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

Re: [PATCH v2] xen/gntdev: Avoid blocking in unmap_grant_pages()



On 30.05.22 19:50, Demi Marie Obenour wrote:
On Mon, May 30, 2022 at 08:41:20AM +0200, Juergen Gross wrote:
On 25.05.22 20:41, Demi Marie Obenour wrote:
unmap_grant_pages() currently waits for the pages to no longer be used.
In https://github.com/QubesOS/qubes-issues/issues/7481, this lead to a
deadlock against i915: i915 was waiting for gntdev's MMU notifier to
finish, while gntdev was waiting for i915 to free its pages.  I also
believe this is responsible for various deadlocks I have experienced in
the past.

Avoid these problems by making unmap_grant_pages async.  This requires
making it return void, as any errors will not be available when the
function returns.  Fortunately, the only use of the return value is a
WARN_ON(), which can be replaced by a WARN_ON when the error is
detected.  Additionally, a failed call will not prevent further calls
from being made, but this is harmless.

Because unmap_grant_pages is now async, the grant handle will be sent to
INVALID_GRANT_HANDLE too late to prevent multiple unmaps of the same
handle.  Instead, a separate bool array is allocated for this purpose.
This wastes memory, but stuffing this information in padding bytes is
too fragile.  Furthermore, it is necessary to grab a reference to the
map before making the asynchronous call, and release the reference when
the call returns.

I think there is even more syncing needed:

- In the error path of gntdev_mmap() unmap_grant_pages() is being called and
   it is assumed, map is available afterwards again. This should be rather easy
   to avoid by adding a counter of active mappings to struct gntdev_grant_map
   (number of pages not being unmapped yet). In case this counter is not zero
   gntdev_mmap() should bail out early.

Is it possible to just unmap the pages directly here?  I don’t think
there can be any other users of these pages yet.  Userspace could race
against the unmap and cause a page fault, but that should just cause
userspace to get SIGSEGV or SIGBUS.  In any case this code can use the
sync version; if it gets blocked that’s userspace’s problem.

- gntdev_put_map() is calling unmap_grant_pages() in case the refcount has
   dropped to zero. This call can set the refcount to 1 again, so there is
   another delay needed before freeing map. I think unmap_grant_pages() should
   return in case the count of mapped pages is zero (see above), thus avoiding
   to increment the refcount of map if nothing is to be done. This would enable
   gntdev_put_map() to just return after the call of unmap_grant_pages() in case
   the refcount has been incremented again.

I will change this in v3, but I do wonder if gntdev is using the wrong
MMU notifier callback.  It seems that the appropriate callback is
actually release(): if I understand correctly, release() is called
precisely when the refcount on the physical page is about to drop to 0,
and that is what we want.

No, I don't think this is correct.

release() is being called when the refcount of the address space is about
to drop to 0. It has no page granularity.


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key

Attachment: OpenPGP_signature
Description: OpenPGP digital signature


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.