i915/intel_guc.c
+++ b/drivers/gpu/drm/i915/intel_guc.c
@@ -69,6 +69,63 @@ void intel_guc_init_early(struct intel_guc *guc)
 	guc->notify = gen8_guc_raise_irq;
 }
 
+int intel_guc_init_wq(struct intel_guc *guc)
+{
+	struct drm_i915_private *dev_priv = guc_to_i915(guc);
+
+	/*
+	 * GuC log buffer flush work item has to do register access to
+	 * send the ack to GuC and this work item, if not synced before
+	 * suspend, can potentially get executed after the GFX device is
+	 * suspended.
+	 * By marking the WQ as freezable, we don't have to bother about
+	 * flushing of this work item from the suspend hooks, the pending
+	 * work item if any will be either executed before the suspend
+	 * or scheduled later on resume. This way the handling of work
+	 * item can be kept same between system suspend & rpm suspend.
+	 */
+	guc->log.runtime.flush_wq = alloc_ordered_workqueue("i915-guc_log",
+						WQ_HIGHPRI | WQ_FREEZABLE);
+	if (!guc->log.runtime.flush_wq)
+		return -ENOMEM;
+
+	/*
+	 * Even though both sending GuC action, and adding a new workitem to
+	 * GuC workqueue are serialized (each with its own locking), since
+	 * we're using mutliple engines, it's possible that we're going to
+	 * issue a preempt request with two (or more - each for different
+	 * engine) workitems in GuC queue. In this situation, GuC may submit
+	 * all of them, which will make us very confused.
+	 * Our preemption contexts may even already be complete - before we
+	 * even had the chance to sent the preempt action to GuC!. Rather
+	 * than introducing yet another lock, we can just use ordered workqueue
+	 * to make sure we're always sending a single preemption request with a
+	 * single workitem.
+	 */
+	if (HAS_LOGICAL_RING_PREEMPTION(dev_priv) &&
+	    USES_GUC_SUBMISSION(dev_priv)) {
+		guc->preempt_wq = alloc_ordered_workqueue("i915-guc_preempt",
+							  WQ_HIGHPRI);
+		if (!guc->preempt_wq) {
+			destroy_workqueue(guc->log.runtime.flush_wq);
+			return -ENOMEM;
+		}
+	}
+
+	return 0;
+}
+
+void intel_guc_fini_wq(struct intel_guc *guc)
+{
+	struct drm_i915_private *dev_priv = guc_to_i915(guc);
+
+	if (HAS_LOGICAL_RING_PREEMPTION(dev_priv) &&
+	    USES_GUC_SUBMISSION(dev_priv))
+		destroy_workqueue(guc->preempt_wq);
+
+	destroy_workqueue(guc->log.runtime.flush_wq);
+}
+
 static int guc_shared_data_create(struct intel_guc *guc)
 {
 	struct i915_vma *vma;
diff --git a/drivers/gpu/drm/i915/intel_guc.h b/drivers/gpu/drm/i915/intel_guc.h
index 81659e223e11..52856a97477d 100644
--- a/drivers/gpu/drm/i915/intel_guc.h
+++ b/drivers/gpu/drm/i915/intel_guc.h
@@ -119,6 +119,8 @@ static inline u32 guc_ggtt_offset(struct i915_vma *vma)
 void intel_guc_init_early(struct intel_guc *guc);
 void intel_guc_init_send_regs(struct intel_guc *guc);
 void intel_guc_init_params(struct intel_guc *guc);
+int intel_guc_init_wq(struct intel_guc *guc);
+void intel_guc_fini_wq(struct intel_guc *guc);
 int intel_guc_init(struct intel_guc *guc);
 void intel_guc_fini(struct intel_guc *guc);
 int intel_guc_send_nop(struct intel_guc *guc, const u32 *action, u32 len);
diff --git a/drivers/gpu/drm/i915/intel_guc_log.c b/drivers/gpu/drm/i915/intel_guc_log.c
index 1a2c5eed9929..eaedd63e3819 100644
--- a/drivers/gpu/drm/i915/intel_guc_log.c
+++ b/drivers/gpu/drm/i915/intel_guc_log.c
@@ -411,30 +411,8 @@ static int guc_log_runtime_create(struct intel_guc *guc)
 	guc->log.runtime.relay_chan = guc_log_relay_chan;
 
