Return-Path: linux-nfs-owner@vger.kernel.org Received: from mail-ig0-f175.google.com ([209.85.213.175]:46196 "EHLO mail-ig0-f175.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751293AbaDYPij convert rfc822-to-8bit (ORCPT ); Fri, 25 Apr 2014 11:38:39 -0400 Received: by mail-ig0-f175.google.com with SMTP id h3so2338513igd.14 for ; Fri, 25 Apr 2014 08:38:39 -0700 (PDT) Content-Type: text/plain; charset=windows-1252 Mime-Version: 1.0 (Mac OS X Mail 7.2 \(1874\)) Subject: Re: [PATCH 15/18 v2] pnfs: support multiple verfs per direct req From: Weston Andros Adamson In-Reply-To: <7192CB82-72A3-466B-88AD-061A8C998AB2@primarydata.com> Date: Fri, 25 Apr 2014 11:38:39 -0400 Cc: Trond Myklebust , linux-nfs list Message-Id: References: <1398363328-7100-1-git-send-email-dros@primarydata.com> <1398363328-7100-16-git-send-email-dros@primarydata.com> <535A6DE9.1040109@gmail.com> <7192CB82-72A3-466B-88AD-061A8C998AB2@primarydata.com> To: Anna Schumaker Sender: linux-nfs-owner@vger.kernel.org List-ID: On Apr 25, 2014, at 11:12 AM, Weston Andros Adamson wrote: > On Apr 25, 2014, at 10:15 AM, Anna Schumaker wrote: > >> On 04/24/2014 02:15 PM, Weston Andros Adamson wrote: >>> Support direct requests that span multiple pnfs data servers by >>> comparing nfs_pgio_header->verf to a cached verf in pnfs_commit_bucket. >>> Continue to use dreq->verf if the MDS is used / non-pNFS. >>> >>> Signed-off-by: Weston Andros Adamson >>> --- >>> fs/nfs/direct.c | 98 ++++++++++++++++++++++++++++++++++++++++++++++--- >>> fs/nfs/nfs4filelayout.c | 6 +++ >>> include/linux/nfs.h | 5 ++- >>> include/linux/nfs_xdr.h | 2 + >>> 4 files changed, 105 insertions(+), 6 deletions(-) >>> >>> diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c >>> index 2c0e08f..9349933 100644 >>> --- a/fs/nfs/direct.c >>> +++ b/fs/nfs/direct.c >>> @@ -108,6 +108,93 @@ static inline int put_dreq(struct nfs_direct_req *dreq) >>> return atomic_dec_and_test(&dreq->io_count); >>> } >>> >>> +/* >>> + * nfs_direct_select_verf - select the right verifier >>> + * @dreq - direct request possibly spanning multiple servers >>> + * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs >>> + * @ds_idx - index of data server in data server list, only valid if ds_clp set >>> + * >>> + * returns the correct verifier to use given the role of the server >>> + */ >>> +static struct nfs_writeverf * >>> +nfs_direct_select_verf(struct nfs_direct_req *dreq, >>> + struct nfs_client *ds_clp, >>> + int ds_idx) >>> +{ >>> + struct nfs_writeverf *verfp = &dreq->verf; >>> + >>> + if (ds_clp) { >>> + /* pNFS is in use, use the DS verf */ >>> + if (ds_idx >= 0 && ds_idx < dreq->ds_cinfo.nbuckets) >> >> The struct pnfs_ds_commit_info is empty if CONFIG_NFS_V4_1=n, so this won't compile. >> >> Anna > > Good catch, I?ll iterate through the patchset and test with v4.1 disabled. > > Time to add some #ifdefs > > -dros That was the only problem I found with 4.1 disabled. Fixed and pushed. -dros > >> >>> + verfp = &dreq->ds_cinfo.buckets[ds_idx].direct_verf; >>> + else >>> + WARN_ON_ONCE(1); >>> + } >>> + return verfp; >>> +} >>> + >>> + >>> +/* >>> + * nfs_direct_set_hdr_verf - set the write/commit verifier >>> + * @dreq - direct request possibly spanning multiple servers >>> + * @hdr - pageio header to validate against previously seen verfs >>> + * >>> + * Set the server's (MDS or DS) "seen" verifier >>> + */ >>> +static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq, >>> + struct nfs_pgio_header *hdr) >>> +{ >>> + struct nfs_writeverf *verfp; >>> + >>> + verfp = nfs_direct_select_verf(dreq, hdr->data->ds_clp, >>> + hdr->data->ds_idx); >>> + WARN_ON_ONCE(verfp->committed >= 0); >>> + memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf)); >>> + WARN_ON_ONCE(verfp->committed < 0); >>> +} >>> + >>> +/* >>> + * nfs_direct_cmp_hdr_verf - compare verifier for pgio header >>> + * @dreq - direct request possibly spanning multiple servers >>> + * @hdr - pageio header to validate against previously seen verf >>> + * >>> + * set the server's "seen" verf if not initialized. >>> + * returns result of comparison between @hdr->verf and the "seen" >>> + * verf of the server used by @hdr (DS or MDS) >>> + */ >>> +static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq, >>> + struct nfs_pgio_header *hdr) >>> +{ >>> + struct nfs_writeverf *verfp; >>> + >>> + verfp = nfs_direct_select_verf(dreq, hdr->data->ds_clp, >>> + hdr->data->ds_idx); >>> + if (verfp->committed < 0) { >>> + nfs_direct_set_hdr_verf(dreq, hdr); >>> + return 0; >>> + } >>> + return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf)); >>> +} >>> + >>> +/* >>> + * nfs_direct_cmp_commit_data_verf - compare verifier for commit data >>> + * @dreq - direct request possibly spanning multiple servers >>> + * @data - commit data to validate against previously seen verf >>> + * >>> + * returns result of comparison between @data->verf and the verf of >>> + * the server used by @data (DS or MDS) >>> + */ >>> +static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq, >>> + struct nfs_commit_data *data) >>> +{ >>> + struct nfs_writeverf *verfp; >>> + >>> + verfp = nfs_direct_select_verf(dreq, data->ds_clp, >>> + data->ds_commit_index); >>> + WARN_ON_ONCE(verfp->committed < 0); >>> + return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf)); >>> +} >>> + >>> /** >>> * nfs_direct_IO - NFS address space operation for direct I/O >>> * @rw: direction (read or write) >>> @@ -168,6 +255,7 @@ static inline struct nfs_direct_req *nfs_direct_req_alloc(void) >>> kref_get(&dreq->kref); >>> init_completion(&dreq->completion); >>> INIT_LIST_HEAD(&dreq->mds_cinfo.list); >>> + dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */ >>> INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); >>> spin_lock_init(&dreq->lock); >>> >>> @@ -602,7 +690,7 @@ static void nfs_direct_commit_complete(struct nfs_commit_data *data) >>> dprintk("NFS: %5u commit failed with error %d.\n", >>> data->task.tk_pid, status); >>> dreq->flags = NFS_ODIRECT_RESCHED_WRITES; >>> - } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) { >>> + } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) { >>> dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid); >>> dreq->flags = NFS_ODIRECT_RESCHED_WRITES; >>> } >>> @@ -811,13 +899,13 @@ static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) >>> if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) >>> bit = NFS_IOHDR_NEED_RESCHED; >>> else if (dreq->flags == 