LU-3259 clio: Revise read ahead implementation

[fs/lustre-release.git] / lustre / doc / clio.txt

               ******************************************************
               * Overview of the Lustre Client I/O (CLIO) subsystem *
               ******************************************************

Original Authors:
=================
Nikita Danilov <Nikita_Danilov@xyratex.com>

Topics
======

1. Overview
        1.1. Goals
        1.2. Terminology
                i.   I/O vs. Transfer
                ii.  Top-{object,lock,page}, Sub-{object,lock,page}
                iii. VVP, llite
        1.3. Main Differences with the Pre-CLIO Client Code
        1.4. Layered objects, Slices, Headers
        1.5. Instantiation
        1.6. Life Cycle
        1.7. State Machines
        1.8. Finalization
        1.9. Code Structure
2. Layers
        2.1. VVP, Echo-client
        2.2. LOV, LOVSUB (layouts)
        2.3. OSC
3. Objects
        3.1. FID, Hashing, Caching, LRU
        3.2. Top-object, Sub-object
        3.3. Object Operations
        3.4. Object Attributes
        3.5. Object Layout
4. Pages
        4.1. Page Indexing
        4.2. Page Ownership
        4.3. Page Transfer Locking
        4.4. Page Operations
        4.5. Page Initialization
5. Locks
        5.1. Lock Life Cycle
        5.2. cl_lock and LDLM Lock
        5.3. Use Case: Lock Invalidation
6. IO
        6.1. Fixed IO Types
        6.2. IO State Machine
        6.3. Parallel IO
        6.4. Data-flow: From Stack to IO Slice
7. Transfer
        7.1. Immediate vs. Opportunistic Transfers
        7.2. Page Lists
        7.3. Transfer States: Prepare, Completion
        7.4. Page Completion Handlers, Synchronous Transfer
8. LU Environment (lu_env)
        8.1. Motivation, Server Environment Usage
        8.2. Client Usage
        8.3. Sub-environments
9. Use Cases
        9.1. Inode Creation
        9.2. First IO to a File
                i. Read, Read-ahead
                ii. Write
        9.3. Lock-less and No-cache IO

================
= 1. Overview =
================

1.1. Goals
==========

CLIO is a re-write of interfaces between layers in the client data-path (read,
write, truncate). Its goals are:

- Reduce the number of bugs in the IO path;

- Introduce more logical layer interfaces instead of current all-in-one OBD
  device interface;

- Define clear and precise semantics for the interface entry points;

- Simplify the structure of the client code.

- Support upcoming features:

        . SNS,
        . p2p caching,
        . parallel non-blocking IO,
        . pNFS;

- Reduce stack consumption.

Restrictions:

- No meta-data changes;
- No support for 2.4 kernels;
- Portable code;
- No changes to recovery;
- The same layers with mostly the same functionality;
- As few changes to the core logic of each Lustre data-stack layer as possible
  (e.g., no changes to the read-ahead or OSC RPC logic).

1.2. Terminology
================

Any discussion of client functionality has to talk about `read' and `write'
system calls on the one hand and about `read' and `write' requests to the
server on the other hand. To avoid confusion, the former high level operations
are called `IO', while the latter are called `transfer'.

Many concepts apply uniformly to pages, locks, files, and IO contexts, such as
reference counting, caching, etc. To describe such situations, a common term is
needed to denote things from any of the above classes. `Object' would be a
natural choice, but files and especially stripes are already called objects, so
the term `entity' is used instead.

Due to the striping it's often a case that some entity is composed of multiple
entities of the same kind: a file is composed of stripe objects, a logical lock
on a file is composed of stripe locks on the file's stripes, etc. In these
cases we shall talk about top-object, top-lock, top-page, top-IO, etc. being
constructed from sub-objects, sub-locks, sub-pages, sub-io's respectively.

The topmost module in the Linux client, is traditionally known as `llite'.  The
corresponding CLIO layer is called `VVP' (VFS, VM, POSIX) to reflect its
functional responsibilities.

1.3. Main Differences with the Pre-CLIO Client Code
===================================================

- Locks on files (as opposed to locks on stripes) are first-class objects (i.e.
  layered entities);

- Sub-objects (stripes) are first-class objects;

- Stripe-related logic is moved out of llite (almost);

- IO control flow is different:

        . Pre-CLIO: llite implements control flow, calling underlying OBD
          methods as necessary;

        . CLIO: generic code (cl_io_loop()) controls IO logic calling all
          layers, including VVP.

  In other words, VVP (or any other top-layer) instead of calling some
  pre-existing `lustre interface', also implements parts of this interface.

- The lu_env allocator from MDT is used on the client.

1.4. Layered Objects
====================

CLIO continues the layered object approach that was found to be useful for the
MDS rewrite in Lustre 2.0. In this approach, instances of key entity types
(files, pages, locks, etc.) are represented as a header, containing attributes
shared by all layers. Each header contains a linked list of per-layer `slices',
each of which contains a pointer to a vector of function pointers. Generic
operations on layered objects are implemented by going through the list of
slices and invoking the corresponding function from the operation vector at
every layer.  In this way generic object behavior is delegated to the layers.

For example, a page entity is represented by struct cl_page, from which hangs
off a list of cl_page_slice structures, one for each layer in the stack.
cl_page_slice contains a pointer to struct cl_page_operations.
cl_page_operations has the field

        void (*cpo_completion)(const struct lu_env *env,
              const struct cl_page_slice *slice, int ioret);

When transfer of a page is finished, ->cpo_completion() methods are called in a
particular order (bottom to top in this case).

Allocation of slices is done during instance creation. If a layer needs some
private state for an object, it embeds the slice into its own data structure.
For example, the OSC layer defines

        struct osc_lock {
                struct cl_lock_slice ols_cl;
                struct ldlm_lock *ols_lock;
                ...
        };

When an operation from cl_lock_operations is called, it is given a pointer to
struct cl_lock_slice, and the layer casts it to its private structure (for
example, struct osc_lock) to access per-layer state.

The following types of layered objects exist in CLIO:

- File system objects (files and stripes): struct cl_object_header, slices
  are of type struct cl_object;

- Cached pages with data: struct cl_page, slices are of type
  cl_page_slice;

- Extent locks: struct cl_lock, slices are of type cl_lock_slice;

- IO content: struct cl_io, slices are of type cl_io_slice;

- Transfer request: struct cl_req, slices are of type cl_req_slice.

