On Thu, Jan 07, 2021 at 08:15:41AM -0500, Mikulas Patocka wrote: > Hi > > I announce a new version of NVFS - a filesystem for persistent memory. > http://people.redhat.com/~mpatocka/nvfs/ Utilities, AFAICS > git://leontynka.twibright.com/nvfs.git Seems to hang on git pull at the moment... Do you have it anywhere else? > I found out that on NVFS, reading a file with the read method has 10% > better performance than the read_iter method. The benchmark just reads the > same 4k page over and over again - and the cost of creating and parsing > the kiocb and iov_iter structures is just that high. Apples and oranges... What happens if you take ssize_t read_iter_locked(struct file *file, struct iov_iter *to, loff_t *ppos) { struct inode *inode = file_inode(file); struct nvfs_memory_inode *nmi = i_to_nmi(inode); struct nvfs_superblock *nvs = inode->i_sb->s_fs_info; ssize_t total = 0; loff_t pos = *ppos; int r; int shift = nvs->log2_page_size; size_t i_size; i_size = inode->i_size; if (pos >= i_size) return 0; iov_iter_truncate(to, i_size - pos); while (iov_iter_count(to)) { void *blk, *ptr; size_t page_mask = (1UL << shift) - 1; unsigned page_offset = pos & page_mask; unsigned prealloc = (iov_iter_count(to) + page_mask) >> shift; unsigned size; blk = nvfs_bmap(nmi, pos >> shift, &prealloc, NULL, NULL, NULL); if (unlikely(IS_ERR(blk))) { r = PTR_ERR(blk); goto ret_r; } size = ((size_t)prealloc << shift) - page_offset; ptr = blk + page_offset; if (unlikely(!blk)) { size = min(size, (unsigned)PAGE_SIZE); ptr = empty_zero_page; } size = copy_to_iter(to, ptr, size); if (unlikely(!size)) { r = -EFAULT; goto ret_r; } pos += size; total += size; } while (iov_iter_count(to)); r = 0; ret_r: *ppos = pos; if (file) file_accessed(file); return total ? total : r; } and use that instead of your nvfs_rw_iter_locked() in your ->read_iter() for DAX read case? Then the same with s/copy_to_iter/_copy_to_iter/, to see how much of that is "hardening" overhead. Incidentally, what's the point of sharing nvfs_rw_iter() for read and write cases? They have practically no overlap - count the lines common for wr and !wr cases. And if you do the same in nvfs_rw_iter_locked(), you'll see that the shared parts _there_ are bloody pointless on the read side. Not that it had been more useful on the write side, really, but that's another story (nvfs_write_pages() handling of copyin is... interesting). Let's figure out what's going on with the read overhead first... lib/iov_iter.c primitives certainly could use massage for better code generation, but let's find out how much of the PITA is due to those and how much comes from you fighing the damn thing instead of using it sanely...
