: extremely well. So well, in fact, that I can have a program which
: mmap()'s a large file and continuously scans it - generating 4MB/sec of
: network traffic, with virtually no effect to the rest of the system.
Oh, clarification: continuously scans it but also calls
madvise(... MADV_DONTNEED) on the pages once it is through, that is.
If you mmap and scan a file without any madvise's it still loads the
system down just like it always has.
The program below simulates scanning a file sequentially via a mmap
which also uses the MADV_DONTNEED feature.
I did find one disadvantage to my patch - easily fixed (not enough to
generate a new patch before this one goes in) - and that is by moving
pages to the cache a lot of unnecessary page faults can be incurred.
What we want to do, in fact, is for the pages we would normally move
to the cache we instead move to the *head* of the inactive list (rather
then the tail, which is what normally happens). This disadvantage
becomes apparent when you run the below program on a small file that
easily fits in main memory.
-Matt
#include <sys/types.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <stdio.h>
#include <fcntl.h>
#include <stdlib.h>
#include <stdarg.h>
#include <unistd.h>
volatile char c;
int
main(int ac, char **av)
{
int fd;
int i;
int j;
int pgsize = getpagesize();
int pgmask = pgsize - 1;
char *ptr;
struct stat st;
if (ac == 1) {
printf("%s filename\n", av[0]);
exit(1);
}
if ((fd = open(av[1], O_RDONLY)) < 0) {
perror("open");
exit(1);
}
fstat(fd, &st);
ptr = mmap(NULL, st.st_size, PROT_READ, MAP_SHARED, fd, 0);
madvise(ptr, st.st_size & ~pgmask, MADV_SEQUENTIAL);
if (ptr == (char *)-1) {
perror("mmap");
exit(1);
}
for (i = j = 0; i < (int)st.st_size; (i += pgsize), ++j) {
c = ptr[i];
if (i && (j & 7) == 0)
madvise(ptr + i - pgsize * 8, pgsize * 8, MADV_DONTNEED);
}
madvise(ptr, st.st_size & ~pgmask, MADV_DONTNEED);
return(0);
}
To Unsubscribe: send mail to majord...@freebsd.org
with "unsubscribe freebsd-hackers" in the body of the message
