Age | Commit message (Collapse) | Author |
|
Avoid crashing if the crypto key payload was NULL, as when it was not correctly
allocated and initialized. Also, avoid leaking it.
Signed-off-by: Sylvain Munaut <tnt@246tNt.com>
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client
Pull Ceph changes from Sage Weil:
"Lots of stuff this time around:
- lots of cleanup and refactoring in the libceph messenger code, and
many hard to hit races and bugs closed as a result.
- lots of cleanup and refactoring in the rbd code from Alex Elder,
mostly in preparation for the layering functionality that will be
coming in 3.7.
- some misc rbd cleanups from Josh Durgin that are finally going
upstream
- support for CRUSH tunables (used by newer clusters to improve the
data placement)
- some cleanup in our use of d_parent that Al brought up a while back
- a random collection of fixes across the tree
There is another patch coming that fixes up our ->atomic_open()
behavior, but I'm going to hammer on it a bit more before sending it."
Fix up conflicts due to commits that were already committed earlier in
drivers/block/rbd.c, net/ceph/{messenger.c, osd_client.c}
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client: (132 commits)
rbd: create rbd_refresh_helper()
rbd: return obj version in __rbd_refresh_header()
rbd: fixes in rbd_header_from_disk()
rbd: always pass ops array to rbd_req_sync_op()
rbd: pass null version pointer in add_snap()
rbd: make rbd_create_rw_ops() return a pointer
rbd: have __rbd_add_snap_dev() return a pointer
libceph: recheck con state after allocating incoming message
libceph: change ceph_con_in_msg_alloc convention to be less weird
libceph: avoid dropping con mutex before fault
libceph: verify state after retaking con lock after dispatch
libceph: revoke mon_client messages on session restart
libceph: fix handling of immediate socket connect failure
ceph: update MAINTAINERS file
libceph: be less chatty about stray replies
libceph: clear all flags on con_close
libceph: clean up con flags
libceph: replace connection state bits with states
libceph: drop unnecessary CLOSED check in socket state change callback
libceph: close socket directly from ceph_con_close()
...
|
|
We drop the lock when calling the ->alloc_msg() con op, which means
we need to (a) not clobber con->in_msg without the mutex held, and (b)
we need to verify that we are still in the OPEN state when we retake
it to avoid causing any mayhem. If the state does change, -EAGAIN
will get us back to con_work() and loop.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
|
|
This function's calling convention is very limiting. In particular,
we can't return any error other than ENOMEM (and only implicitly),
which is a problem (see next patch).
Instead, return an normal 0 or error code, and make the skip a pointer
output parameter. Drop the useless in_hdr argument (we have the con
pointer).
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
|
|
The ceph_fault() function takes the con mutex, so we should avoid
dropping it before calling it. This fixes a potential race with
another thread calling ceph_con_close(), or _open(), or similar (we
don't reverify con->state after retaking the lock).
Add annotation so that lockdep realizes we will drop the mutex before
returning.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
|
|
We drop the con mutex when delivering a message. When we retake the
lock, we need to verify we are still in the OPEN state before
preparing to read the next tag, or else we risk stepping on a
connection that has been closed.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
|
|
Revoke all mon_client messages when we shut down the old connection.
This is mostly moot since we are re-using the same ceph_connection,
but it is cleaner.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
|
|
If the connect() call immediately fails such that sock == NULL, we
still need con_close_socket() to reset our socket state to CLOSED.
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>
|
|
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
Rename flags with CON_FLAG prefix, move the definitions into the c file,
and (better) document their meaning.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
Use a simple set of 6 enumerated values for the socket states (CON_STATE_*)
and use those instead of the state bits. All of the con->state checks are
now under the protection of the con mutex, so this is safe. It also
simplifies many of the state checks because we can check for anything other
than the expected state instead of various bits for races we can think of.
This appears to hold up well to stress testing both with and without socket
failure injection on the server side.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
If we are CLOSED, the socket is closed and we won't get these.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
It is simpler to do this immediately, since we already hold the con mutex.
It also avoids the need to deal with a not-quite-CLOSED socket in con_work.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
If the state is CLOSED or OPENING, we shouldn't have a socket.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
Take the con mutex before checking whether the connection is closed to
avoid racing with someone else closing it.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
Avoid dropping and retaking con->mutex in the ceph_con_send() case by
leaving locking up to the caller.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
If we fault on a lossy connection, we should still close the socket
immediately, and do so under the con mutex.
We should also take the con mutex before printing out the state bits in
the debug output.
