implementation.
The kernel RPC code, which is responsible for the low-level scheduling
of incoming NFS requests, contains a throttling mechanism that
prevents too much kernel memory from being tied up by NFS requests
that are being serviced. When the throttle is engaged, the RPC layer
stops servicing incoming NFS sockets, resulting ultimately in
backpressure on the clients (if they're using TCP). However, this is
a very heavy-handed mechanism as it prevents all clients from making
any requests, regardless of how heavy or light they are. (Thus, when
engaged, the throttle often prevents clients from even mounting the
filesystem.) The throttle mechanism applies specifically to requests
that have been received by the RPC layer (from a TCP or UDP socket)
and are queued waiting to be serviced by one of the nfsd threads; it
does not limit the amount of backlog in the socket buffers.
The original implementation limited the total bytes of queued requests
to the minimum of a quarter of (nmbclusters * MCLBYTES) and 45 MiB.
The former limit seems reasonable, since requests queued in the socket
buffers and replies being constructed to the requests in progress will
all require some amount of network memory, but the 45 MiB limit is
plainly ridiculous for modern memory sizes: when running 256 service
threads on a busy server, 45 MiB would result in just a single
maximum-sized NFS3PROC_WRITE queued per thread before throttling.
Removing this limit exposed integer-overflow bugs in the original
computation, and related bugs in the routines that actually account
for the amount of traffic enqueued for service threads. The old
implementation also attempted to reduce accounting overhead by
batching updates until each queue is fully drained, but this is prone
to livelock, resulting in repeated accumulate-throttle-drain cycles on
a busy server. Various data types are changed to long or unsigned
long; explicit 64-bit types are not used due to the unavailability of
64-bit atomics on many 32-bit platforms, but those platforms also
cannot support nmbclusters large enough to cause overflow.
This code (in a 10.1 kernel) is presently running on production NFS
servers at CSAIL.
Summary of this revision:
* Removes 45 MiB limit on requests queued for nfsd service threads
* Fixes integer-overflow and signedness bugs
* Avoids unnecessary throttling by not deferring accounting for
completed requests
Differential Revision: https://reviews.freebsd.org/D2165
Reviewed by: rmacklem, mav
MFC after: 30 days
Relnotes: yes
Sponsored by: MIT Computer Science & Artificial Intelligence Laboratory
feature is to quisce the system before suspend.
Stop is implemented by reusing the thread_single(9) with the special
mode SINGLE_ALLPROC. SINGLE_ALLPROC differs from the existing
single-threading modes by allowing (requiring) caller to operate on
other process. Interruptible sleeps for !TDF_SBDRY threads are
suspended like SIGSTOP does it, instead of aborting the sleep, like
SINGLE_NO_EXIT, to avoid spurious EINTRs on resume.
Provide debugging sysctl debug.stop_all_proc, which causes total stop
and suspends syncer, while waiting for variable reset for resume. It
is used for debugging; should be removed after the real use of the
interface is added.
In collaboration with: pho
Discussed with: avg
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks
This is not correct at least for the stop requests. Check for stop
conditions and suspend threads if requested.
Reported and tested by: pho
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
sb_cc member of struct sockbuf to a couple of inline functions:
sbavail() and sbused()
Right now they are equal, but once notion of "not ready socket buffer data",
will be checked in, they are going to be different.
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
into head. The code is not believed to have any effect
on the semantics of non-NFSv4.1 server behaviour.
It is a rather large merge, but I am hoping that there will
not be any regressions for the NFS server.
MFC after: 1 month
Old design with unified thread pool was good from the point of thread
utilization. But single pool-wide mutex became huge congestion point
for systems with many CPUs. To reduce the congestion create several
thread groups within a pool (one group for every 6 CPUs and 12 threads),
each group with own mutex. Each connection during its registration is
assigned to one of the groups in round-robin fashion. File affinify
code may still move requests between the groups, but otherwise groups
are self-contained.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
This allows to slightly simplify svc_run_internal() code: if we processed
all the requests in a queue, then we know that new one will not appear.
MFC after: 2 weeks
New algorithm does not create additional lock congestion, while some races
it includes should not be a problem. Those races may keep requests in DRC
cache for some more time by returning ACK position smaller then actual,
but it still should be able to drop thems when proper ACK finally read.
Races of the original algorithm based on TCP seq number were worse because
they happened when reply sequence number were recorded. After that even
correctly read ACKs could not clean DRC sometimes.
- Introduce additional hash to group requests by hash of sockref. This
allows to process TCP acknowledgements without looping though all the cache,
and as result allows to do it every time.
