2.5-18.1

1 files changed, 189 insertions, 64 deletions

diff --git a/nscd/nscd_helper.c b/nscd/nscd_helper.c
index 0e16cb8..71ea53e 100644
--- a/nscd/nscd_helper.c
+++ b/nscd/nscd_helper.c
@@ -1,4 +1,5 @@
-/* Copyright (C) 1998-2002, 2003, 2004 Free Software Foundation, Inc.
+/* Copyright (C) 1998-2002,2003,2004,2005,2006,2007
+   Free Software Foundation, Inc.
    This file is part of the GNU C Library.
    Contributed by Ulrich Drepper <drepper@cygnus.com>, 1998.
 
@@ -21,11 +22,14 @@
 #include <errno.h>
 #include <fcntl.h>
 #include <stdbool.h>
+#include <string.h>
+#include <time.h>
 #include <unistd.h>
 #include <sys/mman.h>
 #include <sys/poll.h>
 #include <sys/socket.h>
 #include <sys/stat.h>
+#include <sys/time.h>
 #include <sys/uio.h>
 #include <sys/un.h>
 #include <not-cancel.h>
@@ -34,6 +38,64 @@
 #include "nscd-client.h"
 
 
+ssize_t
+__readall (int fd, void *buf, size_t len)
+{
+  size_t n = len;
+  ssize_t ret;
+  do
+    {
+      ret = TEMP_FAILURE_RETRY (__read (fd, buf, n));
+      if (ret <= 0)
+	break;
+      buf = (char *) buf + ret;
+      n -= ret;
+    }
+  while (n > 0);
+  return ret < 0 ? ret : len - n;
+}
+
+
+ssize_t
+__readvall (int fd, const struct iovec *iov, int iovcnt)
+{
+  ssize_t ret = TEMP_FAILURE_RETRY (__readv (fd, iov, iovcnt));
+  if (ret <= 0)
+    return ret;
+
+  size_t total = 0;
+  for (int i = 0; i < iovcnt; ++i)
+    total += iov[i].iov_len;
+
+  if (ret < total)
+    {
+      struct iovec iov_buf[iovcnt];
+      ssize_t r = ret;
+
+      struct iovec *iovp = memcpy (iov_buf, iov, iovcnt * sizeof (*iov));
+      do
+	{
+	  while (iovp->iov_len <= r)
+	    {
+	      r -= iovp->iov_len;
+	      --iovcnt;
+	      ++iovp;
+	    }
+	  iovp->iov_base = (char *) iovp->iov_base + r;
+	  iovp->iov_len -= r;
+	  r = TEMP_FAILURE_RETRY (__readv (fd, iovp, iovcnt));
+	  if (r <= 0)
+	    break;
+	  ret += r;
+	}
+      while (ret < total);
+      if (r < 0)
+	ret = r;
+    }
+  return ret;
+}
+
+
 static int
 open_socket (void)
 {
@@ -77,6 +139,36 @@ __nscd_unmap (struct mapped_database *mapped)
 }
 
 
+static int
+wait_on_socket (int sock)
+{
+  struct pollfd fds[1];
+  fds[0].fd = sock;
+  fds[0].events = POLLIN | POLLERR | POLLHUP;
+  int n = __poll (fds, 1, 5 * 1000);
+  if (n == -1 && __builtin_expect (errno == EINTR, 0))
+    {
+      /* Handle the case where the poll() call is interrupted by a
+	 signal.  We cannot just use TEMP_FAILURE_RETRY since it might
+	 lead to infinite loops.  */
+      struct timeval now;
+      (void) __gettimeofday (&now, NULL);
+      long int end = (now.tv_sec + 5) * 1000 + (now.tv_usec + 500) / 1000;
+      while (1)
+	{
+	  long int timeout = end - (now.tv_sec * 1000
+				    + (now.tv_usec + 500) / 1000);
+	  n = __poll (fds, 1, timeout);
+	  if (n != -1 || errno != EINTR)
+	    break;
+	  (void) __gettimeofday (&now, NULL);
+	}
+    }
+
+  return n;
+}
+
+
 /* Try to get a file descriptor for the shared meory segment
    containing the database.  */
 static struct mapped_database *
@@ -86,102 +178,115 @@ get_mapping (request_type type, const char *key,
   struct mapped_database *result = NO_MAPPING;
 #ifdef SCM_RIGHTS
   const size_t keylen = strlen (key) + 1;
-  char resdata[keylen];
   int saved_errno = errno;
 
   int mapfd = -1;
 
