The following race occurs while accessing the dmabuf object exported as
file:
P1 P2
dma_buf_release() dmabuffs_dname()
[say lsof reading /proc/<P1 pid>/fd/<num>]
read dmabuf stored in dentry->d_fsdata
Free the dmabuf object
Start accessing the dmabuf structure
In the above description, the dmabuf object freed in P1 is being
accessed from P2 which is resulting into the use-after-free. Below is
the dump stack reported.
We are reading the dmabuf object stored in the dentry->d_fsdata but
there is no binding between the dentry and the dmabuf which means that
the dmabuf can be freed while it is being read from ->d_fsdata and
inuse. Reviews on the patch V1 says that protecting the dmabuf inuse
with an extra refcount is not a viable solution as the exported dmabuf
is already under file's refcount and keeping the multiple refcounts on
the same object coordinated is not possible.
As we are reading the dmabuf in ->d_fsdata just to get the user passed
name, we can directly store the name in d_fsdata thus can avoid the
reading of dmabuf altogether.
Call Trace:
kasan_report+0x12/0x20
__asan_report_load8_noabort+0x14/0x20
dmabuffs_dname+0x4f4/0x560
tomoyo_realpath_from_path+0x165/0x660
tomoyo_get_realpath
tomoyo_check_open_permission+0x2a3/0x3e0
tomoyo_file_open
tomoyo_file_open+0xa9/0xd0
security_file_open+0x71/0x300
do_dentry_open+0x37a/0x1380
vfs_open+0xa0/0xd0
path_openat+0x12ee/0x3490
do_filp_open+0x192/0x260
do_sys_openat2+0x5eb/0x7e0
do_sys_open+0xf2/0x180
Fixes: bb2bb9030425 ("dma-buf: add DMA_BUF_SET_NAME ioctls")
Reported-by: [email protected]
Cc: <[email protected]> [5.3+]
Signed-off-by: Charan Teja Reddy <[email protected]>
---
Changes in v2:
- Pass the user passed name in ->d_fsdata instead of dmabuf
- Improve the commit message
Changes in v1: (https://patchwork.kernel.org/patch/11514063/)
drivers/dma-buf/dma-buf.c | 17 ++++++++++-------
1 file changed, 10 insertions(+), 7 deletions(-)
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
index 01ce125..0071f7d 100644
--- a/drivers/dma-buf/dma-buf.c
+++ b/drivers/dma-buf/dma-buf.c
@@ -25,6 +25,7 @@
#include <linux/mm.h>
#include <linux/mount.h>
#include <linux/pseudo_fs.h>
+#include <linux/dcache.h>
#include <uapi/linux/dma-buf.h>
#include <uapi/linux/magic.h>
@@ -40,15 +41,13 @@ struct dma_buf_list {
static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen)
{
- struct dma_buf *dmabuf;
char name[DMA_BUF_NAME_LEN];
size_t ret = 0;
- dmabuf = dentry->d_fsdata;
- dma_resv_lock(dmabuf->resv, NULL);
- if (dmabuf->name)
- ret = strlcpy(name, dmabuf->name, DMA_BUF_NAME_LEN);
- dma_resv_unlock(dmabuf->resv);
+ spin_lock(&dentry->d_lock);
+ if (dentry->d_fsdata)
+ ret = strlcpy(name, dentry->d_fsdata, DMA_BUF_NAME_LEN);
+ spin_unlock(&dentry->d_lock);
return dynamic_dname(dentry, buffer, buflen, "/%s:%s",
dentry->d_name.name, ret > 0 ? name : "");
@@ -80,12 +79,16 @@ static int dma_buf_fs_init_context(struct fs_context *fc)
static int dma_buf_release(struct inode *inode, struct file *file)
{
struct dma_buf *dmabuf;
+ struct dentry *dentry = file->f_path.dentry;
if (!is_dma_buf_file(file))
return -EINVAL;
dmabuf = file->private_data;
+ spin_lock(&dentry->d_lock);
+ dentry->d_fsdata = NULL;
+ spin_unlock(&dentry->d_lock);
BUG_ON(dmabuf->vmapping_counter);
/*
@@ -343,6 +346,7 @@ static long dma_buf_set_name(struct dma_buf *dmabuf, const char __user *buf)
}
kfree(dmabuf->name);
dmabuf->name = name;
+ dmabuf->file->f_path.dentry->d_fsdata = name;
out_unlock:
dma_resv_unlock(dmabuf->resv);
@@ -446,7 +450,6 @@ static struct file *dma_buf_getfile(struct dma_buf *dmabuf, int flags)
goto err_alloc_file;
file->f_flags = flags & (O_ACCMODE | O_NONBLOCK);
file->private_data = dmabuf;
- file->f_path.dentry->d_fsdata = dmabuf;
return file;
--
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a
member of the Code Aurora Forum, hosted by The Linux Foundation
On Fri, May 08, 2020 at 12:11:03PM +0530, Charan Teja Reddy wrote:
> The following race occurs while accessing the dmabuf object exported as
> file:
> P1 P2
> dma_buf_release() dmabuffs_dname()
> [say lsof reading /proc/<P1 pid>/fd/<num>]
>
> read dmabuf stored in dentry->d_fsdata
> Free the dmabuf object
> Start accessing the dmabuf structure
>
> In the above description, the dmabuf object freed in P1 is being
> accessed from P2 which is resulting into the use-after-free. Below is
> the dump stack reported.
>
> We are reading the dmabuf object stored in the dentry->d_fsdata but
> there is no binding between the dentry and the dmabuf which means that
> the dmabuf can be freed while it is being read from ->d_fsdata and
> inuse. Reviews on the patch V1 says that protecting the dmabuf inuse
> with an extra refcount is not a viable solution as the exported dmabuf
> is already under file's refcount and keeping the multiple refcounts on
> the same object coordinated is not possible.
>
> As we are reading the dmabuf in ->d_fsdata just to get the user passed
> name, we can directly store the name in d_fsdata thus can avoid the
> reading of dmabuf altogether.
