BTW, I have a WIP patch to xinstall to make it use copy_file_range. The patch works, but I never wrote a fallback path in case copy_file_range fails for some reason. Alex, would you be interested to finish the patch? -Alan
On Wed, Oct 14, 2020 at 7:35 AM Alexander Richardson < arichard...@freebsd.org> wrote: > On Wed, 14 Oct 2020 at 14:29, Mateusz Guzik <mjgu...@gmail.com> wrote: > > > > This should use copy_file_range (also available on Linux). > > > > I agree. I even mentioned this in > https://reviews.freebsd.org/D26041#589287. > This change avoids the two unnecessary syscalls, but I agree that > longer-term install should share the copy_file_range code with cp. > The only thing that copy_file_range won't speed up is the check > whether source and target are already identical. > > Alex > > > On 10/14/20, Alex Richardson <arichard...@freebsd.org> wrote: > > > Author: arichardson > > > Date: Wed Oct 14 12:28:41 2020 > > > New Revision: 366697 > > > URL: https://svnweb.freebsd.org/changeset/base/366697 > > > > > > Log: > > > install(1): Avoid unncessary fstatfs() calls and use mmap() based on > size > > > > > > According to git blame the trymmap() function was added in 1996 to > skip > > > mmap() calls for NFS file systems. However, nowadays mmap() should be > > > perfectly safe even on NFS. Importantly, onl ufs and cd9660 file > systems > > > were whitelisted so we don't use mmap() on ZFS. It also prevents the > use > > > of mmap() when bootstrapping from macOS/Linux since on those systems > the > > > trymmap() function was always returning zero due to the missing > > > MFSNAMELEN > > > define. > > > > > > This change keeps the trymmap() function but changes it to check > whether > > > using mmap() can reduce the number of system calls that are required. > > > Using mmap() only reduces the number of system calls if we need > multiple > > > read() > > > syscalls, i.e. if the file size is > MAXBSIZE. However, mmap() is > more > > > expensive > > > than read() so this sets the threshold at 4 fewer syscalls. > Additionally, > > > for > > > larger file size mmap() can significantly increase the number of page > > > faults, > > > so avoid it in that case. > > > > > > It's unclear whether using mmap() is ever faster than a read with an > > > appropriate > > > buffer size, but this change at least removes two unnecessary system > > > calls > > > for every file that is installed. > > > > > > Reviewed By: markj > > > Differential Revision: https://reviews.freebsd.org/D26041 > > > > > > Modified: > > > head/usr.bin/xinstall/xinstall.c > > > > > > Modified: head/usr.bin/xinstall/xinstall.c > > > > ============================================================================== > > > --- head/usr.bin/xinstall/xinstall.c Wed Oct 14 10:12:39 2020 > (r366696) > > > +++ head/usr.bin/xinstall/xinstall.c Wed Oct 14 12:28:41 2020 > (r366697) > > > @@ -148,7 +148,7 @@ static void metadata_log(const char *, const > char *, s > > > const char *, const char *, off_t); > > > static int parseid(const char *, id_t *); > > > static int strip(const char *, int, const char *, char **); > > > -static int trymmap(int); > > > +static int trymmap(size_t); > > > static void usage(void); > > > > > > int > > > @@ -1087,7 +1087,7 @@ compare(int from_fd, const char *from_name > __unused, > > > s > > > if (do_digest) > > > digest_init(&ctx); > > > done_compare = 0; > > > - if (trymmap(from_fd) && trymmap(to_fd)) { > > > + if (trymmap(from_len) && trymmap(to_len)) { > > > p = mmap(NULL, from_len, PROT_READ, MAP_SHARED, > > > from_fd, (off_t)0); > > > if (p == MAP_FAILED) > > > @@ -1248,13 +1248,8 @@ copy(int from_fd, const char *from_name, int > to_fd, > > > co > > > > > > digest_init(&ctx); > > > > > > - /* > > > - * Mmap and write if less than 8M (the limit is so we don't > totally > > > - * trash memory on big files. This is really a minor hack, but > it > > > - * wins some CPU back. > > > - */ > > > done_copy = 0; > > > - if (size <= 8 * 1048576 && trymmap(from_fd) && > > > + if (trymmap((size_t)size) && > > > (p = mmap(NULL, (size_t)size, PROT_READ, MAP_SHARED, > > > from_fd, (off_t)0)) != MAP_FAILED) { > > > nw = write(to_fd, p, size); > > > @@ -1523,20 +1518,23 @@ usage(void) > > > * return true (1) if mmap should be tried, false (0) if not. > > > */ > > > static int > > > -trymmap(int fd) > > > +trymmap(size_t filesize) > > > { > > > -/* > > > - * The ifdef is for bootstrapping - f_fstypename doesn't exist in > > > - * pre-Lite2-merge systems. > > > - */ > > > -#ifdef MFSNAMELEN > > > - struct statfs stfs; > > > - > > > - if (fstatfs(fd, &stfs) != 0) > > > - return (0); > > > - if (strcmp(stfs.f_fstypename, "ufs") == 0 || > > > - strcmp(stfs.f_fstypename, "cd9660") == 0) > > > - return (1); > > > -#endif > > > - return (0); > > > + /* > > > + * This function existed to skip mmap() for NFS file systems > whereas > > > + * nowadays mmap() should be perfectly safe. Nevertheless, using > mmap() > > > + * only reduces the number of system calls if we need multiple > read() > > > + * syscalls, i.e. if the file size is > MAXBSIZE. However, > mmap() is > > > + * more expensive than read() so set the threshold at 4 fewer > syscalls. > > > + * Additionally, for larger file size mmap() can significantly > increase > > > + * the number of page faults, so avoid it in that case. > > > + * > > > + * Note: the 8MB limit is not based on any meaningful > benchmarking > > > + * results, it is simply reusing the same value that was used > before > > > + * and also matches bin/cp. > > > + * > > > + * XXX: Maybe we shouldn't bother with mmap() at all, since we > use > > > + * MAXBSIZE the syscall overhead of read() shouldn't be too high? > > > + */ > > > + return (filesize > 4 * MAXBSIZE && filesize < 8 * 1024 * 1024); > > > } > > > _______________________________________________ > > > svn-src-all@freebsd.org mailing list > > > https://lists.freebsd.org/mailman/listinfo/svn-src-all > > > To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org" > > > > > > > > > -- > > Mateusz Guzik <mjguzik gmail.com> > _______________________________________________ svn-src-all@freebsd.org mailing list https://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"
