Re: [PATCH] drm: fix double free for gbo in drm_gem_vram_init and drm_gem_vram_create
From: Thomas Zimmermann
Date: Wed Jul 01 2020 - 02:57:04 EST
Hi
Thanks for the patch and apologies for being late with the review.
The fix is good, but I'd like to see different approach. I'd rather have
drm_gem_vram_init() being integrated into drm_gem_vram_create().
Do you prefer to make the patch or shall I type up something? Would you
be able to test?
I have some comments on the resulting changes below.
Am 01.07.20 um 04:32 schrieb Jia Yang:
> Ping...
>
> On 2020/6/20 14:21, Jia Yang wrote:
>> I got a use-after-free report when doing some fuzz test:
>>
>> If ttm_bo_init() fails, the "gbo" and "gbo->bo.base" will be
>> freed by ttm_buffer_object_destroy() in ttm_bo_init(). But
>> then drm_gem_vram_create() and drm_gem_vram_init() will free
>> "gbo" and "gbo->bo.base" again.
>>
>> BUG: KMSAN: use-after-free in drm_vma_offset_remove+0xb3/0x150
>> CPU: 0 PID: 24282 Comm: syz-executor.1 Tainted: G B W 5.7.0-rc4-msan #2
>> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
>> Call Trace:
>> __dump_stack
>> dump_stack+0x1c9/0x220
>> kmsan_report+0xf7/0x1e0
>> __msan_warning+0x58/0xa0
>> drm_vma_offset_remove+0xb3/0x150
>> drm_gem_free_mmap_offset
>> drm_gem_object_release+0x159/0x180
>> drm_gem_vram_init
>> drm_gem_vram_create+0x7c5/0x990
>> drm_gem_vram_fill_create_dumb
>> drm_gem_vram_driver_dumb_create+0x238/0x590
>> drm_mode_create_dumb
>> drm_mode_create_dumb_ioctl+0x41d/0x450
>> drm_ioctl_kernel+0x5a4/0x710
>> drm_ioctl+0xc6f/0x1240
>> vfs_ioctl
>> ksys_ioctl
>> __do_sys_ioctl
>> __se_sys_ioctl+0x2e9/0x410
>> __x64_sys_ioctl+0x4a/0x70
>> do_syscall_64+0xb8/0x160
>> entry_SYSCALL_64_after_hwframe+0x44/0xa9
>> RIP: 0033:0x4689b9
>> Code: fd e0 fa ff c3 66 2e 0f 1f 84 00 00 00 00 00 66 90 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 0f 83 cb e0 fa ff c3 66 2e 0f 1f 84 00 00 00 00
>> RSP: 002b:00007f368fa4dc98 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
>> RAX: ffffffffffffffda RBX: 000000000076bf00 RCX: 00000000004689b9
>> RDX: 0000000020000240 RSI: 00000000c02064b2 RDI: 0000000000000003
>> RBP: 0000000000000004 R08: 0000000000000000 R09: 0000000000000000
>> R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
>> R13: 00000000004d17e0 R14: 00007f368fa4e6d4 R15: 000000000076bf0c
>>
>> Uninit was created at:
>> kmsan_save_stack_with_flags
>> kmsan_internal_poison_shadow+0x66/0xd0
>> kmsan_slab_free+0x6e/0xb0
>> slab_free_freelist_hook
>> slab_free
>> kfree+0x571/0x30a0
>> drm_gem_vram_destroy
>> ttm_buffer_object_destroy+0xc8/0x130
>> ttm_bo_release
>> kref_put
>> ttm_bo_put+0x117d/0x23e0
>> ttm_bo_init_reserved+0x11c0/0x11d0
>> ttm_bo_init+0x289/0x3f0
>> drm_gem_vram_init
>> drm_gem_vram_create+0x775/0x990
>> drm_gem_vram_fill_create_dumb
>> drm_gem_vram_driver_dumb_create+0x238/0x590
>> drm_mode_create_dumb
>> drm_mode_create_dumb_ioctl+0x41d/0x450
>> drm_ioctl_kernel+0x5a4/0x710
>> drm_ioctl+0xc6f/0x1240
>> vfs_ioctl
>> ksys_ioctl
>> __do_sys_ioctl
>> __se_sys_ioctl+0x2e9/0x410
>> __x64_sys_ioctl+0x4a/0x70
>> do_syscall_64+0xb8/0x160
>> entry_SYSCALL_64_after_hwframe+0x44/0xa9
>>
>> If ttm_bo_init() fails, the "gbo" will be freed by
>> ttm_buffer_object_destroy() in ttm_bo_init(). But then
>> drm_gem_vram_create() and drm_gem_vram_init() will free
>> "gbo" again.