>
> Call Trace:
> kasan_report+0x12/0x20
> __asan_report_load8_noabort+0x14/0x20
> dmabuffs_dname+0x4f4/0x560
> tomoyo_realpath_from_path+0x165/0x660
> tomoyo_get_realpath
> tomoyo_check_open_permission+0x2a3/0x3e0
> tomoyo_file_open
> tomoyo_file_open+0xa9/0xd0
> security_file_open+0x71/0x300
> do_dentry_open+0x37a/0x1380
> vfs_open+0xa0/0xd0
> path_openat+0x12ee/0x3490
> do_filp_open+0x192/0x260
> do_sys_openat2+0x5eb/0x7e0
> do_sys_open+0xf2/0x180
>
> Fixes: bb2bb9030425 ("dma-buf: add DMA_BUF_SET_NAME ioctls")
> Reported-by: [email protected]
> Cc: <[email protected]> [5.3+]
> Signed-off-by: Charan Teja Reddy <[email protected]>
> ---
>
> Changes in v2:
>
> - Pass the user passed name in ->d_fsdata instead of dmabuf
> - Improve the commit message
>
> Changes in v1: (https://patchwork.kernel.org/patch/11514063/)
>
> drivers/dma-buf/dma-buf.c | 17 ++++++++++-------
> 1 file changed, 10 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> index 01ce125..0071f7d 100644
> --- a/drivers/dma-buf/dma-buf.c
> +++ b/drivers/dma-buf/dma-buf.c
> @@ -25,6 +25,7 @@
> #include <linux/mm.h>
> #include <linux/mount.h>
> #include <linux/pseudo_fs.h>
> +#include <linux/dcache.h>
>
> #include <uapi/linux/dma-buf.h>
> #include <uapi/linux/magic.h>
> @@ -40,15 +41,13 @@ struct dma_buf_list {
>
> static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen)
> {
> - struct dma_buf *dmabuf;
> char name[DMA_BUF_NAME_LEN];
> size_t ret = 0;
>
> - dmabuf = dentry->d_fsdata;
> - dma_resv_lock(dmabuf->resv, NULL);
> - if (dmabuf->name)
> - ret = strlcpy(name, dmabuf->name, DMA_BUF_NAME_LEN);
> - dma_resv_unlock(dmabuf->resv);
> + spin_lock(&dentry->d_lock);
Are you sure this lock always protects d_fsdata?
> + if (dentry->d_fsdata)
> + ret = strlcpy(name, dentry->d_fsdata, DMA_BUF_NAME_LEN);
> + spin_unlock(&dentry->d_lock);
>
> return dynamic_dname(dentry, buffer, buflen, "/%s:%s",
> dentry->d_name.name, ret > 0 ? name : "");
If the above check fails the name will be what? How could d_name.name
be valid but d_fsdata not be valid?
> @@ -80,12 +79,16 @@ static int dma_buf_fs_init_context(struct fs_context *fc)
> static int dma_buf_release(struct inode *inode, struct file *file)
> {
> struct dma_buf *dmabuf;
> + struct dentry *dentry = file->f_path.dentry;
>
> if (!is_dma_buf_file(file))
> return -EINVAL;
>
> dmabuf = file->private_data;
>
> + spin_lock(&dentry->d_lock);
> + dentry->d_fsdata = NULL;
> + spin_unlock(&dentry->d_lock);
> BUG_ON(dmabuf->vmapping_counter);
>
> /*
> @@ -343,6 +346,7 @@ static long dma_buf_set_name(struct dma_buf *dmabuf, const char __user *buf)
> }
> kfree(dmabuf->name);
> dmabuf->name = name;
> + dmabuf->file->f_path.dentry->d_fsdata = name;
You are just changing the use of d_fsdata from being a pointer to the
dmabuf to being a pointer to the name string? What's to keep that name
string around and not have the same reference counting issues that the
dmabuf structure itself has? Who frees that string memory?
thanks,
greg k-h
Thank you Greg for the comments.
On 5/12/2020 2:22 PM, Greg KH wrote:
> On Fri, May 08, 2020 at 12:11:03PM +0530, Charan Teja Reddy wrote:
>> The following race occurs while accessing the dmabuf object exported as
>> file:
>> P1 P2
>> dma_buf_release() dmabuffs_dname()
>> [say lsof reading /proc/<P1 pid>/fd/<num>]
>>
>> read dmabuf stored in dentry->d_fsdata
>> Free the dmabuf object
>> Start accessing the dmabuf structure
>>
>> In the above description, the dmabuf object freed in P1 is being
>> accessed from P2 which is resulting into the use-after-free. Below is
>> the dump stack reported.
>>
>> We are reading the dmabuf object stored in the dentry->d_fsdata but
>> there is no binding between the dentry and the dmabuf which means that
>> the dmabuf can be freed while it is being read from ->d_fsdata and
>> inuse. Reviews on the patch V1 says that protecting the dmabuf inuse
>> with an extra refcount is not a viable solution as the exported dmabuf
>> is already under file's refcount and keeping the multiple refcounts on
>> the same object coordinated is not possible.
>>
>> As we are reading the dmabuf in ->d_fsdata just to get the user passed
>> name, we can directly store the name in d_fsdata thus can avoid the
>> reading of dmabuf altogether.
>>
>> Call Trace:
>> kasan_report+0x12/0x20
>> __asan_report_load8_noabort+0x14/0x20
>> dmabuffs_dname+0x4f4/0x560
>> tomoyo_realpath_from_path+0x165/0x660
>> tomoyo_get_realpath
>> tomoyo_check_open_permission+0x2a3/0x3e0
>> tomoyo_file_open
>> tomoyo_file_open+0xa9/0xd0
>> security_file_open+0x71/0x300
>> do_dentry_open+0x37a/0x1380
>> vfs_open+0xa0/0xd0
>> path_openat+0x12ee/0x3490
>> do_filp_open+0x192/0x260
>> do_sys_openat2+0x5eb/0x7e0
>> do_sys_open+0xf2/0x180
>>
>> Fixes: bb2bb9030425 ("dma-buf: add DMA_BUF_SET_NAME ioctls")
>> Reported-by: [email protected]
>> Cc: <[email protected]> [5.3+]
>> Signed-off-by: Charan Teja Reddy <[email protected]>
>> ---
>>
>> Changes in v2:
>>
>> - Pass the user passed name in ->d_fsdata instead of dmabuf
>> - Improve the commit message
>>
>> Changes in v1: (https://patchwork.kernel.org/patch/11514063/)
>>
>> drivers/dma-buf/dma-buf.c | 17 ++++++++++-------
>> 1 file changed, 10 insertions(+), 7 deletions(-)
>>
>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
>> index 01ce125..0071f7d 100644
>> --- a/drivers/dma-buf/dma-buf.c
>> +++ b/drivers/dma-buf/dma-buf.c
>> @@ -25,6 +25,7 @@
>> #include <linux/mm.h>
>> #include <linux/mount.h>
>> #include <linux/pseudo_fs.h>
>> +#include <linux/dcache.h>
>>
>> #include <uapi/linux/dma-buf.h>
>> #include <uapi/linux/magic.h>
>> @@ -40,15 +41,13 @@ struct dma_buf_list {
>>
>> static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen)
>> {
>> - struct dma_buf *dmabuf;
>> char name[DMA_BUF_NAME_LEN];
>> size_t ret = 0;
>>
>> - dmabuf = dentry->d_fsdata;
>> - dma_resv_lock(dmabuf->resv, NULL);
>> - if (dmabuf->name)
>> - ret = strlcpy(name, dmabuf->name, DMA_BUF_NAME_LEN);
>> - dma_resv_unlock(dmabuf->resv);
>> + spin_lock(&dentry->d_lock);
>
> Are you sure this lock always protects d_fsdata?
