[Intel-gfx] [PATCH] drm/i915: Fix the use-after-free between i915_active_ref and __active_retire
Liu, Chuansheng
chuansheng.liu at intel.com
Sat Dec 7 01:50:00 UTC 2019
> -----Original Message-----
> From: Chris Wilson <chris at chris-wilson.co.uk>
> Sent: Friday, December 6, 2019 8:15 PM
> To: Liu, Chuansheng <chuansheng.liu at intel.com>; intel-
> gfx at lists.freedesktop.org
> Subject: RE: [Intel-gfx] [PATCH] drm/i915: Fix the use-after-free between
> i915_active_ref and __active_retire
>
> Quoting Liu, Chuansheng (2019-12-06 12:10:25)
> > Chris,
> >
> > Thanks for reviewing, please see below comments.
> >
> > > -----Original Message-----
> > > From: Chris Wilson <chris at chris-wilson.co.uk>
> > > Sent: Friday, December 6, 2019 8:04 PM
> > > To: Liu, Chuansheng <chuansheng.liu at intel.com>; intel-
> > > gfx at lists.freedesktop.org
> > > Subject: Re: [Intel-gfx] [PATCH] drm/i915: Fix the use-after-free between
> > > i915_active_ref and __active_retire
> > >
> > > Quoting Chuansheng Liu (2019-12-06 11:56:35)
> > > > We easily hit drm/i915 panic on TGL when running glmark2, and finally
> > > > caught the race condition of use-after-free with enabling KASAN.
> > > >
> > > > The call stack is below:
> > > > ===
> > > > [ 534.472675] BUG: KASAN: use-after-free in
> > > __i915_active_fence_set+0x26d/0x3d0 [i915]
> > > > [ 534.472679] Write of size 8 at addr ffff8883f0372388 by task
> glmark2/3199
> > > >
> > > > [ 534.472684] CPU: 3 PID: 3199 Comm: glmark2 Tainted: G U E
> 5.4.0-
> > > rc8 #8
> > > > [ 534.472687] Call Trace:
> > > > [ 534.472693] dump_stack+0x95/0xd5
> > > > [ 534.472722] ? __i915_active_fence_set+0x26d/0x3d0 [i915]
> > > > [ 534.472727] print_address_description.constprop.5+0x20/0x320
> > > > [ 534.472751] ? __i915_active_fence_set+0x26d/0x3d0 [i915]
> > > > [ 534.472792] ? __i915_active_fence_set+0x26d/0x3d0 [i915]
> > > > [ 534.472794] __kasan_report+0x149/0x18c
> > > > [ 534.472798] ? _raw_spin_lock+0x1/0xd0
> > > > [ 534.472820] ? __i915_active_fence_set+0x26d/0x3d0 [i915]
> > > > [ 534.472822] kasan_report+0x12/0x20
> > > > [ 534.472825] __asan_report_store8_noabort+0x17/0x20
> > > > [ 534.472847] __i915_active_fence_set+0x26d/0x3d0 [i915]
> > > > [ 534.472870] i915_active_ref+0x2c8/0x530 [i915]
> > > > [ 534.472874] ? dma_resv_add_shared_fence+0x291/0x460
> > > > [ 534.472902] __i915_vma_move_to_active+0x56/0x70 [i915]
> > > > [ 534.472927] i915_vma_move_to_active+0x54/0x420 [i915]
> > > > [ 534.472931] ? mutex_unlock+0x22/0x40
> > > > [ 534.472957] i915_gem_do_execbuffer+0x1d45/0x3e20 [i915]
> > > > [ 534.472959] ? __kmalloc_node+0x12c/0x350
> > > > [ 534.472983] ? eb_relocate_slow+0xb40/0xb40 [i915]
> > > > [ 534.472985] ? _raw_write_trylock+0x110/0x110
> > > > [ 534.472987] ? get_partial_node.isra.72+0x51/0x260
> > > > [ 534.472991] ? unix_stream_read_generic+0x583/0x1a80
> > > > [ 534.472994] ? ___slab_alloc+0x1d8/0x550
> > > > [ 534.472998] ? kvmalloc_node+0x31/0x80
> > > > [ 534.473000] ? kasan_unpoison_shadow+0x35/0x50
