Libiscsi supports to parse two iscsi url schemes: 'iscsi://' and 'iser://'.
Fix the missing iser parsing, introduced from 12222077.
Signed-off-by: Han Han <hhan@redhat.com>
Remove a bunch of duplicate code by sharing a function for source and
destination endpoint initialization.
Signed-off-by: David Disseldorp <ddiss@suse.de>
I don't see any problems that calling the callback
during connect/login in iscsi_cancel_pdus(). So let's
remove this check. Otherwise, we have no way to be aware
of a cancellation during login and cause something like
iscsi_login_sync() hangs.
Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
We should plug the cmdsn gap in order to continue
to use the session when the pdus is cancelled before
sending out.
Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
Since the NUMBER OF LOGICAL BLOCKS field in the COMPARE AND WRITE command
is an 8 bit field, the maximum value that can be encoded is 255.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Assign the NUMBER OF LOGICAL BLOCKS field in the COMPARE AND WRITE PDU
directly. Use the terminology from SBC-4, namely NUMBER OF LOGICAL BLOCKS
instead of TL.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Cmdsn of a data-out pdu struct is less than `expcmdsn` since it's from its
cmd pdu. A data-out pdu doesn't carray a cmdsn on the wire actually, so it
doesn't matter to itself, but if we rewrite the cmdsn of a immediate pdu with
it, it will cause an error.
Related error logs:
libiscsi: iscsi_write_to_socket: outqueue[0]->cmdsn < expcmdsn (3648bab5 < 3648bab9) opcode 00 [iqn.2003-01.org.linux-iscsi.tgt0]
libiscsi: reconnect initiated [iqn.2003-01.org.linux-iscsi.tgt0]
libiscsi: connecting to portal 127.0.0.1 [iqn.2003-01.org.linux-iscsi.tgt0]
libiscsi: connection established (127.0.0.1:62404 -> 127.0.0.1) [iqn.2003-01.org.linux-iscsi.tgt0]
Signed-off-by: wanghonghao <wanghonghao@bytedance.com>
Allocate `iser_pdu` from small allocation pool.
Lifecycle of `iscsi_in_pdu` is inside the function in iSER transport. Allocate
it on stack.
Signed-off-by: wanghonghao <wanghonghao@bytedance.com>
This commit is to fix compatibility with CHAP.
iSER transport only post `login_resp_buf` (which is larger than `rx_desc`) as
work request (WR) once, but there may be multiple requests and responses during
login phase (e.g. when CHAP is used) and login can't be finished in such cases.
Signed-off-by: wanghonghao <wanghonghao@bytedance.com>
tx_desc and memory region buffer assigned to iser pdus should be given back to
tx_desc list and allocator before free all memory regions.
This may happend during reconnecting/disconnecting.
Signed-off-by: wanghonghao <wanghonghao@bytedance.com>
This patch is used to fix the following problems in the current connection
method:
1. iscsi_iser_connect() waits until the connection is established or failed,
and may block the caller for a long time.
2. Although there's a cm_thread handles communication events, but in fact it
has no effects after the connection is established.
3. Resources are not released properly after reconnection failed. And once we
try to reconnect again, the resources will leak permanently.
(see iscsi_reconnect()).
This patch eliminate cm_thread and handle communication events in the caller
thread.
Connection procedure:
1. Create a mock fd by eventfd() (or just use old_iscsi->fd while reconnecting),
and assign it to iscsi->fd.
2. Create communication event channel, make it non-blocking and dup the
notifier fd to iscsi->fd.
3. Handle communication events by iscsi_which_events()/iscsi_service() loop
until connection established or falied.
4. If connection is established successfully, dup the notifier fd of completion
queue (CQ) events to iscsi->fd.
5. Handle completion queue (CQ) events by iscsi_which_events()/iscsi_service()
loop.
The entire procedure is non-blocking.
After established, whenever iscsi_service() is called with revents=0 or
queue_pdu() is called with a NOP pdu, communication events will be checked.
When connection failed, iser transport cleanup itself before callbacks.
Signed-off-by: wanghonghao <wanghonghao@bytedance.com>
Implement an allocator for allocating memory region of different lengths.
The allocator registers 4MB memory chunks as memory regions, and select a
free segment from one of them each time.
4KB is the minimum allocation unit, and free segments in the same chunk can be
merged into a larger free segment by the rule of buddy allocation. As a result,
size of allocated segments will be power of 2, this may waste some space but
produces less fragments.