I think yes. In the dma-buf.c, I have to make sure that d_fsdata should
always be under d_lock thus it will be protected. (In this posted patch
there is one place(in dma_buf_set_name) that is missed, will update this
in V3).
>
>> + if (dentry->d_fsdata)
>> + ret = strlcpy(name, dentry->d_fsdata, DMA_BUF_NAME_LEN);
>> + spin_unlock(&dentry->d_lock);
>>
>> return dynamic_dname(dentry, buffer, buflen, "/%s:%s",
>> dentry->d_name.name, ret > 0 ? name : "");
>
> If the above check fails the name will be what? How could d_name.name
> be valid but d_fsdata not be valid?
In case of check fails, empty string "" is appended to the name by the
code, ret > 0 ? name : "", ret is initialized to zero. Thus the name
string will be like "/dmabuf:".
Regarding the validity of d_fsdata, we are setting the dmabuf's
dentry->d_fsdata to NULL in the dma_buf_release() thus can go invalid if
that dmabuf is in the free path.
>
>
>> @@ -80,12 +79,16 @@ static int dma_buf_fs_init_context(struct fs_context *fc)
>> static int dma_buf_release(struct inode *inode, struct file *file)
>> {
>> struct dma_buf *dmabuf;
>> + struct dentry *dentry = file->f_path.dentry;
>>
>> if (!is_dma_buf_file(file))
>> return -EINVAL;
>>
>> dmabuf = file->private_data;
>>
>> + spin_lock(&dentry->d_lock);
>> + dentry->d_fsdata = NULL;
>> + spin_unlock(&dentry->d_lock);
>> BUG_ON(dmabuf->vmapping_counter);
>>
>> /*
>> @@ -343,6 +346,7 @@ static long dma_buf_set_name(struct dma_buf *dmabuf, const char __user *buf)
>> }
>> kfree(dmabuf->name);
>> dmabuf->name = name;
>> + dmabuf->file->f_path.dentry->d_fsdata = name;
>
> You are just changing the use of d_fsdata from being a pointer to the
> dmabuf to being a pointer to the name string? What's to keep that name
> string around and not have the same reference counting issues that the
> dmabuf structure itself has? Who frees that string memory?
>
Yes, I am just storing the name string in the d_fsdata in place of
dmabuf and this helps to get rid of any extra refcount requirement.
Because the user passed name carried in the d_fsdata is copied to the
local buffer in dmabuffs_dname under spin_lock(d_lock) and the same
d_fsdata is set to NULL(under the d_lock only) when that dmabuf is in
the release path. So, when d_fsdata is NULL, name string is not accessed
from the dmabuffs_dname thus extra count is not required.
String memory, stored in the dmabuf->name, is released from the
dma_buf_release(). Flow will be like, It fist sets d_fsdata=NULL and
then free the dmabuf->name.
However from your comments I have realized that there is a race in this
patch when using the name string between dma_buf_set_name() and
dmabuffs_dname(). But, If the idea of passing the name string inplace of
dmabuf in d_fsdata looks fine, I can update this next patch.
> thanks,
>
> greg k-h
>
--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
Forum, a Linux Foundation Collaborative Project
On Wed, May 13, 2020 at 05:40:26PM +0530, Charan Teja Kalla wrote:
>
> Thank you Greg for the comments.
> On 5/12/2020 2:22 PM, Greg KH wrote:
> > On Fri, May 08, 2020 at 12:11:03PM +0530, Charan Teja Reddy wrote:
> >> The following race occurs while accessing the dmabuf object exported as
> >> file:
> >> P1 P2
> >> dma_buf_release() dmabuffs_dname()
> >> [say lsof reading /proc/<P1 pid>/fd/<num>]
> >>
> >> read dmabuf stored in dentry->d_fsdata
> >> Free the dmabuf object
> >> Start accessing the dmabuf structure
> >>
> >> In the above description, the dmabuf object freed in P1 is being
> >> accessed from P2 which is resulting into the use-after-free. Below is
> >> the dump stack reported.
> >>
> >> We are reading the dmabuf object stored in the dentry->d_fsdata but
> >> there is no binding between the dentry and the dmabuf which means that
> >> the dmabuf can be freed while it is being read from ->d_fsdata and
> >> inuse. Reviews on the patch V1 says that protecting the dmabuf inuse
> >> with an extra refcount is not a viable solution as the exported dmabuf
> >> is already under file's refcount and keeping the multiple refcounts on
> >> the same object coordinated is not possible.
> >>
> >> As we are reading the dmabuf in ->d_fsdata just to get the user passed
> >> name, we can directly store the name in d_fsdata thus can avoid the
> >> reading of dmabuf altogether.
> >>
> >> Call Trace:
> >> kasan_report+0x12/0x20
> >> __asan_report_load8_noabort+0x14/0x20
> >> dmabuffs_dname+0x4f4/0x560
> >> tomoyo_realpath_from_path+0x165/0x660
> >> tomoyo_get_realpath
> >> tomoyo_check_open_permission+0x2a3/0x3e0
> >> tomoyo_file_open
> >> tomoyo_file_open+0xa9/0xd0
> >> security_file_open+0x71/0x300
> >> do_dentry_open+0x37a/0x1380
> >> vfs_open+0xa0/0xd0
> >> path_openat+0x12ee/0x3490
> >> do_filp_open+0x192/0x260
> >> do_sys_openat2+0x5eb/0x7e0
> >> do_sys_open+0xf2/0x180
> >>
> >> Fixes: bb2bb9030425 ("dma-buf: add DMA_BUF_SET_NAME ioctls")
> >> Reported-by: [email protected]
> >> Cc: <[email protected]> [5.3+]
> >> Signed-off-by: Charan Teja Reddy <[email protected]>
> >> ---
> >>
> >> Changes in v2:
> >>
> >> - Pass the user passed name in ->d_fsdata instead of dmabuf
> >> - Improve the commit message
> >>
> >> Changes in v1: (https://patchwork.kernel.org/patch/11514063/)
> >>
> >> drivers/dma-buf/dma-buf.c | 17 ++++++++++-------
> >> 1 file changed, 10 insertions(+), 7 deletions(-)
> >>
> >> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> >> index 01ce125..0071f7d 100644
> >> --- a/drivers/dma-buf/dma-buf.c
> >> +++ b/drivers/dma-buf/dma-buf.c
> >> @@ -25,6 +25,7 @@
> >> #include <linux/mm.h>
> >> #include <linux/mount.h>
> >> #include <linux/pseudo_fs.h>
> >> +#include <linux/dcache.h>
> >>
> >> #include <uapi/linux/dma-buf.h>
> >> #include <uapi/linux/magic.h>
> >> @@ -40,15 +41,13 @@ struct dma_buf_list {
> >>
> >> static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen)
> >> {
> >> - struct dma_buf *dmabuf;
> >> char name[DMA_BUF_NAME_LEN];
> >> size_t ret = 0;
> >>
> >> - dmabuf = dentry->d_fsdata;
> >> - dma_resv_lock(dmabuf->resv, NULL);
> >> - if (dmabuf->name)
> >> - ret = strlcpy(name, dmabuf->name, DMA_BUF_NAME_LEN);
> >> - dma_resv_unlock(dmabuf->resv);
> >> + spin_lock(&dentry->d_lock);
> >
> > Are you sure this lock always protects d_fsdata?
>
> I think yes. In the dma-buf.c, I have to make sure that d_fsdata should
> always be under d_lock thus it will be protected. (In this posted patch
> there is one place(in dma_buf_set_name) that is missed, will update this
> in V3).
>
> >
> >> + if (dentry->d_fsdata)
> >> + ret = strlcpy(name, dentry->d_fsdata, DMA_BUF_NAME_LEN);
> >> + spin_unlock(&dentry->d_lock);
> >>
> >> return dynamic_dname(dentry, buffer, buflen, "/%s:%s",
> >> dentry->d_name.name, ret > 0 ? name : "");
> >
> > If the above check fails the name will be what? How could d_name.name
> > be valid but d_fsdata not be valid?
>
> In case of check fails, empty string "" is appended to the name by the
> code, ret > 0 ? name : "", ret is initialized to zero. Thus the name
> string will be like "/dmabuf:".
So multiple objects can have the same "name" if this happens to multiple
ones at once?
> Regarding the validity of d_fsdata, we are setting the dmabuf's
> dentry->d_fsdata to NULL in the dma_buf_release() thus can go invalid if
> that dmabuf is in the free path.
Why are we allowing the name to be set if the dmabuf is on the free path
at all? Shouldn't that be the real fix here?
> >> @@ -80,12 +79,16 @@ static int dma_buf_fs_init_context(struct fs_context *fc)
> >> static int dma_buf_release(struct inode *inode, struct file *file)
> >> {
> >> struct dma_buf *dmabuf;
> >> + struct dentry *dentry = file->f_path.dentry;
> >>
> >> if (!is_dma_buf_file(file))
> >> return -EINVAL;
> >>
> >> dmabuf = file->private_data;
> >>
> >> + spin_lock(&dentry->d_lock);
> >> + dentry->d_fsdata = NULL;
> >> + spin_unlock(&dentry->d_lock);
> >> BUG_ON(dmabuf->vmapping_counter);
> >>
> >> /*
> >> @@ -343,6 +346,7 @@ static long dma_buf_set_name(struct dma_buf *dmabuf, const char __user *buf)
> >> }
> >> kfree(dmabuf->name);
> >> dmabuf->name = name;
> >> + dmabuf->file->f_path.dentry->d_fsdata = name;
> >
> > You are just changing the use of d_fsdata from being a pointer to the
> > dmabuf to being a pointer to the name string? What's to keep that name
> > string around and not have the same reference counting issues that the
> > dmabuf structure itself has? Who frees that string memory?
> >
>
> Yes, I am just storing the name string in the d_fsdata in place of
> dmabuf and this helps to get rid of any extra refcount requirement.
> Because the user passed name carried in the d_fsdata is copied to the
> local buffer in dmabuffs_dname under spin_lock(d_lock) and the same
> d_fsdata is set to NULL(under the d_lock only) when that dmabuf is in
> the release path. So, when d_fsdata is NULL, name string is not accessed
> from the dmabuffs_dname thus extra count is not required.
>
> String memory, stored in the dmabuf->name, is released from the
> dma_buf_release(). Flow will be like, It fist sets d_fsdata=NULL and
> then free the dmabuf->name.
>
> However from your comments I have realized that there is a race in this
> patch when using the name string between dma_buf_set_name() and
> dmabuffs_dname(). But, If the idea of passing the name string inplace of
> dmabuf in d_fsdata looks fine, I can update this next patch.
I'll leave that to the dmabuf authors/maintainers, but it feels odd to
me...
thanks,
greg k-h
On Wed, May 13, 2020 at 02:51:12PM +0200, Greg KH wrote:
> On Wed, May 13, 2020 at 05:40:26PM +0530, Charan Teja Kalla wrote:
> >
> > Thank you Greg for the comments.
> > On 5/12/2020 2:22 PM, Greg KH wrote:
> > > On Fri, May 08, 2020 at 12:11:03PM +0530, Charan Teja Reddy wrote:
> > >> The following race occurs while accessing the dmabuf object exported as
> > >> file:
> > >> P1 P2
> > >> dma_buf_release() dmabuffs_dname()
> > >> [say lsof reading /proc/<P1 pid>/fd/<num>]
> > >>
> > >> read dmabuf stored in dentry->d_fsdata
> > >> Free the dmabuf object
> > >> Start accessing the dmabuf structure
> > >>
> > >> In the above description, the dmabuf object freed in P1 is being
> > >> accessed from P2 which is resulting into the use-after-free. Below is
> > >> the dump stack reported.
> > >>
> > >> We are reading the dmabuf object stored in the dentry->d_fsdata but
> > >> there is no binding between the dentry and the dmabuf which means that
> > >> the dmabuf can be freed while it is being read from ->d_fsdata and
> > >> inuse. Reviews on the patch V1 says that protecting the dmabuf inuse
> > >> with an extra refcount is not a viable solution as the exported dmabuf
> > >> is already under file's refcount and keeping the multiple refcounts on
> > >> the same object coordinated is not possible.
> > >>
> > >> As we are reading the dmabuf in ->d_fsdata just to get the user passed
> > >> name, we can directly store the name in d_fsdata thus can avoid the
> > >> reading of dmabuf altogether.
