[Patch 0/1] [focal/linux-azure-cvm] hv/bounce buffer: Fix a race that can fail disk detection

[Tim Gardner]

BugLink: https://bugs.launchpad.net/bugs/1971164

SRU Justification

[Impact]

The linux-azure-cvm kernel (e.g. Ubuntu-azure-cvm-5.4.0-1078.81+cvm1) has a race
condition bug in the Linux vmbus bounce buffer code (drivers/hv/hv_bounce.c), and
as a result somtimes the kenrel fails to detect some of the SCSI disks, and the
Linux dmesg log may show one of the 2 messages:

#1: [ 2.995732] sd 3:0:0:3: [sdd] Sector size 0 reported, assuming 512.

#2: [ 3.651293] scsi host3: scsi scan: INQUIRY result too short (5), using 36

Sometimes I see a strange call-trace (the 'order's is 18, if I print it)
2022-04-26T20:10:18,398144+00:00 kmalloc_order_trace+0x1e/0x80
2022-04-26T20:10:18,398147+00:00 __kmalloc+0x3ae/0x4c0
2022-04-26T20:10:18,398150+00:00 __scsi_scan_target+0x283/0x590
2022-04-26T20:10:18,398155+00:00 scsi_scan_channel.part.16+0x62/0x80
2022-04-26T20:10:18,398158+00:00 scsi_scan_host_selected+0xd5/0x150
2022-04-26T20:10:18,398160+00:00 store_scan+0xc8/0xe0
(This is very strange because 'order 18' means (1 << 18) * 4096 bytes = 1GBytes.)

After some investigation, we eventually got the root cause and made a fix:
https://github.com/dcui/linux-azure-cvm/commit/ddde4dc33242794000e1d9667a5f9cfa31c15fdf

With the fix, we no longer see the above strange symptoms.
Please include the fix into the next release of the v5.4 linux-azure-cvm kernel. Thanks!

[Test case]

Microsoft tested

[Where things could go wrong]

Some SCSI drives may continue to go undetected.

[Other Info]

SF: #00335631