Documentation/vfs.txt

Unorganized Notes on the Virtual File System
-------------------------------------
Our VFS is built similarly to Linux's, mostly because I'd like to have somewhat
easy integration with ext2 (or at the very least have our own FS that can
integrate into Linux easily).

There are four main objects in a filesystem: superblock, inode, dentry, and a
file.  
	- Superblock: Specific instance of a mounted filesystem.  All
	  synchronization is done with the one spinlock.
	- Inode: represents a specific file
	- Dentry: in memory object, corresponding to an element of a path.
	  E.g. /, usr, bin, and vim are all dentries.  All have inodes.  Vim
	  happens to be a
	  file instead of a directory.
	- File: represents a file opened by a process.

So far, dentries are the most complicated, so here are some notes about them.
These are probably self-consistent, but may be a bit old (or just wrong) because
I wrote them as I put the VFS together:

A dentry is just a connection between a part of a path and an inode.  We just
want to walk from dentry to dentry, asking each to find the next part in the
path.  When you hit a dentry that's not the end of the desired path, you ask its
inode for the next dentry by using it's operation (lookup).

lookup() takes the inode of the directory (or 0 for the root) and a dentry
(already allocated, with the d_name and i_name filled in), and it will find the
correct inode for the given dentry, as well as filling out any specific FS
things in the dentry (like d_ops).  It will return 0 on failure, or the dentry
pointer passed in on success.  Somehow the nameidata bit will be used too.  This
will probably change a bit...  Note that lookup() needs to read the actual
directory file and also lookup the inode off the disc, which means it will
block.

When the next dentry is a mountpoint, (e.g. /mnt/cdrom), when you ask mnt for
cdrom, lookup() will parse it's file (a directory) and see 'cdrom' there as a child
entry.  Then it will return cdrom's dentry, like always.

But since cdrom was a mountpoint, (which you can see from the d_mount_point),
you need to walk through the structures a bit to get the dentry of the FS rooted
at that mountpoint to keep walking.  The VFS can handle that, so lookup()
doesn't need to worry about it.

Why are there two dentries associated with a vfsmount?  Normally, a dentry for a
directory points to an inode that contains its members.  When a FS is mounted
over it, we don't want to destroy that.  Instead, we mark it as a mountpoint,
and when following, we walk into the next FS.  The mountpoint is the dentry in
the parent mount (using the parent mount's FS superblock), and the root is the
dentry of the mounted FS, using the FS superblock of the mounted FS.

Sometimes lookup() will not be called at all.  We want to look for present
dentries first (ones that exist in memory already, including all of those with a
refcnt).  So we can look in the cache (currently just an SLIST, but will
eventually be a hash table).  When hitting in the cache (hashed by dentry name,
and probably the parent dir), we need to verify that the dentry's parent is our
own (handled by the VFS code).  All vfsmount dentries will be in the cache,
since they all have a refcnt (the vfsmount struct points to them).

Dentries for / and pwd and whatnot have a refcnt (pointed to by fs_struct,
vfsmounts, etc).  Anything with a pointer to it has a refcnt, so it never goes
away.  So if we have an inode in memory, it's entire dentry path is cached,
(just follow the parent pointers back).  Note that when a dentry's refcnt hits
0, we do *not* deallocate (like we do with many other structures).  It will just
be on the LRU list in the dcache system.  Likewise, every dentry points to its
inode, which pins that inode in memory.

Other refcnts: just about everything has a refcnt, and we need to be careful
about when we want to use them and dealloc.  Some things would be a pain in the
ass, like with the super_block refcnt.  Every dentry has a ref to the SB, but
changing that ref every time we add or remove a dentry will probably be an
unnecessary penalty if we can ensure all dentries of that FS are gone before
removing the superblock through another mechanism.  We'll see.  Mostly, this
just means we need to think about what we really want from a refcnt, and whether
or not we want the kref / process style refcnting.

Mounting:
-------------------------------------------
When you mount, you need to read in the super block and connect the relevant
data structures together.  The SB is connected to the vfsmount, which is
connected to the dentry of the mount point and the dentry of the root of the FS.
This means when mounting the FS, we need to create the dentry for "/", which
means we also need the inode, which needs to be read_inode()'d in.  Actually, we
might not need to read the inode in right away - we might be able to get away
with reading them in on-demand.

