gzip receives read error, gives up, unlinks source file

[Andrew Brown]

my disk started throwing read errors last night, as i was trying to
compress some kernel dumps.  gzip (presumably) got a read error, but
didn't notice, and completed "successfully" (ie, it wrote the summary
information to the end of the output file, unlinked the input file,
and exited with zero).

this ain't good.

you can duplicate this behavior by running gzip under gdb with a file
large enough to incur multiple reads.  i used something a little over
130k.

    this# dd if=/dev/zero bs=1k count=132 of=compressme
    132+0 records in
    132+0 records out
    135168 bytes transferred in 0.009 secs (15018666 bytes/sec)
    this# ls -l compressme*
    -rw-r--r--  1 root  wsrc  135168 Mar 28 13:39 compressme
    this# gdb obj.i386/gzip
    ...
    (gdb) b read
    Breakpoint 1 at 0x8048f38
    (gdb) run compressme
    Starting program: /usr/src/gnu/usr.bin/gzip/obj.i386/gzip compressme
    Breakpoint 1 at 0xbdb50754

    Breakpoint 1, 0xbdb50754 in read () from /usr/lib/libc.so.12
    (gdb) c
    Continuing.

    Breakpoint 1, 0xbdb50754 in read () from /usr/lib/libc.so.12
    (gdb) c
    Continuing.

    Breakpoint 1, 0xbdb50754 in read () from /usr/lib/libc.so.12
    (gdb) ret
    Make selected stack frame return now? (y or n) y
    #0  0x804d082 in fill_window ()
        at /usr/src/gnu/usr.bin/gzip/deflate.c:557
    557             n = read_buf((char*)window+strstart+lookahead, more);
    (gdb) n
    Single stepping until exit from function read, 
    which has no line number information.
    558             if (n == 0 || n == (unsigned)EOF) {
    (gdb) set variable n=-1
    (gdb) set variable errno=5
    (gdb) d 1
    (gdb) c
    Continuing.

    Program exited normally.
    (gdb) 
    Suspended
    this# ls compressme*
    compressme.gz
    this# gunzip -l compressme.gz 
      compressed uncompressed  ratio uncompressed_name
             140        98304  99.8% compressme
    this# gunzip compressme.gz
    this# ls -l compressme 
    -rw-r--r--  1 root  wsrc  98304 Mar 28 13:36 compressme

the attached patch checks that isize (the amount read in) is the same
as st_size (from the struct stat information), and makes gzip react
accordingly.

-- 
|-----< "CODE WARRIOR" >-----|
address@hidden             * "ah!  i see you have the internet
address@hidden (Andrew Brown)                that goes *ping*!"
address@hidden       * "information is power -- share the wealth."

gzip.patch
Description: Text document