Merge branch 'silvio' into docs-next

This commit is contained in:
Jonathan Corbet 2016-11-30 17:46:23 -07:00
commit 2fc9254d0c
7 changed files with 930 additions and 914 deletions

View File

@ -1,574 +0,0 @@
========================================
GENERIC ASSOCIATIVE ARRAY IMPLEMENTATION
========================================
Contents:
- Overview.
- The public API.
- Edit script.
- Operations table.
- Manipulation functions.
- Access functions.
- Index key form.
- Internal workings.
- Basic internal tree layout.
- Shortcuts.
- Splitting and collapsing nodes.
- Non-recursive iteration.
- Simultaneous alteration and iteration.
========
OVERVIEW
========
This associative array implementation is an object container with the following
properties:
(1) Objects are opaque pointers. The implementation does not care where they
point (if anywhere) or what they point to (if anything).
[!] NOTE: Pointers to objects _must_ be zero in the least significant bit.
(2) Objects do not need to contain linkage blocks for use by the array. This
permits an object to be located in multiple arrays simultaneously.
Rather, the array is made up of metadata blocks that point to objects.
(3) Objects require index keys to locate them within the array.
(4) Index keys must be unique. Inserting an object with the same key as one
already in the array will replace the old object.
(5) Index keys can be of any length and can be of different lengths.
(6) Index keys should encode the length early on, before any variation due to
length is seen.
(7) Index keys can include a hash to scatter objects throughout the array.
(8) The array can iterated over. The objects will not necessarily come out in
key order.
(9) The array can be iterated over whilst it is being modified, provided the
RCU readlock is being held by the iterator. Note, however, under these
circumstances, some objects may be seen more than once. If this is a
problem, the iterator should lock against modification. Objects will not
be missed, however, unless deleted.
(10) Objects in the array can be looked up by means of their index key.
(11) Objects can be looked up whilst the array is being modified, provided the
RCU readlock is being held by the thread doing the look up.
The implementation uses a tree of 16-pointer nodes internally that are indexed
on each level by nibbles from the index key in the same manner as in a radix
tree. To improve memory efficiency, shortcuts can be emplaced to skip over
what would otherwise be a series of single-occupancy nodes. Further, nodes
pack leaf object pointers into spare space in the node rather than making an
extra branch until as such time an object needs to be added to a full node.
==============
THE PUBLIC API
==============
The public API can be found in <linux/assoc_array.h>. The associative array is
rooted on the following structure:
struct assoc_array {
...
};
The code is selected by enabling CONFIG_ASSOCIATIVE_ARRAY.
EDIT SCRIPT
-----------
The insertion and deletion functions produce an 'edit script' that can later be
applied to effect the changes without risking ENOMEM. This retains the
preallocated metadata blocks that will be installed in the internal tree and
keeps track of the metadata blocks that will be removed from the tree when the
script is applied.
This is also used to keep track of dead blocks and dead objects after the
script has been applied so that they can be freed later. The freeing is done
after an RCU grace period has passed - thus allowing access functions to
proceed under the RCU read lock.
The script appears as outside of the API as a pointer of the type:
struct assoc_array_edit;
There are two functions for dealing with the script:
(1) Apply an edit script.
void assoc_array_apply_edit(struct assoc_array_edit *edit);
This will perform the edit functions, interpolating various write barriers
to permit accesses under the RCU read lock to continue. The edit script
will then be passed to call_rcu() to free it and any dead stuff it points
to.
(2) Cancel an edit script.
void assoc_array_cancel_edit(struct assoc_array_edit *edit);
This frees the edit script and all preallocated memory immediately. If
this was for insertion, the new object is _not_ released by this function,
but must rather be released by the caller.
These functions are guaranteed not to fail.
OPERATIONS TABLE
----------------
Various functions take a table of operations:
struct assoc_array_ops {
...
};
This points to a number of methods, all of which need to be provided:
(1) Get a chunk of index key from caller data:
unsigned long (*get_key_chunk)(const void *index_key, int level);
This should return a chunk of caller-supplied index key starting at the
*bit* position given by the level argument. The level argument will be a
multiple of ASSOC_ARRAY_KEY_CHUNK_SIZE and the function should return
ASSOC_ARRAY_KEY_CHUNK_SIZE bits. No error is possible.
