Re: [PATCH 1/2] uio: Add late_release callback to uio_info

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 11.02.21 07:51, Greg Kroah-Hartman wrote:
On Wed, Feb 10, 2021 at 08:57:11PM +0100, Bodo Stroesser wrote:
On 10.02.21 20:47, Greg Kroah-Hartman wrote:
On Wed, Feb 10, 2021 at 08:40:30PM +0100, Bodo Stroesser wrote:
If uio_unregister_device() is called while userspace daemon
still holds the uio device open or mmap'ed, uio will not call
uio_info->release() on later close / munmap.

At least one user of uio (tcmu) should not free resources (pages
allocated by tcmu which are mmap'ed to userspace) while uio
device still is open, because that could cause userspace daemon
to be killed by SIGSEGV or SIGBUS. Therefore tcmu frees the
pages only after it called uio_unregister_device _and_ the device
was closed.
So, uio not calling uio_info->release causes trouble.
tcmu currently leaks memory in that case.

Just waiting for userspace daemon to exit before calling
uio_unregister_device I think is not the right solution, because
daemon would not become aware of kernel code wanting to destroy
the uio device.
After uio_unregister_device was called, reading or writing the
uio device returns -EIO, which normally results in daemon exit.

This patch adds new callback pointer 'late_release' to struct
uio_info. If uio user sets this callback, it will be called by
uio if userspace closes / munmaps the device after
uio_unregister_device was executed.

That way we can use uio_unregister_device() to notify userspace
that we are going to destroy the device, but still get a call
to late_release when uio device is finally closed.

Signed-off-by: Bodo Stroesser <bostroesser@xxxxxxxxx>
---
   Documentation/driver-api/uio-howto.rst | 10 ++++++++++
   drivers/uio/uio.c                      |  4 ++++
   include/linux/uio_driver.h             |  4 ++++
   3 files changed, 18 insertions(+)

diff --git a/Documentation/driver-api/uio-howto.rst b/Documentation/driver-api/uio-howto.rst
index 907ffa3b38f5..a2d57a7d623a 100644
--- a/Documentation/driver-api/uio-howto.rst
+++ b/Documentation/driver-api/uio-howto.rst
@@ -265,6 +265,16 @@ the members are required, others are optional.
      function. The parameter ``irq_on`` will be 0 to disable interrupts
      and 1 to enable them.
+-  ``int (*late_release)(struct uio_info *info, struct inode *inode)``:
+   Optional. If you define your own :c:func:`open()`, you will
+   in certain cases also want a custom :c:func:`late_release()`
+   function. If uio device is unregistered - by calling
+   :c:func:`uio_unregister_device()` - while it is open or mmap'ed by
+   userspace, the custom :c:func:`release()` function will not be
+   called when userspace later closes the device. An optionally
+   specified :c:func:`late_release()` function will be called in that
+   situation.
+
   Usually, your device will have one or more memory regions that can be
   mapped to user space. For each region, you have to set up a
   ``struct uio_mem`` in the ``mem[]`` array. Here's a description of the
diff --git a/drivers/uio/uio.c b/drivers/uio/uio.c
index ea96e319c8a0..0b2636f8d373 100644
--- a/drivers/uio/uio.c
+++ b/drivers/uio/uio.c
@@ -532,6 +532,8 @@ static int uio_release(struct inode *inode, struct file *filep)
   	mutex_lock(&idev->info_lock);
   	if (idev->info && idev->info->release)
   		ret = idev->info->release(idev->info, inode);
+	else if (idev->late_info && idev->late_info->late_release)
+		ret = idev->late_info->late_release(idev->late_info, inode);
   	mutex_unlock(&idev->info_lock);

Why can't release() be called here?  Why doesn't your driver define a
release() if it cares about this information?  Why do we need 2
different callbacks that fire at exactly the same time?

This feels really wrong.

greg k-h


tcmu has a release callback. But uio can't call it after
uio_unregister_device was executed, because in uio_unregister_device
uio sets the uio_device::info to NULL.

As it should because the driver could then be gone.  It should NEVER
call back into it again.

OTOH, uio does try_module_get(idev->owner) in uio_open before calling
the driver's open callback and module_put(idev_owner) in uio_release
after calling driver's release callback. So driver's release callback
is guaranteed to exist until last release is done.

Apart from that, tcmu also has an uio_info::mmap callback. In that
callback it installs its own vm_operations_struct::fault handler.
This handler also can happen to be called as long as userspace holds
the uio device mmap'ed. I think, this is not a problem due to the
above mentioned mechanism.

tcmu just has to ensure, that the tcmu device, which contains the uio_info - is kept until the final release call happens. Unfortunately
this call will not happen if uio device is open during
uio_unregister_device. That's why tcmu sometimes leaks memory.


So, uio would never call both callbacks for the same release action,
but would call release before uio_unregister_device is executed, and
late_release after that.

That's not ok.

Of course it would be good for tcmu if uio would call uio_info:release even
after uio_unregister_device, but changing this AFAICS could cause
trouble in other drivers using uio.

You are confusing two different lifetime rules here it seems.  One is
the char device and one is the struct device.  They work independently
as different users affect them.
I'm not sure I get your point.


So if one is removed from the system, do not try to keep a callback to
it, otherwise you will crash.

That's why I tried to change uio in a compatible way, so other drivers
based on it are not afflicted by the change. I saw, that some drivers
based on uio free their resources directly after calling
uio_unregister_device. Executing their release callback later would
definitely cause trouble.


And why is scsi using the uio driver in the first place?  That feels
really odd to me.  Why not just make a "real" driver if you want to
somehow tie these two lifetimes together?

Why tcmu driver is based on uio I don't know. I inherited the driver as
it is. Maybe it would have been better to not base it on uio, I don't
know. But changing this now would cause an API change for all existing
userspace apps, e.g. tcmu-runner. I think I should avoid that and
therefore have to find an acceptable solution for the tcmu/uio
combination.


thanks,

greg k-h




[Index of Archives]     [Linux SCSI]     [Kernel Newbies]     [Linux SCSI Target Infrastructure]     [Share Photos]     [IDE]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux ATA RAID]     [Linux IIO]     [Device Mapper]

  Powered by Linux