Whamcloud - gitweb
LU-3333 ptlrpc: Protect request buffer changing 67/6467/5
authorOleg Drokin <oleg.drokin@intel.com>
Mon, 27 May 2013 18:06:09 +0000 (14:06 -0400)
committerOleg Drokin <oleg.drokin@intel.com>
Thu, 17 Apr 2014 15:16:44 +0000 (15:16 +0000)
*_enlarge_reqbuf class of functions can change request body location
for a request that's already in replay list, as such a parallel
traverser of the list (after_reply -> ptlrpc_free_committed) might
access freed and scrambled memory causing assertions.

Separate out the common code in ptlrpc_enlarge_req_buffer and
protect the buffer replacement by imp_lock since all users
that can get to this request must be holding this lock themselves
first.

Change-Id: I051d96196b44faafc15c58e0b2127856929f0ba7
Signed-off-by: Oleg Drokin <oleg.drokin@intel.com>
Reviewed-on: http://review.whamcloud.com/6467
Tested-by: Jenkins
Reviewed-by: Mike Pershin <mike.pershin@intel.com>
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Alexey Lyashkov <alexey_lyashkov@xyratex.com>
lustre/ptlrpc/gss/sec_gss.c
lustre/ptlrpc/pack_generic.c
lustre/ptlrpc/ptlrpc_internal.h
lustre/ptlrpc/sec_null.c
lustre/ptlrpc/sec_plain.c

index eb9debb..69afcf0 100644 (file)
@@ -68,6 +68,7 @@
 #include <lustre_import.h>
 #include <lustre_sec.h>
 
+#include "../ptlrpc_internal.h"
 #include "gss_err.h"
 #include "gss_internal.h"
 #include "gss_api.h"
@@ -1651,9 +1652,9 @@ int gss_enlarge_reqbuf_intg(struct ptlrpc_sec *sec,
                             int svc,
                             int segment, int newsize)
 {
-        struct lustre_msg      *newbuf;
         int                     txtsize, sigsize = 0, i;
         int                     newmsg_size, newbuf_size;
+       int                     rc;
 
         /*
          * gss header is at seg 0;
@@ -1697,19 +1698,10 @@ int gss_enlarge_reqbuf_intg(struct ptlrpc_sec *sec,
         LASSERT(!req->rq_pool || req->rq_reqbuf_len >= newbuf_size);
 
         if (req->rq_reqbuf_len < newbuf_size) {
-                newbuf_size = size_roundup_power2(newbuf_size);
-
-                OBD_ALLOC_LARGE(newbuf, newbuf_size);
-                if (newbuf == NULL)
-                        RETURN(-ENOMEM);
-
-                memcpy(newbuf, req->rq_reqbuf, req->rq_reqbuf_len);
-
-                OBD_FREE_LARGE(req->rq_reqbuf, req->rq_reqbuf_len);
-                req->rq_reqbuf = newbuf;
-                req->rq_reqbuf_len = newbuf_size;
-                req->rq_reqmsg = lustre_msg_buf(req->rq_reqbuf, 1, 0);
-        }
+               rc = ptlrpc_enlarge_req_buffer(req, newbuf_size);
+               if (rc != 0)
+                       RETURN(rc);
+       }
 
         /* do enlargement, from wrapper to embedded, from end to begin */
         if (svc != SPTLRPC_SVC_NULL)
@@ -1769,6 +1761,8 @@ int gss_enlarge_reqbuf_priv(struct ptlrpc_sec *sec,
                 if (newclrbuf_size + newcipbuf_size <= req->rq_reqbuf_len) {
                         void *src, *dst;
 
+                       if (req->rq_import)
+                               spin_lock(&req->rq_import->imp_lock);
                         /* move clear text backward. */
                         src = req->rq_clrbuf;
                         dst = (char *) req->rq_reqbuf + newcipbuf_size;
@@ -1778,6 +1772,9 @@ int gss_enlarge_reqbuf_priv(struct ptlrpc_sec *sec,
                         req->rq_clrbuf = (struct lustre_msg *) dst;
                         req->rq_clrbuf_len = newclrbuf_size;
                         req->rq_reqmsg = lustre_msg_buf(req->rq_clrbuf, 0, 0);
+
+                       if (req->rq_import)
+                               spin_unlock(&req->rq_import->imp_lock);
                 } else {
                         /* sadly we have to split out the clear buffer */
                         LASSERT(req->rq_reqbuf_len >= newcipbuf_size);
@@ -1792,6 +1789,15 @@ int gss_enlarge_reqbuf_priv(struct ptlrpc_sec *sec,
                 if (newclrbuf == NULL)
                         RETURN(-ENOMEM);
 
