drm/i915: Preallocate our mmu notifier workequeu to unbreak cpu hotplug deadlock
authorDaniel Vetter <daniel.vetter@ffwll.ch>
Mon, 9 Oct 2017 16:44:00 +0000 (18:44 +0200)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Tue, 10 Oct 2017 10:57:03 +0000 (12:57 +0200)
4.14-rc1 gained the fancy new cross-release support in lockdep, which
seems to have uncovered a few more rules about what is allowed and
isn't.

This one here seems to indicate that allocating a work-queue while
holding mmap_sem is a no-go, so let's try to preallocate it.

Of course another way to break this chain would be somewhere in the
cpu hotplug code, since this isn't the only trace we're finding now
which goes through msr_create_device.

Full lockdep splat:

======================================================
WARNING: possible circular locking dependency detected
4.14.0-rc1-CI-CI_DRM_3118+ #1 Tainted: G     U
------------------------------------------------------
prime_mmap/1551 is trying to acquire lock:
 (cpu_hotplug_lock.rw_sem){++++}, at: [<ffffffff8109dbb7>] apply_workqueue_attrs+0x17/0x50

but task is already holding lock:
 (&dev_priv->mm_lock){+.+.}, at: [<ffffffffa01a7b2a>] i915_gem_userptr_init__mmu_notifier+0x14a/0x270 [i915]

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #6 (&dev_priv->mm_lock){+.+.}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       __mutex_lock+0x86/0x9b0
       mutex_lock_nested+0x1b/0x20
       i915_gem_userptr_init__mmu_notifier+0x14a/0x270 [i915]
       i915_gem_userptr_ioctl+0x222/0x2c0 [i915]
       drm_ioctl_kernel+0x69/0xb0
       drm_ioctl+0x2f9/0x3d0
       do_vfs_ioctl+0x94/0x670
       SyS_ioctl+0x41/0x70
       entry_SYSCALL_64_fastpath+0x1c/0xb1

-> #5 (&mm->mmap_sem){++++}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       __might_fault+0x68/0x90
       _copy_to_user+0x23/0x70
       filldir+0xa5/0x120
       dcache_readdir+0xf9/0x170
       iterate_dir+0x69/0x1a0
       SyS_getdents+0xa5/0x140
       entry_SYSCALL_64_fastpath+0x1c/0xb1

-> #4 (&sb->s_type->i_mutex_key#5){++++}:
       down_write+0x3b/0x70
       handle_create+0xcb/0x1e0
       devtmpfsd+0x139/0x180
       kthread+0x152/0x190
       ret_from_fork+0x27/0x40

-> #3 ((complete)&req.done){+.+.}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       wait_for_common+0x58/0x210
       wait_for_completion+0x1d/0x20
       devtmpfs_create_node+0x13d/0x160
       device_add+0x5eb/0x620
       device_create_groups_vargs+0xe0/0xf0
       device_create+0x3a/0x40
       msr_device_create+0x2b/0x40
       cpuhp_invoke_callback+0xa3/0x840
       cpuhp_thread_fun+0x7a/0x150
       smpboot_thread_fn+0x18a/0x280
       kthread+0x152/0x190
       ret_from_fork+0x27/0x40

-> #2 (cpuhp_state){+.+.}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       cpuhp_issue_call+0x10b/0x170
       __cpuhp_setup_state_cpuslocked+0x134/0x2a0
       __cpuhp_setup_state+0x46/0x60
       page_writeback_init+0x43/0x67
       pagecache_init+0x3d/0x42
       start_kernel+0x3a8/0x3fc
       x86_64_start_reservations+0x2a/0x2c
       x86_64_start_kernel+0x6d/0x70
       verify_cpu+0x0/0xfb

-> #1 (cpuhp_state_mutex){+.+.}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       __mutex_lock+0x86/0x9b0
       mutex_lock_nested+0x1b/0x20
       __cpuhp_setup_state_cpuslocked+0x52/0x2a0
       __cpuhp_setup_state+0x46/0x60
       page_alloc_init+0x28/0x30
       start_kernel+0x145/0x3fc
       x86_64_start_reservations+0x2a/0x2c
       x86_64_start_kernel+0x6d/0x70
       verify_cpu+0x0/0xfb

-> #0 (cpu_hotplug_lock.rw_sem){++++}:
       check_prev_add+0x430/0x840
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       cpus_read_lock+0x3d/0xb0
       apply_workqueue_attrs+0x17/0x50
       __alloc_workqueue_key+0x1d8/0x4d9
       i915_gem_userptr_init__mmu_notifier+0x1fb/0x270 [i915]
       i915_gem_userptr_ioctl+0x222/0x2c0 [i915]
       drm_ioctl_kernel+0x69/0xb0
       drm_ioctl+0x2f9/0x3d0
       do_vfs_ioctl+0x94/0x670
       SyS_ioctl+0x41/0x70
       entry_SYSCALL_64_fastpath+0x1c/0xb1

other info that might help us debug this:

Chain exists of:
  cpu_hotplug_lock.rw_sem --> &mm->mmap_sem --> &dev_priv->mm_lock

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&dev_priv->mm_lock);
                               lock(&mm->mmap_sem);
                               lock(&dev_priv->mm_lock);
  lock(cpu_hotplug_lock.rw_sem);

 *** DEADLOCK ***

2 locks held by prime_mmap/1551:
 #0:  (&mm->mmap_sem){++++}, at: [<ffffffffa01a7b18>] i915_gem_userptr_init__mmu_notifier+0x138/0x270 [i915]
 #1:  (&dev_priv->mm_lock){+.+.}, at: [<ffffffffa01a7b2a>] i915_gem_userptr_init__mmu_notifier+0x14a/0x270 [i915]

stack backtrace:
CPU: 4 PID: 1551 Comm: prime_mmap Tainted: G     U          4.14.0-rc1-CI-CI_DRM_3118+ #1
Hardware name: Dell Inc. XPS 8300  /0Y2MRG, BIOS A06 10/17/2011
Call Trace:
 dump_stack+0x68/0x9f
 print_circular_bug+0x235/0x3c0
 ? lockdep_init_map_crosslock+0x20/0x20
 check_prev_add+0x430/0x840
 __lock_acquire+0x1420/0x15e0
 ? __lock_acquire+0x1420/0x15e0
 ? lockdep_init_map_crosslock+0x20/0x20
 lock_acquire+0xb0/0x200
 ? apply_workqueue_attrs+0x17/0x50
 cpus_read_lock+0x3d/0xb0
 ? apply_workqueue_attrs+0x17/0x50
 apply_workqueue_attrs+0x17/0x50
 __alloc_workqueue_key+0x1d8/0x4d9
 ? __lockdep_init_map+0x57/0x1c0
 i915_gem_userptr_init__mmu_notifier+0x1fb/0x270 [i915]
 i915_gem_userptr_ioctl+0x222/0x2c0 [i915]
 ? i915_gem_userptr_release+0x140/0x140 [i915]
 drm_ioctl_kernel+0x69/0xb0
 drm_ioctl+0x2f9/0x3d0
 ? i915_gem_userptr_release+0x140/0x140 [i915]
 ? __do_page_fault+0x2a4/0x570
 do_vfs_ioctl+0x94/0x670
 ? entry_SYSCALL_64_fastpath+0x5/0xb1
 ? __this_cpu_preempt_check+0x13/0x20
 ? trace_hardirqs_on_caller+0xe3/0x1b0
 SyS_ioctl+0x41/0x70
 entry_SYSCALL_64_fastpath+0x1c/0xb1
RIP: 0033:0x7fbb83c39587
RSP: 002b:00007fff188dc228 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
RAX: ffffffffffffffda RBX: ffffffff81492963 RCX: 00007fbb83c39587
RDX: 00007fff188dc260 RSI: 00000000c0186473 RDI: 0000000000000003
RBP: ffffc90001487f88 R08: 0000000000000000 R09: 00007fff188dc2ac
R10: 00007fbb83efcb58 R11: 0000000000000246 R12: 0000000000000000
R13: 0000000000000003 R14: 00000000c0186473 R15: 00007fff188dc2ac
 ? __this_cpu_preempt_check+0x13/0x20

Note that this also has the minor benefit of slightly reducing the
critical section where we hold mmap_sem.

v2: Set ret correctly when we raced with another thread.

v3: Use Chris' diff. Attach the right lockdep splat.

v4: Repaint in Tvrtko's colors (aka don't report ENOMEM if we race and
some other thread managed to not also get an ENOMEM and successfully
install the mmu notifier. Note that the kernel guarantees that small
allocations succeed, so this never actually happens).

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sasha Levin <alexander.levin@verizon.com>
Cc: Marta Lofstedt <marta.lofstedt@intel.com>
Cc: Tejun Heo <tj@kernel.org>
References: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_3180/shard-hsw3/igt@prime_mmap@test_userptr.html
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102939
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171009164401.16035-1-daniel.vetter@ffwll.ch
drivers/gpu/drm/i915/i915_gem_userptr.c

index c36a84b..4d712a4 100644 (file)
@@ -164,7 +164,6 @@ static struct i915_mmu_notifier *
 i915_mmu_notifier_create(struct mm_struct *mm)
 {
        struct i915_mmu_notifier *mn;
-       int ret;
 
        mn = kmalloc(sizeof(*mn), GFP_KERNEL);
        if (mn == NULL)
@@ -179,14 +178,6 @@ i915_mmu_notifier_create(struct mm_struct *mm)
                return ERR_PTR(-ENOMEM);
        }
 
-        /* Protected by mmap_sem (write-lock) */
-       ret = __mmu_notifier_register(&mn->mn, mm);
-       if (ret) {
-               destroy_workqueue(mn->wq);
-               kfree(mn);
-               return ERR_PTR(ret);
-       }
-
        return mn;
 }
 
@@ -210,23 +201,40 @@ i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
 static struct i915_mmu_notifier *
 i915_mmu_notifier_find(struct i915_mm_struct *mm)
 {
-       struct i915_mmu_notifier *mn = mm->mn;
+       struct i915_mmu_notifier *mn;
+       int err = 0;
 
        mn = mm->mn;
        if (mn)
                return mn;
 
+       mn = i915_mmu_notifier_create(mm->mm);
+       if (IS_ERR(mn))
+               err = PTR_ERR(mn);
+
        down_write(&mm->mm->mmap_sem);
        mutex_lock(&mm->i915->mm_lock);
-       if ((mn = mm->mn) == NULL) {
-               mn = i915_mmu_notifier_create(mm->mm);
-               if (!IS_ERR(mn))
-                       mm->mn = mn;
+       if (mm->mn == NULL && !err) {
+               /* Protected by mmap_sem (write-lock) */
+               err = __mmu_notifier_register(&mn->mn, mm->mm);
+               if (!err) {
+                       /* Protected by mm_lock */
+                       mm->mn = fetch_and_zero(&mn);
+               }
+       } else {
+               /* someone else raced and successfully installed the mmu
+                * notifier, we can cancel our own errors */
+               err = 0;
        }
        mutex_unlock(&mm->i915->mm_lock);
        up_write(&mm->mm->mmap_sem);
 
-       return mn;
+       if (mn) {
+               destroy_workqueue(mn->wq);
+               kfree(mn);
+       }
+
+       return err ? ERR_PTR(err) : mm->mn;
 }
 
 static int