Smatch report warning as follows:
drivers/dma-buf/dma-buf.c:681 dma_buf_export() warn: '&dmabuf->list_node' not removed from list
If dma_buf_stats_setup() fails in dma_buf_export(), goto err_sysfs and dmabuf will be freed, but dmabuf->list_node will not be removed from db_list.head, then list traversal may cause UAF.
Fix by removeing it from db_list.head before free().
Fixes: ef3a6b70507a ("dma-buf: call dma_buf_stats_setup after dmabuf is in valid list") Signed-off-by: Gaosheng Cui cuigaosheng1@huawei.com --- drivers/dma-buf/dma-buf.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c index b809513b03fe..6848f50226d5 100644 --- a/drivers/dma-buf/dma-buf.c +++ b/drivers/dma-buf/dma-buf.c @@ -675,6 +675,9 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) return dmabuf;
err_sysfs: + mutex_lock(&db_list.lock); + list_del(&dmabuf->list_node); + mutex_unlock(&db_list.lock); /* * Set file->f_path.dentry->d_fsdata to NULL so that when * dma_buf_release() gets invoked by dentry_ops, it exits
Sometime back Dan also reported the same issue[1] where I do mentioned that fput()-->dma_buf_file_release() will remove it from the list.
But it seems that I failed to notice fput() here calls the dma_buf_file_release() asynchronously i.e. dmabuf that is accessed in the close path is already freed. Am I wrong here?
Should we have the __fput_sync(file) here instead of just fput(file) which can solve this problem?
[1]https://lore.kernel.org/all/20220516084704.GG29930@kadam/
Thanks, Charan On 11/17/2022 11:51 AM, Gaosheng Cui wrote:
Smatch report warning as follows:
drivers/dma-buf/dma-buf.c:681 dma_buf_export() warn: '&dmabuf->list_node' not removed from list
If dma_buf_stats_setup() fails in dma_buf_export(), goto err_sysfs and dmabuf will be freed, but dmabuf->list_node will not be removed from db_list.head, then list traversal may cause UAF.
Fix by removeing it from db_list.head before free().
Fixes: ef3a6b70507a ("dma-buf: call dma_buf_stats_setup after dmabuf is in valid list") Signed-off-by: Gaosheng Cui cuigaosheng1@huawei.com
drivers/dma-buf/dma-buf.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c index b809513b03fe..6848f50226d5 100644 --- a/drivers/dma-buf/dma-buf.c +++ b/drivers/dma-buf/dma-buf.c @@ -675,6 +675,9 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) return dmabuf; err_sysfs:
- mutex_lock(&db_list.lock);
- list_del(&dmabuf->list_node);
- mutex_unlock(&db_list.lock); /*
- Set file->f_path.dentry->d_fsdata to NULL so that when
- dma_buf_release() gets invoked by dentry_ops, it exits
Am 17.11.22 um 08:48 schrieb Charan Teja Kalla:
Sometime back Dan also reported the same issue[1] where I do mentioned that fput()-->dma_buf_file_release() will remove it from the list.
But it seems that I failed to notice fput() here calls the dma_buf_file_release() asynchronously i.e. dmabuf that is accessed in the close path is already freed. Am I wrong here?
Should we have the __fput_sync(file) here instead of just fput(file) which can solve this problem?
[1]https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kerne...
That doesn't look like the right solution to me either.
Essentially we have two separate tear down methods for the dma_buf object here:
1. It's not completely initialized and we can call kfree()+module_put() to clean up. There is actually a dma_resv_fini() here. That should probably be fixed.
2. The dma_buf object is fully initialized, but creating the sysfs stats file failed. In this case we should *not* clean it up like we currently do, but rather call fput().
So the right thing to do is a) fix the missing dma_resv_fini() call and b) drop the setting d_fsdata=NULL hack and properly return after the fput().
Regards, Christian.
Thanks, Charan On 11/17/2022 11:51 AM, Gaosheng Cui wrote:
Smatch report warning as follows:
drivers/dma-buf/dma-buf.c:681 dma_buf_export() warn: '&dmabuf->list_node' not removed from list
If dma_buf_stats_setup() fails in dma_buf_export(), goto err_sysfs and dmabuf will be freed, but dmabuf->list_node will not be removed from db_list.head, then list traversal may cause UAF.
Fix by removeing it from db_list.head before free().
Fixes: ef3a6b70507a ("dma-buf: call dma_buf_stats_setup after dmabuf is in valid list") Signed-off-by: Gaosheng Cui cuigaosheng1@huawei.com
drivers/dma-buf/dma-buf.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c index b809513b03fe..6848f50226d5 100644 --- a/drivers/dma-buf/dma-buf.c +++ b/drivers/dma-buf/dma-buf.c @@ -675,6 +675,9 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) return dmabuf; err_sysfs:
- mutex_lock(&db_list.lock);
- list_del(&dmabuf->list_node);
- mutex_unlock(&db_list.lock); /*
- Set file->f_path.dentry->d_fsdata to NULL so that when
- dma_buf_release() gets invoked by dentry_ops, it exits
On Thu, Nov 17, 2022 at 2:16 AM Christian König christian.koenig@amd.com wrote:
Am 17.11.22 um 08:48 schrieb Charan Teja Kalla:
Sometime back Dan also reported the same issue[1] where I do mentioned that fput()-->dma_buf_file_release() will remove it from the list.