> > >>
> > >> Call Trace:
> > >> kasan_report+0x12/0x20
> > >> __asan_report_load8_noabort+0x14/0x20
> > >> dmabuffs_dname+0x4f4/0x560
> > >> tomoyo_realpath_from_path+0x165/0x660
> > >> tomoyo_get_realpath
> > >> tomoyo_check_open_permission+0x2a3/0x3e0
> > >> tomoyo_file_open
> > >> tomoyo_file_open+0xa9/0xd0
> > >> security_file_open+0x71/0x300
> > >> do_dentry_open+0x37a/0x1380
> > >> vfs_open+0xa0/0xd0
> > >> path_openat+0x12ee/0x3490
> > >> do_filp_open+0x192/0x260
> > >> do_sys_openat2+0x5eb/0x7e0
> > >> do_sys_open+0xf2/0x180
> > >>
> > >> Fixes: bb2bb9030425 ("dma-buf: add DMA_BUF_SET_NAME ioctls")
> > >> Reported-by: [email protected]
> > >> Cc: <[email protected]> [5.3+]
> > >> Signed-off-by: Charan Teja Reddy <[email protected]>
> > >> ---
> > >>
> > >> Changes in v2:
> > >>
> > >> - Pass the user passed name in ->d_fsdata instead of dmabuf
> > >> - Improve the commit message
> > >>
> > >> Changes in v1: (https://patchwork.kernel.org/patch/11514063/)
> > >>
> > >> drivers/dma-buf/dma-buf.c | 17 ++++++++++-------
> > >> 1 file changed, 10 insertions(+), 7 deletions(-)
> > >>
> > >> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> > >> index 01ce125..0071f7d 100644
> > >> --- a/drivers/dma-buf/dma-buf.c
> > >> +++ b/drivers/dma-buf/dma-buf.c
> > >> @@ -25,6 +25,7 @@
> > >> #include <linux/mm.h>
> > >> #include <linux/mount.h>
> > >> #include <linux/pseudo_fs.h>
> > >> +#include <linux/dcache.h>
> > >>
> > >> #include <uapi/linux/dma-buf.h>
> > >> #include <uapi/linux/magic.h>
> > >> @@ -40,15 +41,13 @@ struct dma_buf_list {
> > >>
> > >> static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen)
> > >> {
> > >> - struct dma_buf *dmabuf;
> > >> char name[DMA_BUF_NAME_LEN];
> > >> size_t ret = 0;
> > >>
> > >> - dmabuf = dentry->d_fsdata;
> > >> - dma_resv_lock(dmabuf->resv, NULL);
> > >> - if (dmabuf->name)
> > >> - ret = strlcpy(name, dmabuf->name, DMA_BUF_NAME_LEN);
> > >> - dma_resv_unlock(dmabuf->resv);
> > >> + spin_lock(&dentry->d_lock);
> > >
> > > Are you sure this lock always protects d_fsdata?
> >
> > I think yes. In the dma-buf.c, I have to make sure that d_fsdata should
> > always be under d_lock thus it will be protected. (In this posted patch
> > there is one place(in dma_buf_set_name) that is missed, will update this
> > in V3).
> >
> > >
> > >> + if (dentry->d_fsdata)
> > >> + ret = strlcpy(name, dentry->d_fsdata, DMA_BUF_NAME_LEN);
> > >> + spin_unlock(&dentry->d_lock);
> > >>
> > >> return dynamic_dname(dentry, buffer, buflen, "/%s:%s",
> > >> dentry->d_name.name, ret > 0 ? name : "");
> > >
> > > If the above check fails the name will be what? How could d_name.name
> > > be valid but d_fsdata not be valid?
> >
> > In case of check fails, empty string "" is appended to the name by the
> > code, ret > 0 ? name : "", ret is initialized to zero. Thus the name
> > string will be like "/dmabuf:".
>
> So multiple objects can have the same "name" if this happens to multiple
> ones at once?
>
> > Regarding the validity of d_fsdata, we are setting the dmabuf's
> > dentry->d_fsdata to NULL in the dma_buf_release() thus can go invalid if
> > that dmabuf is in the free path.
>
> Why are we allowing the name to be set if the dmabuf is on the free path
> at all? Shouldn't that be the real fix here?
>
> > >> @@ -80,12 +79,16 @@ static int dma_buf_fs_init_context(struct fs_context *fc)
> > >> static int dma_buf_release(struct inode *inode, struct file *file)
> > >> {
> > >> struct dma_buf *dmabuf;
> > >> + struct dentry *dentry = file->f_path.dentry;
> > >>
> > >> if (!is_dma_buf_file(file))
> > >> return -EINVAL;
> > >>
> > >> dmabuf = file->private_data;
> > >>
> > >> + spin_lock(&dentry->d_lock);
> > >> + dentry->d_fsdata = NULL;
> > >> + spin_unlock(&dentry->d_lock);
> > >> BUG_ON(dmabuf->vmapping_counter);
> > >>
> > >> /*
> > >> @@ -343,6 +346,7 @@ static long dma_buf_set_name(struct dma_buf *dmabuf, const char __user *buf)
> > >> }
> > >> kfree(dmabuf->name);
> > >> dmabuf->name = name;
> > >> + dmabuf->file->f_path.dentry->d_fsdata = name;
> > >
> > > You are just changing the use of d_fsdata from being a pointer to the
> > > dmabuf to being a pointer to the name string? What's to keep that name
> > > string around and not have the same reference counting issues that the
> > > dmabuf structure itself has? Who frees that string memory?
> > >
> >
> > Yes, I am just storing the name string in the d_fsdata in place of
> > dmabuf and this helps to get rid of any extra refcount requirement.
> > Because the user passed name carried in the d_fsdata is copied to the
> > local buffer in dmabuffs_dname under spin_lock(d_lock) and the same
> > d_fsdata is set to NULL(under the d_lock only) when that dmabuf is in
> > the release path. So, when d_fsdata is NULL, name string is not accessed
> > from the dmabuffs_dname thus extra count is not required.
> >
> > String memory, stored in the dmabuf->name, is released from the
> > dma_buf_release(). Flow will be like, It fist sets d_fsdata=NULL and
> > then free the dmabuf->name.
> >
> > However from your comments I have realized that there is a race in this
> > patch when using the name string between dma_buf_set_name() and
> > dmabuffs_dname(). But, If the idea of passing the name string inplace of
> > dmabuf in d_fsdata looks fine, I can update this next patch.
>
> I'll leave that to the dmabuf authors/maintainers, but it feels odd to
> me...
I have zero clue about fs internals. This all scares me, that's all. I
know enough about lifetime bugs that if you don't deeply understand a
subsystem, all that's guaranteed is that you will get it wrong.
/me out
Cheers, Daniel
>
> thanks,
>
> greg k-h
> _______________________________________________
> dri-devel mailing list
> [email protected]
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
On Wed, 13 May 2020 at 21:16, Daniel Vetter <[email protected]> wrote:
>
> On Wed, May 13, 2020 at 02:51:12PM +0200, Greg KH wrote:
> > On Wed, May 13, 2020 at 05:40:26PM +0530, Charan Teja Kalla wrote:
> > >
> > > Thank you Greg for the comments.
