On Sat, Oct 29, 2022 at 12:46 PM David Wysochanski <dwysocha@xxxxxxxxxx> wrote:
>
> On Fri, Oct 28, 2022 at 12:59 PM Trond Myklebust <trondmy@xxxxxxxxxx> wrote:
> >
> > On Fri, 2022-10-28 at 07:50 -0400, David Wysochanski wrote:
> > > On Thu, Oct 27, 2022 at 3:16 PM Trond Myklebust <trondmy@xxxxxxxxxx>
> > > wrote:
> > > >
> > > > On Mon, 2022-10-17 at 06:52 -0400, Dave Wysochanski wrote:
> > > > > Convert the NFS buffered read code paths to corresponding netfs
> > > > > APIs,
> > > > > but only when fscache is configured and enabled.
> > > > >
> > > > > The netfs API defines struct netfs_request_ops which must be
> > > > > filled
> > > > > in by the network filesystem. For NFS, we only need to define 5
> > > > > of
> > > > > the functions, the main one being the issue_read() function.
> > > > > The issue_read() function is called by the netfs layer when a
> > > > > read
> > > > > cannot be fulfilled locally, and must be sent to the server
> > > > > (either
> > > > > the cache is not active, or it is active but the data is not
> > > > > available).
> > > > > Once the read from the server is complete, netfs requires a call
> > > > > to
> > > > > netfs_subreq_terminated() which conveys either how many bytes
> > > > > were
> > > > > read
> > > > > successfully, or an error. Note that issue_read() is called with
> > > > > a
> > > > > structure, netfs_io_subrequest, which defines the IO requested,
> > > > > and
> > > > > contains a start and a length (both in bytes), and assumes the
> > > > > underlying
> > > > > netfs will return a either an error on the whole region, or the
> > > > > number
> > > > > of bytes successfully read.
> > > > >
> > > > > The NFS IO path is page based and the main APIs are the pgio APIs
> > > > > defined
> > > > > in pagelist.c. For the pgio APIs, there is no way for the caller
> > > > > to
> > > > > know how many RPCs will be sent and how the pages will be broken
> > > > > up
> > > > > into underlying RPCs, each of which will have their own
> > > > > completion
> > > > > and
> > > > > return code. In contrast, netfs is subrequest based, a single
> > > > > subrequest may contain multiple pages, and a single subrequest is
> > > > > initiated with issue_read() and terminated with
> > > > > netfs_subreq_terminated().
> > > > > Thus, to utilze the netfs APIs, NFS needs some way to accommodate
> > > > > the netfs API requirement on the single response to the whole
> > > > > subrequest, while also minimizing disruptive changes to the NFS
> > > > > pgio layer.
> > > > >
> > > > > The approach taken with this patch is to allocate a small
> > > > > structure
> > > > > for each nfs_netfs_issue_read() call, store the final error and
> > > > > number
> > > > > of bytes successfully transferred in the structure, and update
> > > > > these
> > > > > values
> > > > > as each RPC completes. The refcount on the structure is used as
> > > > > a
> > > > > marker
> > > > > for the last RPC completion, is incremented in
> > > > > nfs_netfs_read_initiate(),
> > > > > and decremented inside nfs_netfs_read_completion(), when a
> > > > > nfs_pgio_header
> > > > > contains a valid pointer to the data. On the final put (which
> > > > > signals
> > > > > the final outstanding RPC is complete) in
> > > > > nfs_netfs_read_completion(),
> > > > > call netfs_subreq_terminated() with either the final error value
> > > > > (if
> > > > > one or more READs complete with an error) or the number of bytes
> > > > > successfully transferred (if all RPCs complete successfully).
> > > > > Note
> > > > > that when all RPCs complete successfully, the number of bytes
> > > > > transferred
> > > > > is capped to the length of the subrequest. Capping the
> > > > > transferred
> > > > > length
> > > > > to the subrequest length prevents "Subreq overread" warnings from
> > > > > netfs.
> > > > > This is due to the "aligned_len" in nfs_pageio_add_page(), and
> > > > > the
> > > > > corner case where NFS requests a full page at the end of the
> > > > > file,
> > > > > even when i_size reflects only a partial page (NFS overread).
> > > > >
> > > > > Signed-off-by: Dave Wysochanski <dwysocha@xxxxxxxxxx>
> > > > > Reviewed-by: Jeff Layton <jlayton@xxxxxxxxxx>
> > > >
> > > >
> > > > This is not doing what I asked for, which was to separate out the
> > > > fscache functionality, so that we can call that if and when it is
> > > > available.
> > > >
> > > I must have misunderstood then.
> > >
> > > The last feedback I have from you was that you wanted it to be
> > > an opt-in feature, and it was a comment on a previous patch
> > > to Kconfig. I was proceeding the best I knew how, but
> > > let me try to get back on track.
> > >
> > > > Instead, it is just wrapping the NFS requests inside netfs
> > > > requests. As
> > > > it stands, that means it is just duplicating information, and
> > > > adding
> > > > unnecessary overhead to the standard I/O path (extra allocations,
> > > > extra
> > > > indirect calls, and extra bloat to the inode).
> > > >
> > > I think I understand what you're saying but I'm not sure. Let me
> > > ask some clarifying questions.
> > >
> > > Are you objecting to the code when CONFIG_NFS_FSCACHE is
> > > configured? Or when it is not? Or both? I think you're objecting
> > > when it's configured, but not enabled (we mount without 'fsc').
> > > Am I right?
> > >
> > > Also, are you objecting to the design that to use fcache we now
> > > have to use netfs, specifically:
> > > - call into netfs via either netfs_read_folio or netfs_readahead
> > > - if fscache is enabled, then the IO can be satisfied from fscache
> > > - if fscache is not enabled, or some of the IO cannot be satisfied
> > > from the cache, then NFS is called back via netfs_issue_read
> > > and we use the normal NFS read pageio interface. This requires
> > > we call netfs_subreq_terminated() when all the RPCs complete,
> > > which is the reason for the small changes to pagelist.c
> >
> > I'm objecting to any middle layer "solution" that adds overhead to the
> > NFS I/O paths.
> >
> Got it.
>
> > I'm willing to consider solutions that are specific only to the fscache
> > use case (i.e. when the 'fsc' mount option is specified). However when
> > I perform a normal NFS mount, and do I/O, then I don't want to see
> > extra memory allocations, extra indirect calls and larger inode
> > footprints.
> >
> > IOW: I want the code to optimise for the case of standard NFS, not for
> > the case of 'NFS with cachefs additions'.
> >
> I agree completely. Are you seeing extra memory allocations
> happen on mounts without 'fsc' or is it more a concern or how
> some of the patches look? We should not be calling any netfs or
> fscache code if 'fsc' is not on the mount and I don't see any in my
> testing. So either there's a misunderstanding here, or there's a
> bug I'm missing.
>
> If fscache is not configured, then nfs_netfs_read_folio() and
> nfs_netfs_readahead() is a wrapper that returns -ENOBUFS.
> If it's configured but not enabled, then the checks for
> netfs_inode(inode)->cache should skip over any netfs code.
> But maybe there's a non-obvious bug you're seeing and
> somehow netfs is still getting called? Because I cannot
> see netfs getting called if 'fsc' is not on the mount in my
> tests.
>
> int nfs_netfs_read_folio(struct file *file, struct folio *folio)
> {
> if (!netfs_inode(folio_inode(folio))->cache)
> return -ENOBUFS;
>
> return netfs_read_folio(file, folio);
> }
>
> int nfs_netfs_readahead(struct readahead_control *ractl)
> {
> struct inode *inode = ractl->mapping->host;
>
> if (!netfs_inode(inode)->cache)
> return -ENOBUFS;
>
> netfs_readahead(ractl);
> return 0;
> }
>
>
> > >
> > > Can you be more specific as to the portions of the patch you don't
> > > like
> > > so I can move it in the right direction?
> > >
> > > This is from patch #2 which you didn't comment on. I'm not sure
> > > you're
> > > ok with it though, since you mention "extra bloat to the inode".
> > > Do you object to this even though it's wrapped in an
> > > #ifdef CONFIG_NFS_FSCACHE? If so, do you require no
> > > extra size be added to nfs_inode?
> > >
> > > @@ -204,9 +208,11 @@ struct nfs_inode {
> > > __u64 write_io;
> > > __u64 read_io;
> > > #ifdef CONFIG_NFS_FSCACHE
> > > - struct fscache_cookie *fscache;
> > > -#endif
> > > + struct netfs_inode netfs; /* netfs context and VFS inode
> > > */
> > > +#else
> > > struct inode vfs_inode;
> > > +#endif
> > > +
> >
> > Ideally, I'd prefer no extra size. I can live with it up to a certain
> > point, however for now NFS is not unconditionally opting into the netfs
> > project. If we're to ever do that, then I want to see streamlined code
> > for the standard I/O case.
> >
> Ok and understood about standard I/O case.
>
> I was thinking how we might not increase the size, but I don't think
> I can make it work.
>
> I thought we could change to something like the below, without an
> embedded struct inode:
>
> @@ -204,9 +208,11 @@ struct nfs_inode {
> __u64 write_io;
> __u64 read_io;
> #ifdef CONFIG_NFS_FSCACHE
> - struct fscache_cookie *fscache;
> -#endif
> + struct netfs_inode *netfs; /* netfs context and VFS inode */
> +#else
> struct inode vfs_inode;
> +#endif
> +
>
> Then I would need to alloc/free a netfs_inode at the time of
> nfs_inode initiation. Unfortunately this has the issue that the NFS_I()
> macro cannot work, because it requires an embedded "struct inode"
> due to "container_of" use:
>
> +#ifdef CONFIG_NFS_FSCACHE
> +static inline struct inode *VFS_I(struct nfs_inode *nfsi)
> +{
> + return &nfsi->netfs.inode;
> +}
> +static inline struct nfs_inode *NFS_I(const struct inode *inode)
> +{
> + return container_of(inode, struct nfs_inode, netfs.inode);
> +}
> +#else
> +static inline struct inode *VFS_I(struct nfs_inode *nfsi)
> +{
> + return &nfsi->vfs_inode;
> +}
> static inline struct nfs_inode *NFS_I(const struct inode *inode)
> {
> return container_of(inode, struct nfs_inode, vfs_inode);
> }
> +#endif
>
>
Actually Trond maybe we can achieve a "0 length increase" of
nfs_inode if dhowells would take a patch to modify the definition
of struct netfs_inode and netfs_inode_init(), something like the WIP
patch below. What do you think?
I think maybe this could be a follow-on patch and if you/dhowells
think it's an ok idea I can try to work out what is needed across
the tree. I thought about it more and I kinda agree that in the
case for NFS where fscache is "configured but not enabled",
then even though we're only adding 24 bytes to the nfs_inode
each time, it will add up so it is worth at least a discussion.
diff --git a/include/linux/netfs.h b/include/linux/netfs.h
index f2402ddeafbf..195714f1c355 100644
--- a/include/linux/netfs.h
+++ b/include/linux/netfs.h
@@ -118,11 +118,7 @@ enum netfs_io_source {
typedef void (*netfs_io_terminated_t)(void *priv, ssize_t transferred_or_error,
bool was_async);
-/*
- * Per-inode context. This wraps the VFS inode.
- */
-struct netfs_inode {
- struct inode inode; /* The VFS inode */
+struct netfs_info {
const struct netfs_request_ops *ops;
#if IS_ENABLED(CONFIG_FSCACHE)
struct fscache_cookie *cache;
@@ -130,6 +126,14 @@ struct netfs_inode {
loff_t remote_i_size; /* Size of the remote file */
};
+/*
+ * Per-inode context. This wraps the VFS inode.
+ */
+struct netfs_inode {
+ struct inode inode; /* The VFS inode */
+ struct netfs_info *netfs; /* Rest of netfs data */
+};
+
/*
* Resources required to do operations on a cache.
*/
@@ -312,10 +316,12 @@ static inline struct netfs_inode
*netfs_inode(struct inode *inode)
static inline void netfs_inode_init(struct netfs_inode *ctx,
const struct netfs_request_ops *ops)
{
- ctx->ops = ops;
- ctx->remote_i_size = i_size_read(&ctx->inode);
+ ctx->netfs = kzalloc(sizeof(struct netfs_info)), GFP_KERNEL);
+ /* FIXME: Check for NULL */
+ ctx->netfs->ops = ops;
+ ctx->netfs->remote_i_size = i_size_read(&ctx->inode);
#if IS_ENABLED(CONFIG_FSCACHE)
- ctx->cache = NULL;
+ ctx->netfs->cache = NULL;
#endif
}
>
> > >
> > >
> > > Are you ok with the stub functions which are placed in fscache.h, and
> > > when CONFIG_NFS_FSCACHE is not set, become either a no-op
> > > or a 1-liner (nfs_netfs_readpage_release)?
> > >
> > > #else /* CONFIG_NFS_FSCACHE */
> > > +static inline void nfs_netfs_inode_init(struct nfs_inode *nfsi) {}
> > > +static inline void nfs_netfs_initiate_read(struct nfs_pgio_header
> > > *hdr) {}
> > > +static inline void nfs_netfs_read_completion(struct nfs_pgio_header
> > > *hdr) {}
> > > +static inline void nfs_netfs_readpage_release(struct nfs_page *req)
> > > +{
> > > + unlock_page(req->wb_page);
> > > +}
> > > static inline void nfs_fscache_release_super_cookie(struct
> > > super_block *sb) {}
> > > static inline void nfs_fscache_init_inode(struct inode *inode) {}
> > >
> > >
> > > Do you object to the below? If so, then do you want
> > > #ifdef CONFIG_NFS_FSCACHE here?
> > >
> > > -- a/fs/nfs/inode.c
> > > +++ b/fs/nfs/inode.c
> > > @@ -2249,6 +2249,8 @@ struct inode *nfs_alloc_inode(struct
> > > super_block *sb)
> > > #ifdef CONFIG_NFS_V4_2
> > > nfsi->xattr_cache = NULL;
> > > #endif
> > > + nfs_netfs_inode_init(nfsi);
> > > +
> > > return VFS_I(nfsi);
> > > }
> > > EXPORT_SYMBOL_GPL(nfs_alloc_i
> > > node);
> > >
> > >
> > > Do you object to the changes in fs/nfs/read.c? Specifically,
> > > how about the below calls to netfs from nfs_read_folio and
> > > nfs_readahead into equivalent netfs calls? So when
> > > NFS_CONFIG_FSCACHE is set, but fscache is not enabled
> > > ('fsc' not on mount), these netfs functions do immediately call
> > > netfs_alloc_request(). But I wonder if we could simply add a
> > > check to see if fscache is enabled on the mount, and skip
> > > over to satisfy what you want. Am I understanding what you
> > > want?
> >
> > Quite frankly, I'd prefer that we just split out the functionality that
> > is needed from the netfs code so that it can be optimised. However I'm
> > not interested enough in the cachefs functionality to work on that
> > myself. ...and as I indicated above, I might be OK with opting into the
> > netfs project, once the overhead can be made to disappear.
> >
> Understood.
>
> If you think it makes more sense, I can move some of the nfs_netfs_*
> functions into a netfs.c file as a starting point. Or that can maybe
> be done in a future patchset?
>
> For now I was equating netfs and fscache together so we can
> move on from the much older and single-page limiting fscache
> interface that is likely to go away soon.
>
> > >
> > > @@ -355,6 +343,10 @@ int nfs_read_folio(struct file *file, struct
> > > folio *folio)
> > > if (NFS_STALE(inode))
> > > goto out_unlock;
> > >
> > > + ret = nfs_netfs_read_folio(file, folio);
> > > + if (!ret)
> > > + goto out;
> > > +
> > >
> > > @@ -405,6 +399,10 @@ void nfs_readahead(struct readahead_control
> > > *ractl)
> > > if (NFS_STALE(inode))
> > > goto out;
> > >
> > > + ret = nfs_netfs_readahead(ractl);
> > > + if (!ret)
> > > + goto out;
> > > +
> > >
> The above wrappers should prevent any additional overhead when fscache
> is not enabled. As far as I know these work to avoid calling netfs
> when 'fsc' is not on the mount.
>
> > >
> > > And how about these calls from different points in the read
> > > path to the earlier mentioned stub functions?
> > >
> > > @@ -110,20 +110,13 @@ EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
> > >
> > > static void nfs_readpage_release(struct nfs_page *req, int error)
> > > {
> > > - struct inode *inode = d_inode(nfs_req_openctx(req)->dentry);
> > > struct page *page = req->wb_page;
> > >
> > > - dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb-
> > > >s_id,
> > > - (unsigned long long)NFS_FILEID(inode), req->wb_bytes,
> > > - (long long)req_offset(req));
> > > -
> > > if (nfs_error_is_fatal_on_server(error) && error != -
> > > ETIMEDOUT)
> > > SetPageError(page);
> > > - if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
> > > - if (PageUptodate(page))
> > > - nfs_fscache_write_page(inode, page);
> > > - unlock_page(page);
> > > - }
> > > + if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE))
> > > + nfs_netfs_readpage_release(req);
> > > +
> >
> > I'm not seeing the value of wrapping unlock_page(), no... That code is
> > going to need to change when we move it to use folios natively anyway.
> >
> Ok, how about I make it conditional on whether fscache is configured
> and enabled then, similar to the nfs_netfs_read_folio() and
> nfs_netfs_readahead()? Below is what that would look like.
> I could inline the code in nfs_netfs_readpage_release() if you
> think it would be clearer.
>
> static void nfs_readpage_release(struct nfs_page *req, int error)
> {
> struct page *page = req->wb_page;
>
> if (nfs_error_is_fatal_on_server(error) && error != -ETIMEDOUT)
> SetPageError(page);
> if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE))
> #ifndef CONFIG_NFS_FSCACHE
> unlock_page(req->wb_page);
> #else
> nfs_netfs_readpage_release(req);
> #endif
> nfs_release_request(req);
> }
>
>
> void nfs_netfs_readpage_release(struct nfs_page *req)
> {
> struct inode *inode = d_inode(nfs_req_openctx(req)->dentry);
>
> /*
> * If fscache is enabled, netfs will unlock pages.
> */
> if (netfs_inode(inode)->cache)
> return;
>
> unlock_page(req->wb_page);
> }
>
>
> > > nfs_release_request(req);
> > > }
> > >
> > > @@ -177,6 +170,8 @@ static void nfs_read_completion(struct
> > > nfs_pgio_header *hdr)
> > > nfs_list_remove_request(req);
> > > nfs_readpage_release(req, error);
> > > }
> > > + nfs_netfs_read_completion(hdr);
> > > +
> > > out:
> > > hdr->release(hdr);
> > > }
> > > @@ -187,6 +182,7 @@ static void nfs_initiate_read(struct
> > > nfs_pgio_header *hdr,
> > > struct rpc_task_setup *task_setup_data,
> > > int how)
> > > {
> > > rpc_ops->read_setup(hdr, msg);
> > > + nfs_netfs_initiate_read(hdr);
> > > trace_nfs_initiate_read(hdr);
> > > }
> > >
> > >
> > > Are you ok with these additions? Something like this would
> > > be required in the case of fscache configured and enabled,
> > > because we could have some of the data in a read in
> > > fscache, and some not. That is the reason for the netfs
> > > design, and why we need to be able to call the normal
> > > NFS read IO path (netfs calls into issue_read, and we call
> > > back via netfs_subreq_terminated)?
> > >
> > > @@ -101,6 +101,9 @@ struct nfs_pageio_descriptor {
> > > struct pnfs_layout_segment *pg_lseg;
> > > struct nfs_io_completion *pg_io_completion;
> > > struct nfs_direct_req *pg_dreq;
> > > +#ifdef CONFIG_NFS_FSCACHE
> > > + void *pg_netfs;
> > > +#endif
> > >
> > > @@ -1619,6 +1619,9 @@ struct nfs_pgio_header {
> > > const struct nfs_rw_ops *rw_ops;
> > > struct nfs_io_completion *io_completion;
> > > struct nfs_direct_req *dreq;
> > > +#ifdef CONFIG_NFS_FSCACHE
> > > + void *netfs;
> > > +#endif
> > >
> > >
> > > And these additions to pagelist.c?
> > >
> > > @@ -68,6 +69,10 @@ void nfs_pgheader_init(struct
> > > nfs_pageio_descriptor *desc,
> > > hdr->good_bytes = mirror->pg_count;
> > > hdr->io_completion = desc->pg_io_completion;
> > > hdr->dreq = desc->pg_dreq;
> > > +#ifdef CONFIG_NFS_FSCACHE
> > > + if (desc->pg_netfs)
> > > + hdr->netfs = desc->pg_netfs;
> > > +#endif
> >
> > Why the conditional?
> >
> Not really needed and I was thinking of removing it, so I'll do that.
>
> > >
> > >
> > > @@ -846,6 +851,9 @@ void nfs_pageio_init(struct nfs_pageio_descriptor
> > > *desc,
> > > desc->pg_lseg = NULL;
> > > desc->pg_io_completion = NULL;
> > > desc->pg_dreq = NULL;
> > > +#ifdef CONFIG_NFS_FSCACHE
> > > + desc->pg_netfs = NULL;
> > > +#endif
> > >
> > >
> > > @@ -1360,6 +1369,9 @@ int nfs_pageio_resend(struct
> > > nfs_pageio_descriptor *desc,
> > >
> > > desc->pg_io_completion = hdr->io_completion;
> > > desc->pg_dreq = hdr->dreq;
> > > +#ifdef CONFIG_NFS_FSCACHE
> > > + desc->pg_netfs = hdr->netfs;
> > > +#endif
> >
> > Those all need wrapper functions instead of embedding #ifdefs.
> >
> Ok.
>
>
>
> > >
> > >
> > > > My expectation is that the standard I/O path should have minimal
> > > > overhead, and should certainly not increase the overhead that we
> > > > already have. Will this be addressed in future iterations of these
> > > > patches?
> > > >
> > >
> > > I will do what I can to satisfy what you want, either by fixing up
> > > this patch or follow-on patches. Hopefully the above questions
> > > will clarify the next steps.
> > >
> >
> > --
> > Trond Myklebust
> > Linux NFS client maintainer, Hammerspace
> > trond.myklebust@xxxxxxxxxxxxxxx
> >
> >
