From: Haolin Xue <[email protected]> In tac_seekable, after the read-ahead memmove that shifts the pending record to make room for newly read data, match_start was set to G_buffer + read_size without checking whether the pending record (saved_record_size) was at least match_length - 1 bytes long. When saved_record_size < match_length - 1 (which can happen when processing a second or later file operand whose final chunk is shorter than the separator), the backward STREQ_LEN scan can read past past_end into leftover bytes from a previous file. If those leftover bytes happen to match the separator, a spurious match is accepted whose match_end exceeds past_end, causing output() to underflow bytes_to_add and read past the G_buffer allocation.
Fix this by capping match_start at past_end - match_length1, so the STREQ_LEN scan never reads past past_end. When saved_record_size >= match_length - 1 (the common case), the cap is a no-op and the existing optimization is preserved. * src/tac.c (tac_seekable): Cap match_start at past_end - match_length1 after the read-ahead memmove. * tests/tac/tac.pl: Add regression test 'oob-multibyte-sep'. Reported at https://bugs.gnu.org/82117. Signed-off-by: Haolin Xue <[email protected]> --- src/tac.c | 10 ++++++++-- tests/tac/tac.pl | 9 +++++++++ 2 files changed, 17 insertions(+), 2 deletions(-) diff --git a/src/tac.c b/src/tac.c index 1845791ab..e7d731231 100644 --- a/src/tac.c +++ b/src/tac.c @@ -341,9 +341,15 @@ tac_seekable (int input_fd, char const *file, off_t file_pos) The source and destination regions probably overlap. */ memmove (G_buffer + read_size, G_buffer, saved_record_size); past_end = G_buffer + read_size + saved_record_size; - /* For non-regexp searches, avoid unnecessary scanning. */ + /* For non-regexp searches, avoid unnecessary scanning. + Also cap match_start so that the backward STREQ_LEN scan + cannot read past past_end: a match starting near the + boundary with the pending record would otherwise read + beyond the valid data when saved_record_size is less + than match_length - 1. */ if (sentinel_length) - match_start = G_buffer + read_size; + match_start = MIN (G_buffer + read_size, + past_end - match_length1); else match_start = past_end; diff --git a/tests/tac/tac.pl b/tests/tac/tac.pl index 99d2d0d50..5a9022191 100755 --- a/tests/tac/tac.pl +++ b/tests/tac/tac.pl @@ -82,6 +82,15 @@ my @Tests = # coreutils-8.5's tac would double-free its primary buffer. ['double-free', {IN=>$long_line}, {OUT=>$long_line}], + + # bug#82117: with a multi-byte separator (length >= 3) and at least + # two file operands, tac could read past the buffer. When the second + # file's pending record (after the read-ahead memmove) is shorter than + # match_length - 1, the backward STREQ_LEN scan reads past past_end, + # and a spurious match causes output() to underflow bytes_to_add. + ['oob-multibyte-sep', '-s XYX', + {IN=>'A' x 8193 . 'X'}, {IN=>'A' x 8191 . 'XY'}, + {OUT=>'A' x 8193 . 'X' . 'A' x 8191 . 'XY'}], ); @Tests = triple_test \@Tests; -- 2.43.0