   /* Send the request.  */
-  struct iovec iov[2];
-  request_header req;
+  struct
+  {
+    request_header req;
+    char key[keylen];
+  } reqdata;
+  size_t real_sizeof_reqdata = sizeof (request_header) + keylen;
 
   int sock = open_socket ();
   if (sock < 0)
     goto out;
 
-  req.version = NSCD_VERSION;
-  req.type = type;
-  req.key_len = keylen;
-
-  iov[0].iov_base = &req;
-  iov[0].iov_len = sizeof (req);
-  iov[1].iov_base = (void *) key;
-  iov[1].iov_len = keylen;
-
-  if (TEMP_FAILURE_RETRY (__writev (sock, iov, 2))
-      != iov[0].iov_len + iov[1].iov_len)
+  reqdata.req.version = NSCD_VERSION;
+  reqdata.req.type = type;
+  reqdata.req.key_len = keylen;
+  memcpy (reqdata.key, key, keylen);
+
+# ifndef MSG_NOSIGNAL
+#  define MSG_NOSIGNAL 0
+# endif
+  if (__builtin_expect (TEMP_FAILURE_RETRY (__send (sock, &reqdata,
+						    real_sizeof_reqdata,
+						    MSG_NOSIGNAL))
+			!= real_sizeof_reqdata, 0))
     /* We cannot even write the request.  */
     goto out_close2;
 
   /* Room for the data sent along with the file descriptor.  We expect
      the key name back.  */
+# define resdata reqdata.key
+  struct iovec iov[1];
   iov[0].iov_base = resdata;
   iov[0].iov_len = keylen;
 
-  char buf[CMSG_SPACE (sizeof (int))];
+  union
+  {
+    struct cmsghdr hdr;
+    char bytes[CMSG_SPACE (sizeof (int))];
+  } buf;
   struct msghdr msg = { .msg_iov = iov, .msg_iovlen = 1,
-			.msg_control = buf, .msg_controllen = sizeof (buf) };
+			.msg_control = buf.bytes,
+			.msg_controllen = sizeof (buf) };
   struct cmsghdr *cmsg = CMSG_FIRSTHDR (&msg);
 
   cmsg->cmsg_level = SOL_SOCKET;
   cmsg->cmsg_type = SCM_RIGHTS;
   cmsg->cmsg_len = CMSG_LEN (sizeof (int));
 
+  /* This access is well-aligned since BUF is correctly aligned for an
+     int and CMSG_DATA preserves this alignment.  */
   *(int *) CMSG_DATA (cmsg) = -1;
 
   msg.msg_controllen = cmsg->cmsg_len;
 
-  struct pollfd fds[1];
-  fds[0].fd = sock;
-  fds[0].events = POLLIN | POLLERR | POLLHUP;
-  if (__poll (fds, 1, 5 * 1000) <= 0)
-    /* Failure or timeout.  */
+  if (wait_on_socket (sock) <= 0)
     goto out_close2;
 
-  if (TEMP_FAILURE_RETRY (__recvmsg (sock, &msg, 0)) != keylen)
+  if (__builtin_expect (TEMP_FAILURE_RETRY (__recvmsg (sock, &msg, 0))
+			!= keylen, 0))
     goto out_close2;
 
-  mapfd = *(int *) CMSG_DATA (cmsg);
+  if (__builtin_expect (CMSG_FIRSTHDR (&msg) == NULL
+			|| (CMSG_FIRSTHDR (&msg)->cmsg_len
+			    != CMSG_LEN (sizeof (int))), 0))
+    goto out_close2;
 
