When ceph_osdc_handle_map() is called to process a new osd map,
kick_requests() is called to ensure all affected requests are
updated if necessary to reflect changes in the osd map. This
happens in two cases: whenever an incremental map update is
processed; and when a full map update (or the last one if there is
more than one) gets processed.
In the former case, the kick_requests() call is followed immediately
by a call to reset_changed_osds() to ensure any connections to osds
affected by the map change are reset. But for full map updates
this isn't done.
Both cases should be doing this osd reset.
Rather than duplicating the reset_changed_osds() call, move it into
the end of kick_requests().
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
The kick_requests() function is called by ceph_osdc_handle_map()
when an osd map change has been indicated. Its purpose is to
re-queue any request whose target osd is different from what it
was when it was originally sent.
It is structured as two loops, one for incomplete but registered
requests, and a second for handling completed linger requests.
As a special case, in the first loop if a request marked to linger
has not yet completed, it is moved from the request list to the
linger list. This is as a quick and dirty way to have the second
loop handle sending the request along with all the other linger
requests.
Because of the way it's done now, however, this quick and dirty
solution can result in these incomplete linger requests never
getting re-sent as desired. The problem lies in the fact that
the second loop only arranges for a linger request to be sent
if it appears its target osd has changed. This is the proper
handling for *completed* linger requests (it avoids issuing
the same linger request twice to the same osd).
But although the linger requests added to the list in the first loop
may have been sent, they have not yet completed, so they need to be
re-sent regardless of whether their target osd has changed.
The first required fix is we need to avoid calling __map_request()
on any incomplete linger request. Otherwise the subsequent
__map_request() call in the second loop will find the target osd
has not changed and will therefore not re-send the request.
Second, we need to be sure that a sent but incomplete linger request
gets re-sent. If the target osd is the same with the new osd map as
it was when the request was originally sent, this won't happen.
This can be fixed through careful handling when we move these
requests from the request list to the linger list, by unregistering
the request *before* it is registered as a linger request. This
works because a side-effect of unregistering the request is to make
the request's r_osd pointer be NULL, and *that* will ensure the
second loop actually re-sends the linger request.
Processing of such a request is done at that point, so continue with
the next one once it's been moved.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
In kick_requests(), we need to register the request before we
unregister the linger request. Otherwise the unregister will
reset the request's osd pointer to NULL.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
The red-black node in the ceph osd request structure is initialized
in ceph_osdc_alloc_request() using rbd_init_node(). We do need to
initialize this, because in __unregister_request() we call
RB_EMPTY_NODE(), which expects the node it's checking to have
been initialized. But rb_init_node() is apparently overkill, and
may in fact be on its way out. So use RB_CLEAR_NODE() instead.
For a little more background, see this commit:
4c199a93 rbtree: empty nodes have no color"
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
The red-black node node in the ceph osd event structure is not
initialized in create_osdc_create_event(). Because this node can
be the subject of a RB_EMPTY_NODE() call later on, we should ensure
the node is initialized properly for that.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
The red-black node node in the ceph osd structure is not initialized
in create_osd(). Because this node can be the subject of a
RB_EMPTY_NODE() call later on, we should ensure the node is
initialized properly for that. Add a call to RB_CLEAR_NODE()
initialize it.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
In __unregister_linger_request(), the request is being removed
from the osd client's req_linger list only when the request
has a non-null osd pointer. It should be done whether or not
the request currently has an osd.
This is most likely a non-issue because I believe the request
will always have an osd when this function is called.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
If an osd has no requests and no linger requests, __reset_osd()
will just remove it with a call to __remove_osd(). That drops
a reference to the osd, and therefore the osd may have been free
by the time __reset_osd() returns. That function offers no
indication this may have occurred, and as a result the osd will
continue to be used even when it's no longer valid.
