Duplicated output when calling recode_perform_task with an output stream and previously buffered output

[oreo639]

This affects enca 1.19 when data is passed through stdin when linked with librecode 3.7.15. (This issue doesn't happen when enca is linked against 3.6 librecode)

https://github.com/nijel/enca/blob/6fc4a2a3e9ec16b892e325f6f53b2501fbcacb17/src/convert_recode.c#L145-L201

When librecode is built against libiconv and enca -VVVVV -L cs -C librecode -x UTF-8 <test-recode.actual >test-recode.tmp is run with the following input, the output to test-recode.tmp is duplicated: https://github.com/nijel/enca/blob/6fc4a2a3e9ec16b892e325f6f53b2501fbcacb17/test/cs-s.iso88592

When performing the two calls individually, the first call outputs the correct conversion once, the second call outputs nothing, however when run in sucession, the first call outputs the correct string and the second call results in it being duplicated, this is despite recode_put_bytes()/recode_put_byte() never getting called for the second conversion. (stepping through it with gdb, it enters transform_ucs2_utf8(), gets the first byte which is EOF and exits, which is correct)

This duplication doesn't seem to happen with glibc's iconv, where the recode_transform_with_iconv() codepath is taken instead.

For whatever reason, calling fflush(stdout); between the two calls appears to stop the duplication.

[oreo639]

Stepping through the program, the first half of the duplicated output gets written when fork () is called. It appears the buffered output ends up getting duplicated after calling fork ().

Adding if (task->output.file) fflush (task->output.file); right before the fork does seem to allow the enca tests to pass, whereas adding it right after the fork results in the output being duplicated.

I can also reproduce the duplicated buffered output when calling recode_perform_task () once to stdout after doing a printf without a newline, which outputted Hello worldHello world<converted text here> despite only doing the Hello world printf once. Adding fflush resulted in Hello world<converted text here> as expected.

This doesn't affect librecode 3.6, since there, the pipe codepath was only used if the consumer explicitly enabled it, which enca doesn't. Building with HAS_PIPE undefined in that file also allows it to work (with the fclose in that path adjusted to fix closing output streams from the parent process).

[rrthomas]

Fantastic, thanks for chasing this down, and for the patch.

[rrthomas]

I have fixed this issue by removing the pipe method.