-  if (CMSG_FIRSTHDR (&msg)->cmsg_len != CMSG_LEN (sizeof (int)))
-    goto out_close;
+  mapfd = *(int *) CMSG_DATA (cmsg);
 
   struct stat64 st;
-  if (strcmp (resdata, key) != 0
-      || fstat64 (mapfd, &st) != 0
-      || st.st_size < sizeof (struct database_pers_head))
+  if (__builtin_expect (strcmp (resdata, key) != 0, 0)
+      || __builtin_expect (fstat64 (mapfd, &st) != 0, 0)
+      || __builtin_expect (st.st_size < sizeof (struct database_pers_head), 0))
     goto out_close;
 
   struct database_pers_head head;
-  if (TEMP_FAILURE_RETRY (__pread (mapfd, &head, sizeof (head), 0))
-      != sizeof (head))
+  if (__builtin_expect (TEMP_FAILURE_RETRY (__pread (mapfd, &head,
+						     sizeof (head), 0))
+			!= sizeof (head), 0))
     goto out_close;
 
-  if (head.version != DB_VERSION || head.header_size != sizeof (head)
+  if (__builtin_expect (head.version != DB_VERSION, 0)
+      || __builtin_expect (head.header_size != sizeof (head), 0)
       /* This really should not happen but who knows, maybe the update
 	 thread got stuck.  */
-      || (! head.nscd_certainly_running
-	  && head.timestamp + MAPPING_TIMEOUT < time (NULL)))
+      || __builtin_expect (! head.nscd_certainly_running
+			   && head.timestamp + MAPPING_TIMEOUT < time (NULL),
+			   0))
     goto out_close;
 
   size_t size = (sizeof (head) + roundup (head.module * sizeof (ref_t), ALIGN)
 		 + head.data_size);
 
-  if (st.st_size < size)
+  if (__builtin_expect (st.st_size < size, 0))
     goto out_close;
 
   /* The file is large enough, map it now.  */
   void *mapping = __mmap (NULL, size, PROT_READ, MAP_SHARED, mapfd, 0);
-  if (mapping != MAP_FAILED)
+  if (__builtin_expect (mapping != MAP_FAILED, 1))
     {
       /* Allocate a record for the mapping.  */
-      struct mapped_database *newp;
-
-      newp = malloc (sizeof (*newp));
+      struct mapped_database *newp = malloc (sizeof (*newp));
       if (newp == NULL)
 	{
 	  /* Ugh, after all we went through the memory allocation failed.  */
-	  __munmap (result, size);
+	  __munmap (mapping, size);
 	  goto out_close;
 	}
 
@@ -189,6 +294,7 @@ get_mapping (request_type type, const char *key,
       newp->data = ((char *) mapping + head.header_size
 		    + roundup (head.module * sizeof (ref_t), ALIGN));
       newp->mapsize = size;
+      newp->datasize = head.data_size;
       /* Set counter to 1 to show it is usable.  */
       newp->counter = 1;
 
@@ -215,17 +321,18 @@ get_mapping (request_type type, const char *key,
 
 struct mapped_database *
 __nscd_get_map_ref (request_type type, const char *name,
-		    struct locked_map_ptr *mapptr, int *gc_cyclep)
+		    volatile struct locked_map_ptr *mapptr, int *gc_cyclep)
 {
   struct mapped_database *cur = mapptr->mapped;
   if (cur == NO_MAPPING)
     return cur;
 
   int cnt = 0;
-  while (atomic_compare_and_exchange_val_acq (&mapptr->lock, 1, 0) != 0)
+  while (__builtin_expect (atomic_compare_and_exchange_val_acq (&mapptr->lock,
+								1, 0) != 0, 0))
     {
       // XXX Best number of rounds?
-      if (++cnt > 5)
+      if (__builtin_expect (++cnt > 5, 0))
 	return NO_MAPPING;
 
