drm/i915: Avoid calling i915_gem_object_unbind holding object lock
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 6 Dec 2019 10:55:27 +0000 (10:55 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Sat, 7 Dec 2019 19:27:36 +0000 (19:27 +0000)
commit8b1c78e06e6167f49beb27be073faa2c6c5c6eff
tree65604e62c759ba6d35922ec3f790beb254b54dab
parenta22198a934e67d461657cde8d79daa305441b9e3
drm/i915: Avoid calling i915_gem_object_unbind holding object lock

In the extreme case, we may wish to wait on an rcu-barrier to reap stale
vm to purge the last of the object bindings. However, we are not allowed
to use rcu_barrier() beneath the dma_resv (i.e. object) lock and do not
take lightly the prospect of unlocking a mutex deep in the bowels of the
routine. i915_gem_object_unbind() itself does not need the object lock,
and it turns out the callers do not need to the unbind as part of a
locked sequence around set-cache-level, so rearrange the code to avoid
taking the object lock in the callers.

<4> [186.816311] ======================================================
<4> [186.816313] WARNING: possible circular locking dependency detected
<4> [186.816316] 5.4.0-rc8-CI-CI_DRM_7486+ #1 Tainted: G     U
<4> [186.816318] ------------------------------------------------------
<4> [186.816320] perf_pmu/1321 is trying to acquire lock:
<4> [186.816322] ffff88849487c4d8 (&mm->mmap_sem#2){++++}, at: __might_fault+0x39/0x90
<4> [186.816331]
but task is already holding lock:
<4> [186.816333] ffffe8ffffa05008 (&cpuctx_mutex){+.+.}, at: perf_event_ctx_lock_nested+0xa9/0x1b0
<4> [186.816339]
which lock already depends on the new lock.