All of this means that for every mount point, the SB, vfsmount, dentry, and
inode are in memory.  Due to the way dentries link, every dentry and inode back
up to the real root are all in memory too.  Having a mount point is like having
a process working in that directory - the chain back up is pinned.

d_subdirs:
-------------------------------------------
Tracking the links between objects can be tricky.  One pain is d_subdirs. Linux
only tracks subdirectories.  We also do this.  I think the reason they do it is
since you should be using the dcache to cache lookups, and not scan the linked
list of children of a dentry for a specific file.  Though it is handy to know
all of your *directory* children.  In KFS, we also track all children in a list.
This is to make our lookups work - instead of having an actual directory file
with name->ino mappings.

KFS and metadata pinning:
-------------------------------------------
KFS pins all metadata ("all paths pinned").  This means that from the root mnt
down to the lowest inode, all dentries and corresponding inodes are pinned in
memory from creation time.   Yeah, this means we have chunks of metadata for
files we aren't using sitting around in RAM.  We also have the *files* sitting
around in RAM too.  Not that concerned, for now.  Plus, I don't want to reparse
the CPIO backing store to figure out inode fields, directory names, etc.

Page Cache:
-------------------------------------------
Every object that has pages, like an inode or the swap (or even direct block
devices) has a page_map tracking which of its pages are currently in memory.
This is called a struct address_space in linux, which is a confusing name.  We
don't have all of the same fields yet, and may be doing things slightly
differently, but the basics are the same.  Together, the page_maps and the
functions to manipulate them make up the Page Cache.  Every page frame that is
in a page mapping can be traced back to its page_map, via pointers in the struct
page.  Note the page_mapping is tracked twice for a file, the f_mapping and the
i_mapping.  We really only need the i_mapping, but this saves a couple cache
misses.  Might go away later.

As a side note, Linux's "address_space" has a great example of the power of
their linked lists.  Their struct has a private_list head.  Due to their list
implementation, they are able to have a generic list head for a list of any
type (it's a struct list_head), and don't need to declare in a generic object
(like the page_map) a specific list type (that would get used by specific
FS's).

Just because a page is in a page_map, it doesn't mean it actually has the
data from the disc in it.  It just means that there is a physical frame
dedicated/mapped to be a page_map's holder for a specific page of an object
(usually a file on disc).  readpage() is called to fill that page in with what
it is supposed to have from its backing store.

This interpretation makes the meaning of "address space" make more sense.  It's
the "address space" of the device, mapping from (file,index) -> page_frame.
Still, calling it an address space just confuses things with the virtual memory
address space.

One of the reasons pages can be in the map without up-to-date data is due to
concurrency issues and outstanding I/O requests.  When the page is first being
filled up, the mapping exists but the data hasn't been brought in yet.  Other
processes can be trying to access the same block/page/byte, and they need to
block but to not try and schedule the operation.  

So here's how a typical page fault (__handle_page_fault(), either on demand or
populated) works on an mmap'd file at a high level.
1. PF is on a virt address -> translated by the vm_region to a file/offset (page).
1b. Weird permission issues?  See below!
2. Look it up in the page_map.
3. If the page is already there, and up-to-date, then great.  Skip to 6.  If
there is one, but it's not up to date, skip to 5.
4. If there is no page, get a free page, tie it (bidirectionally) to the inode
in the page_map.
5. Now there is a page, but it is not up to date, so call readpage().  This will
usually block.
6. Map that page (which has the current contents) into the address space of the
calling process (with the appropriate permissions, RO (MAP_PRIVATE, CoW), or
RW (MAP_SHARED).
Below: Now if in step 1 you had a permissions issue, such as a RW fault on a CoW
MAP_PRIVATE, you'll have to notice the type of error and the type of memory
region, then go through a separate path: get a new page, copy the contents, and
change the mapping.  Note, the page backing that mapping is not backed by the
file - it's just sitting around in the virtual memory of the process.

Also, if we want to use the PG_DIRTY flag, we'll need mark the regions as RO
until we write fault, at which point we dirty the page and change it to RW.

We could have multiple threads trying to fill a page in the page cache at once.
This is handled in file_load_page().  All threads check the page cache.  If two
threads try to add it to the page cache, only one will succeed, and the page
will be locked (PG_LOCKED).  The one who succeeds will readpage().  The one that
didn't will be like any other thread that is checking the page cache - it will
see a page is there, and will check it the page is up to date.  If it isn't, it
will try to lock the page so it can do the IO, with a few extra checks in case
the page had been removed or was filled in while it slept.

A big aspect of this is the use of lock_page() to block.  If the page is locked,
you block until it is unlocked.  (implementation and other issues still
pending).  Even the caller of readpage will lock after submitting the IO
request.  This will cause the caller to sleep until the IO is done.  When the IO
is done, that interrupt handler/thread will mark the page as up-to-date, and
unlock the page, which will wake up any of the waiters.  The caller of
readpage() may not be the first to wake up either.  This all means the IO system
needs to know about marking pages as up-to-date and unlocking them.  This
currently (Jul10) is just sitting in KFS, but can be done later either directly
or with a callback made by
whoever starts the IO.

