On Thu, Apr 21, 2022 at 8:06 AM Borislav Petkov <bp@xxxxxxxxx> wrote:
>
> on AMD zen3
>
> original: 20.11 Gb/s
> rep_good: 34.662 Gb/s
> erms: 36.378 Gb/s
> fsrm: 36.398 Gb/s
Looks good.
Of course, the interesting cases are the "took a page fault in the middle" ones.
A very simple basic test is something like the attached.
It does no error checking or anything else, but doing a 'strace
./a.out' should give you something like
...
openat(AT_FDCWD, "/dev/zero", O_RDONLY) = 3
mmap(NULL, 196608, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS,
-1, 0) = 0x7f10ddfd0000
munmap(0x7f10ddfe0000, 65536) = 0
read(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0", 65536) = 16
exit_group(16) = ?
where that "read(..) = 16" is the important part. It correctly figured
out that it can only do 16 bytes (ok, 17, but we've always allowed the
user accessor functions to block).
With erms/fsrm, presumably you get that optimal "read(..) = 17".
I'm sure we have a test-case for this somewhere, but it was easier for
me to write a few lines of (bad) code than try to find it.
Linus
#include <stddef.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/mman.h>
// Whatever
#define PAGE_SIZE 65536
int main(int argc, char **argv)
{
int fd;
void *map;
fd = open("/dev/zero", O_RDONLY);
map = mmap(NULL, 3*PAGE_SIZE, PROT_READ | PROT_WRITE,
MAP_PRIVATE |MAP_ANONYMOUS , -1, 0);
munmap(map + PAGE_SIZE, PAGE_SIZE);
return read(fd, map + PAGE_SIZE - 17, PAGE_SIZE);
}