(2) Get a chunk of an object's index key.
unsigned long (*get_object_key_chunk)(const void *object, int level);
As the previous function, but gets its data from an object in the array
rather than from a caller-supplied index key.
(3) See if this is the object we're looking for.
bool (*compare_object)(const void *object, const void *index_key);
Compare the object against an index key and return true if it matches and
false if it doesn't.
(4) Diff the index keys of two objects.
int (*diff_objects)(const void *object, const void *index_key);
Return the bit position at which the index key of the specified object
differs from the given index key or -1 if they are the same.
(5) Free an object.
void (*free_object)(void *object);
Free the specified object. Note that this may be called an RCU grace
period after assoc_array_apply_edit() was called, so synchronize_rcu() may
be necessary on module unloading.
MANIPULATION FUNCTIONS
----------------------
There are a number of functions for manipulating an associative array:
(1) Initialise an associative array.
void assoc_array_init(struct assoc_array *array);
This initialises the base structure for an associative array. It can't
fail.
(2) Insert/replace an object in an associative array.
struct assoc_array_edit *
assoc_array_insert(struct assoc_array *array,
const struct assoc_array_ops *ops,
const void *index_key,
void *object);
This inserts the given object into the array. Note that the least
significant bit of the pointer must be zero as it's used to type-mark
pointers internally.
If an object already exists for that key then it will be replaced with the
new object and the old one will be freed automatically.
The index_key argument should hold index key information and is
passed to the methods in the ops table when they are called.
This function makes no alteration to the array itself, but rather returns
an edit script that must be applied. -ENOMEM is returned in the case of
an out-of-memory error.
The caller should lock exclusively against other modifiers of the array.
(3) Delete an object from an associative array.
struct assoc_array_edit *
assoc_array_delete(struct assoc_array *array,
const struct assoc_array_ops *ops,
const void *index_key);
This deletes an object that matches the specified data from the array.
The index_key argument should hold index key information and is
passed to the methods in the ops table when they are called.
This function makes no alteration to the array itself, but rather returns
an edit script that must be applied. -ENOMEM is returned in the case of
an out-of-memory error. NULL will be returned if the specified object is
not found within the array.
The caller should lock exclusively against other modifiers of the array.
(4) Delete all objects from an associative array.
struct assoc_array_edit *
assoc_array_clear(struct assoc_array *array,
const struct assoc_array_ops *ops);
This deletes all the objects from an associative array and leaves it
completely empty.
This function makes no alteration to the array itself, but rather returns
an edit script that must be applied. -ENOMEM is returned in the case of
an out-of-memory error.
The caller should lock exclusively against other modifiers of the array.
(5) Destroy an associative array, deleting all objects.
void assoc_array_destroy(struct assoc_array *array,
const struct assoc_array_ops *ops);
This destroys the contents of the associative array and leaves it
completely empty. It is not permitted for another thread to be traversing
the array under the RCU read lock at the same time as this function is
destroying it as no RCU deferral is performed on memory release -
something that would require memory to be allocated.
The caller should lock exclusively against other modifiers and accessors
of the array.
(6) Garbage collect an associative array.
int assoc_array_gc(struct assoc_array *array,
const struct assoc_array_ops *ops,
bool (*iterator)(void *object, void *iterator_data),
void *iterator_data);
This iterates over the objects in an associative array and passes each one
to iterator(). If iterator() returns true, the object is kept. If it
returns false, the object will be freed. If the iterator() function
returns true, it must perform any appropriate refcount incrementing on the
object before returning.
The internal tree will be packed down if possible as part of the iteration
to reduce the number of nodes in it.
The iterator_data is passed directly to iterator() and is otherwise
ignored by the function.
The function will return 0 if successful and -ENOMEM if there wasn't
enough memory.
It is possible for other threads to iterate over or search the array under
the RCU read lock whilst this function is in progress. The caller should
lock exclusively against other modifiers of the array.