1.5. Instantiation
==================

Entities with different sequences of slices can co-exist. A typical example of
this is a local vs. remote object on the MDS. A local object, based on some
file in the local file system has MDT, MDD, LOD and OSD as its layers, whereas
a remote object (representing an object local to some other MDT) has MDT, MDD,
LOD, and OSP layers.

When the client is being mounted, its device stack is configured according to
llog configuration records. The typical configuration is

                           vvp_device
                                |
                                V
                           lov_device
                                |
                        +---+---+---+---+
                        |   |   |   |   |
                        V   V   V   V   V
                      .... osc_device's ....

In this tree every node knows its descendants. When a new file (inode) is
created, every layer, starting from the top, creates a slice with a state and
an operation vector for this layer, appends this slice to the tail of a list
anchored at the object header, and then calls the corresponding lower layer
device to do the same. That is, the file object structure is determined by the
configuration of devices to which this file belongs.

Pages and locks, in turn, belong to the file objects, and when a new page is
created for a given object, slices of this object are iterated through and
every slice is asked to initialize a new page, which includes (usually)
allocation of a new page slice and its insertion into a list of page slices.
Locks and IO context instantiation is handled similarly.

1.6. Life cycle
===============

All layered objects except locks, IO contexts and transfer requests (which
leaves file objects, pages) are reference counted and cached. They have a
uniform caching mechanism:

- Objects are kept in some sort of an index (global FID hash for file objects,
  per-file radix tree for pages, and per-file list for locks);

- A reference for an object can be acquired by cl_{object,page}_find() functions
  that search the index, and if object is not there, create a new one
  and insert it into the index;

- A reference is released by cl_{object,page}_put() functions. When the last
  reference is released, the object is returned to the cache (still in the
  index), except when the user explicitly set `do not cache' flag for this
  object. In the latter case the object is destroyed immediately.

Locks(cl_lock) are owned by individual IO threads. cl_lock is a container of
lock requirements for underlying DLM locks. cl_lock is allocated by
cl_lock_request() and destroyed by cl_lock_cancel(). DLM locks are cacheable
and can be reused by cl_locks.

IO contexts are owned by a thread (or, potentially a group of threads) doing
IO, and need neither reference counting nor indexing. Similarly, transfer
requests are owned by a OSC device, and their lifetime is from RPC creation
until completion notification.

1.7. State Machines
===================

All types of layered objects contain a state-machine, although for the transfer
requests this machine is trivial (CREATED -> PREPARED -> INFLIGHT ->
COMPLETED), and for the file objects it is very simple.  Page, lock, and IO
state machines are described in more detail below.

As a generic rule, state machine transitions are made under some kind of lock:
VM lock for a page, and LU site spin-lock for an object. After some event that
might cause a state transition, such lock is taken, and the object state is
analysed to check whether transition is possible. If it is, the state machine
is advanced to the new state and the lock is released. IO state transitions do
not require concurrency control.

1.8. Finalization
=================

State machine and reference counting interact during object destruction. In
addition to temporary pointers to an entity (that are counted in its reference
counter), an entity is reachable through

- Indexing structures described above

- Pointers internal to some layer of this entity. For example, a page is
  reachable through a pointer from VM page, and sub-{object, lock, page} might
  be reachable through a pointer from its top-entity.

Entity destruction happens in three phases:

- First, a decision is made to destroy an entity, when, for example, a page is
  truncated from a file, or an inode is destroyed. At this point the `do not
  cache' bit is set in the entity header, and all ways to reach entity from
  internal pointers are severed.

  cl_{page,object}_get() functions never return an entity with the `do not
  cache' bit set, so from this moment no new internal pointers can be
  obtained.

- Pointers `drain' for some time as existing references are released. In
  this phase the entity is reachable through

        . temporary pointers, counted in its reference counter, and
        . possibly a pointer in the indexing structure.

- When the last reference is released, the entity can be safely freed (after
  possibly removing it from the index).

  See lu_object_put(), cl_page_put().

1.9. Code Structure
===================

The CLIO code resides in the following files:

  {llite,lov,osc}/*_{dev,object,lock,page,io}.c
  obdclass/cl_*.c
  include/cl_object.h

All global CLIO data-types are defined in include/cl_object.h header which
contains detailed documentation. Generic clio code is in
obdclass/cl_{object,page,lock,io}.c

An implementation of CLIO interfaces for a layer foo is located in
foo/foo_{dev,object,page,lock,io}.c files.

Definitions of data-structures shared within a layer are in
foo/foo_cl_internal.h.

=============
= 2. Layers =
=============

This section briefly outlines responsibility of every layer in the stack. More
detailed description of functionality is in the following sections on objects,
pages and locks.

2.1. VVP, Echo-client
==========================

There are currently 2 options for the top-most Lustre layer:

- VVP: linux kernel client,
- echo-client: special client used by the Lustre testing sub-system.

Other possibilities are:

- Client ports to other operating systems (OSX, Windows, Solaris),
- pNFS and NFS exports.

The responsibilities of the top-most layer include:

- Definition of the entry points through which Lustre is accessed by the
  applications;
- Interaction with the hosting VM/MM system;
- Interaction with the hosting VFS or equivalent;
- Implementation of the desired semantics of top of Lustre (e.g. POSIX
  or Win32 semantics).

Let's look at VVP in more detail. First, VVP implements VFS entry points
required by the Linux kernel interface: ll_file_{read,write,sendfile}(). Then,
VVP implements VM entry points: ll_{write,invalidate,release}page().

For file objects, VVP slice (vvp_object) contains a pointer to an inode.

For pages, the VVP slice (vvp_page) contains a pointer to the VM page
(struct page), a `defer up to date' flag to track read-ahead hits (similar to
the pre-CLIO client), and fields necessary for synchronous transfer (see
below).  VVP is responsible for implementation of the interaction between
client page (cl_page) and the VM.

There is no special VVP private state for locks.

For IO, VVP implements

- Mapping from Linux specific entry points (readv, writev, sendfile, etc.)
  to Lustre IO loop,

- mmap,

- POSIX features like short reads, O_APPEND atomicity, etc.

- Read-ahead (this is arguably not the best layer in which to implement
  read-ahead, as the existing read-ahead algorithm is network-aware).