> > > > [ 534.473002] ? _raw_spin_lock+0x7b/0xd0
> > > > [ 534.473004] ? radix_tree_lookup+0xd/0x10
> > > > [ 534.473006] ? idr_find+0x3b/0x60
> > > > [ 534.473029] i915_gem_execbuffer2_ioctl+0x634/0x8a0 [i915]
> > > > [ 534.473052] ? i915_gem_execbuffer_ioctl+0xd50/0xd50 [i915]
> > > > [ 534.473054] ? unix_stream_recvmsg+0x97/0xd0
> > > > [ 534.473056] ? unix_stream_splice_read+0x1c0/0x1c0
> > > > [ 534.473058] ? __unix_insert_socket+0x180/0x180
> > > > [ 534.473081] ? i915_gem_execbuffer_ioctl+0xd50/0xd50 [i915]
> > > > [ 534.473094] drm_ioctl_kernel+0x1ed/0x2b0 [drm]
> > > > [ 534.473103] ? drm_setversion+0x8c0/0x8c0 [drm]
> > > > [ 534.473106] ? __kasan_check_write+0x14/0x20
> > > > [ 534.473115] drm_ioctl+0x68b/0xaa0 [drm]
> > > > ...
> > > >
> > > > [ 534.473239] Allocated by task 3199:
> > > > [ 534.473241] save_stack+0x21/0x90
> > > > [ 534.473243] __kasan_kmalloc.constprop.8+0xa7/0xd0
> > > > [ 534.473245] kasan_slab_alloc+0x11/0x20
> > > > [ 534.473246] kmem_cache_alloc+0xce/0x240
> > > > [ 534.473273] i915_active_ref+0xc2/0x530 [i915]
> > > > [ 534.473302] __i915_vma_move_to_active+0x56/0x70 [i915]
> > > > [ 534.473328] i915_vma_move_to_active+0x54/0x420 [i915]
> > > > [ 534.473355] i915_gem_do_execbuffer+0x1d45/0x3e20 [i915]
> > > > [ 534.473381] i915_gem_execbuffer2_ioctl+0x634/0x8a0 [i915]
> > > > [ 534.473392] drm_ioctl_kernel+0x1ed/0x2b0 [drm]
> > > > [ 534.473402] drm_ioctl+0x68b/0xaa0 [drm]
> > > > [ 534.473404] do_vfs_ioctl+0x19a/0xf10
> > > > [ 534.473405] ksys_ioctl+0x75/0x80
> > > > [ 534.473407] __x64_sys_ioctl+0x73/0xb0
> > > > [ 534.473408] do_syscall_64+0x9f/0x3a0
> > > > [ 534.473410] entry_SYSCALL_64_after_hwframe+0x44/0xa9
> > > >
> > > > [ 534.473412] Freed by task 0:
> > > > [ 534.473414] save_stack+0x21/0x90
> > > > [ 534.473415] __kasan_slab_free+0x137/0x190
> > > > [ 534.473417] kasan_slab_free+0xe/0x10
> > > > [ 534.473418] kmem_cache_free+0xeb/0x2c0
> > > > [ 534.473444] __active_retire+0x1f2/0x240 [i915]
> > > > [ 534.473471] active_retire+0x13b/0x1b0 [i915]
> > > > [ 534.473496] node_retire+0x54/0x80 [i915]
> > > > [ 534.473523] intel_engine_breadcrumbs_irq+0x5f0/0xd10 [i915]
> > > > [ 534.473549] cs_irq_handler+0x66/0xb0 [i915]
> > > > [ 534.473575] gen11_gt_irq_handler+0x26c/0x400 [i915]
> > > > [ 534.473600] gen11_irq_handler+0xc3/0x250 [i915]
> > > > [ 534.473603] __handle_irq_event_percpu+0xe0/0x4c0
> > > > [ 534.473605] handle_irq_event_percpu+0x71/0x140
> > > > [ 534.473606] handle_irq_event+0xad/0x140
> > > > [ 534.473608] handle_edge_irq+0x1f6/0x780
> > > > [ 534.473610] do_IRQ+0x9f/0x1f0
> > > >
> > > > [ 534.473612] The buggy address belongs to the object at
> ffff8883f0372380
> > > > which belongs to the cache active_node of size 72
> > > > [ 534.473615] The buggy address is located 8 bytes inside of
> > > >
> > > > ===
> > > >
> > > > The race scenerio is like:
> > > > Initially ref->count is 1, interrupt handler is trying to free the
> > > > node.
