Merge git://git.kernel.org/pub/scm/linux/kernel/git/bunk/trivial
[safe/jmp/linux-2.6] / fs / nfs / write.c
index 875f5b0..bca5734 100644 (file)
@@ -46,7 +46,6 @@
  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
  */
 
-#include <linux/config.h>
 #include <linux/types.h>
 #include <linux/slab.h>
 #include <linux/mm.h>
@@ -86,23 +85,22 @@ static const struct rpc_call_ops nfs_write_full_ops;
 static const struct rpc_call_ops nfs_commit_ops;
 
 static kmem_cache_t *nfs_wdata_cachep;
-mempool_t *nfs_wdata_mempool;
+static mempool_t *nfs_wdata_mempool;
 static mempool_t *nfs_commit_mempool;
 
 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
 
-static inline struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount)
+struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount)
 {
        struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
 
        if (p) {
                memset(p, 0, sizeof(*p));
                INIT_LIST_HEAD(&p->pages);
-               if (pagecount < NFS_PAGEVEC_SIZE)
-                       p->pagevec = &p->page_array[0];
+               if (pagecount <= ARRAY_SIZE(p->page_array))
+                       p->pagevec = p->page_array;
                else {
-                       size_t size = ++pagecount * sizeof(struct page *);
-                       p->pagevec = kzalloc(size, GFP_NOFS);
+                       p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
                        if (!p->pagevec) {
                                mempool_free(p, nfs_commit_mempool);
                                p = NULL;
@@ -112,13 +110,40 @@ static inline struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount)
        return p;
 }
 
-static inline void nfs_commit_free(struct nfs_write_data *p)
+void nfs_commit_free(struct nfs_write_data *p)
 {
        if (p && (p->pagevec != &p->page_array[0]))
                kfree(p->pagevec);
        mempool_free(p, nfs_commit_mempool);
 }
 
+struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
+{
+       struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
+
+       if (p) {
+               memset(p, 0, sizeof(*p));
+               INIT_LIST_HEAD(&p->pages);
+               if (pagecount <= ARRAY_SIZE(p->page_array))
+                       p->pagevec = p->page_array;
+               else {
+                       p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
+                       if (!p->pagevec) {
+                               mempool_free(p, nfs_wdata_mempool);
+                               p = NULL;
+                       }
+               }
+       }
+       return p;
+}
+
+void nfs_writedata_free(struct nfs_write_data *p)
+{
+       if (p && (p->pagevec != &p->page_array[0]))
+               kfree(p->pagevec);
+       mempool_free(p, nfs_wdata_mempool);
+}
+
 void nfs_writedata_release(void *wdata)
 {
        nfs_writedata_free(wdata);
@@ -399,6 +424,7 @@ static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
                if (nfs_have_delegation(inode, FMODE_WRITE))
                        nfsi->change_attr++;
        }
+       SetPagePrivate(req->wb_page);
        nfsi->npages++;
        atomic_inc(&req->wb_count);
        return 0;
@@ -415,6 +441,7 @@ static void nfs_inode_remove_request(struct nfs_page *req)
        BUG_ON (!NFS_WBACK_BUSY(req));
 
        spin_lock(&nfsi->req_lock);
+       ClearPagePrivate(req->wb_page);
        radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
        nfsi->npages--;
        if (!nfsi->npages) {
@@ -469,7 +496,7 @@ nfs_mark_request_dirty(struct nfs_page *req)
        nfs_list_add_request(req, &nfsi->dirty);
        nfsi->ndirty++;
        spin_unlock(&nfsi->req_lock);
-       inc_page_state(nr_dirty);
+       inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
        mark_inode_dirty(inode);
 }
 
@@ -497,7 +524,7 @@ nfs_mark_request_commit(struct nfs_page *req)
        nfs_list_add_request(req, &nfsi->commit);
        nfsi->ncommit++;
        spin_unlock(&nfsi->req_lock);
-       inc_page_state(nr_unstable);
+       inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
        mark_inode_dirty(inode);
 }
 #endif
@@ -507,8 +534,7 @@ nfs_mark_request_commit(struct nfs_page *req)
  *
  * Interruptible by signals only if mounted with intr flag.
  */
-static int
-nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
+static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 {
        struct nfs_inode *nfsi = NFS_I(inode);
        struct nfs_page *req;
@@ -521,7 +547,6 @@ nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int
        else
                idx_end = idx_start + npages - 1;
 
-       spin_lock(&nfsi->req_lock);
        next = idx_start;
        while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
                if (req->wb_index > idx_end)
@@ -534,15 +559,36 @@ nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int
                spin_unlock(&nfsi->req_lock);
                error = nfs_wait_on_request(req);
                nfs_release_request(req);
+               spin_lock(&nfsi->req_lock);
                if (error < 0)
                        return error;
-               spin_lock(&nfsi->req_lock);
                res++;
        }
-       spin_unlock(&nfsi->req_lock);
        return res;
 }
 
