d/freebsd-dev
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
6103bae6ae
freebsd-dev/sys/fs/nfsserver/nfs_nfsdcache.c

1035 lines
28 KiB
C
Raw Normal View History

Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
/*-
* Copyright (c) 1989, 1993
* The Regents of the University of California. All rights reserved.
*
* This code is derived from software contributed to Berkeley by
* Rick Macklem at The University of Guelph.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 4. Neither the name of the University nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
/*
* Here is the basic algorithm:
* First, some design criteria I used:
* - I think a false hit is more serious than a false miss
* - A false hit for an RPC that has Op(s) that order via seqid# must be
* avoided at all cost
* - A valid hit will probably happen a long time after the original reply
* and the TCP socket that the original request was received on will no
* longer be active
* (The long time delay implies to me that LRU is not appropriate.)
* - The mechanism will satisfy the requirements of ordering Ops with seqid#s
* in them as well as minimizing the risk of redoing retried non-idempotent
* Ops.
* Because it is biased towards avoiding false hits, multiple entries with
* the same xid are to be expected, especially for the case of the entry
* in the cache being related to a seqid# sequenced Op.
*
* The basic algorithm I'm about to code up:
* - Null RPCs bypass the cache and are just done
* For TCP
* - key on <xid, NFS version> (as noted above, there can be several
* entries with the same key)
* When a request arrives:
* For all that match key
* - if RPC# != OR request_size !=
* - not a match with this one
* - if NFSv4 and received on same TCP socket OR
* received on a TCP connection created before the
* entry was cached
* - not a match with this one
* (V2,3 clients might retry on same TCP socket)
* - calculate checksum on first N bytes of NFS XDR
* - if checksum !=
* - not a match for this one
* If any of the remaining ones that match has a
* seqid_refcnt > 0
* - not a match (go do RPC, using new cache entry)
* If one match left
* - a hit (reply from cache)
* else
* - miss (go do RPC, using new cache entry)
*
* During processing of NFSv4 request:
* - set a flag when a non-idempotent Op is processed
* - when an Op that uses a seqid# (Open,...) is processed
* - if same seqid# as referenced entry in cache
* - free new cache entry
* - reply from referenced cache entry
* else if next seqid# in order
* - free referenced cache entry
* - increment seqid_refcnt on new cache entry
* - set pointer from Openowner/Lockowner to
* new cache entry (aka reference it)
* else if first seqid# in sequence
* - increment seqid_refcnt on new cache entry
* - set pointer from Openowner/Lockowner to
* new cache entry (aka reference it)
*
* At end of RPC processing:
* - if seqid_refcnt > 0 OR flagged non-idempotent on new
* cache entry
* - save reply in cache entry
* - calculate checksum on first N bytes of NFS XDR
* request
* - note op and length of XDR request (in bytes)
* - timestamp it
* else
* - free new cache entry
* - Send reply (noting info for socket activity check, below)
*
* For cache entries saved above:
* - if saved since seqid_refcnt was > 0
* - free when seqid_refcnt decrements to 0
* (when next one in sequence is processed above, or
* when Openowner/Lockowner is discarded)
* else { non-idempotent Op(s) }
* - free when
* - some further activity observed on same
* socket
* (I'm not yet sure how I'm going to do
* this. Maybe look at the TCP connection
* to see if the send_tcp_sequence# is well
* past sent reply OR K additional RPCs
* replied on same socket OR?)
* OR
* - when very old (hours, days, weeks?)
*
* For UDP (v2, 3 only), pretty much the old way:
* - key on <xid, NFS version, RPC#, Client host ip#>
* (at most one entry for each key)
*
* When a Request arrives:
* - if a match with entry via key
* - if RPC marked In_progress
* - discard request (don't send reply)
* else
* - reply from cache
* - timestamp cache entry
* else
* - add entry to cache, marked In_progress
* - do RPC
* - when RPC done
* - if RPC# non-idempotent
* - mark entry Done (not In_progress)
* - save reply
* - timestamp cache entry
* else
* - free cache entry
* - send reply
*
* Later, entries with saved replies are free'd a short time (few minutes)
* after reply sent (timestamp).
* Reference: Chet Juszczak, "Improving the Performance and Correctness
* of an NFS Server", in Proc. Winter 1989 USENIX Conference,
* pages 53-63. San Diego, February 1989.
* for the UDP case.
* nfsrc_floodlevel is set to the allowable upper limit for saved replies
* for TCP. For V3, a reply won't be saved when the flood level is
* hit. For V4, the non-idempotent Op will return NFSERR_RESOURCE in
* that case. This level should be set high enough that this almost
* never happens.