2.2. LOV, LOVSUB
================

The LOV layer implements RAID-0 striping. It maps top-entities (file objects,
locks, pages, IOs) to one or more sub-entities. LOVSUB is a companion layer
that does the reverse mapping.

2.3. OSC
========

The OSC layer deals with networking stuff:

- It decides when an efficient RPC can be formed from cached data;

- It calls LNET to initiate a transfer and to get notification of completion;

- It calls LDLM to implement distributed cache coherency, and to get
  notifications of lock cancellation requests;

==============
= 3. Objects =
==============

3.1. FID, Hashing, Caching, LRU
===============================

Files and stripes are collectively known as (file system) `objects'. The CLIO
client reuses support for layered objects from the MDT stack. Both client and
MDT objects are based on struct lu_object type, representing a slice of a file
system object. lu_object's for a given object are linked through the
->lo_linkage field into a list hanging off field ->loh_layers of struct
lu_object_header, that represents a whole layered object.

lu_object and lu_object_header provide functionality common between a client
and a server:

- An object is uniquely identified by a FID; all objects are kept in a hash
  table indexed by a FID;

- Objects are reference counted. When the last reference to an object is
  released it is returned back into the cache, unless it has been explicitly
  marked for deletion, in which case it is immediately destroyed;

- Objects in the cache are kept in a LRU list that is scanned to keep cache
  size under control.

On the MDT, lu_object is wrapped into struct md_object where additional state
that all server-side objects have is stored. Similarly, on a client, lu_object
and lu_object_header are embedded into struct cl_object and struct
cl_object_header where additional client state is stored.

3.2. Top-object, Sub-object
===========================

An important distinction from the server side, where md_object and dt_object
are used, is that cl_object "fans out" at the LOV level: depending on the file
layout, a single file is represented as a set of "sub-objects" (stripes). At
the implementation level, struct lov_object contains an array of cl_objects.
Each sub-object is a full-fledged cl_object, having its FID and living in the
LRU and hash table. Each sub-object has its own radix tree of pages, and its
own list of locks.

This leads to the next important difference with the server side: on the
client, it's quite usual to have objects with the different sequence of layers.
For example, typical top-object is composed of the following layers:

- VVP
- LOV

whereas its sub-objects are composed of layers:

- LOVSUB
- OSC

Here "LOVSUB" is a mostly dummy layer, whose purpose is to keep track of the
object-subobject relationship:

                  cl_object_header-+--->object LRU list
                          |
                          V
           inode<----vvp_object
                          |
                          V
                     lov_object
                          |
                  +---+---+---+---+
                  |   |   |   |   |
                  V   |   |   |   |
     cl_object_header |   .   .   .
           |          |   .   .   .
           |          V
           .  cl_object_header-+--->object LRU list
                      |
                      V
                lovsub_object
                      |
                      V
                 osc_object

Sub-objects are not cached independently: when top-object is about to be
discarded from the memory, all its sub-objects are torn-down and destroyed too.

3.3. Object Operations
======================

In addition to the lu_object_operations vector, each cl_object slice has
cl_object_operations. lu_object_operations deals with object creation and
destruction of objects. Client specific cl_object_operations fall into two
categories:

- Creation of dependent entities: these are ->coo_{page,lock,io}_init()
  methods called at every layer when a new page, lock or IO context are being
  created, and

- Object attributes: ->coo_attr_{get,set}() methods that are called to get or
  set common client object attributes (struct cl_attr): size, [mac]times, etc.

3.4. Object Attributes
======================

A cl_object has a set of attributes defined by struct cl_attr. Attributes
include object size, object known-minimum-size (KMS), access, change and
modification times and ownership identifiers. Description of KMS is beyond the
scope of this document, refer to the (non-)existent Lustre documentation on the
subject.

Both top-objects and sub-objects have attributes. Consistency of the attributes
is protected by a lock on the top-object, accessible through
cl_object_attr_{un,}lock() calls. This allows a sub-object and its top-object
attributes to be changed atomically.

Attributes are accessible through cl_object_attr_{g,s}et() functions that call
per-layer ->coo_attr_{s,g}et() object methods. Top-object attributes are
calculated from the sub-object ones by lov_attr_get() that optimizes for the
case when none of sub-object attributes have changed since the last call to
lov_attr_get().

As a further potential optimization, one can recalculate top-object attributes
at the moment when any sub-object attribute is changed. This would allow to
avoid collecting cumulative attributes over all sub-objects. To implement this
optimization _all_ changes of sub-object attributes must go through
cl_object_attr_set().

3.5. Object Layout
==================

Layout of an object decides how data of the file are placed onto OSTs. Object
layout can be changed and if that happens, clients will have to reconfigure
the object layout by calling cl_conf_set() before it can do anything to this
object.

In order to be notified for the layout change, the client has to cache an
inodebits lock: MDS_INODELOCK_LAYOUT in the memory. To change an object's
layout, The MDT holds the EX mode of layout lock therefore all clients having
this object cached will be notified.

Reconfiguring layout of objects is expensive because it has to clean up page
cache and rebuild the sub-objects. There is a field lsm_layout_gen in
lov_stripe_md and it must be increased whenever the layout is changed for an
object. Therefore, if the revocation of layout lock is due to flase sharing of
ibits lock, the lsm_layout_gen won't be changed and the client can reuse the
page cache and subobjects.

CLIO uses ll_refresh_layout() to make sure that valid layout is fetched. This
function must be called before any IO can be started to an object.

============
= 4. Pages =
============

A cl_page represents a portion of a file, cached in the memory. All pages of
the given file are of the same size, and are kept in the radix tree of kernel
VM.

A cl_page is associated with a VM page of the hosting environment (struct page
in the Linux kernel, for example), cfs_page_t. It is assumed that this
association is implemented by one of cl_page layers (top layer in the current
design) that

- intercepts per-VM-page call-backs made by the host environment (e.g., memory
  pressure),

- translates state (page flag bits) and locking between lustre and the host
  environment.

The association between cl_page and cfs_page_t is immutable and established
when cl_page is created. It is possible to imagine a setup where different
pages get their backing VM buffers from different sources. For example, in the
case if pNFS export, some pages might be backed by local DMU buffers, while
others (representing data in remote stripes), by normal VM pages.

Unlike any other entities in CLIO, there is no subpage for cl_page.

4.1. Page Indexing
==================

Pages within a given object are linearly ordered. The page index is stored in
the ->cpl_index field in cl_page_slice. In a typical Lustre setup, a top-object
has an array of sub-objects, and every page in a top-object corresponds to a
page slice in one of its sub-objects.

