the commit (found by bisect)

  commit 18a25da84354c6bb655320de6072c00eda6eb602
  Author: NeilBrown <ne...@suse.com>
  Date:   Wed Sep 6 09:43:28 2017 +1000

    dm: ensure bio submission follows a depth-first tree walk

cause serious regression while reading from DM device.

The reproducer is below, basically it tries to simulate failure we see in 
regression test: we have DM device with error and zero target and try to read
"passphrase" from it (it is test for 64 bit offset error path):

Test device:
# dmsetup table test
0 10000000 error 
10000000 1000000 zero 

We try to run this operation:
  lseek64(fd, 5119999988, SEEK_CUR); // this should seek to error target sector
  read(fd, buf, 13); // this should fail, if we seek to error part of the device

While on 4.15 the read properly fails:
  Seek returned 5119999988.
  Read returned -1.

for 4.16 it actually succeeds returning some random data
(perhaps kernel memory, so this bug is even more dangerous):
  Seek returned 5119999988.
  Read returned 13.

Full reproducer below:

#define _GNU_SOURCE
#include <stdio.h>
#include <stddef.h>
#include <stdint.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <inttypes.h>

int main (int argc, char *argv[])
        char buf[13];
        int fd;
        //uint64_t offset64 = 5119999999;
        uint64_t offset64 =   5119999988;
        off64_t r;
        ssize_t bytes;

        system("echo -e \'0 10000000 error\'\\\\n\'10000000 1000000 zero\' | 
dmsetup create test");

        fd = open("/dev/mapper/test", O_RDONLY);
        if (fd == -1) {
                printf("open fail\n");
                return 1;

        r = lseek64(fd, offset64, SEEK_CUR);
        printf("Seek returned %" PRIu64 ".\n", r);
        if (r < 0) {
                printf("seek fail\n");
                return 2;

        bytes = read(fd, buf, 13);
        printf("Read returned %d.\n", (int)bytes);

        return 0;

Please let me know if you need more info to reproduce it.


Reply via email to