GEOM -- modular disk I/O request transformation framework.
The GEOM framework provides an infrastructure in which "classes" can perform
transformations on disk I/O requests on their path from the upper
kernel to the device drivers and back.
Transformations in a GEOM context range from the simple geometric displacement
performed in typical disk partitioning modules over RAID algorithms
and device multipath resolution to full blown cryptographic protection
of the stored data.
Compared to traditional "volume management", GEOM differs from most and
in some cases all previous implementations in the following ways:
+o GEOM is extensible. It is trivially simple to write a new class of
transformation and it will not be given stepchild treatment. If
someone for some reason wanted to mount IBM MVS diskpacks, a class
recognizing and configuring their VTOC information would be a trivial
matter.
+o GEOM is topologically agnostic. Most volume management implementations
have very strict notions of how classes can fit together, very
often one fixed hierarchy is provided for instance subdisk - plex -
volume.
Being extensible means that new transformations are treated no differently
than existing transformations.
Fixed hierarchies are bad because they make it impossible to express the
intent efficiently. In the fixed hierarchy above it is not possible to
mirror two physical disks and then partition the mirror into subdisks,
instead one is forced to make subdisks on the physical volumes and to
mirror these two and two resulting in a much more complex configuration.
GEOM on the other hand does not care in which order things are done, the
only restriction is that cycles in the graph will not be allowed.
TERMINOLOGY and TOPOLOGY [Toc] [Back] Geom is quite object oriented and consequently the terminology borrows a
lot of context and semantics from the OO vocabulary:
A "class", represented by the data structure g_class implements one particular
kind of transformation. Typical examples are MBR disk partition,
BSD disklabel, and RAID5 classes.
An instance of a class is called a "geom" and represented by the data
structure "g_geom". In a typical i386 FreeBSD system, there will be one
geom of class MBR for each disk.
A "provider", represented by the data structure "g_provider", is the
front gate at which a geom offers service. A provider is "a disk-like
thing which appears in /dev" - a logical disk in other words. All
providers have three main properties: name, sectorsize and size.
A "consumer" is the backdoor through which a geom connects to another
geom provider and through which I/O requests are sent.
The topological relationship between these entities are as follows:
+o A class has zero or more geom instances.
+o A geom has exactly one class it is derived from.
+o A geom has zero or more consumers.
+o A geom has zero or more providers.
+o A consumer can be attached to zero or one providers.
+o A provider can have zero or more consumers attached.
All geoms have a rank-number assigned, which is used to detect and prevent
loops in the acyclic directed graph. This rank number is assigned
as follows:
1. A geom with no attached consumers has rank=1
2. A geom with attached consumers has a rank one higher than the highest
rank of the geoms of the providers its consumers are attached
to.
SPECIAL TOPOLOGICAL MANEUVERS [Toc] [Back] In addition to the straightforward attach, which attaches a consumer to a
provider, and detach, which breaks the bond, a number of special topological
maneuvers exists to facilitate configuration and to improve the
overall flexibility.
TASTING is a process that happens whenever a new class or new provider is
created and it provides the class a chance to automatically configure an
instance on providers, which it recognize as its own. A typical example
is the MBR disk-partition class which will look for the MBR table in the
first sector and if found and validated it will instantiate a geom to
multiplex according to the contents of the MBR.
A new class will be offered to all existing providers in turn and a new
provider will be offered to all classes in turn.
Exactly what a class does to recognize if it should accept the offered
provider is not defined by GEOM, but the sensible set of options are:
+o Examine specific data structures on the disk.
+o Examine properties like sectorsize or mediasize for the provider.
+o Examine the rank number of the provider's geom.
+o Examine the method name of the provider's geom.
ORPHANIZATION is the process by which a provider is removed while it
potentially is still being used.
When a geom orphans a provider, all future I/O requests will "bounce" on
the provider with an error code set by the geom. Any consumers attached
to the provider will receive notification about the orphanization when
the eventloop gets around to it, and they can take appropriate action at
that time.
A geom which came into being as a result of a normal taste operation
should selfdestruct unless it has a way to keep functioning lacking the
orphaned provider. Geoms like diskslicers should therefore selfdestruct
whereas RAID5 or mirror geoms will be able to continue, as long as they
do not loose quorum.
When a provider is orphaned, this does not necessarily result in any
immediate change in the topology: any attached consumers are still
attached, any opened paths are still open, any outstanding I/O requests
are still outstanding.
The typical scenario is
+o A device driver detects a disk has departed and orphans the
provider for it.
+o The geoms on top of the disk receive the orphanization event
and orphans all their providers in turn. Providers, which are
not attached to, will typically self-destruct right away. This
process continues in a quasi-recursive fashion until all relevant
pieces of the tree has heard the bad news.
