In scenarios of continuous and parallel usage of multiple FFS interfaces
and concurrent adb operations (e.g., adb root, adb reboot), there's a
chance that ffs_epfile_async_io_complete() might be processed after
ffs_epfile_release(). This could lead to a NULL pointer dereference of
ffs when accessing the ffs pointer in ffs_epfile_async_io_complete(), as
ffs is freed as part of ffs_epfile_release(). This epfile release is
part of file operation and is triggered when user space daemons restart
themselves or a reboot is initiated.
Fix this issue by adding a NULL pointer check for ffs in
ffs_epfile_async_io_complete().
[ 9981.393115] Unable to handle kernel NULL pointer dereference at virtual address 00000000000001e0
[ 9981.402854] Mem abort info:
..
[ 9981.532540] Hardware name: Qualcomm Technologies,
[ 9981.540579] pstate: 204000c5 (nzCv daIF +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[ 9981.548438] pc : ffs_epfile_async_io_complete+0x38/0x4c
[ 9981.554529] lr : usb_gadget_giveback_request+0x30/0xd0
..
[ 9981.645057] Call trace:
[ 9981.648282] ffs_epfile_async_io_complete+0x38/0x4c
[ 9981.654004] usb_gadget_giveback_request+0x30/0xd0
[ 9981.659637] dwc3_gadget_endpoint_trbs_complete+0x1a8/0x48c
[ 9981.666074] dwc3_process_event_entry+0x378/0x648
[ 9981.671622] dwc3_process_event_buf+0x6c/0x288
[ 9981.676903] dwc3_thread_interrupt+0x3c/0x68
[ 9981.682003] irq_thread_fn+0x2c/0x8c
[ 9981.686388] irq_thread+0x198/0x2ac
[ 9981.690685] kthread+0x154/0x218
[ 9981.694717] ret_from_fork+0x10/0x20
Signed-off-by: Selvarasu Ganesan <[email protected]>
---
drivers/usb/gadget/function/f_fs.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/drivers/usb/gadget/function/f_fs.c b/drivers/usb/gadget/function/f_fs.c
index be3851cffb73..d8c8e88628f9 100644
--- a/drivers/usb/gadget/function/f_fs.c
+++ b/drivers/usb/gadget/function/f_fs.c
@@ -849,7 +849,9 @@ static void ffs_epfile_async_io_complete(struct usb_ep *_ep,
usb_ep_free_request(_ep, req);
INIT_WORK(&io_data->work, ffs_user_copy_worker);
- queue_work(ffs->io_completion_wq, &io_data->work);
+
+ if (ffs && ffs->io_completion_wq)
+ queue_work(ffs->io_completion_wq, &io_data->work);
}
static void __ffs_epfile_read_buffer_free(struct ffs_epfile *epfile)
--
2.17.1
On Thu, Feb 22, 2024 at 09:48:09PM -0800, Selvarasu Ganesan wrote:
> In scenarios of continuous and parallel usage of multiple FFS interfaces
> and concurrent adb operations (e.g., adb root, adb reboot), there's a
> chance that ffs_epfile_async_io_complete() might be processed after
> ffs_epfile_release(). This could lead to a NULL pointer dereference of
> ffs when accessing the ffs pointer in ffs_epfile_async_io_complete(), as
> ffs is freed as part of ffs_epfile_release(). This epfile release is
> part of file operation and is triggered when user space daemons restart
> themselves or a reboot is initiated.
>
> Fix this issue by adding a NULL pointer check for ffs in
> ffs_epfile_async_io_complete().
>
> [ 9981.393115] Unable to handle kernel NULL pointer dereference at virtual address 00000000000001e0
> [ 9981.402854] Mem abort info:
> ...
> [ 9981.532540] Hardware name: Qualcomm Technologies,
> [ 9981.540579] pstate: 204000c5 (nzCv daIF +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
> [ 9981.548438] pc : ffs_epfile_async_io_complete+0x38/0x4c
> [ 9981.554529] lr : usb_gadget_giveback_request+0x30/0xd0
> ...