> > > On 5/12/2020 2:22 PM, Greg KH wrote:
> > > > On Fri, May 08, 2020 at 12:11:03PM +0530, Charan Teja Reddy wrote:
> > > >> The following race occurs while accessing the dmabuf object exported as
> > > >> file:
> > > >> P1 P2
> > > >> dma_buf_release() dmabuffs_dname()
> > > >> [say lsof reading /proc/<P1 pid>/fd/<num>]
> > > >>
> > > >> read dmabuf stored in dentry->d_fsdata
> > > >> Free the dmabuf object
> > > >> Start accessing the dmabuf structure
> > > >>
> > > >> In the above description, the dmabuf object freed in P1 is being
> > > >> accessed from P2 which is resulting into the use-after-free. Below is
> > > >> the dump stack reported.
> > > >>
> > > >> We are reading the dmabuf object stored in the dentry->d_fsdata but
> > > >> there is no binding between the dentry and the dmabuf which means that
> > > >> the dmabuf can be freed while it is being read from ->d_fsdata and
> > > >> inuse. Reviews on the patch V1 says that protecting the dmabuf inuse
> > > >> with an extra refcount is not a viable solution as the exported dmabuf
> > > >> is already under file's refcount and keeping the multiple refcounts on
> > > >> the same object coordinated is not possible.
> > > >>
> > > >> As we are reading the dmabuf in ->d_fsdata just to get the user passed
> > > >> name, we can directly store the name in d_fsdata thus can avoid the
> > > >> reading of dmabuf altogether.
> > > >>
> > > >> Call Trace:
> > > >> kasan_report+0x12/0x20
> > > >> __asan_report_load8_noabort+0x14/0x20
> > > >> dmabuffs_dname+0x4f4/0x560
> > > >> tomoyo_realpath_from_path+0x165/0x660
> > > >> tomoyo_get_realpath
> > > >> tomoyo_check_open_permission+0x2a3/0x3e0
> > > >> tomoyo_file_open
> > > >> tomoyo_file_open+0xa9/0xd0
> > > >> security_file_open+0x71/0x300
> > > >> do_dentry_open+0x37a/0x1380
> > > >> vfs_open+0xa0/0xd0
> > > >> path_openat+0x12ee/0x3490
> > > >> do_filp_open+0x192/0x260
> > > >> do_sys_openat2+0x5eb/0x7e0
> > > >> do_sys_open+0xf2/0x180
> > > >>
> > > >> Fixes: bb2bb9030425 ("dma-buf: add DMA_BUF_SET_NAME ioctls")
> > > >> Reported-by: [email protected]
> > > >> Cc: <[email protected]> [5.3+]
> > > >> Signed-off-by: Charan Teja Reddy <[email protected]>
> > > >> ---
> > > >>
> > > >> Changes in v2:
> > > >>
> > > >> - Pass the user passed name in ->d_fsdata instead of dmabuf
> > > >> - Improve the commit message
> > > >>
> > > >> Changes in v1: (https://patchwork.kernel.org/patch/11514063/)
> > > >>
> > > >> drivers/dma-buf/dma-buf.c | 17 ++++++++++-------
> > > >> 1 file changed, 10 insertions(+), 7 deletions(-)
> > > >>
> > > >> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> > > >> index 01ce125..0071f7d 100644
> > > >> --- a/drivers/dma-buf/dma-buf.c
> > > >> +++ b/drivers/dma-buf/dma-buf.c
> > > >> @@ -25,6 +25,7 @@
> > > >> #include <linux/mm.h>
> > > >> #include <linux/mount.h>
> > > >> #include <linux/pseudo_fs.h>
> > > >> +#include <linux/dcache.h>
> > > >>
> > > >> #include <uapi/linux/dma-buf.h>
> > > >> #include <uapi/linux/magic.h>
> > > >> @@ -40,15 +41,13 @@ struct dma_buf_list {
> > > >>
> > > >> static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen)
> > > >> {
> > > >> - struct dma_buf *dmabuf;
> > > >> char name[DMA_BUF_NAME_LEN];
> > > >> size_t ret = 0;
> > > >>
> > > >> - dmabuf = dentry->d_fsdata;
> > > >> - dma_resv_lock(dmabuf->resv, NULL);
> > > >> - if (dmabuf->name)
> > > >> - ret = strlcpy(name, dmabuf->name, DMA_BUF_NAME_LEN);
> > > >> - dma_resv_unlock(dmabuf->resv);
> > > >> + spin_lock(&dentry->d_lock);
> > > >
> > > > Are you sure this lock always protects d_fsdata?
> > >
> > > I think yes. In the dma-buf.c, I have to make sure that d_fsdata should
> > > always be under d_lock thus it will be protected. (In this posted patch
> > > there is one place(in dma_buf_set_name) that is missed, will update this
> > > in V3).
> > >
> > > >
> > > >> + if (dentry->d_fsdata)
> > > >> + ret = strlcpy(name, dentry->d_fsdata, DMA_BUF_NAME_LEN);
> > > >> + spin_unlock(&dentry->d_lock);
> > > >>
> > > >> return dynamic_dname(dentry, buffer, buflen, "/%s:%s",
> > > >> dentry->d_name.name, ret > 0 ? name : "");
> > > >
> > > > If the above check fails the name will be what? How could d_name.name
> > > > be valid but d_fsdata not be valid?
> > >
> > > In case of check fails, empty string "" is appended to the name by the
> > > code, ret > 0 ? name : "", ret is initialized to zero. Thus the name
> > > string will be like "/dmabuf:".
> >
> > So multiple objects can have the same "name" if this happens to multiple
> > ones at once?
> >
> > > Regarding the validity of d_fsdata, we are setting the dmabuf's
> > > dentry->d_fsdata to NULL in the dma_buf_release() thus can go invalid if
> > > that dmabuf is in the free path.
> >
> > Why are we allowing the name to be set if the dmabuf is on the free path
> > at all? Shouldn't that be the real fix here?
> >
> > > >> @@ -80,12 +79,16 @@ static int dma_buf_fs_init_context(struct fs_context *fc)
> > > >> static int dma_buf_release(struct inode *inode, struct file *file)
> > > >> {
> > > >> struct dma_buf *dmabuf;
> > > >> + struct dentry *dentry = file->f_path.dentry;
> > > >>
> > > >> if (!is_dma_buf_file(file))
> > > >> return -EINVAL;
> > > >>
> > > >> dmabuf = file->private_data;
> > > >>
> > > >> + spin_lock(&dentry->d_lock);
> > > >> + dentry->d_fsdata = NULL;
> > > >> + spin_unlock(&dentry->d_lock);
> > > >> BUG_ON(dmabuf->vmapping_counter);
> > > >>
> > > >> /*
> > > >> @@ -343,6 +346,7 @@ static long dma_buf_set_name(struct dma_buf *dmabuf, const char __user *buf)
> > > >> }
> > > >> kfree(dmabuf->name);
> > > >> dmabuf->name = name;
> > > >> + dmabuf->file->f_path.dentry->d_fsdata = name;
> > > >
> > > > You are just changing the use of d_fsdata from being a pointer to the
> > > > dmabuf to being a pointer to the name string? What's to keep that name
> > > > string around and not have the same reference counting issues that the
> > > > dmabuf structure itself has? Who frees that string memory?