There is a radix tree for pages on the OSC layer. When a LDLM lock is being
cancelled, the OSC will look up the radix tree so that the pages belong to
the corresponding lock extent can be destroyed.

4.2. Page Ownership
===================

A cl_page can be "owned" by a particular cl_io (see below), guaranteeing this
IO an exclusive access to this page with regard to other IO attempts and
various events changing page state (such as transfer completion, or eviction of
the page from memory). Note, that in general a cl_io cannot be identified with
a particular thread, and page ownership is not exactly equal to the current
thread holding a lock on the page. The layer implementing the association
between cl_page and cfs_page_t has to implement ownership on top of available
synchronization mechanisms.

While the Lustre client maintains the notion of page ownership by IO, the
hosting MM/VM usually has its own page concurrency control mechanisms. For
example, in Linux, page access is synchronized by the per-page PG_locked
bit-lock, and generic kernel code (generic_file_*()) takes care to acquire and
release such locks as necessary around the calls to the file system methods
(->readpage(), ->write_begin(), ->write_end(), etc.). This leads to the
situation when there are two different ways to own a page in the client:

- Client code explicitly and voluntary owns the page (cl_page_own());

- The hosting VM locks a page and then calls the client, which has to "assume"
  ownership from the VM (cl_page_assume()).

Dual methods to release ownership are cl_page_disown() and cl_page_unassume().

4.3. Page Transfer Locking
==========================

cl_page implements a simple locking design: as noted above, a page is protected
by a VM lock while IO owns it. The same lock is kept while the page is in
transfer.  Note that this is different from the standard Linux kernel behavior
where page write-out is protected by a lock (PG_writeback) separate from VM
lock (PG_locked). It is felt that this single-lock design is more portable and,
moreover, Lustre cannot benefit much from a separate write-out lock due to LDLM
locking.

4.4. Page Operations
====================

See documentation for cl_object.h:cl_page_operations. See cl_page state
descriptions in documentation for cl_object.h:cl_page_state.

4.5. Page Initialization
========================

cl_page is the most frequently allocated and freed entity in the CLIO stack. In
order to improvement the performance of allocation and free, a cl_page, along
with the corresponding cl_page_slice for each layer, is allocated as a single
memory buffer.

Now that the CLIO can support different type of object layout, and each layout
may lead to different cl_page in size. When an object is initialized, the object
initialization method ->loo_object_init() for each layer will decide the size of
buffer for cl_page_slice by calling cl_object_page_init(). cl_object_page_init()
will add the size to coh_page_bufsize of top cl_object_header and co_slice_off
of the corresponding cl_object is used to remember the offset of page slice for
this object.

============
= 5. Locks =
============

A struct cl_lock represents an extent lock on cached file or stripe data.
cl_lock is used only to collect the lock requirement to complete an IO. Due to
the existence of MMAP, it may require more than one lock to complete an IO.
Except the cases of direct IO and lockless IO, a cl_lock will be attached to a
LDLM lock on the OSC layer.

As locks protect cached data, and the unit of data caching is a page, locks are
of page granularity.

5.1. Lock Life Cycle
====================

The lock requirements are collected in cl_io_lock(). In cl_io_lock(), the
->cio_lock() method for each layers are invoked to decide the lock extent by
IO region, layout, and buffers. For example, in the VVP layer, it has to search
buffers of IO and if the buffers belong to a Lustre file mmap region, the locks
for the corresponding file will be requred.

Once the lock requirements are collected, cl_lock_request() is called to create
and initialize individual locks. In cl_lock_request(), ->clo_enqueue() is called
for each layers. Especially on the OSC layer, osc_lock_enqueue() is called to
match or create LDLM lock to fulfill the lock requirement.

cl_lock is not cacheable. The locks will be destroyed after IO is complete. The
lock destroying process starts from cl_io_unlock() where cl_lock_release() is
called for each cl_lock. In cl_lock_release(), ->clo_cancel() methods are called
for each layer to release the resource held by cl_lock. The most important
resource held by cl_lock is the LDLM lock on the OSC layer. It will be released
by osc_lock_cancel(). LDLM locks can still be cached in memory after being
detached from cl_lock.

5.2. cl_lock and LDLM Lock
==========================

As the result of enqueue, an LDLM lock is attached to a cl_lock_slice on the OSC
layer. The field ols_dlmlock in the osc_lock points to the LDLM lock.

When an LDLM lock is attached to a osc_lock, its use count(l_readers,l_writers)
is increased therefore it can't be revoked during this time. A LDLM lock can be
shared by multiple osc_lock, in that case, each osc_lock will hold the LDLM lock
according to their use type, i.e., increase l_readers or l_writers respectively.

When a cl_lock is cancelled, the corresponding LDLM lock will be released.
Cancellation of cl_lock does not necessarily cause the underlying LDLM lock to
be cancelled. The LDLM lock can be cached in the memory unless it's being
cancelled by OST.

To cache a page in the client memory, the page index must be covered by at
least one LDLM lock's extent. Please refer to section 5.3 for the details of
pages and lock.

5.3. Use Case: Lock Invalidation
================================

To demonstrate how objects, pages and lock data-structures interact, let's look
at the example of stripe lock invalidation.

Imagine that on the client C0 there is a file object F, striped over stripes
S0, S1 and S2 (files and stripes are represented by cl_object_header). Further,
C0 just finished a write IO to file F's offset [a, b] and left some clean and
dirty pages in C0 and corresponding LDLM lock LS0, LS1, and LS2 for the
corresponding stripes of S0, S1, and S2 respectively. From section 4.1, the
caching pages stay in the radix tree of S0, S1 and S2.

Some other client requests a lock that conflicts with LS1. The OST where S1
lives, sends a blocking AST to C0.

C0's LDLM invokes lock->l_blocking_ast(), which is osc_ldlm_blocking_ast(),
which eventually calls osc_cache_writeback_range() with the corresponding LDLM
lock's extent as parameters. To find the pages covered by a LDLM lock,
LDLM lock stores a pointer to osc_object in its l_ast_data.

In osc_cache_writeback_range(), it will check if there exist dirty pages for
the extent of this lock. If that is the case, an OST_WRITE RPC will be issued
to write the dirty pages back.