A note on refcnting.  When a page is added to the page cache, that's a stored
reference.  When you lookup a page in the page cache, you get a refcnt'd
reference back.  When you pull a page from the page cache, you also get a
refcnt'd reference back - specifically it is the ref that was in the page map.

Files with Holes
--------------------------
If a file has a hole, we'll still try to look up the page in the page cache.
When that doesn't happen, we'll create and add a page, then call readpage().
Readpage will realize there is no page on disk/backing store for that part of
the file (since it was a hole) and just memset the page to 0.  In the future, we
can consider getting a CoW 0-page, but that's a bit premature and a bookkeeping
pain.

This also applies to trying to write to a block beyond the EOF.  If the request
hits the page cache and readpage(), it's because it was already checked and
cleared in another part of the VFS, such as in generic_file_write().

Files That End Before a Page Boundary
--------------------------
So what if we have a file that is only 1024 bytes.  When we read it in, we'll
have a whole page added to the page cache, with the extra 3K 0'd.  When we go to
write it back, it will write back the 1K, and ignore the rest.  But what if we
extend the file later?  That page is already in the page cache, but the buffer
heads aren't tracking the new 3K.  When that page gets evicted, only the 1K will
write back.

There are two ways to deal with this: 1) When we extend the file, check and see
if it is in the middle of a page, and if so, alloc the blocks (all FS
specific!), and adjust the BHs to map the new data.  2) Preallocate the
remaining blocks and stitch up the BH mapping at readpage (which is in the
specific FS).  This way, we reserve the blocks (but don't have to actually read
them in - we just mark the buffer as dirty).  When we grow a file, we don't need
to worry about any of these details.  We just make sure the page map has the
page, and not whether or not it's a half-page that needs a FS-specific solution.
I'm going with #2 for now.  Note that while the blocks are allocated, the file's
size is still 1024B.

Kref, Dentries, Inodes, and Files (or "I don't see why it's like X, but..."
--------------------------
There are multiple dentries pointing to an inode.  The dentries are (or will be)
cached too, but that is irrelevant.  The dentries pin the inodes in memory.
However, files pin inodes in memory (or they did) briefly.  After running around
in circles a bit, I asked, why doesn't the file pin the dentry instead of the
inode?  The answer: it is supposed to.  Linux does that, and I didn't because
pinning the inode made more sense at the time.

The heart of the issue is about understanding what files are and how they
relate to the rest of the VFS.  A 'file' in the OS is a structure to track an
open FS-disk-file, which is managed by the inode.  Given this, it makes sense
that a dentry (which is a name on a path) would be pinned by the corresponding
inode, and the file would pin the inode.  It doesn't, but it is believable.  In
reality, there are many names (dentries) for a given disk file, and the OS file
that you open corresponds to *one* of those names, and thus a dentry, and not to
the inode/specific file.  You need to go through the dentry to pin the inode.

In short, it's not: file -> inode -> dentry -> parent_dentry -> ...
It's file -> dentry -> parent_dentry ->
             |-> inode      |-> parent_inode
Another dentry and file (both OS and disk) can point to the same inode.  If you
don't do it this way, you can't pin up past the multi-dentry-point in the inode,
and your system doesn't really make sense.

So here is how it works: files pin dentries.  Dentries can pin other dentries,
on up the directory hierarchy.  Independently of the files, dentries pin their
inode.  There are many dentries per inode (or can be).  Since each dentry
doesn't know if it is the last dentry to decref the inode, we use a kref on
i_kref.  The inodes are storing references to the dentries, but they are the
kref "internal" / weak references.  Even if we did decref them, we don't trigger
anything with it.

The moral of the story is that if you don't fully understand something, you are
not in as good of a position to recommend changes or criticize as if you did
your homework.  Not that you can't, just that you should 'do your homework.'

Musings on path_lookup()
--------------------------
Things can get tricky with path lookup, especially with ., .., and symlinks.
When doing a LOOKUP_PARENT on a . or .., we give the parent of whatever the path
would have resolved too.  So /dir1/dir2/dir3/.'s parent is dir2.
/dir1/dir2/dir3/..'s parent is dir1.  I don't think Linux does this (note the
parent lookup is for internal kernel stuff, like when you want to edit
metadata).  When you try to make a . or .. file, you should get some sort of
error anyways.  We'll see how this works out.

