pgio: fix buffered write retry path

Change nfs_find_and_lock_request so nfs_page_async_flush can handle multiple requests in a page. There is only one request for a page the first time nfs_page_async_flush is called, but if a write or commit fails, async_flush is called again and there may be multiple requests associated with the page. The solution is to merge all the requests in a page group into a single request before calling nfs_pageio_add_request.

Rename nfs_find_and_lock_request to nfs_lock_and_join_requests and change it to first lock all requests for the page, then cancel and merge all subrequests into the head request.

Signed-off-by Weston Andros Adamson <[email protected]>

---
 fs/nfs/internal.h |   1 +
 fs/nfs/pagelist.c |   4 +-
 fs/nfs/write.c    | 254 +++++++++++++++++++++++++++++++++++++++++++++++++-----
 3 files changed, 234 insertions(+), 25 deletions(-)

fs/nfs/internal.h +1 -0

index da36257..2f19e83 100644
@@ -244,6 +244,7 @@ void nfs_pgio_data_destroy(struct nfs_pgio_header *);
 int nfs_generic_pgio(struct nfs_pageio_descriptor *, struct nfs_pgio_header *);
 int nfs_initiate_pgio(struct rpc_clnt *, struct nfs_pgio_header *,
 		      const struct rpc_call_ops *, int, int);
+void nfs_free_request(struct nfs_page *req);
 
 static inline void nfs_iocounter_init(struct nfs_io_counter *c)
 {

fs/nfs/pagelist.c +1 -3

index 8b074da..a22c130 100644
@@ -29,8 +29,6 @@
 static struct kmem_cache *nfs_page_cachep;
 static const struct rpc_call_ops nfs_pgio_common_ops;
 
-static void nfs_free_request(struct nfs_page *);
-
 static bool nfs_pgarray_set(struct nfs_page_array *p, unsigned int pagecount)
 {
 	p->npages = pagecount;

@@ -406,7 +404,7 @@ static void nfs_clear_request(struct nfs_page *req)
  *
  * Note: Should never be called with the spinlock held!
  */
-static void nfs_free_request(struct nfs_page *req)
+void nfs_free_request(struct nfs_page *req)
 {
 	WARN_ON_ONCE(req->wb_this_page != req);
 

fs/nfs/write.c +25 -1

index 2e2b9f1..4dab432 100644
@@ -46,6 +46,7 @@ static const struct rpc_call_ops nfs_commit_ops;
 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
 static const struct nfs_rw_ops nfs_rw_write_ops;
+static void nfs_clear_request_commit(struct nfs_page *req);
 
 static struct kmem_cache *nfs_wdata_cachep;
 static mempool_t *nfs_wdata_mempool;

@@ -289,36 +290,245 @@ static void nfs_end_page_writeback(struct nfs_page *req)
 		clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 }
 
-static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
+
+/* nfs_page_group_clear_bits
+ *   @req - an nfs request
+ * clears all page group related bits from @req
+ */
+static void
+nfs_page_group_clear_bits(struct nfs_page *req)
+{
+	clear_bit(PG_TEARDOWN, &req->wb_flags);
+	clear_bit(PG_UNLOCKPAGE, &req->wb_flags);
+	clear_bit(PG_UPTODATE, &req->wb_flags);
+	clear_bit(PG_WB_END, &req->wb_flags);
+	clear_bit(PG_REMOVE, &req->wb_flags);
+}
+
+
+/*
+ * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
+ *
+ * this is a helper function for nfs_lock_and_join_requests
+ *
+ * @inode - inode associated with request page group, must be holding inode lock
+ * @head  - head request of page group, must be holding head lock
+ * @req   - request that couldn't lock and needs to wait on the req bit lock

