Bug 539

What GNU dd does with conv=noerror (without sync, that is)?

sync seems to be a dangerous option for image creation, it will result in extra NUL bytes if there are any short reads.

Typically there are no short reads, but I imagine if you use bs=1M and you hit a bad sector, there will be a short read and then a failed read. You do not want to pad short read result, you want it to be copied as-is, and then next read fails and dd seeks over 1M, writing 1M NULs to output. This at least preserves file size.

With conv=sync,noerror, there will be extra bytes: 1st 1M block is <data up to bad sector><NUL padding to 1M>, 2nd 1M block is 1M of NULs.

I plan to fix the bug in dd.c as follows:

                n = safe_read(ifd, ibuf, ibs);
                if (n == 0)
                        break;
                if (n < 0) {
                        if (!(flags & FLAG_NOERROR))
                                goto die_infile;
                        n = ibs;
                        bb_simple_perror_msg(infile);
==>                     /* GNU dd with conv=noerror skips over "bad blocks" */
==>                     xlseek(ifd, ibs, SEEK_CUR);
                }

Please try it.

I disagree.
With noerror:
If you get a short read you should seek to the end of the current block and then continue. The short-read block is padded with nulls (standard noerror operation).
Where sync comes in is that blocks for which no data is read at all are correctly inserted as nul blocks in the output (otherwise you break data structures).

From GNU dd - when a partial read occurs, a partial seek is performed:
"advance_input_after_read_error (input_blocksize - partread)"
Your suggestion may advance to a non-block start byte after a partial read.

I do not understand your last comment

Fixed in git, including correct behavior with and without conv=sync