Change__reset_osd() so it returns an error (ENODEV) when it
deletes the osd being reset. And change __kick_osd_requests() so it
returns immediately (before referencing osd again) if __reset_osd()
returns *any* error.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
In __unregister_request(), there is a call to list_del_init()
referencing a request that was the subject of a call to
ceph_osdc_put_request() on the previous line. This is not
safe, because the request structure could have been freed
by the time we reach the list_del_init().
Fix this by reversing the order of these lines.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-off-by: Sage Weil <sage@inktank.com>
This would reset a connection with any OSD that had an outstanding
request that was taking more than N seconds. The idea was that if the
OSD was buggy, the client could compensate by resending the request.
In reality, this only served to hide server bugs, and we haven't
actually seen such a bug in quite a while. Moreover, the userspace
client code never did this.
More importantly, often the request is taking a long time because the
OSD is trying to recover, or overloaded, and killing the connection
and retrying would only make the situation worse by giving the OSD
more work to do.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
If we are creating an osd request and get an invalid layout, return
an EINVAL to the caller. We switch up the return to have an error
code instead of NULL implying -ENOMEM.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
If we encounter an invalid (e.g., zeroed) mapping, return an error
and avoid a divide by zero.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
There are many (normal) conditions that can lead to us getting
unexpected replies, include cluster topology changes, osd failures,
and timeouts. There's no need to spam the console about it.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
This is a trivial fix for the debug output, as it is inconsistent
with the function name so may confuse people when debugging.
[elder@inktank.com: switched to use __func__]
Signed-off-by: Jiaju Zhang <jjzhang@suse.de>
Reviewed-by: Alex Elder <elder@inktank.com>
The linger op registration (i.e., watch) modifies the object state. As
such, the OSD will reply with success if it has already applied without
doing the associated side-effects (setting up the watch session state).
If we lose the ACK and resubmit, we will see success but the watch will not
be correctly registered and we won't get notifies.
To fix this, always resubmit the linger op with a new tid. We accomplish
this by re-registering as a linger (i.e., 'registered') if we are not yet
registered. Then the second loop will treat this just like a normal
case of re-registering.
This mirrors a similar fix on the userland ceph.git, commit 5dd68b95, and
ceph bug #2796.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
These don't strictly need to be initialized based on how they are used, but
it is good practice to do so.
Reported-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
Initialize the type field for messages in a msgpool. The caller was doing
this for osd ops, but not for the reply messages.
Reported-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
ceph_con_revoke_message() is passed both a message and a ceph
connection. A ceph_msg allocated for incoming messages on a
connection always has a pointer to that connection, so there's no
need to provide the connection when revoking such a message.
Note that the existing logic does not preclude the message supplied
being a null/bogus message pointer. The only user of this interface
is the OSD client, and the only value an osd client passes is a
request's r_reply field. That is always non-null (except briefly in
an error path in ceph_osdc_alloc_request(), and that drops the
only reference so the request won't ever have a reply to revoke).
So we can safely assume the passed-in message is non-null, but add a
BUG_ON() to make it very obvious we are imposing this restriction.
Rename the function ceph_msg_revoke_incoming() to reflect that it is
really an operation on an incoming message.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
ceph_con_revoke() is passed both a message and a ceph connection.
Now that any message associated with a connection holds a pointer
to that connection, there's no need to provide the connection when
revoking a message.
This has the added benefit of precluding the possibility of the
providing the wrong connection pointer. If the message's connection
pointer is null, it is not being tracked by any connection, so
revoking it is a no-op. This is supported as a convenience for
upper layers, so they can revoke a message that is not actually
"in flight."
Rename the function ceph_msg_revoke() to reflect that it is really
an operation on a message, not a connection.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
The function ceph_alloc_msg() is only used to allocate a message
that will be assigned to a connection's in_msg pointer. Rename the
function so this implied usage is more clear.