ACCESS FUNCTIONS
----------------
There are two functions for accessing an associative array:
(1) Iterate over all the objects in an associative array.
int assoc_array_iterate(const struct assoc_array *array,
int (*iterator)(const void *object,
void *iterator_data),
void *iterator_data);
This passes each object in the array to the iterator callback function.
iterator_data is private data for that function.
This may be used on an array at the same time as the array is being
modified, provided the RCU read lock is held. Under such circumstances,
it is possible for the iteration function to see some objects twice. If
this is a problem, then modification should be locked against. The
iteration algorithm should not, however, miss any objects.
The function will return 0 if no objects were in the array or else it will
return the result of the last iterator function called. Iteration stops
immediately if any call to the iteration function results in a non-zero
return.
(2) Find an object in an associative array.
void *assoc_array_find(const struct assoc_array *array,
const struct assoc_array_ops *ops,
const void *index_key);
This walks through the array's internal tree directly to the object
specified by the index key..
This may be used on an array at the same time as the array is being
modified, provided the RCU read lock is held.
The function will return the object if found (and set *_type to the object
type) or will return NULL if the object was not found.
INDEX KEY FORM
--------------
The index key can be of any form, but since the algorithms aren't told how long
the key is, it is strongly recommended that the index key includes its length
very early on before any variation due to the length would have an effect on
comparisons.
This will cause leaves with different length keys to scatter away from each
other - and those with the same length keys to cluster together.
It is also recommended that the index key begin with a hash of the rest of the
key to maximise scattering throughout keyspace.
The better the scattering, the wider and lower the internal tree will be.
Poor scattering isn't too much of a problem as there are shortcuts and nodes
can contain mixtures of leaves and metadata pointers.
The index key is read in chunks of machine word. Each chunk is subdivided into
one nibble (4 bits) per level, so on a 32-bit CPU this is good for 8 levels and
on a 64-bit CPU, 16 levels. Unless the scattering is really poor, it is
unlikely that more than one word of any particular index key will have to be
used.
=================
INTERNAL WORKINGS
=================
The associative array data structure has an internal tree. This tree is
constructed of two types of metadata blocks: nodes and shortcuts.
A node is an array of slots. Each slot can contain one of four things:
(*) A NULL pointer, indicating that the slot is empty.
(*) A pointer to an object (a leaf).
(*) A pointer to a node at the next level.
(*) A pointer to a shortcut.
BASIC INTERNAL TREE LAYOUT
--------------------------
Ignoring shortcuts for the moment, the nodes form a multilevel tree. The index
key space is strictly subdivided by the nodes in the tree and nodes occur on
fixed levels. For example:
Level: 0 1 2 3
=============== =============== =============== ===============
NODE D
NODE B NODE C +------>+---+
+------>+---+ +------>+---+ | | 0 |
NODE A | | 0 | | | 0 | | +---+
+---+ | +---+ | +---+ | : :
| 0 | | : : | : : | +---+
+---+ | +---+ | +---+ | | f |
| 1 |---+ | 3 |---+ | 7 |---+ +---+
+---+ +---+ +---+
: : : : | 8 |---+
+---+ +---+ +---+ | NODE E
| e |---+ | f | : : +------>+---+
+---+ | +---+ +---+ | 0 |
| f | | | f | +---+
+---+ | +---+ : :
| NODE F +---+
+------>+---+ | f |
| 0 | NODE G +---+
+---+ +------>+---+
: : | | 0 |
+---+ | +---+
| 6 |---+ : :
+---+ +---+
: : | f |
+---+ +---+
| f |
+---+
In the above example, there are 7 nodes (A-G), each with 16 slots (0-f).
Assuming no other meta data nodes in the tree, the key space is divided thusly:
KEY PREFIX NODE
========== ====
137* D
138* E
13[0-69-f]* C
1[0-24-f]* B
e6* G
e[0-57-f]* F
[02-df]* A
So, for instance, keys with the following example index keys will be found in
the appropriate nodes:
INDEX KEY PREFIX NODE
=============== ======= ====
13694892892489 13 C
13795289025897 137 D
13889dde88793 138 E
138bbb89003093 138 E
1394879524789 12 C
1458952489 1 B
9431809de993ba - A
b4542910809cd - A
e5284310def98 e F
e68428974237 e6 G
e7fffcbd443 e F
f3842239082 - A
To save memory, if a node can hold all the leaves in its portion of keyspace,
then the node will have all those leaves in it and will not have any metadata
pointers - even if some of those leaves would like to be in the same slot.
