[2/3] Add system-wide tunables: cache ld.so.cache

Message ID xnbk662es0.fsf@greed.delorie.com
State Changes Requested
Headers
Series [1/3] Add system-wide tunables: ldconfig part |

Checks

Context Check Description
linaro-tcwg-bot/tcwg_glibc_build--master-aarch64 success Testing passed
linaro-tcwg-bot/tcwg_glibc_check--master-aarch64 success Testing passed
linaro-tcwg-bot/tcwg_glibc_build--master-arm success Testing passed
linaro-tcwg-bot/tcwg_glibc_check--master-arm success Testing passed
redhat-pt-bot/TryBot-apply_patch success Patch applied to master at the time it was sent

Commit Message

DJ Delorie April 19, 2024, 3:49 a.m. UTC
  The purpose of this change is twofold:

1. The ld.so.cache is cached in memory and only re-read if/when
   it changes on disk.  This allows us to have much more intensive
   security checks in the future, without impacting performance as
   much.  It also allows for cases where the cache is corrupted -
   we continue using the last valid one.

2. We break out the load/check logic so that the cache can be
   loaded independently of the library lookup, such as for
   code that only needs to look at the extensions.
---
 elf/dl-cache.c | 252 ++++++++++++++++++++++++++++++++-----------------
 1 file changed, 167 insertions(+), 85 deletions(-)
  

Patch

diff --git a/elf/dl-cache.c b/elf/dl-cache.c
index 85f3f179ed..d90278889d 100644
--- a/elf/dl-cache.c
+++ b/elf/dl-cache.c
@@ -26,6 +26,8 @@ 
 #include <_itoa.h>
 #include <dl-hwcaps.h>
 #include <dl-isa-level.h>
+#include <sys/types.h>
+#include <sys/stat.h>
 
 #ifndef _DL_PLATFORMS_COUNT
 # define _DL_PLATFORMS_COUNT 0
@@ -35,6 +37,10 @@ 
 static struct cache_file *cache;
 static struct cache_file_new *cache_new;
 static size_t cachesize;
+static struct cache_extension_all_loaded ext;
+
+static struct stat cache_file_time;
+static struct stat new_cache_file_time;
 
 #ifdef SHARED
 /* This is used to cache the priorities of glibc-hwcaps
@@ -57,6 +63,7 @@  glibc_hwcaps_priorities_free (void)
     free (glibc_hwcaps_priorities);
   glibc_hwcaps_priorities = NULL;
   glibc_hwcaps_priorities_allocated = 0;
+  glibc_hwcaps_priorities_length = 0;
 }
 
 /* Ordered comparison of a hwcaps string from the cache on the left
@@ -88,10 +95,6 @@  glibc_hwcaps_compare (uint32_t left_index, struct dl_hwcaps_priority *right)
 static void
 glibc_hwcaps_priorities_init (void)
 {
-  struct cache_extension_all_loaded ext;
-  if (!cache_extension_load (cache_new, cache, cachesize, &ext))
-    return;
-
   uint32_t length = (ext.sections[cache_extension_tag_glibc_hwcaps].size
 		     / sizeof (uint32_t));
   if (length > glibc_hwcaps_priorities_allocated)
@@ -390,6 +393,159 @@  _dl_cache_libcmp (const char *p1, const char *p2)
   return *p1 - *p2;
 }
 
