BUG() in 2.4: sg direct IO + exit()

["Tony Battersby" <tonyb@cybernetics.com>]

The following BUG() is triggered in 2.4.x when a program calls exit()
immediately after sending a SCSI command that uses direct IO:

kernel BUG at page_alloc.c:98!
invalid operand: 0000
sg sym53c8xx_2 e1000
CPU:    0
EIP:    0010:[<c0133d22>]    Not tainted
EFLAGS: 00010202
EIP is at __free_pages_ok+0x32/0x2b0 [kernel]
eax: 00000001   ebx: c12aceb0   ecx: c12aceb0   edx: 00000000
esi: 00000000   edi: 00000000   ebp: 00000001   esp: c02bbe94
ds: 0018   es: 0018   ss: 0018
Process swapper (pid: 0, stackpage=c02bb000)
Stack: cfefc640 cfdf2de8 cfbe93c0 cfefc600 00000001 00000000 cfefc600
c12aceb0
       cf926fa0 00000000 cfb8b040 c0127cb0 cfb8b04c cfb8b04c cfb8b000
d0885db6
       cf926fa0 00000000 cfb8b040 d0885cc9 cfb8b04c 00000001 cfd08680
cfd086c8
Call Trace:
 [<c0127cb0>] unmap_kiobuf+0x30/0x50 [kernel]
 [<d0885db6>] sg_unmap_and+0x26/0x50 [sg]
 [<d0885cc9>] sg_finish_rem_req+0x39/0x70 [sg]
 [<d0885451>] sg_cmd_done_bh+0x281/0x380 [sg]
 [<c01ab65a>] scsi_finish_command+0xda/0xe0 [kernel]
 [<c01ab380>] scsi_bottom_half_handler+0xc0/0x230 [kernel]
 [<c011dacb>] bh_action+0x4b/0x90 [kernel]
 [<c011d971>] tasklet_hi_action+0x61/0xa0 [kernel]
 [<c011d6fb>] do_softirq+0x6b/0xd0 [kernel]
 [<c0108baf>] do_IRQ+0xdf/0xf0 [kernel]
 [<c01052c0>] default_idle+0x0/0x40 [kernel]
 [<c01052c0>] default_idle+0x0/0x40 [kernel]
 [<c010b388>] call_do_IRQ+0x5/0xd [kernel]
 [<c01052c0>] default_idle+0x0/0x40 [kernel]
 [<c01052c0>] default_idle+0x0/0x40 [kernel]
 [<c01052ec>] default_idle+0x2c/0x40 [kernel]
 [<c0105372>] cpu_idle+0x52/0x70 [kernel]
 [<c0105000>] stext+0x0/0x50 [kernel]

Code: 0f 0b 62 00 ab 04 23 c0 89 d8 e8 2f ed ff ff 8b 7b 28 85 ff
 <0>Kernel panic: Aiee, killing interrupt handler!
In interrupt handler - not syncing

The following program reproduces the problem:

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <unistd.h>
#include <scsi/sg.h>
#include <limits.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>

unsigned char mode_sense_cdb[] = { 0x1a, 0x00, 0x3f, 0x00, 0xff, 0x00 };

int main(int argc, char *argv[])
{
    int fd;
    unsigned char sense_data[256];
    unsigned char buf[256];
    sg_io_hdr_t io;

    fd = open("/dev/sg0", O_RDWR);
    if (fd == -1) {
        perror("open(/dev/sg0)");
        exit(1);
    }

    memset(&io, 0, sizeof(sg_io_hdr_t));
    io.interface_id = 'S';
    io.timeout = UINT_MAX;
    io.flags = SG_FLAG_DIRECT_IO;
    io.sbp = sense_data;
    io.mx_sb_len = 0xff;
    io.cmdp = mode_sense_cdb;
    io.cmd_len = 6;
    io.dxfer_direction = SG_DXFER_FROM_DEV;
    io.dxferp = buf;
    io.dxfer_len = 0xff;
    write(fd, &io, sizeof(sg_io_hdr_t));

    return 0;
}

Just compile, echo 1 > /proc/scsi/sg/allow_dio, and run.  The SCSI
device that I am using is a tape drive, but it shouldn't matter.

I have tested this with 2.4.21+kksymoops and 2.4.26-pre5+kksymoops with
the same results.  The above stack trace is from 2.4.26-pre5.  I haven't
tried 2.6.

I am running a SMP kernel on a UP Intel P4 with 256 MB RAM.  The test
machine does not have any swap space configured (swapoff -a).

Anthony J. Battersby
Cybernetics