Symlinks can be a bit tricky.  We handle ours a bit differently too, especially
regarding PARENT lookups.  Ultimately, you can do the same things in ROS that
you can do in Linux - if you try to create a file that is a dangling symlink,
you'll correctly create the destination file.  We handle this in
link_path_walk().  It will return the PARENT of whatever you would resolve to -
instead of trying to handle this in do_file_open() (which I think linux does).

Also, our handling of symlinks differs a bit from linux.  Eventually, it has
become clear we're going to need to manually port ext2, and we do some things
differently in our core VFS already.  Might as well do more thing differently -
like getting rid of follow_link and put_link from the FS specific sections.  Our
FSs just need to know how to return a char* for a symname - and not do any of
the actual link following.  Or any of the other stuff they do.  We'll see if
that turns out to be an issue or not...

Unlinking and other Link Stuff
-------------------------
Unlinking is just disconnecting a dentry-inode pair from the directory tree, and
decreasing the inode's i_nlink.  Nothing else happens yet, since we need to keep
the FS-file (controlled by the dentry/inode) so long as any OS-files have it
open.  They have it opened via open or mmap - any way that there is a reference
to a file, which then pins the dentry and inode.  When the OS-files close,
eventually the dentry's refcnt hits 0.  When it does, it normally would be up
for caching, but we can check nlinks and just drop it.  When that happens, it
releases the inode, which will see its nlinks is 0.  That will trigger the
underlying FS to clear out the FS-file.

For directories, you can only have one hardlink to a directory - meaning you are
only in the directory tree in one place.  However, all of your children can get
to you by going '../'.  We'll count these as hardlinks too.  This means that
every child increments its parent-dir's nlink.  This is the on-disk links, not
to be confused with the dentry->d_parent and kref() business that goes on for
the in-memory objects.  A directory cannot be removed if nlinks > 1.  If it is
1, then you can rmdir it, which will set its nlinks to 0.  Then its inode's
storage space will get freed when it is deleted, like any other inode.  In
theory.

Inodes: drop? delete? dealloc?
--------------------------
Inodes exist both in memory and on disk, but in different manners.  When a file
(F_WHATEVER, could be DIR) exists in an FS, it'll have an inode on disk.  When
it is time to delete that file, we call _delete_inode().  When we want to free
the memory associated with an in-memory (VFS) inode, we call _dealloc_inode().

What about drop_inode?  For now, we don't use it.  We have inode_release() in
the VFS.  If we need an FS specific one (like for ext2), or have FS-specific
work that needs to be done in inode_release(), we'll use it later.

Either way, inode_release() is called when we no longer use the in-memory inode.
If there are no hard links, it will delete the inode.  Either way, it will just
free the in-memory inode (after deleting the disc version).

Example: I want to unlink (rm) a file.  There are two cases: the file is already
open (with a dentry and the inode in memory) or the file is not.  Both cases are
handled the same way!  In either case, we eventually call do_lookup on the item
in question, getting both a dentry and its inode read in.  (We read it in for a
couple reasons: convenient to check the type, and we need to manipulate the
nlink).  If a process has the file open, or even if it is sitting in the cache,
we will get the same inode (from the inode cache, might not be implemented yet).
When we decref the dentry and it is done, it will decref the inode.  This
dentry's final decref will be deferred until any open files are closed.  Note,
this requires a working dentry/inode-cache - otherwise we'll have multiple
copies of the same FS/disk-inode (and possibly dentry).  Anyway, when this is
done, the release function will delete the inode, then dealloc it.

Another example:  We simply close a file.  When that happens, we decref the
dentry, which decrefs the inode.  It may remain cached for a bit - not a big
deal.  When it is finally removed, nlinks is positive, so the inode's in memory
copy is written back (if it was dirty) and the structure is deallocated.

Side notes: dentry cached inodes should be removed after their lookup in unlink.
Also, since multiple dentries point to the same inode, it's not enough to just
cache dentries - we need to be able to find inodes too so that we get the one
inode regardless of which dentry we use (which may be uncached).

Dentry and Inode Caches
--------------------------
The dentry caches dentry lookups - we need the parent and a hash of the name to
do a lookup.  The dcache consists of a hash table for the lookups, as well as an
extra list of entries that are unused (their kref is 0).  The dcache also caches
negative entries - entries that were wrong.  This still speeds up future
requests.  Most uses of the system just need to use dcache_get and dcache put.
Not all of this is implemented yet.

The inode cache is similar, though we can't just have the inodes hang off the
dentry cache (more than one dentry points to the same inode).  We don't need to
worry about unused lists or anything like that - once the kref hits 0, we're
done and we can rip it out of the hash.

Both hashes hang off the superblock, with concurrent access protected by locks
in the SB.

The dentry cache is the weirdest of them all - for normal entries, its key and
value are the same thing.  The actual hashing of a dentry is done by the qstr
value, and to determine equality, we need to compare parents (compared to the
inode cache, where the only thing that matters is the i_ino).  Put another way,
we need elements of the whole dentry to get a unique key (d_parent and d_name).

As stated above, the dcache also caches negative entries.  This is to prevent a
lookup on disk.  These negative entries are in the dcache and on the LRU list
(their refcnt is 0, the are not USED).  When we dcache_get, we don't bother with
returning the actual dentry (after increffing) and then decref it again.
Instead, we just return the negative result (via the query dentry,
incidentally).

Freeing of dentries happens in one of two ways: call __dentry_free() directly,
which is appropriate when you have the only copy (like in do_lookup()), or it
will get freed when the dcache gets reaped (the LRU entries are freed).  When it
is decref'd, it simply goes into a state where it is ready to be reaped, but
kept around for future lookups - most usages throughout the vfs can just decref
when they are done using it.

One complication is cached negative dentries.  These are only referenced once
(in the dcache), so they can get __dentry_free()d directly.  This gets tricky
with rmdir and unlink.  Initially, those functions marked the dentry as negative
and unused, and would let them stay in the dcache (returning negative results on
future lookups).  The problem with this is that now the dcache could have a
negative dentry that was a real, formerly used dentry - one with a refcnt that
needs to be decref'd and released.  

There are two solutions: one is to change the dcache to not assume that negative
entries are unreferenced (which also means on the LRU).  The other is to just
remove the dentry from the dcache on rmdir/unlink.  It won't be negative - and
that won't matter, since it is un-lookup-able.  And it will die nicely when it
gets decref'd.  All we'll do is add a DENTRY_DYING flag, and dentry_release()
will avoid LRU and unusing it.  The dcache can continue to assume that negative
entries are unused/LRU/dentry_freeable/ref==0, and not worry about calling
kref_put().

X: Buffer Cache, Page Cache, Buffer Heads, WTH?
-------------------------------
X.1: Page vs Buffer Cache
--------------------
So we discussed the page cache, but as described, it does not satisfy all of
our disk caching needs.  Traditionally, there would also be a 'buffer cache.'
Buffers usually refer to memory used to hold data from the disk (or network).
I can think of a couple different ways someone could think of the buffer
cache, and to understand them, we first need to understand what we need to be
caching.

There are two main types of FS data: file data and metadata.  This metadata
is FS-specific data accessed by the VFS to get to a file's contents.  For
example, the superblock, inodes, inode indirect blocks, and to a certain
extent the directory's contents are all metadata.  There isn't an FS file (not
to be confused with an OS file) that corresponds to this data, but it needs to
be read in and cached.  Furthermore, this cache needs to be managed and
written back when dirty.  Note that file data can be broken up into two
different types: read()/write() data and mmap'd data.  When people talk about
buffer caches versus page caches, they are talking about these two different
types of file data (at least NetBSD did
http://www.usenix.org/event/usenix2000/freenix/full_papers/silvers/silvers_html/).

A very simple buffer cache would include anything *ever* read from disk.  It
would then get copied into the page cache in PGSIZE chunks for the page cache.
This would suck, since we now have two or three copies.  We obviously want to
use the page cache for both mmap() and read/write().  It's not clear about the
metadata.

Another usage of a buffer cache would be to cache only the disk blocks needed
for metadata.  I think this is what Linux did before it unified its buffer and
page caches (implied in UTLK).  The main issue with this is that you have two
different systems for managing essentially similar data - we only want to deal
with flushing, syncing, and writebacks of one subsystem, not in multiple
different subsystems.

The solution is to use the page cache to cache the metadata blocks' buffers.
We do this by having the block device be 'mapped' (but not read) in PGSIZE
chunks through its own struct page_mapping (a.k.a. struct address_space in
Linux).  This way, both files and the block device are mapped in PGSIZE chunks
via the same page_mapping structure, and will be managed by the same code.

Sort of.  We don't actually read in PGSIZE chunks for the block buffer cache.
If blocks will be in the bdev's cache, then they will be adjacent and on the
same page.  It is possible some adjacent blocks (which would be on the same
page) are not both metadata.

A more accurate way to describe what we do is that metadata blocks are copied
into a 'buffer cache' that is mapped and managed similarly to the page cache.
Pages are made of buffer heads, which hold data, and the reclaiming of pages of
memory from either the page cache or the buffer cache will use the same code -
since they are both just made of buffer pages.