+o Eventually the buck stops when it reaches geom_dev at the top
of the stack.
+o Geom_dev will call destroy_dev(9) to stop any more request from
coming in. It will sleep until all (if any) outstanding I/O
requests have been returned. It will explicitly close (ie:
zero the access counts), a change which will propagate all the
way down through the mesh. It will then detach and destroy its
geom.
+o The geom whose provider is now attached will destroy the
provider, detach and destroy its consumer and destroy its geom.
+o This process percolates all the way down through the mesh,
until the cleanup is complete.
While this approach seems byzantine, it does provide the maximum flexibility
and robustness in handling disappearing devices.
The one absolutely crucial detail to be aware is that if the device
driver does not return all I/O requests, the tree will not unravel.
SPOILING is a special case of orphanization used to protect against stale
metadata. It is probably easiest to understand spoiling by going through
an example.
Imagine a disk, "da0" on top of which a MBR geom provides "da0s1" and
"da0s2" and on top of "da0s1" a BSD geom provides "da0s1a" through
"da0s1e", both the MBR and BSD geoms have autoconfigured based on data
structures on the disk media. Now imagine the case where "da0" is opened
for writing and those data structures are modified or overwritten: Now
the geoms would be operating on stale metadata unless some notification
system can inform them otherwise.
To avoid this situation, when the open of "da0" for write happens, all
attached consumers are told about this, and geoms like MBR and BSD will
selfdestruct as a result. When "da0" is closed again, it will be offered
for tasting again and if the data structures for MBR and BSD are still
there, new geoms will instantiate themselves anew.
Now for the fine print:
If any of the paths through the MBR or BSD module were open, they would
have opened downwards with an exclusive bit rendering it impossible to
open "da0" for writing in that case and conversely the requested exclusive
bit would render it impossible to open a path through the MBR geom
while "da0" is open for writing.
From this it also follows that changing the size of open geoms can only
be done with their cooperation.
Finally: the spoiling only happens when the write count goes from zero to
non-zero and the retasting only when the write count goes from non-zero
to zero.
INSERT/DELETE are a very special operation which allows a new geom to be
instantiated between a consumer and a provider attached to each other and
to remove it again.
To understand the utility of this, imagine a provider with being mounted
as a file system. Between the DEVFS geoms consumer and its provider we
insert a mirror module which configures itself with one mirror copy and
consequently is transparent to the I/O requests on the path. We can now
configure yet a mirror copy on the mirror geom, request a synchronization,
and finally drop the first mirror copy. We have now in essence
moved a mounted file system from one disk to another while it was being
used. At this point the mirror geom can be deleted from the path again,
it has served its purpose.
CONFIGURE is the process where the administrator issues instructions for
a particular class to instantiate itself. There are multiple ways to
express intent in this case, a particular provider can be specified with
a level of override forcing for instance a BSD disklabel module to attach
to a provider which was not found palatable during the TASTE operation.
Finally IO is the reason we even do this: it concerns itself with sending
I/O requests through the graph.
I/O REQUESTS represented by struct bio, originate at a consumer, are
scheduled on its attached provider, and when processed, returned to the
consumer. It is important to realize that the struct bio which enters
through the provider of a particular geom does not "come out on the other
side". Even simple transformations like MBR and BSD will clone the
struct bio, modify the clone, and schedule the clone on their own consumer.
Note that cloning the struct bio does not involve cloning the
actual data area specified in the IO request.
In total four different IO requests exist in GEOM: read, write, delete,
and get attribute.
Read and write are self explanatory.
Delete indicates that a certain range of data is no longer used and that
it can be erased or freed as the underlying technology supports. Technologies
like flash adaptation layers can arrange to erase the relevant
blocks before they will become reassigned and cryptographic devices may
want to fill random bits into the range to reduce the amount of data
available for attack.
It is important to recognize that a delete indication is not a request
and consequently there is no guarantee that the data actually will be
erased or made unavailable unless guaranteed by specific geoms in the
graph. If "secure delete" semantics are required, a geom should be
pushed which converts delete indications into (a sequence of) write
requests.
Get attribute supports inspection and manipulation of out-of-band
attributes on a particular provider or path. Attributes are named by
ascii strings and they will be discussed in a separate section below.
(stay tuned while the author rests his brain and fingers: more to come.)
This software was developed for the FreeBSD Project by Poul-Henning Kamp
and NAI Labs, the Security Research Division of Network Associates, Inc.
under DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
DARPA CHATS research program.
The first precursor for GEOM was a gruesome hack to Minix 1.2 and was
never distributed. An earlier attempt to implement a less general scheme
in FreeBSD never succeeded.
Poul-Henning Kamp <[email protected]>
FreeBSD 5.2.1 March 27, 2002 FreeBSD 5.2.1 [ Back ] |
