| Message ID | dc92d8ac746eaa95e5c22ca5e366b824c210a3f4.1628248828.git.asml.silence@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | fs: optimise generic_write_check_limits() | expand |
On Fri, Aug 06, 2021 at 12:22:10PM +0100, Pavel Begunkov wrote: > Even though ->s_maxbytes is used by generic_write_check_limits() only in > case of O_LARGEFILE, the value is loaded unconditionally, which is heavy > and takes 4 indirect loads. Optimise it by not touching ->s_maxbytes, > if it's not going to be used. Is this "optimisation" actually worth anything? Look at how force_o_largefile() is used. I would suggest that on the vast majority of machines, O_LARGEFILE is always set.
On Fri, Aug 06, 2021 at 12:22:10PM +0100, Pavel Begunkov wrote: > Even though ->s_maxbytes is used by generic_write_check_limits() only in > case of O_LARGEFILE, the value is loaded unconditionally, which is heavy > and takes 4 indirect loads. Optimise it by not touching ->s_maxbytes, > if it's not going to be used. Out of curiosity - how much of improvement have you actually seen on that? I can't say I hate the patch, but I'd like to see the data...
On 8/6/21 2:28 PM, Matthew Wilcox wrote: > On Fri, Aug 06, 2021 at 12:22:10PM +0100, Pavel Begunkov wrote: >> Even though ->s_maxbytes is used by generic_write_check_limits() only in >> case of O_LARGEFILE, the value is loaded unconditionally, which is heavy >> and takes 4 indirect loads. Optimise it by not touching ->s_maxbytes, >> if it's not going to be used. > > Is this "optimisation" actually worth anything? Look at how > force_o_largefile() is used. I would suggest that on the vast majority > of machines, O_LARGEFILE is always set. Makes sense to leave it alone then, thanks
From: Matthew Wilcox > Sent: 06 August 2021 14:28 > > On Fri, Aug 06, 2021 at 12:22:10PM +0100, Pavel Begunkov wrote: > > Even though ->s_maxbytes is used by generic_write_check_limits() only in > > case of O_LARGEFILE, the value is loaded unconditionally, which is heavy > > and takes 4 indirect loads. Optimise it by not touching ->s_maxbytes, > > if it's not going to be used. > > Is this "optimisation" actually worth anything? Look at how > force_o_largefile() is used. I would suggest that on the vast majority > of machines, O_LARGEFILE is always set. An option would be to only determine ->s_maxbytes when the size if larger than MAX_NON_LFS. So you'd end up with something like: if (pos >= max_size) { if (!(file->f_flags & O_LARGEFILE)) return -EFBIG; inode = file->f_mapping->host; if (pos >= inode->i_sb->s_maxbytes) return -EFBIG; } David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
On Sun, Aug 08, 2021 at 02:41:13PM +0000, David Laight wrote: > From: Matthew Wilcox > > Sent: 06 August 2021 14:28 > > > > On Fri, Aug 06, 2021 at 12:22:10PM +0100, Pavel Begunkov wrote: > > > Even though ->s_maxbytes is used by generic_write_check_limits() only in > > > case of O_LARGEFILE, the value is loaded unconditionally, which is heavy > > > and takes 4 indirect loads. Optimise it by not touching ->s_maxbytes, > > > if it's not going to be used. > > > > Is this "optimisation" actually worth anything? Look at how > > force_o_largefile() is used. I would suggest that on the vast majority > > of machines, O_LARGEFILE is always set. > > An option would be to only determine ->s_maxbytes when the size > if larger than MAX_NON_LFS. > > So you'd end up with something like: > > if (pos >= max_size) { > if (!(file->f_flags & O_LARGEFILE)) > return -EFBIG; > inode = file->f_mapping->host; > if (pos >= inode->i_sb->s_maxbytes) > return -EFBIG; > } You're optimising the part of the function that you can see in the diff instead of the whole function. And there's no evidence that there's much win to be had here ...
diff --git a/fs/read_write.c b/fs/read_write.c index 9db7adf160d2..db662d0c3cfa 100644 --- a/fs/read_write.c +++ b/fs/read_write.c @@ -1609,9 +1609,8 @@ SYSCALL_DEFINE6(copy_file_range, int, fd_in, loff_t __user *, off_in, */ int generic_write_check_limits(struct file *file, loff_t pos, loff_t *count) { - struct inode *inode = file->f_mapping->host; - loff_t max_size = inode->i_sb->s_maxbytes; loff_t limit = rlimit(RLIMIT_FSIZE); + loff_t max_size = MAX_NON_LFS; if (limit != RLIM_INFINITY) { if (pos >= limit) { @@ -1621,8 +1620,11 @@ int generic_write_check_limits(struct file *file, loff_t pos, loff_t *count) *count = min(*count, limit - pos); } - if (!(file->f_flags & O_LARGEFILE)) - max_size = MAX_NON_LFS; + if (file->f_flags & O_LARGEFILE) { + struct inode *inode = file->f_mapping->host; + + max_size = inode->i_sb->s_maxbytes; + } if (unlikely(pos >= max_size)) return -EFBIG;
Even though ->s_maxbytes is used by generic_write_check_limits() only in case of O_LARGEFILE, the value is loaded unconditionally, which is heavy and takes 4 indirect loads. Optimise it by not touching ->s_maxbytes, if it's not going to be used. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> --- fs/read_write.c | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-)