X.2: Mapping Blockdev Data vs File Data
--------------------
An important distinction between file data (as managed by an inode) and the
bdev is that PGSIZE chunks of data for the bdev *must* be made of contiguous
disk blocks.  Ideally, file data is also contiguous/sequential, but that is
not always the case - hence the need for the inode's block pointers.  This
means that the chunk at the bottom of the page_mapping radix tree is a page,
and on that page there may be several buffers, holding the data of
nonsequential disk blocks - but that not all pages are like this.  The bdev
pages are made up of sequential blocks due to the very nature of what we are
mapping; there's no inode abstraction in between.

There are a couple ways we could handle this.  We adopt the Linux approach of
using something called a buffer head (BH), which describes the mapping from
in-memory buffer to block device / block number.  These are slab-allocated,
and exist for each buffer of a page.  The page itself points to the first of
its BHs, all of which exist in a LL.  The maximum number of them is determined
by PGSIZE / blocksize.  Whenever there is a page in the page cache (meaning, in
a page_mapping), that is up to date, it will have a BH.

Another way would be to not have BHs at all, and just figure out (at
operation-time) what the n blocks on disk are for any given page, and submit
the IO operations for those blocks.  The BHs serve as a cache of that info.
They also serve as a location to store data about the mapping, such as whether
it is dirty or not, whether the data is up to date or not (has it been read
in), etc.  The caching ought to be rather beneficial, since it means we do not
need to access the disk-inode and indirect blocks whenever we want to perform
an operation (which may be frequent - imagine the periodic writeback of an
mmap'd file undergoing writes).  The per-buffer dirty tracking will also help:
no need to write unrelated blocks if only one is edited (though this will not
help with mmap(), since we don't usually know which part of the page is
written).

X.4: Do we always need BHs?
------------------
But what about those pages made up of contiguous blocks?  We don't have a
bunch of independent, non-sequential blocks that we need to map.  Do we need a
bunch of BHs for that?  Do we need any?  It really seems excessive to break it
up into separate buffers for no reason.  At the other extreme, we could get by
without having a BH at all, though this gets back to the other issue of
caching.  What we do (or will do) is have one BH for the entire page of
contiguous blocks.  If the page is a "buffer page," in Linux terms (meaning it
has separate buffers), it will have n BHs in a LL.  Either way, we'll always
have the mapping handy.  We wouldn't need to re-verify the contiguous nature of
the blocks anyways, since the fact that the page was up to date and didn't need
a BH would mean it was contiguous.  Further benefits to using the one BH
include: 1) we are likely to make BHs be the unit of IO *submission*, and having
one handy will simplify that code. 2) some code paths within the VFS may get BHs
as a return value, which they can then dirty.  Always having a BH makes this
easier (no need to find out if it's a buffer page, then decide how to dirty it).

Another compliation with this is certain code will want a block in the middle
of a page (very common for metadata).  That code will get the BH for the
entire page back.  It will need to determine the starting block and then the
offset of the block it wants.  Note, this usage of the BHs is finding the
buffer corresponding to a block number.  The BH is a bidirectional mapping
between buffers and blockdev:blocknum.  These calculations are a minor
complication, but easy enough to do, and will probably be worth it.  The
tradeoff is against having multiple BHs, which would mean multiple block IO
requests for writing a single page (and writing the whole page may be a common
operation).

X.3: What about opening a blockdev as a file?
--------------------
Basically, don't do it for now.  The way we actually do things is a buffer cache
page is "up to date", but has no BHs or data until a specific block is
requested.  This is because we don't always know if all the blocks mapped by a
page are actually metadata blocks.  If they aren't we'll have issues where we
read in extra blocks that exist in both a file's page cache and the block
device's buffer cache.

A safe invariant is that a given block will only ever be in one cache: either a
file's page mapping or the buffer cache's page mapping.  When these blocks are
freed, we'll need to rip them out of their mapping (and possibly flush them).

There is one easy way to avoid this: don't open a bdev file if a file system is
mounted on top of it.  If you do, don't be surprised about inconsistencies.
Ideally, the FS will never leave the actual disk in an inconsistent state, but
the bdev's page cache could read things at different times and get some weird
results.  Just don't do this (for now - not like I plan to make this possible
any time soon).

Could we use the same buffers for both the blockdev-file page mapping and the
inode-file page mapping?  No - because the inode-file has the inode
indirection in between, which means a PGSIZE chunk of the file might not be
contiguous (as mentioned above).