> > > >
> > >
> > > Yes, I am just storing the name string in the d_fsdata in place of
> > > dmabuf and this helps to get rid of any extra refcount requirement.
> > > Because the user passed name carried in the d_fsdata is copied to the
> > > local buffer in dmabuffs_dname under spin_lock(d_lock) and the same
> > > d_fsdata is set to NULL(under the d_lock only) when that dmabuf is in
> > > the release path. So, when d_fsdata is NULL, name string is not accessed
> > > from the dmabuffs_dname thus extra count is not required.
> > >
> > > String memory, stored in the dmabuf->name, is released from the
> > > dma_buf_release(). Flow will be like, It fist sets d_fsdata=NULL and
> > > then free the dmabuf->name.
> > >
> > > However from your comments I have realized that there is a race in this
> > > patch when using the name string between dma_buf_set_name() and
> > > dmabuffs_dname(). But, If the idea of passing the name string inplace of
> > > dmabuf in d_fsdata looks fine, I can update this next patch.
> >
> > I'll leave that to the dmabuf authors/maintainers, but it feels odd to
> > me...
>
> I have zero clue about fs internals. This all scares me, that's all. I
> know enough about lifetime bugs that if you don't deeply understand a
> subsystem, all that's guaranteed is that you will get it wrong.
Likewise, and that made me realise that the 'fix' may not be as
innocuous or quick.
I will try to check with some folks more experienced than me in the fs
domain and see what is the logical way to handle it.
>
> /me out
>
> Cheers, Daniel
>
Best,
Sumit.
Thank you for the reply.
On 5/13/2020 9:33 PM, Sumit Semwal wrote:
> On Wed, 13 May 2020 at 21:16, Daniel Vetter <[email protected]> wrote:
>>
>> On Wed, May 13, 2020 at 02:51:12PM +0200, Greg KH wrote:
>>> On Wed, May 13, 2020 at 05:40:26PM +0530, Charan Teja Kalla wrote:
>>>>
>>>> Thank you Greg for the comments.
>>>> On 5/12/2020 2:22 PM, Greg KH wrote:
>>>>> On Fri, May 08, 2020 at 12:11:03PM +0530, Charan Teja Reddy wrote:
>>>>>> The following race occurs while accessing the dmabuf object exported as
>>>>>> file:
>>>>>> P1 P2
>>>>>> dma_buf_release() dmabuffs_dname()
>>>>>> [say lsof reading /proc/<P1 pid>/fd/<num>]
>>>>>>
>>>>>> read dmabuf stored in dentry->d_fsdata
>>>>>> Free the dmabuf object
>>>>>> Start accessing the dmabuf structure
>>>>>>
>>>>>> In the above description, the dmabuf object freed in P1 is being
>>>>>> accessed from P2 which is resulting into the use-after-free. Below is
>>>>>> the dump stack reported.
>>>>>>
>>>>>> We are reading the dmabuf object stored in the dentry->d_fsdata but
>>>>>> there is no binding between the dentry and the dmabuf which means that
>>>>>> the dmabuf can be freed while it is being read from ->d_fsdata and
>>>>>> inuse. Reviews on the patch V1 says that protecting the dmabuf inuse
>>>>>> with an extra refcount is not a viable solution as the exported dmabuf
>>>>>> is already under file's refcount and keeping the multiple refcounts on
>>>>>> the same object coordinated is not possible.
>>>>>>
>>>>>> As we are reading the dmabuf in ->d_fsdata just to get the user passed
>>>>>> name, we can directly store the name in d_fsdata thus can avoid the
>>>>>> reading of dmabuf altogether.
>>>>>>
>>>>>> Call Trace:
>>>>>> kasan_report+0x12/0x20
>>>>>> __asan_report_load8_noabort+0x14/0x20
>>>>>> dmabuffs_dname+0x4f4/0x560
>>>>>> tomoyo_realpath_from_path+0x165/0x660
>>>>>> tomoyo_get_realpath
>>>>>> tomoyo_check_open_permission+0x2a3/0x3e0
>>>>>> tomoyo_file_open
>>>>>> tomoyo_file_open+0xa9/0xd0
>>>>>> security_file_open+0x71/0x300
>>>>>> do_dentry_open+0x37a/0x1380
>>>>>> vfs_open+0xa0/0xd0
>>>>>> path_openat+0x12ee/0x3490
>>>>>> do_filp_open+0x192/0x260
>>>>>> do_sys_openat2+0x5eb/0x7e0
>>>>>> do_sys_open+0xf2/0x180
>>>>>>
>>>>>> Fixes: bb2bb9030425 ("dma-buf: add DMA_BUF_SET_NAME ioctls")
>>>>>> Reported-by: [email protected]
>>>>>> Cc: <[email protected]> [5.3+]
>>>>>> Signed-off-by: Charan Teja Reddy <[email protected]>
>>>>>> ---
>>>>>>
>>>>>> Changes in v2:
>>>>>>
>>>>>> - Pass the user passed name in ->d_fsdata instead of dmabuf
>>>>>> - Improve the commit message
>>>>>>
>>>>>> Changes in v1: (https://patchwork.kernel.org/patch/11514063/)
>>>>>>
>>>>>> drivers/dma-buf/dma-buf.c | 17 ++++++++++-------
>>>>>> 1 file changed, 10 insertions(+), 7 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
>>>>>> index 01ce125..0071f7d 100644
>>>>>> --- a/drivers/dma-buf/dma-buf.c
>>>>>> +++ b/drivers/dma-buf/dma-buf.c
>>>>>> @@ -25,6 +25,7 @@
>>>>>> #include <linux/mm.h>
>>>>>> #include <linux/mount.h>
>>>>>> #include <linux/pseudo_fs.h>
>>>>>> +#include <linux/dcache.h>
>>>>>>
>>>>>> #include <uapi/linux/dma-buf.h>
>>>>>> #include <uapi/linux/magic.h>
>>>>>> @@ -40,15 +41,13 @@ struct dma_buf_list {
>>>>>>
>>>>>> static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen)
>>>>>> {
>>>>>> - struct dma_buf *dmabuf;
>>>>>> char name[DMA_BUF_NAME_LEN];
>>>>>> size_t ret = 0;
>>>>>>
>>>>>> - dmabuf = dentry->d_fsdata;
>>>>>> - dma_resv_lock(dmabuf->resv, NULL);
>>>>>> - if (dmabuf->name)
>>>>>> - ret = strlcpy(name, dmabuf->name, DMA_BUF_NAME_LEN);
>>>>>> - dma_resv_unlock(dmabuf->resv);
>>>>>> + spin_lock(&dentry->d_lock);
>>>>>
>>>>> Are you sure this lock always protects d_fsdata?