> [ 9981.645057] Call trace:
> [ 9981.648282] ffs_epfile_async_io_complete+0x38/0x4c
> [ 9981.654004] usb_gadget_giveback_request+0x30/0xd0
> [ 9981.659637] dwc3_gadget_endpoint_trbs_complete+0x1a8/0x48c
> [ 9981.666074] dwc3_process_event_entry+0x378/0x648
> [ 9981.671622] dwc3_process_event_buf+0x6c/0x288
> [ 9981.676903] dwc3_thread_interrupt+0x3c/0x68
> [ 9981.682003] irq_thread_fn+0x2c/0x8c
> [ 9981.686388] irq_thread+0x198/0x2ac
> [ 9981.690685] kthread+0x154/0x218
> [ 9981.694717] ret_from_fork+0x10/0x20
>
> Signed-off-by: Selvarasu Ganesan <[email protected]>
What commit id does this fix? Should it go to stable kernels?
> ---
> drivers/usb/gadget/function/f_fs.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/usb/gadget/function/f_fs.c b/drivers/usb/gadget/function/f_fs.c
> index be3851cffb73..d8c8e88628f9 100644
> --- a/drivers/usb/gadget/function/f_fs.c
> +++ b/drivers/usb/gadget/function/f_fs.c
> @@ -849,7 +849,9 @@ static void ffs_epfile_async_io_complete(struct usb_ep *_ep,
> usb_ep_free_request(_ep, req);
>
> INIT_WORK(&io_data->work, ffs_user_copy_worker);
> - queue_work(ffs->io_completion_wq, &io_data->work);
> +
> + if (ffs && ffs->io_completion_wq)
> + queue_work(ffs->io_completion_wq, &io_data->work);
What happens if ffs->io_compleation_wq goes away right after you test
it but before you call queue_work()?
Where is the locking here to prevent that?
thanks,
greg k-h
On 2/23/2024 11:28 AM, Greg KH wrote:
> On Thu, Feb 22, 2024 at 09:48:09PM -0800, Selvarasu Ganesan wrote:
>> In scenarios of continuous and parallel usage of multiple FFS interfaces
>> and concurrent adb operations (e.g., adb root, adb reboot), there's a
>> chance that ffs_epfile_async_io_complete() might be processed after
>> ffs_epfile_release(). This could lead to a NULL pointer dereference of
>> ffs when accessing the ffs pointer in ffs_epfile_async_io_complete(), as
>> ffs is freed as part of ffs_epfile_release(). This epfile release is
>> part of file operation and is triggered when user space daemons restart
>> themselves or a reboot is initiated.
>>
>> Fix this issue by adding a NULL pointer check for ffs in
>> ffs_epfile_async_io_complete().
>>
>> [ 9981.393115] Unable to handle kernel NULL pointer dereference at virtual address 00000000000001e0
>> [ 9981.402854] Mem abort info:
>> ...
>> [ 9981.532540] Hardware name: Qualcomm Technologies,
>> [ 9981.540579] pstate: 204000c5 (nzCv daIF +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>> [ 9981.548438] pc : ffs_epfile_async_io_complete+0x38/0x4c
>> [ 9981.554529] lr : usb_gadget_giveback_request+0x30/0xd0
>> ...
>> [ 9981.645057] Call trace:
>> [ 9981.648282] ffs_epfile_async_io_complete+0x38/0x4c
>> [ 9981.654004] usb_gadget_giveback_request+0x30/0xd0
>> [ 9981.659637] dwc3_gadget_endpoint_trbs_complete+0x1a8/0x48c
>> [ 9981.666074] dwc3_process_event_entry+0x378/0x648
>> [ 9981.671622] dwc3_process_event_buf+0x6c/0x288
>> [ 9981.676903] dwc3_thread_interrupt+0x3c/0x68
>> [ 9981.682003] irq_thread_fn+0x2c/0x8c
>> [ 9981.686388] irq_thread+0x198/0x2ac
>> [ 9981.690685] kthread+0x154/0x218
>> [ 9981.694717] ret_from_fork+0x10/0x20
>>
>> Signed-off-by: Selvarasu Ganesan <[email protected]>
>
> What commit id does this fix? Should it go to stable kernels?
