The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com --- drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT; - dev->dev = device_create(vduse_class, NULL, - MKDEV(MAJOR(vduse_major), dev->minor), - dev, "%s", config->name); + dev->dev = device_create_with_groups(vduse_class, NULL, + MKDEV(MAJOR(vduse_major), dev->minor), + dev, vduse_dev_groups, "%s", config->name); if (IS_ERR(dev->dev)) { ret = PTR_ERR(dev->dev); goto err_dev; @@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode; - vduse_class->dev_groups = vduse_dev_groups;
ret = alloc_chrdev_region(&vduse_major, 0, VDUSE_DEV_MAX, "vduse"); if (ret)
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config, dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
- dev->dev = device_create(vduse_class, NULL,
MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
- dev->dev = device_create_with_groups(vduse_class, NULL,
MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name);@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class); vduse_class->devnode = vduse_devnode;
- vduse_class->dev_groups = vduse_dev_groups;
Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples: - empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok. - __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design. - no Documentation/ABI/ entries for the sysfs files here. I think it's burried in some other documentation file but that's not the correct place for it and if you run scripts/get_abi.pl with the code loaded it will rightly complain about this.
Do you want to address these, or do you want patches for them?
thanks,
greg k-h
On Tue, Apr 26, 2022 at 4:07 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
dev->dev = device_create(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
dev->dev = device_create_with_groups(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name); if (IS_ERR(dev->dev)) { ret = PTR_ERR(dev->dev); goto err_dev;@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode;
vduse_class->dev_groups = vduse_dev_groups;Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples: - empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok.
Sorry, I failed to find the documentation. Do you mean we should remove the empty release() callbacks?
- __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design.
I don't find a good way to remove it. We have to make sure the module can't be removed until all vduse devices are destroyed.
And I think __module_get(THIS_MODULE) should be safe in our case since we always call it when we have a reference from open().
- no Documentation/ABI/ entries for the sysfs files here. I think it's burried in some other documentation file but that's not the correct place for it and if you run scripts/get_abi.pl with the code loaded it will rightly complain about this.
OK, I will add one.
Do you want to address these, or do you want patches for them?
Let me send some individual patches for them.
Thanks, Yongji
On Tue, Apr 26, 2022 at 05:41:15PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 4:07 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
dev->dev = device_create(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
dev->dev = device_create_with_groups(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name); if (IS_ERR(dev->dev)) { ret = PTR_ERR(dev->dev); goto err_dev;@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode;
vduse_class->dev_groups = vduse_dev_groups;Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples: - empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok.
Sorry, I failed to find the documentation. Do you mean we should remove the empty release() callbacks?
Yes, why are they needed?
(hint, retorical question, you added them to remove the driver core warning when the device is removed, which means someone added them just because they thought that their code could ignore the hints that the driver core was telling them.)
Please properly free the memory here.
- __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design.I don't find a good way to remove it. We have to make sure the module can't be removed until all vduse devices are destroyed.
That will happen automatically when the module is removed.
And I think __module_get(THIS_MODULE) should be safe in our case since we always call it when we have a reference from open().
What happened if someone removed the module _right before_ this was called? You can not grab your own reference count safely.
Please just remove it, it's not needed and is broken. There should not be any reason that the module can not be unloaded, UNLESS a file handle is open, and you properly handle that already.
thanks,
greg k-h
On Tue, Apr 26, 2022 at 6:26 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 05:41:15PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 4:07 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
dev->dev = device_create(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
dev->dev = device_create_with_groups(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name); if (IS_ERR(dev->dev)) { ret = PTR_ERR(dev->dev); goto err_dev;@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode;
vduse_class->dev_groups = vduse_dev_groups;Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples: - empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok.
Sorry, I failed to find the documentation. Do you mean we should remove the empty release() callbacks?
Yes, why are they needed?
(hint, retorical question, you added them to remove the driver core warning when the device is removed, which means someone added them just because they thought that their code could ignore the hints that the driver core was telling them.)
OK, I see.
Please properly free the memory here.
One question is how to deal with the case if the device/kobject is defined as a static variable. We should not need to free any resources in this case. Or do you suggest just using dynamic allocation here?
- __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design.I don't find a good way to remove it. We have to make sure the module can't be removed until all vduse devices are destroyed.
That will happen automatically when the module is removed.
And I think __module_get(THIS_MODULE) should be safe in our case since we always call it when we have a reference from open().
What happened if someone removed the module _right before_ this was called? You can not grab your own reference count safely.
I don't get you here. We should already grab a reference count from open() before calling this. So it should fail if someone tries to remove the module at this time.
Please just remove it, it's not needed and is broken. There should not be any reason that the module can not be unloaded, UNLESS a file handle is open, and you properly handle that already.
But in our case, I think we should prevent unloading the module If we already created some vduse devices via /dev/vduse/control (note that the control device's file handle could be closed after device creation). Otherwise, we might get some crashes when accessing those created vduse devices.
