mmap (2) vs read (2)

mmap (2) vs read (2)

[Linh Dang]

Hi, how come reading memory from /dev/mem using pread(2) or mmap(2)
will give diffent results?

I run the little prog below 10 times and it always give the following
results

# for i in `seq 1 10`; do ./bar 0x4000000 ; done
04000000: 00000000
pread(04000000): 00cc7d83
04000000: 00000000
pread(04000000): 00cc7d83
04000000: 00000000
pread(04000000): 00cc7d83
04000000: 00000000
pread(04000000): 00cc7d83
04000000: 00000000
pread(04000000): 00cc7d83
04000000: 00000000
pread(04000000): 00cc7d83
04000000: 00000000
pread(04000000): 00cc7d83
04000000: 00000000
pread(04000000): 00cc7d83
04000000: 00000000
pread(04000000): 00cc7d83
04000000: 00000000
pread(04000000): 00cc7d83

thanx


----------------------------------- bar.c ---------------------------------
#include <unistd.h>
#include <stdio.h>
#include <inttypes.h>
#include <sys/types.h>
#include <sys/mman.h>
#include <fcntl.h>
#include <errno.h>


int
main(int argc, char** argv)
{
        uint32_t *addr, item;
        uint32_t offset, val;
        int fd, update = 0;
        if (argc < 2)
                return 1;

        offset  = strtoul(argv[1], NULL, 0);
        if (argc > 2) {
                val = strtoul(argv[2], NULL, 0);
                update = 1;
        } 

        fd   = open("/dev/mem", O_RDWR, 0666);
        addr = mmap(NULL, 4096, PROT_READ|PROT_WRITE, 0x8080|MAP_SHARED, fd, offset);

        if (addr)
        {
                printf("%8.8x: %8.8x", offset, *addr);
                if (update) {
                        *addr = val;
                        msync(addr, 4096, MS_SYNC);
                        printf("-> %8.8x\n", *addr);
                } else
                        printf("\n");

                pread(fd, &item, sizeof(item), offset);
                printf("pread(%8.8x): %8.8x\n", offset, item);

                munmap(addr, 4096);
        }

        close(fd);

}

Re: mmap (2) vs read (2)

[Arjan van de Ven]

[-- Attachment #1: Type: text/plain, Size: 451 bytes --]

On Sat, 2005-09-17 at 12:10 -0400, Linh Dang wrote:
> Hi, how come reading memory from /dev/mem using pread(2) or mmap(2)
> will give diffent results?

because you're being evil ;)

mmap of /dev/mem for *ram* is special. To avoid cache aliases and other
evils, you can only mmap non-ram realistically on /dev/mem.

Why are you using /dev/mem in the first place, it's a sure sign that
you're doing something really wrong in your design...


[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: mmap (2) vs read (2)

[Linh Dang]

Arjan van de Ven <arjanv@redhat.com> wrote:

> On Sat, 2005-09-17 at 12:10 -0400, Linh Dang wrote:
>> Hi, how come reading memory from /dev/mem using pread(2) or mmap(2)
>> will give diffent results?
>
> because you're being evil ;)
>
> mmap of /dev/mem for *ram* is special. To avoid cache aliases and
> other evils, you can only mmap non-ram realistically on /dev/mem.
>
> Why are you using /dev/mem in the first place, it's a sure sign that
> you're doing something really wrong in your design...

Thanx for the reply, what I'm doing is writing a driver (based on
mem.c) to export a block of ram to (other masters on) the PCI bus. The
driver does:

        1. get a contiguous block of ram using alloc_pages()
        2. export (via host-bridge hw setting) the block to the pci
           bus
        3. provide the .mmap() method in the driver to let userspace
           to mmap the device

In doing so, I encountered the inconsistencies of mmap(2) vs
read(2)/write(2). The work around I found is to SetPageReserved() on
all the pages got from alloc_pages(). But unfortunately I have no
clues why it's so. The vm code is not the easiest one to read
(compared to let's say the network code.)

-- 
Linh Dang