But it seems that I failed to notice fput() here calls the dma_buf_file_release() asynchronously i.e. dmabuf that is accessed in the close path is already freed. Am I wrong here?
Should we have the __fput_sync(file) here instead of just fput(file) which can solve this problem?
[1]https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kerne...
That doesn't look like the right solution to me either.
Essentially we have two separate tear down methods for the dma_buf object here:
- It's not completely initialized and we can call kfree()+module_put()
to clean up. There is actually a dma_resv_fini() here. That should probably be fixed.
- The dma_buf object is fully initialized, but creating the sysfs stats
file failed. In this case we should *not* clean it up like we currently do, but rather call fput().
So the right thing to do is a) fix the missing dma_resv_fini() call and b) drop the setting d_fsdata=NULL hack and properly return after the fput().
This looks right to me if by properly return you mean return ERR_PTR(ret); at the end of err_sysfs after the fput. (letting dma_buf_file_release and dma_buf_release do the full cleanup)
If we still want to avoid calling dmabuf->ops->release(dmabuf) in dma_buf_release like the comment says I guess we could use sysfs_entry and ERR_PTR to flag that, otherwise it looks like we'd need a bit somewhere.
Regards, Christian.
Thanks, Charan On 11/17/2022 11:51 AM, Gaosheng Cui wrote:
Smatch report warning as follows:
drivers/dma-buf/dma-buf.c:681 dma_buf_export() warn: '&dmabuf->list_node' not removed from list
If dma_buf_stats_setup() fails in dma_buf_export(), goto err_sysfs and dmabuf will be freed, but dmabuf->list_node will not be removed from db_list.head, then list traversal may cause UAF.
Fix by removeing it from db_list.head before free().
Fixes: ef3a6b70507a ("dma-buf: call dma_buf_stats_setup after dmabuf is in valid list") Signed-off-by: Gaosheng Cui cuigaosheng1@huawei.com
drivers/dma-buf/dma-buf.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c index b809513b03fe..6848f50226d5 100644 --- a/drivers/dma-buf/dma-buf.c +++ b/drivers/dma-buf/dma-buf.c @@ -675,6 +675,9 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) return dmabuf;
err_sysfs:
- mutex_lock(&db_list.lock);
- list_del(&dmabuf->list_node);
- mutex_unlock(&db_list.lock); /*
- Set file->f_path.dentry->d_fsdata to NULL so that when
- dma_buf_release() gets invoked by dentry_ops, it exits
Am 18.11.22 um 03:36 schrieb T.J. Mercier:
On Thu, Nov 17, 2022 at 2:16 AM Christian König christian.koenig@amd.com wrote:
Am 17.11.22 um 08:48 schrieb Charan Teja Kalla:
Sometime back Dan also reported the same issue[1] where I do mentioned that fput()-->dma_buf_file_release() will remove it from the list.
But it seems that I failed to notice fput() here calls the dma_buf_file_release() asynchronously i.e. dmabuf that is accessed in the close path is already freed. Am I wrong here?
Should we have the __fput_sync(file) here instead of just fput(file) which can solve this problem?
[1]https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kerne...
That doesn't look like the right solution to me either.
Essentially we have two separate tear down methods for the dma_buf object here:
- It's not completely initialized and we can call kfree()+module_put()
to clean up. There is actually a dma_resv_fini() here. That should probably be fixed.
- The dma_buf object is fully initialized, but creating the sysfs stats
file failed. In this case we should *not* clean it up like we currently do, but rather call fput().
So the right thing to do is a) fix the missing dma_resv_fini() call and b) drop the setting d_fsdata=NULL hack and properly return after the fput().
This looks right to me if by properly return you mean return ERR_PTR(ret); at the end of err_sysfs after the fput. (letting dma_buf_file_release and dma_buf_release do the full cleanup)
Yes, exactly that's the idea.
The only alternatives I can see would be to either move allocating the file and so completing the dma_buf initialization last again or just ignore errors from sysfs.
If we still want to avoid calling dmabuf->ops->release(dmabuf) in dma_buf_release like the comment says I guess we could use sysfs_entry and ERR_PTR to flag that, otherwise it looks like we'd need a bit somewhere.
No, this should be dropped as far as I can see. The sysfs cleanup code looks like it can handle not initialized kobj pointers.
Regards, Christian.
Regards, Christian.