A node can contain a heterogeneous mix of leaves and metadata pointers.
Metadata pointers must be in the slots that match their subdivisions of key
space. The leaves can be in any slot not occupied by a metadata pointer. It
is guaranteed that none of the leaves in a node will match a slot occupied by a
metadata pointer. If the metadata pointer is there, any leaf whose key matches
the metadata key prefix must be in the subtree that the metadata pointer points
to.
In the above example list of index keys, node A will contain:
SLOT CONTENT INDEX KEY (PREFIX)
==== =============== ==================
1 PTR TO NODE B 1*
any LEAF 9431809de993ba
any LEAF b4542910809cd
e PTR TO NODE F e*
any LEAF f3842239082
and node B:
3 PTR TO NODE C 13*
any LEAF 1458952489
SHORTCUTS
---------
Shortcuts are metadata records that jump over a piece of keyspace. A shortcut
is a replacement for a series of single-occupancy nodes ascending through the
levels. Shortcuts exist to save memory and to speed up traversal.
It is possible for the root of the tree to be a shortcut - say, for example,
the tree contains at least 17 nodes all with key prefix '1111'. The insertion
algorithm will insert a shortcut to skip over the '1111' keyspace in a single
bound and get to the fourth level where these actually become different.
SPLITTING AND COLLAPSING NODES
------------------------------
Each node has a maximum capacity of 16 leaves and metadata pointers. If the
insertion algorithm finds that it is trying to insert a 17th object into a
node, that node will be split such that at least two leaves that have a common
key segment at that level end up in a separate node rooted on that slot for
that common key segment.
If the leaves in a full node and the leaf that is being inserted are
sufficiently similar, then a shortcut will be inserted into the tree.
When the number of objects in the subtree rooted at a node falls to 16 or
fewer, then the subtree will be collapsed down to a single node - and this will
ripple towards the root if possible.
NON-RECURSIVE ITERATION
-----------------------
Each node and shortcut contains a back pointer to its parent and the number of
slot in that parent that points to it. None-recursive iteration uses these to
proceed rootwards through the tree, going to the parent node, slot N + 1 to
make sure progress is made without the need for a stack.
The backpointers, however, make simultaneous alteration and iteration tricky.
SIMULTANEOUS ALTERATION AND ITERATION
-------------------------------------
There are a number of cases to consider:
(1) Simple insert/replace. This involves simply replacing a NULL or old
matching leaf pointer with the pointer to the new leaf after a barrier.
The metadata blocks don't change otherwise. An old leaf won't be freed
until after the RCU grace period.
(2) Simple delete. This involves just clearing an old matching leaf. The
metadata blocks don't change otherwise. The old leaf won't be freed until
after the RCU grace period.
(3) Insertion replacing part of a subtree that we haven't yet entered. This
may involve replacement of part of that subtree - but that won't affect
the iteration as we won't have reached the pointer to it yet and the
ancestry blocks are not replaced (the layout of those does not change).
(4) Insertion replacing nodes that we're actively processing. This isn't a
problem as we've passed the anchoring pointer and won't switch onto the
new layout until we follow the back pointers - at which point we've
already examined the leaves in the replaced node (we iterate over all the
leaves in a node before following any of its metadata pointers).
We might, however, re-see some leaves that have been split out into a new
branch that's in a slot further along than we were at.
(5) Insertion replacing nodes that we're processing a dependent branch of.
This won't affect us until we follow the back pointers. Similar to (4).
(6) Deletion collapsing a branch under us. This doesn't affect us because the
back pointers will get us back to the parent of the new node before we
could see the new node. The entire collapsed subtree is thrown away
unchanged - and will still be rooted on the same slot, so we shouldn't
process it a second time as we'll go back to slot + 1.
