Re: [PATCH] drm/ttm: stop passing NULL fence in ttm_bo_move_sync_cleanup

[Matthew Auld <matthew.auld@intel.com>]

On 11/04/2022 10:56, Thomas Hellström wrote:
> Hi, Matthew
> 
> On 4/11/22 10:56, Matthew Auld wrote:
>> If we hit the sync case, like when skipping clearing for kernel internal
>> objects, or when falling back to cpu clearing, like in i915, we end up
>> trying to add a NULL fence, but with some recent changes in this area
>> this now just results in NULL deref in dma_resv_add_fence:
>>
>> <1>[    5.466383] BUG: kernel NULL pointer dereference, address: 
>> 0000000000000008
>> <1>[    5.466384] #PF: supervisor read access in kernel mode
>> <1>[    5.466385] #PF: error_code(0x0000) - not-present page
>> <6>[    5.466386] PGD 0 P4D 0
>> <4>[    5.466387] Oops: 0000 [#1] PREEMPT SMP NOPTI
>> <4>[    5.466389] CPU: 5 PID: 267 Comm: modprobe Not tainted 
>> 5.18.0-rc2-CI-CI_DRM_11481+ #1
>> <4>[    5.466391] RIP: 0010:dma_resv_add_fence+0x63/0x260
>> <4>[    5.466395] Code: 38 85 c0 0f 84 df 01 00 00 0f 88 e8 01 00 00 
>> 83 c0 01 0f 88 df 01 00 00 8b 05 35 89 10 01 49 8d 5e 68 85 c0 0f 85 
>> 45 01 00 00 <48> 8b 45 08 48 3d c0 a5 0a 82 0f 84 5c 01 00 00 48 3d 60 
>> a5 0a 82
>> <4>[    5.466396] RSP: 0018:ffffc90000e974f8 EFLAGS: 00010202
>> <4>[    5.466397] RAX: 0000000000000001 RBX: ffff888123e88b28 RCX: 
>> 00000000ffffffff
>> <4>[    5.466398] RDX: 0000000000000001 RSI: ffffffff822e4f50 RDI: 
>> ffffffff8233f087
>> <4>[    5.466399] RBP: 0000000000000000 R08: ffff8881313dbc80 R09: 
>> 0000000000000001
>> <4>[    5.466399] R10: 0000000000000001 R11: 00000000da354294 R12: 
>> 0000000000000000
>> <4>[    5.466400] R13: ffff88810927dc58 R14: ffff888123e88ac0 R15: 
>> ffff88810a88d600
>> <4>[    5.466401] FS:  00007f5fa1193540(0000) 
>> GS:ffff88845d880000(0000) knlGS:0000000000000000
>> <4>[    5.466402] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>> <4>[    5.466402] CR2: 0000000000000008 CR3: 0000000106dd6003 CR4: 
>> 00000000003706e0
>> <4>[    5.466403] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 
>> 0000000000000000
>> <4>[    5.466404] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 
>> 0000000000000400
>> <4>[    5.466404] Call Trace:
>> <4>[    5.466405]  <TASK>
>> <4>[    5.466406]  ttm_bo_move_accel_cleanup+0x62/0x270 [ttm]
>> <4>[    5.466411]  ? i915_rsgt_from_buddy_resource+0x185/0x1e0 [i915]
>> <4>[    5.466529]  i915_ttm_move+0xfd/0x430 [i915]
>> <4>[    5.466833]  ? dma_resv_reserve_fences+0x4e/0x320
>> <4>[    5.466836]  ? ttm_bo_add_move_fence.constprop.20+0xf7/0x140 [ttm]
>> <4>[    5.466841]  ttm_bo_handle_move_mem+0xa1/0x140 [ttm]
>> <4>[    5.466845]  ttm_bo_validate+0xee/0x160 [ttm]
>> <4>[    5.466849]  __i915_ttm_get_pages+0x4f/0x210 [i915]
>> <4>[    5.466976]  i915_ttm_get_pages+0xad/0x140 [i915]
>> <4>[    5.467094]  ____i915_gem_object_get_pages+0x32/0xf0 [i915]
>> <4>[    5.467210]  __i915_gem_object_get_pages+0x89/0xa0 [i915]
>> <4>[    5.467323]  i915_vma_get_pages+0x114/0x1d0 [i915]
>> <4>[    5.467446]  i915_vma_pin_ww+0xd3/0xa90 [i915]
>> <4>[    5.467570]  i915_vma_pin.constprop.10+0x119/0x1b0 [i915]
>> <4>[    5.467700]  ? __mutex_unlock_slowpath+0x3e/0x2b0
>> <4>[    5.467704]  intel_alloc_initial_plane_obj.isra.6+0x1a9/0x390 
>> [i915]
>> <4>[    5.467833]  intel_crtc_initial_plane_config+0x83/0x340 [i915]
>>
>> In the ttm_bo_move_sync_cleanup() case it seems we only really care
>> about calling ttm_bo_wait_free_node(), so let's instead just call that
>> directly.
>>
>> Signed-off-by: Matthew Auld <matthew.auld@intel.com>
>> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
>> Cc: Christian König <christian.koenig@amd.com>
>> Cc: Lucas De Marchi <lucas.demarchi@intel.com>
>> Cc: Nirmoy Das <nirmoy.das@linux.intel.com>
>> ---
>>   drivers/gpu/drm/ttm/ttm_bo_util.c | 15 +++++++++++++++
>>   include/drm/ttm/ttm_bo_driver.h   | 11 +++--------
>>   2 files changed, 18 insertions(+), 8 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c 
>> b/drivers/gpu/drm/ttm/ttm_bo_util.c
>> index bc5190340b9c..1cbfb00c1d65 100644
>> --- a/drivers/gpu/drm/ttm/ttm_bo_util.c
>> +++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
>> @@ -572,6 +572,21 @@ int ttm_bo_move_accel_cleanup(struct 
>> ttm_buffer_object *bo,
>>   }
>>   EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
>> +void ttm_bo_move_sync_cleanup(struct ttm_buffer_object *bo,
>> +                  struct ttm_resource *new_mem)
>> +{
>> +    struct ttm_device *bdev = bo->bdev;
>> +    struct ttm_resource_manager *man = ttm_manager_type(bdev, 
>> new_mem->mem_type);
>> +    int ret;
>> +
>> +    ret = ttm_bo_wait_free_node(bo, man->use_tt);
>> +    if (WARN_ON(ret))
>> +        return;
>> +
>> +    ttm_bo_assign_mem(bo, new_mem);
>> +}
>> +EXPORT_SYMBOL(ttm_bo_move_sync_cleanup);
> 
> I don't think this will help in the case where we call 
> ttm_bo_move_accel_cleanup() with a NULL fence....