 	INIT_WORK(&guc->log.runtime.flush_work, capture_logs_work);
-
-	/*
-	 * GuC log buffer flush work item has to do register access to
-	 * send the ack to GuC and this work item, if not synced before
-	 * suspend, can potentially get executed after the GFX device is
-	 * suspended.
-	 * By marking the WQ as freezable, we don't have to bother about
-	 * flushing of this work item from the suspend hooks, the pending
-	 * work item if any will be either executed before the suspend
-	 * or scheduled later on resume. This way the handling of work
-	 * item can be kept same between system suspend & rpm suspend.
-	 */
-	guc->log.runtime.flush_wq = alloc_ordered_workqueue("i915-guc_log",
-						WQ_HIGHPRI | WQ_FREEZABLE);
-	if (!guc->log.runtime.flush_wq) {
-		DRM_ERROR("Couldn't allocate the wq for GuC logging\n");
-		ret = -ENOMEM;
-		goto err_relaychan;
-	}
-
 	return 0;
 
-err_relaychan:
-	relay_close(guc->log.runtime.relay_chan);
 err_vaddr:
 	i915_gem_object_unpin_map(guc->log.vma->obj);
 	guc->log.runtime.buf_addr = NULL;
@@ -450,7 +428,6 @@ static void guc_log_runtime_destroy(struct intel_guc *guc)
 	if (!guc_log_has_runtime(guc))
 		return;
 
-	destroy_workqueue(guc->log.runtime.flush_wq);
 	relay_close(guc->log.runtime.relay_chan);
 	i915_gem_object_unpin_map(guc->log.vma->obj);
 	guc->log.runtime.buf_addr = NULL;
diff --git a/drivers/gpu/drm/i915/intel_guc_submission.c b/drivers/gpu/drm/i915/intel_guc_submission.c
index c020560c395e..1c0d0b09ac9e 100644
--- a/drivers/gpu/drm/i915/intel_guc_submission.c
+++ b/drivers/gpu/drm/i915/intel_guc_submission.c
@@ -1187,57 +1187,15 @@ static void guc_ads_destroy(struct intel_guc *guc)
 	i915_vma_unpin_and_release(&guc->ads_vma);
 }
 
-static int guc_preempt_work_create(struct intel_guc *guc)
-{
-	struct drm_i915_private *dev_priv = guc_to_i915(guc);
-	struct intel_engine_cs *engine;
-	enum intel_engine_id id;
-
-	/*
-	 * Even though both sending GuC action, and adding a new workitem to
-	 * GuC workqueue are serialized (each with its own locking), since
-	 * we're using mutliple engines, it's possible that we're going to
-	 * issue a preempt request with two (or more - each for different
-	 * engine) workitems in GuC queue. In this situation, GuC may submit
-	 * all of them, which will make us very confused.
-	 * Our preemption contexts may even already be complete - before we
-	 * even had the chance to sent the preempt action to GuC!. Rather
-	 * than introducing yet another lock, we can just use ordered workqueue
-	 * to make sure we're always sending a single preemption request with a
-	 * single workitem.
-	 */
-	guc->preempt_wq = alloc_ordered_workqueue("i915-guc_preempt",
-						  WQ_HIGHPRI);
-	if (!guc->preempt_wq)
-		return -ENOMEM;
-
-	for_each_engine(engine, dev_priv, id) {
-		guc->preempt_work[id].engine = engine;
-		INIT_WORK(&guc->preempt_work[id].work, inject_preempt_context);
-	}
-
-	return 0;
-}
-
-static void guc_preempt_work_destroy(struct intel_guc *guc)
-{
-	struct drm_i915_private *dev_priv = guc_to_i915(guc);
-	struct intel_engine_cs *engine;
-	enum intel_engine_id id;
-
-	for_each_engine(engine, dev_priv, id)
-		cancel_work_sync(&guc->preempt_work[id].work);
-
-	destroy_workqueue(guc->preempt_wq);
-	guc->preempt_wq = NULL;
-}
-
 /*
  * Set up the memory resources to be shared with the GuC (via the GGTT)
  * at firmware loading time.
  */
 int intel_guc_submission_init(struct intel_guc *guc)
 {
+	struct drm_i915_private *dev_priv = guc_to_i915(guc);
+	struct intel_engine_cs *engine;
+	enum intel_engine_id id;
 	int ret;
 
 	if (guc->stage_desc_pool)
@@ -1256,20 +1214,20 @@ int intel_guc_submission_init(struct intel_guc *guc)
 	if (ret < 0)
 		goto err_stage_desc_pool;
 
