drm/amdgpu: fix fdinfo race with process exit
Subsystems (2)
PATCH v2 1/1 drm/amdgpu: fix fdinfo race with process exit
Get process vm root BO ref in case process is exiting and root BO is freed, to avoid NULL pointer dereference backtrace:
BUG: unable to handle kernel NULL pointer dereference at 0000000000000000 Call Trace: amdgpu_show_fdinfo+0xfe/0x2a0 [amdgpu] seq_show+0x12c/0x180 seq_read+0x153/0x410 vfs_read+0x91/0x140[ 3427.206183] ksys_read+0x4f/0xb0 do_syscall_64+0x5b/0x1a0 entry_SYSCALL_64_after_hwframe+0x65/0xca
v2: rebase to staging
--- drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c +9 -2
@@ -59,6 +59,7 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f)uint64_t vram_mem =0 , gtt_mem =0 , cpu_mem =0 ;struct drm_file *file = f->private_data;struct amdgpu_device *adev =drm_to_adev (file->minor->dev); +struct amdgpu_bo *root;int ret; ret =amdgpu_file_to_fpriv (f, &fpriv);
@@ -69,13 +70,19 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f) dev =PCI_SLOT (adev->pdev->devfn); fn =PCI_FUNC (adev->pdev->devfn); - ret =amdgpu_bo_reserve (fpriv->vm.root.bo, false); + root =amdgpu_bo_ref (fpriv->vm.root.bo); +if (!root) +return ; + + ret =amdgpu_bo_reserve (root, false);if (ret) {DRM_ERROR ("Fail to reserve bo\n" );return ; }amdgpu_vm_get_memory (&fpriv->vm, &vram_mem, >t_mem, &cpu_mem); -amdgpu_bo_unreserve (fpriv->vm.root.bo); +amdgpu_bo_unreserve (root); +amdgpu_bo_unref (&root); +seq_printf (m,"pdev:\t%04x:%02x:%02x.%d\npasid:\t%u\n" , domain, bus, dev, fn, fpriv->vm.pasid);seq_printf (m,"vram mem:\t%llu kB\n" , vram_mem/1024UL );
Signature · 6 lines
2.17.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
ping?
On 2021-07-29 10:13 p.m., Philip Yang wrote:
Get process vm root BO ref in case process is exiting and root BO is freed, to avoid NULL pointer dereference backtrace: BUG: unable to handle kernel NULL pointer dereference at 0000000000000000 Call Trace: amdgpu_show_fdinfo+0xfe/0x2a0 [amdgpu] seq_show+0x12c/0x180 seq_read+0x153/0x410 vfs_read+0x91/0x140[ 3427.206183] ksys_read+0x4f/0xb0 do_syscall_64+0x5b/0x1a0 entry_SYSCALL_64_after_hwframe+0x65/0xca v2: rebase to staging Signed-off-by: Philip Yang <[email protected]> --- drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c index d94c5419ec25..5a6857c44bb6 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c @@ -59,6 +59,7 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f) uint64_t vram_mem = 0, gtt_mem = 0, cpu_mem = 0; struct drm_file *file = f->private_data; struct amdgpu_device *adev = drm_to_adev(file->minor->dev); + struct amdgpu_bo *root; int ret; ret = amdgpu_file_to_fpriv(f, &fpriv); @@ -69,13 +70,19 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f) dev = PCI_SLOT(adev->pdev->devfn); fn = PCI_FUNC(adev->pdev->devfn); - ret = amdgpu_bo_reserve(fpriv->vm.root.bo, false); + root = amdgpu_bo_ref(fpriv->vm.root.bo); + if (!root) + return; + + ret = amdgpu_bo_reserve(root, false); if (ret) { DRM_ERROR("Fail to reserve bo\n"); return; } amdgpu_vm_get_memory(&fpriv->vm, &vram_mem, >t_mem, &cpu_mem); - amdgpu_bo_unreserve(fpriv->vm.root.bo); + amdgpu_bo_unreserve(root); + amdgpu_bo_unref(&root); + seq_printf(m, "pdev:\t%04x:%02x:%02x.%d\npasid:\t%u\n", domain, bus, dev, fn, fpriv->vm.pasid); seq_printf(m, "vram mem:\t%llu kB\n", vram_mem/1024UL);
Sorry I'm on vacation and can't reply immediately.
