[PATCH 1/5] NFS: Memory allocation failures are not server fatal errors

[PATCH 1/5] NFS: Memory allocation failures are not server fatal errors

[trondmy]

From: Trond Myklebust <trond.myklebust@hammerspace.com>

We need to filter out ENOMEM in nfs_error_is_fatal_on_server(), because
running out of memory on our client is not a server error.

Reported-by: Olga Kornievskaia <aglo@umich.edu>
Fixes: 2dc23afffbca ("NFS: ENOMEM should also be a fatal error.")
Cc: stable@vger.kernel.org
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
---
 fs/nfs/internal.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/fs/nfs/internal.h b/fs/nfs/internal.h
index 7eefa16ed381..8f8cd6e2d4db 100644
--- a/fs/nfs/internal.h
+++ b/fs/nfs/internal.h
@@ -841,6 +841,7 @@ static inline bool nfs_error_is_fatal_on_server(int err)
 	case 0:
 	case -ERESTARTSYS:
 	case -EINTR:
+	case -ENOMEM:
 		return false;
 	}
 	return nfs_error_is_fatal(err);
-- 
2.36.1

[PATCH 2/5] NFSv4/pNFS: Do not fail I/O when we fail to allocate the pNFS layout

[trondmy]

From: Trond Myklebust <trond.myklebust@hammerspace.com>

Commit 587f03deb69b caused pnfs_update_layout() to stop returning ENOMEM
when the memory allocation fails, and hence causes it to fall back to
trying to do I/O through the MDS. There is no guarantee that this will
fare any better. If we're failing the pNFS layout allocation, then we
should just redirty the page and retry later.

Reported-by: Olga Kornievskaia <aglo@umich.edu>
Fixes: 587f03deb69b ("pnfs: refactor send_layoutget")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
---
 fs/nfs/pnfs.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c
index 856c962273c7..68a87be3e6f9 100644
--- a/fs/nfs/pnfs.c
+++ b/fs/nfs/pnfs.c
@@ -2000,6 +2000,7 @@ pnfs_update_layout(struct inode *ino,
 	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
 	if (lo == NULL) {
 		spin_unlock(&ino->i_lock);
+		lseg = ERR_PTR(-ENOMEM);
 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
 				 PNFS_UPDATE_LAYOUT_NOMEM);
 		goto out;
@@ -2128,6 +2129,7 @@ pnfs_update_layout(struct inode *ino,
 
 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
 	if (!lgp) {
+		lseg = ERR_PTR(-ENOMEM);
 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
 					 PNFS_UPDATE_LAYOUT_NOMEM);
 		nfs_layoutget_end(lo);
-- 
2.36.1

[PATCH 3/5] NFS: Further fixes to the writeback error handling

[trondmy]

From: Trond Myklebust <trond.myklebust@hammerspace.com>

When we handle an error by redirtying the page, we're not corrupting the
mapping, so we don't want the error to be recorded in the mapping.
If the caller has specified a sync_mode of WB_SYNC_NONE, we can just
return AOP_WRITEPAGE_ACTIVATE. However if we're dealing with
WB_SYNC_ALL, we need to ensure that retries happen when the errors are
non-fatal.

Reported-by: Olga Kornievskaia <aglo@umich.edu>
Fixes: 8fc75bed96bb ("NFS: Fix up return value on fatal errors in nfs_page_async_flush()")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
---
 fs/nfs/write.c | 39 ++++++++++++++++++---------------------
 1 file changed, 18 insertions(+), 21 deletions(-)