*/
#ifndef APPLEKEXT
#include <fs/nfs/nfsport.h>
extern struct nfsstats newnfsstats;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
extern struct mtx nfsrc_udpmtx;
extern struct nfsrchash_bucket nfsrchash_table[NFSRVCACHE_HASHSIZE];
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
extern struct nfsrchash_bucket nfsrcahash_table[NFSRVCACHE_HASHSIZE];
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
int nfsrc_floodlevel = NFSRVCACHE_FLOODLEVEL, nfsrc_tcpsavedreplies = 0;
#endif /* !APPLEKEXT */
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
SYSCTL_DECL(_vfs_nfsd);
static u_int nfsrc_tcphighwater = 0;
static int
sysctl_tcphighwater(SYSCTL_HANDLER_ARGS)
{
int error, newhighwater;
newhighwater = nfsrc_tcphighwater;
error = sysctl_handle_int(oidp, &newhighwater, 0, req);
if (error != 0 || req->newptr == NULL)
return (error);
if (newhighwater < 0)
return (EINVAL);
if (newhighwater >= nfsrc_floodlevel)
nfsrc_floodlevel = newhighwater + newhighwater / 5;
nfsrc_tcphighwater = newhighwater;
return (0);
}
SYSCTL_PROC(_vfs_nfsd, OID_AUTO, tcphighwater, CTLTYPE_UINT | CTLFLAG_RW, 0,
sizeof(nfsrc_tcphighwater), sysctl_tcphighwater, "IU",
"High water mark for TCP cache entries");
static u_int nfsrc_udphighwater = NFSRVCACHE_UDPHIGHWATER;
SYSCTL_UINT(_vfs_nfsd, OID_AUTO, udphighwater, CTLFLAG_RW,
&nfsrc_udphighwater, 0,
"High water mark for UDP cache entries");
static u_int nfsrc_tcptimeout = NFSRVCACHE_TCPTIMEOUT;
SYSCTL_UINT(_vfs_nfsd, OID_AUTO, tcpcachetimeo, CTLFLAG_RW,
&nfsrc_tcptimeout, 0,
"Timeout for TCP entries in the DRC");
static u_int nfsrc_tcpnonidempotent = 1;
SYSCTL_UINT(_vfs_nfsd, OID_AUTO, cachetcp, CTLFLAG_RW,
&nfsrc_tcpnonidempotent, 0,
"Enable the DRC for NFS over TCP");
static int nfsrc_udpcachesize = 0;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
static TAILQ_HEAD(, nfsrvcache) nfsrvudplru;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
static struct nfsrvhashhead nfsrvudphashtbl[NFSRVCACHE_HASHSIZE];
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
/*
* and the reverse mapping from generic to Version 2 procedure numbers
*/
static int newnfsv2_procid[NFS_V3NPROCS] = {
NFSV2PROC_NULL,
NFSV2PROC_GETATTR,
NFSV2PROC_SETATTR,
NFSV2PROC_LOOKUP,
NFSV2PROC_NOOP,
NFSV2PROC_READLINK,
NFSV2PROC_READ,
NFSV2PROC_WRITE,
NFSV2PROC_CREATE,
NFSV2PROC_MKDIR,
NFSV2PROC_SYMLINK,
NFSV2PROC_CREATE,
NFSV2PROC_REMOVE,
NFSV2PROC_RMDIR,
NFSV2PROC_RENAME,
NFSV2PROC_LINK,
NFSV2PROC_READDIR,
NFSV2PROC_NOOP,
NFSV2PROC_STATFS,
NFSV2PROC_NOOP,
NFSV2PROC_NOOP,
NFSV2PROC_NOOP,
};
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
#define nfsrc_hash(xid) (((xid) + ((xid) >> 24)) % NFSRVCACHE_HASHSIZE)
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
#define NFSRCUDPHASH(xid) \
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
(&nfsrvudphashtbl[nfsrc_hash(xid)])
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
#define NFSRCHASH(xid) \
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
(&nfsrchash_table[nfsrc_hash(xid)].tbl)
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
#define NFSRCAHASH(xid) (&nfsrcahash_table[nfsrc_hash(xid)])
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
#define TRUE 1
#define FALSE 0
#define NFSRVCACHE_CHECKLEN 100
/* True iff the rpc reply is an nfs status ONLY! */
static int nfsv2_repstat[NFS_V3NPROCS] = {
FALSE,
FALSE,
FALSE,
FALSE,
FALSE,
FALSE,
FALSE,
FALSE,
FALSE,
FALSE,
TRUE,
TRUE,
TRUE,
TRUE,
FALSE,
TRUE,
FALSE,
FALSE,
FALSE,
FALSE,
FALSE,
FALSE,
};
/*
* Will NFS want to work over IPv6 someday?
*/
#define NETFAMILY(rp) \
(((rp)->rc_flag & RC_INETIPV6) ? AF_INET6 : AF_INET)
/* local functions */
static int nfsrc_getudp(struct nfsrv_descript *nd, struct nfsrvcache *newrp);
static int nfsrc_gettcp(struct nfsrv_descript *nd, struct nfsrvcache *newrp);
static void nfsrc_lock(struct nfsrvcache *rp);
static void nfsrc_unlock(struct nfsrvcache *rp);
static void nfsrc_wanted(struct nfsrvcache *rp);
static void nfsrc_freecache(struct nfsrvcache *rp);
static int nfsrc_getlenandcksum(mbuf_t m1, u_int16_t *cksum);
static void nfsrc_marksametcpconn(u_int64_t);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
/*
* Return the correct mutex for this cache entry.
*/
static __inline struct mtx *
nfsrc_cachemutex(struct nfsrvcache *rp)
{
if ((rp->rc_flag & RC_UDP) != 0)
return (&nfsrc_udpmtx);
return (&nfsrchash_table[nfsrc_hash(rp->rc_xid)].mtx);
}
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
/*
* Initialize the server request cache list
*/
APPLESTATIC void
nfsrvd_initcache(void)
{
int i;
static int inited = 0;
if (inited)
return;
inited = 1;
for (i = 0; i < NFSRVCACHE_HASHSIZE; i++) {
LIST_INIT(&nfsrvudphashtbl[i]);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
LIST_INIT(&nfsrchash_table[i].tbl);
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
LIST_INIT(&nfsrcahash_table[i].tbl);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
TAILQ_INIT(&nfsrvudplru);
nfsrc_tcpsavedreplies = 0;
nfsrc_udpcachesize = 0;
newnfsstats.srvcache_tcppeak = 0;
newnfsstats.srvcache_size = 0;
}
/*
* Get a cache entry for this request. Basically just malloc a new one
* and then call nfsrc_getudp() or nfsrc_gettcp() to do the rest.