- Indroduce additional callbacks to notify application layer about sockets
disconnection. Without this last few requests processed just before socket
disconnection never processed their ACKs and stuck in cache for many hours.
- Implement transport-specific method for tracking reply acknowledgements.
New implementation does not cross multiple stack layers to get the data and
does not have race conditions that previously made some requests stuck
in cache. This could be done more efficiently at sockbuf layer, but that
would broke some KBIs, while I don't know other consumers for it aside NFS.
- Instead of traversing all DRC twice per request, run cleaning only once
per request, and except in some conditions traverse only single hash slot
at a time.
Together this limits NFS DRC growth only to situations of real connectivity
problems. If network is working well, and so all replies are acknowledged,
cache remains almost empty even after hours of heavy load. Without this
change on the same test cache was growing to many thousand requests even
with perfectly working local network.
As another result this reduces CPU time spent on the DRC handling during
SPEC NFS benchmark from about 10% to 0.5%.
Sponsored by: iXsystems, Inc.
Do not insert active ports into pool->sp_active list if they are success-
fully assigned to some thread. This makes that list include only ports that
really require attention, and so traversal can be reduced to simple taking
the first one.
Remove idle thread from pool->sp_idlethreads list when assigning some
work (port of requests) to it. That again makes possible to replace list
traversals with simple taking the first element.
When processing receive buffer, write the amount of data, expected
in present request record, into socket's so_rcv.sb_lowat to make stack
aware about our needs. When processing following upcalls, ignore them
until socket collect enough data to be read and processed in one turn.
This change reduces number of context switches and other operations
in RPC stack during large NFS writes (especially via non-Jumbo networks)
by order of magnitude.
After precessing current packet, take another look into the pending
buffer to find out whether the next packet had been already received.
If not, deactivate this port right there without making RPC code to
push this port to another thread just to find that there is nothing.
If the next packet is received partially, also deactivate the port, but
also update socket's so_rcv.sb_lowat to not be woken up prematurely.
This change additionally reduces number of context switches per NFS
request about in half.
3-clause BSD license as specified by Oracle America, Inc. in 2010.
This license change was approved by Wim Coekaerts, Senior Vice
President, Linux and Virtualization at Oracle Corporation.
- close cosmetic race in svc_exit();
- do not set wait timeout for idle threads if we have no use for wakeups;
- create new requested thread sooner, not only after some another thread
wakeup, that may happen later under constant load.
krpc client side UDP was observed as way out of range and
caused the rpc.lockd daemon to hang trying to do an RPC.
Inspection of the code found two places where the RPC request
is re-queued, but the value of cu_sent was not incremented.
Since cu_sent is always decremented when the RPC request is
dequeued, I think this could have caused cu_sent to go out of
range. This patch adds lines to increment cu_sent for these
two cases.
Reported by: dwhite@ixsystems.com
Discussed with: dwhite@ixsystems.com
MFC after: 2 weeks
credentials to the kernel rpc. Modify the NFSv4 client to add
support for the gssname and allgssname mount options to use this
capability. Requires the gssd daemon to be running with the "-h" option.
Reviewed by: jhb
connection after it was accepted by the userland nfsd process but before
it was handled off to svc_vc_create() in the kernel, then svc_vc_create()
would see it as a new listen socket and try to listen on it leaving a
dangling reference to the socket. Instead, check for disconnected sockets
and treat them like a connected socket. The call to pru_getaddr() should
fail and cause svc_vc_create() to fail. Note that we need to lock the
socket to get a consistent snapshot of so_state since there is a window
in soisdisconnected() where both flags are clear.
Reviewed by: dfr, rmacklem
MFC after: 1 week
are used by NFSv4.1 for callbacks. A backchannel is a connection
established by the client, but used for RPCs done by the server
on the client (callbacks). As a result, this patch mixes some
client side calls in the server side and vice versa. Some
definitions in the .c files were extracted out into a file called
krpc.h, so that they could be included in multiple .c files.
This code has been in projects/nfsv4.1-client for some time.
Although no one has given it a formal review, I believe kib@
has taken a look at it.
After further discussion, instead of pretending to use
uid_t and gid_t as upstream Solaris and linux try to, we
are better using u_int, which is in fact what the code
can handle and best approaches the range of values used
by uid and gid.
Discussed with: bde
Reviewed by: bde
When creating a client with clnt_tli_create, it uses strdup to copy
strings for these fields if nconf is passed in. clnt_dg_destroy frees
these strings already. Make sure clnt_vc_destroy frees them in the same
way.
This change matches the reference (OpenSolaris) implementation.