0) { >>> - memcpy(&dreq->verf, &hdr->verf, >>> - sizeof(dreq->verf)); >>> + nfs_direct_set_hdr_verf(dreq, hdr); >>> bit = NFS_IOHDR_NEED_COMMIT; >>> dreq->flags = NFS_ODIRECT_DO_COMMIT; >>> } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { >>> - if (memcmp(&dreq->verf, &hdr->verf, sizeof(dreq->verf))) { >>> - dreq->flags = NFS_ODIRECT_RESCHED_WRITES; >>> + if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr)) { >>> + dreq->flags = >>> + NFS_ODIRECT_RESCHED_WRITES; >>> bit = NFS_IOHDR_NEED_RESCHED; >>> } else >>> bit = NFS_IOHDR_NEED_COMMIT; >>> diff --git a/fs/nfs/nfs4filelayout.c b/fs/nfs/nfs4filelayout.c >>> index 7a665e0..0ebc521 100644 >>> --- a/fs/nfs/nfs4filelayout.c >>> +++ b/fs/nfs/nfs4filelayout.c >>> @@ -560,6 +560,7 @@ filelayout_read_pagelist(struct nfs_pgio_data *data) >>> /* No multipath support. Use first DS */ >>> atomic_inc(&ds->ds_clp->cl_count); >>> data->ds_clp = ds->ds_clp; >>> + data->ds_idx = idx; >>> fh = nfs4_fl_select_ds_fh(lseg, j); >>> if (fh) >>> data->args.fh = fh; >>> @@ -603,6 +604,7 @@ filelayout_write_pagelist(struct nfs_pgio_data *data, int sync) >>> data->pgio_done_cb = filelayout_write_done_cb; >>> atomic_inc(&ds->ds_clp->cl_count); >>> data->ds_clp = ds->ds_clp; >>> + data->ds_idx = idx; >>> fh = nfs4_fl_select_ds_fh(lseg, j); >>> if (fh) >>> data->args.fh = fh; >>> @@ -875,6 +877,8 @@ filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg, >>> for (i = 0; i < size; i++) { >>> INIT_LIST_HEAD(&buckets[i].written); >>> INIT_LIST_HEAD(&buckets[i].committing); >>> + /* mark direct verifier as unset */ >>> + buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW; >>> } >>> >>> spin_lock(cinfo->lock); >>> @@ -885,6 +889,8 @@ filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg, >>> &buckets[i].written); >>> list_splice(&cinfo->ds->buckets[i].committing, >>> &buckets[i].committing); >>> + buckets[i].direct_verf.committed = >>> + cinfo->ds->buckets[i].direct_verf.committed; >>> buckets[i].wlseg = cinfo->ds->buckets[i].wlseg; >>> buckets[i].clseg = cinfo->ds->buckets[i].clseg; >>> } >>> diff --git a/include/linux/nfs.h b/include/linux/nfs.h >>> index 3e794c1..610af51 100644 >>> --- a/include/linux/nfs.h >>> +++ b/include/linux/nfs.h >>> @@ -46,6 +46,9 @@ static inline void nfs_copy_fh(struct nfs_fh *target, const struct nfs_fh *sourc >>> enum nfs3_stable_how { >>> NFS_UNSTABLE = 0, >>> NFS_DATA_SYNC = 1, >>> - NFS_FILE_SYNC = 2 >>> + NFS_FILE_SYNC = 2, >>> + >>> + /* used by direct.c to mark verf as invalid */ >>> + NFS_INVALID_STABLE_HOW = -1 >>> }; >>> #endif /* _LINUX_NFS_H */ >>> diff --git a/include/linux/nfs_xdr.h b/include/linux/nfs_xdr.h >>> index 29828c7..bb9fb88 100644 >>> --- a/include/linux/nfs_xdr.h >>> +++ b/include/linux/nfs_xdr.h >>> @@ -1111,6 +1111,7 @@ struct pnfs_commit_bucket { >>> struct list_head committing; >>> struct pnfs_layout_segment *wlseg; >>> struct pnfs_layout_segment *clseg; >>> + struct nfs_writeverf direct_verf; >>> }; >>> >>> struct pnfs_ds_commit_info { >>> @@ -1294,6 +1295,7 @@ struct nfs_pgio_data { >>> __u64 mds_offset; /* Filelayout dense stripe */ >>> struct nfs_page_array pages; >>> struct nfs_client *ds_clp; /* pNFS data server */ >>> + int ds_idx; /* ds index if ds_clp is set */ >>> }; >>> >>> struct nfs_rw_header {