pandora-kernel.git
12 years agocrush: fix memory leak when destroying tree buckets
Sage Weil [Mon, 7 May 2012 22:37:05 +0000 (15:37 -0700)]
crush: fix memory leak when destroying tree buckets

Reflects ceph.git commit 46d63d98434b3bc9dad2fc9ab23cbaedc3bcb0e4.

Reported-by: Alexander Lyakas <alex.bolshoy@gmail.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agocrush: fix tree node weight lookup
Sage Weil [Mon, 7 May 2012 22:36:49 +0000 (15:36 -0700)]
crush: fix tree node weight lookup

Fix the node weight lookup for tree buckets by using a correct accessor.

Reflects ceph.git commit d287ade5bcbdca82a3aef145b92924cf1e856733.

Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agocrush: remove parent maps
Sage Weil [Mon, 7 May 2012 22:36:35 +0000 (15:36 -0700)]
crush: remove parent maps

These were used for the ill-fated forcefeed feature.  Remove them.

Reflects ceph.git commit ebdf80edfecfbd5a842b71fbe5732857994380c1.

Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agocrush: remove forcefeed functionality
Sage Weil [Mon, 7 May 2012 22:39:29 +0000 (15:39 -0700)]
crush: remove forcefeed functionality

Remove forcefeed functionality from CRUSH.  This is an ugly misfeature that
is mostly useless and unused.  Remove it.

Reflects ceph.git commit ed974b5000f2851207d860a651809af4a1867942.

Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
Conflicts:

net/ceph/crush/mapper.c

12 years agocrush: use a temporary variable to simplify crush_do_rule
Sage Weil [Mon, 7 May 2012 22:35:48 +0000 (15:35 -0700)]
crush: use a temporary variable to simplify crush_do_rule

Use a temporary variable here to avoid repeated array lookups and clean up
the code a bit.

This reflects ceph.git commit 6b5be27634ad307b471a5bf0db85c4f5c834885f.

Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agocrush: be more tolerant of nonsensical crush maps
Sage Weil [Mon, 7 May 2012 22:35:24 +0000 (15:35 -0700)]
crush: be more tolerant of nonsensical crush maps

If we get a map that doesn't make sense, error out or ignore the badness
instead of BUGging out.  This reflects the ceph.git commits
9895f0bff7dc68e9b49b572613d242315fb11b6c and
8ded26472058d5205803f244c2f33cb6cb10de79.

Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agocrush: adjust local retry threshold
Sage Weil [Mon, 7 May 2012 22:35:09 +0000 (15:35 -0700)]
crush: adjust local retry threshold

This small adjustment reflects a change that was made in ceph.git commit
af6a9f30696c900a2a8bd7ae24e8ed15fb4964bb, about 6 months ago.  An N-1
search is not exhaustive.  Fixed ceph.git bug #1594.

Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agocrush: clean up types, const-ness
Sage Weil [Mon, 7 May 2012 22:38:35 +0000 (15:38 -0700)]
crush: clean up types, const-ness

Move various types from int -> __u32 (or similar), and add const as
appropriate.

This reflects changes that have been present in the userland implementation
for some time.

Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agoceph: refactor SETLAYOUT and SETDIRLAYOUT ioctl checks into common helper
Sage Weil [Mon, 7 May 2012 22:34:35 +0000 (15:34 -0700)]
ceph: refactor SETLAYOUT and SETDIRLAYOUT ioctl checks into common helper

Both of these methods perform similar checks; move that code to a helper
so that we can ensure the checks are consistent.

Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agoceph: drop support for preferred_osd pgs
Sage Weil [Mon, 7 May 2012 22:33:36 +0000 (15:33 -0700)]
ceph: drop support for preferred_osd pgs

This was an ill-conceived feature that has been removed from Ceph.  Do
this gracefully:

 - reject attempts to specify a preferred_osd via the ioctl
 - stop exposing this information via virtual xattrs
 - always fill in -1 for requests, in case we talk to an older server
 - don't calculate preferred_osd placements/pgids

Reviewed-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agorbd: don't hold spinlock during messenger flush
Alex Elder [Wed, 4 Apr 2012 18:35:44 +0000 (13:35 -0500)]
rbd: don't hold spinlock during messenger flush

A recent change made changes to the rbd_client_list be protected by
a spinlock.  Unfortunately in rbd_put_client(), the lock is taken
before possibly dropping the last reference to an rbd_client, and on
the last reference that eventually calls flush_workqueue() which can
sleep.

The problem was flagged by a debug spinlock warning:
    BUG: spinlock wrong CPU on CPU#3, rbd/27814

The solution is to move the spinlock acquisition and release inside
rbd_client_release(), which is the spot where it's really needed for
protecting the removal of the rbd_client from the client list.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
12 years agorbd: move snap_rwsem to the device, rename to header_rwsem
Josh Durgin [Tue, 22 Nov 2011 01:11:12 +0000 (17:11 -0800)]
rbd: move snap_rwsem to the device, rename to header_rwsem

A new temporary header is allocated each time the header changes, but
only the changed properties are copied over. We don't need a new
semaphore for each header update.

This addresses http://tracker.newdream.net/issues/2174

Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>
Reviewed-by: Alex Elder <elder@dreamhost.com>
12 years agoceph: fix three bugs, two in ceph_vxattrcb_file_layout()
Alex Elder [Thu, 8 Mar 2012 22:50:09 +0000 (16:50 -0600)]
ceph: fix three bugs, two in ceph_vxattrcb_file_layout()

In ceph_vxattrcb_file_layout(), there is a check to determine
whether a preferred PG should be formatted into the output buffer.
That check assumes that a preferred PG number of 0 indicates "no
preference," but that is wrong.  No preference is indicated by a
negative (specifically, -1) PG number.

In addition, if that condition yields true, the preferred value
is formatted into a sized buffer, but the size consumed by the
earlier snprintf() call is not accounted for, opening up the
possibilty of a buffer overrun.

Finally, in ceph_vxattrcb_dir_rctime() where the nanoseconds part of
the time displayed did not include leading 0's, which led to
erroneous (sub-second portion of) time values being shown.

This fixes these three issues:
    http://tracker.newdream.net/issues/2155
    http://tracker.newdream.net/issues/2156
    http://tracker.newdream.net/issues/2157

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
12 years agolibceph: isolate kmap() call in write_partial_msg_pages()
Alex Elder [Wed, 7 Mar 2012 17:40:08 +0000 (11:40 -0600)]
libceph: isolate kmap() call in write_partial_msg_pages()

In write_partial_msg_pages(), every case now does an identical call
to kmap(page).  Instead, just call it once inside the CRC-computing
block where it's needed.  Move the definition of kaddr inside that
block, and make it a (char *) to ensure portable pointer arithmetic.

We still don't kunmap() it until after the sendpage() call, in case
that also ends up needing to use the mapping.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
12 years agolibceph: rename "page_shift" variable to something sensible
Alex Elder [Wed, 7 Mar 2012 17:40:08 +0000 (11:40 -0600)]
libceph: rename "page_shift" variable to something sensible

In write_partial_msg_pages() there is a local variable used to
track the starting offset within a bio segment to use.  Its name,
"page_shift" defies the Linux convention of using that name for
log-base-2(page size).

