For write/discard obj_requests that involved a copyup method call, the
opcode of the first op is CEPH_OSD_OP_CALL and the ->callback is
rbd_img_obj_copyup_callback().  The latter frees copyup pages, sets
->xferred and delegates to rbd_img_obj_callback(), the "normal" image
object callback, for reporting to block layer and putting refs.

rbd_osd_req_callback() however treats CEPH_OSD_OP_CALL as a trivial op,
which means obj_request is marked done in rbd_osd_trivial_callback(),
*before* ->callback is invoked and rbd_img_obj_copyup_callback() has
a chance to run.  Marking obj_request done essentially means giving
rbd_img_obj_callback() a license to end it at any moment, so if another
obj_request from the same img_request is being completed concurrently,
rbd_img_obj_end_request() may very well be called on such prematurally
marked done request:

<obj_request-1/2 reply>
handle_reply()
  rbd_osd_req_callback()
    rbd_osd_trivial_callback()
    rbd_obj_request_complete()
    rbd_img_obj_copyup_callback()
    rbd_img_obj_callback()
                                    <obj_request-2/2 reply>
                                    handle_reply()
                                      rbd_osd_req_callback()
                                        rbd_osd_trivial_callback()
      for_each_obj_request(obj_request->img_request) {
        rbd_img_obj_end_request(obj_request-1/2)
        rbd_img_obj_end_request(obj_request-2/2) <--
      }

Calling rbd_img_obj_end_request() on such a request leads to trouble,
in particular because its ->xfferred is 0.  We report 0 to the block
layer with blk_update_request(), get back 1 for "this request has more
data in flight" and then trip on

    rbd_assert(more ^ (which == img_request->obj_request_count));

with rhs (which == ...) being 1 because rbd_img_obj_end_request() has
been called for both requests and lhs (more) being 1 because we haven't
got a chance to set ->xfferred in rbd_img_obj_copyup_callback() yet.

To fix this, leverage that rbd wants to call class methods in only two
cases: one is a generic method call wrapper (obj_request is standalone)
and the other is a copyup (obj_request is part of an img_request).  So
make a dedicated handler for CEPH_OSD_OP_CALL and directly invoke
rbd_img_obj_copyup_callback() from it if obj_request is part of an
img_request, similar to how CEPH_OSD_OP_READ handler invokes
rbd_img_obj_request_read_callback().

Since rbd_img_obj_copyup_callback() is now being called from the OSD
request callback (only), it is renamed to rbd_osd_copyup_callback().

Cc: Alex Elder <elder@linaro.org>
Cc: stable@vger.kernel.org # 3.10+, needs backporting for < 3.18
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>

rbd_obj_request_create() is called on the main I/O path, so we need to
use GFP_NOIO to make sure allocation doesn't blow back on us.  Not all
callers need this, but I'm still hardcoding the flag inside rather than
making it a parameter because a) this is going to stable, and b) those
callers shouldn't really use rbd_obj_request_create() and will be fixed
in the future.

More memory allocation fixes will follow.

Cc: stable@vger.kernel.org # 3.10+
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>

nr_requests (/sys/block/rbd<id>/queue/nr_requests) is pretty much
irrelevant in blk-mq case because each driver sets its own max depth
that it can handle and that's the number of tags that gets preallocated
on setup.  Users can't increase queue depth beyond that value via
writing to nr_requests.

For rbd we are happy with the default BLKDEV_MAX_RQ (128) for most
cases but we want to give users the opportunity to increase it.
Introduce a new per-device queue_depth option to do just that:

    $ sudo rbd map -o queue_depth=1024 ...

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>

Also nuke useless Opt_last_bool and don't break lines unnecessarily.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>

The default queue_limits::max_segments value (BLK_MAX_SEGMENTS = 128)
unnecessarily limits bio sizes to 512k (assuming 4k pages).  rbd, being
a virtual block device, doesn't have any restrictions on the number of
physical segments, so bump max_segments to max_hw_sectors, in theory
allowing a sector per segment (although the only case this matters that
I can think of is some readv/writev style thing).  In practice this is
going to give us 1M bios - the number of segments in a bio is limited
in bio_get_nr_vecs() by BIO_MAX_PAGES = 256.

Note that this doesn't result in any improvement on a typical direct
sequential test.  This is because on a box with a not too badly
fragmented memory the default BLK_MAX_SEGMENTS is enough to see nice
rbd object size sized requests.  The only difference is the size of
bios being merged - 512k vs 1M for something like

    $ dd if=/dev/zero of=/dev/rbd0 oflag=direct bs=$RBD_OBJ_SIZE
    $ dd if=/dev/rbd0 iflag=direct of=/dev/null bs=$RBD_OBJ_SIZE

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>

As part of unmap sequence, kernel client has to talk to the OSDs to
teardown watch on the header object.  If none of the OSDs are available
it would hang forever, until interrupted by a signal - when that
happens we follow through with the rest of unmap procedure (i.e.
unregister the device and put all the data structures) and the unmap is
still considired successful (rbd cli tool exits with 0).  The watch on
the userspace side should eventually timeout so that's fine.

This isn't very nice, because various userspace tools (pacemaker rbd
resource agent, for example) then have to worry about setting up their
own timeouts.  Timeout it with mount_timeout (60 seconds by default).

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Reviewed-by: Sage Weil <sage@redhat.com>

There are currently three libceph-level timeouts that the user can
specify on mount: mount_timeout, osd_idle_ttl and osdkeepalive.  All of
these are in seconds and no checking is done on user input: negative
values are accepted, we multiply them all by HZ which may or may not
overflow, arbitrarily large jiffies then get added together, etc.

There is also a bug in the way mount_timeout=0 is handled.  It's
supposed to mean "infinite timeout", but that's not how wait.h APIs
treat it and so __ceph_open_session() for example will busy loop
without much chance of being interrupted if none of ceph-mons are
there.

Fix all this by verifying user input, storing timeouts capped by
msecs_to_jiffies() in jiffies and using the new ceph_timeout_jiffies()
helper for all user-specified waits to handle infinite timeouts
correctly.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>

Signed-off-by: Yan, Zheng <zyan@redhat.com>
Reviewed-by: Alex Elder <elder@linaro.org>

When we end I/O struct request with error, we need to pass
obj_request->length as @nr_bytes so that the entire obj_request worth
of bytes is completed.  Otherwise block layer ends up confused and we
trip on

    rbd_assert(more ^ (which == img_request->obj_request_count));

in rbd_img_obj_callback() due to more being true no matter what.  We
already do it in most cases but we are missing some, in particular
those where we don't even get a chance to submit any obj_requests, due
to an early -ENOMEM for example.

A number of obj_request->xferred assignments seem to be redundant but
I haven't touched any of obj_request->xferred stuff to keep this small
and isolated.

Cc: Alex Elder <elder@linaro.org>
Cc: stable@vger.kernel.org # 3.10+
Reported-by: Shawn Edwards <lesser.evil@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>