| Message ID | 20180228231922.15796-1-linux@rasmusvillemoes.dk (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 2018-03-01 00:19, Rasmus Villemoes wrote: > I noticed that offsetof(struct filename, iname) is actually 28 on 64 > bit platforms, so we always pass an unaligned pointer to > strncpy_from_user. This is mostly a problem for those 64 bit platforms > without HAVE_EFFICIENT_UNALIGNED_ACCESS, but even on x86_64, unaligned > accesses carry a penalty. > > A user-space microbenchmark doing nothing but strncpy_from_user from the > same (aligned) source string runs about 5% faster when the destination > is aligned. That number increases to 20% when the string is long > enough (~32 bytes) that we cross a cache line boundary - that's for > example the case for about half the files a "git status" in a kernel > tree ends up stat'ing. > > This won't make any real-life workloads 5%, or even 1%, faster, but path > lookup is common enough that cutting even a few cycles should be > worthwhile. So ensure we always pass an aligned destination pointer to > strncpy_from_user. Instead of explicit padding, simply swap the refcnt > and aname members, as suggested by Al Viro. polite ping...
diff --git a/fs/namei.c b/fs/namei.c index 921ae32dbc80..5a66e7ca5d60 100644 --- a/fs/namei.c +++ b/fs/namei.c @@ -39,6 +39,7 @@ #include <linux/bitops.h> #include <linux/init_task.h> #include <linux/uaccess.h> +#include <linux/build_bug.h> #include "internal.h" #include "mount.h" @@ -130,6 +131,7 @@ getname_flags(const char __user *filename, int flags, int *empty) struct filename *result; char *kname; int len; + BUILD_BUG_ON(offsetof(struct filename, iname) % sizeof(long) != 0); result = audit_reusename(filename); if (result) diff --git a/include/linux/fs.h b/include/linux/fs.h index 2a815560fda0..d7b2caadb292 100644 --- a/include/linux/fs.h +++ b/include/linux/fs.h @@ -2380,8 +2380,8 @@ struct audit_names; struct filename { const char *name; /* pointer to actual string */ const __user char *uptr; /* original userland pointer */ - struct audit_names *aname; int refcnt; + struct audit_names *aname; const char iname[]; };
I noticed that offsetof(struct filename, iname) is actually 28 on 64 bit platforms, so we always pass an unaligned pointer to strncpy_from_user. This is mostly a problem for those 64 bit platforms without HAVE_EFFICIENT_UNALIGNED_ACCESS, but even on x86_64, unaligned accesses carry a penalty. A user-space microbenchmark doing nothing but strncpy_from_user from the same (aligned) source string runs about 5% faster when the destination is aligned. That number increases to 20% when the string is long enough (~32 bytes) that we cross a cache line boundary - that's for example the case for about half the files a "git status" in a kernel tree ends up stat'ing. This won't make any real-life workloads 5%, or even 1%, faster, but path lookup is common enough that cutting even a few cycles should be worthwhile. So ensure we always pass an aligned destination pointer to strncpy_from_user. Instead of explicit padding, simply swap the refcnt and aname members, as suggested by Al Viro. Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk> --- fs/namei.c | 2 ++ include/linux/fs.h | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-)