Once all pages are written, they are removed from radix-trees and destroyed.
osc_lock_discard_pages() is called for this purpose. It will look up radix tree
and then discard every page the extent covers.

=========
= 6. IO =
=========

An IO context (struct cl_io) is a layered object describing the state of an
ongoing IO operation (such as a system call).

6.1. Fixed IO Types
===================

There are two classes of IO contexts, represented by cl_io:

- An IO for a specific type of client activity, enumerated by enum cl_io_type:

        . CIT_READ: read system call including read(2), readv(2), pread(2),
          sendfile(2);
        . CIT_WRITE: write system call;
        . CIT_SETATTR: truncate and utime system call;
        . CIT_FAULT: page fault handling;
        . CIT_FSYNC: fsync(2) system call, ->writepages() writeback request;

- A `catch-all' CIT_MISC IO type for all other IO activity:

        . cancellation of an extent lock,
        . VM induced page write-out,
        . glimpse,
        . other miscellaneous stuff.

The difference between CIT_MISC and other IO types is that CIT_MISC IO is
merely a context in which pages are owned and locks are enqueued, whereas
other IO types, in addition to being a context, are also state machines.

6.2. IO State Machine
=====================

The idea behind the cl_io state machine is that initial `work' that has to be
done (e.g., writing a 3MB user buffer into a given file) is done as a sequence
of `iterations', and an iteration is executed as following an idiomatic
sequence of steps:

- Prepare: determine what work is to be done at this iteration;

- Lock: enqueue and acquire all locks necessary to perform this iteration;

- Start: either perform iteration work synchronously, or post it
  asynchronously, or both;

- End: wait for the completion of asynchronous work;

- Unlock: release locks, acquired at the "lock" step;

- Finalize: finalize iteration state.

cl_io is a layered entity and each step above is performed by invoking the
corresponding cl_io_operations method on every layer. As will be explained
below, this is especially important in the `prepare' step, as it allows layers
to cooperate in determining the scope of the current iteration.

Before an IO can be started, the client has to make sure that the object's
layout is valid. The client will check there exists a valid layout lock being
cached in the client side memory, otherwise, ll_layout_refresh() has to be
called to fetch uptodate layout from the MDT side.

For CIT_READ or CIT_WRITE IO, a typical scenario is splitting the original user
buffer into chunks that map completely inside of a single stripe in the target
file, and processing each chunk as a separate iteration. In this case, it is
the LOV layer that (in lov_io_rw_iter_init() function) determines the extent of
the current iteration.

Once the iteration is prepared, the `lock' step acquires all necessary DLM
locks to cover the region of a file that is affected by the current iteration.
The `start' step does the actual processing, which for write means placing
pages from the user buffer into the cache, and for read means fetching pages
from the server, including read-ahead pages (see `immediate transfer' below).
Truncate and page fault are executed in one iteration (currently that is, it's
easy to change truncate implementation to, for instance, truncate each stripe
in a separate iteration, should the need arise).

6.3. Parallel IO
================

One important planned generalization of this model is an out of order execution
of iterations.

A motivating example for this is a write of a large user level buffer,
overlapping with multiple stripes. Typically, a busy Lustre client has its
per-OSC caches for the dirty pages nearly full, which means that the write
often has to block, waiting for the cache to drain. Instead of blocking the
whole IO operation, CIT_WRITE might switch to the next stripe and try to do IO
there.  Without such a `non-blocking' IO, a slow OST or an unfair network
degrades the performance of the whole cluster.

Another example is a legacy single-threaded application running on a multi-core
client machine, where IO throughput is limited by the single thread copying
data between the user buffer to the kernel pages. Multiple concurrent IO
iterations that can be scheduled independently on the available processors
eliminate this bottleneck by copying the data in parallel.

Obviously, parallel IO is not compatible with the usual `sequential IO'
semantics. For example, POSIX read and write have a very simple failure model,
where some initial (possibly empty) segment of a user buffer is processed
successfully, and none of the remaining bytes were read and written. Parallel
IO can fail in much more complex ways.

For now, only sequential iterations are supported.

6.4. Data-flow: From Stack to IO Slice
======================================

The parallel IO design outlined above implies that an ongoing IO can be
preempted by other IO and later resumed, all potentially in the same thread.
This means that IO state cannot be kept on a stack, as it is customarily done
in UNIX file system drivers. Instead, the layered cl_io is used to store
information about the current iteration and progress within it.  Coincidentally
(almost) this is similar to the way IO requests are used by the Windows driver
stack.

A set of common fields in struct cl_io describe the IO and are shared by all
layers.  Important properties so described include:

- The IO type;

- A file (struct cl_object) against which this IO is executed;

- A position in a file where the read or write is taking place, and a count of
  bytes remaining to be processed (for CIT_READ and CIT_WRITE);

- A size to which file is being truncated or expanded (for CIT_SETATTR);

- A list of locks acquired for this IO;

Each layer keeps IO state in its `IO slice', described below, with all slices
chained to the list hanging off of struct cl_io:

- vvp_io is used by the top-most layer of the Linux kernel client.

  The most important state in vvp_io is an array of struct iovec, describing
  user space buffers from or to which IO is taking place. Note that other
  layers in the IO stack have no idea that data actually came from user space.

  vvp_io contains kernel specific fields, such as VM information describing a
  page fault, or the sendfile target.

- lov_io: IO state private for the LOV layer is kept here. The most important IO
  state at the LOV layer is an array of sub-IO's. Each sub-IO is a normal
  struct cl_io, representing a part of the IO process for a given iteration.
  With current sequential iterations, only one sub-IO is active at a time.

- osc_io: this slice stores IO state private to the OSC layer that exists within
  each sub-IO created by LOV.

=================
= 7. RPC Engine =
=================

7.1. Immediate vs. Opportunistic Transfers
==========================================

There are two possible modes of transfer initiation on the client:

- Immediate transfer: this is started when a high level IO wants a page or a
  collection of pages to be transferred right away. Examples: read-ahead,
  a synchronous read in the case of non-page aligned write, page write-out as
  part of an extent lock cancellation, page write-out as a part of memory
  cleansing. Immediate transfer can be both cl_req_type::CRT_READ and
  cl_req_type::CRT_WRITE;

- Opportunistic transfer (cl_req_type::CRT_WRITE only), that happens when IO
  wants to transfer a page to the server some time later, when it can be done
  efficiently. Example: pages dirtied by the write(2) path. Pages submitted for
  an opportunistic transfer are kept in a "staging area".