Hmm, do you know if that case actually exists? I thought the only 
current user passing fence == NULL was ttm_bo_move_sync_cleanup().

> 
> Perhaps we need to fix it there.
> 
> /Thomas
> 
> 
> 
>> +
>>   /**
>>    * ttm_bo_pipeline_gutting - purge the contents of a bo
>>    * @bo: The buffer object
>> diff --git a/include/drm/ttm/ttm_bo_driver.h 
>> b/include/drm/ttm/ttm_bo_driver.h
>> index 059a595e14e5..897b88f0bd59 100644
>> --- a/include/drm/ttm/ttm_bo_driver.h
>> +++ b/include/drm/ttm/ttm_bo_driver.h
>> @@ -245,7 +245,7 @@ int ttm_bo_move_accel_cleanup(struct 
>> ttm_buffer_object *bo,
>>                     struct ttm_resource *new_mem);
>>   /**
>> - * ttm_bo_move_accel_cleanup.
>> + * ttm_bo_move_sync_cleanup.
>>    *
>>    * @bo: A pointer to a struct ttm_buffer_object.
>>    * @new_mem: struct ttm_resource indicating where to move.
>> @@ -253,13 +253,8 @@ int ttm_bo_move_accel_cleanup(struct 
>> ttm_buffer_object *bo,
>>    * Special case of ttm_bo_move_accel_cleanup where the bo is guaranteed
>>    * by the caller to be idle. Typically used after memcpy buffer moves.
>>    */
>> -static inline void ttm_bo_move_sync_cleanup(struct ttm_buffer_object 
>> *bo,
>> -                        struct ttm_resource *new_mem)
>> -{
>> -    int ret = ttm_bo_move_accel_cleanup(bo, NULL, true, false, new_mem);
>> -
>> -    WARN_ON(ret);
>> -}
>> +void ttm_bo_move_sync_cleanup(struct ttm_buffer_object *bo,
>> +                  struct ttm_resource *new_mem);
>>   /**
>>    * ttm_bo_pipeline_gutting.