Fixes: 2e4c7553cd6f9 ("usb: gadget: f_fs: add aio support"). Yes it's
required to propagate to stable kernel as well.
>
>> ---
>> drivers/usb/gadget/function/f_fs.c | 4 +++-
>> 1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/usb/gadget/function/f_fs.c b/drivers/usb/gadget/function/f_fs.c
>> index be3851cffb73..d8c8e88628f9 100644
>> --- a/drivers/usb/gadget/function/f_fs.c
>> +++ b/drivers/usb/gadget/function/f_fs.c
>> @@ -849,7 +849,9 @@ static void ffs_epfile_async_io_complete(struct usb_ep *_ep,
>> usb_ep_free_request(_ep, req);
>>
>> INIT_WORK(&io_data->work, ffs_user_copy_worker);
>> - queue_work(ffs->io_completion_wq, &io_data->work);
>> +
>> + if (ffs && ffs->io_completion_wq)
>> + queue_work(ffs->io_completion_wq, &io_data->work);
>
> What happens if ffs->io_compleation_wq goes away right after you test
> it but before you call queue_work()?
>
> Where is the locking here to prevent that?
>
> thanks,
>
> greg k-h
Hi Greg,
Thank you for your feedback. I understand your concern about the
potential race condition with ffs->io_completion_wq. I’m considering
introducing a lock to protect this section of the code, but I wanted to
get your opinion on this.
In the f_fs.c driver, there are pre-existing locks. Would it be suitable
to utilize these locks, or do you suggest the creation of a new lock
specifically for ffs->io_completion_wq? We anticipate a performance
impact if we use the existing lock, as it might be held by different
threads. What are your thoughts on this?"
Here’s what the code might look like with a new lock:
static void ffs_epfile_async_io_complete(struct usb_ep *_ep,
struct usb_request *req)
{
...
spin_lock(&ffs->new_lock);
if (ffs && ffs->io_completion_wq)
queue_work(ffs->io_completion_wq, &io_data->work);
spin_unlock(&ffs->new_lock);
...
}
static void ffs_data_put(struct ffs_data *ffs) {
..
destroy_workqueue(ffs->io_completion_wq);
kfree(ffs->dev_name);
spin_lock(&ffs->new_lock);
kfree(ffs);
spin_unlock(&ffs->new_lock);
..
}
Thanks,
Selva
On Fri, Feb 23, 2024 at 05:05:59PM +0530, Selvarasu Ganesan wrote:
>
> On 2/23/2024 11:28 AM, Greg KH wrote:
> > On Thu, Feb 22, 2024 at 09:48:09PM -0800, Selvarasu Ganesan wrote:
> > > In scenarios of continuous and parallel usage of multiple FFS interfaces
> > > and concurrent adb operations (e.g., adb root, adb reboot), there's a
> > > chance that ffs_epfile_async_io_complete() might be processed after
> > > ffs_epfile_release(). This could lead to a NULL pointer dereference of
> > > ffs when accessing the ffs pointer in ffs_epfile_async_io_complete(), as
> > > ffs is freed as part of ffs_epfile_release(). This epfile release is
> > > part of file operation and is triggered when user space daemons restart
> > > themselves or a reboot is initiated.
> > >
> > > Fix this issue by adding a NULL pointer check for ffs in
> > > ffs_epfile_async_io_complete().
> > >
> > > [ 9981.393115] Unable to handle kernel NULL pointer dereference at virtual address 00000000000001e0
> > > [ 9981.402854] Mem abort info:
> > > ...
> > > [ 9981.532540] Hardware name: Qualcomm Technologies,
> > > [ 9981.540579] pstate: 204000c5 (nzCv daIF +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
> > > [ 9981.548438] pc : ffs_epfile_async_io_complete+0x38/0x4c
> > > [ 9981.554529] lr : usb_gadget_giveback_request+0x30/0xd0
> > > ...