In any case, a transfer takes place in the form of a cl_req, which is a
representation for a network RPC.

Pages queued for an opportunistic transfer are placed into a staging area
(represented as a set of per-object and per-device queues at the OSC layer)
until it is decided that an efficient RPC can be composed of them. This
decision is made by "a req-formation engine", currently implemented as part of
the OSC layer. Req-formation depends on many factors: the size of the resulting
RPC, RPC alignment, whether or not multi-object RPCs are supported by the
server, max-RPC-in-flight limitations, size of the staging area, etc. CLIO uses
osc_extent to group pages for req-formation. osc_extent are further managed in
a per-object red-black tree for efficient RPC formatting.

For an immediate transfer the IO submits a cl_page_list which the req-formation
engine slices into cl_req's, possibly adding cached pages to some of the
resulting req's.

Whenever a page from cl_page_list is added to a newly constructed req, its
cl_page_operations::cpo_prep() layer methods are called. At that moment, the
page state is atomically changed from cl_page_state::CPS_OWNED to
cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner is
zeroed, and cl_page::cp_req is set to the req. cl_page_operations::cpo_prep()
method at a particular layer might return -EALREADY to indicate that it does
not need to submit this page at all. This is possible, for example, if a page
submitted for read became up-to-date in the meantime; and for write, if the
page don't have dirty bit set. See cl_io_submit_rw() for details.

Whenever a staged page is added to a newly constructed req, its
cl_page_operations::cpo_make_ready() layer methods are called. At that moment,
the page state is atomically changed from cl_page_state::CPS_CACHED to
cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to req. The
cl_page_operations::cpo_make_ready() method at a particular layer might return
-EAGAIN to indicate that this page is not currently eligible for the transfer.

The RPC engine guarantees that once the ->cpo_prep() or ->cpo_make_ready()
method has been called, the page completion routine (->cpo_completion() layer
method) will eventually be called (either as a result of successful page
transfer completion, or due to timeout).

To summarize, there are two main entry points into transfer sub-system:

- cl_io_submit_rw(): submits a list of pages for immediate transfer;

- cl_io_commit_async(): places a list of pages into staging area for future
  opportunistic transfer.

7.2. Page Lists
===============

To submit a group of pages for immediate transfer struct cl_2queue is used. It
contains two page lists: qin (input queue) and qout (output queue). Pages are
linked into these queues by cl_page::cp_batch list heads. Qin is populated with
the pages to be submitted to the transfer, and pages that were actually
submitted are placed onto qout. Not all pages from qin might end up on qout due
to

- ->cpo_prep() methods deciding that page should not be transferred, or

- unrecoverable submission error.

Pages not moved to qout remain on qin. It is up to the transfer submitter to
decide when to remove pages from qin and qout. Remaining pages on qin are
usually removed from this list right after (partially unsuccessful) transfer
submission. Pages are usually left on qout until transfer completion. This way
the caller can determine when all pages from the list were transferred.

The association between a page and an immediate transfer queue is protected by
cl_page::cl_mutex. This mutex is acquired when a cl_page is added in a
cl_page_list and released when a page is removed from the list.

When an RPC is formed, all of its constituent pages are linked together through
cl_page::cp_flight list hanging off of cl_req::crq_pages. Pages are removed
from this list just before the transfer completion method is invoked. No
special lock protects this list, as pages in transfer are under a VM lock.

7.3. Transfer States: Prepare, Completion
=========================================

The transfer (cl_req) state machine is trivial, and it is not explicitly coded.
A newly created transfer is in the "prepare" state while pages are collected.
When all pages are gathered, the transfer enters the "in-flight" state where it
remains until it reaches the "completion" state where page completion handlers
are invoked.

The per-layer ->cro_prep() transfer method is called when transfer preparation
is completed and transfer is about to enter the in-flight state. Similarly, the
per-layer ->cro_completion() method is called when the transfer completes
before per-page completion methods are called.

Additionally, before moving a transfer out of the prepare state, the RPC engine
calls the cl_req_attr_set() function.  This function invokes ->cro_attr_set()
methods on every layer to fill in RPC header that server uses to determine
where to get or put data. This replaces the old ->ap_{update,fill}_obdo()
methods.

Further, cl_req's are not reference counted and access to them is not
synchronized. This is because they are accessed only by the RPC engine in OSC
which fully controls RPC life-time, and it uses an internal OSC lock
(client_obd::cl_loi_list_lock spin-lock) for serialization.

7.4. Page Completion Handlers, Synchronous Transfer
===================================================

When a transfer completes, cl_req completion methods are called on every layer.
Then, for every transfer page, per-layer page completion methods
->cpo_completion() are invoked. The page is still under the VM lock at this
moment.  Completion methods are called bottom-to-top and it is responsibility
of the last of them (i.e., the completion method of the top-most layer---VVP)
to release the VM lock.

Both immediate and opportunistic transfers are asynchronous in the sense that
control can return to the caller before the transfer completes. CLIO doesn't
provide a synchronous transfer interface at all and it is up to a particular
caller to implement it if necessary. The simplest way to wait for the transfer
completion is wait on a page VM lock. This approach is used implicitly by the
Linux kernel. There is a case, though, where one wants to do transfer
completely synchronously without releasing the page VM lock: when
->prepare_write() method determines that a write goes from a non page-aligned
buffer into a not up-to-date page, a portion of a page has to be fetched from
the server. The VM page lock cannot be used to synchronize transfer completion
in this case, because it is used to mark the page as owned by IO. To handle
this, VVP attaches struct cl_sync_io to struct vvp_page. cl_sync_io contains a
number of pages still in IO and a synchronization primitive (struct completion)
which is signalled when transfer of the last page completes. The VVP page
completion handler (vvp_page_completion_common()) checks for attached
cl_sync_io and if it is there, decreases the number of in-flight pages and
signals completion when that number drops to 0. A similar mechanism is used for
direct-IO.

=============
= 8. lu_env =
=============

8.1. Motivation, Server Environment Usage
=========================================

lu_env and related data-types (struct lu_context and struct lu_context_key)
together implement a memory pre-allocation interface that Lustre uses to
decrease stack consumption without resorting to fully dynamic allocation.

