Reads and writes are pretty similar in some ways probably actually:

there's a proc_read_setup(struct nfs_read_data *).  When the operation
completes, an nfs{?|3|4}_read_done is called, which calls....

nfs_readpage_result(struct rpc_task *): loops over singly-linked list of
nfs_pages data->pages, does some magic to tell the mm what we've found out
about these pages (memclear_highpage_flush, SetPageUptodate, unlock_page,
etc.), does magic to release this nfs_page.

Now on 2.6.7-CITI_NFS4_ALL-8, tracing starting from nfs_readpages.

nfs_readpages first gets an nfs_open_context, which is where we get the rpc
credential and nfs4 state from.

Then it does read_cache_pages():
	Add to radix tree of pages associated with the vma.
	call the passed-in callback (readpage_async_filler in this case)
	adds the page to the "inactive list" (what's that?)
		(subtlety: actually it batches the addition of pages
		to the list using __pagevec_lru_add())
	In case of an error return from the callback, just releases the rest
		of the pages?

readpage_async_filler():
	nfs_wb_page(): calls nfs_sync_inode, which in our case just does
		nfs_flush_inode() and nfs_commit_inode() until the latter
		returns 0.
		nfs_flush_inode():
			nfs_scan_dirty():