Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S935287AbZLQCZ2 (ORCPT ); Wed, 16 Dec 2009 21:25:28 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1758420AbZLQCZY (ORCPT ); Wed, 16 Dec 2009 21:25:24 -0500 Received: from mail.nec-labs.com ([138.15.200.209]:46193 "EHLO mail.nec-labs.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1757982AbZLQCZX (ORCPT ); Wed, 16 Dec 2009 21:25:23 -0500 X-Greylist: delayed 1298 seconds by postgrey-1.27 at vger.kernel.org; Wed, 16 Dec 2009 21:25:22 EST Subject: [PATCH] improve the performance of large sequential write NFS workloads From: Steve Rago To: linux-nfs@vger.kernel.org Cc: linux-kernel@vger.kernel.org, Trond.Myklebust@netapp.com Content-Type: text/plain Date: Wed, 16 Dec 2009 21:03:40 -0500 Message-Id: <1261015420.1947.54.camel@serenity> Mime-Version: 1.0 X-Mailer: Evolution 2.12.3 (2.12.3-5.fc8) Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 14110 Lines: 403 Eager Writeback for NFS Clients ------------------------------- Prevent applications that write large sequential streams of data (like backup, for example) from entering into a memory pressure state, which degrades performance by falling back to synchronous operations (both synchronous writes and additional commits). This is accomplished by preventing the client application from dirtying pages faster than they can be written to the server: clients write pages eagerly instead of lazily. The eager writeback is controlled by a sysctl: fs.nfs.nfs_max_woutstanding set to 0 disables the feature. Otherwise it contains the maximum number of outstanding NFS writes that can be in flight for a given file. This is used to block the application from dirtying more pages until the writes are complete. This patch is based heavily (okay, almost entirely) on a prior patch by Peter Staubach. For the original patch, see http://article.gmane.org/gmane.linux.nfs/24323. The patch below applies to linux-2.6.32-rc7, but it should apply cleanly to vanilla linux-2.6.32. Performance data and tuning notes can be found on my web site (http://www.nec-labs.com/~sar). With iozone, I see about 50% improvement for large sequential write workloads over a 1Gb Ethernet. With an in-house micro-benchmark, I see 80% improvement for large, single-stream, sequential workloads (where "large" is defined to be greater than the memory size on the client). Signed-off-by: Steve Rago --- diff -rupN -X linux-2.6.32-rc7/Documentation/dontdiff linux-2.6.32-rc7-orig/fs/fs-writeback.c linux-2.6.32-rc7/fs/fs-writeback.c --- linux-2.6.32-rc7-orig/fs/fs-writeback.c 2009-11-12 19:46:07.000000000 -0500 +++ linux-2.6.32-rc7/fs/fs-writeback.c 2009-11-30 15:36:30.735453450 -0500 @@ -771,6 +771,8 @@ static long wb_writeback(struct bdi_writ wbc.range_start = 0; wbc.range_end = LLONG_MAX; } + if (args->for_background || wbc.for_kupdate) + wbc.nonblocking = 1; for (;;) { /* @@ -859,6 +861,8 @@ static long wb_check_old_data_flush(stru unsigned long expired; long nr_pages; + if (dirty_writeback_interval == 0) + return 0; expired = wb->last_old_flush + msecs_to_jiffies(dirty_writeback_interval * 10); if (time_before(jiffies, expired)) @@ -954,7 +958,11 @@ int bdi_writeback_task(struct bdi_writeb break; } - wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10); + if (dirty_writeback_interval == 0) + wait_jiffies = msecs_to_jiffies(5000); /* default */ + else + wait_jiffies = + msecs_to_jiffies(dirty_writeback_interval * 10); schedule_timeout_interruptible(wait_jiffies); try_to_freeze(); } diff -rupN -X linux-2.6.32-rc7/Documentation/dontdiff