*/
APPLESTATIC int
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
nfsrvd_getcache(struct nfsrv_descript *nd)
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
{
struct nfsrvcache *newrp;
int ret;
if (nd->nd_procnum == NFSPROC_NULL)
panic("nfsd cache null");
MALLOC(newrp, struct nfsrvcache *, sizeof (struct nfsrvcache),
M_NFSRVCACHE, M_WAITOK);
NFSBZERO((caddr_t)newrp, sizeof (struct nfsrvcache));
if (nd->nd_flag & ND_NFSV4)
newrp->rc_flag = RC_NFSV4;
else if (nd->nd_flag & ND_NFSV3)
newrp->rc_flag = RC_NFSV3;
else
newrp->rc_flag = RC_NFSV2;
newrp->rc_xid = nd->nd_retxid;
newrp->rc_proc = nd->nd_procnum;
newrp->rc_sockref = nd->nd_sockref;
newrp->rc_cachetime = nd->nd_tcpconntime;
if (nd->nd_flag & ND_SAMETCPCONN)
newrp->rc_flag |= RC_SAMETCPCONN;
if (nd->nd_nam2 != NULL) {
newrp->rc_flag |= RC_UDP;
ret = nfsrc_getudp(nd, newrp);
} else {
ret = nfsrc_gettcp(nd, newrp);
}
Add DEXITCODE plumbing to NFS. Isilon has the concept of an in-memory exit-code ring that saves the last exit code of a function and allows for stack tracing. This is very helpful when debugging tough issues. This patch is essentially a no-op for BSD at this point, until we upstream the dexitcode logic itself. The patch adds DEXITCODE calls to every NFS function that returns an errno error code. A number of code paths were also reorganized to have single exit paths, to reduce code duplication. Submitted by: David Kwan <dkwan@isilon.com> Reviewed by: rmacklem Approved by: zml (mentor) MFC after: 2 weeks
2011-07-16 08:51:09 +00:00
NFSEXITCODE2(0, nd);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
return (ret);
}
/*
* For UDP (v2, v3):
* - key on <xid, NFS version, RPC#, Client host ip#>
* (at most one entry for each key)
*/
static int
nfsrc_getudp(struct nfsrv_descript *nd, struct nfsrvcache *newrp)
{
struct nfsrvcache *rp;
struct sockaddr_in *saddr;
struct sockaddr_in6 *saddr6;
struct nfsrvhashhead *hp;
int ret = 0;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
struct mtx *mutex;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mutex = nfsrc_cachemutex(newrp);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
hp = NFSRCUDPHASH(newrp->rc_xid);
loop:
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_lock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
LIST_FOREACH(rp, hp, rc_hash) {
if (newrp->rc_xid == rp->rc_xid &&
newrp->rc_proc == rp->rc_proc &&
(newrp->rc_flag & rp->rc_flag & RC_NFSVERS) &&
nfsaddr_match(NETFAMILY(rp), &rp->rc_haddr, nd->nd_nam)) {
if ((rp->rc_flag & RC_LOCKED) != 0) {
rp->rc_flag |= RC_WANTED;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
(void)mtx_sleep(rp, mutex, (PZERO - 1) | PDROP,
"nfsrc", 10 * hz);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
goto loop;
}
if (rp->rc_flag == 0)
panic("nfs udp cache0");
rp->rc_flag |= RC_LOCKED;
TAILQ_REMOVE(&nfsrvudplru, rp, rc_lru);
TAILQ_INSERT_TAIL(&nfsrvudplru, rp, rc_lru);
if (rp->rc_flag & RC_INPROG) {
newnfsstats.srvcache_inproghits++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
ret = RC_DROPIT;
} else if (rp->rc_flag & RC_REPSTATUS) {
/*
* V2 only.
*/
newnfsstats.srvcache_nonidemdonehits++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nfsrvd_rephead(nd);
*(nd->nd_errp) = rp->rc_status;
ret = RC_REPLY;
rp->rc_timestamp = NFSD_MONOSEC +
NFSRVCACHE_UDPTIMEOUT;
} else if (rp->rc_flag & RC_REPMBUF) {
newnfsstats.srvcache_nonidemdonehits++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nd->nd_mreq = m_copym(rp->rc_reply, 0,
Mechanically substitute flags from historic mbuf allocator with malloc(9) flags within sys. Exceptions: - sys/contrib not touched - sys/mbuf.h edited manually
2012-12-05 08:04:20 +00:00
M_COPYALL, M_WAITOK);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
ret = RC_REPLY;
rp->rc_timestamp = NFSD_MONOSEC +
NFSRVCACHE_UDPTIMEOUT;
} else {
panic("nfs udp cache1");
}
nfsrc_unlock(rp);
free((caddr_t)newrp, M_NFSRVCACHE);
Add DEXITCODE plumbing to NFS. Isilon has the concept of an in-memory exit-code ring that saves the last exit code of a function and allows for stack tracing. This is very helpful when debugging tough issues. This patch is essentially a no-op for BSD at this point, until we upstream the dexitcode logic itself. The patch adds DEXITCODE calls to every NFS function that returns an errno error code. A number of code paths were also reorganized to have single exit paths, to reduce code duplication. Submitted by: David Kwan <dkwan@isilon.com> Reviewed by: rmacklem Approved by: zml (mentor) MFC after: 2 weeks
2011-07-16 08:51:09 +00:00
goto out;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
}
newnfsstats.srvcache_misses++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
atomic_add_int(&newnfsstats.srvcache_size, 1);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nfsrc_udpcachesize++;
newrp->rc_flag |= RC_INPROG;
saddr = NFSSOCKADDR(nd->nd_nam, struct sockaddr_in *);
if (saddr->sin_family == AF_INET)
newrp->rc_inet = saddr->sin_addr.s_addr;
else if (saddr->sin_family == AF_INET6) {
saddr6 = (struct sockaddr_in6 *)saddr;
This fixes the experimental NFS server so that it won't crash in the caching code for IPv6 by fixing a typo that used the incorrect variable. It also fixes the indentation of the statement above it. Reported by: simon AT comsys.ntu-kpi.kiev.ua MFC after: 5 days
2010-02-13 23:56:19 +00:00
NFSBCOPY((caddr_t)&saddr6->sin6_addr, (caddr_t)&newrp->rc_inet6,
sizeof (struct in6_addr));
newrp->rc_flag |= RC_INETIPV6;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
LIST_INSERT_HEAD(hp, newrp, rc_hash);
TAILQ_INSERT_TAIL(&nfsrvudplru, newrp, rc_lru);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nd->nd_rp = newrp;
Add DEXITCODE plumbing to NFS. Isilon has the concept of an in-memory exit-code ring that saves the last exit code of a function and allows for stack tracing. This is very helpful when debugging tough issues. This patch is essentially a no-op for BSD at this point, until we upstream the dexitcode logic itself. The patch adds DEXITCODE calls to every NFS function that returns an errno error code. A number of code paths were also reorganized to have single exit paths, to reduce code duplication. Submitted by: David Kwan <dkwan@isilon.com> Reviewed by: rmacklem Approved by: zml (mentor) MFC after: 2 weeks