+static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
+{
+       struct nfs_inode *nfsi = NFS_I(inode);
+       int ret;
+
+       spin_lock(&nfsi->req_lock);
+       ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
+       spin_unlock(&nfsi->req_lock);
+       return ret;
+}
+
+static void nfs_cancel_requests(struct list_head *head)
+{
+       struct nfs_page *req;
+       while(!list_empty(head)) {
+               req = nfs_list_entry(head->next);
+               nfs_list_remove_request(req);
+               nfs_inode_remove_request(req);
+               nfs_clear_page_writeback(req);
+       }
+}
+
 /*
  * nfs_scan_dirty - Scan an inode for dirty requests
  * @inode: NFS inode to scan
@@ -562,7 +608,6 @@ nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_sta
        if (nfsi->ndirty != 0) {
                res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
                nfsi->ndirty -= res;
-               sub_page_state(nr_dirty,res);
                if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
                        printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
        }
@@ -587,13 +632,18 @@ nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_st
        int res = 0;
 
        if (nfsi->ncommit != 0) {
-               res = nfs_scan_list(&nfsi->commit, dst, idx_start, npages);
+               res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
                nfsi->ncommit -= res;
                if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
                        printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
        }
        return res;
 }
+#else
+static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
+{
+       return 0;
+}
 #endif
 
 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
@@ -931,7 +981,7 @@ static void nfs_execute_write(struct nfs_write_data *data)
  * Generate multiple small requests to write out a single
  * contiguous dirty area on one page.
  */
-static int nfs_flush_multi(struct list_head *head, struct inode *inode, int how)
+static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 {
        struct nfs_page *req = nfs_list_entry(head->next);
        struct page *page = req->wb_page;
@@ -1000,16 +1050,13 @@ out_bad:
  * This is the case if nfs_updatepage detects a conflicting request
  * that has been written but not committed.
  */
-static int nfs_flush_one(struct list_head *head, struct inode *inode, int how)
+static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
 {
        struct nfs_page         *req;
        struct page             **pages;
        struct nfs_write_data   *data;
        unsigned int            count;
 
-       if (NFS_SERVER(inode)->wsize < PAGE_CACHE_SIZE)
-               return nfs_flush_multi(head, inode, how);
-
        data = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
        if (!data)
                goto out_bad;
@@ -1042,24 +1089,32 @@ static int nfs_flush_one(struct list_head *head, struct inode *inode, int how)
        return -ENOMEM;
 }
 
-static int
-nfs_flush_list(struct list_head *head, int wpages, int how)
+static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
 {
        LIST_HEAD(one_request);
-       struct nfs_page         *req;
-       int                     error = 0;
-       unsigned int            pages = 0;
+       int (*flush_one)(struct inode *, struct list_head *, int);
+       struct nfs_page *req;
+       int wpages = NFS_SERVER(inode)->wpages;
+       int wsize = NFS_SERVER(inode)->wsize;
+       int error;
 
-       while (!list_empty(head)) {
-               pages += nfs_coalesce_requests(head, &one_request, wpages);
+       flush_one = nfs_flush_one;
+       if (wsize < PAGE_CACHE_SIZE)
+               flush_one = nfs_flush_multi;
+       /* For single writes, FLUSH_STABLE is more efficient */
+       if (npages <= wpages && npages == NFS_I(inode)->npages
+                       && nfs_list_entry(head->next)->wb_bytes <= wsize)
+               how |= FLUSH_STABLE;
+
+       do {
+               nfs_coalesce_requests(head, &one_request, wpages);
                req = nfs_list_entry(one_request.next);
-               error = nfs_flush_one(&one_request, req->wb_context->dentry->d_inode, how);
+               error = flush_one(inode, &one_request, how);
                if (error < 0)
-                       break;
-       }
-       if (error >= 0)
-               return pages;
-
+                       goto out_err;
+       } while (!list_empty(head));
+       return 0;
+out_err:
        while (!list_empty(head)) {
                req = nfs_list_entry(head->next);
                nfs_list_remove_request(req);
@@ -1337,7 +1392,6 @@ static void nfs_commit_done(struct rpc_task *task, void *calldata)
 {
        struct nfs_write_data   *data = calldata;
        struct nfs_page         *req;
-       int res = 0;
 
         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
                                 task->tk_pid, task->tk_status);
@@ -1349,6 +1403,7 @@ static void nfs_commit_done(struct rpc_task *task, void *calldata)
        while (!list_empty(&data->pages)) {
                req = nfs_list_entry(data->pages.next);
                nfs_list_remove_request(req);
+               dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 
                dprintk("NFS: commit (%s/%Ld %d@%Ld)",
                        req->wb_context->dentry->d_inode->i_sb->s_id,
@@ -1375,15 +1430,18 @@ static void nfs_commit_done(struct rpc_task *task, void *calldata)
                nfs_mark_request_dirty(req);
        next:
                nfs_clear_page_writeback(req);
-               res++;
        }
-       sub_page_state(nr_unstable,res);
 }
 
 static const struct rpc_call_ops nfs_commit_ops = {
        .rpc_call_done = nfs_commit_done,
        .rpc_release = nfs_commit_release,
 };
+#else
+static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
+{
+       return 0;
+}
 #endif
 