>>
>> Reported-by: Hulk Robot <hulkci@xxxxxxxxxx>
>> Signed-off-by: Jia Yang <jiayang5@xxxxxxxxxx>
>> ---
>> drivers/gpu/drm/drm_gem_vram_helper.c | 28 +++++++++++++++------------
>> 1 file changed, 16 insertions(+), 12 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/drm_gem_vram_helper.c b/drivers/gpu/drm/drm_gem_vram_helper.c
>> index 8b2d5c945c95..1d85af9a481a 100644
>> --- a/drivers/gpu/drm/drm_gem_vram_helper.c
>> +++ b/drivers/gpu/drm/drm_gem_vram_helper.c
>> @@ -175,6 +175,10 @@ static void drm_gem_vram_placement(struct drm_gem_vram_object *gbo,
>> }
>> }
>>
>> +/*
>> + * Note that on error, drm_gem_vram_init will free the buffer object.
>> + */
>> +
This comment could go away.
>> static int drm_gem_vram_init(struct drm_device *dev,
>> struct drm_gem_vram_object *gbo,
>> size_t size, unsigned long pg_align)
>> @@ -184,15 +188,19 @@ static int drm_gem_vram_init(struct drm_device *dev,
>> int ret;
>> size_t acc_size;
>>
>> - if (WARN_ONCE(!vmm, "VRAM MM not initialized"))
>> + if (WARN_ONCE(!vmm, "VRAM MM not initialized")) {
>> + kfree(gbo);
>> return -EINVAL;
>> + }
This test can go to the top of drm_gem_vram_create()
>> bdev = &vmm->bdev;
>>
>> gbo->bo.base.funcs = &drm_gem_vram_object_funcs;
>>
>> ret = drm_gem_object_init(dev, &gbo->bo.base, size);
>> - if (ret)
>> + if (ret) {
>> + kfree(gbo);
>> return ret;
>> + }
>>
>> acc_size = ttm_bo_dma_acc_size(bdev, size, sizeof(*gbo));
>>
>> @@ -203,13 +211,13 @@ static int drm_gem_vram_init(struct drm_device *dev,
>> &gbo->placement, pg_align, false, acc_size,
>> NULL, NULL, ttm_buffer_object_destroy);
>> if (ret)
>> - goto err_drm_gem_object_release;
>> + /*
>> + * A failing ttm_bo_init will call ttm_buffer_object_destroy
>> + * to release gbo->bo.base and kfree gbo.
>> + */
>> + return ret;
The rest of this function would be inlined where the call to
drm_gem_vram_init() currently happens.
>>
>> return 0;
>> -
>> -err_drm_gem_object_release:
>> - drm_gem_object_release(&gbo->bo.base);
>> - return ret;
>> }
>>
>> /**
>> @@ -243,13 +251,9 @@ struct drm_gem_vram_object *drm_gem_vram_create(struct drm_device *dev,
>>
>> ret = drm_gem_vram_init(dev, gbo, size, pg_align);
>> if (ret < 0)
>> - goto err_kfree;
>> + return ERR_PTR(ret);
>>
>> return gbo;
>> -
>> -err_kfree:
>> - kfree(gbo);
>> - return ERR_PTR(ret);
All inlined code up to ttm_bo_init() would jump to error labels on
errors. ttm_bo_init() would return directly.
Best regards
Thomas
>> }
>> EXPORT_SYMBOL(drm_gem_vram_create);
>>
>>
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 NÃrnberg, Germany
(HRB 36809, AG NÃrnberg)
GeschÃftsfÃhrer: Felix ImendÃrffer
Attachment:
signature.asc
Description: OpenPGP digital signature