Re: [RFC v2 PATCH] mm: shmem: make stat.st_blksize return huge page size if THP is on

[Yang Shi]

On 4/22/18 6:47 PM, Michal Hocko wrote:
> On Sat 21-04-18 00:33:59, Yang Shi wrote:
> > Since tmpfs THP was supported in 4.8, hugetlbfs is not the only
> > filesystem with huge page support anymore. tmpfs can use huge page via
> > THP when mounting by "huge=" mount option.
> >
> > When applications use huge page on hugetlbfs, it just need check the
> > filesystem magic number, but it is not enough for tmpfs. Make
> > stat.st_blksize return huge page size if it is mounted by appropriate
> > "huge=" option.
> >
> > Some applications could benefit from this change, for example QEMU.
> > When use mmap file as guest VM backend memory, QEMU typically mmap the
> > file size plus one extra page. If the file is on hugetlbfs the extra
> > page is huge page size (i.e. 2MB), but it is still 4KB on tmpfs even
> > though THP is enabled. tmpfs THP requires VMA is huge page aligned, so
> > if 4KB page is used THP will not be used at all. The below /proc/meminfo
> > fragment shows the THP use of QEMU with 4K page:
> >
> > ShmemHugePages:   679936 kB
> > ShmemPmdMapped:        0 kB
> >
> > By reading st_blksize, tmpfs can use huge page, then /proc/meminfo looks
> > like:
> >
> > ShmemHugePages:    77824 kB
> > ShmemPmdMapped:     6144 kB
> >
> > statfs.f_bsize still returns 4KB for tmpfs since THP could be split, and it
> > also may fallback to 4KB page silently if there is not enough huge page.
> > Furthermore, different f_bsize makes max_blocks and free_blocks
> > calculation harder but without too much benefit. Returning huge page
> > size via stat.st_blksize sounds good enough.
> I am not sure I understand the above. So does QEMU or other tmpfs users
> rely on f_bsize to do mmap alignment tricks? Also I thought that THP

QEMU doesn't. It just check filesystem magic number now, if it is 
hugetlbfs, then it do mmap on huge page size alignment.

> will be used on the first aligned address even when the initial/last
> portion of the mapping is not THP aligned.

No, my test shows it is not. And, transhuge_vma_suitable() does check 
the virtual address alignment. If it is not huge page size aligned, it 
will not set PMD for huge page.

> And more importantly
> [...]
> > --- a/mm/shmem.c
> > +++ b/mm/shmem.c
> > @@ -39,6 +39,7 @@
> >   #include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
> >   
> >   static struct vfsmount *shm_mnt;
> > +static bool is_huge = false;
> >   
> >   #ifdef CONFIG_SHMEM
> >   /*
> > @@ -995,6 +996,8 @@ static int shmem_getattr(const struct path *path, struct kstat *stat,
> >   		spin_unlock_irq(&info->lock);
> >   	}
> >   	generic_fillattr(inode, stat);
> > +	if (is_huge)
> > +		stat->blksize = HPAGE_PMD_SIZE;
> >   	return 0;
> >   }
> >   
> > @@ -3574,6 +3577,7 @@ static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
> >   					huge != SHMEM_HUGE_NEVER)
> >   				goto bad_val;
> >   			sbinfo->huge = huge;
> > +			is_huge = true;
> Huh! How come this is a global flag. What if we have multiple shmem
> mounts some with huge pages enabled and some without? Btw. we seem to
> already have that information stored in the supperblock
> 		} else if (!strcmp(this_char, "huge")) {
> 			int huge;
> 			huge = shmem_parse_huge(value);
> 			if (huge < 0)
> 				goto bad_val;
> 			if (!has_transparent_hugepage() &&
> 					huge != SHMEM_HUGE_NEVER)
> 				goto bad_val;
> 			sbinfo->huge = huge;

Aha, my bad. I should used SHMEM_SB(inode->i_sb) to get shmem_sb_info 
then check the huge. Thanks a lot for catching this. Will fix in new 
version.

Yang