Thanks, Charan On 11/17/2022 11:51 AM, Gaosheng Cui wrote:
Smatch report warning as follows:
drivers/dma-buf/dma-buf.c:681 dma_buf_export() warn: '&dmabuf->list_node' not removed from list
If dma_buf_stats_setup() fails in dma_buf_export(), goto err_sysfs and dmabuf will be freed, but dmabuf->list_node will not be removed from db_list.head, then list traversal may cause UAF.
Fix by removeing it from db_list.head before free().
Fixes: ef3a6b70507a ("dma-buf: call dma_buf_stats_setup after dmabuf is in valid list") Signed-off-by: Gaosheng Cui cuigaosheng1@huawei.com
drivers/dma-buf/dma-buf.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c index b809513b03fe..6848f50226d5 100644 --- a/drivers/dma-buf/dma-buf.c +++ b/drivers/dma-buf/dma-buf.c @@ -675,6 +675,9 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) return dmabuf;
err_sysfs:
- mutex_lock(&db_list.lock);
- list_del(&dmabuf->list_node);
- mutex_unlock(&db_list.lock); /*
- Set file->f_path.dentry->d_fsdata to NULL so that when
- dma_buf_release() gets invoked by dentry_ops, it exits
On Fri, Nov 18, 2022 at 12:27 AM Christian König christian.koenig@amd.com wrote:
Am 18.11.22 um 03:36 schrieb T.J. Mercier:
On Thu, Nov 17, 2022 at 2:16 AM Christian König christian.koenig@amd.com wrote:
Am 17.11.22 um 08:48 schrieb Charan Teja Kalla:
Sometime back Dan also reported the same issue[1] where I do mentioned that fput()-->dma_buf_file_release() will remove it from the list.
But it seems that I failed to notice fput() here calls the dma_buf_file_release() asynchronously i.e. dmabuf that is accessed in the close path is already freed. Am I wrong here?
Should we have the __fput_sync(file) here instead of just fput(file) which can solve this problem?
[1]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kerne...
That doesn't look like the right solution to me either.
Essentially we have two separate tear down methods for the dma_buf object here:
- It's not completely initialized and we can call kfree()+module_put()
to clean up. There is actually a dma_resv_fini() here. That should probably be fixed.
- The dma_buf object is fully initialized, but creating the sysfs stats
file failed. In this case we should *not* clean it up like we currently do, but rather call fput().
So the right thing to do is a) fix the missing dma_resv_fini() call and b) drop the setting d_fsdata=NULL hack and properly return after the
fput().
This looks right to me if by properly return you mean return ERR_PTR(ret); at the end of err_sysfs after the fput. (letting dma_buf_file_release and dma_buf_release do the full cleanup)
Yes, exactly that's the idea.
The only alternatives I can see would be to either move allocating the file and so completing the dma_buf initialization last again or just ignore errors from sysfs.
If we still want to avoid calling dmabuf->ops->release(dmabuf) in dma_buf_release like the comment says I guess we could use sysfs_entry and ERR_PTR to flag that, otherwise it looks like we'd need a bit somewhere.
No, this should be dropped as far as I can see. The sysfs cleanup code looks like it can handle not initialized kobj pointers.
Yeah there is also the null check in dma_buf_stats_teardown() that would prevent it from running, but I understood the comment to be referring to the release() dma_buf_ops call into the exporter which comes right after the teardown call. That looks like it's preventing the fput task work calling back into the exporter after the exporter already got an error from dma_buf_export(). Otherwise the exporter sees a release() for a buffer that it doesn't know about / thinks shouldn't exist. So I could imagine an exporter trying to double free: once for the failed dma_buf_export() call, and again when the release() op is called later.
Regards, Christian.
Regards, Christian.
Thanks, Charan On 11/17/2022 11:51 AM, Gaosheng Cui wrote:
Smatch report warning as follows:
drivers/dma-buf/dma-buf.c:681 dma_buf_export() warn: '&dmabuf->list_node' not removed from list
If dma_buf_stats_setup() fails in dma_buf_export(), goto err_sysfs and dmabuf will be freed, but dmabuf->list_node will not be removed from db_list.head, then list traversal may cause UAF.
Fix by removeing it from db_list.head before free().