2011-07-16 08:51:09 +00:00
ret = RC_DOIT;
out:
NFSEXITCODE2(0, nd);
return (ret);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
/*
* Update a request cache entry after the rpc has been done
*/
APPLESTATIC struct nfsrvcache *
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
nfsrvd_updatecache(struct nfsrv_descript *nd)
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
{
struct nfsrvcache *rp;
struct nfsrvcache *retrp = NULL;
Fix a number of places where the new NFS server did not lock the mutex when manipulating rc_flag in the DRC cache. This is believed to fix a hung server that was reported to the freebsd-fs@ list on June 9 under the subject heading "New NFS server stress test hang", where all the threads were waiting for the RC_LOCKED flag to clear. Tested by: jwd at slowblink.com MFC after: 2 weeks
2011-06-19 23:54:01 +00:00
mbuf_t m;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
struct mtx *mutex;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
rp = nd->nd_rp;
if (!rp)
panic("nfsrvd_updatecache null rp");
nd->nd_rp = NULL;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mutex = nfsrc_cachemutex(rp);
mtx_lock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nfsrc_lock(rp);
if (!(rp->rc_flag & RC_INPROG))
panic("nfsrvd_updatecache not inprog");
rp->rc_flag &= ~RC_INPROG;
if (rp->rc_flag & RC_UDP) {
TAILQ_REMOVE(&nfsrvudplru, rp, rc_lru);
TAILQ_INSERT_TAIL(&nfsrvudplru, rp, rc_lru);
}
/*
* Reply from cache is a special case returned by nfsrv_checkseqid().
*/
if (nd->nd_repstat == NFSERR_REPLYFROMCACHE) {
newnfsstats.srvcache_nonidemdonehits++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nd->nd_repstat = 0;
if (nd->nd_mreq)
mbuf_freem(nd->nd_mreq);
if (!(rp->rc_flag & RC_REPMBUF))
panic("reply from cache");
nd->nd_mreq = m_copym(rp->rc_reply, 0,
Mechanically substitute flags from historic mbuf allocator with malloc(9) flags within sys. Exceptions: - sys/contrib not touched - sys/mbuf.h edited manually
2012-12-05 08:04:20 +00:00
M_COPYALL, M_WAITOK);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
rp->rc_timestamp = NFSD_MONOSEC + nfsrc_tcptimeout;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nfsrc_unlock(rp);
Add DEXITCODE plumbing to NFS. Isilon has the concept of an in-memory exit-code ring that saves the last exit code of a function and allows for stack tracing. This is very helpful when debugging tough issues. This patch is essentially a no-op for BSD at this point, until we upstream the dexitcode logic itself. The patch adds DEXITCODE calls to every NFS function that returns an errno error code. A number of code paths were also reorganized to have single exit paths, to reduce code duplication. Submitted by: David Kwan <dkwan@isilon.com> Reviewed by: rmacklem Approved by: zml (mentor) MFC after: 2 weeks
2011-07-16 08:51:09 +00:00
goto out;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
/*
* If rc_refcnt > 0, save it
* For UDP, save it if ND_SAVEREPLY is set
* For TCP, save it if ND_SAVEREPLY and nfsrc_tcpnonidempotent is set
*/
if (nd->nd_repstat != NFSERR_DONTREPLY &&
(rp->rc_refcnt > 0 ||
((nd->nd_flag & ND_SAVEREPLY) && (rp->rc_flag & RC_UDP)) ||
((nd->nd_flag & ND_SAVEREPLY) && !(rp->rc_flag & RC_UDP) &&
nfsrc_tcpsavedreplies <= nfsrc_floodlevel &&
nfsrc_tcpnonidempotent))) {
if (rp->rc_refcnt > 0) {
if (!(rp->rc_flag & RC_NFSV4))
panic("update_cache refcnt");
rp->rc_flag |= RC_REFCNT;
}
if ((nd->nd_flag & ND_NFSV2) &&
nfsv2_repstat[newnfsv2_procid[nd->nd_procnum]]) {
rp->rc_status = nd->nd_repstat;
rp->rc_flag |= RC_REPSTATUS;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
} else {
if (!(rp->rc_flag & RC_UDP)) {
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
atomic_add_int(&nfsrc_tcpsavedreplies, 1);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
if (nfsrc_tcpsavedreplies >
newnfsstats.srvcache_tcppeak)
newnfsstats.srvcache_tcppeak =
nfsrc_tcpsavedreplies;
}
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Mechanically substitute flags from historic mbuf allocator with malloc(9) flags within sys. Exceptions: - sys/contrib not touched - sys/mbuf.h edited manually
2012-12-05 08:04:20 +00:00
m = m_copym(nd->nd_mreq, 0, M_COPYALL, M_WAITOK);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_lock(mutex);
Fix a number of places where the new NFS server did not lock the mutex when manipulating rc_flag in the DRC cache. This is believed to fix a hung server that was reported to the freebsd-fs@ list on June 9 under the subject heading "New NFS server stress test hang", where all the threads were waiting for the RC_LOCKED flag to clear. Tested by: jwd at slowblink.com MFC after: 2 weeks
2011-06-19 23:54:01 +00:00
rp->rc_reply = m;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
rp->rc_flag |= RC_REPMBUF;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
if (rp->rc_flag & RC_UDP) {
rp->rc_timestamp = NFSD_MONOSEC +
NFSRVCACHE_UDPTIMEOUT;
nfsrc_unlock(rp);
} else {
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
rp->rc_timestamp = NFSD_MONOSEC + nfsrc_tcptimeout;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
if (rp->rc_refcnt > 0)
nfsrc_unlock(rp);
else
retrp = rp;
}
} else {
nfsrc_freecache(rp);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
Add DEXITCODE plumbing to NFS. Isilon has the concept of an in-memory exit-code ring that saves the last exit code of a function and allows for stack tracing. This is very helpful when debugging tough issues. This patch is essentially a no-op for BSD at this point, until we upstream the dexitcode logic itself. The patch adds DEXITCODE calls to every NFS function that returns an errno error code. A number of code paths were also reorganized to have single exit paths, to reduce code duplication. Submitted by: David Kwan <dkwan@isilon.com> Reviewed by: rmacklem Approved by: zml (mentor) MFC after: 2 weeks
2011-07-16 08:51:09 +00:00
out:
NFSEXITCODE2(0, nd);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
return (retrp);
}
/*
* Invalidate and, if possible, free an in prog cache entry.