Stack space is severely limited in the Linux kernel. Lustre traditionally
allocated a lot of automatic variables, resulting in spurious stack overflows
that are hard to trigger (they usually need a certain combination of driver
calls and interrupts to happen, making them extremely difficult to reproduce)
and debug (as stack overflow can easily result in corruption of thread-related
data-structures in the kernel memory, confusing the debugger).

The simplest way to handle this is to replace automatic variables with calls
to the generic memory allocator, but

- The generic allocator has scalability problems, and

- Additional code to free allocated memory is needed.

The lu_env interface was originally introduced in the MDS rewrite for Lustre
2.0 and matches server-side threading model very well. Roughly speaking,
lu_context represents a context in which computation is executed and
lu_context_key is a description of per-context data. In the simplest case
lu_context corresponds to a server thread; then lu_context_key is effectively a
thread-local storage (TLS). For a similar idea see the user-level pthreads
interface pthread_key_create().

More formally, lu_context_key defines a constructor-destructor pair and a tags
bit-mask. When lu_context is initialized (with a given tag bit-mask), a global
array of all registered lu_context_keys is scanned, constructors for all keys
with matching tags are invoked and their return values are stored in
lu_context.

Once lu_context has been initialized, a value of any key allocated for this
context can be retrieved very efficiently by indexing in the per-context
array. lu_context_key_get() function is used for this.

When context is finalized, destructors are called for all keys allocated in
this context.

The typical server usage is to have a lu_context for every server thread,
initialized when the thread is started. To reduce stack consumption by the
code running in this thread, a lu_context_key is registered that allocates in
its constructor a struct containing as fields values otherwise allocated on
the stack. See {mdt,osd,cmm,mdd}_thread_info for examples. Instead of doing

        int function(args) {
                /* structure "bar" in module "foo" */
                struct foo_bar bar;
                ...

the code roughly does

        struct foo_thread_info {
                struct foo_bar fti_bar;
                ...
        };

        int function(const struct lu_env *env, args) {
                struct foo_bar *bar;
                ...
                bar = &lu_context_key_get(&env->le_ctx, &foo_thread_key)->fti_

etc.

struct lu_env contains 2 contexts:

- le_ctx: this context is embedded in lu_env. By convention, this context is
  used _only_ to avoid allocations on the stack, and it should never be used to
  pass parameters between functions or layers. The reason for this restriction
  is that using contexts for implicit state sharing leads to a code that is
  difficult to understand and modify.

- le_ses: this is a pointer to a context shared by all threads handling given
  RPC. Context itself is embedded into struct ptlrpc_request. Currently a
  request is always processed by a single thread, but this might change in the
  future in a design where a small pool of threads processes RPCs
  asynchronously.

Additionally, state kept in env->le_ses context is shared by multiple layers.
For example, remote user credentials are stored there.

8.2. Client Environment Usage
=============================

On a client there is a lu_env associated with every thread executing Lustre
code. Again, it contains &env->le_ctx context used to reduce stack consumption.
env->le_ses is used to share state between all threads handling a given IO.
Again, currently an IO is processed by a single thread. env->le_ses is used to
efficiently allocate cl_io slices ({vvp,lov,osc}_io).

There are three important differences with lu_env usage on the server:

- While on the server there is a fixed pool of threads, any client thread can
  execute Lustre code. This makes it impractical to pre-allocate and
  pre-initialize lu_context for every thread. Instead, contexts are constructed
  on demand and after use returned into a global cache that amortizes creation
  cost;

- Client call-chains frequentyly cross Lustre-VFS and Lustre-VM boundaries.
  This means that just passing lu_env as a first parameter to every Lustre
  function and method is not enough. To work around this problem, a pointer to
  lu_env is stored in a field in the kernel data-structure associated with the
  current thread (task_struct::journal_info), from where it is recovered when
  Lustre code is re-entered from VFS or VM;

- Sometimes client code is re-entered in a fashion that precludes re-use of the
  higher level lu _env. For example, when a read or write incurs a page fault
  in the user space buffer memory-mapped from a Lustre file, page fault
  handling is a separate IO, independent of the already ongoing system call.
  The Lustre page fault handler allocates a new lu_env (by calling
  lu_env_get_nested()) in which the nested IO is going on. A similar situation
  occurs when client DLM lock LRU shrinking code is invoked in the context of a
  system call.

8.3. Sub-environments
=====================

As described above, lu_env (specifically, lu_env->le_ses) is used on a client
to allocate per-IO state, including foo_io data on every layer. This leads to a
complication at the LOV layer, which maintains multiple sub-IOs. As layers
below LOV allocate their IO slices in lu_env->le_ses, LOV has to allocate an
lu_env for every sub-IO and to carefully juggle them when invoking lower layer
methods. The case of a single IO is optimized by re-using the top-environment.

================
= 9. Use cases =
================

9.1. Inode Creation
===================

Lookup ends up calling ll_update_inode() to setup a new inode with a given
meta-data descriptor (obtained from the meta-data path). cl_inode_init() calls
cl_object_find() eventually calling lu_object_find_try() that either finds a
cl_object in the cache or allocates a new one, calling
lu_device_operations::ldo_object_{alloc,init}() methods on every layer top to
bottom. Every layer allocates its private data structure ({vvp,lov}_object) and
links it into an object header (cl_object_header) by calling lu_object_add().
At the VVP layer, vvp_object contains a pointer to the inode. The LOV layer
allocates a lov_object containing an array of pointers to sub-objects that are
found in the cache or allocated by calling cl_object_find (recursively). These
sub-objects have LOVSUB and OSC layer data.

A top-object and its sub-objects are inserted into a global FID-based hash
table and a global LRU list.

9.2. First IO to a File
=======================

After an object is instantiated as described in the previous use case, the
first IO call against this object has to create DLM locks. The following
operations re-use cached locks (see below).

A read call starts at ll_file_readv() which eventually calls
ll_file_io_generic(). This function calls cl_io_init() to initialize an IO
context, which calls the cl_object_operations::coo_io_init() method on every
layer.  As in the case of object instantiation, these methods allocate
layer-private IO state ({vvp,lov}_io) and add it to the list hanging off of the
IO context header cl_io by calling cl_io_add(). At the VVP layer, vvp_io_init()
handles special cases (like count == 0), updates statistic counters, and in the
case of write it takes a per-inode semaphore to avoid possible deadlock.