Fixes: ef3a6b70507a ("dma-buf: call dma_buf_stats_setup after dmabuf
is in valid list")
Signed-off-by: Gaosheng Cui cuigaosheng1@huawei.com
drivers/dma-buf/dma-buf.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c index b809513b03fe..6848f50226d5 100644 --- a/drivers/dma-buf/dma-buf.c +++ b/drivers/dma-buf/dma-buf.c @@ -675,6 +675,9 @@ struct dma_buf *dma_buf_export(const struct
dma_buf_export_info *exp_info)
return dmabuf;err_sysfs:
- mutex_lock(&db_list.lock);
- list_del(&dmabuf->list_node);
- mutex_unlock(&db_list.lock); /*
- Set file->f_path.dentry->d_fsdata to NULL so that when
- dma_buf_release() gets invoked by dentry_ops, it exits
Am 18.11.22 um 18:05 schrieb T.J. Mercier:
On Fri, Nov 18, 2022 at 12:27 AM Christian König christian.koenig@amd.com wrote:
Am 18.11.22 um 03:36 schrieb T.J. Mercier: > On Thu, Nov 17, 2022 at 2:16 AM Christian König > <christian.koenig@amd.com> wrote: >> Am 17.11.22 um 08:48 schrieb Charan Teja Kalla: >>> Sometime back Dan also reported the same issue[1] where I do mentioned >>> that fput()-->dma_buf_file_release() will remove it from the list. >>> >>> But it seems that I failed to notice fput() here calls the >>> dma_buf_file_release() asynchronously i.e. dmabuf that is accessed in >>> the close path is already freed. Am I wrong here? >>> >>> Should we have the __fput_sync(file) here instead of just fput(file) >>> which can solve this problem? >>> >>> [1]https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fall%2F20220516084704.GG29930%40kadam%2F&data=05%7C01%7Cchristian.koenig%40amd.com%7C7d87a302d300479ecfa608dac90dc9f4%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C638043358319479671%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=erPl1hGdfLbfCxK3J3xiIR9boJbgj6hPUnCBvZFobog%3D&reserved=0 <https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fall%2F20220516084704.GG29930%40kadam%2F&data=05%7C01%7Cchristian.koenig%40amd.com%7Ca7d0f99c9d8440a4c6d208dac98717cd%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C638043879326367851%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=V38k91cCy446dyGpmRQWTNn5U8u2zSfCAgIRmACukq4%3D&reserved=0> >> That doesn't look like the right solution to me either. >> >> Essentially we have two separate tear down methods for the dma_buf >> object here: >> >> 1. It's not completely initialized and we can call kfree()+module_put() >> to clean up. >> There is actually a dma_resv_fini() here. That should probably be >> fixed. >> >> 2. The dma_buf object is fully initialized, but creating the sysfs stats >> file failed. >> In this case we should *not* clean it up like we currently do, but >> rather call fput(). >> >> So the right thing to do is a) fix the missing dma_resv_fini() call and >> b) drop the setting d_fsdata=NULL hack and properly return after the fput(). >> > This looks right to me if by properly return you mean return > ERR_PTR(ret); at the end of err_sysfs after the fput. (letting > dma_buf_file_release and dma_buf_release do the full cleanup) Yes, exactly that's the idea. The only alternatives I can see would be to either move allocating the file and so completing the dma_buf initialization last again or just ignore errors from sysfs. > If we still want to avoid calling dmabuf->ops->release(dmabuf) in > dma_buf_release like the comment says I guess we could use sysfs_entry > and ERR_PTR to flag that, otherwise it looks like we'd need a bit > somewhere. No, this should be dropped as far as I can see. The sysfs cleanup code looks like it can handle not initialized kobj pointers.Yeah there is also the null check in dma_buf_stats_teardown() that would prevent it from running, but I understood the comment to be referring to the release() dma_buf_ops call into the exporter which comes right after the teardown call. That looks like it's preventing the fput task work calling back into the exporter after the exporter already got an error from dma_buf_export(). Otherwise the exporter sees a release() for a buffer that it doesn't know about / thinks shouldn't exist. So I could imagine an exporter trying to double free: once for the failed dma_buf_export() call, and again when the release() op is called later.
Oh, very good point as well. Yeah, then creating the file should probably come last.
Regards, Christian.
Regards, Christian. > > > >> Regards, >> Christian. >> >>> Thanks, >>> Charan >>> On 11/17/2022 11:51 AM, Gaosheng Cui wrote: >>>> Smatch report warning as follows: >>>> >>>> drivers/dma-buf/dma-buf.c:681 dma_buf_export() warn: >>>> '&dmabuf->list_node' not removed from list >>>> >>>> If dma_buf_stats_setup() fails in dma_buf_export(), goto err_sysfs >>>> and dmabuf will be freed, but dmabuf->list_node will not be removed >>>> from db_list.head, then list traversal may cause UAF. >>>> >>>> Fix by removeing it from db_list.head before free(). >>>> >>>> Fixes: ef3a6b70507a ("dma-buf: call dma_buf_stats_setup after dmabuf is in valid list") >>>> Signed-off-by: Gaosheng Cui <cuigaosheng1@huawei.com> >>>> --- >>>> drivers/dma-buf/dma-buf.c | 3 +++ >>>> 1 file changed, 3 insertions(+) >>>> >>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c >>>> index b809513b03fe..6848f50226d5 100644 >>>> --- a/drivers/dma-buf/dma-buf.c >>>> +++ b/drivers/dma-buf/dma-buf.c >>>> @@ -675,6 +675,9 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) >>>> return dmabuf; >>>> >>>> err_sysfs: >>>> + mutex_lock(&db_list.lock); >>>> + list_del(&dmabuf->list_node); >>>> + mutex_unlock(&db_list.lock); >>>> /* >>>> * Set file->f_path.dentry->d_fsdata to NULL so that when >>>> * dma_buf_release() gets invoked by dentry_ops, it exits
Thanks T.J and Christian for the inputs.