Thanks, Yongji
On Tue, Apr 26, 2022 at 08:30:38PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 6:26 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 05:41:15PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 4:07 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
dev->dev = device_create(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
dev->dev = device_create_with_groups(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name); if (IS_ERR(dev->dev)) { ret = PTR_ERR(dev->dev); goto err_dev;@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode;
vduse_class->dev_groups = vduse_dev_groups;Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples: - empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok.
Sorry, I failed to find the documentation. Do you mean we should remove the empty release() callbacks?
Yes, why are they needed?
(hint, retorical question, you added them to remove the driver core warning when the device is removed, which means someone added them just because they thought that their code could ignore the hints that the driver core was telling them.)
OK, I see.
Please properly free the memory here.
One question is how to deal with the case if the device/kobject is defined as a static variable. We should not need to free any resources in this case. Or do you suggest just using dynamic allocation here?
A kobject can NEVER be a static variable[1]. That's not how the driver model works at all. If this is how this code is written, it needs to be fixed.
[1] Ok, yes, drivers and busses and classes have static kobjects, ignore them...
- __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design.I don't find a good way to remove it. We have to make sure the module can't be removed until all vduse devices are destroyed.
That will happen automatically when the module is removed.
And I think __module_get(THIS_MODULE) should be safe in our case since we always call it when we have a reference from open().
What happened if someone removed the module _right before_ this was called? You can not grab your own reference count safely.
I don't get you here. We should already grab a reference count from open() before calling this. So it should fail if someone tries to remove the module at this time.
Then why are you trying to grab the module reference again?
Please just remove it, it's not needed and is broken. There should not be any reason that the module can not be unloaded, UNLESS a file handle is open, and you properly handle that already.
But in our case, I think we should prevent unloading the module If we already created some vduse devices via /dev/vduse/control (note that the control device's file handle could be closed after device creation). Otherwise, we might get some crashes when accessing those created vduse devices.
Then the code is written incorrectly, this should not be an issue at all. Your devices will all be cleaned up properly before your code is unloaded from the system.
Note that no other driver or bus does this, what makes this different?
thanks,
greg k-h
On Tue, Apr 26, 2022 at 8:44 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 08:30:38PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 6:26 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 05:41:15PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 4:07 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
dev->dev = device_create(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
dev->dev = device_create_with_groups(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name); if (IS_ERR(dev->dev)) { ret = PTR_ERR(dev->dev); goto err_dev;@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode;
vduse_class->dev_groups = vduse_dev_groups;Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples: - empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok.
Sorry, I failed to find the documentation. Do you mean we should remove the empty release() callbacks?
Yes, why are they needed?
(hint, retorical question, you added them to remove the driver core warning when the device is removed, which means someone added them just because they thought that their code could ignore the hints that the driver core was telling them.)
OK, I see.
Please properly free the memory here.
One question is how to deal with the case if the device/kobject is defined as a static variable. We should not need to free any resources in this case. Or do you suggest just using dynamic allocation here?
A kobject can NEVER be a static variable[1]. That's not how the driver model works at all. If this is how this code is written, it needs to be fixed.
OK, I see.
[1] Ok, yes, drivers and busses and classes have static kobjects, ignore them...
- __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design.I don't find a good way to remove it. We have to make sure the module can't be removed until all vduse devices are destroyed.
That will happen automatically when the module is removed.
And I think __module_get(THIS_MODULE) should be safe in our case since we always call it when we have a reference from open().
What happened if someone removed the module _right before_ this was called? You can not grab your own reference count safely.
I don't get you here. We should already grab a reference count from open() before calling this. So it should fail if someone tries to remove the module at this time.
Then why are you trying to grab the module reference again?
Please just remove it, it's not needed and is broken. There should not be any reason that the module can not be unloaded, UNLESS a file handle is open, and you properly handle that already.
But in our case, I think we should prevent unloading the module If we already created some vduse devices via /dev/vduse/control (note that the control device's file handle could be closed after device creation). Otherwise, we might get some crashes when accessing those created vduse devices.
Then the code is written incorrectly, this should not be an issue at all. Your devices will all be cleaned up properly before your code is unloaded from the system.
In current design, the vduse device can't be cleaned up properly until it is unbinded from the vDPA bus explicitly. So I use the extra __module_get() to make sure we can't unload the module until the device is cleaned up properly.
Note that no other driver or bus does this, what makes this different?
I can see some similar behavior in loop and rbd modules.