+               /* Must lock this, so that otherwise unprotected change of
+                * rq_reqmsg is not racing with parallel processing of
+                * imp_replay_list traversing threads. See LU-3333
+                * This is a bandaid at best, we really need to deal with this
+                * in request enlarging code before unpacking that's already
+                * there */
+               if (req->rq_import)
+                       spin_lock(&req->rq_import->imp_lock);
+
                 memcpy(newclrbuf, req->rq_clrbuf, req->rq_clrbuf_len);
 
                 if (req->rq_reqbuf == NULL ||
@@ -1804,6 +1810,9 @@ int gss_enlarge_reqbuf_priv(struct ptlrpc_sec *sec,
                 req->rq_clrbuf = newclrbuf;
                 req->rq_clrbuf_len = newclrbuf_size;
                 req->rq_reqmsg = lustre_msg_buf(req->rq_clrbuf, 0, 0);
+
+               if (req->rq_import)
+                       spin_unlock(&req->rq_import->imp_lock);
         }
 
         _sptlrpc_enlarge_msg_inplace(req->rq_clrbuf, 0, newmsg_size);
index d12885d..262ee3b 100644 (file)
@@ -96,6 +96,46 @@ int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
 }
 EXPORT_SYMBOL(ptlrpc_buf_need_swab);
 
+/* This enlarges the req buffer of request \a req to the next power of 2
+ * multiple of \a newbuf_size.
+ * Returns zero on success or ENOMEM if it failed to allocate the new buffer.
+ *
+ * This is used in the reply path on the client if the server responded
+ * with a bigger message than we expected so we can save the new state for
+ * a possible future replay where we'll need to present this new info
+ * (usually striping that's not available at create time) */
+int ptlrpc_enlarge_req_buffer(struct ptlrpc_request *req, int newbuf_size)
+{
+       struct lustre_msg *newbuf;
+
+       newbuf_size = size_roundup_power2(newbuf_size);
+
+       OBD_ALLOC_LARGE(newbuf, newbuf_size);
+       if (newbuf == NULL)
+               return -ENOMEM;
+
+       /* Must lock this, so that otherwise unprotected change of
+        * rq_reqmsg is not racing with parallel processing of
+        * imp_replay_list traversing threads. See LU-3333
+        * This is a bandaid at best, we really need to deal with this
+        * in request enlarging code before unpacking what's already
+        * there */
+       if (req->rq_import)
+               spin_lock(&req->rq_import->imp_lock);
+
+       memcpy(newbuf, req->rq_reqbuf, req->rq_reqbuf_len);
+
+       OBD_FREE_LARGE(req->rq_reqbuf, req->rq_reqbuf_len);
+       req->rq_reqbuf = newbuf;
+       req->rq_reqbuf_len = newbuf_size;
+       req->rq_reqmsg = lustre_msg_buf(req->rq_reqbuf, 1, 0);
+
+       if (req->rq_import)
+               spin_unlock(&req->rq_import->imp_lock);
+
+       return 0;
+}
+
 static inline int lustre_msg_check_version_v2(struct lustre_msg_v2 *msg,
                                               __u32 version)
 {
index ab8a1dd..f21a0a7 100644 (file)
@@ -238,6 +238,7 @@ void ptlrpc_add_bulk_page(struct ptlrpc_bulk_desc *desc, struct page *page,
 struct ptlrpc_reply_state *
 lustre_get_emerg_rs(struct ptlrpc_service_part *svcpt);
 void lustre_put_emerg_rs(struct ptlrpc_reply_state *rs);
+int ptlrpc_enlarge_req_buffer(struct ptlrpc_request *req, int newbuf_size);
 