Tested by: David Wolfskill
Obtained from: Bull GNU/Linux NFSv4 Project (libtirpc)
MFC after: 2 weeks
w.r.t. a Linux NFS client doing a krb5 NFS mount against the
FreeBSD server. We determined this was a Linux bug:
http://www.spinics.net/lists/linux-nfs/msg32466.html, however
the mount failed to work, because the Destroy operation with a
bogus encrypted checksum destroyed the authenticator handle.
This patch changes the rpcsec_gss code so that it doesn't
Destroy the authenticator handle for this case and, as such,
the Linux mount will work.
Tested by: Attila Bogar and Herbert Poeckl
MFC after: 2 weeks
The attempt to merge changes from the linux libtirpc caused
rpc.lockd to exit after startup under unclear conditions.
After many hours of selective experiments and inconsistent results
the conclusion is that it's better to just revert everything and
restart in a future time with a much smaller subset of the
changes.
____
MFC after: 3 days
Reported by: David Wolfskill
Tested by: David Wolfskill
The following change caused rpc.lockd to exit after startup:
____
libtirpc: be sure to free cl_netid and cl_tp
When creating a client with clnt_tli_create, it uses strdup to copy
strings for these fields if nconf is passed in. clnt_dg_destroy frees
these strings already. Make sure clnt_vc_destroy frees them in the
same way.
____
MFC after: 3 days
Reported by: David Wolfskill
Tested by: David Wolfskill
C++ mangling will cause trouble with variables like __rpc_xdr
in xdr.h so rename this to XDR.
While here add proper C++ guards to RPC headers.
PR: 137443
MFC after: 2 weeks
We especifically ignored the glibc compatibility changes
but this should help interaction with Solaris and Linux.
____
Fixed infinite loop in svc_run()
author Steve Dickson
Tue, 10 Jun 2008 12:35:52 -0500 (13:35 -0400)
Fixed infinite loop in svc_run()
____
__rpc_taddr2uaddr_af() assumes the netbuf to always have a
non-zero data. This is a bad assumption and can lead to a
seg-fault. This patch adds a check for zero length and returns
NULL when found.
author Steve Dickson
Mon, 27 Oct 2008 11:46:54 -0500 (12:46 -0400)
____
Changed clnt_spcreateerror() to return clearer
and more concise error messages.
author Steve Dickson
Thu, 20 Nov 2008 08:55:31 -0500 (08:55 -0500)
____
Converted all uid and gid variables of the type uid_t and gid_t.
author Steve Dickson
Wed, 28 Jan 2009 12:44:46 -0500 (12:44 -0500)
____
libtirpc: set r_netid and r_owner in __rpcb_findaddr_timed
These fields in the rpcbind GETADDR call are being passed uninitialized
to CLNT_CALL. In the case of x86_64 at least, this usually leads to a
segfault. On x86, it sometimes causes segfaults and other times causes
garbage to be sent on the wire.
rpcbind generally ignores the r_owner field for calls that come in over
the wire, so it really doesn't matter what we send in that slot. We just
need to send something. The reference implementation from Sun seems to
send a blank string. Have ours follow suit.
author Jeff Layton
Fri, 13 Mar 2009 11:44:16 -0500 (12:44 -0400)
____
libtirpc: be sure to free cl_netid and cl_tp
When creating a client with clnt_tli_create, it uses strdup to copy
strings for these fields if nconf is passed in. clnt_dg_destroy frees
these strings already. Make sure clnt_vc_destroy frees them in the same
way.
author Jeff Layton
Fri, 13 Mar 2009 11:47:36 -0500 (12:47 -0400)
Obtained from: Bull GNU/Linux NFSv4 Project
MFC after: 3 weeks
subject heading "mtx_lock() of destroyed mutex on NFS" and
PR# 156168 appear to be caused by clnt_dg_destroy() closing
down the socket prematurely. When to close down the socket
is controlled by a reference count (cs_refs), but clnt_dg_create()
checks for sb_upcall being non-NULL to decide if a new socket
is needed. I believe the crashes were caused by the following race:
clnt_dg_destroy() finds cs_refs == 0 and decides to delete socket
clnt_dg_destroy() then loses race with clnt_dg_create() for
acquisition of the SOCKBUF_LOCK()
clnt_dg_create() finds sb_upcall != NULL and increments cs_refs to 1
clnt_dg_destroy() then acquires SOCKBUF_LOCK(), sets sb_upcall to
NULL and destroys socket
This patch fixes the above race by changing clnt_dg_destroy() so
that it acquires SOCKBUF_LOCK() before testing cs_refs.
Tested by: bz
PR: 156168
Reviewed by: dfr
MFC after: 2 weeks