Note:
(*) Under some circumstances, we need to simultaneously change the parent
pointer and the parent slot pointer on a node (say, for example, we
inserted another node before it and moved it up a level). We cannot do
this without locking against a read - so we have to replace that node too.
However, when we're changing a shortcut into a node this isn't a problem
as shortcuts only have one slot and so the parent slot number isn't used
when traversing backwards over one. This means that it's okay to change
the slot number first - provided suitable barriers are used to make sure
the parent slot number is read after the back pointer.
Obsolete blocks and leaves are freed up after an RCU grace period has passed,
so as long as anyone doing walking or iteration holds the RCU read lock, the
old superstructure should not go away on them.

View File

@ -0,0 +1,551 @@
========================================
Generic Associative Array Implementation
========================================
Overview
========
This associative array implementation is an object container with the following
properties:
1. Objects are opaque pointers. The implementation does not care where they
point (if anywhere) or what they point to (if anything).
.. note:: Pointers to objects _must_ be zero in the least significant bit.**
2. Objects do not need to contain linkage blocks for use by the array. This
permits an object to be located in multiple arrays simultaneously.
Rather, the array is made up of metadata blocks that point to objects.
3. Objects require index keys to locate them within the array.
4. Index keys must be unique. Inserting an object with the same key as one
already in the array will replace the old object.
5. Index keys can be of any length and can be of different lengths.
6. Index keys should encode the length early on, before any variation due to
length is seen.
7. Index keys can include a hash to scatter objects throughout the array.
8. The array can iterated over. The objects will not necessarily come out in
key order.
9. The array can be iterated over whilst it is being modified, provided the
RCU readlock is being held by the iterator. Note, however, under these
circumstances, some objects may be seen more than once. If this is a
problem, the iterator should lock against modification. Objects will not
be missed, however, unless deleted.
10. Objects in the array can be looked up by means of their index key.
11. Objects can be looked up whilst the array is being modified, provided the
RCU readlock is being held by the thread doing the look up.
The implementation uses a tree of 16-pointer nodes internally that are indexed
on each level by nibbles from the index key in the same manner as in a radix
tree. To improve memory efficiency, shortcuts can be emplaced to skip over
what would otherwise be a series of single-occupancy nodes. Further, nodes
pack leaf object pointers into spare space in the node rather than making an
extra branch until as such time an object needs to be added to a full node.
The Public API
==============
The public API can be found in ``<linux/assoc_array.h>``. The associative
array is rooted on the following structure::
struct assoc_array {
...
};
The code is selected by enabling ``CONFIG_ASSOCIATIVE_ARRAY`` with::
./script/config -e ASSOCIATIVE_ARRAY
Edit Script
-----------
The insertion and deletion functions produce an 'edit script' that can later be
applied to effect the changes without risking ``ENOMEM``. This retains the
preallocated metadata blocks that will be installed in the internal tree and
keeps track of the metadata blocks that will be removed from the tree when the
script is applied.
This is also used to keep track of dead blocks and dead objects after the
script has been applied so that they can be freed later. The freeing is done
after an RCU grace period has passed - thus allowing access functions to
proceed under the RCU read lock.
The script appears as outside of the API as a pointer of the type::
struct assoc_array_edit;
There are two functions for dealing with the script:
1. Apply an edit script::
void assoc_array_apply_edit(struct assoc_array_edit *edit);
This will perform the edit functions, interpolating various write barriers
to permit accesses under the RCU read lock to continue. The edit script
will then be passed to ``call_rcu()`` to free it and any dead stuff it points
to.
2. Cancel an edit script::
void assoc_array_cancel_edit(struct assoc_array_edit *edit);
This frees the edit script and all preallocated memory immediately. If
this was for insertion, the new object is _not_ released by this function,
but must rather be released by the caller.
These functions are guaranteed not to fail.
Operations Table
----------------
Various functions take a table of operations::
struct assoc_array_ops {
...