* Must not sleep.
*/
APPLESTATIC void
nfsrvd_delcache(struct nfsrvcache *rp)
{
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
struct mtx *mutex;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mutex = nfsrc_cachemutex(rp);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
if (!(rp->rc_flag & RC_INPROG))
panic("nfsrvd_delcache not in prog");
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_lock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
rp->rc_flag &= ~RC_INPROG;
if (rp->rc_refcnt == 0 && !(rp->rc_flag & RC_LOCKED))
nfsrc_freecache(rp);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
/*
* Called after nfsrvd_updatecache() once the reply is sent, to update
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
* the entry's sequence number and unlock it. The argument is
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
* the pointer returned by nfsrvd_updatecache().
*/
APPLESTATIC void
Fix lock leak in purely hypothetical case of TCP connection without SVC_ACK method. This change should be NOP now, but it is better to be future safe. Reported by: rmacklem
2014-01-14 20:18:38 +00:00
nfsrvd_sentcache(struct nfsrvcache *rp, int have_seq, uint32_t seq)
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
{
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
struct nfsrchash_bucket *hbp;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
KASSERT(rp->rc_flag & RC_LOCKED, ("nfsrvd_sentcache not locked"));
Fix lock leak in purely hypothetical case of TCP connection without SVC_ACK method. This change should be NOP now, but it is better to be future safe. Reported by: rmacklem
2014-01-14 20:18:38 +00:00
if (have_seq) {
hbp = NFSRCAHASH(rp->rc_sockref);
mtx_lock(&hbp->mtx);
rp->rc_tcpseq = seq;
if (rp->rc_acked != RC_NO_ACK)
LIST_INSERT_HEAD(&hbp->tbl, rp, rc_ahash);
rp->rc_acked = RC_NO_ACK;
mtx_unlock(&hbp->mtx);
}
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nfsrc_unlock(rp);
}
/*
* Get a cache entry for TCP
* - key on <xid, nfs version>
* (allow multiple entries for a given key)
*/
static int
nfsrc_gettcp(struct nfsrv_descript *nd, struct nfsrvcache *newrp)
{
struct nfsrvcache *rp, *nextrp;
int i;
struct nfsrvcache *hitrp;
struct nfsrvhashhead *hp, nfsrc_templist;
int hit, ret = 0;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
struct mtx *mutex;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mutex = nfsrc_cachemutex(newrp);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
hp = NFSRCHASH(newrp->rc_xid);
newrp->rc_reqlen = nfsrc_getlenandcksum(nd->nd_mrep, &newrp->rc_cksum);
tryagain:
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_lock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
hit = 1;
LIST_INIT(&nfsrc_templist);
/*
* Get all the matches and put them on the temp list.
*/
rp = LIST_FIRST(hp);
while (rp != LIST_END(hp)) {
nextrp = LIST_NEXT(rp, rc_hash);
if (newrp->rc_xid == rp->rc_xid &&
(!(rp->rc_flag & RC_INPROG) ||
((newrp->rc_flag & RC_SAMETCPCONN) &&
newrp->rc_sockref == rp->rc_sockref)) &&
(newrp->rc_flag & rp->rc_flag & RC_NFSVERS) &&
newrp->rc_proc == rp->rc_proc &&
((newrp->rc_flag & RC_NFSV4) &&
newrp->rc_sockref != rp->rc_sockref &&
newrp->rc_cachetime >= rp->rc_cachetime)
&& newrp->rc_reqlen == rp->rc_reqlen &&
newrp->rc_cksum == rp->rc_cksum) {
LIST_REMOVE(rp, rc_hash);
LIST_INSERT_HEAD(&nfsrc_templist, rp, rc_hash);
}
rp = nextrp;
}
/*
* Now, use nfsrc_templist to decide if there is a match.
*/
i = 0;
LIST_FOREACH(rp, &nfsrc_templist, rc_hash) {
i++;
if (rp->rc_refcnt > 0) {
hit = 0;
break;
}
}
/*
* Can be a hit only if one entry left.
* Note possible hit entry and put nfsrc_templist back on hash
* list.
*/
if (i != 1)
hit = 0;
hitrp = rp = LIST_FIRST(&nfsrc_templist);
while (rp != LIST_END(&nfsrc_templist)) {
nextrp = LIST_NEXT(rp, rc_hash);
LIST_REMOVE(rp, rc_hash);
LIST_INSERT_HEAD(hp, rp, rc_hash);
rp = nextrp;
}
if (LIST_FIRST(&nfsrc_templist) != LIST_END(&nfsrc_templist))
panic("nfs gettcp cache templist");
if (hit) {
rp = hitrp;
if ((rp->rc_flag & RC_LOCKED) != 0) {
rp->rc_flag |= RC_WANTED;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
(void)mtx_sleep(rp, mutex, (PZERO - 1) | PDROP,
"nfsrc", 10 * hz);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
goto tryagain;
}
if (rp->rc_flag == 0)
panic("nfs tcp cache0");
rp->rc_flag |= RC_LOCKED;
if (rp->rc_flag & RC_INPROG) {
newnfsstats.srvcache_inproghits++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
if (newrp->rc_sockref == rp->rc_sockref)
nfsrc_marksametcpconn(rp->rc_sockref);
ret = RC_DROPIT;
} else if (rp->rc_flag & RC_REPSTATUS) {
/*
* V2 only.