> > > >
> > > > ===
> > > > CPUA in interrupt context CPUB in i915_gem_execbuffer2_ioctl
> > > > __active_retire -->
> > > > spin_lock(&ref->tree_lock)
> > > > decrease ref->count to 0
> > > > i915_active_ref -->
> > > > increase ref->count to 1
> > > > (i915_active_acquire)
> > > >
> > > > get the dirty ref->cache
> > > > (READ_ONCE(ref->cache))
> > > >
> > > > return the dirty node
> > > >
> > > > set ref->cache to NULL
> > > > spin_unlock(&ref->tree_lock)
> > > > free the node
> > > >
> > > > hit use-after-free in
> > > > __i915_active_fence_set()
> > > >
> > > > ===
> > > >
> > > > Here we need to use spinlock ref->tree_lock to protect the access
> > > > of READ_ONCE(ref->cache), then the race scenerio can be resolved.
> > > >
> > > > with this patch, it passed our stress test for a very long time.
> > > >
> > > > Signed-off-by: Chuansheng Liu <chuansheng.liu at intel.com>
> > > > ---
> > > > drivers/gpu/drm/i915/i915_active.c | 3 +++
> > > > 1 file changed, 3 insertions(+)
> > > >
> > > > diff --git a/drivers/gpu/drm/i915/i915_active.c
> > > b/drivers/gpu/drm/i915/i915_active.c
> > > > index dca15ace88f6..3d68b910e949 100644
> > > > --- a/drivers/gpu/drm/i915/i915_active.c
> > > > +++ b/drivers/gpu/drm/i915/i915_active.c
> > > > @@ -214,7 +214,10 @@ active_instance(struct i915_active *ref, struct
> > > intel_timeline *tl)
> > > > * after the previous activity has been retired, or if it matches the
> > > > * current timeline.
> > > > */
> > > > + spin_lock_irq(&ref->tree_lock);
> > > > node = READ_ONCE(ref->cache);
> > > > + spin_unlock_irq(&ref->tree_lock);
> > >
> > > Incorrect. The serialisation with __active_retire is required at
> > > i915_active_acquire.
> > You suggest the change can be made in i915_active_acquire()?
> > So that we can play ref->count closely together with tree_lock
> > and ODEBUG stuff.
> >
> > If so, I can make a new patch😊
>
> See bbca083de291, the same race was reported last night.
Thanks for your patch bbca083de291, we will try it on our platforms.
Looking into the code in i915_active_acquire(), I think the overhead
of spinlock may be introduced in case of multi callers of i915_active_acquire()
at the same time, to avoid such overhead, I suggest below change, does
it make sense? Thanks.
- if (!atomic_read(&ref->count) && ref->active)
- err = ref->active(ref);
- if (!err) {
- spin_lock_irq(&ref->tree_lock); /* vs __active_retire() */
- debug_active_activate(ref);
- atomic_inc(&ref->count);
- spin_unlock_irq(&ref->tree_lock);
+ if (!atomic_add_unless(&ref->count, 1, 0)) {
+ if (ref->active)
+ err = ref->active(ref);
+ if (!err) {
+ /* vs __active_retire() */
+ spin_lock_irq(&ref->tree_lock);
+ debug_active_activate(ref);
+ atomic_inc(&ref->count);
+ spin_unlock_irq(&ref->tree_lock);
+ }
}
mutex_unlock(&ref->mutex);
More information about the Intel-gfx
mailing list