Since it's only used in the bio case rename it "bio_offset".  Use it
along with the page_pos field to compute the memory offset when
computing CRC's in that function.  This makes the bio case match the
others more closely.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
12 years agolibceph: get rid of zero_page_address
Alex Elder [Wed, 7 Mar 2012 17:40:08 +0000 (11:40 -0600)]
libceph: get rid of zero_page_address

There's not a lot of benefit to zero_page_address, which basically
holds a mapping of the zero page through the life of the messenger
module.  Even with our own mapping, the sendpage interface where
it's used may need to kmap() it again.  It's almost certain to
be in low memory anyway.

So stop treating the zero page specially in write_partial_msg_pages()
and just get rid of zero_page_address entirely.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
12 years agolibceph: only call kernel_sendpage() via helper
Alex Elder [Wed, 7 Mar 2012 17:40:08 +0000 (11:40 -0600)]
libceph: only call kernel_sendpage() via helper

Make ceph_tcp_sendpage() be the only place kernel_sendpage() is
used, by using this helper in write_partial_msg_pages().

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
12 years agolibceph: use kernel_sendpage() for sending zeroes
Alex Elder [Wed, 7 Mar 2012 17:40:08 +0000 (11:40 -0600)]
libceph: use kernel_sendpage() for sending zeroes

If a message queued for send gets revoked, zeroes are sent over the
wire instead of any unsent data.  This is done by constructing a
message and passing it to kernel_sendmsg() via ceph_tcp_sendmsg().

Since we are already working with a page in this case we can use
the sendpage interface instead.  Create a new ceph_tcp_sendpage()
helper that sets up flags to match the way ceph_tcp_sendmsg()
does now.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
12 years agolibceph: fix inverted crc option logic
Alex Elder [Wed, 7 Mar 2012 17:40:08 +0000 (11:40 -0600)]
libceph: fix inverted crc option logic

CRC's are computed for all messages between ceph entities.  The CRC
computation for the data portion of message can optionally be
disabled using the "nocrc" (common) ceph option.  The default is
for CRC computation for the data portion to be enabled.

Unfortunately, the code that implements this feature interprets the
feature flag wrong, meaning that by default the CRC's have *not*
been computed (or checked) for the data portion of messages unless
the "nocrc" option was supplied.

Fix this, in write_partial_msg_pages() and read_partial_message().
Also change the flag variable in write_partial_msg_pages() to be
"no_datacrc" to match the usage elsewhere in the file.

This fixes http://tracker.newdream.net/issues/2064

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
12 years agolibceph: some simple changes
Alex Elder [Wed, 15 Feb 2012 13:43:55 +0000 (07:43 -0600)]
libceph: some simple changes

Nothing too big here.
    - define the size of the buffer used for consuming ignored
      incoming data using a symbolic constant
    - simplify the condition determining whether to unmap the page
      in write_partial_msg_pages(): do it for crc but not if the
      page is the zero page

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: small refactor in write_partial_kvec()
Alex Elder [Wed, 15 Feb 2012 13:43:54 +0000 (07:43 -0600)]
libceph: small refactor in write_partial_kvec()

Make a small change in the code that counts down kvecs consumed by
a ceph_tcp_sendmsg() call.  Same functionality, just blocked out
a little differently.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: do crc calculations outside loop
Alex Elder [Wed, 15 Feb 2012 13:43:54 +0000 (07:43 -0600)]
libceph: do crc calculations outside loop

Move blocks of code out of loops in read_partial_message_section()
and read_partial_message().  They were only was getting called at
the end of the last iteration of the loop anyway.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: separate CRC calculation from byte swapping
Alex Elder [Wed, 15 Feb 2012 13:43:54 +0000 (07:43 -0600)]
libceph: separate CRC calculation from byte swapping

Calculate CRC in a separate step from rearranging the byte order
of the result, to improve clarity and readability.

Use offsetof() to determine the number of bytes to include in the
CRC calculation.

In read_partial_message(), switch which value gets byte-swapped,
since the just-computed CRC is already likely to be in a register.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: use "do" in CRC-related Boolean variables
Alex Elder [Wed, 15 Feb 2012 13:43:54 +0000 (07:43 -0600)]
libceph: use "do" in CRC-related Boolean variables

Change the name (and type) of a few CRC-related Boolean local
variables so they contain the word "do", to distingish their purpose
from variables used for holding an actual CRC value.

Note that in the process of doing this I identified a fairly serious
logic error in write_partial_msg_pages():  the value of "do_crc"
assigned appears to be the opposite of what it should be.  No
attempt to fix this is made here; this change preserves the
erroneous behavior.  The problem I found is documented here:
    http://tracker.newdream.net/issues/2064

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: ensure Boolean options support both senses
Alex Elder [Wed, 15 Feb 2012 13:43:54 +0000 (07:43 -0600)]
ceph: ensure Boolean options support both senses

Many ceph-related Boolean options offer the ability to both enable
and disable a feature.  For all those that don't offer this, add
a new option so that they do.

Note that ceph_show_options()--which reports mount options currently
in effect--only reports the option if it is different from the
default value.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: a few small changes
Alex Elder [Tue, 14 Feb 2012 20:05:33 +0000 (14:05 -0600)]
libceph: a few small changes

This gathers a number of very minor changes:
    - use %hu when formatting the a socket address's address family
    - null out the ceph_msgr_wq pointer after the queue has been
      destroyed
    - drop a needless cast in ceph_write_space()
    - add a WARN() call in ceph_state_change() in the event an
      unrecognized socket state is encountered
    - rearrange the logic in ceph_con_get() and ceph_con_put() so
      that:
        - the reference counts are only atomically read once
- the values displayed via dout() calls are known to
  be meaningful at the time they are formatted

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: make ceph_tcp_connect() return int
Alex Elder [Tue, 14 Feb 2012 20:05:33 +0000 (14:05 -0600)]
libceph: make ceph_tcp_connect() return int

There is no real need for ceph_tcp_connect() to return the socket
pointer it creates, since it already assigns it to con->sock, which
is visible to the caller.  Instead, have it return an error code,
which tidies things up a bit.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: encapsulate some messenger cleanup code
Alex Elder [Tue, 14 Feb 2012 20:05:33 +0000 (14:05 -0600)]
libceph: encapsulate some messenger cleanup code

Define a helper function to perform various cleanup operations.  Use
it both in the exit routine and in the init routine in the event of
an error.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: make ceph_msgr_wq private
Alex Elder [Tue, 14 Feb 2012 20:05:33 +0000 (14:05 -0600)]
libceph: make ceph_msgr_wq private

The messenger workqueue has no need to be public.  So give it static
scope.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: encapsulate connection kvec operations
Alex Elder [Tue, 14 Feb 2012 20:05:33 +0000 (14:05 -0600)]
libceph: encapsulate connection kvec operations

Encapsulate the operation of adding a new chunk of data to the next
open slot in a ceph_connection's out_kvec array.  Also add a "reset"
operation to make subsequent add operations start at the beginning
of the array again.

Use these routines throughout, avoiding duplicate code and ensuring
all calls are handled consistently.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: move prepare_write_banner()
Alex Elder [Tue, 14 Feb 2012 20:05:33 +0000 (14:05 -0600)]
libceph: move prepare_write_banner()

One of the arguments to prepare_write_connect() indicates whether it
is being called immediately after a call to prepare_write_banner().
Move the prepare_write_banner() call inside prepare_write_connect(),
and reinterpret (and rename) the "after_banner" argument so it
indicates that prepare_write_connect() should *make* the call
rather than should know it has already been made.

