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>
This patch fixes the errors made in the users of the crypto layer during
the sg_init_table conversion. It also adds a few conversions that were
missing altogether.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Most drivers need to set length and offset as well, so may as well fold
those three lines into one.
Add sg_assign_page() for those two locations that only needed to set
the page, where the offset/length is set outside of the function context.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
It was found by LSI that on setups with large amounts of memory
we were bouncing buffers when we did not need to. If the iscsi tcp
code touches the data buffer (or a helper does),
it will kmap the buffer. iscsi_tcp also does not interact with hardware,
so it does not have any hw dma restrictions. This patch sets the bounce
buffer settings for our device queue so buffers should not be bounced
because of a driver limit.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
This prevents the iscsi modules from being unloaded while
there are active mounts from an iscsi target.
Signed-off-by: Olaf Kirch <olaf.kirch@oracle.com>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
- remove the unnecessary map_single path.
- convert to use the new accessors for the sg lists and the
parameters.
TODO: use scsi_for_each_sg().
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Acked-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
iSCSI must support software iscsi (iscsi_tcp, iser), hardware iscsi (qla4xxx),
and partial offload (broadcom). To be able to allow each stack or driver
or port (virtual or physical) to be able to log into the same target portal
we use the initiator tuple [[HWADDRESS | NETDEVNAME], INITIATOR_NAME] and
the target tuple [TARGETNAME, CONN_ADDRESS, CONN_PORT] to id a session.
This patch adds the netdev name, which is used by software iscsi when
it binds a session to a netdevice using the SO_BINDTODEVICE sock opt.
It cannot use HWADDRESS because if someone did vlans then the same netdevice
will have the same mac and the initiator,target id will not be unique.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: David C Somayajulu <david.somayajulu@qlogic.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
This patch exports the local address for the session. For
qla4xxx this is the ip of the hba's port. For software
this is the src addr of the socket.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: David C Somayajulu <david.somayajulu@qlogic.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
This patch should fix the file descriptor leak problem. A quick look
through the kernel shows that users of sockfd_lookup use sockfd_put to
release their handle. We were using sock_release which from the comments
and code look like it does not release the get() on the file from the
lookup.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
Add a slave_configure function to iSCSI TCP to remove any DMA
alignment restriction. This permits the use of direct IO from
arbitrary addresses.
Signed-off-by: Pete Wyckoff <pw@osc.edu>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.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>
This patch allows us to set can_queue and cmds_per_lun from userspace
when we create the session/host. From there we can set it on a per
target basis. The patch fully converts iscsi_tcp, but only hooks
up ib_iser for cmd_per_lun since it currently has a lots of preallocations
based on can_queue.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Roland Dreier <rdreier@cisco.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
The cmdsn allocation and pdu transmit code can race, and we can end
up sending a pdu with cmdsn 10 before a pdu with 5. The target will
then fail the connection/session. This patch fixes the problem by
delaying the cmdsn allocation until we are about to send the pdu.
This also removes the xmitmutex. We were using the connection xmitmutex
during error handling to handle races with mtask and ctask cleanup and
completion. For ctasks we now have nice refcounting and for the mtask,
if we hit the case where the mtask timesout and it is floating
around somewhere in the driver, we end up dropping the session.
And to handle session level cleanup, we use the xmit suspend bit
along with scsi_flush_queue and the session lock to make sure
that the xmit thread is not possibly transmitting a task while
we are trying to kill it.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Roland Dreier <rdreier@cisco.com>
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>
The attached patches add sysfs files for the chap settings
to the iscsi transport class, iscsi_tcp and ib_iser. This is
needed for software iscsi because there are times when iscsid
can die and it will need to reread the values it was using.
And it is needed by qla4xxx for basic management opertaions.
This patch does not hook in qla4xxx yet, because I am not sure
the mbx command to use.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Roland Dreier <rdreier@cisco.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
- Remove shadow of request length from struct iscsi_cmd_task.
- change all users to use scsi_cmnd->request_bufflen directly
(With bidi we will use scsi-ml API to retrieve in/out length)
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>
Cc: Roland Dreier <rdreier@cisco.com>
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>
For iscsi root boot, software iscsi needs to know what the BIOS/OF
initiator used for the initiator name so this puts it in sysfs
for userspace to be able to pick up.
For hw iscsi, it is nice to see what the card is using.