linux-2.6.32-rc7-orig/fs/nfs/file.c linux-2.6.32-rc7/fs/nfs/file.c --- linux-2.6.32-rc7-orig/fs/nfs/file.c 2009-11-12 19:46:07.000000000 -0500 +++ linux-2.6.32-rc7/fs/nfs/file.c 2009-11-30 15:21:22.635101295 -0500 @@ -589,11 +589,17 @@ static int nfs_need_sync_write(struct fi return 0; } +static int nfs_is_seqwrite(struct inode *inode, loff_t pos) +{ + return NFS_I(inode)->wrpos == pos; +} + static ssize_t nfs_file_write(struct kiocb *iocb, const struct iovec *iov, unsigned long nr_segs, loff_t pos) { struct dentry * dentry = iocb->ki_filp->f_path.dentry; struct inode * inode = dentry->d_inode; + struct nfs_inode *nfsi = NFS_I(inode); ssize_t result; size_t count = iov_length(iov, nr_segs); @@ -607,6 +613,12 @@ static ssize_t nfs_file_write(struct kio result = -EBUSY; if (IS_SWAPFILE(inode)) goto out_swapfile; + + result = count; + if (!count) + goto out; + nfs_wait_woutstanding(inode); + /* * O_APPEND implies that we must revalidate the file length. */ @@ -623,10 +635,21 @@ static ssize_t nfs_file_write(struct kio nfs_add_stats(inode, NFSIOS_NORMALWRITTENBYTES, count); result = generic_file_aio_write(iocb, iov, nr_segs, pos); /* Return error values for O_SYNC and IS_SYNC() */ - if (result >= 0 && nfs_need_sync_write(iocb->ki_filp, inode)) { - int err = nfs_do_fsync(nfs_file_open_context(iocb->ki_filp), inode); - if (err < 0) - result = err; + if (result >= 0) { + if (nfs_need_sync_write(iocb->ki_filp, inode)) { + int err; + + err = nfs_do_fsync(nfs_file_open_context(iocb->ki_filp), + inode); + if (err < 0) + result = err; + } else if (nfs_max_woutstanding != 0 && + nfs_is_seqwrite(inode, pos) && + atomic_read(&nfsi->ndirty) >= NFS_SERVER(inode)->wpages) { + nfs_wb_eager(inode); + } + if (result > 0) + nfsi->wrpos = pos + result; } out: return result; diff -rupN -X linux-2.6.32-rc7/Documentation/dontdiff linux-2.6.32-rc7-orig/fs/nfs/inode.c linux-2.6.32-rc7/fs/nfs/inode.c --- linux-2.6.32-rc7-orig/fs/nfs/inode.c 2009-11-12 19:46:07.000000000 -0500 +++ linux-2.6.32-rc7/fs/nfs/inode.c 2009-11-13 11:36:43.888410914 -0500 @@ -508,7 +508,9 @@ void nfs_setattr_update_inode(struct ino int nfs_getattr(struct vfsmount *mnt, struct dentry *dentry, struct kstat *stat) { struct inode *inode = dentry->d_inode; - int need_atime = NFS_I(inode)->cache_validity & NFS_INO_INVALID_ATIME; + struct nfs_inode *nfsi = NFS_I(inode); + int need_atime = nfsi->cache_validity & NFS_INO_INVALID_ATIME; + int woutstanding = nfs_max_woutstanding; int err; /* @@ -519,9 +521,8 @@ int nfs_getattr(struct vfsmount *mnt, st * nfs_wb_nocommit. */ if (S_ISREG(inode->i_mode)) { - mutex_lock(&inode->i_mutex); + atomic_add(woutstanding, &nfsi->writes); nfs_wb_nocommit(inode); - mutex_unlock(&inode->i_mutex); } /* @@ -545,6 +546,11 @@ int nfs_getattr(struct vfsmount *mnt, st generic_fillattr(inode, stat); stat->ino = nfs_compat_user_ino64(NFS_FILEID(inode)); } + + if (S_ISREG(inode->i_mode)) { + atomic_sub(woutstanding, &nfsi->writes); + wake_up(&nfsi->writes_wq); + } return err; } @@ -1418,9 +1424,13 @@ static void init_once(void *foo) INIT_LIST_HEAD(&nfsi->access_cache_inode_lru); INIT_RADIX_TREE(&nfsi->nfs_page_tree, GFP_ATOMIC); nfsi->npages = 0; + atomic_set(&nfsi->ndirty, 0); atomic_set(&nfsi->silly_count, 1); INIT_HLIST_HEAD(&nfsi->silly_list); init_waitqueue_head(&nfsi->waitqueue); + atomic_set(&nfsi->writes, 0); + init_waitqueue_head(&nfsi->writes_wq); + nfsi->wrpos = 0; nfs4_init_once(nfsi); } diff -rupN -X linux-2.6.32-rc7/Documentation/dontdiff linux-2.6.32-rc7-orig/fs/nfs/sysctl.c linux-2.6.32-rc7/fs/nfs/sysctl.c --- linux-2.6.32-rc7-orig/fs/nfs/sysctl.c 2009-11-12 19:46:07.000000000 -0500 +++ linux-2.6.32-rc7/fs/nfs/sysctl.c 2009-11-13 11:36:43.895459044 -0500 @@ -58,6 +58,14 @@ static ctl_table nfs_cb_sysctls[] = { .mode = 0644, .proc_handler = &proc_dointvec, }, + { + .ctl_name = CTL_UNNUMBERED, + .procname = "nfs_max_woutstanding", + .data = &nfs_max_woutstanding, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = &proc_dointvec, + }, { .ctl_name = 0 } }; diff -rupN -X linux-2.6.32-rc7/Documentation/dontdiff linux-2.6.32-rc7-orig/fs/nfs/write.c linux-2.6.32-rc7/fs/nfs/write.c --- linux-2.6.32-rc7-orig/fs/nfs/write.c 2009-11-12 19:46:07.000000000 -0500 +++ linux-2.6.32-rc7/fs/nfs/write.c 2009-12-08 13:26:35.416629518 -0500 @@ -176,6 +176,8 @@ static void nfs_mark_uptodate(struct pag static int wb_priority(struct writeback_control *wbc) { + if (nfs_max_woutstanding != 0) + return 0; if (wbc->for_reclaim) return FLUSH_HIGHPRI | FLUSH_STABLE; if (wbc->for_kupdate) @@ -200,7 +202,9 @@ static int nfs_set_page_writeback(struct if (!ret) { struct inode *inode = page->mapping->host; struct nfs_server *nfss = NFS_SERVER(inode); + struct nfs_inode *nfsi = NFS_I(inode); + atomic_dec(&nfsi->ndirty); if (atomic_long_inc_return(&nfss->writeback) > NFS_CONGESTION_ON_THRESH) { set_bdi_congested(&nfss->backing_dev_info, @@ -325,6 +329,39 @@ static int nfs_writepages_callback(struc return ret; } +int nfs_max_woutstanding = 16; + +static void nfs_inc_woutstanding(struct inode *inode) +{ + struct nfs_inode *nfsi = NFS_I(inode); + atomic_inc(&nfsi->writes); +} + +static void nfs_dec_woutstanding(struct inode *inode) +{ + struct nfs_inode *nfsi = NFS_I(inode); + if (atomic_dec_return(&nfsi->writes) < nfs_max_woutstanding) + wake_up(&nfsi->writes_wq); +} + +void nfs_wait_woutstanding(struct inode *inode) +{ + if (nfs_max_woutstanding != 0) { + unsigned long background_thresh; + unsigned long dirty_thresh; + long npages; + struct nfs_inode *nfsi = NFS_I(inode); + + get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL); + npages = global_page_state(NR_FILE_DIRTY) + + global_page_state(NR_UNSTABLE_NFS) + + global_page_state(NR_WRITEBACK); + if (npages >= background_thresh) + wait_event(nfsi->writes_wq, + atomic_read(&nfsi->writes) < nfs_max_woutstanding); + } +} + int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc) { struct inode *inode = mapping->host; @@ -420,6 +457,9 @@ static void nfs_inode_remove_request(str static void nfs_mark_request_dirty(struct nfs_page *req) { + struct inode *inode = req->wb_context->path.dentry->d_inode; + struct nfs_inode *nfsi = NFS_I(inode); + atomic_inc(&nfsi->ndirty); __set_page_dirty_nobuffers(req->wb_page); } @@ -682,16 +722,18 @@ static struct nfs_page * nfs_setup_write req = nfs_try_to_update_request(inode, page, offset, bytes); if (req != NULL) - goto out; + return req; req = nfs_create_request(ctx, inode, page, offset, bytes); if (IS_ERR(req)) - goto out; + return req; error = nfs_inode_add_request(inode, req); if (error != 0) { nfs_release_request(req); req = ERR_PTR(error); + } else { + struct nfs_inode *nfsi = NFS_I(inode); + atomic_inc(&nfsi->ndirty); } -out: return req; } @@ -877,6 +919,7 @@ static int nfs_write_rpcsetup(struct nfs count, (unsigned long long)data->args.offset); + nfs_inc_woutstanding(inode); task = rpc_run_task(&task_setup_data); if (IS_ERR(task)) return PTR_ERR(task); @@ -1172,7 +1215,7 @@ int nfs_writeback_done(struct rpc_task * */ status = NFS_PROTO(data->inode)->write_done(task, data); if (status != 0) - return status; + goto out; nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count); #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) @@ -1229,6 +1272,8 @@ int nfs_writeback_done(struct rpc_task * task->tk_status = -EIO; } nfs4_sequence_free_slot(server->nfs_client, &data->res.seq_res); +out: + nfs_dec_woutstanding(data->inode); return 0; } @@ -1591,6 +1636,24 @@ int nfs_wb_page(struct inode *inode, str return nfs_wb_page_priority(inode, page, FLUSH_STABLE); } +int nfs_wb_eager(struct inode *inode) +{ + struct address_space *mapping = inode->i_mapping; + struct writeback_control wbc = { + .bdi = mapping->backing_dev_info, + .sync_mode = WB_SYNC_NONE, + .nr_to_write = LONG_MAX, + .range_start = 0, + .range_end = LLONG_MAX, + }; + int ret; + + ret = nfs_writepages(mapping, &wbc); + if (ret < 0) + __mark_inode_dirty(inode, I_DIRTY_PAGES); + return ret; +} + #ifdef CONFIG_MIGRATION int nfs_migrate_page(struct address_space *mapping, struct page *newpage, struct page *page) @@ -1674,4 +1737,3 @@ void nfs_destroy_writepagecache(void) mempool_destroy(nfs_wdata_mempool); kmem_cache_destroy(nfs_wdata_cachep); } - diff -rupN -X linux-2.6.32-rc7/Documentation/dontdiff linux-2.6.32-rc7-orig/include/linux/nfs_fs.h linux-2.6.32-rc7/include/linux/nfs_fs.h --- linux-2.6.32-rc7-orig/include/linux/nfs_fs.h 2009-11-12 19:46:07.000000000 -0500 +++ linux-2.6.32-rc7/include/linux/nfs_fs.h 2009-11-13 11:36:43.982136105 -0500 @@ -166,6 +166,7 @@ struct nfs_inode { struct radix_tree_root nfs_page_tree; unsigned long npages; + atomic_t ndirty; /* Open contexts for shared mmap writes */ struct list_head open_files; @@ -187,6 +188,11 @@ struct nfs_inode { #ifdef CONFIG_NFS_FSCACHE struct fscache_cookie *fscache; #endif + + loff_t wrpos; + atomic_t writes; + wait_queue_head_t writes_wq; + struct inode vfs_inode; }; @@ -467,11 +473,13 @@ extern void nfs_unblock_sillyrename(stru * linux/fs/nfs/write.c */ extern int nfs_congestion_kb; +extern int nfs_max_woutstanding; extern int nfs_writepage(struct page *page, struct writeback_control *wbc); extern int nfs_writepages(struct address_space *, struct writeback_control *); extern int nfs_flush_incompatible(struct file *file, struct page *page); extern int nfs_updatepage(struct file *, struct page *, unsigned int, unsigned int); extern int nfs_writeback_done(struct rpc_task *, struct nfs_write_data *); +extern void nfs_wait_woutstanding(struct inode *); /* * Try to write back everything synchronously (but check the @@ -482,6 +490,7 @@ extern int nfs_wb_all(struct inode *inod extern int nfs_wb_nocommit(struct inode *inode); extern int nfs_wb_page(struct inode *inode, struct page* page); extern int nfs_wb_page_cancel(struct inode *inode, struct page* page); +extern int nfs_wb_eager(struct inode *inode); #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) extern int nfs_commit_inode(struct inode *, int); extern struct nfs_write_data *nfs_commitdata_alloc(void); diff -rupN -X linux-2.6.32-rc7/Documentation/dontdiff linux-2.6.32-rc7-orig/mm/page-writeback.c linux-2.6.32-rc7/mm/page-writeback.c --- linux-2.6.32-rc7-orig/mm/page-writeback.c 2009-11-12 19:46:07.000000000 -0500 +++ linux-2.6.32-rc7/mm/page-writeback.c 2009-11-18 10:05:22.314373138 -0500 @@ -536,7 +536,7 @@ static void balance_dirty_pages(struct a * threshold otherwise wait until the disk writes catch * up. */ - if (bdi_nr_reclaimable > bdi_thresh) { + if (bdi_nr_reclaimable != 0) { writeback_inodes_wbc(&wbc); pages_written += write_chunk - wbc.nr_to_write; get_dirty_limits(&background_thresh, &dirty_thresh, -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/