fs/nfs/write.c +197 -21

+{
+	struct nfs_page *tmp;
+	int ret;
+
+	/* relinquish all the locks successfully grabbed this run */
+	for (tmp = head ; tmp != req; tmp = tmp->wb_this_page)
+		nfs_unlock_request(tmp);
+
+	WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
+
+	/* grab a ref on the request that will be waited on */
+	kref_get(&req->wb_kref);
+
+	nfs_page_group_unlock(head);
+	spin_unlock(&inode->i_lock);
+
+	/* release ref from nfs_page_find_head_request_locked */
+	nfs_release_request(head);
+
+	if (!nonblock)
+		ret = nfs_wait_on_request(req);
+	else
+		ret = -EAGAIN;
+	nfs_release_request(req);
+
+	return ret;
+}
+
+/*
+ * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
+ *
+ * @destroy_list - request list (using wb_this_page) terminated by @old_head
+ * @old_head - the old head of the list
+ *
+ * All subrequests must be locked and removed from all lists, so at this point
+ * they are only "active" in this function, and possibly in nfs_wait_on_request
+ * with a reference held by some other context.
+ */
+static void
+nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
+				 struct nfs_page *old_head)
+{
+	while (destroy_list) {
+		struct nfs_page *subreq = destroy_list;
+
+		destroy_list = (subreq->wb_this_page == old_head) ?
+				   NULL : subreq->wb_this_page;
+
+		WARN_ON_ONCE(old_head != subreq->wb_head);
+
+		/* make sure old group is not used */
+		subreq->wb_head = subreq;
+		subreq->wb_this_page = subreq;
+
+		nfs_clear_request_commit(subreq);
+
+		/* subreq is now totally disconnected from page group or any
+		 * write / commit lists. last chance to wake any waiters */
+		nfs_unlock_request(subreq);
+
+		if (!test_bit(PG_TEARDOWN, &subreq->wb_flags)) {
+			/* release ref on old head request */
+			nfs_release_request(old_head);
+
+			nfs_page_group_clear_bits(subreq);
+
+			/* release the PG_INODE_REF reference */
+			if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags))
+				nfs_release_request(subreq);
+			else
+				WARN_ON_ONCE(1);
+		} else {
+			WARN_ON_ONCE(test_bit(PG_CLEAN, &subreq->wb_flags));
+			/* zombie requests have already released the last
+			 * reference and were waiting on the rest of the
+			 * group to complete. Since it's no longer part of a
+			 * group, simply free the request */
+			nfs_page_group_clear_bits(subreq);
+			nfs_free_request(subreq);
+		}
+	}
+}
+
+/*
+ * nfs_lock_and_join_requests - join all subreqs to the head req and return
+ *                              a locked reference, cancelling any pending
+ *                              operations for this page.
+ *
+ * @page - the page used to lookup the "page group" of nfs_page structures
+ * @nonblock - if true, don't block waiting for request locks
+ *
+ * This function joins all sub requests to the head request by first
+ * locking all requests in the group, cancelling any pending operations
+ * and finally updating the head request to cover the whole range covered by
+ * the (former) group.  All subrequests are removed from any write or commit
+ * lists, unlinked from the group and destroyed.
+ *
+ * Returns a locked, referenced pointer to the head request - which after
+ * this call is guaranteed to be the only request associated with the page.
+ * Returns NULL if no requests are found for @page, or a ERR_PTR if an
+ * error was encountered.
+ */
+static struct nfs_page *
+nfs_lock_and_join_requests(struct page *page, bool nonblock)
 {
 	struct inode *inode = page_file_mapping(page)->host;
-	struct nfs_page *req;
+	struct nfs_page *head, *subreq;
+	struct nfs_page *destroy_list = NULL;
+	unsigned int total_bytes;
 	int ret;
 