We could have tried to avoid this bdev file problem by having the page mapping
radix trees point to a set of BHs that describes that page worth of buffers,
so that the bdev and an inode pointing to the same data will use the same
buffers and BHs.  That won't work, since the files buffers/blocks aren't in
the same order as they are on disk within a page.  Instead, we'd need to have
the page mapping go down to the FS's blocksize granularity, not to PGSIZE, so
that we could have independent leaves point to wherever they want.  This would
push the specific block device's blocksize into the VFS (which I don't like).

But the blocksize-infecting-the-VFS alone isn't too bad.  The real issue is
what is at the end of the page mapping.  Is it a page or a buffer/BH?  We want
pages for two related reasons: higher levels of the kernel's IO systems deal
with pages:
1. mmap().  Good old mmap() requires at least a page of contiguous block
buffers, so that the physical page frame can be mapped directly into a
process's address space.
2. Fast IO.  We want to use page remapping for fast IO.  This means that IO
has to be in PGSIZE chunks - not blocksize chunks.

x.4: Managing Dirty Buffers
--------------------
Many (some/all?) of the block functions will return a BH to describe the
mapping.  One reason for doing this (mentioned above) is to allow the caller
to manage if a buffer is dirty or not.  The tracking of dirty pages is done by
the page cache.  The dirtying is done by the caller; it can simply mark the BH
dirty and forget about it.  The writeback is handled usually by the page cache
or is triggered by an FS event.  Eventually, we'll have a kernel thread that
periodically writes dirty buffers back to disk.  Processes can also do this by
calling fsync.  FSs themselves will trigger syncs of metadata.  This will come
from having dirty SBs and inodes in the VFS.

Note, the issue of whether or not we pin metadata blocks, such as the inode's
indirect blocks (or other related blocks), in the page cache is independent of
all these issues.  If they are not cached / pinned, we would just have to
block and reread the inode's structures back into memory and proceed (for
instance, we'd do this when identifying blocks for a page mapping on a file
read).  The reason we wouldn't want to pin them is to save memory.

x.5: Reference Counting BHs and Pages
--------------------
There are a lot of complications with this, and if there is a good reason,
we'll change this later.

x.5.1: Basics
-----------
So we talk about passing around BHs, both when submitting IO ops and when
returning from things like get_buffer().  However, currently, we do not kref
or reference count BHs in any way.  Instead, we kref pages.  We do this (for
now) for a couple reasons:
1) Pages are the unit of memory management in the kernel.  Higher levels of
the kernel will pin/incref the page (such as in an mmap()).
2) BHs hang off of a page, but exist only to be expressive about the
management of the buffers on the page.  It's not like how a file hangs off a
dentry, where the dentry doesn't know (or care) about the file.
3) We already refcount pages.  While we could do the same for the BHs, it is a
bit redundant.  Any place that would be kreffing the BH can just kref the
whole page.  Code that receives a BH as a return value is actually getting a
page under the covers (though should use put_buffer() to drop its reference).

x.5.2: How we refcnt pages in a page mapping
-----------
When a page is in the page cache, we give it one kref.  Whenever a function or
subsystem is using one of these pages (IO, using the data, etc), there needs
to be a kref.  When it is time to remove the page from the page mapping, the
code needs to remove it from the radix tree, then kref_put it.  When the final
user is done with it (ideally, there is none, but we need to handle the case)
the release function will clean up the page - including freeing its BHs.

This does suck in that we could be removing an item from the page cache while
it is being used, violating some LRU policy.  The actual decision to remove it
should use those policies (when possible); the kreffing is simply to avoid
issues from race conditions.  (A reader starts using a page right before it is
ripped from the mapping).

x.5.3: More issues with Evictions
-----------
One issue with this is that dirty pages/buffers will need to be written back.
If someone tries to read while the page is removed from the page_mapping, but
before it is written back, they could get an old version if the read happens
before the write.  This is only an issue if the page is dirty.  One option
would be to writeback the page/buffer, then later remove it from the page
cache when it is read.  There's issues with concurrent writers, though if that
happens, we probably don't really want to remove it (it was LRU).  Note this
is an issue regardless of whether or not BHs are refcounted.  Also, this is
not an issue for when we kick a dentry/inode out of the cache - there should
be no one else trying to use it (since their refcnt was 0).  This is just a
concern when the system runs low on memory and wants to reclaim potentially
memory held by caches.

Also note that writeback of pages will happen regardless of eviction plans
(fsync, every n sec, etc).

This refcounting pattern is very similar to unlinking, where you can continue
to access a file once it is removed from the directory.  The difference here
is that future requests will get the same object (the page), unlike in the FS,
where you get ENOENT.  The page mapping is a cache, and you need to get the
same old data that was in there before the eviction.