Thanks, Yongji
On Tue, Apr 26, 2022 at 09:17:23PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 8:44 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 08:30:38PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 6:26 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 05:41:15PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 4:07 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote: > The control device has no drvdata. So we will get a > NULL pointer dereference when accessing control > device's msg_timeout attribute via sysfs: > > [ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 > [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) > [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) > [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) > [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) > [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) > [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) > [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) > [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) > [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) > [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) > [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) > [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) > [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) > [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) > [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) > [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419) > > To fix it, don't create the unneeded attribute for > control device anymore. > > Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") > Reported-by: kernel test robot oliver.sang@intel.com > Cc: stable@vger.kernel.org > Signed-off-by: Xie Yongji xieyongji@bytedance.com > --- > drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- > 1 file changed, 3 insertions(+), 4 deletions(-) > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c > index f85d1a08ed87..160e40d03084 100644 > --- a/drivers/vdpa/vdpa_user/vduse_dev.c > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c > @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config, > > dev->minor = ret; > dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT; > - dev->dev = device_create(vduse_class, NULL, > - MKDEV(MAJOR(vduse_major), dev->minor), > - dev, "%s", config->name); > + dev->dev = device_create_with_groups(vduse_class, NULL, > + MKDEV(MAJOR(vduse_major), dev->minor), > + dev, vduse_dev_groups, "%s", config->name); > if (IS_ERR(dev->dev)) { > ret = PTR_ERR(dev->dev); > goto err_dev; > @@ -1595,7 +1595,6 @@ static int vduse_init(void) > return PTR_ERR(vduse_class); > > vduse_class->devnode = vduse_devnode; > - vduse_class->dev_groups = vduse_dev_groups;
Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples: - empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok.
Sorry, I failed to find the documentation. Do you mean we should remove the empty release() callbacks?
Yes, why are they needed?
(hint, retorical question, you added them to remove the driver core warning when the device is removed, which means someone added them just because they thought that their code could ignore the hints that the driver core was telling them.)
OK, I see.
Please properly free the memory here.
One question is how to deal with the case if the device/kobject is defined as a static variable. We should not need to free any resources in this case. Or do you suggest just using dynamic allocation here?
A kobject can NEVER be a static variable[1]. That's not how the driver model works at all. If this is how this code is written, it needs to be fixed.
OK, I see.
[1] Ok, yes, drivers and busses and classes have static kobjects, ignore them...
- __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design.I don't find a good way to remove it. We have to make sure the module can't be removed until all vduse devices are destroyed.
That will happen automatically when the module is removed.
And I think __module_get(THIS_MODULE) should be safe in our case since we always call it when we have a reference from open().
What happened if someone removed the module _right before_ this was called? You can not grab your own reference count safely.
I don't get you here. We should already grab a reference count from open() before calling this. So it should fail if someone tries to remove the module at this time.
Then why are you trying to grab the module reference again?
Please just remove it, it's not needed and is broken. There should not be any reason that the module can not be unloaded, UNLESS a file handle is open, and you properly handle that already.
But in our case, I think we should prevent unloading the module If we already created some vduse devices via /dev/vduse/control (note that the control device's file handle could be closed after device creation). Otherwise, we might get some crashes when accessing those created vduse devices.
Then the code is written incorrectly, this should not be an issue at all. Your devices will all be cleaned up properly before your code is unloaded from the system.
In current design, the vduse device can't be cleaned up properly until it is unbinded from the vDPA bus explicitly. So I use the extra __module_get() to make sure we can't unload the module until the device is cleaned up properly.
Then something is wrong, it should not work that way.
Note that no other driver or bus does this, what makes this different?
I can see some similar behavior in loop and rbd modules.
Never treat the loop code as a good example :)
This should not be needed, when your module is unloaded, all devices it handled should be properly removed by it.
thanks,
greg k-h
On Tue, Apr 26, 2022 at 9:34 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 09:17:23PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 8:44 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 08:30:38PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 6:26 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 05:41:15PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 4:07 PM Greg KH gregkh@linuxfoundation.org wrote: > > On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote: > > The control device has no drvdata. So we will get a > > NULL pointer dereference when accessing control > > device's msg_timeout attribute via sysfs: > > > > [ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 > > [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) > > [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) > > [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) > > [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) > > [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) > > [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) > > [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) > > [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) > > [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) > > [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) > > [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) > > [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) > > [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) > > [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) > > [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) > > [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419) > > > > To fix it, don't create the unneeded attribute for > > control device anymore. > > > > Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") > > Reported-by: kernel test robot oliver.sang@intel.com > > Cc: stable@vger.kernel.org > > Signed-off-by: Xie Yongji xieyongji@bytedance.com > > --- > > drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- > > 1 file changed, 3 insertions(+), 4 deletions(-) > > > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c > > index f85d1a08ed87..160e40d03084 100644 > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c > > @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config, > > > > dev->minor = ret; > > dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT; > > - dev->dev = device_create(vduse_class, NULL, > > - MKDEV(MAJOR(vduse_major), dev->minor), > > - dev, "%s", config->name); > > + dev->dev = device_create_with_groups(vduse_class, NULL, > > + MKDEV(MAJOR(vduse_major), dev->minor), > > + dev, vduse_dev_groups, "%s", config->name); > > if (IS_ERR(dev->dev)) { > > ret = PTR_ERR(dev->dev); > > goto err_dev; > > @@ -1595,7 +1595,6 @@ static int vduse_init(void) > > return PTR_ERR(vduse_class); > > > > vduse_class->devnode = vduse_devnode; > > - vduse_class->dev_groups = vduse_dev_groups; > > Ok, this looks much better. > > But wow, there are some problems in this code overall. I see a number > of flat-out-wrong things in there that should have been caught by code > reviews. Some examples: > - empty release() callbacks. That is a huge sign the code > design is wrong and broken and you are just trying to make the > driver core quiet for some reason. The documentation in the > kernel explains why this is not ok.
