[PATCH] [TSan] Allocate fd table in user heap instead of using internal allocator

[Alexey Samsonov]

Hi dvyukov,

We catch races on file descriptors assuming fd table is located in the
user memory (as we use functions like MemoryImitateWrite). Thus, it's wrong to allocate
this table with internal_alloc() - we were lucky up to now because internal_alloc()
us currently libc malloc.

http://llvm-reviews.chandlerc.com/D663

Files:
  lib/tsan/rtl/tsan_fd.cc

Index: lib/tsan/rtl/tsan_fd.cc
===================================================================
--- lib/tsan/rtl/tsan_fd.cc
+++ lib/tsan/rtl/tsan_fd.cc
@@ -74,13 +74,14 @@
   uptr l1 = atomic_load(pl1, memory_order_consume);
   if (l1 == 0) {
     uptr size = kTableSizeL2 * sizeof(FdDesc);
-    void *p = internal_alloc(MBlockFD, size);
+    // We need this to reside in user memory to properly catch races on it.
+    void *p = user_alloc(thr, pc, size);
     internal_memset(p, 0, size);
     MemoryResetRange(thr, (uptr)&fddesc, (uptr)p, size);
     if (atomic_compare_exchange_strong(pl1, &l1, (uptr)p, memory_order_acq_rel))
       l1 = (uptr)p;
     else
-      internal_free(p);
+      user_free(thr, pc, p);
   }
   return &((FdDesc*)l1)[fd % kTableSizeL2];  // NOLINT
 }
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D663.1.patch
Type: text/x-patch
Size: 784 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130412/37d4263f/attachment.bin>