In each chunk, a complete binary tree (which is actully an array) is used
to maintain free segments. Each node records the order of the largest segment
can be allocated from its subtree. Here's a miniature example.
A chunk with all segments free:
level 4 4(0x1)
level 3 3(0x2) 3(0x3)
level 2 2(0x4) 2(0x5) 2(0x6) 2(0x7)
level 1 1(0x8) 1(0x9) 1(0xa) 1(0xb) 1(0xc) 1(0xd) 1(0xe) 1(0xf)
After allocate a 16KB(order=3) memory region:
level 4 3(0x1)
level 3 0(0x2) 3(0x3)
level 2 2(0x4) 2(0x5) 2(0x6) 2(0x7)
level 1 1(0x8) 1(0x9) 1(0xa) 1(0xb) 1(0xc) 1(0xd) 1(0xe) 1(0xf)
It tooks 1 comparison to determine if a chunk can satisfy and at most 11
loops to find the leftmost free segment meets the requirments.
The value of each node is not more than 11, and a 8-bit integer is enough
to store it, so only 2048 bytes is required for each tree. And since the
entire tree is in a contiguous piece of memory and no rotations are needed,
it's far more efficient than self-balancing trees of the same size.
Different 4MB chunks are linked as a list, and the selection order is from
head to tail each time. If no existing chunks can satisfy the allocation,
the allocator will register another 4M chunk and add it to the tail.
+---------+ +---------+ +---------+
|4MB chunk| ---> |4MB chunk| ---> |4MB chunk|
+---------+ +---------+ +---------+
In most cases, smaller IOs can always get memory regions from the first or
second chunk and never traverse the list too much, and if we really send a lot
of large IOs, the cost of the traversal is rarely critical.
At last, obviously, the chunks can only allocate a maximum of 4MB memory region,
if a larger memory region is needed, the allocater registers/deregisters a
memory region directly regardless of buffer.
Signed-off-by: wanghonghao <wanghonghao@bytedance.com>
Check for 0x0A descriptor support via SCSI_COPY_RESULTS_OP_PARAMS. If
supported, perform a 0x0A type XCOPY using target's
maximum_segment_length (or 256M if larger).
Signed-off-by: David Disseldorp <ddiss@suse.de>
Type 0x0A segment descriptors provide a 32-bit wide number-of-bytes
field, allowing for much larger copy offloads with a single segment
descriptor.
Signed-off-by: David Disseldorp <ddiss@suse.de>
Hit iser hang in rdma_destroy_id with trace:
#0 pthread_cond_wait@@GLIBC_2.3.2 () at ../sysdeps/unix/sysv/linux/x86_64/pthread_cond_wait.S:185
#1 0x00007f96ecbbcbb3 in rdma_destroy_id () from /usr/lib/librdmacm.so.1
#2 0x00005632027311d4 in iser_conn_release (iser_conn=iser_conn@entry=0x7f96d4027440) at iser.c:261
#3 0x0000563202731428 in iscsi_iser_connect (iscsi=0x563205206c70, sa=<optimized out>, ai_family=<optimized out>)
at iser.c:1516
#4 0x000056320273dd3c in iscsi_connect_async (iscsi=iscsi@entry=0x563205206c70,
portal=portal@entry=0x563205207084 "210.32.124.205:3260", cb=cb@entry=0x56320272b220 <iscsi_connect_cb>,
private_data=private_data@entry=0x7f96d4008b00) at socket.c:389
#5 0x000056320272b325 in iscsi_full_connect_async (iscsi=0x563205206c70,
portal=0x563205207084 "210.32.124.205:3260", lun=1, cb=cb@entry=0x56320272aef0 <iscsi_reconnect_cb>,
private_data=private_data@entry=0x0) at connect.c:230
#6 0x000056320272b711 in iscsi_reconnect (iscsi=<optimized out>) at connect.c:473
#7 0x00005632026810a8 in iscsi_timed_check_events (opaque=0x563205206ae0) at block/iscsi.c:387
Currently use pthread_cancel to kill cmthread forcefully, cmthread may
exits without rdma_ack_cm_event, then unacknowledged event will be
remained in librdmacm. rdma_destroy_id hangs until uplayer ack all
the cm event.
Since destroying qp, cm thread will handle DISCONNECTED event, and
exits by itself. Joining cm thread to wait cm thread to exit
gracefully.