This patch adds the new param, and hooks in qla4xxx, iscsi_tcp, and ib_iser.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: David C Somayajulu <david.somayajulu@qlogic.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
iscsid and udev need to key off the hw address being
used so add some helpers for iser and iscsi tcp.
Also convert them
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Roland Dreier <rdreier@cisco.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
People do not read the README and seem to like to
unselect the crc32c module even though iscsi_tcp selects
it for them. This patch spits a error that tells the user
that they really do need the module. Hopefully, we will
get fewer people asking about this now.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
For a while now, the block layer has seperated max sectors
and max hw sectors. Software iscsi has no limit so this patch
increases max hw sectors, so we can support large pass through
commands.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
This patch renames DEFAULT_MAX_RECV_DATA_SEGMENT_LENGTH to avoid
confusion with the drivers default values (DEFAULT_MAX_RECV_DATA_SEGMENT_LENGTH
is the iscsi RFC specific default).
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
The return value of crypto_alloc_hash() should be checked by
IS_ERR().
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
The transition from crypto_digest_*() to the crypto_hash_*() family
introduced a bug into the data digest calculation: crypto_hash_update() is
called with the number of S/G elements instead of the S/G lists data size.
Signed-off-by: Arne Redlich <arne.redlich@xiranet.com>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
XMSTATE_SOL_HDR could be set when the xmit thread tests it, but there may
not be anything on the r2tqueue yet. Move the XMSTATE_SOL_HDR set
before the addition to the queue to make sure that when we pull something
off it it is valid. This does not add locks around the xmstate test or make
that a atmoic_t because this is a fast path and if it is set when we test it
we can handle it there without the overhead. Later on we check the xmitqueue
for all requests with the session lock so we will not miss it.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
Unconditionally free crypto state, as it is always allocated during
TCP connection creation. Without this, crypto structures leak and
crc32c module refcounts grow as connections are created and
destroyed.
Signed-off-by: Pete Wyckoff <pw@osc.edu>
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>
When a digest is spread across two network buffers, we currently
ignore this and try to check the digest with the partial buffer.
Or course this fails. This patch has use iscsi_tcp_copy to
copy the whole digest before testing it.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
When we relogin to a target, we have not yet negotiated digests
so we must reset the hdr_size var.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
This patch built over the last ones fixes a bug in the partial header
resend code, where we add on another 4 bytes to the send length on the resend.
We want just the header plus digest.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
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>
A couple targets like string bean and MDS, send r2ts with
a data len greater than the max burst we agreed to. We
were being strict in our enforcing of the iscsi rfc in that
code path, but there is no driver limitation that prevents
us from fullfilling the request. To allow those targets
to work we will ignore the max_burst length and send as
much data as the target asks for assuming it has consciously
decided to override its max burst length.
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>
The version info is useful for iscsi tcp, iser and qla4xxx so move to
transport class.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
Must pass ISCSI_ERR values from the recv path and propogate them
upwards.
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>
When we enter recovery and flush the running commands
we cannot freee the connection before flushing the commands.
Some commands may have a reference to the connection
that needs to be released before. iscsi_stop was forcing
the term and suspend too early and was causing a oops
in iser, so this patch removes those callbacks all together
and allows the LLD to handle that detail.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
if iscsi_data_rsp fails we must bail out. Since the pdu values like
data length are invalid we cannot continue to process the data since
it could over run buffers.
This fixes a bug with cisco 5428s where that target is sending
too much data.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
The iscsi tcp code can pluck multiple rt2s from the tasks's r2tqueue
in the xmit code. This can result in the task being queued on the xmit queue
but gettting completed at the same time.
This patch fixes the above bug by making the fifo a list so
we always remove the entry on the list del.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
In the xmit patch we are sending a -EXXX value to iscsi_conn_failure
which is causing userspace to get confused.
We should be sending a ISCSI_ERR_* value that userspace understands.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
Convert iscsi_tcp to new lib functions.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
We can race and misset the suspend bit if iscsi_write_space is
called then iscsi_send returns with a failure indicating
there is no space.
To handle this this patch returns a error upwards allowing xmitworker
to decide if we need to try and transmit again. For the no
write space case xmitworker will not retry, and instead
let iscsi_write_space queue it back up if needed (this relies
on the work queue code to properly requeue us if needed).
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
If recovery failed or we are in recovery only overwrite the state
if we are going to terminate the session or if we logged back in.
STOP_CONN_SUSPEND and conn_cnt are not used. We only support
a single connection session ATM, so cleanup that code while
we are working around it.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
Tony Battersby