Date: Mon, 20 Aug 2012 08:24:38 -0400 From: John Baldwin <jhb@freebsd.org> To: freebsd-current@freebsd.org Cc: Poul-Henning Kamp <phk@phk.freebsd.dk>, Matt Jacob <mjacob@freebsd.org> Subject: Re: BUFSIZ = 1024, still ? Message-ID: <201208200824.38500.jhb@freebsd.org> In-Reply-To: <6800.1345323911@critter.freebsd.dk> References: <6800.1345323911@critter.freebsd.dk>
next in thread | previous in thread | raw e-mail | index | archive | help
On Saturday, August 18, 2012 5:05:11 pm Poul-Henning Kamp wrote:
> In message <5030033B.4060705@feral.com>, Matthew Jacob writes:
> >On 8/18/2012 1:32 PM, Poul-Henning Kamp wrote:
> >> Shouldn't we at least increase it to pagesize ?
> >>
> >What data suggests to you it would be better at pagesize?
> 
> The number of system calls to fwrite() a big file ?
Have you looked at an actual ktrace? :)  stdio doesn't use BUFSIZ for
regular files:
head/lib/libc/stdio/makebuf.c:
/*
 * Internal routine to determine `proper' buffering for a file.
 */
int
__swhatbuf(fp, bufsize, couldbetty)
        FILE *fp;
        size_t *bufsize;
        int *couldbetty;
{
        struct stat st;
        if (fp->_file < 0 || _fstat(fp->_file, &st) < 0) {
               ...
                *bufsize = BUFSIZ;
                return (__SNPT);
        }
        ...
        if (st.st_blksize <= 0) {
                *bufsize = BUFSIZ;
                return (__SNPT);
        }
        *bufsize = st.st_blksize;
        ...
}
For a regular file stdio will use the filesystem's block size, not BUFSIZ.
Test program:
#include <stdio.h>
#include <err.h>
int
main(int ac, char **av)
{
        char junk;
        FILE *fp;
        int i;
        fp = fopen("/tmp/junk", "w");
        if (fp == NULL)
                err(1, "fopen");
        for (i = 0; i < 1024 * 1024; i++)
                if (fwrite(&junk, sizeof(junk), 1, fp) != 1)
                        errx(1, "fwrite failed");
        return (0);
}
ktrace excerpt:
 42599 a.out    CALL  write(0x3,0x800a04000,0x4000)
 42599 a.out    RET   write 16384/0x4000
 42599 a.out    CALL  write(0x3,0x800a04000,0x4000)
 42599 a.out    RET   write 16384/0x4000
 42599 a.out    CALL  write(0x3,0x800a04000,0x4000)
 42599 a.out    RET   write 16384/0x4000
 42599 a.out    CALL  write(0x3,0x800a04000,0x4000)
 42599 a.out    RET   write 16384/0x4000
 42599 a.out    CALL  write(0x3,0x800a04000,0x4000)
 42599 a.out    RET   write 16384/0x4000
 42599 a.out    CALL  write(0x3,0x800a04000,0x4000)
 42599 a.out    RET   write 16384/0x4000
This hint also works for pipes (they set st_blksize to PAGE_SIZE).  Given 
that, I think BUFSIZ should just be left alone.
-- 
John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201208200824.38500.jhb>
