[PATCH] Add user-defined callback on write() calls to labeled data.

[Sam Kerner]

================
Comment at: test/dfsan/write_callback.c:1
@@ +1,2 @@
+// RUN: %clang_dfsan                        -m64 %s -o %t && %t %T/file1.txt
+// RUN: %clang_dfsan -mllvm -dfsan-args-abi -m64 %s -o %t && %t %T/file2.txt
----------------
Sam Kerner wrote:
> Peter Collingbourne wrote:
> > Maybe we could check that these files have the contents we expect?
> It would be reasonable to assume that calls to write() do not fail in
> tests.  But write() may only write some of the bytes it is passed.
> For example:
> 
> char buf[] = "123456789";
> int bytes_written = write(fd, buf, 10);
> 
> If (for whatever reason) only 5 bytes are written, write may return 5.
>  I doubt it would do so on a local file system, but I don't want to
> assume something that is not in the spec.
> 
> I considered adding retries to writes:
> 
> char buf[] = "123456789";
> 
> char* cur = buf;
> int bytes_left = strlen(buf);
> while ( bytes_left > 0 ) {
>   int result = write(fd, cur, bytes_left);
> 
>   if (result < 0) ERROR HANDLER
> 
>   bytes_left -= result;
> }
> 
> ... but now the number of callbacks tests must expect changes based on
> how many times write() ends up being called.
> 
> 
Please ignore the above comment.  I wrote it before realizing that its premise is flawed, and wrote the rest in a way that does check that the file is written.  


http://reviews.llvm.org/D3268