This is the wrong approach. The fdinfo should have grabbed a reference to the fd it prints the info for.
So we should never race here. Can you double check how this happens?
Thanks,
Christian.
Am 03.08.21 um 16:06 schrieb philip yang:
Quoted 58 lines
ping?
Quoted 54 lines · Philip Yang
Get process vm root BO ref in case process is exiting and root BO is freed, to avoid NULL pointer dereference backtrace:
BUG: unable to handle kernel NULL pointer dereference at 0000000000000000 Call Trace: amdgpu_show_fdinfo+0xfe/0x2a0 [amdgpu] seq_show+0x12c/0x180 seq_read+0x153/0x410 vfs_read+0x91/0x140[ 3427.206183] ksys_read+0x4f/0xb0 do_syscall_64+0x5b/0x1a0 entry_SYSCALL_64_after_hwframe+0x65/0xcav2: rebase to staging
Signed-off-by Philip Yang<[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c +9
-2@@ -59,6 +59,7 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f)uint64_t vram_mem =0 , gtt_mem =0 , cpu_mem =0 ;struct drm_file *file = f->private_data;struct amdgpu_device *adev =drm_to_adev (file->minor->dev); +struct amdgpu_bo *root;int ret; ret =amdgpu_file_to_fpriv (f, &fpriv);@@ -69,13 +70,19 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f) dev =PCI_SLOT (adev->pdev->devfn); fn =PCI_FUNC (adev->pdev->devfn); - ret =amdgpu_bo_reserve (fpriv->vm.root.bo, false); + root =amdgpu_bo_ref (fpriv->vm.root.bo); +if (!root) +return ; + + ret =amdgpu_bo_reserve (root, false);if (ret) {DRM_ERROR ("Fail to reserve bo\n" );return ; }amdgpu_vm_get_memory (&fpriv->vm, &vram_mem, >t_mem, &cpu_mem); -amdgpu_bo_unreserve (fpriv->vm.root.bo); +amdgpu_bo_unreserve (root); +amdgpu_bo_unref (&root); +seq_printf (m,"pdev:\t%04x:%02x:%02x.%d\npasid:\t%u\n" , domain, bus, dev, fn, fpriv->vm.pasid);seq_printf (m,"vram mem:\t%llu kB\n" , vram_mem/1024UL );
On 2021-08-04 5:04 a.m., Christian König wrote:
Sorry I'm on vacation and can't reply immediately.
This is the wrong approach. The fdinfo should have grabbed a reference to the fd it prints the info for.
So we should never race here. Can you double check how this happens?
This backtrace happened once, from /var/crash/..$date../vmcode-dmesg.log on the server machine, I can not repro the issue, grep app folder, there are python scripts accessing /proc/pid/node_id/fdinfo. This happened after app crash segmentation fault killed.
fdinfo grab fpriv reference, but not fpriv->vm.root.bo reference, I think this is needed, otherwise amdgpu_bo_reserve(fpriv->vm.root.bo) may deference NULL pointer.
Regards,
Philip
Thanks,
Christian.
Am 03.08.21 um 16:06 schrieb philip yang:
ping?