};
This points to a number of methods, all of which need to be provided:
1. Get a chunk of index key from caller data::
unsigned long (*get_key_chunk)(const void *index_key, int level);
This should return a chunk of caller-supplied index key starting at the
*bit* position given by the level argument. The level argument will be a
multiple of ``ASSOC_ARRAY_KEY_CHUNK_SIZE`` and the function should return
``ASSOC_ARRAY_KEY_CHUNK_SIZE bits``. No error is possible.
2. Get a chunk of an object's index key::
unsigned long (*get_object_key_chunk)(const void *object, int level);
As the previous function, but gets its data from an object in the array
rather than from a caller-supplied index key.
3. See if this is the object we're looking for::
bool (*compare_object)(const void *object, const void *index_key);
Compare the object against an index key and return ``true`` if it matches and
``false`` if it doesn't.
4. Diff the index keys of two objects::
int (*diff_objects)(const void *object, const void *index_key);
Return the bit position at which the index key of the specified object
differs from the given index key or -1 if they are the same.
5. Free an object::
void (*free_object)(void *object);
Free the specified object. Note that this may be called an RCU grace period
after ``assoc_array_apply_edit()`` was called, so ``synchronize_rcu()`` may be
necessary on module unloading.
Manipulation Functions
----------------------
There are a number of functions for manipulating an associative array:
1. Initialise an associative array::
void assoc_array_init(struct assoc_array *array);
This initialises the base structure for an associative array. It can't fail.
2. Insert/replace an object in an associative array::
struct assoc_array_edit *
assoc_array_insert(struct assoc_array *array,
const struct assoc_array_ops *ops,
const void *index_key,
void *object);
This inserts the given object into the array. Note that the least
significant bit of the pointer must be zero as it's used to type-mark
pointers internally.
If an object already exists for that key then it will be replaced with the
new object and the old one will be freed automatically.
The ``index_key`` argument should hold index key information and is
passed to the methods in the ops table when they are called.
This function makes no alteration to the array itself, but rather returns
an edit script that must be applied. ``-ENOMEM`` is returned in the case of
an out-of-memory error.
The caller should lock exclusively against other modifiers of the array.
3. Delete an object from an associative array::
struct assoc_array_edit *
assoc_array_delete(struct assoc_array *array,
const struct assoc_array_ops *ops,
const void *index_key);
This deletes an object that matches the specified data from the array.
The ``index_key`` argument should hold index key information and is
passed to the methods in the ops table when they are called.
This function makes no alteration to the array itself, but rather returns
an edit script that must be applied. ``-ENOMEM`` is returned in the case of
an out-of-memory error. ``NULL`` will be returned if the specified object is
not found within the array.
The caller should lock exclusively against other modifiers of the array.
4. Delete all objects from an associative array::
struct assoc_array_edit *
assoc_array_clear(struct assoc_array *array,
const struct assoc_array_ops *ops);
This deletes all the objects from an associative array and leaves it
completely empty.
This function makes no alteration to the array itself, but rather returns
an edit script that must be applied. ``-ENOMEM`` is returned in the case of
an out-of-memory error.
The caller should lock exclusively against other modifiers of the array.
5. Destroy an associative array, deleting all objects::
void assoc_array_destroy(struct assoc_array *array,
const struct assoc_array_ops *ops);
This destroys the contents of the associative array and leaves it
completely empty. It is not permitted for another thread to be traversing
the array under the RCU read lock at the same time as this function is
destroying it as no RCU deferral is performed on memory release -
something that would require memory to be allocated.
The caller should lock exclusively against other modifiers and accessors
of the array.
6. Garbage collect an associative array::
int assoc_array_gc(struct assoc_array *array,
const struct assoc_array_ops *ops,
bool (*iterator)(void *object, void *iterator_data),
void *iterator_data);
This iterates over the objects in an associative array and passes each one to
``iterator()``. If ``iterator()`` returns ``true``, the object is kept. If it
returns ``false``, the object will be freed. If the ``iterator()`` function
returns ``true``, it must perform any appropriate refcount incrementing on the
object before returning.
The internal tree will be packed down if possible as part of the iteration
to reduce the number of nodes in it.
The ``iterator_data`` is passed directly to ``iterator()`` and is otherwise
ignored by the function.