On 11/19/2022 7:00 PM, Christian König wrote:
Yes, exactly that's the idea.
The only alternatives I can see would be to either move allocating the file and so completing the dma_buf initialization last again or just ignore errors from sysfs.
> If we still want to avoid calling dmabuf->ops->release(dmabuf) in > dma_buf_release like the comment says I guess we could use sysfs_entry > and ERR_PTR to flag that, otherwise it looks like we'd need a bit > somewhere.
No, this should be dropped as far as I can see. The sysfs cleanup code looks like it can handle not initialized kobj pointers.
Yeah there is also the null check in dma_buf_stats_teardown() that would prevent it from running, but I understood the comment to be referring to the release() dma_buf_ops call into the exporter which comes right after the teardown call. That looks like it's preventing the fput task work calling back into the exporter after the exporter already got an error from dma_buf_export(). Otherwise the exporter sees a release() for a buffer that it doesn't know about / thinks shouldn't exist. So I could imagine an exporter trying to double free: once for the failed dma_buf_export() call, and again when the release() op is called later.
Oh, very good point as well. Yeah, then creating the file should probably come last.
@Gaosheng: Could you please make these changes or you let me to do?
Regards, Christian.
Thanks T.J and Christian, thanks everyone for taking time to review this patch.
Charan, actually I don't have a good patch to to fix it, if you can submit a new patch to solve it, please feel free to do it.
By the way, I'd appreciate it if you could send to me the new patch when you submit it.
Thanks again!
Gaosheng.
On 2022/11/24 13:56, Charan Teja Kalla wrote:
Thanks T.J and Christian for the inputs.
On 11/19/2022 7:00 PM, Christian König wrote:
Yes, exactly that's the idea.
The only alternatives I can see would be to either move allocating the file and so completing the dma_buf initialization last again or just ignore errors from sysfs.
> If we still want to avoid calling dmabuf->ops->release(dmabuf) in > dma_buf_release like the comment says I guess we could use sysfs_entry > and ERR_PTR to flag that, otherwise it looks like we'd need a bit > somewhere.
No, this should be dropped as far as I can see. The sysfs cleanup code looks like it can handle not initialized kobj pointers.
Yeah there is also the null check in dma_buf_stats_teardown() that would prevent it from running, but I understood the comment to be referring to the release() dma_buf_ops call into the exporter which comes right after the teardown call. That looks like it's preventing the fput task work calling back into the exporter after the exporter already got an error from dma_buf_export(). Otherwise the exporter sees a release() for a buffer that it doesn't know about / thinks shouldn't exist. So I could imagine an exporter trying to double free: once for the failed dma_buf_export() call, and again when the release() op is called later.
Oh, very good point as well. Yeah, then creating the file should probably come last.
@Gaosheng: Could you please make these changes or you let me to do?
Regards, Christian.
.
Some tips: Before we call the dma_buf_stats_setup(), we have to finish creating the file, otherwise dma_buf_stats_setup() will return -EINVAL, maybe we need to think about this when making a new patch.
Hope these tips are useful, thanks!
On 2022/11/24 13:56, Charan Teja Kalla wrote:
Thanks T.J and Christian for the inputs.
On 11/19/2022 7:00 PM, Christian König wrote:
Yes, exactly that's the idea.
The only alternatives I can see would be to either move allocating the file and so completing the dma_buf initialization last again or just ignore errors from sysfs.
> If we still want to avoid calling dmabuf->ops->release(dmabuf) in > dma_buf_release like the comment says I guess we could use sysfs_entry > and ERR_PTR to flag that, otherwise it looks like we'd need a bit > somewhere.
No, this should be dropped as far as I can see. The sysfs cleanup code looks like it can handle not initialized kobj pointers.
Yeah there is also the null check in dma_buf_stats_teardown() that would prevent it from running, but I understood the comment to be referring to the release() dma_buf_ops call into the exporter which comes right after the teardown call. That looks like it's preventing the fput task work calling back into the exporter after the exporter already got an error from dma_buf_export(). Otherwise the exporter sees a release() for a buffer that it doesn't know about / thinks shouldn't exist. So I could imagine an exporter trying to double free: once for the failed dma_buf_export() call, and again when the release() op is called later.
Oh, very good point as well. Yeah, then creating the file should probably come last.
@Gaosheng: Could you please make these changes or you let me to do?
Regards, Christian.
.
Am 24.11.22 um 13:05 schrieb cuigaosheng:
Some tips: Before we call the dma_buf_stats_setup(), we have to finish creating the file, otherwise dma_buf_stats_setup() will return -EINVAL, maybe we need to think about this when making a new patch.
