[patch] libstdc++/67173 Fix filesystem::canonical for Solaris 10.

Jonathan Wakely jwakely@redhat.com
Thu Sep 17 14:38:00 GMT 2015


On 17/09/15 12:16 +0100, Jonathan Wakely wrote:
>On 16/09/15 17:42 -0600, Martin Sebor wrote:
>>I see now the first exists test will detect symlink loops in
>>the original path. But I'm not convinced there isn't a corner
>>case that's subject to a TOCTOU race condition between the first
>>exists test and the while loop during which a symlink loop can
>>be introduced.
>>
>>Suppose we call the function with /foo/bar as an argument and
>>the path exists and contains no symlinks. result is / and cmpts
>>is set to { foo, bar }. Just as the loop is entered, /foo/bar
>>is replaced with a symlink containing /foo/bar. The loop then
>>proceeds like so:
>>
>>1. The first iteration removes foo from cmpts and sets result
>>to /foo. cmpts is { bar }.
>>
>>2. The second iteration removes bar from cmpts, sets result to
>>/foo/bar, determines it's a symlink, reads its contents, sees
>>it's an absolute pathname and replaces result with /. It then
>>inserts the symlink's components { foo, bar } into cmpts. cmpts
>>becomes { foo, bar }. exists(result) succeeds.
>>
>>3. The next iteration of the loop has the same initial state
>>as the first.
>>
>>But I could have very easily missed something that takes care
>>of this corner case. If I did, sorry for the false alarm!
>
>No, you're right. The TS says such filesystem races are undefined:
>http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4099.html#fs.race.behavior
>but it would be nice to fail gracefully rather than DOS the
>application.
>
>The simplest approach would be to increment a counter every time we
>follow a symlink, and if it reaches some limit decide something is
>wrong and fail with ELOOP.
>
>I don't see how anything else can be 100% bulletproof, because a truly
>evil attacker could just keep altering the contents of symlinks so we
>keep ping-ponging between two or more paths. If we keep track of paths
>we've seen before the attacker could just keep changing the contents
>to a unique path each time, that initially exists as a file, but by
>the time we get to is_symlink() its become a symlink to a new path.
>
>So if we use a counter, what's a sane maximum? Is MAXSYMLINKS in
><sys/param.h> the value the kernel uses? 20 seems quite low, I was
>thinking of a much higher number.

This patch sets ELOOP after following 40 symlinks.

I can also move the exists(result, ec) check to the end, because the
is_symlink(result, ec) call will already check for existence on every
iteration that adds a component to the result.

I've also simplified the error handling (when exists(p, ec) fails it
sets ENOENT anyway) and moved !ec into the loop condition, rather than
using 'fail(e); break;' on errors.

I'm quite happy with this version now.


-------------- next part --------------
diff --git a/libstdc++-v3/src/filesystem/ops.cc b/libstdc++-v3/src/filesystem/ops.cc
index b5c8eb9..95146bf 100644
--- a/libstdc++-v3/src/filesystem/ops.cc
+++ b/libstdc++-v3/src/filesystem/ops.cc
@@ -98,6 +98,7 @@ fs::canonical(const path& p, const path& base, error_code& ec)
 {
   const path pa = absolute(p, base);
   path result;
+
 #ifdef _GLIBCXX_USE_REALPATH
   char_ptr buf{ nullptr };
 # if _XOPEN_VERSION < 700
@@ -119,18 +120,9 @@ fs::canonical(const path& p, const path& base, error_code& ec)
     }
 #endif
 
-  auto fail = [&ec, &result](int e) mutable {
-      if (!ec.value())
-	ec.assign(e, std::generic_category());
-      result.clear();
-  };
-
   if (!exists(pa, ec))
-    {
-      fail(ENOENT);
-      return result;
-    }
-  // else we can assume no unresolvable symlink loops
+    return result;
+  // else: we know there are (currently) no unresolvable symlink loops
 
   result = pa.root_path();
 
@@ -138,18 +130,17 @@ fs::canonical(const path& p, const path& base, error_code& ec)
   for (auto& f : pa.relative_path())
     cmpts.push_back(f);
 
-  while (!cmpts.empty())
+  int max_allowed_symlinks = 40;
+
+  while (!cmpts.empty() && !ec)
     {
       path f = std::move(cmpts.front());
       cmpts.pop_front();
 
       if (f.compare(".") == 0)
 	{
-	  if (!is_directory(result, ec))
-	    {
-	      fail(ENOTDIR);
-	      break;
-	    }
+	  if (!is_directory(result, ec) && !ec)
+	    ec.assign(ENOTDIR, std::generic_category());
 	}
       else if (f.compare("..") == 0)
 	{
@@ -166,27 +157,30 @@ fs::canonical(const path& p, const path& base, error_code& ec)
 	  if (is_symlink(result, ec))
 	    {
 	      path link = read_symlink(result, ec);
-	      if (!ec.value())
+	      if (!ec)
 		{
-		  if (link.is_absolute())
+		  if (--max_allowed_symlinks == 0)
+		    ec.assign(ELOOP, std::generic_category());
+		  else
 		    {
-		      result = link.root_path();
-		      link = link.relative_path();
+		      if (link.is_absolute())
+			{
+			  result = link.root_path();
+			  link = link.relative_path();
+			}
+		      else
+			result.remove_filename();
+
+		      cmpts.insert(cmpts.begin(), link.begin(), link.end());
 		    }
-		  else
-		    result.remove_filename();
-
-		  cmpts.insert(cmpts.begin(), link.begin(), link.end());
 		}
 	    }
-
-	  if (ec.value() || !exists(result, ec))
-	    {
-	      fail(ENOENT);
-	      break;
-	    }
 	}
     }
+
+  if (ec || !exists(result, ec))
+    result.clear();
+
   return result;
 }
 


More information about the Gcc-patches mailing list