The commit handling code is not safe against memory-pressure deadlocks
when writing to swap. In particular, nfs_commitdata_alloc() blocks
indefinitely waiting for memory, and this can consume all available
workqueue threads.
swap-out most likely uses STABLE writes anyway as COND_STABLE indicates
that a stable write should be used if the write fits in a single
request, and it normally does. However if we ever swap with a small
wsize, or gather unusually large numbers of pages for a single write,
this might change.
For safety, make it explicit in the code that direct writes used for swap
must always use FLUSH_COND_STABLE.
Signed-off-by: NeilBrown <[email protected]>
---
fs/nfs/direct.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
index 43a956d7fd62..29c007b2a17a 100644
--- a/fs/nfs/direct.c
+++ b/fs/nfs/direct.c
@@ -791,7 +791,7 @@ static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
*/
static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
struct iov_iter *iter,
- loff_t pos)
+ loff_t pos, int ioflags)
{
struct nfs_pageio_descriptor desc;
struct inode *inode = dreq->inode;
@@ -799,7 +799,7 @@ static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
size_t requested_bytes = 0;
size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
- nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
+ nfs_pageio_init_write(&desc, inode, ioflags, false,
&nfs_direct_write_completion_ops);
desc.pg_dreq = dreq;
get_dreq(dreq);
@@ -905,6 +905,7 @@ ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
struct nfs_direct_req *dreq;
struct nfs_lock_context *l_ctx;
loff_t pos, end;
+ int ioflags = swap ? FLUSH_COND_STABLE : FLUSH_STABLE;
dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
file, iov_iter_count(iter), (long long) iocb->ki_pos);
@@ -947,7 +948,7 @@ ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
if (!swap)
nfs_start_io_direct(inode);
- requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
+ requested = nfs_direct_write_schedule_iovec(dreq, iter, pos, ioflags);
if (mapping->nrpages) {
invalidate_inode_pages2_range(mapping,
On Wed, 26 Jan 2022, Trond Myklebust wrote:
> On Mon, 2022-01-24 at 14:48 +1100, NeilBrown wrote:
> > The commit handling code is not safe against memory-pressure
> > deadlocks
> > when writing to swap. In particular, nfs_commitdata_alloc() blocks
> > indefinitely waiting for memory, and this can consume all available
> > workqueue threads.
> >
> > swap-out most likely uses STABLE writes anyway as COND_STABLE
> > indicates
> > that a stable write should be used if the write fits in a single
> > request, and it normally does. However if we ever swap with a small
> > wsize, or gather unusually large numbers of pages for a single write,
> > this might change.
> >
> > For safety, make it explicit in the code that direct writes used for
> > swap
> > must always use FLUSH_COND_STABLE.
>
> OK. Your explanation above has me extremely confused.
>
> If you want to avoid commit, then you should be using FLUSH_STABLE,
> since that forces the writes to be synchronous. FLUSH_COND_STABLE can
> and will use unstable writes if it sees that there are more writes to
> come.
>
> >
> > Signed-off-by: NeilBrown <[email protected]>
> > ---
> > fs/nfs/direct.c | 7 ++++---
> > 1 file changed, 4 insertions(+), 3 deletions(-)
> >
> > diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
> > index 43a956d7fd62..29c007b2a17a 100644
> > --- a/fs/nfs/direct.c
> > +++ b/fs/nfs/direct.c
> > @@ -791,7 +791,7 @@ static const struct nfs_pgio_completion_ops
> > nfs_direct_write_completion_ops = {
> > */
> > static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req
> > *dreq,
> > struct iov_iter *iter,
> > - loff_t pos)
> > + loff_t pos, int
> > ioflags)
> > {
> > struct nfs_pageio_descriptor desc;
> > struct inode *inode = dreq->inode;
> > @@ -799,7 +799,7 @@ static ssize_t
> > nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
> > size_t requested_bytes = 0;
> > size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize,
> > PAGE_SIZE);
> >
> > - nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
> > + nfs_pageio_init_write(&desc, inode, ioflags, false,
> > &nfs_direct_write_completion_ops);
> > desc.pg_dreq = dreq;
> > get_dreq(dreq);
> > @@ -905,6 +905,7 @@ ssize_t nfs_file_direct_write(struct kiocb *iocb,
> > struct iov_iter *iter,
> > struct nfs_direct_req *dreq;
> > struct nfs_lock_context *l_ctx;
> > loff_t pos, end;
> > + int ioflags = swap ? FLUSH_COND_STABLE : FLUSH_STABLE;
>
> This is an unacceptable change in behaviour for the non-swap case, so
> NACK.
>
Hi Trond,
thanks for the review.
You are right - I had that test exactly backwards. I've fixed for the
next version.
Thanks,
NeilBrown