I was already wondering why the order is this way.
Why is dma_buf_stats_setup() needing the file in the first place?
Thanks, Christian.
Hope these tips are useful, thanks!
On 2022/11/24 13:56, Charan Teja Kalla wrote:
Thanks T.J and Christian for the inputs.
On 11/19/2022 7:00 PM, Christian König wrote:
Yes, exactly that's the idea.
The only alternatives I can see would be to either move allocating the file and so completing the dma_buf initialization last again or just ignore errors from sysfs.
> If we still want to avoid calling dmabuf->ops->release(dmabuf) in > dma_buf_release like the comment says I guess we could use sysfs_entry > and ERR_PTR to flag that, otherwise it looks like we'd need a bit > somewhere.
No, this should be dropped as far as I can see. The sysfs cleanup code looks like it can handle not initialized kobj pointers.
Yeah there is also the null check in dma_buf_stats_teardown() that would prevent it from running, but I understood the comment to be referring to the release() dma_buf_ops call into the exporter which comes right after the teardown call. That looks like it's preventing the fput task work calling back into the exporter after the exporter already got an error from dma_buf_export(). Otherwise the exporter sees a release() for a buffer that it doesn't know about / thinks shouldn't exist. So I could imagine an exporter trying to double free: once for the failed dma_buf_export() call, and again when the release() op is called later.
Oh, very good point as well. Yeah, then creating the file should probably come last.
@Gaosheng: Could you please make these changes or you let me to do?
Regards, Christian.
.
I was already wondering why the order is this way.
Why is dma_buf_stats_setup() needing the file in the first place?
dmabuf->file will be used in dma_buf_stats_setup(), the dma_buf_stats_setup() as follows:
171 int dma_buf_stats_setup(struct dma_buf *dmabuf) 172 { 173 struct dma_buf_sysfs_entry *sysfs_entry; 174 int ret; 175 176 if (!dmabuf || !dmabuf->file) 177 return -EINVAL; 178 179 if (!dmabuf->exp_name) { 180 pr_err("exporter name must not be empty if stats needed\n"); 181 return -EINVAL; 182 } 183 184 sysfs_entry = kzalloc(sizeof(struct dma_buf_sysfs_entry), GFP_KERNEL); 185 if (!sysfs_entry) 186 return -ENOMEM; 187 188 sysfs_entry->kobj.kset = dma_buf_per_buffer_stats_kset; 189 sysfs_entry->dmabuf = dmabuf; 190 191 dmabuf->sysfs_entry = sysfs_entry; 192 193 /* create the directory for buffer stats */ 194 ret = kobject_init_and_add(&sysfs_entry->kobj, &dma_buf_ktype, NULL, 195 "%lu", file_inode(dmabuf->file)->i_ino); 196 if (ret) 197 goto err_sysfs_dmabuf; 198 199 return 0; 200 201 err_sysfs_dmabuf: 202 kobject_put(&sysfs_entry->kobj); 203 dmabuf->sysfs_entry = NULL; 204 return ret; 205 }
Did I miss something?
Thanks.
On 2022/11/24 20:37, Christian König wrote:
Am 24.11.22 um 13:05 schrieb cuigaosheng:
Some tips: Before we call the dma_buf_stats_setup(), we have to finish creating the file, otherwise dma_buf_stats_setup() will return -EINVAL, maybe we need to think about this when making a new patch.
I was already wondering why the order is this way.
Why is dma_buf_stats_setup() needing the file in the first place?
Thanks, Christian.
Hope these tips are useful, thanks!
On 2022/11/24 13:56, Charan Teja Kalla wrote:
Thanks T.J and Christian for the inputs.
On 11/19/2022 7:00 PM, Christian König wrote:
Yes, exactly that's the idea.
The only alternatives I can see would be to either move allocating the file and so completing the dma_buf initialization last again or just ignore errors from sysfs.
> If we still want to avoid calling dmabuf->ops->release(dmabuf) in > dma_buf_release like the comment says I guess we could use sysfs_entry > and ERR_PTR to flag that, otherwise it looks like we'd need a bit > somewhere.
No, this should be dropped as far as I can see. The sysfs cleanup code looks like it can handle not initialized kobj pointers.
Yeah there is also the null check in dma_buf_stats_teardown() that would prevent it from running, but I understood the comment to be referring to the release() dma_buf_ops call into the exporter which comes right after the teardown call. That looks like it's preventing the fput task work calling back into the exporter after the exporter already got an error from dma_buf_export(). Otherwise the exporter sees a release() for a buffer that it doesn't know about / thinks shouldn't exist. So I could imagine an exporter trying to double free: once for the failed dma_buf_export() call, and again when the release() op is called later.
Oh, very good point as well. Yeah, then creating the file should probably come last.
@Gaosheng: Could you please make these changes or you let me to do?
Regards, Christian.
.
.
Am 24.11.22 um 13:49 schrieb cuigaosheng:
I was already wondering why the order is this way.
