There is a race condition in iscsi_tcp.c that may cause it to forget
that it received a R2T from the target. This race may cause a data-out
command (such as a write) to lock up. The race occurs here:
static int
iscsi_send_unsol_pdu(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
{
struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
int rc;
if (tcp_ctask->xmstate & XMSTATE_UNS_HDR) {
BUG_ON(!ctask->unsol_count);
tcp_ctask->xmstate &= ~XMSTATE_UNS_HDR; <---- RACE
...
static int
iscsi_r2t_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
{
...
tcp_ctask->xmstate |= XMSTATE_SOL_HDR_INIT; <---- RACE
...
While iscsi_xmitworker() (called from scsi_queue_work()) is preparing to
send unsolicited data, iscsi_tcp_data_recv() (called from
tcp_read_sock()) interrupts it upon receipt of a R2T from the target.
Both contexts do read-modify-write of tcp_ctask->xmstate. Usually, gcc
on x86 will make &= and |= atomic on UP (not guaranteed of course), but
in this case iscsi_send_unsol_pdu() reads the value of xmstate before
clearing the bit, which causes gcc to read xmstate into a CPU register,
test it, clear the bit, and then store it back to memory. If the recv
interrupt happens during this sequence, then the XMSTATE_SOL_HDR_INIT
bit set by the recv interrupt will be lost, and the R2T will be
forgotten.
The patch below (against 2.6.24-rc1) converts accesses of xmstate to use
set_bit, clear_bit, and test_bit instead of |= and &=. I have tested
this patch and verified that it fixes the problem. Another possible
approach would be to hold a lock during most of the rx/tx setup and
post-processing, and drop the lock only for the actual rx/tx.
Signed-off-by: Tony Battersby <tonyb@cybernetics.com>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
If we got the padding, data and header in different skbs,
we were not handling the padding correctly because we attributed it
to the data's skb. This resulted in the initiator reading from
pad bytes + skb offset instead of the correct offset.
If you could not connect with the open solaris target, this
will fix the lock up problem you were hitting.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
If iscsi_tcp partially sends a header, it would recalculate the
header size and readd the size of the digest (if header digests
are used).This would cause us to send sizeof(digest) extra bytes
when we sent the rest of the header.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
This patch fixes handling of expected datasn/r2tsn as received from
target. It is done according to: T10 rfc3720 section 3.2.2.3. Data Sequencing.
. unify expected datasn/r2tsn into one counter
. calculate than check expected datasn/r2tsn. On error print a message
and fail the request. (TODO use iscsi retransmits)
. remove the FIXME ;)
. avoid zero length memset
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>
Signed-off-by: Benny Halevy <bhalevy@panasas.com>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
This patch converts ISCSI to use the new crypto_hash interface instead
of crypto_digest. It's a fairly straightforward substitution.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
We currently allocated seperate tfms for data and header digests. There
is no reason for this since we can never calculate a rx header and
digest at the same time. Same for sends. So this patch removes the data
tfms and has the send and recv sides use the rx_tfm or tx_tfm.
I also made the connection creation code preallocate the tfms because I
thought I hit a bug where I changed the digests settings during a
relogin but could not allocate the tfm and then we just failed.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
iscsi_tcp calculates padding by using the expected transfer length. This
has the problem where if we have immediate data = no and initial R2T =
yes, and the transfer length ended up needing padding then we send:
1. header
2. padding which should have gone after data
3. data
Besides this bug, we also assume the target will always ask for nice
transfer lengths and the first burst length will always be a nice value.
As far as I can tell form the RFC this is not a requirement. It would be
silly to do this, but if someone did it we will end doing bad things.
Finally the last bug in that bit of code is in our handling of the
recalculation of data digests when we do not send a whole iscsi_buf in
one try. The bug here is that we call crypto_digest_final on a
iscsi_sendpage error, then when we send the rest of the iscsi_buf, we
doiscsi_data_digest_init and this causes the previous data digest to be
lost.
And to make matters worse, some of these bugs are replicated over and
over and over again for immediate data, solicited data and unsolicited
data. So the attached patch made over the iscsi git tree (see
kernel.org/git for details) which I updated today to include the patches
I said I merged, consolidates the sending of data, padding and digests
and calculation of data digests and fixes the above bugs.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
iSCSI RFC states that the first burst length must be smaller than the
max burst length. We currently assume targets will be good, but that may
not be the case, so this patch adds a check.