Sorry, I failed to find the documentation. Do you mean we should remove the empty release() callbacks?
Yes, why are they needed?
(hint, retorical question, you added them to remove the driver core warning when the device is removed, which means someone added them just because they thought that their code could ignore the hints that the driver core was telling them.)
OK, I see.
Please properly free the memory here.
One question is how to deal with the case if the device/kobject is defined as a static variable. We should not need to free any resources in this case. Or do you suggest just using dynamic allocation here?
A kobject can NEVER be a static variable[1]. That's not how the driver model works at all. If this is how this code is written, it needs to be fixed.
OK, I see.
[1] Ok, yes, drivers and busses and classes have static kobjects, ignore them...
> - __module_get(THIS_MODULE); That's racy, buggy, and doesn't do > what you think it does. Please never ever ever do that. It > too is a sign of a broken design.
I don't find a good way to remove it. We have to make sure the module can't be removed until all vduse devices are destroyed.
That will happen automatically when the module is removed.
And I think __module_get(THIS_MODULE) should be safe in our case since we always call it when we have a reference from open().
What happened if someone removed the module _right before_ this was called? You can not grab your own reference count safely.
I don't get you here. We should already grab a reference count from open() before calling this. So it should fail if someone tries to remove the module at this time.
Then why are you trying to grab the module reference again?
Please just remove it, it's not needed and is broken. There should not be any reason that the module can not be unloaded, UNLESS a file handle is open, and you properly handle that already.
But in our case, I think we should prevent unloading the module If we already created some vduse devices via /dev/vduse/control (note that the control device's file handle could be closed after device creation). Otherwise, we might get some crashes when accessing those created vduse devices.
Then the code is written incorrectly, this should not be an issue at all. Your devices will all be cleaned up properly before your code is unloaded from the system.
In current design, the vduse device can't be cleaned up properly until it is unbinded from the vDPA bus explicitly. So I use the extra __module_get() to make sure we can't unload the module until the device is cleaned up properly.
Then something is wrong, it should not work that way.
Note that no other driver or bus does this, what makes this different?
I can see some similar behavior in loop and rbd modules.
Never treat the loop code as a good example :)
OK.
This should not be needed, when your module is unloaded, all devices it handled should be properly removed by it.
I see. But it's not easy to achieve that currently. Maybe we need something like DEVICE_NEEDS_RESET support in virtio core.
Thanks, Yongji
On Tue, Apr 26, 2022 at 10:02:02PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 9:34 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 09:17:23PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 8:44 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 08:30:38PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 6:26 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 05:41:15PM +0800, Yongji Xie wrote: > On Tue, Apr 26, 2022 at 4:07 PM Greg KH gregkh@linuxfoundation.org wrote: > > > > On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote: > > > The control device has no drvdata. So we will get a > > > NULL pointer dereference when accessing control > > > device's msg_timeout attribute via sysfs: > > > > > > [ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 > > > [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) > > > [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) > > > [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) > > > [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) > > > [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) > > > [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) > > > [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) > > > [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) > > > [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) > > > [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) > > > [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) > > > [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) > > > [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) > > > [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) > > > [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) > > > [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419) > > > > > > To fix it, don't create the unneeded attribute for > > > control device anymore. > > > > > > Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") > > > Reported-by: kernel test robot oliver.sang@intel.com > > > Cc: stable@vger.kernel.org > > > Signed-off-by: Xie Yongji xieyongji@bytedance.com > > > --- > > > drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- > > > 1 file changed, 3 insertions(+), 4 deletions(-) > > > > > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c > > > index f85d1a08ed87..160e40d03084 100644 > > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c > > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c > > > @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config, > > > > > > dev->minor = ret; > > > dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT; > > > - dev->dev = device_create(vduse_class, NULL, > > > - MKDEV(MAJOR(vduse_major), dev->minor), > > > - dev, "%s", config->name); > > > + dev->dev = device_create_with_groups(vduse_class, NULL, > > > + MKDEV(MAJOR(vduse_major), dev->minor), > > > + dev, vduse_dev_groups, "%s", config->name); > > > if (IS_ERR(dev->dev)) { > > > ret = PTR_ERR(dev->dev); > > > goto err_dev; > > > @@ -1595,7 +1595,6 @@ static int vduse_init(void) > > > return PTR_ERR(vduse_class); > > > > > > vduse_class->devnode = vduse_devnode; > > > - vduse_class->dev_groups = vduse_dev_groups; > > > > Ok, this looks much better. > > > > But wow, there are some problems in this code overall. I see a number > > of flat-out-wrong things in there that should have been caught by code > > reviews. Some examples: > > - empty release() callbacks. That is a huge sign the code > > design is wrong and broken and you are just trying to make the > > driver core quiet for some reason. The documentation in the > > kernel explains why this is not ok. > > Sorry, I failed to find the documentation. Do you mean we should > remove the empty release() callbacks?