> > > [ 9981.645057] Call trace:
> > > [ 9981.648282] ffs_epfile_async_io_complete+0x38/0x4c
> > > [ 9981.654004] usb_gadget_giveback_request+0x30/0xd0
> > > [ 9981.659637] dwc3_gadget_endpoint_trbs_complete+0x1a8/0x48c
> > > [ 9981.666074] dwc3_process_event_entry+0x378/0x648
> > > [ 9981.671622] dwc3_process_event_buf+0x6c/0x288
> > > [ 9981.676903] dwc3_thread_interrupt+0x3c/0x68
> > > [ 9981.682003] irq_thread_fn+0x2c/0x8c
> > > [ 9981.686388] irq_thread+0x198/0x2ac
> > > [ 9981.690685] kthread+0x154/0x218
> > > [ 9981.694717] ret_from_fork+0x10/0x20
> > >
> > > Signed-off-by: Selvarasu Ganesan <[email protected]>
> >
> > What commit id does this fix? Should it go to stable kernels?
>
> Fixes: 2e4c7553cd6f9 ("usb: gadget: f_fs: add aio support"). Yes it's
> required to propagate to stable kernel as well.
Great, when you resend the next version, please include both proper
tags.
> > > ---
> > > drivers/usb/gadget/function/f_fs.c | 4 +++-
> > > 1 file changed, 3 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/usb/gadget/function/f_fs.c b/drivers/usb/gadget/function/f_fs.c
> > > index be3851cffb73..d8c8e88628f9 100644
> > > --- a/drivers/usb/gadget/function/f_fs.c
> > > +++ b/drivers/usb/gadget/function/f_fs.c
> > > @@ -849,7 +849,9 @@ static void ffs_epfile_async_io_complete(struct usb_ep *_ep,
> > > usb_ep_free_request(_ep, req);
> > > INIT_WORK(&io_data->work, ffs_user_copy_worker);
> > > - queue_work(ffs->io_completion_wq, &io_data->work);
> > > +
> > > + if (ffs && ffs->io_completion_wq)
> > > + queue_work(ffs->io_completion_wq, &io_data->work);
> >
> > What happens if ffs->io_compleation_wq goes away right after you test
> > it but before you call queue_work()?
> >
> > Where is the locking here to prevent that?
> >
> > thanks,
> >
> > greg k-h
>
> Hi Greg,
>
> Thank you for your feedback. I understand your concern about the
> potential race condition with ffs->io_completion_wq. I’m considering
> introducing a lock to protect this section of the code, but I wanted to
> get your opinion on this.
> In the f_fs.c driver, there are pre-existing locks. Would it be suitable to
> utilize these locks, or do you suggest the creation of a new lock
> specifically for ffs->io_completion_wq? We anticipate a performance impact
> if we use the existing lock, as it might be held by different
> threads. What are your thoughts on this?"
Test it out yourself and see what works best!
thanks,
greg k-h
On 2/23/24 4:35 AM, Selvarasu Ganesan wrote:
> Here?s what the code might look like with a new lock:
>
> static void ffs_epfile_async_io_complete(struct usb_ep *_ep,
> struct usb_request *req)
> {
> ....
> spin_lock(&ffs->new_lock);
> if (ffs && ffs->io_completion_wq)
> queue_work(ffs->io_completion_wq, &io_data->work);
> spin_unlock(&ffs->new_lock);
> ....
> }
>
>
>
> static void ffs_data_put(struct ffs_data *ffs) {
> ...
> destroy_workqueue(ffs->io_completion_wq);
> kfree(ffs->dev_name);
> spin_lock(&ffs->new_lock);
> kfree(ffs);
> spin_unlock(&ffs->new_lock);
> ...
> }
This obviously won't work at all, and it's not the right way to fix it
at all. It needs a ref count.
--
Jens Axboe