This was split out from the next patch to highlight this change in
logic.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: don't drop the rbd_id too early
Alex Elder [Wed, 8 Feb 2012 22:11:14 +0000 (16:11 -0600)]
rbd: don't drop the rbd_id too early

Currently an rbd device's id is released when it is removed, but it
is done before the code is run to clean up sysfs-related files (such
as /sys/bus/rbd/devices/1).

It's possible that an rbd is still in use after the rbd_remove()
call has been made.  It's essentially the same as an active inode
that stays around after it has been removed--until its final close
operation.  This means that the id shows up as free for reuse at a
time it should not be.

The effect of this was seen by Jens Rehpoehler, who:
    - had a filesystem mounted on an rbd device
    - unmapped that filesystem (without unmounting)
    - found that the mount still worked properly
    - but hit a panic when he attempted to re-map a new rbd device

This re-map attempt found the previously-unmapped id available.
The subsequent attempt to reuse it was met with a panic while
attempting to (re-)install the sysfs entry for the new mapped
device.

Fix this by holding off "putting" the rbd id, until the rbd_device
release function is called--when the last reference is finally
dropped.

Note: This fixes: http://tracker.newdream.net/issues/1907

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: small changes
Alex Elder [Tue, 7 Feb 2012 18:03:37 +0000 (12:03 -0600)]
rbd: small changes

Here is another set of small code tidy-ups:
    - Define SECTOR_SHIFT and SECTOR_SIZE, and use these symbolic
      names throughout.  Tell the blk_queue system our physical
      block size, in the (unlikely) event we want to use something
      other than the default.
    - Delete the definition of struct rbd_info, which is never used.
    - Move the definition of dev_to_rbd() down in its source file,
      just above where it gets first used, and change its name to
      dev_to_rbd_dev().
    - Replace an open-coded operation in rbd_dev_release() to use
      dev_to_rbd_dev() instead.
    - Calculate the segment size for a given rbd_device just once in
      rbd_init_disk().
    - Use the '%zd' conversion specifier in rbd_snap_size_show(),
      since the value formatted is a size_t.
    - Switch to the '%llu' conversion specifier in rbd_snap_id_show().
      since the value formatted is unsigned.

Signed-off-by: Alex Elder <elder@dreamhost.com>
12 years agorbd: do some refactoring
Alex Elder [Tue, 7 Feb 2012 18:03:36 +0000 (12:03 -0600)]
rbd: do some refactoring

A few blocks of code are rearranged a bit here:
    - In rbd_header_from_disk():
- Don't bother computing snap_count until we're sure the
  on-disk header starts with a good signature.
- Move a few independent lines of code so they are *after* a
  check for a failed memory allocation.
- Get rid of unnecessary local variable "ret".
    - Make a few other changes in rbd_read_header(), similar to the
      above--just moving things around a bit while preserving the
      functionality.
    - In rbd_rq_fn(), just assign rq in the while loop's controlling
      expression rather than duplicating it before and at the end of
      the loop body.  This allows the use of "continue" rather than
      "goto next" in a number of spots.
    - Rearrange the logic in snap_by_name().  End result is the same.

Signed-off-by: Alex Elder <elder@dreamhost.com>
12 years agorbd: fix module sysfs setup/teardown code
Alex Elder [Tue, 7 Feb 2012 18:03:36 +0000 (12:03 -0600)]
rbd: fix module sysfs setup/teardown code

Once rbd_bus_type is registered, it allows an "add" operation via
the /sys/bus/rbd/add bus attribute, and adding a new rbd device that
way establishes a connection between the device and rbd_root_dev.
But rbd_root_dev is not registered until after the rbd_bus_type
registration is complete.  This could (in principle anyway) result
in an invalid state.

Since rbd_root_dev has no tie to rbd_bus_type we can reorder these
two initializations and never be faced with this scenario.

In addition, unregister the device in the event the bus registration
fails at module init time.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: don't allocate mon_addrs buffer in rbd_add()
Alex Elder [Thu, 2 Feb 2012 14:13:30 +0000 (08:13 -0600)]
rbd: don't allocate mon_addrs buffer in rbd_add()

The mon_addrs buffer in rbd_add is used to hold a copy of the
monitor IP addresses supplied via /sys/bus/rbd/add.  That is
passed to rbd_get_client(), which never modifies it (nor do
any of the functions it gets passed to thereafter)--the mon_addr
parameter to rbd_get_client() is a pointer to constant data, so it
can't be modifed.  Furthermore, rbd_get_client() has the length of
the mon_addrs buffer and that is used to ensure nothing goes beyond
its end.

Based on all this, there is no reason that a buffer needs to
be used to hold a copy of the mon_addrs provided via
/sys/bus/rbd/add.   Instead, the location within that passed-in
buffer can be provided, along with the length of the "token"
therein which represents the monitor IP's.

A small change to rbd_add_parse_args() allows the address within the
buffer to be passed back, and the length is already returned.  This
now means that, at least from the perspective of this interface,
there is no such thing as a list of monitor addresses that is too
long.

Signed-off-by: Alex Elder <elder@dreamhost.com>
12 years agorbd: have rbd_parse_args() report found mon_addrs size
Alex Elder [Thu, 2 Feb 2012 14:13:30 +0000 (08:13 -0600)]
rbd: have rbd_parse_args() report found mon_addrs size

The argument parsing routine already computes the size of the
mon_addrs buffer it extracts from the "command."  Pass it to the
caller so it can use it to provide the length to rbd_get_client().

Signed-off-by: Alex Elder <elder@dreamhost.com>
12 years agorbd: do a few checks at build time
Alex Elder [Thu, 2 Feb 2012 14:13:30 +0000 (08:13 -0600)]
rbd: do a few checks at build time

This is a bit gratuitous, but there are a few things that can be
verified at build time rather than run time, so do that.

Signed-off-by: Alex Elder <elder@dreamhost.com>
12 years agorbd: don't use sscanf() in rbd_add_parse_args()
Alex Elder [Thu, 2 Feb 2012 14:13:30 +0000 (08:13 -0600)]
rbd: don't use sscanf() in rbd_add_parse_args()

Make use of a few simple helper routines to parse the arguments
rather than sscanf().  This will treat both missing and too-long
arguments as invalid input (rather than silently truncating the
input in the too-long case).  In time this can also be used by
rbd_add() to use the passed-in buffer in place, rather than copying
its contents into new buffers.

It appears to me that the sscanf() previously used would not
correctly handle a supplied snapshot--the two final "%s" conversion
specifications were not separated by a space, and I'm not sure
how sscanf() handles that situation.  It may not be well-defined.
So that may be a bug this change fixes (but I didn't verify that).

The sizes of the mon_addrs and options buffers are now passed to
rbd_add_parse_args(), so they can be supplied to copy_token().

Signed-off-by: Alex Elder <elder@dreamhost.com>
12 years agorbd: encapsulate argument parsing for rbd_add()
Alex Elder [Thu, 2 Feb 2012 14:13:30 +0000 (08:13 -0600)]
rbd: encapsulate argument parsing for rbd_add()

Move the code that parses the arguments provided to rbd_add() (which
are supplied via /sys/bus/rbd/add) into a separate function.