Yes, why are they needed?
(hint, retorical question, you added them to remove the driver core warning when the device is removed, which means someone added them just because they thought that their code could ignore the hints that the driver core was telling them.)
OK, I see.
Please properly free the memory here.
One question is how to deal with the case if the device/kobject is defined as a static variable. We should not need to free any resources in this case. Or do you suggest just using dynamic allocation here?
A kobject can NEVER be a static variable[1]. That's not how the driver model works at all. If this is how this code is written, it needs to be fixed.
OK, I see.
[1] Ok, yes, drivers and busses and classes have static kobjects, ignore them...
> > - __module_get(THIS_MODULE); That's racy, buggy, and doesn't do > > what you think it does. Please never ever ever do that. It > > too is a sign of a broken design. > > I don't find a good way to remove it. We have to make sure the module > can't be removed until all vduse devices are destroyed.
That will happen automatically when the module is removed.
> And I think __module_get(THIS_MODULE) should be safe in our case since > we always call it when we have a reference from open().
What happened if someone removed the module _right before_ this was called? You can not grab your own reference count safely.
I don't get you here. We should already grab a reference count from open() before calling this. So it should fail if someone tries to remove the module at this time.
Then why are you trying to grab the module reference again?
Please just remove it, it's not needed and is broken. There should not be any reason that the module can not be unloaded, UNLESS a file handle is open, and you properly handle that already.
But in our case, I think we should prevent unloading the module If we already created some vduse devices via /dev/vduse/control (note that the control device's file handle could be closed after device creation). Otherwise, we might get some crashes when accessing those created vduse devices.
Then the code is written incorrectly, this should not be an issue at all. Your devices will all be cleaned up properly before your code is unloaded from the system.
In current design, the vduse device can't be cleaned up properly until it is unbinded from the vDPA bus explicitly. So I use the extra __module_get() to make sure we can't unload the module until the device is cleaned up properly.
Then something is wrong, it should not work that way.
Note that no other driver or bus does this, what makes this different?
I can see some similar behavior in loop and rbd modules.
Never treat the loop code as a good example :)
OK.
This should not be needed, when your module is unloaded, all devices it handled should be properly removed by it.
I see. But it's not easy to achieve that currently. Maybe we need something like DEVICE_NEEDS_RESET support in virtio core.
This shouldn't be that hard, again, no other subsystem needs this. How about fix the other issues first and then we can work on that one.
thanks,
greg k-h
On Tue, Apr 26, 2022 at 10:09 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 10:02:02PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 9:34 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 09:17:23PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 8:44 PM Greg KH gregkh@linuxfoundation.org wrote:
On Tue, Apr 26, 2022 at 08:30:38PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 6:26 PM Greg KH gregkh@linuxfoundation.org wrote: > > On Tue, Apr 26, 2022 at 05:41:15PM +0800, Yongji Xie wrote: > > On Tue, Apr 26, 2022 at 4:07 PM Greg KH gregkh@linuxfoundation.org wrote: > > > > > > On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote: > > > > The control device has no drvdata. So we will get a > > > > NULL pointer dereference when accessing control > > > > device's msg_timeout attribute via sysfs: > > > > > > > > [ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 > > > > [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) > > > > [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) > > > > [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) > > > > [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) > > > > [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) > > > > [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) > > > > [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) > > > > [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) > > > > [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) > > > > [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) > > > > [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) > > > > [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) > > > > [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) > > > > [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) > > > > [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) > > > > [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419) > > > > > > > > To fix it, don't create the unneeded attribute for > > > > control device anymore. > > > > > > > > Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") > > > > Reported-by: kernel test robot oliver.sang@intel.com > > > > Cc: stable@vger.kernel.org > > > > Signed-off-by: Xie Yongji xieyongji@bytedance.com > > > > --- > > > > drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- > > > > 1 file changed, 3 insertions(+), 4 deletions(-) > > > > > > > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c > > > > index f85d1a08ed87..160e40d03084 100644 > > > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c > > > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c > > > > @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config, > > > > > > > > dev->minor = ret; > > > > dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT; > > > > - dev->dev = device_create(vduse_class, NULL, > > > > - MKDEV(MAJOR(vduse_major), dev->minor), > > > > - dev, "%s", config->name); > > > > + dev->dev = device_create_with_groups(vduse_class, NULL, > > > > + MKDEV(MAJOR(vduse_major), dev->minor), > > > > + dev, vduse_dev_groups, "%s", config->name); > > > > if (IS_ERR(dev->dev)) { > > > > ret = PTR_ERR(dev->dev); > > > > goto err_dev; > > > > @@ -1595,7 +1595,6 @@ static int vduse_init(void) > > > > return PTR_ERR(vduse_class); > > > > > > > > vduse_class->devnode = vduse_devnode; > > > > - vduse_class->dev_groups = vduse_dev_groups; > > > > > > Ok, this looks much better. > > > > > > But wow, there are some problems in this code overall. I see a number > > > of flat-out-wrong things in there that should have been caught by code > > > reviews. Some examples: > > > - empty release() callbacks. That is a huge sign the code > > > design is wrong and broken and you are just trying to make the > > > driver core quiet for some reason. The documentation in the > > > kernel explains why this is not ok. > > > > Sorry, I failed to find the documentation. Do you mean we should > > remove the empty release() callbacks? > > Yes, why are they needed? > > (hint, retorical question, you added them to remove the driver core > warning when the device is removed, which means someone added them just > because they thought that their code could ignore the hints that the > driver core was telling them.) >
OK, I see.
