Re: write() _will_ fail on Mac OS X/XNU if nbytes > INT_MAX

Filipe Cabecinhas [off-list ref] writes:

I think it is better to fix it in much lower level of the stack, as
other codepaths would call write(2), either thru xwrite() or
directly.

Ideally the fix should go to the lowest level, i.e. the write(2).  I
do not care if it is done in the kernel or in the libc wrapping
code; the above does not belong to our code (in an ideal world, that
is).

Otherwise you would have to patch everything in /usr/bin, no?

But you do not live in an ideal world and neither do we.  I think
the least ugly way may be to add compat/clipped-write.c that
implements a loop like the above in a helper function:

	ssize_t clipped_write(int, const void *, size_t);

and have a

	#ifdef NEED_CLIPPED_WRITE
	#define write(x,y,z) clipped_write((x),(y),(z))
        #endif

in git-compat-util.h, or something.  Makefile needs to get adjusted
to link with compat/clipped-write.o when NEED_CLIPPED_WRITE is
defined as well.

Hi Junio,

Sorry for the delay. I've updated the patch to work as you suggested (I think).
It's attached.

Thank you,

  Filipe
  F


On Tue, Apr 9, 2013 at 3:50 PM, Junio C Hamano [off-list ref] wrote:

Filipe Cabecinhas [off-list ref] writes:

quoted

Testing with dd bs=INT_MAX+1 count=1 also gets me an “Invalid
argument” error, while bs=INT_MAX will do what's expected.

I have a preliminary patch that fixes it, but it may not be the
preferred way. The code is not ifdef'ed out and I'm doing the fix in
write_in_full(), while it may be preferred to do the fix in xwrite().

A radar bug has been submitted to Apple about this, but I think git
could tolerate the bug while it isn't fixed, by working around it.

Thank you,

  Filipe

diff --git a/wrapper.c b/wrapper.c
index bac59d2..474d760 100644
--- a/wrapper.c
+++ b/wrapper.c

@@ -187,7 +187,12 @@ ssize_t write_in_full(int fd, const void *buf, size_t count)
      ssize_t total = 0;

      while (count > 0) {
-             ssize_t written = xwrite(fd, p, count);
+             ssize_t written = 0;
+        if (count >= INT_MAX)
+                     written = xwrite(fd, p, INT_MAX-1);
+        else
+                     written = xwrite(fd, p, count);

I think it is better to fix it in much lower level of the stack, as
other codepaths would call write(2), either thru xwrite() or
directly.

Ideally the fix should go to the lowest level, i.e. the write(2).  I
do not care if it is done in the kernel or in the libc wrapping
code; the above does not belong to our code (in an ideal world, that
is).

Otherwise you would have to patch everything in /usr/bin, no?

But you do not live in an ideal world and neither do we.  I think
the least ugly way may be to add compat/clipped-write.c that
implements a loop like the above in a helper function:

        ssize_t clipped_write(int, const void *, size_t);

and have a

        #ifdef NEED_CLIPPED_WRITE
        #define write(x,y,z) clipped_write((x),(y),(z))
        #endif

in git-compat-util.h, or something.  Makefile needs to get adjusted
to link with compat/clipped-write.o when NEED_CLIPPED_WRITE is
defined as well.