At the LOV layer, lov_io_init_raid0() allocates a struct lov_io and stores in
it the original IO parameters (starting offset and byte count). This is needed
because LOV is going to modify these parameters. Sub-IOs are not allocated at
this point---they are lazily instantiated later.

Once the top-IO has been initialized, ll_file_io_generic() enters the main IO
loop cl_io_loop() that drives IO iterations, going through

- cl_io_iter_init() calling cl_io_operations::cio_iter_init() top-to-bottom
- cl_io_lock() calling cl_io_operations::cio_lock() top-to-bottom
- cl_io_start() calling cl_io_operations::cio_start() top-to-bottom
- cl_io_end() calling cl_io_operations::cio_end() bottom-to-top
- cl_io_unlock() calling cl_io_operations::cio_unlock() bottom-to-top
- cl_io_iter_fini() calling cl_io_operations::cio_iter_fini() bottom-to-top
- cl_io_rw_advance() calling cl_io_operations::cio_advance() bottom-to-top

repeatedly until cl_io::ci_continue remains 0 after an iteration. These "IO
iterations" move an IO context through consecutive states (see enum
cl_io_state).  ->cio_iter_init() decides at each layer what part of the
remaining IO is to be done during current iteration. Currently,
lov_io_rw_iter_init() is the only non-trivial implementation of this method. It
does the following:

- Except for the cases of truncate and O_APPEND write, it shrinks the IO extent
  recorded in the top-IO (starting offset and bytes count) so that this extent
  is fully contained within a single stripe. This avoids "cascading evictions";

- It allocates sub-IOs for all stripes intersecting with the resulting IO range
  (which, in case of non-append write or read means creating single sub-io) by
  calling cl_io_init() that (as above) creates a cl_io context with lovsub_io
  and osc_io layers. The initialized cl_io is primed from the top-IO
  (lov_io_sub_inherit()) and cl_io_iter_init() is called against it;

- Finally all sub-ios for the current iteration are linked together into a
  lov_io::lis_active list.

Now we have a top-IO and its sub-IO in CIS_IT_STARTED state. cl_io_lock()
collects locks on all layers without actually enqueuing them: vvp_io_rw_lock()
requests a lock on the IO extent (possibly shrunk by LOV, see above) and
optionally on extents of Lustre files that happen to be memory-mapped onto the
user-level buffer used for this IO. In the future layers like SNS might request
additional locks, e.g., to protect parity blocks.

Locks requested by ->cio_lock() methods are added to the cl_lockset embedded
into top cl_io. The lockset contains 2 lock queues: "todo" and "done". Locks
are initially placed in the todo queue. Once locks from all layers have been
collected, they are sorted to avoid deadlocks (cl_io_locks_sort()) and then
enqueued by cl_lockset_lock(). The locks will be moved from todo list into
"done" list when they are granted.

At this stage we have top- and sub-IO in the CIS_LOCKED state with all needed
locks held. cl_io_start() moves cl_io into CIS_IO_GOING mode and calls
->cio_start() method. In the VVP layer this method invokes some version of
generic_file_{read,write}() function.

In the case of read, generic_file_read() calls for every non-up-to-date page
the a_ops->readpage() method that eventually (after obtaining cl_page
corresponding to the VM page supplied to it) calls ll_io_read_page() where it
decides if it's necessary to read ahead more pages by calling ll_readahead().
The number of pages to be read ahead is determined by the read pattern, also
it will factor in the requirements from different layers in CLIO stack, for
example, stripe alignment on the LOV layer and DLM lock coverage on the OSC
layer. The callback ->cio_read_ahead() is used to gather the requirements from
each layer. Please refer to lov_io_read_ahead() and osc_io_read_ahead() for
details.

ll_readahead() populates a queue by a target page and pages from read-ahead
window. The resulting queue is then submitted for immediate transfer by calling
cl_io_submit_rw() which ends up calling osc_io_submit_page() for every
not-up-to-date page in the queue.

->readpage() returns at this point, and the VM waits on a VM page lock, which
is released by the transfer completion handler before copying page data to the
user buffer.

In the case of write, generic_file_write() calls the a_ops->write_begin() and
a_ops->write_end() address space methods that end up calling ll_write_begin()
and ll_write_end() respectively. These functions follow the normal Linux
protocol for write, including a possible synchronous read of a non-overwritten
part of a page (ll_page_sync_io() call in ll_prepare_partial_page()). The pages
are placed into a list of vui_queue of vvp_io. In the normal case the pages will
be committed after all pages are handled by calling vvp_io_write_commit(). In
vvp_io_commit_write(), it calls cl_io_commit_async() to submit the dirty pages
into OSC writeback cache, where grant are allocated and the pages are added into
a red-black tree of osc_extent. In case there is no enough grant on the client
side, cl_io_commit_async will fail with -EDQUOT and the pages are transferred
immediately by calling ll_page_sync_io().

9.3. Lock-less and No-cache IO
==============================

IO context has a "locking mode" selected from MAYBE, NEVER or MANDATORY set
(enum cl_io_lock_dmd), that specifies what degree of distributed cache
coherency is assumed by this IO. MANDATORY mode requires all caches accessed by
this IO to be protected by distributed locks. In NEVER mode no distributed
coherency is needed at the expense of not caching the data. This mode is
required for the cases where client can not or will not participate in the
cache coherency protocol (e.g., a liblustre client that cannot respond to the
lock blocking call-backs while in the compute phase). In MAYBE mode some of the
caches involved in this IO are used and are globally coherent, and some other
caches are bypassed.

O_APPEND writes and truncates are always executed in MANDATORY mode. All other
calls are executed in NEVER mode by liblustre (see below) and in MAYBE mode by
a normal Linux client.

In MAYBE mode every OSC individually decides whether to use DLM. An OST might
return -EUSERS to an enqueue RPC indicating that the stripe in question is
contended and that the client should switch to the lockless IO mode. If this
happens, OSC, instead of using ldlm_lock, creates a special "lockless OSC lock"
that is not backed up by a DLM lock. This lock conflicts with any other lock in
its range and self-cancels when its last user is removed. As a result, when IO
proceeds to the stripe that is in lockless mode, all conflicting extent locks
are cancelled, purging the cache. When IO against this stripe ends, the lock is
cancelled, sending dirty pages (just placed in the cache by IO) back to the
server and invalidating the cache again. "Lockless locks" allow lockless and
no-cache IO mode to be implemented by the same code paths as cached IO.

* * * END * * *