Hello, I have been having issues using tgtd with the AIO backing store.
When booting an Ubuntu 18.04 server image over iscsi the client
attempts to use the WRITE_SAME(16) operation, which tgtd's AIO backing
store does not support. This results in the following errors on the
client:
kernel: [ 66.542374] sd 11:0:0:1: [sda] tag#0 FAILED Result:
hostbyte=DID_OK driverbyte=DRIVER_SENSE
kernel: [ 66.542378] sd 11:0:0:1: [sda] tag#0 Sense Key : Hardware
Error [current]
kernel: [ 66.542382] sd 11:0:0:1: [sda] tag#0 Add. Sense: Internal
target failure
kernel: [ 66.542385] sd 11:0:0:1: [sda] tag#0 CDB: Write same(16) 93
00 00 00 00 00 00 42 90 68 00 00 3f a0 00 00
kernel: [ 66.542388] print_req_error: critical target error, dev sda,
sector 4362344
And on the server running tgtd:
tgtd[10965]: tgtd: bs_aio_cmd_submit(218) WRITE_SAME not yet supported
for AIO backend.
I see that other backing stores call bs_create_opcode_map() before
registering their template but not AIO. When applying the following
patch I no longer see these errors.
diff --git a/usr/bs_aio.c b/usr/bs_aio.c
index 4d84db4..00a232a 100644
--- a/usr/bs_aio.c
+++ b/usr/bs_aio.c
@@ -424,6 +424,49 @@ static struct backingstore_template aio_bst = {
__attribute__((constructor)) static void register_bs_module(void)
{
+ unsigned char opcodes[] = {
+ ALLOW_MEDIUM_REMOVAL,
+ COMPARE_AND_WRITE,
+ FORMAT_UNIT,
+ INQUIRY,
+ MAINT_PROTOCOL_IN,
+ MODE_SELECT,
+ MODE_SELECT_10,
+ MODE_SENSE,
+ MODE_SENSE_10,
+ ORWRITE_16,
+ PERSISTENT_RESERVE_IN,
+ PERSISTENT_RESERVE_OUT,
+ PRE_FETCH_10,
+ PRE_FETCH_16,
+ READ_10,
+ READ_12,
+ READ_16,
+ READ_6,
+ READ_CAPACITY,
+ RELEASE,
+ REPORT_LUNS,
+ REQUEST_SENSE,
+ RESERVE,
+ SEND_DIAGNOSTIC,
+ SERVICE_ACTION_IN,
+ START_STOP,
+ SYNCHRONIZE_CACHE,
+ SYNCHRONIZE_CACHE_16,
+ TEST_UNIT_READY,
+ UNMAP,
+ VERIFY_10,
+ VERIFY_12,
+ VERIFY_16,
+ WRITE_10,
+ WRITE_12,
+ WRITE_16,
+ WRITE_6,
+ WRITE_VERIFY,
+ WRITE_VERIFY_12,
+ WRITE_VERIFY_16
+ };
+ bs_create_opcode_map(&aio_bst, opcodes, ARRAY_SIZE(opcodes));
register_backingstore_template(&aio_bst);
}
My understanding is that this results in the server informing the
client that the SCSI WRITE_SAME* opcodes are not supported, and the
client then resorts to using multiple WRITE_* opcodes instead. What
happens currently is the server fails to execute the operation in an
ungraceful manner, hence the errors.
I have a few questions/comments:
1. Is there a reason why bs_create_opcode_map() is not already used for AIO?
2. Is my understanding of the situation somewhat correct?
3. I just copy-pasted opcodes from bs_rdwr.c's sbc_opcodes[] and
removed the WRITE_SAME* operations because I did not know what other
opcodes the AIO block store supports. This is not more permissive than
what is on master, but a proper patch would include just the opcodes
that are supported...sadly I do not know how to find this out.
4. I did not run unit tests on these changes.
This is my first submission to a mailing list, thank you for your
patience and looking forward to having this issue fixed, either via
this patch or something else entirely.