In addition, make that assignment inside the function (again, since
that's precisely what it's intended to be used for). This allows us
to return what is now provided via the passed-in address of a "skip"
variable. The return type is now Boolean to be explicit that there
are only two possible outcomes.
Make sure the result of an ->alloc_msg method call always sets the
value of *skip properly.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Move the initialization of a ceph connection's private pointer,
operations vector pointer, and peer name information into
ceph_con_init(). Rearrange the arguments so the connection pointer
is first. Hide the byte-swapping of the peer entity number inside
ceph_con_init()
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
There were a few direct calls to ceph_con_{get,put}() instead of the con
ops from osd_client.c. This is a bug since those ops aren't defined to
be ceph_con_get/put.
This breaks refcounting on the ceph_osd structs that contain the
ceph_connections, and could lead to all manner of strangeness.
The purpose of the ->get and ->put methods in a ceph connection are
to allow the connection to indicate it has a reference to something
external to the messaging system, *not* to indicate something
external has a reference to the connection.
[elder@inktank.com: added that last sentence]
Signed-off-by: Sage Weil <sage@newdream.net>
Reviewed-by: Alex Elder <elder@inktank.com>
In ceph_osdc_release_request(), a reference to the r_reply message
is dropped. But just after that, that same message is revoked if it
was in use to receive an incoming reply. Reorder these so we are
sure we hold a reference until we're actually done with the message.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Pass the osd number to the create_osd() routine, and move the
initialization of fields that depend on it therein.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
A ceph client has a pointer to a ceph messenger structure in it.
There is always exactly one ceph messenger for a ceph client, so
there is no need to allocate it separate from the ceph client
structure.
Switch the ceph_client structure to embed its ceph_messenger
structure.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Pull ceph updates from Sage Weil:
"There are some updates and cleanups to the CRUSH placement code, a bug
fix with incremental maps, several cleanups and fixes from Josh Durgin
in the RBD block device code, a series of cleanups and bug fixes from
Alex Elder in the messenger code, and some miscellaneous bounds
checking and gfp cleanups/fixes."
Fix up trivial conflicts in net/ceph/{messenger.c,osdmap.c} due to the
networking people preferring "unsigned int" over just "unsigned".
* git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client: (45 commits)
libceph: fix pg_temp updates
libceph: avoid unregistering osd request when not registered
ceph: add auth buf in prepare_write_connect()
ceph: rename prepare_connect_authorizer()
ceph: return pointer from prepare_connect_authorizer()
ceph: use info returned by get_authorizer
ceph: have get_authorizer methods return pointers
ceph: ensure auth ops are defined before use
ceph: messenger: reduce args to create_authorizer
ceph: define ceph_auth_handshake type
ceph: messenger: check return from get_authorizer
ceph: messenger: rework prepare_connect_authorizer()
ceph: messenger: check prepare_write_connect() result
ceph: don't set WRITE_PENDING too early
ceph: drop msgr argument from prepare_write_connect()
ceph: messenger: send banner in process_connect()
ceph: messenger: reset connection kvec caller
libceph: don't reset kvec in prepare_write_banner()
ceph: ignore preferred_osd field
ceph: fully initialize new layout
...
There is a race between two __unregister_request() callers: the
reply path and the ceph_osdc_wait_request(). If we get a reply
*and* the timeout expires at roughly the same time, both callers
will try to unregister the request, and the second one will do bad
things.
Simply check if the request is still already unregistered; if so,
return immediately and do nothing.
Fixes http://tracker.newdream.net/issues/2420
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Rather than passing a bunch of arguments to be filled in with the
content of the ceph_auth_handshake buffer now returned by the
get_authorizer method, just use the returned information in the
caller, and drop the unnecessary arguments.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Have the get_authorizer auth_client method return a ceph_auth
pointer rather than an integer, pointer-encoding any returned
error value. This is to pave the way for making use of the
returned value in an upcoming patch.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
In the create_authorizer method for both the mds and osd clients,
the auth_client->ops pointer is blindly dereferenced. There is no
obvious guarantee that this pointer has been assigned. And
furthermore, even if the ops pointer is non-null there is definitely
no guarantee that the create_authorizer or destroy_authorizer
methods are defined.