+/* Set the cache back to the "no cache" state, which may include
+   cleaning up a loaded cache.  */
+static void
+_dl_maybe_unload_ldsocache (void)
+{
+  if (cache != NULL)
+    __munmap (cache, cachesize);
+
+  cache = NULL;
+  cache_new = NULL;
+  cachesize = 0;
+
+#ifdef SHARED
+  glibc_hwcaps_priorities_free ();
+#endif
+}
+
+/* Returns TRUE if for any reason the cache needs to be reloaded
+   (including, the first time, loaded).  */
+static bool
+_dl_check_ldsocache_needs_loading (void)
+{
+  int rv;
+  static int copy_old_time = 0;
+
+  /* Save the previous stat every time.  We only care when this
+     changes, and we only stat it here, so we can get away with doing
+     the copy now instead of at every single return statement in this
+     function.  However, we only need to copy it if the previous stat
+     succeeded.  The only way this could be subverted is if the admin
+     moves the file aside, then moves it back, but CACHE would be set
+     to NULL in the interim so that would be detected.  */
+  if (copy_old_time)
+    cache_file_time = new_cache_file_time;
+  rv = __stat (LD_SO_CACHE, &new_cache_file_time);
+  copy_old_time = (rv >= 0);
+
+  /* No file to load, but there used to be.  Assume user intentionally
+     deleted the cache and act accordingly.  */
+  if (rv < 0 && cache != NULL)
+    {
+      _dl_maybe_unload_ldsocache ();
+      return false;
+    }
+
+  /* No file to load and no loaded cache, so nothing to do.  */
+  if (rv < 0)
+    return false;
+
+  /* Any file is better than no file (likely the first time
+     through).  */
+  if (cache == NULL)
+    return true;
+
+  /* At this point, NEW_CACHE_FILE_TIME is valid as well as
+     CACHE_FILE_TIME, so we compare them.  We list fields in the order
+     they're most likely to be different in.  */
+  return ((new_cache_file_time.st_mtime != cache_file_time.st_mtime)
+	  || (new_cache_file_time.st_ino != cache_file_time.st_ino)
+	  || (new_cache_file_time.st_size != cache_file_time.st_size)
+	  || (new_cache_file_time.st_dev != cache_file_time.st_dev)
+	   );
+}
+
+/* Attemps to load and validate the cache.  On return, CACHE is either
+   unchanged (still loaded or still not loaded) or valid.  */
+static void
+_dl_maybe_load_ldsocache (void)
+{
+  struct cache_file *tmp_cache = NULL;
+  struct cache_file_new *tmp_cache_new = NULL;
+  size_t tmp_cachesize = 0;
+  
+  /* Read the contents of the file.  */
+  void *file = _dl_sysdep_read_whole_file (LD_SO_CACHE, &tmp_cachesize,
+					   PROT_READ);
+
+  /* We can handle three different cache file formats here:
+     - only the new format
+     - the old libc5/glibc2.0/2.1 format
+     - the old format with the new format in it
+     The following checks if the cache contains any of these formats.  */
+  if (file != MAP_FAILED && tmp_cachesize > sizeof *cache_new
+      && memcmp (file, CACHEMAGIC_VERSION_NEW,
+		 sizeof CACHEMAGIC_VERSION_NEW - 1) == 0
+      /* Check for corruption, avoiding overflow.  */
+      && ((tmp_cachesize - sizeof *cache_new) / sizeof (struct file_entry_new)
+	  >= ((struct cache_file_new *) file)->nlibs))
+    {
+      if (! cache_file_new_matches_endian (file))
+	{
+	  __munmap (file, tmp_cachesize);
+	  return;
+	}
+
+      tmp_cache_new = file;
+      tmp_cache = file;
+    }
+  else if (file != MAP_FAILED && cachesize > sizeof *cache
+	   && memcmp (file, CACHEMAGIC, sizeof CACHEMAGIC - 1) == 0
+	   /* Check for corruption, avoiding overflow.  */
+	   && ((tmp_cachesize - sizeof *cache) / sizeof (struct file_entry)
+	       >= ((struct cache_file *) file)->nlibs))
+    {
+      size_t offset;
+      /* Looks ok.  */
+      tmp_cache = file;
+
+      /* Check for new version.  */
+      offset = ALIGN_CACHE (sizeof (struct cache_file)
+			    + cache->nlibs * sizeof (struct file_entry));
+
+      tmp_cache_new = (struct cache_file_new *) ((void *) tmp_cache + offset);
+      if (tmp_cachesize < (offset + sizeof (struct cache_file_new))
+	  || memcmp (tmp_cache_new->magic, CACHEMAGIC_VERSION_NEW,
+		     sizeof CACHEMAGIC_VERSION_NEW - 1) != 0)
+	tmp_cache_new = NULL;
+      else
+	{
+	  if (! cache_file_new_matches_endian (tmp_cache_new))
+	    /* The old-format part of the cache is bogus as well
+	       if the endianness does not match.  (But it is
+	       unclear how the new header can be located if the
+	       endianness does not match.)  */
+	    {
+	      __munmap (file, tmp_cachesize);
+	      return;
+	    }
+	}
+    }
+  else
+    {
+      if (file != MAP_FAILED)
+	__munmap (file, tmp_cachesize);
+      return;
+    }
+
+  struct cache_extension_all_loaded tmp_ext;
+  if (!cache_extension_load (tmp_cache_new, tmp_cache, tmp_cachesize, &tmp_ext))
+    /* The extension is corrupt, so the cache is corrupt.  */
+    return;
+
+  /* If we've gotten here, the loaded cache is good and we need to
+     save it.  */
+  _dl_maybe_unload_ldsocache ();
+  cache = tmp_cache;
+  cache_new = tmp_cache_new;
+  cachesize = tmp_cachesize;
+  ext = tmp_ext;
+
+  assert (cache != NULL);
+}
+
 