On 2021-07-29 10:13 p.m., Philip Yang wrote:
Get process vm root BO ref in case process is exiting and root BO is freed, to avoid NULL pointer dereference backtrace:
BUG: unable to handle kernel NULL pointer dereference at 0000000000000000 Call Trace: amdgpu_show_fdinfo+0xfe/0x2a0 [amdgpu] seq_show+0x12c/0x180 seq_read+0x153/0x410 vfs_read+0x91/0x140[ 3427.206183] ksys_read+0x4f/0xb0 do_syscall_64+0x5b/0x1a0 entry_SYSCALL_64_after_hwframe+0x65/0xca
v2: rebase to staging
---
drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c | 11 +++++++++--
1 file changed, 9 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c +0 -0
@@ -59,6 +59,7 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f)
uint64_t vram_mem = 0, gtt_mem = 0, cpu_mem = 0;
struct drm_file *file = f->private_data;
struct amdgpu_device *adev = drm_to_adev(file->minor->dev); + struct amdgpu_bo *root;
int ret;
ret = amdgpu_file_to_fpriv(f, &fpriv);
@@ -69,13 +70,19 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f) dev = PCI_SLOT(adev->pdev->devfn);
fn = PCI_FUNC(adev->pdev->devfn);
- ret = amdgpu_bo_reserve(fpriv->vm.root.bo, false);
+ root = amdgpu_bo_ref(fpriv->vm.root.bo);
+ if (!root)
+ return;
+
+ ret = amdgpu_bo_reserve(root, false);
if (ret) {
DRM_ERROR("Fail to reserve bo\n");
return;
}
amdgpu_vm_get_memory(&fpriv->vm, &vram_mem, >t_mem, &cpu_mem); - amdgpu_bo_unreserve(fpriv->vm.root.bo);
+ amdgpu_bo_unreserve(root);
+ amdgpu_bo_unref(&root);
+
seq_printf(m, "pdev:\t%04x:%02x:%02x.%d\npasid:\t%u\n", domain, bus, dev, fn, fpriv->vm.pasid);
seq_printf(m, "vram mem:\t%llu kB\n", vram_mem/1024UL);
Am 2021-07-29 um 10:13 p.m. schrieb Philip Yang:
Quoted 13 lines
Get process vm root BO ref in case process is exiting and root BO is freed, to avoid NULL pointer dereference backtrace:
BUG: unable to handle kernel NULL pointer dereference at 0000000000000000 Call Trace: amdgpu_show_fdinfo+0xfe/0x2a0 [amdgpu] seq_show+0x12c/0x180 seq_read+0x153/0x410 vfs_read+0x91/0x140[ 3427.206183] ksys_read+0x4f/0xb0 do_syscall_64+0x5b/0x1a0 entry_SYSCALL_64_after_hwframe+0x65/0xca
Quoted 3 lines
v2: rebase to staging
Signed-off-by Philip Yang <[email protected]>
The patch is
Quoted 38 lines
--- drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c +9
-2@@ -59,6 +59,7 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f)uint64_t vram_mem =0 , gtt_mem =0 , cpu_mem =0 ;struct drm_file *file = f->private_data;struct amdgpu_device *adev =drm_to_adev (file->minor->dev); +struct amdgpu_bo *root;int ret; ret =amdgpu_file_to_fpriv (f, &fpriv);@@ -69,13 +70,19 @@ void amdgpu_show_fdinfo(struct seq_file *m, struct file *f) dev =PCI_SLOT (adev->pdev->devfn); fn =PCI_FUNC (adev->pdev->devfn); - ret =amdgpu_bo_reserve (fpriv->vm.root.bo, false); + root =amdgpu_bo_ref (fpriv->vm.root.bo); +if (!root) +return ; + + ret =amdgpu_bo_reserve (root, false);if (ret) {DRM_ERROR ("Fail to reserve bo\n" );return ; }amdgpu_vm_get_memory (&fpriv->vm, &vram_mem, >t_mem, &cpu_mem); -amdgpu_bo_unreserve (fpriv->vm.root.bo); +amdgpu_bo_unreserve (root); +amdgpu_bo_unref (&root); +seq_printf (m,"pdev:\t%04x:%02x:%02x.%d\npasid:\t%u\n" , domain, bus, dev, fn, fpriv->vm.pasid);seq_printf (m,"vram mem:\t%llu kB\n" , vram_mem/1024UL );