Why is dma_buf_stats_setup() needing the file in the first place?
dmabuf->file will be used in dma_buf_stats_setup(), the dma_buf_stats_setup() as follows:
171 int dma_buf_stats_setup(struct dma_buf *dmabuf) 172 { 173 struct dma_buf_sysfs_entry *sysfs_entry; 174 int ret; 175 176 if (!dmabuf || !dmabuf->file) 177 return -EINVAL; 178 179 if (!dmabuf->exp_name) { 180 pr_err("exporter name must not be empty if stats needed\n"); 181 return -EINVAL; 182 } 183 184 sysfs_entry = kzalloc(sizeof(struct dma_buf_sysfs_entry), GFP_KERNEL); 185 if (!sysfs_entry) 186 return -ENOMEM; 187 188 sysfs_entry->kobj.kset = dma_buf_per_buffer_stats_kset; 189 sysfs_entry->dmabuf = dmabuf; 190 191 dmabuf->sysfs_entry = sysfs_entry; 192 193 /* create the directory for buffer stats */ 194 ret = kobject_init_and_add(&sysfs_entry->kobj, &dma_buf_ktype, NULL, 195 "%lu", file_inode(dmabuf->file)->i_ino);
Ah, so it uses the i_ino of the file for the sysfs unique name.
I'm going to take another look how to properly clean this up.
Thanks for pointing this out, Christian.
196 if (ret) 197 goto err_sysfs_dmabuf; 198 199 return 0; 200 201 err_sysfs_dmabuf: 202 kobject_put(&sysfs_entry->kobj); 203 dmabuf->sysfs_entry = NULL; 204 return ret; 205 }
Did I miss something?
Thanks.
On 2022/11/24 20:37, Christian König wrote:
Am 24.11.22 um 13:05 schrieb cuigaosheng:
Some tips: Before we call the dma_buf_stats_setup(), we have to finish creating the file, otherwise dma_buf_stats_setup() will return -EINVAL, maybe we need to think about this when making a new patch.
I was already wondering why the order is this way.
Why is dma_buf_stats_setup() needing the file in the first place?
Thanks, Christian.
Hope these tips are useful, thanks!
On 2022/11/24 13:56, Charan Teja Kalla wrote:
Thanks T.J and Christian for the inputs.
On 11/19/2022 7:00 PM, Christian König wrote:
Yes, exactly that's the idea.
The only alternatives I can see would be to either move allocating the file and so completing the dma_buf initialization last again or just ignore errors from sysfs.
> If we still want to avoid calling dmabuf->ops->release(dmabuf) in > dma_buf_release like the comment says I guess we could use sysfs_entry > and ERR_PTR to flag that, otherwise it looks like we'd need a bit > somewhere.
No, this should be dropped as far as I can see. The sysfs cleanup code looks like it can handle not initialized kobj pointers.
Yeah there is also the null check in dma_buf_stats_teardown() that would prevent it from running, but I understood the comment to be referring to the release() dma_buf_ops call into the exporter which comes right after the teardown call. That looks like it's preventing the fput task work calling back into the exporter after the exporter already got an error from dma_buf_export(). Otherwise the exporter sees a release() for a buffer that it doesn't know about / thinks shouldn't exist. So I could imagine an exporter trying to double free: once for the failed dma_buf_export() call, and again when the release() op is called later.
Oh, very good point as well. Yeah, then creating the file should probably come last.
@Gaosheng: Could you please make these changes or you let me to do?
Regards, Christian.
.
.
Linaro-mm-sig mailing list -- linaro-mm-sig@lists.linaro.org To unsubscribe send an email to linaro-mm-sig-leave@lists.linaro.org
Thanks Christian/TJ for all your inputs!!
On 11/24/2022 6:25 PM, Christian König wrote:
I was already wondering why the order is this way.
Why is dma_buf_stats_setup() needing the file in the first place?
dmabuf->file will be used in dma_buf_stats_setup(), the dma_buf_stats_setup() as follows:
171 int dma_buf_stats_setup(struct dma_buf *dmabuf) 172 { 173 struct dma_buf_sysfs_entry *sysfs_entry; 174 int ret; 175 176 if (!dmabuf || !dmabuf->file) 177 return -EINVAL; 178 179 if (!dmabuf->exp_name) { 180 pr_err("exporter name must not be empty if stats needed\n"); 181 return -EINVAL; 182 } 183 184 sysfs_entry = kzalloc(sizeof(struct dma_buf_sysfs_entry), GFP_KERNEL); 185 if (!sysfs_entry) 186 return -ENOMEM; 187 188 sysfs_entry->kobj.kset = dma_buf_per_buffer_stats_kset; 189 sysfs_entry->dmabuf = dmabuf; 190 191 dmabuf->sysfs_entry = sysfs_entry; 192 193 /* create the directory for buffer stats */ 194 ret = kobject_init_and_add(&sysfs_entry->kobj, &dma_buf_ktype, NULL, 195 "%lu", file_inode(dmabuf->file)->i_ino);
Ah, so it uses the i_ino of the file for the sysfs unique name.