 /* Look up NAME in ld.so.cache and return the file name stored there, or null
    if none is found.  The cache is loaded if it was not already.  If loading
@@ -405,81 +561,14 @@  _dl_load_cache_lookup (const char *name)
   if (__glibc_unlikely (GLRO(dl_debug_mask) & DL_DEBUG_LIBS))
     _dl_debug_printf (" search cache=%s\n", LD_SO_CACHE);
 
-  if (cache == NULL)
-    {
-      /* Read the contents of the file.  */
-      void *file = _dl_sysdep_read_whole_file (LD_SO_CACHE, &cachesize,
-					       PROT_READ);
-
-      /* We can handle three different cache file formats here:
-	 - only the new format
-	 - the old libc5/glibc2.0/2.1 format
-	 - the old format with the new format in it
-	 The following checks if the cache contains any of these formats.  */
-      if (file != MAP_FAILED && cachesize > sizeof *cache_new
-	  && memcmp (file, CACHEMAGIC_VERSION_NEW,
-		     sizeof CACHEMAGIC_VERSION_NEW - 1) == 0
-	  /* Check for corruption, avoiding overflow.  */
-	  && ((cachesize - sizeof *cache_new) / sizeof (struct file_entry_new)
-	      >= ((struct cache_file_new *) file)->nlibs))
-	{
-	  if (! cache_file_new_matches_endian (file))
-	    {
-	      __munmap (file, cachesize);
-	      file = (void *) -1;
-	    }
-	  cache_new = file;
-	  cache = file;
-	}
-      else if (file != MAP_FAILED && cachesize > sizeof *cache
-	       && memcmp (file, CACHEMAGIC, sizeof CACHEMAGIC - 1) == 0
-	       /* Check for corruption, avoiding overflow.  */
-	       && ((cachesize - sizeof *cache) / sizeof (struct file_entry)
-		   >= ((struct cache_file *) file)->nlibs))
-	{
-	  size_t offset;
-	  /* Looks ok.  */
-	  cache = file;
-
-	  /* Check for new version.  */
-	  offset = ALIGN_CACHE (sizeof (struct cache_file)
-				+ cache->nlibs * sizeof (struct file_entry));
-
-	  cache_new = (struct cache_file_new *) ((void *) cache + offset);
-	  if (cachesize < (offset + sizeof (struct cache_file_new))
-	      || memcmp (cache_new->magic, CACHEMAGIC_VERSION_NEW,
-			 sizeof CACHEMAGIC_VERSION_NEW - 1) != 0)
-	      cache_new = (void *) -1;
-	  else
-	    {
-	      if (! cache_file_new_matches_endian (cache_new))
-		{
-		  /* The old-format part of the cache is bogus as well
-		     if the endianness does not match.  (But it is
-		     unclear how the new header can be located if the
-		     endianness does not match.)  */
-		  cache = (void *) -1;
-		  cache_new = (void *) -1;
-		  __munmap (file, cachesize);
-		}
-	    }
-	}
-      else
-	{
-	  if (file != MAP_FAILED)
-	    __munmap (file, cachesize);
-	  cache = (void *) -1;
-	}
+  if (_dl_check_ldsocache_needs_loading ())
+    _dl_maybe_load_ldsocache ();
 
-      assert (cache != NULL);
-    }
-
-  if (cache == (void *) -1)
-    /* Previously looked for the cache file and didn't find it.  */
+  if (cache == NULL)
     return NULL;
 
   const char *best;
-  if (cache_new != (void *) -1)
+  if (cache_new != NULL)
     {
       const char *string_table = (const char *) cache_new;
       best = search_cache (string_table, cachesize,
@@ -505,7 +594,7 @@  _dl_load_cache_lookup (const char *name)
     return NULL;
 
   /* The double copy is *required* since malloc may be interposed
-     and call dlopen itself whose completion would unmap the data
+     and call dlopen itself whose completion may unmap the data
      we are accessing. Therefore we must make the copy of the
      mapping data without using malloc.  */
   char *temp;
@@ -523,14 +612,7 @@  _dl_load_cache_lookup (const char *name)
 void
 _dl_unload_cache (void)
 {
-  if (cache != NULL && cache != (struct cache_file *) -1)
-    {
-      __munmap (cache, cachesize);
-      cache = NULL;
-    }
-#ifdef SHARED
-  /* This marks the glibc_hwcaps_priorities array as out-of-date.  */
-  glibc_hwcaps_priorities_length = 0;
-#endif
+  /* Functionality is no longer needed, but kept for internal ABI for
+     now.  */
 }
 #endif