diff --git a/fs/nfs/write.c b/fs/nfs/write.c
index f00d45cf80ef..a8eb348947a6 100644
--- a/fs/nfs/write.c
+++ b/fs/nfs/write.c
@@ -603,8 +603,9 @@ static void nfs_write_error(struct nfs_page *req, int error)
  * Find an associated nfs write request, and prepare to flush it out
  * May return an error if the user signalled nfs_wait_on_request().
  */
-static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
-				struct page *page)
+static int nfs_page_async_flush(struct page *page,
+				struct writeback_control *wbc,
+				struct nfs_pageio_descriptor *pgio)
 {
 	struct nfs_page *req;
 	int ret = 0;
@@ -630,11 +631,11 @@ static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 		/*
 		 * Remove the problematic req upon fatal errors on the server
 		 */
-		if (nfs_error_is_fatal(ret)) {
-			if (nfs_error_is_fatal_on_server(ret))
-				goto out_launder;
-		} else
-			ret = -EAGAIN;
+		if (nfs_error_is_fatal_on_server(ret))
+			goto out_launder;
+		if (wbc->sync_mode == WB_SYNC_NONE)
+			ret = AOP_WRITEPAGE_ACTIVATE;
+		redirty_page_for_writepage(wbc, page);
 		nfs_redirty_request(req);
 		pgio->pg_error = 0;
 	} else
@@ -650,15 +651,8 @@ static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 			    struct nfs_pageio_descriptor *pgio)
 {
-	int ret;
-
 	nfs_pageio_cond_complete(pgio, page_index(page));
-	ret = nfs_page_async_flush(pgio, page);
-	if (ret == -EAGAIN) {
-		redirty_page_for_writepage(wbc, page);
-		ret = AOP_WRITEPAGE_ACTIVATE;
-	}
-	return ret;
+	return nfs_page_async_flush(page, wbc, pgio);
 }
 
 /*
@@ -737,12 +731,15 @@ int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 		priority = wb_priority(wbc);
 	}
 
-	nfs_pageio_init_write(&pgio, inode, priority, false,
-				&nfs_async_write_completion_ops);
-	pgio.pg_io_completion = ioc;
-	err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
-	pgio.pg_error = 0;
-	nfs_pageio_complete(&pgio);
+	do {
+		nfs_pageio_init_write(&pgio, inode, priority, false,
+				      &nfs_async_write_completion_ops);
+		pgio.pg_io_completion = ioc;
+		err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
+					&pgio);
+		pgio.pg_error = 0;
+		nfs_pageio_complete(&pgio);
+	} while (err < 0 && !nfs_error_is_fatal(err));
 	nfs_io_completion_put(ioc);
 
 	if (err < 0)
-- 
2.36.1

[PATCH 4/5] pNFS/files: Fall back to I/O through the MDS on non-fatal layout errors

[trondmy]

From: Trond Myklebust <trond.myklebust@hammerspace.com>

Only report the error when the server is returning a fatal error, such
as ESTALE, EIO, etc...

Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
---
 fs/nfs/filelayout/filelayout.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/fs/nfs/filelayout/filelayout.c b/fs/nfs/filelayout/filelayout.c
index 76deddab0a8f..2b2661582bbe 100644
--- a/fs/nfs/filelayout/filelayout.c
+++ b/fs/nfs/filelayout/filelayout.c
@@ -839,7 +839,12 @@ fl_pnfs_update_layout(struct inode *ino,
 
 	lseg = pnfs_update_layout(ino, ctx, pos, count, iomode, strict_iomode,
 				  gfp_flags);
-	if (IS_ERR_OR_NULL(lseg))
+	if (IS_ERR(lseg)) {
+		/* Fall back to MDS on recoverable errors */
+		if (!nfs_error_is_fatal_on_server(PTR_ERR(lseg)))
+			lseg = NULL;
+		goto out;
+	} else if (!lseg)
 		goto out;
 
 	lo = NFS_I(ino)->layout;
-- 
2.36.1

[PATCH 5/5] NFSv4: Don't hold the layoutget locks across multiple RPC calls

[trondmy]

From: Trond Myklebust <trond.myklebust@hammerspace.com>

When doing layoutget as part of the open() compound, we have to be
careful to release the layout locks before we can call any further RPC
calls, such as setattr(). The reason is that those calls could trigger
a recall, which could deadlock.

Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
---
 fs/nfs/nfs4proc.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
index a79f66432bd3..bf3ba541b9fb 100644
--- a/fs/nfs/nfs4proc.c
+++ b/fs/nfs/nfs4proc.c
@@ -3098,6 +3098,10 @@ static int _nfs4_open_and_get_state(struct nfs4_opendata *opendata,
 	}
 
 out:
+	if (opendata->lgp) {
+		nfs4_lgopen_release(opendata->lgp);
+		opendata->lgp = NULL;
+	}
 	if (!opendata->cancelled)
 		nfs4_sequence_free_slot(&opendata->o_res.seq_res);
 	return ret;
-- 
2.36.1