+try_again:
+	total_bytes = 0;
+
+	WARN_ON_ONCE(destroy_list);
+
 	spin_lock(&inode->i_lock);
-	for (;;) {
-		req = nfs_page_find_head_request_locked(NFS_I(inode), page);
-		if (req == NULL)
-			break;
-		if (nfs_lock_request(req))
-			break;
-		/* Note: If we hold the page lock, as is the case in nfs_writepage,
-		 *	 then the call to nfs_lock_request() will always
-		 *	 succeed provided that someone hasn't already marked the
-		 *	 request as dirty (in which case we don't care).
-		 */
+
+	/*
+	 * A reference is taken only on the head request which acts as a
+	 * reference to the whole page group - the group will not be destroyed
+	 * until the head reference is released.
+	 */
+	head = nfs_page_find_head_request_locked(NFS_I(inode), page);
+
+	if (!head) {
 		spin_unlock(&inode->i_lock);
-		if (!nonblock)
-			ret = nfs_wait_on_request(req);
-		else
-			ret = -EAGAIN;
-		nfs_release_request(req);
-		if (ret != 0)
+		return NULL;
+	}
+
+	/* lock each request in the page group */
+	nfs_page_group_lock(head);
+	subreq = head;
+	do {
+		/*
+		 * Subrequests are always contiguous, non overlapping
+		 * and in order. If not, it's a programming error.
+		 */
+		WARN_ON_ONCE(subreq->wb_offset !=
+		     (head->wb_offset + total_bytes));
+
+		/* keep track of how many bytes this group covers */
+		total_bytes += subreq->wb_bytes;
+
+		if (!nfs_lock_request(subreq)) {
+			/* releases page group bit lock and
+			 * inode spin lock and all references */
+			ret = nfs_unroll_locks_and_wait(inode, head,
+				subreq, nonblock);
+
+			if (ret == 0)
+				goto try_again;
+
 			return ERR_PTR(ret);
-		spin_lock(&inode->i_lock);
+		}
+
+		subreq = subreq->wb_this_page;
+	} while (subreq != head);
+
+	/* Now that all requests are locked, make sure they aren't on any list.
+	 * Commit list removal accounting is done after locks are dropped */
+	subreq = head;
+	do {
+		nfs_list_remove_request(subreq);
+		subreq = subreq->wb_this_page;
+	} while (subreq != head);
+
+	/* unlink subrequests from head, destroy them later */
+	if (head->wb_this_page != head) {
+		/* destroy list will be terminated by head */
+		destroy_list = head->wb_this_page;
+		head->wb_this_page = head;
+
+		/* change head request to cover whole range that
+		 * the former page group covered */
+		head->wb_bytes = total_bytes;
 	}
+
+	/*
+	 * prepare head request to be added to new pgio descriptor
+	 */
+	nfs_page_group_clear_bits(head);
+
+	/*
+	 * some part of the group was still on the inode list - otherwise
+	 * the group wouldn't be involved in async write.
+	 * grab a reference for the head request, iff it needs one.
+	 */
+	if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags))
+		kref_get(&head->wb_kref);
+
+	nfs_page_group_unlock(head);
+
+	/* drop lock to clear_request_commit the head req and clean up
+	 * requests on destroy list */
 	spin_unlock(&inode->i_lock);
-	return req;
+
+	nfs_destroy_unlinked_subrequests(destroy_list, head);
+
+	/* clean up commit list state */
+	nfs_clear_request_commit(head);
+
+	/* still holds ref on head from nfs_page_find_head_request_locked
+	 * and still has lock on head from lock loop */
+	return head;
 }
 
 /*

@@ -331,7 +541,7 @@ static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 	struct nfs_page *req;
 	int ret = 0;
 
-	req = nfs_find_and_lock_request(page, nonblock);
+	req = nfs_lock_and_join_requests(page, nonblock);
 	if (!req)
 		goto out;
 	ret = PTR_ERR(req);

My recent pgio work added the ability to split requests into sub-page regions, but didn't handle a few places in the writeback code where requests are looked up by struct page and may already be split into multiple requests.

This patchset adds a function "nfs_lock_and_join_requests" in patch "nfs: handle multiple reqs in nfs_page_async_flush", which:
- takes mutex lock
- looks up head request
- grabs request lock for each subrequest
- if unsuccessful, unrolls old locks and waits on subrequest
- removes all requests from commit lists
- merges range of subrequests into the head requests
- unlinks and destroys the old subrequests.

The other patches are related fixes.

The problem showed up when mounting with wsize < PAGE_SIZE - this would cause multiple requests per page. If a commit failed, nfs_page_async_flush would operate just on the head request, leading to a hang.

The nfs_wb_page_cancel patch leverages the same function - nfs_lock_and_join_requests cancels all operations on the page group. I've had a really hard time testing nfs_wb_page_cancel, I've only hit it once in weeks of testing. Any ideas on how to reliably trigger this is appreciated - it's not as easy as just kicking off a ton of writeback then truncating. The one time I did see it was with a ton of i/o on a VM with 256M of RAM, which was swapping like crazy, along with restarting the server repeatedly (to get commit verifier mismatch).