Also rename the "mon_dev_name" variable in rbd_add() to be
"mon_addrs".   The variable represents a list of one or more
comma-separated monitor IP addresses, each with an optional port
number.  I think "mon_addrs" captures that notion a little better.

Signed-off-by: Alex Elder <elder@dreamhost.com>
12 years agorbd: simplify error handling in rbd_add()
Alex Elder [Thu, 2 Feb 2012 14:13:30 +0000 (08:13 -0600)]
rbd: simplify error handling in rbd_add()

If a couple pointers are initialized to NULL then a single
"out_nomem" label can be used for all of the memory allocation
failure cases in rbd_add().

Also, get rid of the "irc" local variable there.  There is no
real need for "rc" to be type ssize_t, and it can be used in
the spot "irc" was.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: reduce memory used for rbd_dev fields
Alex Elder [Thu, 2 Feb 2012 14:13:30 +0000 (08:13 -0600)]
rbd: reduce memory used for rbd_dev fields

The length of the string containing the monitor address
specification(s) will never exceed the length of the string passed
in to rbd_add().  The same holds true for the ceph + rbd options
string.  So reduce the amount of memory allocated for these to
that length rather than the maximum (1024 bytes).

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: have rbd_get_client() return a rbd_client
Alex Elder [Thu, 2 Feb 2012 14:13:30 +0000 (08:13 -0600)]
rbd: have rbd_get_client() return a rbd_client

Since rbd_get_client() currently returns an error code.  It assigns
the rbd_client field of the rbd_device structure it is passed if
successful.  Instead, have it return the created rbd_client
structure and return a pointer-coded error if there is an error.
This makes the assignment of the client pointer more obvious at the
call site.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: a few simple changes
Alex Elder [Sun, 29 Jan 2012 19:57:44 +0000 (13:57 -0600)]
rbd: a few simple changes

Here are a few very simple cleanups:
    - Add a "RBD_" prefix to the two driver name string definitions.
    - Move the definition of struct rbd_request below struct rbd_req_coll
      to avoid the need for an empty declaration of the latter.
    - Move and group the definitions of rbd_root_dev_release() and
      rbd_root_dev, as well as rbd_bus_type and rbd_bus_attrs[],
      close to the top of the file.  Arrange the latter so
      rbd_bus_type.bus_attrs can be initialized statically.
    - Get rid of an unnecessary local variable in rbd_open().
    - Rework some hokey logic in rbd_bus_add_dev(), so the value of
      "ret" at the end is either 0 or -ENOENT to avoid the need for
      the code duplication that was there.
    - Rename a goto target in rbd_add().

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: rename "node_lock"
Alex Elder [Sun, 29 Jan 2012 19:57:44 +0000 (13:57 -0600)]
rbd: rename "node_lock"

The spinlock used to protect rbd_client_list is named "node_lock".
Rename it to "rbd_client_list_lock" to make it more obvious what
it's for.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: move ctl_mutex lock inside rbd_client_create()
Alex Elder [Sun, 29 Jan 2012 19:57:44 +0000 (13:57 -0600)]
rbd: move ctl_mutex lock inside rbd_client_create()

Since rbd_client_create() is only called in one place, move the
acquisition of the mutex around that call inside that function.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: move ctl_mutex lock inside rbd_get_client()
Alex Elder [Sun, 29 Jan 2012 19:57:44 +0000 (13:57 -0600)]
rbd: move ctl_mutex lock inside rbd_get_client()

Since rbd_get_client() is only called in one place, move the
acquisition of the mutex around that call inside that function.

Furthermore, within rbd_get_client(), it appears the mutex only
needs to be held while calling rbd_client_create().  (Moving
the lock inside that function will wait for the next patch.)

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: release client list lock sooner
Alex Elder [Sun, 29 Jan 2012 19:57:44 +0000 (13:57 -0600)]
rbd: release client list lock sooner

In rbd_get_client(), if a client is reused, a number of things
get done while still holding the list lock unnecessarily.

This just moves a few things that need no lock protection outside
the lock.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: restore previous rbd id sequence behavior
Alex Elder [Sun, 29 Jan 2012 19:57:44 +0000 (13:57 -0600)]
rbd: restore previous rbd id sequence behavior

It used to be that selecting a new unique identifier for an added
rbd device required searching all existing ones to find the highest
id is used.  A recent change made that unnecessary, but made it
so that id's used were monotonically non-decreasing.  It's a bit
more pleasant to have smaller rbd id's though, and this change
makes ids get allocated as they were before--each new id is one more
than the maximum currently in use.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: tie rbd_dev_list changes to rbd_id operations
Alex Elder [Thu, 2 Feb 2012 14:13:29 +0000 (08:13 -0600)]
rbd: tie rbd_dev_list changes to rbd_id operations

The only time entries are added to or removed from the global
rbd_dev_list is exactly when a "put" or "get" operation is being
performed on a rbd_dev's id.  So just move the list management code
into get/put routines.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: protect the rbd_dev_list with a spinlock
Alex Elder [Sun, 29 Jan 2012 19:57:44 +0000 (13:57 -0600)]
rbd: protect the rbd_dev_list with a spinlock

The rbd_dev_list is just a simple list of all the current
rbd_devices.  Using the ctl_mutex as a concurrency guard is
overkill.  Instead, use a spinlock for that specific purpose.

This also reduces the window that the ctl_mutex needs to be held in
rbd_add().

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: rework calculation of new rbd id's
Alex Elder [Sun, 29 Jan 2012 19:57:44 +0000 (13:57 -0600)]
rbd: rework calculation of new rbd id's

In order to select a new unique identifier for an added rbd device,
the list of all existing ones is searched and a value one greater
than the highest id is used.

The list search can be avoided by using an atomic variable that
keeps track of the current highest id.  Using a get/put model for
id's we can limit the boundless growth of id numbers a bit by
arranging to reuse the current highest id once it gets released.
Add these calls to "put" the id when an rbd is getting removed.

Note that this changes the pattern of device id's used--new values
will never be below the highest one seen so far (even if there
exists an unused lower one).  I assert this is OK because the key
property of an rbd id is its uniqueness, not its magnitude.

Regardless, a follow-on patch will restore the old way of doing
things, I just think this commit just makes the incremental change
to atomics a little easier to understand.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: encapsulate new rbd id selection
Alex Elder [Sun, 29 Jan 2012 19:57:43 +0000 (13:57 -0600)]
rbd: encapsulate new rbd id selection

Move the loop that finds a new unique rbd id to use into
its own helper function.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: use a single value of snap_name to mean no snap
Josh Durgin [Tue, 22 Nov 2011 02:19:13 +0000 (18:19 -0800)]
rbd: use a single value of snap_name to mean no snap

There's already a constant for this anyway.

Since rbd_header_set_snap() is only used to set the rbd device
snap_name field, just do that within that function rather than
having it take the snap_name as an argument.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
v2: Changed interface rbd_header_set_snap() so it explicitly updates
    the snap_name in the rbd_device.  Also added a BUILD_BUG_ON()
    to verify the size of the snap_name field is sufficient for
    SNAP_HEAD_NAME.