Add checks in both routines to make sure they are defined (non-null)
before use. Add similar checks in a few other spots in these files
while we're at it.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Make use of the new ceph_auth_handshake structure in order to reduce
the number of arguments passed to the create_authorizor method in
ceph_auth_client_ops. Use a local variable of that type as a
shorthand in the get_authorizer method definitions.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
The definitions for the ceph_mds_session and ceph_osd both contain
five fields related only to "authorizers." Encapsulate those fields
into their own struct type, allowing for better isolation in some
upcoming patches.
Fix the #includes in "linux/ceph/osd_client.h" to lay out their more
complete canonical path.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
From Al Viro <viro@zeniv.linux.org.uk>
Al Viro noticed that we were using a non-cpu-encoded value in
a switch statement in osd_req_encode_op(). The result would
clearly not work correctly on a big-endian machine.
Signed-off-by: Alex Elder <elder@dreamhost.com>
Use of "unsigned int" is preferred to bare "unsigned" in net tree.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
ceph_osd_request struct allocates a 40-byte buffer for object names.
RBD image names can be up to 96 chars long (100 with the .rbd suffix),
which results in the object name for the image being truncated, and a
subsequent map failure.
Increase the oid buffer in request messages, in order to avoid the
truncation.
Signed-off-by: Stratos Psomadakis <psomas@grnet.gr>
Signed-off-by: Sage Weil <sage@newdream.net>
If we skip over one or more map epochs, we need to resend all osd requests
because it is possible they remapped to other servers and then back.
Signed-off-by: Sage Weil <sage@newdream.net>
The pool allocation failures are masked by the pool; there is no need to
spam the console about them. (That's the whole point of having the pool
in the first place.)
Mark msg allocations whose failure is safely handled as such.
Signed-off-by: Sage Weil <sage@newdream.net>
The r_req_lru_item list node moves between several lists, and that cycle
is not directly related (and does not begin) with __register_request().
Initialize it in the request constructor, not __register_request(). This
fixes later badness (below) when OSDs restart underneath an rbd mount.
Crashes we've seen due to this include:
[ 213.974288] kernel BUG at net/ceph/messenger.c:2193!
and
[ 144.035274] BUG: unable to handle kernel NULL pointer dereference at 0000000000000048
[ 144.035278] IP: [<ffffffffa036c053>] con_work+0x1463/0x2ce0 [libceph]
Signed-off-by: Sage Weil <sage@newdream.net>
Keep track of when an outgoing message is ACKed (i.e., the server fully
received it and, presumably, queued it for processing). Time out OSD
requests only if it's been too long since they've been received.
This prevents timeouts and connection thrashing when the OSDs are simply
busy and are throttling the requests they read off the network.
Reviewed-by: Yehuda Sadeh <yehuda@hq.newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
Set the page count correctly for non-page-aligned IO. We were already
doing this correctly for alignment, but not the page count. Fixes
DIRECT_IO writes from unaligned pages.
Signed-off-by: Sage Weil <sage@newdream.net>
If we cancel a write, trigger the safe completions to prevent a sync from
blocking indefinitely in ceph_osdc_sync().
Signed-off-by: Sage Weil <sage@newdream.net>
When the cluster is marked full, subscribe to subsequent map updates to
ensure we find out promptly when it is no longer full. This will prevent
us from spewing ENOSPC for (much) longer than necessary.
Signed-off-by: Sage Weil <sage@newdream.net>
Since we pass the nofail arg, we should never get an error; BUG if we do.
(And fix the function to not return an error if __map_request fails.)
