From 10f8f58d4a6ddb230532e096e34aa9ff0982793f Mon Sep 17 00:00:00 2001 From: Alexander Motin Date: Mon, 23 Dec 2013 08:43:16 +0000 Subject: [PATCH] Fix RPC server threads file handle affinity to work better with ZFS. Instead of taking 8 specific bytes of file handle to identify file during RPC thread affitinity handling, use trivial hash of the full file handle. ZFS's struct zfid_short does not have padding field after the length field, as result, originally picked 8 bytes are loosing lower 16 bits of object ID, causing many false matches and unneeded requests affinity to same thread. This fix substantially improves NFS server latency and scalability in SPEC NFS benchmark by more flexible use of multiple NFS threads. Sponsored by: iXsystems, Inc. --- sys/fs/nfsserver/nfs_fha_new.c | 18 +++++++++++------- sys/nfs/nfs_fha.c | 5 +---- sys/nfs/nfs_fha.h | 2 +- sys/nfsserver/nfs_fha_old.c | 30 +++++++++++++++++++++++++++--- 4 files changed, 40 insertions(+), 15 deletions(-) diff --git a/sys/fs/nfsserver/nfs_fha_new.c b/sys/fs/nfsserver/nfs_fha_new.c index c05ad282f63a..2e7115fd9fd3 100644 --- a/sys/fs/nfsserver/nfs_fha_new.c +++ b/sys/fs/nfsserver/nfs_fha_new.c @@ -41,7 +41,7 @@ static void fhanew_init(void *foo); static void fhanew_uninit(void *foo); rpcproc_t fhanew_get_procnum(rpcproc_t procnum); int fhanew_realign(struct mbuf **mb, int malloc_flags); -int fhanew_get_fh(fhandle_t *fh, int v3, struct mbuf **md, caddr_t *dpos); +int fhanew_get_fh(uint64_t *fh, int v3, struct mbuf **md, caddr_t *dpos); int fhanew_is_read(rpcproc_t procnum); int fhanew_is_write(rpcproc_t procnum); int fhanew_get_offset(struct mbuf **md, caddr_t *dpos, int v3, @@ -128,11 +128,13 @@ fhanew_realign(struct mbuf **mb, int malloc_flags) } int -fhanew_get_fh(fhandle_t *fh, int v3, struct mbuf **md, caddr_t *dpos) +fhanew_get_fh(uint64_t *fh, int v3, struct mbuf **md, caddr_t *dpos) { struct nfsrv_descript lnd, *nd; uint32_t *tl; - int error, len; + uint8_t *buf; + uint64_t t; + int error, len, i; error = 0; len = 0; @@ -151,11 +153,13 @@ fhanew_get_fh(fhandle_t *fh, int v3, struct mbuf **md, caddr_t *dpos) len = NFSX_V2FH; } + t = 0; if (len != 0) { - NFSM_DISSECT_NONBLOCK(tl, uint32_t *, len); - bcopy(tl, fh, len); - } else - bzero(fh, sizeof(*fh)); + NFSM_DISSECT_NONBLOCK(buf, uint8_t *, len); + for (i = 0; i < len; i++) + t ^= ((uint64_t)buf[i] << (i & 7) * 8); + } + *fh = t; nfsmout: *md = nd->nd_md; diff --git a/sys/nfs/nfs_fha.c b/sys/nfs/nfs_fha.c index 229080471171..18927292d5ec 100644 --- a/sys/nfs/nfs_fha.c +++ b/sys/nfs/nfs_fha.c @@ -130,7 +130,6 @@ fha_extract_info(struct svc_req *req, struct fha_info *i, struct fha_callbacks *cb) { struct mbuf *md; - fhandle_t fh; caddr_t dpos; static u_int64_t random_fh = 0; int error; @@ -177,12 +176,10 @@ fha_extract_info(struct svc_req *req, struct fha_info *i, dpos = mtod(md, caddr_t); /* Grab the filehandle. */ - error = cb->get_fh(&fh, v3, &md, &dpos); + error = cb->get_fh(&i->fh, v3, &md, &dpos); if (error) goto out; - bcopy(fh.fh_fid.fid_data, &i->fh, sizeof(i->fh)); - /* Content ourselves with zero offset for all but reads. */ if (cb->is_read(procnum) || cb->is_write(procnum)) cb->get_offset(&md, &dpos, v3, i); diff --git a/sys/nfs/nfs_fha.h b/sys/nfs/nfs_fha.h index ef5ada0f1df9..d3dd0394ab59 100644 --- a/sys/nfs/nfs_fha.h +++ b/sys/nfs/nfs_fha.h @@ -82,7 +82,7 @@ struct fha_info { struct fha_callbacks { rpcproc_t (*get_procnum)(rpcproc_t procnum); int (*realign)(struct mbuf **mb, int malloc_flags); - int (*get_fh)(fhandle_t *fh, int v3, struct mbuf **md, caddr_t *dpos); + int (*get_fh)(uint64_t *fh, int v3, struct mbuf **md, caddr_t *dpos); int (*is_read)(rpcproc_t procnum); int (*is_write)(rpcproc_t procnum); int (*get_offset)(struct mbuf **md, caddr_t *dpos, int v3, struct diff --git a/sys/nfsserver/nfs_fha_old.c b/sys/nfsserver/nfs_fha_old.c index 82149b867b20..a9282785fef4 100644 --- a/sys/nfsserver/nfs_fha_old.c +++ b/sys/nfsserver/nfs_fha_old.c @@ -49,7 +49,7 @@ static void fhaold_init(void *foo); static void fhaold_uninit(void *foo); rpcproc_t fhaold_get_procnum(rpcproc_t procnum); int fhaold_realign(struct mbuf **mb, int malloc_flags); -int fhaold_get_fh(fhandle_t *fh, int v3, struct mbuf **md, caddr_t *dpos); +int fhaold_get_fh(uint64_t *fh, int v3, struct mbuf **md, caddr_t *dpos); int fhaold_is_read(rpcproc_t procnum); int fhaold_is_write(rpcproc_t procnum); int fhaold_get_offset(struct mbuf **md, caddr_t *dpos, int v3, @@ -135,9 +135,33 @@ fhaold_realign(struct mbuf **mb, int malloc_flags) } int -fhaold_get_fh(fhandle_t *fh, int v3, struct mbuf **md, caddr_t *dpos) +fhaold_get_fh(uint64_t *fh, int v3, struct mbuf **md, caddr_t *dpos) { - return (nfsm_srvmtofh_xx(fh, v3, md, dpos)); + u_int32_t *tl; + uint8_t *buf; + uint64_t t; + int fhlen, i; + + if (v3) { + tl = nfsm_dissect_xx_nonblock(NFSX_UNSIGNED, md, dpos); + if (tl == NULL) + return EBADRPC; + fhlen = fxdr_unsigned(int, *tl); + if (fhlen != 0 && fhlen != NFSX_V3FH) + return EBADRPC; + } else { + fhlen = NFSX_V2FH; + } + t = 0; + if (fhlen != 0) { + buf = nfsm_dissect_xx_nonblock(fhlen, md, dpos); + if (buf == NULL) + return EBADRPC; + for (i = 0; i < fhlen; i++) + t ^= ((uint64_t)buf[i] << (i & 7) * 8); + } + *fh = t; + return 0; } int