-	ret = guc_preempt_work_create(guc);
-	if (ret)
-		goto err_log;
-	GEM_BUG_ON(!guc->preempt_wq);
-
 	ret = guc_ads_create(guc);
 	if (ret < 0)
-		goto err_wq;
+		goto err_log;
 	GEM_BUG_ON(!guc->ads_vma);
 
+	if (HAS_LOGICAL_RING_PREEMPTION(dev_priv)) {
+		for_each_engine(engine, dev_priv, id) {
+			guc->preempt_work[id].engine = engine;
+			INIT_WORK(&guc->preempt_work[id].work, inject_preempt_context);
+		}
+	}
+
 	return 0;
 
-err_wq:
-	guc_preempt_work_destroy(guc);
 err_log:
 	intel_guc_log_destroy(guc);
 err_stage_desc_pool:
@@ -1279,8 +1237,14 @@ int intel_guc_submission_init(struct intel_guc *guc)
 
 void intel_guc_submission_fini(struct intel_guc *guc)
 {
+	struct drm_i915_private *dev_priv = guc_to_i915(guc);
+	struct intel_engine_cs *engine;
+	enum intel_engine_id id;
+
+	for_each_engine(engine, dev_priv, id)
+		cancel_work_sync(&guc->preempt_work[id].work);
+
 	guc_ads_destroy(guc);
-	guc_preempt_work_destroy(guc);
 	intel_guc_log_destroy(guc);
 	guc_stage_desc_pool_destroy(guc);
 }
diff --git a/drivers/gpu/drm/i915/intel_guc_submission.h b/drivers/gpu/drm/i915/intel_guc_submission.h
index 021fe85c8f71..fb081cefef93 100644
--- a/drivers/gpu/drm/i915/intel_guc_submission.h
+++ b/drivers/gpu/drm/i915/intel_guc_submission.h
@@ -77,5 +77,7 @@ int intel_guc_submission_init(struct intel_guc *guc);
 int intel_guc_submission_enable(struct intel_guc *guc);
 void intel_guc_submission_disable(struct intel_guc *guc);
 void intel_guc_submission_fini(struct intel_guc *guc);
+int intel_guc_preempt_work_create(struct intel_guc *guc);
+void intel_guc_preempt_work_destroy(struct intel_guc *guc);
 
 #endif
diff --git a/drivers/gpu/drm/i915/intel_uc.c b/drivers/gpu/drm/i915/intel_uc.c
index 3040a0e00142..785850838a44 100644
--- a/drivers/gpu/drm/i915/intel_uc.c
+++ b/drivers/gpu/drm/i915/intel_uc.c
@@ -188,6 +188,32 @@ static void guc_disable_communication(struct intel_guc *guc)
 	guc->send = intel_guc_send_nop;
 }
 
+int intel_uc_init_wq(struct drm_i915_private *dev_priv)
+{
+	int ret;
+
+	if (!USES_GUC(dev_priv))
+		return 0;
+
+	ret = intel_guc_init_wq(&dev_priv->guc);
+	if (ret) {
+		DRM_ERROR("Couldn't allocate workqueues for GuC\n");
+		return ret;
+	}
+
+	return 0;
+}
+
+void intel_uc_fini_wq(struct drm_i915_private *dev_priv)
+{
+	if (!USES_GUC(dev_priv))
+		return;
+
+	GEM_BUG_ON(!HAS_GUC(dev_priv));
+
+	intel_guc_fini_wq(&dev_priv->guc);
+}
+
 int intel_uc_init_hw(struct drm_i915_private *dev_priv)
 {
 	struct intel_guc *guc = &dev_priv->guc;
diff --git a/drivers/gpu/drm/i915/intel_uc.h b/drivers/gpu/drm/i915/intel_uc.h
index 7a59e2486e9e..53edfeaf56b0 100644
--- a/drivers/gpu/drm/i915/intel_uc.h
+++ b/drivers/gpu/drm/i915/intel_uc.h
@@ -33,6 +33,8 @@ void intel_uc_init_early(struct drm_i915_private *dev_priv);
 void intel_uc_init_mmio(struct drm_i915_private *dev_priv);
 void intel_uc_init_fw(struct drm_i915_private *dev_priv);
 void intel_uc_fini_fw(struct drm_i915_private *dev_priv);
+int intel_uc_init_wq(struct drm_i915_private *dev_priv);
+void intel_uc_fini_wq(struct drm_i915_private *dev_priv);
 int intel_uc_init_hw(struct drm_i915_private *dev_priv);
 void intel_uc_fini_hw(struct drm_i915_private *dev_priv);
 
-- 
2.14.3



More information about the Intel-gfx mailing list