Signed-off-by: Sage Weil <sage@newdream.net>
Fix the request transition from linger -> normal request. The key is to
preserve r_osd and requeue on the same OSD. Reregister as a normal request,
add the request to the proper queues, then unregister the linger. Fix the
unregister helper to avoid clearing r_osd (and also simplify the parallel
check in __unregister_request()).
Reported-by: Henry Chang <henry.cy.chang@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>
We should only clear r_osd if we are neither registered as a linger or a
regular request. We may unregister as a linger while still registered as
a regular request (e.g., in reset_osd). Incorrectly clearing r_osd there
leads to a null pointer dereference in __send_request.
Also simplify the parallel check in __unregister_request() where we just
removed r_osd_item and know it's empty.
Signed-off-by: Sage Weil <sage@newdream.net>
There was a missing unlock on the error path if __map_request() failed.
Signed-off-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>
This patch fixes 'event_work' dereference before it is checked for NULL.
Signed-off-by: Mariusz Kozlowski <mk@lab.zgora.pl>
Signed-off-by: Sage Weil <sage@newdream.net>
Lingering requests are requests that are sent to the OSD normally but
tracked also after we get a successful request. This keeps the OSD
connection open and resends the original request if the object moves to
another OSD. The OSD can then send notification messages back to us
if another client initiates a notify.
This framework will be used by RBD so that the client gets notification
when a snapshot is created by another node or tool.
Signed-off-by: Yehuda Sadeh <yehuda@hq.newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
If we send a request to osd A, and the request's pg remaps to osd B and
then back to A in quick succession, we need to resend the request to A. The
old code was only calling kick_requests after processing all incremental
maps in a message, so it was very possible to not resend a request that
needed to be resent. This would make the osd eventually time out (at least
with the current default of osd timeouts enabled).
The correct approach is to scan requests on every map incremental. This
patch refactors the kick code in a few ways:
- all requests are either on req_lru (in flight), req_unsent (ready to
send), or req_notarget (currently map to no up osd)
- mapping always done by map_request (previous map_osds)
- if the mapping changes, we requeue. requests are resent only after all
map incrementals are processed.
- some osd reset code is moved out of kick_requests into a separate
function
- the "kick this osd" functionality is moved to kick_osd_requests, as it
is unrelated to scanning for request->pg->osd mapping changes
Signed-off-by: Sage Weil <sage@newdream.net>
The alignment used for reading data into or out of pages used to be taken
from the data_off field in the message header. This only worked as long
as the page alignment matched the object offset, breaking direct io to
non-page aligned offsets.
Instead, explicitly specify the page alignment next to the page vector
in the ceph_msg struct, and use that instead of the message header (which
probably shouldn't be trusted). The alloc_msg callback is responsible for
filling in this field properly when it sets up the page vector.
Signed-off-by: Sage Weil <sage@newdream.net>
We used to infer alignment of IOs within a page based on the file offset,
which assumed they matched. This broke with direct IO that was not aligned
to pages (e.g., 512-byte aligned IO). We were also trusting the alignment
specified in the OSD reply, which could have been adjusted by the server.
Explicitly specify the page alignment when setting up OSD IO requests.
Signed-off-by: Sage Weil <sage@newdream.net>
This factors out protocol and low-level storage parts of ceph into a
separate libceph module living in net/ceph and include/linux/ceph. This
is mostly a matter of moving files around. However, a few key pieces
of the interface change as well:
- ceph_client becomes ceph_fs_client and ceph_client, where the latter
captures the mon and osd clients, and the fs_client gets the mds client
and file system specific pieces.
- Mount option parsing and debugfs setup is correspondingly broken into
two pieces.
- The mon client gets a generic handler callback for otherwise unknown
messages (mds map, in this case).
- The basic supported/required feature bits can be expanded (and are by
ceph_fs_client).
No functional change, aside from some subtle error handling cases that got
cleaned up in the refactoring process.
Signed-off-by: Sage Weil <sage@newdream.net>