 /* pinger.c */
 int ptlrpc_start_pinger(void);
index c62ab83..4c01682 100644 (file)
@@ -49,6 +49,7 @@
 #include <obd_class.h>
 #include <lustre_net.h>
 #include <lustre_sec.h>
+#include "ptlrpc_internal.h"
 
 static struct ptlrpc_sec_policy null_policy;
 static struct ptlrpc_sec        null_sec;
@@ -238,9 +239,9 @@ int null_enlarge_reqbuf(struct ptlrpc_sec *sec,
                         struct ptlrpc_request *req,
                         int segment, int newsize)
 {
-        struct lustre_msg      *newbuf;
         struct lustre_msg      *oldbuf = req->rq_reqmsg;
-        int                     oldsize, newmsg_size, alloc_size;
+        int                     oldsize, newmsg_size;
+       int rc;
 
         LASSERT(req->rq_reqbuf);
         LASSERT(req->rq_reqbuf == req->rq_reqmsg);
@@ -257,18 +258,10 @@ int null_enlarge_reqbuf(struct ptlrpc_sec *sec,
         LASSERT(!req->rq_pool || req->rq_reqbuf_len >= newmsg_size);
 
         if (req->rq_reqbuf_len < newmsg_size) {
-                alloc_size = size_roundup_power2(newmsg_size);
-
-                OBD_ALLOC_LARGE(newbuf, alloc_size);
-                if (newbuf == NULL)
-                        return -ENOMEM;
-
-                memcpy(newbuf, req->rq_reqbuf, req->rq_reqlen);
-
-                OBD_FREE_LARGE(req->rq_reqbuf, req->rq_reqbuf_len);
-                req->rq_reqbuf = req->rq_reqmsg = newbuf;
-                req->rq_reqbuf_len = alloc_size;
-        }
+               rc = ptlrpc_enlarge_req_buffer(req, newmsg_size);
+               if (rc != 0)
+                       return rc;
+       }
 
         _sptlrpc_enlarge_msg_inplace(req->rq_reqmsg, segment, newsize);
         req->rq_reqlen = newmsg_size;
index bf02263..45c8c9f 100644 (file)
@@ -49,6 +49,7 @@
 #include <obd_class.h>
 #include <lustre_net.h>
 #include <lustre_sec.h>
+#include "ptlrpc_internal.h"
 
 struct plain_sec {
         struct ptlrpc_sec       pls_base;
@@ -671,9 +672,9 @@ int plain_enlarge_reqbuf(struct ptlrpc_sec *sec,
                          struct ptlrpc_request *req,
                          int segment, int newsize)
 {
-        struct lustre_msg      *newbuf;
         int                     oldsize;
         int                     newmsg_size, newbuf_size;
+       int                     rc;
         ENTRY;
 
         LASSERT(req->rq_reqbuf);
@@ -699,20 +700,10 @@ int plain_enlarge_reqbuf(struct ptlrpc_sec *sec,
         LASSERT(!req->rq_pool || req->rq_reqbuf_len >= newbuf_size);
 
         if (req->rq_reqbuf_len < newbuf_size) {
-                newbuf_size = size_roundup_power2(newbuf_size);
-
-                OBD_ALLOC_LARGE(newbuf, newbuf_size);
-                if (newbuf == NULL)
-                        RETURN(-ENOMEM);
-
-                memcpy(newbuf, req->rq_reqbuf, req->rq_reqbuf_len);
-
-                OBD_FREE_LARGE(req->rq_reqbuf, req->rq_reqbuf_len);
-                req->rq_reqbuf = newbuf;
-                req->rq_reqbuf_len = newbuf_size;
-                req->rq_reqmsg = lustre_msg_buf(req->rq_reqbuf,
-                                                PLAIN_PACK_MSG_OFF, 0);
-        }
+               rc = ptlrpc_enlarge_req_buffer(req, newbuf_size);
+               if (rc != 0)
+                       RETURN(rc);
+       }
 
         _sptlrpc_enlarge_msg_inplace(req->rq_reqbuf, PLAIN_PACK_MSG_OFF,
                                      newmsg_size);