[ECOS] RedBoot xyModem bug

[Rutger Hofman <rutger at cs dot vu dot nl>]

Hello list,

I ran into a bug in RedBoot xyModem.c. It is in lines 420..434 in the 
current version (cvs 1.21):

    while (!xyz.at_eof && (size > 0)) {
... code for each block in the xyzModem protocols: ...

#if defined(xyzModem_zmodem) || defined(USE_YMODEM_LENGTH)
                        if (xyz.mode == xyzModem_xmodem || 
xyz.file_length == 0) {
#else
                        if (1) {
#endif
                            // Data blocks can be padded with ^Z (EOF) 
characters
                            // This code tries to detect and remove them
                            if ((xyz.bufp[xyz.len-1] == EOF) &&
                                (xyz.bufp[xyz.len-2] == EOF) &&
                                (xyz.bufp[xyz.len-3] == EOF)) {
                                while (xyz.len && (xyz.bufp[xyz.len-1] 
== EOF)) {
                                    xyz.len--;
                                }
                            }
                        }

What this does, is inspect each uploaded block for trailing ^Z = 0x1a 
bytes, and strip them. Rationale: xmodem or ymodem may signal 
end-of-file with ^Z a.k.a. 'CP/M EOF', and may optionally pad the block 
with more ^Z chars. This is live behaviour: sx and sy (0.12.20) under 
Linux do ^Z padding in the last block. The xyModem.c code above is 
wrong. ^Z chars are stripped at the end of *each block* in stead of at 
the end of the file.

This problem surfaced when I enabled zlib in eCos. In one of its tables, 
there is a sequence of bytes 26 = 0x1a = ^Z (in 
packages/services/compress/zlib/current/src/trees.h, lines 90, 91, 
112-114), which may trigger the test above. Depending on the offset in 
the file, these 0x1a's hit a block boundary or not, and the file is 
telescope-truncated or not. (I noticed that my executable crashed or not 
depending on code size, even depending on a number of 'nop' assembly 
instructions that I inserted anywhere in the code; it was a beast of a 
bug to pin down...)

The fix is nonobvious to me. For ymodem, if #define USE_YMODEM_LENGTH is 
on and a file length is specified, any trailing characters are removed 
elsewhere: this works correctly. However, for xmodem or ymodem without 
file length, any trailing ^Z at the end-of-file must still be removed. 
The code above is inherently unaware whether this is the last block or 
not: xmodem sends end-of-file control info, but that is consumed only in 
the *next* block read. The solution might be to maintain one more 
read-ahead block buffer to see if the current block is the last normal 
block.

Some random remarks:

As a workaround (I dare not flash the RedBoot in my device, because we 
bought it ready-made; I use the code base for my eCos application), I 
tried disabling the table in zlib. That lead to another problem, which I 
will report separately.

I am in the dark what must be done if a file to be xmodem'd actually 
contains trailing ^Z characters. There is nothing that forbids that 
outside CP/M or MS-DOS, I'd say.

I also wondered how this bug can have lived for so long in RedBoot. Or 
doesn't anyone use xyModem nowadays?

Rutger Hofman
VU Amsterdam

--
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss