tree feb64bf33b46816cd39f71edb8cacdd1a987cc6b
parent 6b77a338da8cc00ebcf6f7c7cf92f67362d3e8e5
author David Howells <dhowells@redhat.com> 1775773636 +0100
committer David Howells <dhowells@redhat.com> 1776120707 +0100

netfs: Combine prepare and issue ops and grab the buffers on request

Modify the way subrequests are generated in netfslib to try and simplify
the code.  The issue, primarily, is in writeback: the code has to create
multiple streams of write requests to disparate targets with different
properties (e.g. server and fscache), where not every folio needs to go to
every target (e.g. data just read from the server may only need writing to
the cache).

The current model in writeback, at least, is to go carefully through every
folio, preparing a subrequest for each stream when it was detected that
part of the current folio needed to go to that stream, and repeating this
within and across contiguous folios; then to issue subrequests as they
become full or hit boundaries after first setting up the buffer.  However,
this is quite difficult to follow - and makes it tricky to handle
discontiguous folios in a request.

This is changed such that netfs now accumulates buffers and attaches them
to each stream when they become valid for that stream, then flushes the
stream when a limit or a boundary is hit.  The issuing code in netfs then
loops around creating and issuing subrequests without calling a separate
prepare stage (though a function is provided to get an estimate of when
flushing should occur).  The filesystem (or cache) then gets to take a
slice of the master bvec chain as its I/O buffer for each subrequest,
including discontiguities if it can support a sparse/vectored RPC (as Ceph
can).

Similar-ish changes also apply to buffered read and unbuffered read and
write, though in each of those cases there is only a single contiguous
stream.  Though for buffered read this consists of interwoven requests from
multiple sources (server or cache).

To this end, netfslib is changed in the following ways:

 (1) ->prepare_xxx(), buffer selection and ->issue_xxx() are now collapsed
     together such that one ->issue_xxx() call is made with the subrequest
     defined to the maximum extent; the filesystem/cache then reduces the
     length of the subrequest and calls back to netfslib to grab a slice of
     the buffer, which may reduce the subrequest further if a maximum
     segment limit is set.  The filesystem/cache then dispatches the
     operation.

 (2) Retry buffer tracking is added to the netfs_io_request struct.  This
     is then selected by the subrequest retry counter being non-zero.

 (3) The use of iov_iter is pushed down to the filesystem.  Netfslib now
     provides the filesystem with a bvecq holding the buffer rather than an
     iov_iter.  The bvecq can be duplicated and headers/trailers attached
     to hold protocol and several bvecqs can be linked together to create a
     compound operation.

 (4) If the ->issue_xxx() functions terminate with -ENOMEM, a flag is set
     on the request to abort further subrequest generation/retrying.

 (5) During writeback, netfslib now builds up an accumulation of buffered
     data before issuing writes on each stream (one server, one cache).  It
     asks each stream for an estimate of how much data to accumulate before
     it next generates subrequests on the stream.  The filesystem or cache
     is not required to use up all the data accumulated on a stream at that
     time unless the end of the pagecache is hit.

 (6) During read-gaps, in which there are two gaps on either end of a dirty
     streaming write page that need to be filled, a buffer is constructed
     consisting of the two ends plus a sink page repeated to cover the
     middle portion.  This is passed to the server as a single write.  For
     something like Ceph, this should probably be done either as a
     vectored/sparse read or as two separate reads (if different Ceph
     objects are involved).

 (7) During unbuffered/DIO read/write, there is a single contiguous file
     region to be read or written as a single stream.  The dispatching
     function just creates subrequests and calls ->issue_xxx() repeatedly
     to eat through the bufferage.

 (8) At the start of buffered read, the entire set of folios allocated by
     VM readahead is loaded into a bvecq chain, rather than trying to do it
     piecemeal as-needed.  As the pages were already added and locked by
     the VM, this is slightly more efficient than loading piecemeal as only
     a single iteration of the xarray is required.

 (9) During buffered read, there is a single contiguous file region, to
     read as a single stream - however, this stream may be stitched
     together from subrequests to multiple sources.  Which sources are used
     where is now determined by querying the cache to find the next couple
     of extents in which it has data; netfslib uses this to direct the
     subrequests towards the appropriate sources.

     Each subrequest is given the maximum length in the current extent and
     then ->issue_read() is called.  The filesystem then limits the size
     and slices off a piece of the buffer for that extent.

(10) Cachefiles now provides an estimation function that indicates the
     standard maxima for doing DIO (MAX_RW_COUNT and BIO_MAX_VECS).

Note that sparse cachefiles still rely on the backing filesystem for
content mapping.  That will need to be addressed in a future patch and is
not trivial to fix.

Signed-off-by: David Howells <dhowells@redhat.com>
cc: Paulo Alcantara <pc@manguebit.org>
cc: Matthew Wilcox <willy@infradead.org>
cc: Christoph Hellwig <hch@infradead.org>
cc: netfs@lists.linux.dev
cc: linux-fsdevel@vger.kernel.org