Signed-off-by: Sage Weil <sage@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>
|
|
We exponentially back off when we encounter connection errors. If several
errors accumulate, we will eventually wait ages before even trying to
reconnect.
Fix this by resetting the backoff counter after a successful negotiation/
connection with the remote node. Fixes ceph issue #2802.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
|
|
Take the con mutex while we are initiating a ceph open. This is necessary
because the may have previously been in use and then closed, which could
result in a racing workqueue running con_work().
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
|
|
Previously, we were opportunistically initializing the bio_iter if it
appeared to be uninitialized in the middle of the read path. The problem
is that a sequence like:
- start reading message
- initialize bio_iter
- read half a message
- messenger fault, reconnect
- restart reading message
- ** bio_iter now non-NULL, not reinitialized **
- read past end of bio, crash
Instead, initialize the bio_iter unconditionally when we allocate/claim
the message for read.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@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>
|
|
Hold the mutex while twiddling all of the state bits to avoid possible
races. While we're here, make not of why we cannot close the socket
directly.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
|
|
We need to set error_msg to something useful before calling ceph_fault();
do so here for try_{read,write}(). This is more informative than
libceph: osd0 192.168.106.220:6801 (null)
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
|
|
The server side recently added support for tuning some magic
crush variables. Decode these variables if they are present, or use the
default values if they are not present.
Corresponds to ceph.git commit 89af369c25f274fe62ef730e5e8aad0c54f1e5a5.
Signed-off-by: caleb miles <caleb.miles@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
|
|
This is simply cleanup that will keep things more closely synced with the
userland code.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
|
|
Add an atomic variable 'stopping' as flag in struct ceph_messenger,
set this flag to 1 in function ceph_destroy_client(), and add the condition code
in function ceph_data_ready() to test the flag value, if true(1), just return.
Signed-off-by: Guanjun He <gjhe@suse.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
In ancient times, the messenger could both initiate and accept connections.
An artifact if that was data structures to store/process an incoming
ceph_msg_connect request and send an outgoing ceph_msg_connect_reply.
Sadly, the negotiation code was referencing those structures and ignoring
important information (like the peer's connect_seq) from the correct ones.
Among other things, this fixes tight reconnect loops where the server sends
RETRY_SESSION and we (the client) retries with the same connect_seq as last
time. This bug pretty easily triggered by injecting socket failures on the
MDS and running some fs workload like workunits/direct_io/test_sync_io.
Signed-off-by: Sage Weil <sage@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>
|
|
Pull networking changes from David S Miller:
1) Remove the ipv4 routing cache. Now lookups go directly into the FIB
trie and use prebuilt routes cached there.
No more garbage collection, no more rDOS attacks on the routing
cache. Instead we now get predictable and consistent performance,
no matter what the pattern of traffic we service.
This has been almost 2 years in the making. Special thanks to
Julian Anastasov, Eric Dumazet, Steffen Klassert, and others who
have helped along the way.
I'm sure that with a change of this magnitude there will be some
kind of fallout, but such things ought the be simple to fix at this
point. Luckily I'm not European so I'll be around all of August to
fix things :-)
The major stages of this work here are each fronted by a forced
merge commit whose commit message contains a top-level description
of the motivations and implementation issues.
2) Pre-demux of established ipv4 TCP sockets, saves a route demux on
input.
3) TCP SYN/ACK performance tweaks from Eric Dumazet.
4) Add namespace support for netfilter L4 conntrack helpers, from Gao
Feng.
5) Add config mechanism for Energy Efficient Ethernet to ethtool, from
Yuval Mintz.
6) Remove quadratic behavior from /proc/net/unix, from Eric Dumazet.
7) Support for connection tracker helpers in userspace, from Pablo
Neira Ayuso.
8) Allow userspace driven TX load balancing functions in TEAM driver,
from Jiri Pirko.
9) Kill off NLMSG_PUT and RTA_PUT macros, more gross stuff with
embedded gotos.
10) TCP Small Queues, essentially minimize the amount of TCP data queued
up in the packet scheduler layer. Whereas the existing BQL (Byte
Queue Limits) limits the pkt_sched --> netdevice queuing levels,
this controls the TCP --> pkt_sched queueing levels.
From Eric Dumazet.
11) Reduce the number of get_page/put_page ops done on SKB fragments,
from Alexander Duyck.
12) Implement protection against blind resets in TCP (RFC 5961), from
Eric Dumazet.
13) Support the client side of TCP Fast Open, basically the ability to
send data in the SYN exchange, from Yuchung Cheng.
Basically, the sender queues up data with a sendmsg() call using
MSG_FASTOPEN, then they do the connect() which emits the queued up
fastopen data.
14) Avoid all the problems we get into in TCP when timers or PMTU events
hit a locked socket. The TCP Small Queues changes added a
tcp_release_cb() that allows us to queue work up to the
release_sock() caller, and that's what we use here too. From Eric
Dumazet.
15) Zero copy on TX support for TUN driver, from Michael S. Tsirkin.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1870 commits)
genetlink: define lockdep_genl_is_held() when CONFIG_LOCKDEP
r8169: revert "add byte queue limit support".
ipv4: Change rt->rt_iif encoding.
net: Make skb->skb_iif always track skb->dev
ipv4: Prepare for change of rt->rt_iif encoding.
ipv4: Remove all RTCF_DIRECTSRC handliing.
ipv4: Really ignore ICMP address requests/replies.
decnet: Don't set RTCF_DIRECTSRC.
net/ipv4/ip_vti.c: Fix __rcu warnings detected by sparse.
ipv4: Remove redundant assignment
rds: set correct msg_namelen
openvswitch: potential NULL deref in sample()
tcp: dont drop MTU reduction indications
bnx2x: Add new 57840 device IDs
tcp: avoid oops in tcp_metrics and reset tcpm_stamp
niu: Change niu_rbr_fill() to use unlikely() to check niu_rbr_add_page() return value
niu: Fix to check for dma mapping errors.
net: Fix references to out-of-scope variables in put_cmsg_compat()
net: ethernet: davinci_emac: add pm_runtime support
net: ethernet: davinci_emac: Remove unnecessary #include
...
|
|
In ancient times, the messenger could both initiate and accept connections.
An artifact if that was data structures to store/process an incoming
ceph_msg_connect request and send an outgoing ceph_msg_connect_reply.
Sadly, the negotiation code was referencing those structures and ignoring
important information (like the peer's connect_seq) from the correct ones.
Among other things, this fixes tight reconnect loops where the server sends
RETRY_SESSION and we (the client) retries with the same connect_seq as last
time. This bug pretty easily triggered by injecting socket failures on the
MDS and running some fs workload like workunits/direct_io/test_sync_io.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
It is possible to close a socket that is in the OPENING state. For
example, it can happen if ceph_con_close() is called on the con before
the TCP connection is established. con_work() will come around and shut
down the socket.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
Do not re-initialize the con on every connection attempt. When we
ceph_con_close, there may still be work queued on the socket (e.g., to
close it), and re-initializing will clobber the work_struct state.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
The peer name may change on each open attempt, even when the connection is
reused.
Signed-off-by: Sage Weil <sage@inktank.com>
|
|
Sage liked the state diagram I put in my commit description so
I'm putting it in with the code.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
This patch gathers a few small changes in "net/ceph/messenger.c":
out_msg_pos_next()
- small logic change that mostly affects indentation
write_partial_msg_pages().
- use a local variable trail_off to represent the offset into
a message of the trail portion of the data (if present)
- once we are in the trail portion we will always be there, so we
don't always need to check against our data position
- avoid computing len twice after we've reached the trail
- get rid of the variable tmpcrc, which is not needed
- trail_off and trail_len never change so mark them const
- update some comments
read_partial_message_bio()
- bio_iovec_idx() will never return an error, so don't bother
checking for it
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
Currently a ceph connection enters a "CONNECTING" state when it
begins the process of (re-)connecting with its peer. Once the two
ends have successfully exchanged their banner and addresses, an
additional NEGOTIATING bit is set in the ceph connection's state to
indicate the connection information exhange has begun. The
CONNECTING bit/state continues to be set during this phase.
Rather than have the CONNECTING state continue while the NEGOTIATING
bit is set, interpret these two phases as distinct states. In other
words, when NEGOTIATING is set, clear CONNECTING. That way only
one of them will be active at a time.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
There are two phases in the process of linking together the two ends
of a ceph connection. The first involves exchanging a banner and
IP addresses, and if that is successful a second phase exchanges
some detail about each side's connection capabilities.
When initiating a connection, the client side now queues to send
its information for both phases of this process at the same time.
This is probably a bit more efficient, but it is slightly messier
from a layering perspective in the code.
So rearrange things so that the client doesn't send the connection
information until it has received and processed the response in the
initial banner phase (in process_banner()).
