Internally, UBD treats each physical IO segment as a separate command to
be submitted in the execution pipe.  If the pipe returns a transient
error after a few segments have already been written, UBD will tell the
block layer to requeue the request, but there is no way to reclaim the
segments already submitted.  When a new attempt to dispatch the request
is done, those segments already submitted will get duplicated, causing
the WARN_ON below in the best case, and potentially data corruption.

In my system, running a UML instance with 2GB of RAM and a 50M UBD disk,
I can reproduce the WARN_ON by simply running mkfs.fvat against the
disk on a freshly booted system.

There are a few ways to workaround this, like reducing the pressure on
the pipe by reducing the queue depth, which almost eliminates the
occurrence of the problem, increasing the pipe buffer size on the host
system, or by limiting the request to one physical segment, which causes
the block layer to submit way more requests to resolve a single
operation, which can have theoretical performance implications, that I
was *unable* to observe using fio.

Instead, this patch modifies the format of a UBD command, such that all
segments are sent through a single element in the communication pipe,
turning the command submission atomic from the point of view of the
block layer.  The new format has a variable size, depending on the
number of elements, and looks like this:

+------------+-----------+-----------+------------
| cmd_header | segment 0 | segment 1 | segment ...
+------------+-----------+-----------+------------

With this format, we push a pointer to cmd_header in the submission
pipe.

This has the advantage of reducing the memory footprint of executing a
single request, since it allow us to merge some fields in the header.
It is possible to reduce even further each segment memory footprint, by
merging bitmap_words and cow_offset, for instance, but this is not the
focus of this patch and is left as future work.  One issue with the
patch is that for a big number of segments, we now perform one big
memory allocation instead of multiple small ones, but I wasn't able to
trigger any real issues or -ENOMEM because of this change, that wouldn't
be reproduced otherwise.

This was tested using fio with the verify-crc32 option, and by running a
ext4 filesystem over this UBD device.

The observed WARN_ON was:

------------[ cut here ]------------
WARNING: CPU: 0 PID: 0 at lib/refcount.c:28 refcount_warn_saturate+0x13f/0x141
refcount_t: underflow; use-after-free.
Modules linked in:
CPU: 0 PID: 0 Comm: swapper Not tainted 5.5.0-rc6-00002-g2a5bb2cf

 #346
Stack:
 6084eed0 6063dc77 00000009 6084ef60
 00000000 604b8d9f 6084eee0 6063dcbc
 6084ef40 6006ab8d e013d780 1c00000000
Call Trace:
 [<600a0c1c>] ? printk+0x0/0x94
 [<6004a888>] show_stack+0x13b/0x155
 [<6063dc77>] ? dump_stack_print_info+0xdf/0xe8
 [<604b8d9f>] ? refcount_warn_saturate+0x13f/0x141
 [<6063dcbc>] dump_stack+0x2a/0x2c
 [<6006ab8d>] __warn+0x107/0x134
 [<6008da6c>] ? wake_up_process+0x17/0x19
 [<60487628>] ? blk_queue_max_discard_sectors+0x0/0xd
 [<6006b05f>] warn_slowpath_fmt+0xd1/0xdf
 [<6006af8e>] ? warn_slowpath_fmt+0x0/0xdf
 [<600acc14>] ? raw_read_seqcount_begin.constprop.0+0x0/0x15
 [<600619ae>] ? os_nsecs+0x1d/0x2b
 [<604b8d9f>] refcount_warn_saturate+0x13f/0x141
 [<6048bc8f>] refcount_sub_and_test.constprop.0+0x2f/0x37
 [<6048c8de>] blk_mq_free_request+0xf1/0x10d
 [<6048ca06>] __blk_mq_end_request+0x10c/0x114
 [<6005ac0f>] ubd_intr+0xb5/0x169
 [<600a1a37>] __handle_irq_event_percpu+0x6b/0x17e
 [<600a1b70>] handle_irq_event_percpu+0x26/0x69
 [<600a1bd9>] handle_irq_event+0x26/0x34
 [<600a1bb3>] ? handle_irq_event+0x0/0x34
 [<600a5186>] ? unmask_irq+0x0/0x37
 [<600a57e6>] handle_edge_irq+0xbc/0xd6
 [<600a131a>] generic_handle_irq+0x21/0x29
 [<60048f6e>] do_IRQ+0x39/0x54
 [...]
---[ end trace c6e7444e55386c0f ]---

Reported-by: Martyn Welch <martyn@collabora.com>
Signed-off-by: Gabriel Krisman Bertazi <krisman@collabora.com>