12 years agorbd: do not duplicate ceph_client pointer in rbd_device
Alex Elder [Tue, 24 Jan 2012 16:08:37 +0000 (10:08 -0600)]
rbd: do not duplicate ceph_client pointer in rbd_device

The rbd_device structure maintains a duplicate copy of the
ceph_client pointer maintained in its rbd_client structure.  There
appears to be no good reason for this, and its presence presents a
risk of them getting out of synch or otherwise misused.  So kill it
off, and use the rbd_client copy only.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: make ceph_parse_options() return a pointer
Alex Elder [Tue, 24 Jan 2012 16:08:36 +0000 (10:08 -0600)]
rbd: make ceph_parse_options() return a pointer

ceph_parse_options() takes the address of a pointer as an argument
and uses it to return the address of an allocated structure if
successful.  With this interface is not evident at call sites that
the pointer is always initialized.  Change the interface to return
the address instead (or a pointer-coded error code) to make the
validity of the returned pointer obvious.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agorbd: a few small cleanups
Alex Elder [Tue, 24 Jan 2012 16:08:36 +0000 (10:08 -0600)]
rbd: a few small cleanups

Some minor cleanups in "drivers/block/rbd.c:
    - Use the more meaningful "RBD_MAX_OBJ_NAME_LEN" in place if "96"
      in the definition of RBD_MAX_MD_NAME_LEN.
    - Use DEFINE_SPINLOCK() to define and initialize node_lock.
    - Drop a needless (char *) cast in parse_rbd_opts_token().
    - Make a few minor formatting changes.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: make ceph_setxattr() and ceph_removexattr() more alike
Alex Elder [Mon, 23 Jan 2012 21:49:28 +0000 (15:49 -0600)]
ceph: make ceph_setxattr() and ceph_removexattr() more alike

This patch just rearranges a few bits of code to make more
portions of ceph_setxattr() and ceph_removexattr() identical.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: avoid repeatedly computing the size of constant vxattr names
Alex Elder [Mon, 23 Jan 2012 21:49:28 +0000 (15:49 -0600)]
ceph: avoid repeatedly computing the size of constant vxattr names

All names defined in the directory and file virtual extended
attribute tables are constant, and the size of each is known at
compile time.  So there's no need to compute their length every
time any file's attribute is listed.

Record the length of each string and use it when needed to determine
the space need to represent them.  In addition, compute the
aggregate size of strings in each table just once at initialization
time.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: encode type in vxattr callback routines
Alex Elder [Mon, 23 Jan 2012 21:49:28 +0000 (15:49 -0600)]
ceph: encode type in vxattr callback routines

The names of the callback functions used for virtual extended
attributes are based only on the last component of the attribute
name.  Because of the way these are defined, this precludes allowing
a single (lowest) attribute name for different callbacks, dependent
on the type of file being operated on.  (For example, it might be
nice to support both "ceph.dir.layout" and "ceph.file.layout".)

Just change the callback names to avoid this problem.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: drop "_cb" from name of struct ceph_vxattr_cb
Alex Elder [Mon, 23 Jan 2012 21:49:28 +0000 (15:49 -0600)]
ceph: drop "_cb" from name of struct ceph_vxattr_cb

A struct ceph_vxattr_cb does not represent a callback at all, but
rather a virtual extended attribute itself.  Drop the "_cb" suffix
from its name to reflect that.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: use macros to normalize vxattr table definitions
Alex Elder [Mon, 23 Jan 2012 21:49:28 +0000 (15:49 -0600)]
ceph: use macros to normalize vxattr table definitions

Entries in the ceph virtual extended attribute tables all follow a
distinct pattern in their definition.  Enforce this pattern through
the use of a macro.

Also, a null name field signals the end of the table, so make that
be the first field in the ceph_vxattr_cb structure.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: use a symbolic name for "ceph." extended attribute namespace
Alex Elder [Mon, 23 Jan 2012 21:49:28 +0000 (15:49 -0600)]
ceph: use a symbolic name for "ceph." extended attribute namespace

Use symbolic constants to define the top-level prefix for "ceph."
extended attribute names.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: pass inode rather than table to ceph_match_vxattr()
Alex Elder [Mon, 23 Jan 2012 21:49:27 +0000 (15:49 -0600)]
ceph: pass inode rather than table to ceph_match_vxattr()

All callers of ceph_match_vxattr() determine what to pass as the
first argument by calling ceph_inode_vxattrs(inode).  Just do that
inside ceph_match_vxattr() itself, changing it to take an inode
rather than the vxattr pointer as its first argument.

Also ensure the function works correctly for an empty table (i.e.,
containing only a terminating null entry).

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: don't null-terminate xattr values
Alex Elder [Mon, 23 Jan 2012 21:49:27 +0000 (15:49 -0600)]
ceph: don't null-terminate xattr values

For some reason, ceph_setxattr() allocates an extra byte in which a
'\0' is stored past the end of an extended attribute value.  This is
not needed, and is potentially misleading, so get rid of it.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: eliminate some abusive casts
Alex Elder [Mon, 23 Jan 2012 21:49:27 +0000 (15:49 -0600)]
ceph: eliminate some abusive casts

This fixes some spots where a type cast to (void *) was used as
as a universal type hiding mechanism.  Instead, properly cast the
type to the intended target type.

Signed-off-by: Alex Elder <elder@newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: eliminate some needless casts
Alex Elder [Mon, 23 Jan 2012 21:49:27 +0000 (15:49 -0600)]
ceph: eliminate some needless casts

This eliminates type casts in some places where they are not
required.

Signed-off-by: Alex Elder <elder@newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: kill addr_str_lock spinlock; use atomic instead
Alex Elder [Mon, 23 Jan 2012 21:49:27 +0000 (15:49 -0600)]
ceph: kill addr_str_lock spinlock; use atomic instead

A spinlock is used to protect a value used for selecting an array
index for a string used for formatting a socket address for human
consumption.  The index is reset to 0 if it ever reaches the maximum
index value.

Instead, use an ever-increasing atomic variable as a sequence
number, and compute the array index by masking off all but the
sequence number's lowest bits.  Make the number of entries in the
array a power of two to allow the use of such a mask (to avoid jumps
in the index value when the sequence number wraps).

The length of these strings is somewhat arbitrarily set at 60 bytes.
The worst-case length of a string produced is 54 bytes, for an IPv6
address that can't be shortened, e.g.:
    [1234:5678:9abc:def0:1111:2222:123.234.210.100]:32767
Change it so we arbitrarily use 64 bytes instead; if nothing else
it will make the array of these line up better in hex dumps.

Rename a few things to reinforce the distinction between the number
of strings in the array and the length of individual strings.

Signed-off-by: Alex Elder <elder@newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: make use of "else" where appropriate
Alex Elder [Mon, 23 Jan 2012 21:49:27 +0000 (15:49 -0600)]
ceph: make use of "else" where appropriate

Rearrange ceph_tcp_connect() a bit, making use of "else" rather than
re-testing a value with consecutive "if" statements.  Don't record a
connection's socket pointer unless the connect operation is
successful.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: use a shared zero page rather than one per messenger
Alex Elder [Mon, 23 Jan 2012 21:49:27 +0000 (15:49 -0600)]
ceph: use a shared zero page rather than one per messenger

Each messenger allocates a page to be used when writing zeroes
out in the event of error or other abnormal condition.  Instead,
use the kernel ZERO_PAGE() for that purpose.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: fix overflow check in build_snap_context()
Xi Wang [Thu, 16 Feb 2012 16:56:29 +0000 (11:56 -0500)]
ceph: fix overflow check in build_snap_context()

The overflow check for a + n * b should be (n > (ULONG_MAX - a) / b),
rather than (n > ULONG_MAX / b - a).