A final issue is when the VFS aggressively pins blockdev (metadata) buffers.
Ideally, we'd like to be able to expel pages/buffers even if they are
refcnt'd.  The subsystems will always want to keep stuff in RAM.  This
also/especially applies to mmap().  One solution would be to keep them in RAM,
but have the BH keep track of who is holding its reference.  Then we could
unmap the page, which would need to get read back in on its next access.  We'd
need (or ought to have) some sort of callbacks.  This will get solved later
when we deal with unmapping mmap'd files, since it is the same problem - just
with different code and actors.

x.5.4: What about buffers inside pages?
-----------
For a while, I thought about refcounting BHs/buffers.  The issue that drives
it is the buffer cache (block dev page mapping holding metadata blocks of a
FS).  We had been operating on the cache in page-sized chunks, which
erroneously was reading in blocks adjacent to metadata blocks.  This would
have been an issue when we write back pages that have dirty blocks; blocks
were erroneously in the metadata cache and would overwrite potentially
file-realted blocks that were in an incoherent cache (the file/inode's page
mapping).

We broke the bdev's buffer cache up into smaller BHs, so that only metadata
blocks get read in, but we eventually will have to get rid of a metadata block
(and not the entire page from the cache) (ex: an inode is removed from the
disk - its indirect blocks need to go, and they could be next to anything on
disk).  The desire for the BH refcnt came from wanting to rip the BHs out of
the list when it was time to evict them from the cache (in case they became
file blocks later).  It isn't clear what the best way to do this is.  Probably
we'd have all users refcnt the BHs, which refcnt the pages.  However, some
users (like mmap and the file page cache) operate in page chunks - so this
would require them to incref and decref the BHs.  Triggering the page reclaim
might also be tricky.  One option would be to just rip a page from the cache,
callback to whoever has it loaded so they fault it back in later, and
sleep-block everyone who interferes with the operation.  Yikes.

Another option was to forget about the BH <-> page crap for the buffer cache,
and just have a standalone buffer cache with BHs as its unit of operation
(instead of a page), and replicate the algorithms/code for the buffer cache.
There is still a notion of BHs in a page, and page_release() / page_free()
would probably have to be a little different since its page isn't really a
PG_BUFFER (but it really is).

Instead, here's what we do: we refcnt at page granularity, since all users can
be considered users of a page.  While it's not fine-grained, it does represent
the idea that someone doesn't want the page to be freed.  It's similar to
having a dentry be the refcnt source when someone really wants the inode/file.
When we want to remove a page from the block buffer cache, all we do is mark
it as not dirty and not up-to-date.  Now, whenever the page gets evicted from
the cache, we only write back those block buffers that are dirty, so the
recently evicted block will not be written back.  If we attempt to read the
block in the future (perhaps it is reallocated as a metablock), then the BH is
still there for the mapping, but is simply not up to date.  The code already
knows how to handle this (since it could happen during a race condition), and
it will simply read the buffer back in.  This new reading is necessary, since
it is possible that the block was used for file IO in between the uses of it
as a metablock.

If there are more than one thread operating on a block buffer in the page
cache, then at the level of the cache, there is a race.  One could be marking
it as not dirty while another is dirtying it, etc.  However, no one should be
removing a buffer (aka, deallocating a block) while it is in use.  This sort
of concurrency problem should be sorted higher in the software stack (like at
the inode).

On a similar note, no one should ever remove a block's buffer from the
metadata buffer cache if it is dirty.  When those removals happen, it means
the block should be dealloced on the block device - meaning no one cares what
happens to it.  It's not meant to have data preserved.

x.6: What about Directories?  Inodes or metadata?
--------------------
Directories have inodes that have blocks scattered around the disk.  The blocks
making up the body of a directory are not sequential, like when you read blocks
from a bdev.  If you wanted a "page" of a directory, then you'd need to use the
i_mapping (page cache of a file) to access it, like any other file.

However, we don't want a page of a directory - at least not yet.  The main
reason we can get away with this is due to the lack of a mmap() or a desire to
page-remap for directory-contents IO.  It's all for the kernel's internal use.
At no point does anyone call generic_file_read() or _write() on it.

That being said, we treat the blocks of a directory as metadata blocks.  We do
figure out which blocks they are by walking the inode (also made of metadata
blocks), but don't bother to set up a page map for the directory itself.  We
just use the inode to figure out which metadata blocks we want, then read them
in (out of the blockdev's page cache).

Note that userspace issues reads on the directory.  This is mostly a convenience
thing (or an inconvenience thing), which just ends up being a wrapper around
readdir() (via generic_dir_read()).