write(2) and >=4GB buffer ... ubc_uiomove: error=14

[Nicolas Joly <njoly%pasteur.fr@localhost>]

Hi,

I noticed a few copies of the same log message on one of my -current
NetBSD/amd64 boxes: `ubc_uiomove: error=14'.

After some researches, i discovered that they appear with write(2)
calls + large buffers (>=4GB). It does not seems to be a write(2)
syscall limitation (up to SSIZE_MAX), but rather a filesystem problem.

The attached testcase, fails for if writing to a `dummy.big' file on
ffs; but succeed if writing to the `/dev/null' device.

   575      1 bigwrite  CALL  open(0x400bbf,0x602,0x1b6)
   575      1 bigwrite  NAMI  "dummy.big"
   575      1 bigwrite  RET   open 3
   575      1 bigwrite  CALL  write(3,0x7f7efdb00000,0x100000000)
   575      1 bigwrite  RET   write -1 errno 14 Bad address

   558      1 bigwrite  CALL  open(0x400bbf,0x602,0x1b6)
   558      1 bigwrite  NAMI  "/dev/null"
   558      1 bigwrite  RET   open 3
   558      1 bigwrite  CALL  write(3,0x7f7efdb00000,0x100000000)
   558      1 bigwrite  RET   write 4294967296/0x100000000

I'm convinced that trying to write such many bytes at once is not
ideal, but at the same time i wonder if we support such cases ...

Thanks in advance,
Regards.

NB: Linux/x86_64 has no problems with this testcase on ext3.

-- 
Nicolas Joly

Biological Software and Databanks.
Institut Pasteur, Paris.

#include <err.h>
#include <fcntl.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>

int main() {
  int fd, res;
  size_t len;
  ssize_t cnt;
  void *buf;

  len = 4L * 1024 * 1024 * 1024;
  buf = malloc(len);
  if (buf == NULL)
    err(1, "malloc failed");

  fd = open("dummy.big", O_RDWR|O_CREAT|O_TRUNC, 0666);
  if (fd == -1)
    err(1, "open failed");

  cnt = write(fd, buf, len);
  if (cnt == -1)
    err(1, "write failed");
  if ((size_t)cnt != len)
    err(1, "write bogus data");

  res = close(fd);
  if (res == -1)
    err(1, "close failed");

  return 0; }