> Please properly free the memory here. >
One question is how to deal with the case if the device/kobject is defined as a static variable. We should not need to free any resources in this case. Or do you suggest just using dynamic allocation here?
A kobject can NEVER be a static variable[1]. That's not how the driver model works at all. If this is how this code is written, it needs to be fixed.
OK, I see.
[1] Ok, yes, drivers and busses and classes have static kobjects, ignore them...
> > > - __module_get(THIS_MODULE); That's racy, buggy, and doesn't do > > > what you think it does. Please never ever ever do that. It > > > too is a sign of a broken design. > > > > I don't find a good way to remove it. We have to make sure the module > > can't be removed until all vduse devices are destroyed. > > That will happen automatically when the module is removed. > > > And I think __module_get(THIS_MODULE) should be safe in our case since > > we always call it when we have a reference from open(). > > What happened if someone removed the module _right before_ this was > called? You can not grab your own reference count safely. >
I don't get you here. We should already grab a reference count from open() before calling this. So it should fail if someone tries to remove the module at this time.
Then why are you trying to grab the module reference again?
> Please just remove it, it's not needed and is broken. There should not > be any reason that the module can not be unloaded, UNLESS a file handle > is open, and you properly handle that already. >
But in our case, I think we should prevent unloading the module If we already created some vduse devices via /dev/vduse/control (note that the control device's file handle could be closed after device creation). Otherwise, we might get some crashes when accessing those created vduse devices.
Then the code is written incorrectly, this should not be an issue at all. Your devices will all be cleaned up properly before your code is unloaded from the system.
In current design, the vduse device can't be cleaned up properly until it is unbinded from the vDPA bus explicitly. So I use the extra __module_get() to make sure we can't unload the module until the device is cleaned up properly.
Then something is wrong, it should not work that way.
Note that no other driver or bus does this, what makes this different?
I can see some similar behavior in loop and rbd modules.
Never treat the loop code as a good example :)
OK.
This should not be needed, when your module is unloaded, all devices it handled should be properly removed by it.
I see. But it's not easy to achieve that currently. Maybe we need something like DEVICE_NEEDS_RESET support in virtio core.
This shouldn't be that hard, again, no other subsystem needs this. How about fix the other issues first and then we can work on that one.
Looks good to me.
Thanks, Yongji
On Tue, Apr 26, 2022 at 10:02:02PM +0800, Yongji Xie wrote:
This should not be needed, when your module is unloaded, all devices it handled should be properly removed by it.
I see. But it's not easy to achieve that currently. Maybe we need something like DEVICE_NEEDS_RESET support in virtio core.
Not sure what the connection is.
On Tue, Apr 26, 2022 at 10:18 PM Michael S. Tsirkin mst@redhat.com wrote:
On Tue, Apr 26, 2022 at 10:02:02PM +0800, Yongji Xie wrote:
This should not be needed, when your module is unloaded, all devices it handled should be properly removed by it.
I see. But it's not easy to achieve that currently. Maybe we need something like DEVICE_NEEDS_RESET support in virtio core.
Not sure what the connection is.
If we want to force remove all working vduse devices during module unload, we might need to send a DEVICE_NEEDS_RESET notification to device driver to do some cleanup before, e.g., return error for all inflight I/Os.
Thanks, Yongji
On Tue, Apr 26, 2022 at 10:37:17PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 10:18 PM Michael S. Tsirkin mst@redhat.com wrote:
On Tue, Apr 26, 2022 at 10:02:02PM +0800, Yongji Xie wrote:
This should not be needed, when your module is unloaded, all devices it handled should be properly removed by it.