Signed-off-by: Xi Wang <xi.wang@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agolibceph: fix overflow check in crush_decode()
Xi Wang [Thu, 16 Feb 2012 16:55:48 +0000 (11:55 -0500)]
libceph: fix overflow check in crush_decode()

The existing overflow check (n > ULONG_MAX / b) didn't work, because
n = ULONG_MAX / b would both bypass the check and still overflow the
allocation size a + n * b.

The correct check should be (n > (ULONG_MAX - a) / b).

Signed-off-by: Xi Wang <xi.wang@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: avoid panic with mismatched symlink sizes in fill_inode()
Xi Wang [Fri, 3 Feb 2012 14:55:36 +0000 (09:55 -0500)]
ceph: avoid panic with mismatched symlink sizes in fill_inode()

Return -EINVAL rather than panic if iinfo->symlink_len and inode->i_size
do not match.

Also use kstrndup rather than kmalloc/memcpy.

Signed-off-by: Xi Wang <xi.wang@gmail.com>
Reviewed-by: Alex Elder <elder@dreamhost.com>
12 years agoceph: use 2 instead of 1 as fallback for 32-bit inode number
Amon Ott [Mon, 23 Jan 2012 17:25:23 +0000 (09:25 -0800)]
ceph: use 2 instead of 1 as fallback for 32-bit inode number

The root directory of the Ceph mount has inode number 1, so falling back
to 1 always creates a collision. 2 is unused on my test systems and seems
less likely to collide.

Signed-off-by: Amon Ott <ao@m-privacy.de>
Signed-off-by: Sage Weil <sage@newdream.net>
12 years agoceph: don't reset s_cap_ttl to zero
Alex Elder [Fri, 13 Jan 2012 01:48:11 +0000 (17:48 -0800)]
ceph: don't reset s_cap_ttl to zero

Avoid the need to check for a special zero s_cap_ttl value by just
using (jiffies - 1) as the value assigned to indicate "sometime in
the past."

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
12 years agonet/ceph: Only clear SOCK_NOSPACE when there is sufficient space in the socket buffer
Jim Schutt [Wed, 29 Feb 2012 15:30:58 +0000 (08:30 -0700)]
net/ceph: Only clear SOCK_NOSPACE when there is sufficient space in the socket buffer

The Ceph messenger would sometimes queue multiple work items to write
data to a socket when the socket buffer was full.

Fix this problem by making ceph_write_space() use SOCK_NOSPACE in the
same way that net/core/stream.c:sk_stream_write_space() does, i.e.,
clearing it only when sufficient space is available in the socket buffer.

Signed-off-by: Jim Schutt <jaschut@sandia.gov>
Reviewed-by: Alex Elder <elder@dreamhost.com>
12 years agoLinux 3.3 v3.3
Linus Torvalds [Sun, 18 Mar 2012 23:15:34 +0000 (16:15 -0700)]
Linux 3.3

12 years agoDon't limit non-nested epoll paths
Jason Baron [Fri, 16 Mar 2012 20:34:03 +0000 (16:34 -0400)]
Don't limit non-nested epoll paths

Commit 28d82dc1c4ed ("epoll: limit paths") that I did to limit the
number of possible wakeup paths in epoll is causing a few applications
to longer work (dovecot for one).

The original patch is really about limiting the amount of epoll nesting
(since epoll fds can be attached to other fds). Thus, we probably can
allow an unlimited number of paths of depth 1. My current patch limits
it at 1000. And enforce the limits on paths that have a greater depth.

This is captured in: https://bugzilla.redhat.com/show_bug.cgi?id=681578

Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
12 years agoMerge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
Linus Torvalds [Sun, 18 Mar 2012 02:22:24 +0000 (19:22 -0700)]
Merge git://git./linux/kernel/git/davem/net

Pull networking changes from David Miller:
 "1) icmp6_dst_alloc() returns NULL instead of ERR_PTR() leading to
     crashes, particularly during shutdown.  Reported by Dave Jones and
     fixed by Eric Dumazet.

  2) hyperv and wimax/i2400m return NETDEV_TX_BUSY when they have
     already freed the SKB, which causes crashes as to the caller this
     means requeue the packet.  Fixes from Eric Dumazet.

  3) usbnet driver doesn't allocate the right amount of headroom on
     fresh RX SKBs, fix from Eric Dumazet.

  4) Fix regression in ip6_mc_find_dev_rcu(), as an RCU lookup it
     abolutely should not take a reference to 'dev', this leads to
     leaks.  Fix from RonQing Li.

  5) Fix netfilter ctnetlink race between delete and timeout expiration.
     From Pablo Neira Ayuso.

  6) Revert SFQ change which causes regressions, specifically queueing
     to tail can lead to unavoidable flow starvation.  From Eric
     Dumazet.

  7) Fix a memory leak and a crash on corrupt firmware files in bnx2x,
     from Michal Schmidt."

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net:
  netfilter: ctnetlink: fix race between delete and timeout expiration
  ipv6: Don't dev_hold(dev) in ip6_mc_find_dev_rcu.
  wimax/i2400m: fix erroneous NETDEV_TX_BUSY use
  net/hyperv: fix erroneous NETDEV_TX_BUSY use
  net/usbnet: reserve headroom on rx skbs
  bnx2x: fix memory leak in bnx2x_init_firmware()
  bnx2x: fix a crash on corrupt firmware file
  sch_sfq: revert dont put new flow at the end of flows
  ipv6: fix icmp6_dst_alloc()

12 years agoMerge branch 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
Linus Torvalds [Sat, 17 Mar 2012 16:54:16 +0000 (09:54 -0700)]
Merge branch 'perf-urgent-for-linus' of git://git./linux/kernel/git/tip/tip

Pull perf fixes from Ingo Molnar.

* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  perf tools, x86: Build perf on older user-space as well
  perf tools: Use scnprintf where applicable
  perf tools: Incorrect use of snprintf results in SEGV

12 years agonetfilter: ctnetlink: fix race between delete and timeout expiration
Pablo Neira Ayuso [Fri, 16 Mar 2012 02:00:34 +0000 (02:00 +0000)]
netfilter: ctnetlink: fix race between delete and timeout expiration

Kerin Millar reported hardlockups while running `conntrackd -c'
in a busy firewall. That system (with several processors) was
acting as backup in a primary-backup setup.

After several tries, I found a race condition between the deletion
operation of ctnetlink and timeout expiration. This patch fixes
this problem.

Tested-by: Kerin Millar <kerframil@gmail.com>
Reported-by: Kerin Millar <kerframil@gmail.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 years agoipv6: Don't dev_hold(dev) in ip6_mc_find_dev_rcu.
RongQing.Li [Thu, 15 Mar 2012 22:54:14 +0000 (22:54 +0000)]
ipv6: Don't dev_hold(dev) in ip6_mc_find_dev_rcu.

ip6_mc_find_dev_rcu() is called with rcu_read_lock(), so don't
need to dev_hold().
With dev_hold(), not corresponding dev_put(), will lead to leak.

