Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Anna Schumaker <anna@kernel.org>
To:     linux-nfs@vger.kernel.org, trond.myklebust@hammerspace.com
Cc:     anna@kernel.org
Subject: [PATCH v2 3/3] NFSv4.2: Rework scratch handling for READ_PLUS (again)
Date:   Fri,  9 Jun 2023 15:46:13 -0400
Message-ID: <20230609194613.848590-3-anna@kernel.org>
In-Reply-To: <20230609194613.848590-1-anna@kernel.org>
References: <20230609194613.848590-1-anna@kernel.org>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

From: Anna Schumaker <Anna.Schumaker@Netapp.com>

I found that the read code might send multiple requests using the same
nfs_pgio_header, but nfs4_proc_read_setup() is only called once. This is
how we ended up occasionally double-freeing the scratch buffer, but also
means we set a NULL pointer but non-zero length to the xdr scratch
buffer. This results in an oops the first time decoding needs to copy
something to scratch, which frequently happens when decoding READ_PLUS
hole segments.

I fix this by moving scratch handling into the pageio read code. I
provide a function to allocate scratch space for decoding read replies,
and free the scratch buffer when the nfs_pgio_header is freed.

Krzysztof Kozlowski hit a bug a while ago with similar symptoms,
and I'm hopeful that this patch fixes his issue.

Reported-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Fixes: fbd2a05f29a9 (NFSv4.2: Rework scratch handling for READ_PLUS)
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>

pick 7de817656b9b NFSD: Repeal and replace the READ_PLUS implementation
---
 fs/nfs/internal.h |  1 +
 fs/nfs/nfs42.h    |  1 +
 fs/nfs/nfs42xdr.c |  2 +-
 fs/nfs/nfs4proc.c | 13 +------------
 fs/nfs/read.c     | 10 ++++++++++
 5 files changed, 14 insertions(+), 13 deletions(-)

diff --git a/fs/nfs/internal.h b/fs/nfs/internal.h
index 3cc027d3bd58..1607c23f68d4 100644
--- a/fs/nfs/internal.h
+++ b/fs/nfs/internal.h
@@ -489,6 +489,7 @@ extern const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
 extern void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
 			struct inode *inode, bool force_mds,
 			const struct nfs_pgio_completion_ops *compl_ops);
+extern bool nfs_read_alloc_scratch(struct nfs_pgio_header *hdr, size_t size);
 extern int nfs_read_add_folio(struct nfs_pageio_descriptor *pgio,
 			       struct nfs_open_context *ctx,
 			       struct folio *folio);
diff --git a/fs/nfs/nfs42.h b/fs/nfs/nfs42.h
index 0fe5aacbcfdf..b59876b01a1e 100644
--- a/fs/nfs/nfs42.h
+++ b/fs/nfs/nfs42.h
@@ -13,6 +13,7 @@
  * more? Need to consider not to pre-alloc too much for a compound.
  */
 #define PNFS_LAYOUTSTATS_MAXDEV (4)
+#define READ_PLUS_SCRATCH_SIZE (16)
 
 /* nfs4.2proc.c */
 #ifdef CONFIG_NFS_V4_2
diff --git a/fs/nfs/nfs42xdr.c b/fs/nfs/nfs42xdr.c
index 75765382cc0e..20aa5e746497 100644
--- a/fs/nfs/nfs42xdr.c
+++ b/fs/nfs/nfs42xdr.c
@@ -1351,7 +1351,7 @@ static int nfs4_xdr_dec_read_plus(struct rpc_rqst *rqstp,
 	struct compound_hdr hdr;
 	int status;
 
-	xdr_set_scratch_buffer(xdr, res->scratch, sizeof(res->scratch));
+	xdr_set_scratch_buffer(xdr, res->scratch, READ_PLUS_SCRATCH_SIZE);
 
 	status = decode_compound_hdr(xdr, &hdr);
 	if (status)
diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
index d3665390c4cb..73dc8a793ae9 100644
--- a/fs/nfs/nfs4proc.c
+++ b/fs/nfs/nfs4proc.c
@@ -5437,18 +5437,8 @@ static bool nfs4_read_plus_not_supported(struct rpc_task *task,
 	return false;
 }
 
-static inline void nfs4_read_plus_scratch_free(struct nfs_pgio_header *hdr)
-{
-	if (hdr->res.scratch) {
-		kfree(hdr->res.scratch);
-		hdr->res.scratch = NULL;
-	}
-}
-
 static int nfs4_read_done(struct rpc_task *task, struct nfs_pgio_header *hdr)
 {
-	nfs4_read_plus_scratch_free(hdr);
-
 	if (!nfs4_sequence_done(task, &hdr->res.seq_res))
 		return -EAGAIN;
 	if (nfs4_read_stateid_changed(task, &hdr->args))
@@ -5468,8 +5458,7 @@ static bool nfs42_read_plus_support(struct nfs_pgio_header *hdr,
 	/* Note: We don't use READ_PLUS with pNFS yet */
 	if (nfs_server_capable(hdr->inode, NFS_CAP_READ_PLUS) && !hdr->ds_clp) {
 		msg->rpc_proc = &nfs4_procedures[NFSPROC4_CLNT_READ_PLUS];
-		hdr->res.scratch = kmalloc(32, GFP_KERNEL);
-		return hdr->res.scratch != NULL;
+		return nfs_read_alloc_scratch(hdr, READ_PLUS_SCRATCH_SIZE);
 	}
 	return false;
 }
diff --git a/fs/nfs/read.c b/fs/nfs/read.c
index f71eeee67e20..7dc21a48e3e7 100644
--- a/fs/nfs/read.c
+++ b/fs/nfs/read.c
@@ -47,6 +47,8 @@ static struct nfs_pgio_header *nfs_readhdr_alloc(void)
 
 static void nfs_readhdr_free(struct nfs_pgio_header *rhdr)
 {
+	if (rhdr->res.scratch != NULL)
+		kfree(rhdr->res.scratch);
 	kmem_cache_free(nfs_rdata_cachep, rhdr);
 }
 
@@ -108,6 +110,14 @@ void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
 }
 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
 
+bool nfs_read_alloc_scratch(struct nfs_pgio_header *hdr, size_t size)
+{
+	WARN_ON(hdr->res.scratch != NULL);
+	hdr->res.scratch = kmalloc(size, GFP_KERNEL);
+	return hdr->res.scratch != NULL;
+}
+EXPORT_SYMBOL_GPL(nfs_read_alloc_scratch);
+
 static void nfs_readpage_release(struct nfs_page *req, int error)
 {
 	struct folio *folio = nfs_page_to_folio(req);
-- 
2.41.0