fseek does not discard ungetc buffer

[bhaible]

ISO C 99 section 7.19.9.2 specifies that "a successful call to the fseek function undoes any effects of the ungetc function on the stream". Same in POSIX, see

​http://www.opengroup.org/susv3/functions/fseek.html

Attached is a test case showing that this is not fulfilled in Haiku.

[bhaible]

test case

[bhaible]

screenshot of execution of the test case

[bhaible]

a second test case

[bhaible]

screenshot of execution of the second test case

[scottmc]

Perhaps a newer fseek has fixed this? ​http://svn.freebsd.org/viewvc/base/head/lib/libc/stdio/fseek.c?view=log

[hamish]

Fixes ungetc/fseek bug as described in #3145

[hamish]

With these changes, both tests work fine. I observed no repercussions elsewhere, but more testing is recommended.

[zooey]

Replying to hamish:

With these changes, both tests work fine. I observed no repercussions elsewhere, but more testing is recommended.

Hm, is there some reasoning behind this patch, or did you get there by trial and error?

Please elaborate, as I don't quite understand the patch.

[hamish]

_IO_buf_base points to the start of the entire stream buffer, while _IO_read_base points to the start of the read part of the stream buffer. ungetc() and friends store characters in the bytes between _IO_buf_base and _IO_read_base, so the seek function must take these into account and skip them.

[zooey]

Replying to hamish:

_IO_buf_base points to the start of the entire stream buffer, while _IO_read_base points to the start of the read part of the stream buffer. ungetc() and friends store characters in the bytes between _IO_buf_base and _IO_read_base, so the seek function must take these into account and skip them.

Not quite, no: ungetc() invokes _IO_sputbackc(), which puts back characters between _IO_read_ptr and _IO_read_base - _IO_buf_base doesn't play any role there.

That's why I still fail to understand what exactly your patch does that causes the good effect. More importantly, I'm not at all sure it doesn't actually break other use cases. Have you tried running all gnulib tests to see if applying your patch doesn't introduce new problems?

[hamish]

Before putting back the characters it switches to the backup buffer, which moves _IO_read_base back into that area. Take a look at _IO_default_pbackfail.

I got tst-fseek.c from the latest glibc. After disabling all the stats-related tests (stat64 stuff missing), one fread test fails. Without my patch, four fseek tests also fail.

[zooey]

The proposed patch indeed fixes the problem and I have been unable to find any negative side effects.

Applied in hrev40055 - thanks a lot!