On Mon, Apr 08, 2024 at 04:02:17PM +0100, Al Viro wrote:
> On Sun, Apr 07, 2024 at 09:11:20PM +0100, Matthew Wilcox (Oracle) wrote:
> > -static inline bool is_file_hugepages(struct file *file)
> > +static inline bool is_file_hugepages(const struct file *file)
> > {
> > - if (file->f_op == &hugetlbfs_file_operations)
> > - return true;
> > -
> > - return is_file_shm_hugepages(file);
> > + return file->f_op->fop_flags & FOP_HUGE_PAGES;
> > }
>
> Extra cacheline to pull can be costly on a sufficiently hot path...
Sure, but so can a function call, particularly with the $%^&@#
CPU mitigations. Yes, is_file_hugepages() is inline, but
is_file_shm_hugepages() is not. The cacheline in question should be
shared since it's part of fops, which is const.
I'm not seeing is_file_hugepages() called on any partcularly hot paths.
Most of the callers are in mmap() equivalent paths.