This patch also moves the unsol data out offset to the lib so the LLDs
do not have to track it.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
We currently try to allocate a max_recv_data_segment_length
which can be very large (default is 64K), and common uses
are up to 1MB. It is very very difficult to allocte this
much contiguous memory and it turns out we never even use it.
We really only need a couple of pages, so this patch has us
allocates just what we know what we need today.
Later if vendors start adding vendor specific data and
we need to handle large buffers we can do this, but for
the last 4 years we have not seen anyone do this or request
it.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
Discovered by steven@hayter.me.uk and patch by michaelc@cs.wisc.edu
The dtask mempool is reserving 261120 items per session! Since we are now
sending headers with sendmsg there is no reason for the mempool and that
was causing us to us carzy amounts of mem. We can preallicate a header in
the r2t and task struct and reuse them
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
This just converts iscsi_tcp to the lib
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
The current iscsi_tcp eh is not nicely setup for dm-multipath
and performs some extra task management functions when they
are not needed.
The attached patch:
- Fixes the TMF issues. If a session is rebuilt
then we do not send aborts.
- Fixes the problem where if the host reset fired, we would
return SUCCESS even though we had not really done anything
yet. This ends up causing problem with scsi_error.c's TUR.
- If someone has turned on the userspace nop daemon code to try
and detect network problems before the scsi command timeout
we can now drop and clean up the session before the scsi command
timesout and fires the eh speeding up the time it takes for a
command to go from one patch to another. For network problems
we fail the command with DID_BUS_BUSY so if failfast is set
scsi_decide_disposition fails the command up to dm for it to
try on another path.
- And we had to add some basic iscsi session block code. Previously
if we were trying to repair a session we would retrun a MLQUEUE code
in the queuecommand. This worked but it was not the most efficient
or pretty thing to do since it would take a while to relogin
to the target. For iscsi_tcp/open-iscsi a lot of the iscsi error handler
is in userspace the block code is pretty bare. We will be
adding to that for qla4xxx.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
>From erezz@voltaire.com:
rm conn->lock since it is not used anymore. The dataqueue is protected
by the session lock and xmitmutex.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: Alex Aizman <itn780@yahoo.com>
Signed-off-by: Dmitry Yusupov <dmitry_yus@yahoo.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
From:
michaelc@cs.wisc.edufujita.tomonori@lab.ntt.co.jpda-x@monatomic.org
and err path fixup from:
ogerlitz@voltaire.com
This patch cleans up that interface by having the lld and class
pass a iscsi_cls_session or iscsi_cls_conn between each other when
the function is used by HW and SW iscsi llds. This way the lld
does not have to remember if it has to send a handle or pointer
and a handle or pointer to connection, session or host.
This also has the class verify the session handle that gets passed from
userspace instead of using the pointer passed into the kernel directly.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: Alex Aizman <itn780@yahoo.com>
Signed-off-by: Dmitry Yusupov <dmitry_yus@yahoo.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
From Mike Christie <michaelc@cs.wisc.edu> and FUJITA Tomonori <tomof@acm.org>:
We cannot use page_address becuase some pages could be highmem.
Instead, we can use sock_no_sendpage which does kmap for us.
Signed-off-by: Alex Aizman <itn780@yahoo.com>
Signed-off-by: Dmitry Yusupov <dmitry_yus@yahoo.com>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
Users can write to a page while we are sending it and making
digest calculations. This ends up causing us to retry the command
when a digest error is later reported. By using sock_no_sendpage
when data digests are calculated we can avoid a lot of (not all but it
helps) the retries becuase sock_no_sendpage is not zero copy.
Signed-off-by: Alex Aizman <itn780@yahoo.com>
Signed-off-by: Dmitry Yusupov <dmitry_yus@yahoo.com>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
the scsi layer is using semaphores in a mutex way, this patch converts
these into using mutexes instead
Signed-off-by: Arjan van de Ven <arjan@infradead.org>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
From Wang Zhenyu:
High queue depth was a problem for some targets so make queue_depth adjustable
From Mike Christie
Make default queue_depth a little lower
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: Alex Aizman <itn780@yahoo.com>
Signed-off-by: Dmitry Yusupov <dmitry_yus@yahoo.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