*/
newnfsstats.srvcache_nonidemdonehits++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
if (newrp->rc_sockref == rp->rc_sockref)
nfsrc_marksametcpconn(rp->rc_sockref);
ret = RC_REPLY;
nfsrvd_rephead(nd);
*(nd->nd_errp) = rp->rc_status;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
rp->rc_timestamp = NFSD_MONOSEC + nfsrc_tcptimeout;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
} else if (rp->rc_flag & RC_REPMBUF) {
newnfsstats.srvcache_nonidemdonehits++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
if (newrp->rc_sockref == rp->rc_sockref)
nfsrc_marksametcpconn(rp->rc_sockref);
ret = RC_REPLY;
nd->nd_mreq = m_copym(rp->rc_reply, 0,
Mechanically substitute flags from historic mbuf allocator with malloc(9) flags within sys. Exceptions: - sys/contrib not touched - sys/mbuf.h edited manually
2012-12-05 08:04:20 +00:00
M_COPYALL, M_WAITOK);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
rp->rc_timestamp = NFSD_MONOSEC + nfsrc_tcptimeout;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
} else {
panic("nfs tcp cache1");
}
nfsrc_unlock(rp);
free((caddr_t)newrp, M_NFSRVCACHE);
Add DEXITCODE plumbing to NFS. Isilon has the concept of an in-memory exit-code ring that saves the last exit code of a function and allows for stack tracing. This is very helpful when debugging tough issues. This patch is essentially a no-op for BSD at this point, until we upstream the dexitcode logic itself. The patch adds DEXITCODE calls to every NFS function that returns an errno error code. A number of code paths were also reorganized to have single exit paths, to reduce code duplication. Submitted by: David Kwan <dkwan@isilon.com> Reviewed by: rmacklem Approved by: zml (mentor) MFC after: 2 weeks
2011-07-16 08:51:09 +00:00
goto out;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
newnfsstats.srvcache_misses++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
atomic_add_int(&newnfsstats.srvcache_size, 1);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
/*
* For TCP, multiple entries for a key are allowed, so don't
* chain it into the hash table until done.
*/
newrp->rc_cachetime = NFSD_MONOSEC;
newrp->rc_flag |= RC_INPROG;
LIST_INSERT_HEAD(hp, newrp, rc_hash);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nd->nd_rp = newrp;
Add DEXITCODE plumbing to NFS. Isilon has the concept of an in-memory exit-code ring that saves the last exit code of a function and allows for stack tracing. This is very helpful when debugging tough issues. This patch is essentially a no-op for BSD at this point, until we upstream the dexitcode logic itself. The patch adds DEXITCODE calls to every NFS function that returns an errno error code. A number of code paths were also reorganized to have single exit paths, to reduce code duplication. Submitted by: David Kwan <dkwan@isilon.com> Reviewed by: rmacklem Approved by: zml (mentor) MFC after: 2 weeks
2011-07-16 08:51:09 +00:00
ret = RC_DOIT;
out:
NFSEXITCODE2(0, nd);
return (ret);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
/*
* Lock a cache entry.
*/
static void
nfsrc_lock(struct nfsrvcache *rp)
{
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
struct mtx *mutex;
mutex = nfsrc_cachemutex(rp);
mtx_assert(mutex, MA_OWNED);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
while ((rp->rc_flag & RC_LOCKED) != 0) {
rp->rc_flag |= RC_WANTED;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
(void)mtx_sleep(rp, mutex, PZERO - 1, "nfsrc", 0);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
rp->rc_flag |= RC_LOCKED;
}
/*
* Unlock a cache entry.
*/
static void
nfsrc_unlock(struct nfsrvcache *rp)
{
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
struct mtx *mutex;
Fix a number of places where the new NFS server did not lock the mutex when manipulating rc_flag in the DRC cache. This is believed to fix a hung server that was reported to the freebsd-fs@ list on June 9 under the subject heading "New NFS server stress test hang", where all the threads were waiting for the RC_LOCKED flag to clear. Tested by: jwd at slowblink.com MFC after: 2 weeks
2011-06-19 23:54:01 +00:00
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mutex = nfsrc_cachemutex(rp);
mtx_lock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
rp->rc_flag &= ~RC_LOCKED;
nfsrc_wanted(rp);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
/*
* Wakeup anyone wanting entry.
*/
static void
nfsrc_wanted(struct nfsrvcache *rp)
{
if (rp->rc_flag & RC_WANTED) {
rp->rc_flag &= ~RC_WANTED;
wakeup((caddr_t)rp);
}
}
/*
* Free up the entry.
* Must not sleep.
*/
static void
nfsrc_freecache(struct nfsrvcache *rp)
{
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
struct nfsrchash_bucket *hbp;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
LIST_REMOVE(rp, rc_hash);
if (rp->rc_flag & RC_UDP) {
TAILQ_REMOVE(&nfsrvudplru, rp, rc_lru);
nfsrc_udpcachesize--;
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
} else if (rp->rc_acked != RC_NO_SEQ) {
hbp = NFSRCAHASH(rp->rc_sockref);
mtx_lock(&hbp->mtx);
if (rp->rc_acked == RC_NO_ACK)
LIST_REMOVE(rp, rc_ahash);
mtx_unlock(&hbp->mtx);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
nfsrc_wanted(rp);
if (rp->rc_flag & RC_REPMBUF) {
mbuf_freem(rp->rc_reply);
if (!(rp->rc_flag & RC_UDP))
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
atomic_add_int(&nfsrc_tcpsavedreplies, -1);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
FREE((caddr_t)rp, M_NFSRVCACHE);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
atomic_add_int(&newnfsstats.srvcache_size, -1);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
/*
Clean up the experimental NFS server replay cache when the module is unloaded. Reviewed by: rmacklem Approved by: zml (mentor)
2011-01-12 23:34:09 +00:00
* Clean out the cache. Called when nfsserver module is unloaded.
