Package: linux-image-3.16.0-4-amd64 Version: 3.16.7-ckt11-1 Initially discovered inside a QEMU guest, by doing the following:
# hdparm -t /dev/vda Under the Wheezy 3.2 kernel: Timing buffered disk reads: 384 MB in 3.01 seconds = 127.50 MB/sec Under the Jessie 3.16 kernel: Timing buffered disk reads: 46 MB in 3.07 seconds = 14.97 MB/sec After some work swapping kernels, I discovered that this behaviour exists in the 3.16 and 3.17 kernels, but not in 3.2->3.15 or 3.18+ Watching iostat as the I/O is happening indicated that the 3.16/3.17 guests were performing the I/O in 8-sector (512 bytes per sector) chunks; in the other kernels, the request sizes were 254 sectors instead. Some further work identified this patch in 3.18: https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/diff/?id=447f05bb488bff4282088259b04f47f0f9f76760 diff --git a/fs/block_dev.c b/fs/block_dev.c index 6d72746..e2f3ad08 100644 --- a/fs/block_dev.c +++ b/fs/block_dev.c @@ -304,6 +304,12 @@ static int blkdev_readpage(struct file * file, struct page * page) return block_read_full_page(page, blkdev_get_block); } +static int blkdev_readpages(struct file *file, struct address_space *mapping, + struct list_head *pages, unsigned nr_pages) +{ + return mpage_readpages(mapping, pages, nr_pages, blkdev_get_block); +} + static int blkdev_write_begin(struct file *file, struct address_space *mapping, loff_t pos, unsigned len, unsigned flags, struct page **pagep, void **fsdata) @@ -1622,6 +1628,7 @@ static int blkdev_releasepage(struct page *page, gfp_t wait) static const struct address_space_operations def_blk_aops = { .readpage = blkdev_readpage, + .readpages = blkdev_readpages, .writepage = blkdev_writepage, .write_begin = blkdev_write_begin, .write_end = blkdev_write_end, It applies cleanly to 3.16 and 3.17; with the patch applied, the larger request sizes are again seen and performance returns to the previous level. Could we apply this to the 3.16 kernel in Jessie? -- To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