>>>>
>>>> I think yes. In the dma-buf.c, I have to make sure that d_fsdata should
>>>> always be under d_lock thus it will be protected. (In this posted patch
>>>> there is one place(in dma_buf_set_name) that is missed, will update this
>>>> in V3).
>>>>
>>>>>
>>>>>> + if (dentry->d_fsdata)
>>>>>> + ret = strlcpy(name, dentry->d_fsdata, DMA_BUF_NAME_LEN);
>>>>>> + spin_unlock(&dentry->d_lock);
>>>>>>
>>>>>> return dynamic_dname(dentry, buffer, buflen, "/%s:%s",
>>>>>> dentry->d_name.name, ret > 0 ? name : "");
>>>>>
>>>>> If the above check fails the name will be what? How could d_name.name
>>>>> be valid but d_fsdata not be valid?
>>>>
>>>> In case of check fails, empty string "" is appended to the name by the
>>>> code, ret > 0 ? name : "", ret is initialized to zero. Thus the name
>>>> string will be like "/dmabuf:".
>>>
>>> So multiple objects can have the same "name" if this happens to multiple
>>> ones at once?
Yes that it can happen.
>>>
>>>> Regarding the validity of d_fsdata, we are setting the dmabuf's
>>>> dentry->d_fsdata to NULL in the dma_buf_release() thus can go invalid if
>>>> that dmabuf is in the free path.
>>>
>>> Why are we allowing the name to be set if the dmabuf is on the free path
>>> at all? Shouldn't that be the real fix here?
I don't think that user setting the name is the problem but accessing
the ->name while dmabuf is on the free path. And given a dmabuf we don't
know If that is already in the free path.
>>>
>>>>>> @@ -80,12 +79,16 @@ static int dma_buf_fs_init_context(struct fs_context *fc)
>>>>>> static int dma_buf_release(struct inode *inode, struct file *file)
>>>>>> {
>>>>>> struct dma_buf *dmabuf;
>>>>>> + struct dentry *dentry = file->f_path.dentry;
>>>>>>
>>>>>> if (!is_dma_buf_file(file))
>>>>>> return -EINVAL;
>>>>>>
>>>>>> dmabuf = file->private_data;
>>>>>>
>>>>>> + spin_lock(&dentry->d_lock);
>>>>>> + dentry->d_fsdata = NULL;
>>>>>> + spin_unlock(&dentry->d_lock);
>>>>>> BUG_ON(dmabuf->vmapping_counter);
>>>>>>
>>>>>> /*
>>>>>> @@ -343,6 +346,7 @@ static long dma_buf_set_name(struct dma_buf *dmabuf, const char __user *buf)
>>>>>> }
>>>>>> kfree(dmabuf->name);
>>>>>> dmabuf->name = name;
>>>>>> + dmabuf->file->f_path.dentry->d_fsdata = name;
>>>>>
>>>>> You are just changing the use of d_fsdata from being a pointer to the
>>>>> dmabuf to being a pointer to the name string? What's to keep that name
>>>>> string around and not have the same reference counting issues that the
>>>>> dmabuf structure itself has? Who frees that string memory?
>>>>>
>>>>
>>>> Yes, I am just storing the name string in the d_fsdata in place of
>>>> dmabuf and this helps to get rid of any extra refcount requirement.
>>>> Because the user passed name carried in the d_fsdata is copied to the
>>>> local buffer in dmabuffs_dname under spin_lock(d_lock) and the same
>>>> d_fsdata is set to NULL(under the d_lock only) when that dmabuf is in
>>>> the release path. So, when d_fsdata is NULL, name string is not accessed
>>>> from the dmabuffs_dname thus extra count is not required.
>>>>
>>>> String memory, stored in the dmabuf->name, is released from the
>>>> dma_buf_release(). Flow will be like, It fist sets d_fsdata=NULL and
>>>> then free the dmabuf->name.
>>>>
>>>> However from your comments I have realized that there is a race in this
>>>> patch when using the name string between dma_buf_set_name() and
>>>> dmabuffs_dname(). But, If the idea of passing the name string inplace of
>>>> dmabuf in d_fsdata looks fine, I can update this next patch.
>>>
>>> I'll leave that to the dmabuf authors/maintainers, but it feels odd to
>>> me...
>>
>> I have zero clue about fs internals. This all scares me, that's all. I
>> know enough about lifetime bugs that if you don't deeply understand a
>> subsystem, all that's guaranteed is that you will get it wrong.
>
> Likewise, and that made me realise that the 'fix' may not be as
> innocuous or quick.
>
> I will try to check with some folks more experienced than me in the fs
> domain and see what is the logical way to handle it.
>
BTW, I also would like to bring your notice that we have seen
sleep-while-atomic bug in the dmabuffs_dname() because of the mutex
used. It is caused from the SELinux permissions checks. I think SELinux
tries to validate the inherited files from fork + exec and in doing so,
it has to traverse the files. So, it relies on iterate_fd() (which
traverse files under spin_lock) and call
match_file(security/selinux/hooks.c) where the permission checks happen
and then dumps the information logs using dump_common_audit_data(). This
function is what actually trying to get the file path name and thus use
d_path(). If the file check happen on the dmabuf's fd, then it ends up
in ->dmabuffs_dname() under spin_lock(). So, flow will be like:
flush_unauthorized_files()
iterate_fd()
spin_lock()
match_file()
file_has_perm()
avc_has_perm()
avc_audit()
slow_avc_audit()
common_lsm_audit()
dump_common_audit_data()
audit_log_d_path()
d_path()
dmabuffs_dname()
mutex_lock()--> Sleep while atomic.
So, we have to remove the use of mutex in the dmabuffs_dname(which is
another bug in the existing code).
Call trace captured is as below:
___might_sleep+0x204/0x208
__might_sleep+0x50/0x88
__mutex_lock_common+0x5c/0x1068
__mutex_lock_common+0x5c/0x1068
mutex_lock_nested+0x40/0x50
dmabuffs_dname+0xa0/0x170
d_path+0x84/0x290
audit_log_d_path+0x74/0x130
common_lsm_audit+0x334/0x6e8
slow_avc_audit+0xb8/0xf8
avc_has_perm+0x154/0x218
file_has_perm+0x70/0x180
match_file+0x60/0x78
iterate_fd+0x128/0x168
selinux_bprm_committing_creds+0x178/0x248
security_bprm_committing_creds+0x30/0x48
install_exec_creds+0x1c/0x68
load_elf_binary+0x3a4/0x14e0
search_binary_handler+0xb0/0x1e0
>>
>> /me out
>>
>> Cheers, Daniel
>>
>
> Best,
> Sumit.
>
--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
Forum, a Linux Foundation Collaborative Project