I see. But it's not easy to achieve that currently. Maybe we need something like DEVICE_NEEDS_RESET support in virtio core.
Not sure what the connection is.
If we want to force remove all working vduse devices during module unload, we might need to send a DEVICE_NEEDS_RESET notification to device driver to do some cleanup before, e.g., return error for all inflight I/Os.
Thanks, Yongji
IMHO DEVICE_NEEDS_RESET won't help much with that, it's more in case device is still there but needs a reset to start working.
On Tue, Apr 26, 2022 at 10:38 PM Michael S. Tsirkin mst@redhat.com wrote:
On Tue, Apr 26, 2022 at 10:37:17PM +0800, Yongji Xie wrote:
On Tue, Apr 26, 2022 at 10:18 PM Michael S. Tsirkin mst@redhat.com wrote:
On Tue, Apr 26, 2022 at 10:02:02PM +0800, Yongji Xie wrote:
This should not be needed, when your module is unloaded, all devices it handled should be properly removed by it.
I see. But it's not easy to achieve that currently. Maybe we need something like DEVICE_NEEDS_RESET support in virtio core.
Not sure what the connection is.
If we want to force remove all working vduse devices during module unload, we might need to send a DEVICE_NEEDS_RESET notification to device driver to do some cleanup before, e.g., return error for all inflight I/Os.
Thanks, Yongji
IMHO DEVICE_NEEDS_RESET won't help much with that, it's more in case device is still there but needs a reset to start working.
OK, we just want a reset without restarting work in our case. Looks like we might need some other mechanism to achieve that.
Thanks, Yongji
On Tue, Apr 26, 2022 at 10:07:38AM +0200, Greg KH wrote:
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config, dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
- dev->dev = device_create(vduse_class, NULL,
MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
- dev->dev = device_create_with_groups(vduse_class, NULL,
MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name);@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class); vduse_class->devnode = vduse_devnode;
- vduse_class->dev_groups = vduse_dev_groups;
Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples:
- empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok.
- __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design.
- no Documentation/ABI/ entries for the sysfs files here. I think it's burried in some other documentation file but that's not the correct place for it and if you run scripts/get_abi.pl with the code loaded it will rightly complain about this.
Do you want to address these, or do you want patches for them?
thanks,
greg k-h
So, any patches?
On Wed, Jun 8, 2022 at 8:52 PM Michael S. Tsirkin mst@redhat.com wrote:
On Tue, Apr 26, 2022 at 10:07:38AM +0200, Greg KH wrote:
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
- dev->dev = device_create(vduse_class, NULL,
MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
- dev->dev = device_create_with_groups(vduse_class, NULL,
MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name);@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode;
- vduse_class->dev_groups = vduse_dev_groups;
Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples: - empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok. - __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design. - no Documentation/ABI/ entries for the sysfs files here. I think it's burried in some other documentation file but that's not the correct place for it and if you run scripts/get_abi.pl with the code loaded it will rightly complain about this.
Do you want to address these, or do you want patches for them?
thanks,
greg k-h
So, any patches?
For empty release() callbacks, I think Parav is working on it based on the discussion [1]. I can help test and send the patch if Parav wants.
For Documentation, I have sent a patch [2].
[1] https://www.spinics.net/lists/linux-virtualization/msg56518.html [2] https://lore.kernel.org/all/CACGkMEuJeU6c1z8+_FqGtovbF+Sq8w_eQUcG8SHm_GXV5q7...
Thanks, Yongji
From: Yongji Xie xieyongji@bytedance.com Sent: Wednesday, June 8, 2022 9:26 AM
On Wed, Jun 8, 2022 at 8:52 PM Michael S. Tsirkin mst@redhat.com wrote:
On Tue, Apr 26, 2022 at 10:07:38AM +0200, Greg KH wrote:
On Tue, Apr 26, 2022 at 03:36:56PM +0800, Xie Yongji wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
- dev->dev = device_create(vduse_class, NULL,
MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
- dev->dev = device_create_with_groups(vduse_class, NULL,
MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s",- config->name); if (IS_ERR(dev->dev)) { ret = PTR_ERR(dev->dev); goto err_dev;
@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode;
- vduse_class->dev_groups = vduse_dev_groups;
Ok, this looks much better.
But wow, there are some problems in this code overall. I see a number of flat-out-wrong things in there that should have been caught by code reviews. Some examples: - empty release() callbacks. That is a huge sign the code design is wrong and broken and you are just trying to make the driver core quiet for some reason. The documentation in the kernel explains why this is not ok. - __module_get(THIS_MODULE); That's racy, buggy, and doesn't do what you think it does. Please never ever ever do that. It too is a sign of a broken design. - no Documentation/ABI/ entries for the sysfs files here. I think it's burried in some other documentation file but that's not the correct place for it and if you run scripts/get_abi.pl with the code loaded it will rightly complain about this.