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
*/
APPLESTATIC void
nfsrvd_cleancache(void)
{
struct nfsrvcache *rp, *nextrp;
int i;
for (i = 0; i < NFSRVCACHE_HASHSIZE; i++) {
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_lock(&nfsrchash_table[i].mtx);
LIST_FOREACH_SAFE(rp, &nfsrchash_table[i].tbl, rc_hash, nextrp)
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nfsrc_freecache(rp);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(&nfsrchash_table[i].mtx);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_lock(&nfsrc_udpmtx);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
for (i = 0; i < NFSRVCACHE_HASHSIZE; i++) {
LIST_FOREACH_SAFE(rp, &nfsrvudphashtbl[i], rc_hash, nextrp) {
nfsrc_freecache(rp);
}
}
newnfsstats.srvcache_size = 0;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(&nfsrc_udpmtx);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nfsrc_tcpsavedreplies = 0;
}
Slightly simplify expiration logic introduced in r254337. - Do not update the histogram for items we are any way deleting from cache. - Do not update the histogram if nfsrc_tcphighwater is not set. - Remove some extra math operations.
2013-12-25 16:58:42 +00:00
#define HISTSIZE 16
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
/*
* The basic rule is to get rid of entries that are expired.
*/
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
void
nfsrc_trimcache(u_int64_t sockref, uint32_t snd_una, int final)
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
{
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
struct nfsrchash_bucket *hbp;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
struct nfsrvcache *rp, *nextrp;
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
int force, lastslot, i, j, k, tto, time_histo[HISTSIZE];
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
time_t thisstamp;
static time_t udp_lasttrim = 0, tcp_lasttrim = 0;
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
static int onethread = 0, oneslot = 0;
if (sockref != 0) {
hbp = NFSRCAHASH(sockref);
mtx_lock(&hbp->mtx);
LIST_FOREACH_SAFE(rp, &hbp->tbl, rc_ahash, nextrp) {
if (sockref == rp->rc_sockref) {
if (SEQ_GEQ(snd_una, rp->rc_tcpseq)) {
rp->rc_acked = RC_ACK;
LIST_REMOVE(rp, rc_ahash);
} else if (final) {
rp->rc_acked = RC_NACK;
LIST_REMOVE(rp, rc_ahash);
}
}
}
mtx_unlock(&hbp->mtx);
}
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
if (atomic_cmpset_acq_int(&onethread, 0, 1) == 0)
return;
if (NFSD_MONOSEC != udp_lasttrim ||
nfsrc_udpcachesize >= (nfsrc_udphighwater +
nfsrc_udphighwater / 2)) {
mtx_lock(&nfsrc_udpmtx);
udp_lasttrim = NFSD_MONOSEC;
TAILQ_FOREACH_SAFE(rp, &nfsrvudplru, rc_lru, nextrp) {
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
if (!(rp->rc_flag & (RC_INPROG|RC_LOCKED|RC_WANTED))
&& rp->rc_refcnt == 0
&& ((rp->rc_flag & RC_REFCNT) ||
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
udp_lasttrim > rp->rc_timestamp ||
nfsrc_udpcachesize > nfsrc_udphighwater))
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
nfsrc_freecache(rp);
}
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(&nfsrc_udpmtx);
}
if (NFSD_MONOSEC != tcp_lasttrim ||
nfsrc_tcpsavedreplies >= nfsrc_tcphighwater) {
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
force = nfsrc_tcphighwater / 4;
if (force > 0 &&
nfsrc_tcpsavedreplies + force >= nfsrc_tcphighwater) {
for (i = 0; i < HISTSIZE; i++)
time_histo[i] = 0;
i = 0;
Fix off-by-one error in r260229. Coverity CID: 1148955
2014-01-07 11:43:51 +00:00
lastslot = NFSRVCACHE_HASHSIZE - 1;
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
} else {
force = 0;
if (NFSD_MONOSEC != tcp_lasttrim) {
i = 0;
lastslot = NFSRVCACHE_HASHSIZE - 1;
} else {
lastslot = i = oneslot;
if (++oneslot >= NFSRVCACHE_HASHSIZE)
oneslot = 0;
}
}
Slightly simplify expiration logic introduced in r254337. - Do not update the histogram for items we are any way deleting from cache. - Do not update the histogram if nfsrc_tcphighwater is not set. - Remove some extra math operations.
2013-12-25 16:58:42 +00:00
tto = nfsrc_tcptimeout;
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
tcp_lasttrim = NFSD_MONOSEC;
for (; i <= lastslot; i++) {
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_lock(&nfsrchash_table[i].mtx);
LIST_FOREACH_SAFE(rp, &nfsrchash_table[i].tbl, rc_hash,
nextrp) {
if (!(rp->rc_flag &
(RC_INPROG|RC_LOCKED|RC_WANTED))
&& rp->rc_refcnt == 0) {
Slightly simplify expiration logic introduced in r254337. - Do not update the histogram for items we are any way deleting from cache. - Do not update the histogram if nfsrc_tcphighwater is not set. - Remove some extra math operations.
2013-12-25 16:58:42 +00:00
if ((rp->rc_flag & RC_REFCNT) ||
tcp_lasttrim > rp->rc_timestamp ||
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
rp->rc_acked == RC_ACK) {
Slightly simplify expiration logic introduced in r254337. - Do not update the histogram for items we are any way deleting from cache. - Do not update the histogram if nfsrc_tcphighwater is not set. - Remove some extra math operations.
2013-12-25 16:58:42 +00:00
nfsrc_freecache(rp);
continue;
}
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
if (force == 0)
Slightly simplify expiration logic introduced in r254337. - Do not update the histogram for items we are any way deleting from cache. - Do not update the histogram if nfsrc_tcphighwater is not set. - Remove some extra math operations.
2013-12-25 16:58:42 +00:00
continue;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
/*
* The timestamps range from roughly the
* present (tcp_lasttrim) to the present
* + nfsrc_tcptimeout. Generate a simple
* histogram of where the timeouts fall.
*/
j = rp->rc_timestamp - tcp_lasttrim;
Slightly simplify expiration logic introduced in r254337. - Do not update the histogram for items we are any way deleting from cache. - Do not update the histogram if nfsrc_tcphighwater is not set. - Remove some extra math operations.