I'm going to take another look how to properly clean this up.
How about deleting the dmabuf from the db_list directly in the error path (which is usually done by the fput()) and then continue with the normal fput() here.
Just compile tested the below code and If the logic make sense for you, will send the final tested patch. ----------------------><---------------------------------------------
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c index e6f36c0..10a1727 100644 --- a/drivers/dma-buf/dma-buf.c +++ b/drivers/dma-buf/dma-buf.c @@ -87,19 +87,28 @@ static void dma_buf_release(struct dentry *dentry) kfree(dmabuf); }
-static int dma_buf_file_release(struct inode *inode, struct file *file) +static void dma_buf_db_list_remove(struct file *file) { struct dma_buf *dmabuf;
- if (!is_dma_buf_file(file)) - return -EINVAL; - dmabuf = file->private_data; + if (!dmabuf) + return;
mutex_lock(&db_list.lock); list_del(&dmabuf->list_node); mutex_unlock(&db_list.lock);
+ file->private_data = NULL; +} + +static int dma_buf_file_release(struct inode *inode, struct file *file) +{ + if (!is_dma_buf_file(file)) + return -EINVAL; + + dma_buf_db_list_remove(file); + return 0; }
@@ -688,6 +697,8 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) * early before calling the release() dma_buf op. */ file->f_path.dentry->d_fsdata = NULL; + + dma_buf_db_list_remove(file); fput(file); err_dmabuf: kfree(dmabuf);
--------------------><-----------------------------
Thanks for pointing this out, Christian.
Am 06.12.22 um 13:55 schrieb Charan Teja Kalla:
Thanks Christian/TJ for all your inputs!!
On 11/24/2022 6:25 PM, Christian König wrote:
I was already wondering why the order is this way.
Why is dma_buf_stats_setup() needing the file in the first place?
dmabuf->file will be used in dma_buf_stats_setup(), the dma_buf_stats_setup() as follows:
171 int dma_buf_stats_setup(struct dma_buf *dmabuf) 172 { 173 struct dma_buf_sysfs_entry *sysfs_entry; 174 int ret; 175 176 if (!dmabuf || !dmabuf->file) 177 return -EINVAL; 178 179 if (!dmabuf->exp_name) { 180 pr_err("exporter name must not be empty if stats needed\n"); 181 return -EINVAL; 182 } 183 184 sysfs_entry = kzalloc(sizeof(struct dma_buf_sysfs_entry), GFP_KERNEL); 185 if (!sysfs_entry) 186 return -ENOMEM; 187 188 sysfs_entry->kobj.kset = dma_buf_per_buffer_stats_kset; 189 sysfs_entry->dmabuf = dmabuf; 190 191 dmabuf->sysfs_entry = sysfs_entry; 192 193 /* create the directory for buffer stats */ 194 ret = kobject_init_and_add(&sysfs_entry->kobj, &dma_buf_ktype, NULL, 195 "%lu", file_inode(dmabuf->file)->i_ino);
Ah, so it uses the i_ino of the file for the sysfs unique name.
I'm going to take another look how to properly clean this up.
How about deleting the dmabuf from the db_list directly in the error path (which is usually done by the fput()) and then continue with the normal fput() here.
No, that's not really clean either.
Give me 10 Minutes, going to come up with something.
Regards, Christian.
Just compile tested the below code and If the logic make sense for you, will send the final tested patch. ----------------------><---------------------------------------------
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c index e6f36c0..10a1727 100644 --- a/drivers/dma-buf/dma-buf.c +++ b/drivers/dma-buf/dma-buf.c @@ -87,19 +87,28 @@ static void dma_buf_release(struct dentry *dentry) kfree(dmabuf); }
-static int dma_buf_file_release(struct inode *inode, struct file *file) +static void dma_buf_db_list_remove(struct file *file) { struct dma_buf *dmabuf;
if (!is_dma_buf_file(file))return -EINVAL;dmabuf = file->private_data;
if (!dmabuf)return; mutex_lock(&db_list.lock); list_del(&dmabuf->list_node); mutex_unlock(&db_list.lock);file->private_data = NULL;+}
+static int dma_buf_file_release(struct inode *inode, struct file *file) +{
if (!is_dma_buf_file(file))return -EINVAL;dma_buf_db_list_remove(file);return 0;@@ -688,6 +697,8 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) * early before calling the release() dma_buf op. */ file->f_path.dentry->d_fsdata = NULL;
dma_buf_db_list_remove(file); fput(file);--------------------><-----------------------------
Thanks for pointing this out, Christian.
linaro-mm-sig@lists.linaro.org
-
Charan Teja Kalla -
Christian König -
Christian König -
cuigaosheng -
Gaosheng Cui
-
T.J. Mercier