On Tue, 16 Sep 2014 21:32:43 -0400 Trond Myklebust <trond.mykleb...@primarydata.com> wrote:
> On Tue, Sep 16, 2014 at 9:10 PM, NeilBrown <ne...@suse.de> wrote:
> >
> > However ... something else occurs to me. We could use the bdi congestion
> > markers to guide the timeout.
> > When the wait for PG_private times out, or when a connection
> > re-establishment
> > is required (and maybe other similar times) we could set_bdi_congested().
> > Then in nfs_release_page() we could completely avoid the wait if
> > bdi_write_congested().
> >
> > The congestion setting should encourage vmscan away from the filesystem so
> > it
> > won't keep calling nfs_release_page() which is a bonus.
> >
> > Setting bdi_congestion from the RPC layer might be awkward from a layering
> > perspective, but probably isn't necessary.
> >
> > Would the following allay your concerns? The change to
> > nfs_inode_remove_request ensures that any congestion is removed when a
> > 'commit' completes.
> >
> > We certainly could keep the PF_FSTRANS setting in the SUNRPC layer - that
> > was
> > why it was a separate patch. It would be nice to find a uniform solution
> > though.
> >
> > Thanks,
> > NeilBrown
> >
> >
> >
> > diff --git a/fs/nfs/file.c b/fs/nfs/file.c
> > index 5949ca37cd18..bc674ad250ce 100644
> > --- a/fs/nfs/file.c
> > +++ b/fs/nfs/file.c
> > @@ -477,10 +477,15 @@ static int nfs_release_page(struct page *page, gfp_t
> > gfp)
> > * benefit that someone else can worry about the freezer.
> > */
> > if (mapping) {
> > + struct nfs_server *nfss = NFS_SERVER(mapping->host);
> > nfs_commit_inode(mapping->host, 0);
> > - if ((gfp & __GFP_WAIT))
> > + if ((gfp & __GFP_WAIT) &&
> > + !bdi_write_congested(&nfss->backing_dev_info))
> > wait_on_page_bit_killable_timeout(page, PG_private,
> > HZ);
> > + if (PagePrivate(page))
> > + set_bdi_congested(&nfss->backing_dev_info,
> > + BLK_RW_ASYNC);
> > }
> > /* If PagePrivate() is set, then the page is not freeable */
> > if (PagePrivate(page))
> > diff --git a/fs/nfs/write.c b/fs/nfs/write.c
> > index 700e7a865e6d..3ab122e92c9d 100644
> > --- a/fs/nfs/write.c
> > +++ b/fs/nfs/write.c
> > @@ -726,6 +726,7 @@ static void nfs_inode_remove_request(struct nfs_page
> > *req)
> > struct inode *inode = req->wb_context->dentry->d_inode;
> > struct nfs_inode *nfsi = NFS_I(inode);
> > struct nfs_page *head;
> > + struct nfs_server *nfss = NFS_SERVER(inode);
> >
> > if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
> > head = req->wb_head;
> > @@ -742,6 +743,9 @@ static void nfs_inode_remove_request(struct nfs_page
> > *req)
> > spin_unlock(&inode->i_lock);
> > }
> >
> > + if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
> > + clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
>
> Hmm.... We already have this equivalent functionality in
> nfs_end_page_writeback(), so adding it to nfs_inode_remove_request()
> is just causing duplication as far as the stable writeback path is
> concerned. How about adding it to nfs_commit_release_pages() instead?
>
> Otherwise, yes, the above does indeed look at if it has merit. Have
> you got a good test?
>
Altered patch below. I'll post a proper one after some testing.
For testing I create a memory pressure load with:
=============================
#!/bin/bash
umount /mnt/ramdisk
umount /mnt/ramdisk
mount -t tmpfs -o size=4G none /mnt/ramdisk
#swapoff -a
i=0
while [ $i -le 10000 ]; do
i=$(($i+1))
dd if=/dev/zero of=/mnt/ramdisk/testdata.dd bs=1M count=6500
date
done
==============================
Where the '4G' matches memory size, and then write out to an NFS file with
=================================
#!/bin/bash
umount /mnt2 /mnt3
umount /mnt2 /mnt3
mount /dev/sdd /mnt2
exportfs -avu
exportfs -av
mount $* 127.0.0.1:/mnt2 /mnt3
for j in {1..100}; do
i=1
while [ $i -le 10000 ]; do
echo "Step $i"
date +%H:%M:%S
i=$(($i+1))
zcat /boot/vmlinux-3.13.3-1-desktop.gz | uuencode -
date +%H:%M:%S
done | dd of=/mnt3/testdat.file bs=1M
done
====================================
Pick your own way to create a large file of random data .. though
probably /dev/zero would do.
With both those going for a few hours the current kernel will deadlock.
With my patches it doesn't.
I'll see if I can come up with some way to measure maximum delay in
try_to_free_pages() and see how the 'congestion' change affects that.
Thanks,
NeilBrown
----------------------
diff --git a/fs/nfs/file.c b/fs/nfs/file.c
index 5949ca37cd18..bc674ad250ce 100644
--- a/fs/nfs/file.c
+++ b/fs/nfs/file.c
@@ -477,10 +477,15 @@ static int nfs_release_page(struct page *page, gfp_t gfp)
* benefit that someone else can worry about the freezer.
*/
if (mapping) {
+ struct nfs_server *nfss = NFS_SERVER(mapping->host);
nfs_commit_inode(mapping->host, 0);
- if ((gfp & __GFP_WAIT))
+ if ((gfp & __GFP_WAIT) &&
+ !bdi_write_congested(&nfss->backing_dev_info))
wait_on_page_bit_killable_timeout(page, PG_private,
HZ);
+ if (PagePrivate(page))
+ set_bdi_congested(&nfss->backing_dev_info,
+ BLK_RW_ASYNC);
}
/* If PagePrivate() is set, then the page is not freeable */
if (PagePrivate(page))
diff --git a/fs/nfs/write.c b/fs/nfs/write.c
index 700e7a865e6d..8d4aae9d977a 100644
--- a/fs/nfs/write.c
+++ b/fs/nfs/write.c
@@ -1641,6 +1641,7 @@ static void nfs_commit_release_pages(struct
nfs_commit_data *data)
struct nfs_page *req;
int status = data->task.tk_status;
struct nfs_commit_info cinfo;
+ struct nfs_server *nfss;
while (!list_empty(&data->pages)) {
req = nfs_list_entry(data->pages.next);
@@ -1674,6 +1675,10 @@ static void nfs_commit_release_pages(struct
nfs_commit_data *data)
next:
nfs_unlock_and_release_request(req);
}
+ nfss = NFS_SERVER(data->inode);
+ if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
+ clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
+
nfs_init_cinfo(&cinfo, data->inode, data->dreq);
if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
nfs_commit_clear_lock(NFS_I(data->inode));
signature.asc
Description: PGP signature
