On Mon, May 15, 2023 at 09:20:44AM +0800, xiubli@xxxxxxxxxx wrote:
> From: Xiubo Li <xiubli@xxxxxxxxxx>
>
> Blindly expanding the readahead windows will cause unneccessary
> pagecache thrashing and also will introdue the network workload.
s/introdue/introduce/
> We should disable expanding the windows if the readahead is disabled
> and also shouldn't expand the windows too much.
>
> Expanding forward firstly instead of expanding backward for possible
> sequential reads.
>
> Bound `rreq->len` to the actual file size to restore the previous page
> cache usage.
>
> The posix_fadvise may change the maximum size of a file readahead.
>
> Cc: stable@xxxxxxxxxxxxxxx
> Fixes: 49870056005c ("ceph: convert ceph_readpages to ceph_readahead")
> URL: https://lore.kernel.org/ceph-devel/20230504082510.247-1-sehuww@xxxxxxxxxxxxxxxx
> URL: https://www.spinics.net/lists/ceph-users/msg76183.html
> Cc: Hu Weiwen <sehuww@xxxxxxxxxxxxxxxx>
> Signed-off-by: Xiubo Li <xiubli@xxxxxxxxxx>
> ---
> fs/ceph/addr.c | 40 +++++++++++++++++++++++++++++++++-------
> 1 file changed, 33 insertions(+), 7 deletions(-)
>
> diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> index 93fff1a7373f..4b29777c01d7 100644
> --- a/fs/ceph/addr.c
> +++ b/fs/ceph/addr.c
> @@ -188,16 +188,42 @@ static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
> struct inode *inode = rreq->inode;
> struct ceph_inode_info *ci = ceph_inode(inode);
> struct ceph_file_layout *lo = &ci->i_layout;
> + unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
> + loff_t end = rreq->start + rreq->len, new_end;
> + struct ceph_netfs_request_data *priv = rreq->netfs_priv;
> + unsigned long max_len;
> u32 blockoff;
> - u64 blockno;
>
> - /* Expand the start downward */
> - blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
> - rreq->start = blockno * lo->stripe_unit;
> - rreq->len += blockoff;
> + if (priv) {
> + /* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
> + if (priv->file_ra_disabled)
> + max_pages = 0;
> + else
> + max_pages = priv->file_ra_pages;
> +
> + }
> +
> + /* Readahead is disabled */
> + if (!max_pages)
> + return;
>
> - /* Now, round up the length to the next block */
> - rreq->len = roundup(rreq->len, lo->stripe_unit);
> + max_len = max_pages << PAGE_SHIFT;
> +
> + /*
> + * Try to expand the length forward by rounding up it to the next
An extra space between "rounding up".
Apart from above two typo, LGTM.
Reviewed-by: Hu Weiwen <sehuww@xxxxxxxxxxxxxxxx>
I also tested this patch with our workload. Reading the first 16k images
from ImageNet dataset (1.69GiB) takes about 1.8Gi page cache (as
reported by `free -h'). This is expected.
For the fadvise use-case, I use `fio' to do the test:
$ fio --name=rand --size=32M --fadvise_hint=1 --ioengine=libaio --iodepth=128 --rw=randread --bs=4k --filesize=2G
after the test, page cache increased by about 35Mi, which is expected.
So if appropriate:
Tested-by: Hu Weiwen <sehuww@xxxxxxxxxxxxxxxx>
However, also note random reading to a large file without fadvise still
suffers from degradation. e.g., this test:
$ fio --name=rand --size=32M --fadvise_hint=0 --ioengine=libaio --iodepth=128 --rw=randread --bs=4k --filesize=2G
will load nearly every page of the 2Gi test file into page cache,
although I only need 32Mi of them.
> + * block, but do not exceed the file size, unless the original
> + * request already exceeds it.
> + */
> + new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
> + if (new_end > end && new_end <= rreq->start + max_len)
> + rreq->len = new_end - rreq->start;
> +
> + /* Try to expand the start downward */
> + div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
> + if (rreq->len + blockoff <= max_len) {
> + rreq->start -= blockoff;
> + rreq->len += blockoff;
> + }
> }
>
> static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
> --
> 2.40.1
>