Signed-off-by: zhenwei pi <pizhenwei@bytedance.com>
Hit the crash stack:
#0 iser_initialize_headers (iser_pdu=0x7f1a3404ef50, iser_conn=0x0) at iser.c:514
#1 iscsi_iser_send_pdu (iscsi=0x7f1a3406d700, pdu=0x7f1a3404ef50) at iser.c:714
#2 0x000055e3160f0157 in iscsi_scsi_command_async (iscsi=0x7f1a3406d700, iscsi@entry=0x55e317fbcc70,
lun=lun@entry=1, task=task@entry=0x7f1a34026610, cb=cb@entry=0x55e316044c10 <iscsi_co_generic_cb>,
d=d@entry=0x7f15feeb7710, private_data=private_data@entry=0x7f15feeb77e0) at iscsi-command.c:282
#3 0x000055e3160f1616 in iscsi_write10_iov_task (iscsi=0x55e317fbcc70, lun=1, lba=lba@entry=10401896,
data=data@entry=0x0, datalen=4096, blocksize=<optimized out>, wrprotect=0, dpo=0, fua=0, fua_nv=0,
group_number=0, cb=0x55e316044c10 <iscsi_co_generic_cb>, private_data=0x7f15feeb77e0, iov=0x7f1a34042090,
niov=1) at iscsi-command.c:1107
#4 0x000055e31604680f in iscsi_co_writev (bs=<optimized out>, sector_num=<optimized out>,
nb_sectors=<optimized out>, iov=0x7f1a3404e380, flags=<optimized out>) at block/iscsi.c:640
#5 0x000055e31601e89c in bdrv_driver_pwritev (bs=bs@entry=0x55e317fb6570, offset=offset@entry=5325770752,
bytes=bytes@entry=4096, qiov=qiov@entry=0x7f1a3404e380, qiov_offset=qiov_offset@entry=0, flags=flags@entry=0)
at block/io.c:1220
The reason is that during async reconnection, before reconnecting
call back function gets woked, we have closed the old connection,
and the new connection is not ready.
At the same time, up layer still sends pdu to the old iscsi context.
In this patch, before reconnecting successfully, just add the pdu to
waitpdu without sending.
Suggested by Bart, do not show iser related log here.
Signed-off-by: zhenwei pi <pizhenwei@bytedance.com>
[ bvanassche: reformatted patch ]
The SPC4r37 specification defines XCOPY Segment Descriptor lengths for
each SD type. The spec defined SD lengths don't account for the four SD
header bytes. To make it easier to match the spec to the implementation,
use the spec defined lengths in get_desc_len() with a four byte SD
header addition.
Signed-off-by: David Disseldorp <ddiss@suse.de>
Checking whether a pointer is NULL after it has been dereferenced is not
useful. This was detected by Coverity.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Reduce the size of struct iscsi_context by reordering the members of this
data structure. Additionally, change the rdma_ack_timeout value from
'unsigned char' into 'uint8_t' to make it clear that this variable
represents an integer.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Since 2c1619edef61a03cb516efaa81750784c3071d10 for linux kernel and
55843c4ab8f559679d28c559cc4d681836be769b for rdma-core, rdma cma
supports RDMA_OPTION_ID_ACK_TIMEOUT. It's useful for RDMA out of
sequence case. Because this feature is added recently, we have to
check this in autogen.sh before building source code.
Depend on production enviroument, tunning rdma ack timeout could get
the best performance. Suggested by Bart and Ronnie, instead of using
a fixed timeout value, add two methods to set rdma ack timeout value.
1, add URL variable 'LIBISCSI_RDMA_ACK_TIMEOUT'. This could works
for a specified connection.
2, add env argument 'LIBISCSI_RDMA_ACK_TIMEOUT'. This works as a
common setting for all the connection of a process.
Test under different packet loss rate and different ack timeout, run
fio (iodepth=1) in a guest os, I got this result:
latency under packet loss rate 0.00001:
timeout 19: avg 170.22, pct99.9 215
timeout 10: avg 160.08, pct99.9 215
timeout 8 : avg 146.39, pct99.9 177
timeout 7 : avg 148.37, pct99.9 211
latency under packet loss rate 0.0001:
timeout 19: avg 949.23, pct99.9 306
timeout 10: avg 818.53, pct99.9 378
timeout 8 : avg 615.84, pct99.9 189
timeout 7 : avg 618.89, pct99.9 310
Base on this test report, setting ack timeout to 8(1048.576 usec) is
a good choice in my test enviroument.
Signed-off-by: zhenwei pi <pizhenwei@bytedance.com>