 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
@@ -1391,24 +1449,16 @@ static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
 {
        struct nfs_inode *nfsi = NFS_I(inode);
        LIST_HEAD(head);
-       int                     res,
-                               error = 0;
+       int res;
 
        spin_lock(&nfsi->req_lock);
        res = nfs_scan_dirty(inode, &head, idx_start, npages);
        spin_unlock(&nfsi->req_lock);
        if (res) {
-               struct nfs_server *server = NFS_SERVER(inode);
-
-               /* For single writes, FLUSH_STABLE is more efficient */
-               if (res == nfsi->npages && nfsi->npages <= server->wpages) {
-                       if (res > 1 || nfs_list_entry(head.next)->wb_bytes <= server->wsize)
-                               how |= FLUSH_STABLE;
-               }
-               error = nfs_flush_list(&head, server->wpages, how);
+               int error = nfs_flush_list(inode, &head, res, how);
+               if (error < 0)
+                       return error;
        }
-       if (error < 0)
-               return error;
        return res;
 }
 
@@ -1417,14 +1467,13 @@ int nfs_commit_inode(struct inode *inode, int how)
 {
        struct nfs_inode *nfsi = NFS_I(inode);
        LIST_HEAD(head);
-       int                     res,
-                               error = 0;
+       int res;
 
        spin_lock(&nfsi->req_lock);
        res = nfs_scan_commit(inode, &head, 0, 0);
        spin_unlock(&nfsi->req_lock);
        if (res) {
-               error = nfs_commit_list(inode, &head, how);
+               int error = nfs_commit_list(inode, &head, how);
                if (error < 0)
                        return error;
        }
@@ -1432,31 +1481,51 @@ int nfs_commit_inode(struct inode *inode, int how)
 }
 #endif
 
-int nfs_sync_inode(struct inode *inode, unsigned long idx_start,
-                 unsigned int npages, int how)
+int nfs_sync_inode_wait(struct inode *inode, unsigned long idx_start,
+               unsigned int npages, int how)
 {
+       struct nfs_inode *nfsi = NFS_I(inode);
+       LIST_HEAD(head);
        int nocommit = how & FLUSH_NOCOMMIT;
-       int wait = how & FLUSH_WAIT;
-       int error;
-
-       how &= ~(FLUSH_WAIT|FLUSH_NOCOMMIT);
+       int pages, ret;
 
+       how &= ~FLUSH_NOCOMMIT;
+       spin_lock(&nfsi->req_lock);
        do {
-               if (wait) {
-                       error = nfs_wait_on_requests(inode, idx_start, npages);
-                       if (error != 0)
-                               continue;
+               ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
+               if (ret != 0)
+                       continue;
+               pages = nfs_scan_dirty(inode, &head, idx_start, npages);
+               if (pages != 0) {
+                       spin_unlock(&nfsi->req_lock);
+                       if (how & FLUSH_INVALIDATE)
+                               nfs_cancel_requests(&head);
+                       else
+                               ret = nfs_flush_list(inode, &head, pages, how);
+                       spin_lock(&nfsi->req_lock);
+                       continue;
                }
-               error = nfs_flush_inode(inode, idx_start, npages, how);
-               if (error != 0)
+               if (nocommit)
+                       break;
+               pages = nfs_scan_commit(inode, &head, idx_start, npages);
+               if (pages == 0)
+                       break;
+               if (how & FLUSH_INVALIDATE) {
+                       spin_unlock(&nfsi->req_lock);
+                       nfs_cancel_requests(&head);
+                       spin_lock(&nfsi->req_lock);
                        continue;
-               if (!nocommit)
-                       error = nfs_commit_inode(inode, how);
-       } while (error > 0);
-       return error;
+               }
+               pages += nfs_scan_commit(inode, &head, 0, 0);
+               spin_unlock(&nfsi->req_lock);
+               ret = nfs_commit_list(inode, &head, how);
+               spin_lock(&nfsi->req_lock);
+       } while (ret >= 0);
+       spin_unlock(&nfsi->req_lock);
+       return ret;
 }
 
-int nfs_init_writepagecache(void)
+int __init nfs_init_writepagecache(void)
 {
        nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
                                             sizeof(struct nfs_write_data),
@@ -1465,17 +1534,13 @@ int nfs_init_writepagecache(void)
        if (nfs_wdata_cachep == NULL)
                return -ENOMEM;
 
-       nfs_wdata_mempool = mempool_create(MIN_POOL_WRITE,
-                                          mempool_alloc_slab,
-                                          mempool_free_slab,
-                                          nfs_wdata_cachep);
+       nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
+                                                    nfs_wdata_cachep);
        if (nfs_wdata_mempool == NULL)
                return -ENOMEM;
 
-       nfs_commit_mempool = mempool_create(MIN_POOL_COMMIT,
-                                          mempool_alloc_slab,
-                                          mempool_free_slab,
-                                          nfs_wdata_cachep);
+       nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
+                                                     nfs_wdata_cachep);
        if (nfs_commit_mempool == NULL)
                return -ENOMEM;