Do you want to address these, or do you want patches for them?
thanks,
greg k-h
So, any patches?
For empty release() callbacks, I think Parav is working on it based on the discussion [1]. I can help test and send the patch if Parav wants.
As there are no more comments, I will send the patch for vduse this week based on [1] which was previously in email form.
For Documentation, I have sent a patch [2].
[1] https://www.spinics.net/lists/linux-virtualization/msg56518.html [2] https://lore.kernel.org/all/CACGkMEuJeU6c1z8+_FqGtovbF+Sq8w_eQUcG8 SHm_GXV5q7yNA@mail.gmail.com/
Thanks, Yongji
Ping.
On Tue, Apr 26, 2022 at 3:36 PM Xie Yongji xieyongji@bytedance.com wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
dev->dev = device_create(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
dev->dev = device_create_with_groups(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name); if (IS_ERR(dev->dev)) { ret = PTR_ERR(dev->dev); goto err_dev;@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode;
vduse_class->dev_groups = vduse_dev_groups; ret = alloc_chrdev_region(&vduse_major, 0, VDUSE_DEV_MAX, "vduse"); if (ret)-- 2.20.1
On Thu, Jun 02, 2022 at 12:55:02PM +0800, Yongji Xie wrote:
Ping.
Thanks for the reminder! Will queue for rc2, rc1 has too much stuff already.
On Tue, Apr 26, 2022 at 3:36 PM Xie Yongji xieyongji@bytedance.com wrote:
The control device has no drvdata. So we will get a NULL pointer dereference when accessing control device's msg_timeout attribute via sysfs:
[ 132.841881][ T3644] BUG: kernel NULL pointer dereference, address: 00000000000000f8 [ 132.850619][ T3644] RIP: 0010:msg_timeout_show (drivers/vdpa/vdpa_user/vduse_dev.c:1271) [ 132.869447][ T3644] dev_attr_show (drivers/base/core.c:2094) [ 132.870215][ T3644] sysfs_kf_seq_show (fs/sysfs/file.c:59) [ 132.871164][ T3644] ? device_remove_bin_file (drivers/base/core.c:2088) [ 132.872082][ T3644] kernfs_seq_show (fs/kernfs/file.c:164) [ 132.872838][ T3644] seq_read_iter (fs/seq_file.c:230) [ 132.873578][ T3644] ? __vmalloc_area_node (mm/vmalloc.c:3041) [ 132.874532][ T3644] kernfs_fop_read_iter (fs/kernfs/file.c:238) [ 132.875513][ T3644] __kernel_read (fs/read_write.c:440 (discriminator 1)) [ 132.876319][ T3644] kernel_read (fs/read_write.c:459) [ 132.877129][ T3644] kernel_read_file (fs/kernel_read_file.c:94) [ 132.877978][ T3644] kernel_read_file_from_fd (include/linux/file.h:45 fs/kernel_read_file.c:186) [ 132.879019][ T3644] __do_sys_finit_module (kernel/module.c:4207) [ 132.879930][ T3644] __ia32_sys_finit_module (kernel/module.c:4189) [ 132.880930][ T3644] do_int80_syscall_32 (arch/x86/entry/common.c:112 arch/x86/entry/common.c:132) [ 132.881847][ T3644] entry_INT80_compat (arch/x86/entry/entry_64_compat.S:419)
To fix it, don't create the unneeded attribute for control device anymore.
Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace") Reported-by: kernel test robot oliver.sang@intel.com Cc: stable@vger.kernel.org Signed-off-by: Xie Yongji xieyongji@bytedance.com
drivers/vdpa/vdpa_user/vduse_dev.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c index f85d1a08ed87..160e40d03084 100644 --- a/drivers/vdpa/vdpa_user/vduse_dev.c +++ b/drivers/vdpa/vdpa_user/vduse_dev.c @@ -1344,9 +1344,9 @@ static int vduse_create_dev(struct vduse_dev_config *config,
dev->minor = ret; dev->msg_timeout = VDUSE_MSG_DEFAULT_TIMEOUT;
dev->dev = device_create(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, "%s", config->name);
dev->dev = device_create_with_groups(vduse_class, NULL,MKDEV(MAJOR(vduse_major), dev->minor),dev, vduse_dev_groups, "%s", config->name); if (IS_ERR(dev->dev)) { ret = PTR_ERR(dev->dev); goto err_dev;@@ -1595,7 +1595,6 @@ static int vduse_init(void) return PTR_ERR(vduse_class);
vduse_class->devnode = vduse_devnode;
vduse_class->dev_groups = vduse_dev_groups; ret = alloc_chrdev_region(&vduse_major, 0, VDUSE_DEV_MAX, "vduse"); if (ret)-- 2.20.1
linux-stable-mirror@lists.linaro.org
-
Greg KH -
Michael S. Tsirkin -
Parav Pandit -
Xie Yongji -
Yongji Xie