On 1/1/2025 11:36 PM, Jiang Liu wrote:
If some GPU device failed to probe, `rmmod amdgpu` will trigger a use
after free bug related to amdgpu_driver_release_kms() as:
2024-12-26 16:17:45 [16002.085540] BUG: kernel NULL pointer dereference, address: 0000000000000000
2024-12-26 16:17:45 [16002.093792] #PF: supervisor read access in kernel mode
2024-12-26 16:17:45 [16002.099993] #PF: error_code(0x0000) - not-present page
2024-12-26 16:17:45 [16002.106188] PGD 0 P4D 0
2024-12-26 16:17:45 [16002.109464] Oops: Oops: 0000 [#1] PREEMPT SMP NOPTI
2024-12-26 16:17:45 [16002.115372] CPU: 2 PID: 14375 Comm: rmmod Kdump: loaded Tainted: G W E 6.10.0+ #2
2024-12-26 16:17:45 [16002.125577] Hardware name: Alibaba Alibaba Cloud ECS/Alibaba Cloud ECS, BIOS 3.0.ES.AL.P.087.05 04/07/2024
2024-12-26 16:17:45 [16002.136858] RIP: 0010:drm_sched_fini+0x3f/0xe0 [gpu_sched]
2024-12-26 16:17:45 [16002.143463] Code: 60 c6 87 be 00 00 00 01 e8 ce e0 90 d8 48 8d bb 80 00 00 00 e8 c2 e0 90 d8 8b 43 20 85 c0 74 51 45 31 e4 48 8b
43 28 4d 63 ec <4a> 8b 2c e8 48 89 ef e8 f5 0e 59 d9 48 8b 45 10 48 8d 55 10 48 39
2024-12-26 16:17:45 [16002.164992] RSP: 0018:ffffb570dbbb7da8 EFLAGS: 00010246
2024-12-26 16:17:45 [16002.171316] RAX: 0000000000000000 RBX: ffff96b0fdadc878 RCX: 0000000000000000
2024-12-26 16:17:46 [16002.179784] RDX: 000fffffffe00000 RSI: 0000000000000000 RDI: ffff96b0fdadc8f8
2024-12-26 16:17:46 [16002.188252] RBP: ffff96b0fdadc800 R08: ffff97abbd035040 R09: ffffffff9ac52540
2024-12-26 16:17:46 [16002.196722] R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000000
2024-12-26 16:17:46 [16002.205179] R13: 0000000000000000 R14: ffff96b0fdadfc00 R15: 0000000000000000
2024-12-26 16:17:46 [16002.213648] FS: 00007f2737000740(0000) GS:ffff97abbd100000(0000) knlGS:0000000000000000
2024-12-26 16:17:46 [16002.223189] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
2024-12-26 16:17:46 [16002.230103] CR2: 0000000000000000 CR3: 000000011be3a005 CR4: 0000000000f70ef0
2024-12-26 16:17:46 [16002.238581] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
2024-12-26 16:17:46 [16002.247053] DR3: 0000000000000000 DR6: 00000000fffe07f0 DR7: 0000000000000400
e024se+0x3c/0x90 [amdxcp]
2024-12-26 16:17:46 [16002.337645] __do_sys_delete_module.constprop.0+0x176/0x310
2024-12-26 16:17:46 [16002.344324] do_syscall_64+0x5d/0x170
2024-12-26 16:17:46 [16002.348858] entry_SYSCALL_64_after_hwframe+0x76/0x7e
2024-12-26 16:17:46 [16002.354956] RIP: 0033:0x7f2736a620cb-12-26
Fix it by unplugging xcp drm devices when failed to probe GPU devices.
Signed-off-by: Jiang Liu <gerry@xxxxxxxxxxxxxxxxx>
Tested-by: Shuo Liu <shuox.liu@xxxxxxxxxxxxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 4 +++-
drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c | 1 +
2 files changed, 4 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
index 5ffe1dad9622..e7f35e3a6d2d 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
@@ -164,8 +164,10 @@ int amdgpu_driver_load_kms(struct amdgpu_device *adev, unsigned long flags)
DRM_WARN("smart shift update failed\n");
out:
- if (r)
+ if (r) {
+ amdgpu_xcp_dev_unplug(adev);
amdgpu_driver_unload_kms(dev);
+ }
I wonder if you can call amdgpu_xcp_dev_unplug
here. It will call drm_dev_unplug that unplugs a hotpluggable DRM
device. In you case amdgpu_driver_load_kms
failed during probe time. We need unwind amdgpu_driver_load_kms. Why need do unplug a DRM device? The backtrace show:
2024-12-26 16:17:45 [16002.136858] RIP: 0010:drm_sched_fini+0x3f/0xe0 [gpu_sched] has:
2024-12-26 16:17:45 [16002.085540] BUG: kernel NULL pointer dereference, address: 0000000000000000 How this change is related to the invalid access above? Can you
explain more?
Per my understanding, `amdgpu_xcp_dev_unplug(adev)` is a workaround for design flaw of the amdgpu_xcp driver. Currently the amdgpu_xcp driver assumes everything will go as expected and there’s no failure or GPU hot-lug operations. So it only provides an interface `amdgpu_xcp_drm_dev_alloc()` to create xcp devices for a GPU instance, and another interface `amdgpu_xcp_drv_release()` to destroy xcp devices for GPU instances. There’s no interface to undo what done by `amdgpu_xcp_drm_dev_alloc()`. And I found `amdgpu_xcp_dev_unplug(adev)` can undo work done by `amdgpu_xcp_drm_dev_alloc()`.
So it’s a workaround, the fundamental solution should be to enhance amdgpu_xcp driver to provide interface to unroll work done by `amdgpu_xcp_drm_dev_alloc()`.
The code paths to trigger the use after free are: 1) in function amdgpu_xcp_dev_alloc(), we allocate a drm_device by calling amdgpu_xcp_drm_dev_alloc(), and then change the device’s driver to amdgpu_partition_driver. 2) in function amdgpu_xcp_dev_unplug(), it restores drm_device’s driver to the original device driver.
In function amdgpu_driver_load_kms(), if we don’t call amdgpu_xcp_dev_unplug() on error recover path, the `xcp_dev[idx].driver` will still points to amdgpu_partition_driver. Later when amdgpu_xcp_drv_release() gets called, it will call platform_device_unregister() -> amdgpu_partition_driver.release -> amdgpu_driver_release_kms(). Here when amdgpu_driver_release_kms() gets called, the corresponding amdgpu_device object has already been released on error recovery path in function amdgpu_pci_probe().
So just like amdgpu_pci_remove(), we call amdgpu_xcp_gpu_dev_unplug() before calling amdkcl_drm_dev_release(). Regards Xiaogang
return r;
}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c
index a6d456ec6aeb..ef4eaacf67f6 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c
@@ -382,6 +382,7 @@ void amdgpu_xcp_dev_unplug(struct amdgpu_device *adev)
p_ddev->primary->dev = adev->xcp_mgr->xcp[i].pdev;
p_ddev->driver = adev->xcp_mgr->xcp[i].driver;
p_ddev->vma_offset_manager = adev->xcp_mgr->xcp[i].vma_offset_manager;
+ adev->xcp_mgr->xcp[i].ddev = NULL;
}
}
| 