2013-12-25 16:58:42 +00:00
if (j >= tto)
j = HISTSIZE - 1;
else if (j < 0)
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
j = 0;
Slightly simplify expiration logic introduced in r254337. - Do not update the histogram for items we are any way deleting from cache. - Do not update the histogram if nfsrc_tcphighwater is not set. - Remove some extra math operations.
2013-12-25 16:58:42 +00:00
else
j = j * HISTSIZE / tto;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
time_histo[j]++;
}
}
mtx_unlock(&nfsrchash_table[i].mtx);
}
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
if (force) {
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
/*
* Trim some more with a smaller timeout of as little
* as 20% of nfsrc_tcptimeout to try and get below
* 80% of the nfsrc_tcphighwater.
*/
k = 0;
Slightly simplify expiration logic introduced in r254337. - Do not update the histogram for items we are any way deleting from cache. - Do not update the histogram if nfsrc_tcphighwater is not set. - Remove some extra math operations.
2013-12-25 16:58:42 +00:00
for (i = 0; i < (HISTSIZE - 2); i++) {
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
k += time_histo[i];
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
if (k > force)
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
break;
}
Slightly simplify expiration logic introduced in r254337. - Do not update the histogram for items we are any way deleting from cache. - Do not update the histogram if nfsrc_tcphighwater is not set. - Remove some extra math operations.
2013-12-25 16:58:42 +00:00
k = tto * (i + 1) / HISTSIZE;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
if (k < 1)
k = 1;
thisstamp = tcp_lasttrim + k;
for (i = 0; i < NFSRVCACHE_HASHSIZE; i++) {
mtx_lock(&nfsrchash_table[i].mtx);
LIST_FOREACH_SAFE(rp, &nfsrchash_table[i].tbl,
rc_hash, nextrp) {
if (!(rp->rc_flag &
(RC_INPROG|RC_LOCKED|RC_WANTED))
&& rp->rc_refcnt == 0
&& ((rp->rc_flag & RC_REFCNT) ||
thisstamp > rp->rc_timestamp ||
Rework NFS Duplicate Request Cache cleanup logic. - 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.
2014-01-03 15:09:59 +00:00
rp->rc_acked == RC_ACK))
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
nfsrc_freecache(rp);
}
mtx_unlock(&nfsrchash_table[i].mtx);
}
}
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
atomic_store_rel_int(&onethread, 0);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
/*
* Add a seqid# reference to the cache entry.
*/
APPLESTATIC void
nfsrvd_refcache(struct nfsrvcache *rp)
{
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
struct mtx *mutex;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mutex = nfsrc_cachemutex(rp);
mtx_lock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
if (rp->rc_refcnt < 0)
panic("nfs cache refcnt");
rp->rc_refcnt++;
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
/*
* Dereference a seqid# cache entry.
*/
APPLESTATIC void
nfsrvd_derefcache(struct nfsrvcache *rp)
{
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
struct mtx *mutex;
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mutex = nfsrc_cachemutex(rp);
mtx_lock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
if (rp->rc_refcnt <= 0)
panic("nfs cache derefcnt");
rp->rc_refcnt--;
if (rp->rc_refcnt == 0 && !(rp->rc_flag & (RC_LOCKED | RC_INPROG)))
nfsrc_freecache(rp);
Fix several performance related issues in the new NFS server's DRC for NFS over TCP. - Increase the size of the hash tables. - Create a separate mutex for each hash list of the TCP hash table. - Single thread the code that deletes stale cache entries. - Add a tunable called vfs.nfsd.tcphighwater, which can be increased to allow the cache to grow larger, avoiding the overhead of frequent scans to delete stale cache entries. (The default value will result in frequent scans to delete stale cache entries, analagous to what the pre-patched code does.) - Add a tunable called vfs.nfsd.cachetcp that can be used to disable DRC caching for NFS over TCP, since the old NFS server didn't DRC cache TCP. It also adjusts the size of nfsrc_floodlevel dynamically, so that it is always greater than vfs.nfsd.tcphighwater. For UDP the algorithm remains the same as the pre-patched code, but the tunable vfs.nfsd.udphighwater can be used to allow the cache to grow larger and reduce the overhead caused by frequent scans for stale entries. UDP also uses a larger hash table size than the pre-patched code. Reported by: wollman Tested by: wollman (earlier version of patch) Submitted by: ivoras (earlier patch) Reviewed by: jhb (earlier version of patch) MFC after: 1 month
2013-08-14 21:11:26 +00:00
mtx_unlock(mutex);
Add the experimental nfs subtree to the kernel, that includes support for NFSv4 as well as NFSv2 and 3. It lives in 3 subdirs under sys/fs: nfs - functions that are common to the client and server nfsclient - a mutation of sys/nfsclient that call generic functions to do RPCs and handle state. As such, it retains the buffer cache handling characteristics and vnode semantics that are found in sys/nfsclient, for the most part. nfsserver - the server. It includes a DRC designed specifically for NFSv4, that is used instead of the generic DRC in sys/rpc. The build glue will be checked in later, so at this point, it consists of 3 new subdirs that should not affect kernel building. Approved by: kib (mentor)
2009-05-04 15:23:58 +00:00
}
/*
* Calculate the length of the mbuf list and a checksum on the first up to
* NFSRVCACHE_CHECKLEN bytes.
*/
static int
nfsrc_getlenandcksum(mbuf_t m1, u_int16_t *cksum)
{
int len = 0, cklen;
mbuf_t m;
m = m1;
while (m) {
len += mbuf_len(m);
m = mbuf_next(m);
}
cklen = (len > NFSRVCACHE_CHECKLEN) ? NFSRVCACHE_CHECKLEN : len;
*cksum = in_cksum(m1, cklen);
return (len);
}
/*
* Mark a TCP connection that is seeing retries. Should never happen for
* NFSv4.
*/
static void
nfsrc_marksametcpconn(u_int64_t sockref)
{
}
Reference in New Issue Copy Permalink