       atomic_delay ();
@@ -238,8 +345,10 @@ __nscd_get_map_ref (request_type type, const char *name,
       /* If not mapped or timestamp not updated, request new map.  */
       if (cur == NULL
 	  || (cur->head->nscd_certainly_running == 0
-	      && cur->head->timestamp + MAPPING_TIMEOUT < time (NULL)))
-	cur = get_mapping (type, name, &mapptr->mapped);
+	      && cur->head->timestamp + MAPPING_TIMEOUT < time (NULL))
+	  || cur->head->data_size > cur->datasize)
+	cur = get_mapping (type, name,
+			   (struct mapped_database **) &mapptr->mapped);
 
       if (__builtin_expect (cur != NO_MAPPING, 1))
 	{
@@ -257,28 +366,50 @@ __nscd_get_map_ref (request_type type, const char *name,
 }
 
 
-const struct datahead *
+/* Don't return const struct datahead *, as eventhough the record
+   is normally constant, it can change arbitrarily during nscd
+   garbage collection.  */
+struct datahead *
 __nscd_cache_search (request_type type, const char *key, size_t keylen,
 		     const struct mapped_database *mapped)
 {
   unsigned long int hash = __nis_hash (key, keylen) % mapped->head->module;
+  size_t datasize = mapped->datasize;
 
   ref_t work = mapped->head->array[hash];
-  while (work != ENDREF)
+  while (work != ENDREF && work + sizeof (struct hashentry) <= datasize)
     {
       struct hashentry *here = (struct hashentry *) (mapped->data + work);
 
-      if (type == here->type && keylen == here->len
-	  && memcmp (key, mapped->data + here->key, keylen) == 0)
+#ifndef _STRING_ARCH_unaligned
+      /* Although during garbage collection when moving struct hashentry
+	 records around we first copy from old to new location and then
+	 adjust pointer from previous hashentry to it, there is no barrier
+	 between those memory writes.  It is very unlikely to hit it,
+	 so check alignment only if a misaligned load can crash the
+	 application.  */
+      if ((uintptr_t) here & (__alignof__ (*here) - 1))
+	return NULL;
+#endif
+
+      if (type == here->type
+	  && keylen == here->len
+	  && here->key + keylen <= datasize
+	  && memcmp (key, mapped->data + here->key, keylen) == 0
+	  && here->packet + sizeof (struct datahead) <= datasize)
 	{
 	  /* We found the entry.  Increment the appropriate counter.  */
-	  const struct datahead *dh
+	  struct datahead *dh
 	    = (struct datahead *) (mapped->data + here->packet);
 
+#ifndef _STRING_ARCH_unaligned
+	  if ((uintptr_t) dh & (__alignof__ (*dh) - 1))
+	    return NULL;
+#endif
+
 	  /* See whether we must ignore the entry or whether something
 	     is wrong because garbage collection is in progress.  */
-	  if (dh->usable && ((char *) dh + dh->allocsize
-			     <= (char *) mapped->head + mapped->mapsize))
+	  if (dh->usable && here->packet + dh->allocsize <= datasize)
 	    return dh;
 	}
 
@@ -311,19 +442,13 @@ __nscd_open_socket (const char *key, size_t keylen, request_type type,
       vec[1].iov_len = keylen;
 
       ssize_t nbytes = TEMP_FAILURE_RETRY (__writev (sock, vec, 2));
-      if (nbytes == (ssize_t) (sizeof (request_header) + keylen))
-	{
+      if (nbytes == (ssize_t) (sizeof (request_header) + keylen)
 	  /* Wait for data.  */
-	  struct pollfd fds[1];
-	  fds[0].fd = sock;
-	  fds[0].events = POLLIN | POLLERR | POLLHUP;
-	  if (__poll (fds, 1, 5 * 1000) > 0)
-	    {
-	      nbytes = TEMP_FAILURE_RETRY (__read (sock, response,
-						   responselen));
-	      if (nbytes == (ssize_t) responselen)
-		return sock;
-	    }
+	  && wait_on_socket (sock) > 0)
+	{
+	  nbytes = TEMP_FAILURE_RETRY (__read (sock, response, responselen));
+	  if (nbytes == (ssize_t) responselen)
+	    return sock;
 	}
 
       close_not_cancel_no_status (sock);