xen: mark local pages as FOREIGN in the m2p_override
authorStefano Stabellini <stefano.stabellini@eu.citrix.com>
Wed, 23 May 2012 17:57:20 +0000 (18:57 +0100)
committerKonrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Thu, 14 Jun 2012 18:03:41 +0000 (14:03 -0400)
commitb9e0d95c041ca2d7ad297ee37c2e9cfab67a188f
tree0b055245bfb1493545b5f4d5e7ba3a9fdade81cf
parent5e626254206a709c6e937f3dda69bf26c7344f6f
xen: mark local pages as FOREIGN in the m2p_override

When the frontend and the backend reside on the same domain, even if we
add pages to the m2p_override, these pages will never be returned by
mfn_to_pfn because the check "get_phys_to_machine(pfn) != mfn" will
always fail, so the pfn of the frontend will be returned instead
(resulting in a deadlock because the frontend pages are already locked).

INFO: task qemu-system-i38:1085 blocked for more than 120 seconds.
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
qemu-system-i38 D ffff8800cfc137c0     0  1085      1 0x00000000
 ffff8800c47ed898 0000000000000282 ffff8800be4596b0 00000000000137c0
 ffff8800c47edfd8 ffff8800c47ec010 00000000000137c0 00000000000137c0
 ffff8800c47edfd8 00000000000137c0 ffffffff82213020 ffff8800be4596b0
Call Trace:
 [<ffffffff81101ee0>] ? __lock_page+0x70/0x70
 [<ffffffff81a0fdd9>] schedule+0x29/0x70
 [<ffffffff81a0fe80>] io_schedule+0x60/0x80
 [<ffffffff81101eee>] sleep_on_page+0xe/0x20
 [<ffffffff81a0e1ca>] __wait_on_bit_lock+0x5a/0xc0
 [<ffffffff81101ed7>] __lock_page+0x67/0x70
 [<ffffffff8106f750>] ? autoremove_wake_function+0x40/0x40
 [<ffffffff811867e6>] ? bio_add_page+0x36/0x40
 [<ffffffff8110b692>] set_page_dirty_lock+0x52/0x60
 [<ffffffff81186021>] bio_set_pages_dirty+0x51/0x70
 [<ffffffff8118c6b4>] do_blockdev_direct_IO+0xb24/0xeb0
 [<ffffffff811e71a0>] ? ext3_get_blocks_handle+0xe00/0xe00
 [<ffffffff8118ca95>] __blockdev_direct_IO+0x55/0x60
 [<ffffffff811e71a0>] ? ext3_get_blocks_handle+0xe00/0xe00
 [<ffffffff811e91c8>] ext3_direct_IO+0xf8/0x390
 [<ffffffff811e71a0>] ? ext3_get_blocks_handle+0xe00/0xe00
 [<ffffffff81004b60>] ? xen_mc_flush+0xb0/0x1b0
 [<ffffffff81104027>] generic_file_aio_read+0x737/0x780
 [<ffffffff813bedeb>] ? gnttab_map_refs+0x15b/0x1e0
 [<ffffffff811038f0>] ? find_get_pages+0x150/0x150
 [<ffffffff8119736c>] aio_rw_vect_retry+0x7c/0x1d0
 [<ffffffff811972f0>] ? lookup_ioctx+0x90/0x90
 [<ffffffff81198856>] aio_run_iocb+0x66/0x1a0
 [<ffffffff811998b8>] do_io_submit+0x708/0xb90
 [<ffffffff81199d50>] sys_io_submit+0x10/0x20
 [<ffffffff81a18d69>] system_call_fastpath+0x16/0x1b

The explanation is in the comment within the code:

We need to do this because the pages shared by the frontend
(xen-blkfront) can be already locked (lock_page, called by
do_read_cache_page); when the userspace backend tries to use them
with direct_IO, mfn_to_pfn returns the pfn of the frontend, so
do_blockdev_direct_IO is going to try to lock the same pages
again resulting in a deadlock.

A simplified call graph looks like this:

pygrub                          QEMU
-----------------------------------------------
do_read_cache_page              io_submit
  |                              |
lock_page                       ext3_direct_IO
                                 |
                                bio_add_page
                                 |
                                lock_page

Internally the xen-blkback uses m2p_add_override to swizzle (temporarily)
a 'struct page' to have a different MFN (so that it can point to another
guest). It also can easily find out whether another pfn corresponding
to the mfn exists in the m2p, and can set the FOREIGN bit
in the p2m, making sure that mfn_to_pfn returns the pfn of the backend.

This allows the backend to perform direct_IO on these pages, but as a
side effect prevents the frontend from using get_user_pages_fast on
them while they are being shared with the backend.

Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
arch/x86/xen/p2m.c