<4> [186.816341]
the existing dependency chain (in reverse order) is:
<4> [186.816343]
-> #6 (&cpuctx_mutex){+.+.}:
<4> [186.816349]        __mutex_lock+0x9a/0x9d0
<4> [186.816352]        perf_event_init_cpu+0xa4/0x140
<4> [186.816357]        perf_event_init+0x19d/0x1cd
<4> [186.816362]        start_kernel+0x372/0x4f4
<4> [186.816365]        secondary_startup_64+0xa4/0xb0
<4> [186.816381]
-> #5 (pmus_lock){+.+.}:
<4> [186.816385]        __mutex_lock+0x9a/0x9d0
<4> [186.816387]        perf_event_init_cpu+0x6b/0x140
<4> [186.816404]        cpuhp_invoke_callback+0x9b/0x9d0
<4> [186.816406]        _cpu_up+0xa2/0x140
<4> [186.816409]        do_cpu_up+0x61/0xa0
<4> [186.816411]        smp_init+0x57/0x96
<4> [186.816413]        kernel_init_freeable+0xac/0x1c7
<4> [186.816416]        kernel_init+0x5/0x100
<4> [186.816419]        ret_from_fork+0x24/0x50
<4> [186.816421]
-> #4 (cpu_hotplug_lock.rw_sem){++++}:
<4> [186.816424]        cpus_read_lock+0x34/0xd0
<4> [186.816427]        rcu_barrier+0xaa/0x190
<4> [186.816429]        kernel_init+0x21/0x100
<4> [186.816431]        ret_from_fork+0x24/0x50
<4> [186.816433]
-> #3 (rcu_state.barrier_mutex){+.+.}:
<4> [186.816436]        __mutex_lock+0x9a/0x9d0
<4> [186.816438]        rcu_barrier+0x23/0x190
<4> [186.816502]        i915_gem_object_unbind+0x3a6/0x400 [i915]
<4> [186.816537]        i915_gem_object_set_cache_level+0x32/0x90 [i915]
<4> [186.816571]        i915_gem_object_pin_to_display_plane+0x5d/0x160 [i915]
<4> [186.816612]        intel_pin_and_fence_fb_obj+0x9e/0x200 [i915]
<4> [186.816679]        intel_plane_pin_fb+0x3f/0xd0 [i915]
<4> [186.816717]        intel_prepare_plane_fb+0x130/0x520 [i915]
<4> [186.816722]        drm_atomic_helper_prepare_planes+0x85/0x110
<4> [186.816761]        intel_atomic_commit+0xc6/0x350 [i915]
<4> [186.816764]        drm_atomic_helper_update_plane+0xed/0x110
<4> [186.816768]        setplane_internal+0x97/0x190
<4> [186.816770]        drm_mode_setplane+0xcd/0x190
<4> [186.816773]        drm_ioctl_kernel+0xa7/0xf0
<4> [186.816775]        drm_ioctl+0x2e1/0x390
<4> [186.816778]        do_vfs_ioctl+0xa0/0x6f0
<4> [186.816780]        ksys_ioctl+0x35/0x60
<4> [186.816782]        __x64_sys_ioctl+0x11/0x20
<4> [186.816785]        do_syscall_64+0x4f/0x210
<4> [186.816787]        entry_SYSCALL_64_after_hwframe+0x49/0xbe
<4> [186.816789]
-> #2 (reservation_ww_class_mutex){+.+.}:
<4> [186.816793]        __ww_mutex_lock.constprop.15+0xc3/0x1090
<4> [186.816795]        ww_mutex_lock+0x39/0x70
<4> [186.816798]        dma_resv_lockdep+0x10e/0x1f7
<4> [186.816800]        do_one_initcall+0x58/0x2ff
<4> [186.816802]        kernel_init_freeable+0x137/0x1c7
<4> [186.816804]        kernel_init+0x5/0x100
<4> [186.816806]        ret_from_fork+0x24/0x50
<4> [186.816808]
-> #1 (reservation_ww_class_acquire){+.+.}:
<4> [186.816811]        dma_resv_lockdep+0xec/0x1f7
<4> [186.816813]        do_one_initcall+0x58/0x2ff
<4> [186.816815]        kernel_init_freeable+0x137/0x1c7
<4> [186.816817]        kernel_init+0x5/0x100
<4> [186.816819]        ret_from_fork+0x24/0x50
<4> [186.816820]
-> #0 (&mm->mmap_sem#2){++++}:
<4> [186.816824]        __lock_acquire+0x1328/0x15d0
<4> [186.816826]        lock_acquire+0xa7/0x1c0
<4> [186.816828]        __might_fault+0x63/0x90
<4> [186.816831]        _copy_to_user+0x1e/0x80
<4> [186.816834]        perf_read+0x200/0x2b0
<4> [186.816836]        vfs_read+0x96/0x160
<4> [186.816838]        ksys_read+0x9f/0xe0
<4> [186.816839]        do_syscall_64+0x4f/0x210
<4> [186.816841]        entry_SYSCALL_64_after_hwframe+0x49/0xbe
<4> [186.816843]
other info that might help us debug this:

<4> [186.816846] Chain exists of:
  &mm->mmap_sem#2 --> pmus_lock --> &cpuctx_mutex

<4> [186.816849]  Possible unsafe locking scenario:

<4> [186.816851]        CPU0                    CPU1
<4> [186.816853]        ----                    ----
<4> [186.816854]   lock(&cpuctx_mutex);
<4> [186.816856]                                lock(pmus_lock);
<4> [186.816858]                                lock(&cpuctx_mutex);
<4> [186.816860]   lock(&mm->mmap_sem#2);
<4> [186.816861]
 *** DEADLOCK ***

Closes: https://gitlab.freedesktop.org/drm/intel/issues/728
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191206105527.1130413-5-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/display/intel_display.c
drivers/gpu/drm/i915/display/intel_overlay.c
drivers/gpu/drm/i915/gem/i915_gem_domain.c