[ bug introduced in 96b52e61be1 (ipv6: mcast: RCU conversions) ]

Signed-off-by: RongQing.Li <roy.qing.li@gmail.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 years agoMerge branch 'akpm' (more patches from Andrew)
Linus Torvalds [Sat, 17 Mar 2012 00:14:55 +0000 (17:14 -0700)]
Merge branch 'akpm' (more patches from Andrew)

Merge some more email patches from Andrew Morton:
 "A couple of nilfs fixes"

* emailed from Andrew Morton <akpm@linux-foundation.org>:
  nilfs2: fix NULL pointer dereference in nilfs_load_super_block()
  nilfs2: clamp ns_r_segments_percentage to [1, 99]

12 years agonilfs2: fix NULL pointer dereference in nilfs_load_super_block()
Ryusuke Konishi [Sat, 17 Mar 2012 00:08:39 +0000 (17:08 -0700)]
nilfs2: fix NULL pointer dereference in nilfs_load_super_block()

According to the report from Slicky Devil, nilfs caused kernel oops at
nilfs_load_super_block function during mount after he shrank the
partition without resizing the filesystem:

 BUG: unable to handle kernel NULL pointer dereference at 00000048
 IP: [<d0d7a08e>] nilfs_load_super_block+0x17e/0x280 [nilfs2]
 *pde = 00000000
 Oops: 0000 [#1] PREEMPT SMP
 ...
 Call Trace:
  [<d0d7a87b>] init_nilfs+0x4b/0x2e0 [nilfs2]
  [<d0d6f707>] nilfs_mount+0x447/0x5b0 [nilfs2]
  [<c0226636>] mount_fs+0x36/0x180
  [<c023d961>] vfs_kern_mount+0x51/0xa0
  [<c023ddae>] do_kern_mount+0x3e/0xe0
  [<c023f189>] do_mount+0x169/0x700
  [<c023fa9b>] sys_mount+0x6b/0xa0
  [<c04abd1f>] sysenter_do_call+0x12/0x28
 Code: 53 18 8b 43 20 89 4b 18 8b 4b 24 89 53 1c 89 43 24 89 4b 20 8b 43
 20 c7 43 2c 00 00 00 00 23 75 e8 8b 50 68 89 53 28 8b 54 b3 20 <8b> 72
 48 8b 7a 4c 8b 55 08 89 b3 84 00 00 00 89 bb 88 00 00 00
 EIP: [<d0d7a08e>] nilfs_load_super_block+0x17e/0x280 [nilfs2] SS:ESP 0068:ca9bbdcc
 CR2: 0000000000000048

This turned out due to a defect in an error path which runs if the
calculated location of the secondary super block was invalid.

This patch fixes it and eliminates the reported oops.

Reported-by: Slicky Devil <slicky.dvl@gmail.com>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Tested-by: Slicky Devil <slicky.dvl@gmail.com>
Cc: <stable@vger.kernel.org> [2.6.30+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
12 years agonilfs2: clamp ns_r_segments_percentage to [1, 99]
Haogang Chen [Sat, 17 Mar 2012 00:08:38 +0000 (17:08 -0700)]
nilfs2: clamp ns_r_segments_percentage to [1, 99]

ns_r_segments_percentage is read from the disk.  Bogus or malicious
value could cause integer overflow and malfunction due to meaningless
disk usage calculation.  This patch reports error when mounting such
bogus volumes.

Signed-off-by: Haogang Chen <haogangchen@gmail.com>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
12 years agoMerge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jmorris...
Linus Torvalds [Sat, 17 Mar 2012 00:04:02 +0000 (17:04 -0700)]
Merge branch 'for-linus' of git://git./linux/kernel/git/jmorris/linux-security

Pull maintainer update from James Morris:
 "Please pull this patch which adds Serge as maintainer of the
  capabilities code, as discussed on lwn and the lsm list.

  New capabilities must be signed off by the maintainer, and new uses of
  any capabilities should at be cc'd to the maintainer."

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jmorris/linux-security:
  MAINTAINERS: Add Serge as maintainer of capabilities

12 years agoMerge tag 'for-linus' of git://linux-c6x.org/git/projects/linux-c6x-upstreaming
Linus Torvalds [Sat, 17 Mar 2012 00:03:15 +0000 (17:03 -0700)]
Merge tag 'for-linus' of git://linux-c6x.org/git/projects/linux-c6x-upstreaming

Pull c6x bugfix from Mark Salter:
 "Remove dead code from entry.S which causes a build failure when using
  a newer assembler (v2.22 complains about it, v2.20 ignores it)."

* tag 'for-linus' of git://linux-c6x.org/git/projects/linux-c6x-upstreaming:
  C6X: remove dead code from entry.S

12 years agoafs: Remote abort can cause BUG in rxrpc code
Anton Blanchard [Fri, 16 Mar 2012 10:28:19 +0000 (10:28 +0000)]
afs: Remote abort can cause BUG in rxrpc code

When writing files to afs I sometimes hit a BUG:

kernel BUG at fs/afs/rxrpc.c:179!

With a backtrace of:

afs_free_call
afs_make_call
afs_fs_store_data
afs_vnode_store_data
afs_write_back_from_locked_page
afs_writepages_region
afs_writepages

The cause is:

ASSERT(skb_queue_empty(&call->rx_queue));

Looking at a tcpdump of the session the abort happens because we
are exceeding our disk quota:

rx abort fs reply store-data error diskquota exceeded (32)

So the abort error is valid. We hit the BUG because we haven't
freed all the resources for the call.

By freeing any skbs in call->rx_queue before calling afs_free_call
we avoid hitting leaking memory and avoid hitting the BUG.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: David Howells <dhowells@redhat.com>
Cc: <stable@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
12 years agoafs: Read of file returns EBADMSG
Anton Blanchard [Fri, 16 Mar 2012 10:28:07 +0000 (10:28 +0000)]
afs: Read of file returns EBADMSG

A read of a large file on an afs mount failed:

# cat junk.file > /dev/null
cat: junk.file: Bad message

Looking at the trace, call->offset wrapped since it is only an
unsigned short. In afs_extract_data:

        _enter("{%u},{%zu},%d,,%zu", call->offset, len, last, count);
...

        if (call->offset < count) {
                if (last) {
                        _leave(" = -EBADMSG [%d < %zu]", call->offset, count);
                        return -EBADMSG;
                }

Which matches the trace:

[cat   ] ==> afs_extract_data({65132},{524},1,,65536)
[cat   ] <== afs_extract_data() = -EBADMSG [0 < 65536]

call->offset went from 65132 to 0. Fix this by making call->offset an
unsigned int.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: David Howells <dhowells@redhat.com>
Cc: <stable@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
12 years agoC6X: remove dead code from entry.S
Mark Salter [Fri, 16 Mar 2012 13:27:57 +0000 (09:27 -0400)]
C6X: remove dead code from entry.S

The ENDPROC() on sys_fadvise64_c6x() in arch/c6x/kernel/entry.S is
outside of the conditional block with the matching ENTRY() macro. This
leads a newer (v2.22 vs. v2.20) assembler to complain:

  /tmp/ccGZBaPT.s: Assembler messages:
  /tmp/ccGZBaPT.s: Error: .size expression for sys_fadvise64_c6x does not evaluate to a constant

The conditional block became dead code when c6x switched to generic
unistd.h and should be removed along with the offending ENDPROC().