Move the code (in the (con->sock == NULL) case in try_write()) that
prepares for writing the connection information, delaying doing that
until the banner exchange has completed. Move the code that begins
the transition to this second "NEGOTIATING" phase out of
process_banner() and into its caller, so preparing to write the
connection information and preparing to read the response are
adjacent to each other.
Finally, preparing to write the connection information now requires
the output kvec to be reset in all cases, so move that into the
prepare_write_connect() and delete it from all callers.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
There is no state explicitly defined when a ceph connection is fully
operational. So define one.
It's set when the connection sequence completes successfully, and is
cleared when the connection gets closed.
Be a little more careful when examining the old state when a socket
disconnect event is reported.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
A connection state's NEGOTIATING bit gets set while in CONNECTING
state after we have successfully exchanged a ceph banner and IP
addresses with the connection's peer (the server). But that bit
is not cleared again--at least not until another connection attempt
is initiated.
Instead, clear it as soon as the connection is fully established.
Also, clear it when a socket connection gets prematurely closed
in the midst of establishing a ceph connection (in case we had
reached the point where it was set).
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
A connection that is closed will no longer be connecting. So
clear the CONNECTING state bit in ceph_con_close(). Similarly,
if the socket has been closed we no longer are in connecting
state (a new connect sequence will need to be initiated).
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
In con_close_socket(), a connection's SOCK_CLOSED flag gets set and
then cleared while its shutdown method is called and its reference
gets dropped.
Previously, that flag got set only if it had not already been set,
so setting it in con_close_socket() might have prevented additional
processing being done on a socket being shut down. We no longer set
SOCK_CLOSED in the socket event routine conditionally, so setting
that bit here no longer provides whatever benefit it might have
provided before.
A race condition could still leave the SOCK_CLOSED bit set even
after we've issued the call to con_close_socket(), so we still clear
that bit after shutting the socket down. Add a comment explaining
the reason for this.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
When a TCP_CLOSE or TCP_CLOSE_WAIT event occurs, the SOCK_CLOSED
connection flag bit is set, and if it had not been previously set
queue_con() is called to ensure con_work() will get a chance to
handle the changed state.
con_work() atomically checks--and if set, clears--the SOCK_CLOSED
bit if it was set. This means that even if the bit were set
repeatedly, the related processing in con_work() only gets called
once per transition of the bit from 0 to 1.
What's important then is that we ensure con_work() gets called *at
least* once when a socket close event occurs, not that it gets
called *exactly* once.
The work queue mechanism already takes care of queueing work
only if it is not already queued, so there's no need for us
to call queue_con() conditionally.
So this patch just makes it so the SOCK_CLOSED flag gets set
unconditionally in ceph_sock_state_change().
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
Currently the socket state change event handler records an error
message on a connection to distinguish a close while connecting from
a close while a connection was already established.
Changing connection information during handling of a socket event is
not very clean, so instead move this assignment inside con_work(),
where it can be done during normal connection-level processing (and
under protection of the connection mutex as well).
Move the handling of a socket closed event up to the top of the
processing loop in con_work(); there's no point in handling backoff
etc. if we have a newly-closed socket to take care of.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
The following commit changed it so SOCK_CLOSED bit was stored in
a connection's new "flags" field rather than its "state" field.
libceph: start separating connection flags from state
commit 928443cd
That bit is used in con_close_socket() to protect against setting an
error message more than once in the socket event handler function.
Unfortunately, the field being operated on in that function was not
updated to be "flags" as it should have been. This fixes that
error.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
Recently a bug was fixed in which the bio_iter field in a ceph
message was not being properly re-initialized when a message got
re-transmitted:
commit 43643528cce60ca184fe8197efa8e8da7c89a037
Author: Yan, Zheng <zheng.z.yan@intel.com>
rbd: Clear ceph_msg->bio_iter for retransmitted message
We are now only initializing the bio_iter field when we are about to
start to write message data (in prepare_write_message_data()),
rather than every time we are attempting to write any portion of the
message data (in write_partial_msg_pages()). This means we no
longer need to use the msg->bio_iter field as a flag.
So just don't do that any more. Trust prepare_write_message_data()
to ensure msg->bio_iter is properly initialized, every time we are
about to begin writing (or re-writing) a message's bio data.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|
|
If a message has a non-null bio pointer, its bio_iter field is
initialized in write_partial_msg_pages() if this has not been done
already. This is really a one-time setup operation for sending a
message's (bio) data, so move that initialization code into
prepare_write_message_data() which serves that purpose.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
|