Signed-off-by: Mark Salter <msalter@redhat.com>
Acked-by: David Howells <dhowells@redhat.com>
12 years agowimax/i2400m: fix erroneous NETDEV_TX_BUSY use
Eric Dumazet [Wed, 14 Mar 2012 09:21:44 +0000 (09:21 +0000)]
wimax/i2400m: fix erroneous NETDEV_TX_BUSY use

A driver start_xmit() method cannot free skb and return NETDEV_TX_BUSY,
since caller is going to reuse freed skb.

In fact netif_tx_stop_queue() / netif_stop_queue() is needed before
returning NETDEV_TX_BUSY or you can trigger a ksoftirqd fatal loop.

In case of memory allocation error, only safe way is to drop the packet
and return NETDEV_TX_OK

Also increments tx_dropped counter

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Inaky Perez-Gonzalez <inaky.perez-gonzalez@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 years agonet/hyperv: fix erroneous NETDEV_TX_BUSY use
Eric Dumazet [Wed, 14 Mar 2012 08:53:34 +0000 (08:53 +0000)]
net/hyperv: fix erroneous NETDEV_TX_BUSY use

A driver start_xmit() method cannot free skb and return NETDEV_TX_BUSY,
since caller is going to reuse freed skb.

This is mostly a revert of commit bf769375c (staging: hv: fix the return
status of netvsc_start_xmit())

In fact netif_tx_stop_queue() / netif_stop_queue() is needed before
returning NETDEV_TX_BUSY or you can trigger a ksoftirqd fatal loop.

In case of memory allocation error, only safe way is to drop the packet
and return NETDEV_TX_OK

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 years agonet/usbnet: reserve headroom on rx skbs
Eric Dumazet [Wed, 14 Mar 2012 06:56:25 +0000 (06:56 +0000)]
net/usbnet: reserve headroom on rx skbs

network drivers should reserve some headroom on incoming skbs so that we
dont need expensive reallocations, eg forwarding packets in tunnels.

This NET_SKB_PAD padding is done in various helpers, like
__netdev_alloc_skb_ip_align() in this patch, combining NET_SKB_PAD and
NET_IP_ALIGN magic.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Oliver Neukum <oneukum@suse.de>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Acked-by: Oliver Neukum <oneukum@suse.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 years agobnx2x: fix memory leak in bnx2x_init_firmware()
Michal Schmidt [Thu, 15 Mar 2012 14:08:29 +0000 (14:08 +0000)]
bnx2x: fix memory leak in bnx2x_init_firmware()

When cycling the interface down and up, bnx2x_init_firmware() knows that
the firmware is already loaded, but nevertheless it allocates certain
arrays anew (init_data, init_ops, init_ops_offsets, iro_arr). The old
arrays are leaked.

Fix the leaks by returning early if the firmware was already loaded.
Because if the firmware is loaded, so are the arrays.

Signed-off-by: Michal Schmidt <mschmidt@redhat.com>
Acked-by: Eilon Greenstein <eilong@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 years agobnx2x: fix a crash on corrupt firmware file
Michal Schmidt [Thu, 15 Mar 2012 14:08:28 +0000 (14:08 +0000)]
bnx2x: fix a crash on corrupt firmware file

If the requested firmware is deemed corrupt and then released, reset the
pointer to NULL in order to avoid double-freeing it in
bnx2x_release_firmware() or dereferencing it in bnx2x_init_firmware().

Signed-off-by: Michal Schmidt <mschmidt@redhat.com>
Acked-by: Eilon Greenstein <eilong@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 years agosch_sfq: revert dont put new flow at the end of flows
Eric Dumazet [Tue, 13 Mar 2012 18:04:25 +0000 (18:04 +0000)]
sch_sfq: revert dont put new flow at the end of flows

This reverts commit d47a0ac7b6 (sch_sfq: dont put new flow at the end of
flows)

As Jesper found out, patch sounded great but has bad side effects.

In stress situation, pushing new flows in front of the queue can prevent
old flows doing any progress. Packets can stay in SFQ queue for
unlimited amount of time.

It's possible to add heuristics to limit this problem, but this would
add complexity outside of SFQ scope.

A more sensible answer to Dave Taht concerns (who reported the issued I
tried to solve in original commit) is probably to use a qdisc hierarchy
so that high prio packets dont enter a potentially crowded SFQ qdisc.

Reported-by: Jesper Dangaard Brouer <jdb@comx.dk>
Cc: Dave Taht <dave.taht@gmail.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 years agoipv6: fix icmp6_dst_alloc()
Eric Dumazet [Wed, 14 Mar 2012 21:13:11 +0000 (21:13 +0000)]
ipv6: fix icmp6_dst_alloc()

commit 87a115783 ( ipv6: Move xfrm_lookup() call down into
icmp6_dst_alloc().) forgot to convert one error path, leading
to crashes in mld_sendpack()

Many thanks to Dave Jones for providing a very complete bug report.

Reported-by: Dave Jones <davej@redhat.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 years agoMAINTAINERS: Add Serge as maintainer of capabilities
James Morris [Fri, 16 Mar 2012 01:05:48 +0000 (12:05 +1100)]
MAINTAINERS: Add Serge as maintainer of capabilities

Add Serge as maintainer of capabilities, per suggestion on LWN:
http://lwn.net/Articles/486306/

Signed-off-by: James Morris <james.l.morris@oracle.com>
12 years agoMerge branch 'akpm' (Andrew's patch-bomb)
Linus Torvalds [Fri, 16 Mar 2012 00:16:22 +0000 (17:16 -0700)]
Merge branch 'akpm' (Andrew's patch-bomb)

Merge patches from Andrew Morton:
 "Nine patches - some bug fixes and some MAINTAINERS fiddling."

* emailed from Andrew Morton <akpm@linux-foundation.org>:
  drivers/video/backlight/s6e63m0.c: fix corruption storing gamma mode
  MAINTAINERS: add entry for exynos mipi display drivers
  MAINTAINERS: fix link to Gustavo Padovans tree
  MAINTAINERS: add Johan to Bluetooth maintainers
  MAINTAINERS: Gustavo has moved
  prctl: use CAP_SYS_RESOURCE for PR_SET_MM option
  rapidio/tsi721: fix bug in register offset definitions
  MAINTAINERS: update ST's Mailing list for SPEAr
  memcg: free mem_cgroup by RCU to fix oops

12 years agoMerge branch 'i2c-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jdelvar...
Linus Torvalds [Fri, 16 Mar 2012 00:14:35 +0000 (17:14 -0700)]
Merge branch 'i2c-for-linus' of git://git./linux/kernel/git/jdelvare/staging

Pull i2c subsystem fixes from Jean Delvare.

* 'i2c-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jdelvare/staging:
  i2c-algo-bit: Fix spurious SCL timeouts under heavy load
  i2c-core: Comment says "transmitted" but means "received"