The function will return ``0`` if successful and ``-ENOMEM`` if there wasn't
enough memory.
It is possible for other threads to iterate over or search the array under
the RCU read lock whilst this function is in progress. The caller should
lock exclusively against other modifiers of the array.
Access Functions
----------------
There are two functions for accessing an associative array:
1. Iterate over all the objects in an associative array::
int assoc_array_iterate(const struct assoc_array *array,
int (*iterator)(const void *object,
void *iterator_data),
void *iterator_data);
This passes each object in the array to the iterator callback function.
``iterator_data`` is private data for that function.
This may be used on an array at the same time as the array is being
modified, provided the RCU read lock is held. Under such circumstances,
it is possible for the iteration function to see some objects twice. If
this is a problem, then modification should be locked against. The
iteration algorithm should not, however, miss any objects.
The function will return ``0`` if no objects were in the array or else it will
return the result of the last iterator function called. Iteration stops
immediately if any call to the iteration function results in a non-zero
return.
2. Find an object in an associative array::
void *assoc_array_find(const struct assoc_array *array,
const struct assoc_array_ops *ops,
const void *index_key);
This walks through the array's internal tree directly to the object
specified by the index key..
This may be used on an array at the same time as the array is being
modified, provided the RCU read lock is held.
The function will return the object if found (and set ``*_type`` to the object
type) or will return ``NULL`` if the object was not found.
Index Key Form
--------------
The index key can be of any form, but since the algorithms aren't told how long
the key is, it is strongly recommended that the index key includes its length
very early on before any variation due to the length would have an effect on
comparisons.
This will cause leaves with different length keys to scatter away from each
other - and those with the same length keys to cluster together.
It is also recommended that the index key begin with a hash of the rest of the
key to maximise scattering throughout keyspace.
The better the scattering, the wider and lower the internal tree will be.
Poor scattering isn't too much of a problem as there are shortcuts and nodes
can contain mixtures of leaves and metadata pointers.
The index key is read in chunks of machine word. Each chunk is subdivided into
one nibble (4 bits) per level, so on a 32-bit CPU this is good for 8 levels and
on a 64-bit CPU, 16 levels. Unless the scattering is really poor, it is
unlikely that more than one word of any particular index key will have to be
used.
Internal Workings
=================
The associative array data structure has an internal tree. This tree is
constructed of two types of metadata blocks: nodes and shortcuts.
A node is an array of slots. Each slot can contain one of four things:
* A NULL pointer, indicating that the slot is empty.
* A pointer to an object (a leaf).
* A pointer to a node at the next level.
* A pointer to a shortcut.
Basic Internal Tree Layout
--------------------------
Ignoring shortcuts for the moment, the nodes form a multilevel tree. The index
key space is strictly subdivided by the nodes in the tree and nodes occur on
fixed levels. For example::
Level: 0 1 2 3
=============== =============== =============== ===============
NODE D
NODE B NODE C +------>+---+
+------>+---+ +------>+---+ | | 0 |
NODE A | | 0 | | | 0 | | +---+
+---+ | +---+ | +---+ | : :
| 0 | | : : | : : | +---+
+---+ | +---+ | +---+ | | f |
| 1 |---+ | 3 |---+ | 7 |---+ +---+
+---+ +---+ +---+
: : : : | 8 |---+
+---+ +---+ +---+ | NODE E
| e |---+ | f | : : +------>+---+
+---+ | +---+ +---+ | 0 |
| f | | | f | +---+
+---+ | +---+ : :
| NODE F +---+
+------>+---+ | f |
| 0 | NODE G +---+
+---+ +------>+---+
: : | | 0 |
+---+ | +---+
| 6 |---+ : :
+---+ +---+
: : | f |
+---+ +---+
| f |
+---+
In the above example, there are 7 nodes (A-G), each with 16 slots (0-f).
Assuming no other meta data nodes in the tree, the key space is divided
thusly::
KEY PREFIX NODE
========== ====
137* D
138* E
13[0-69-f]* C
1[0-24-f]* B
e6* G
e[0-57-f]* F
[02-df]* A
So, for instance, keys with the following example index keys will be found in
the appropriate nodes::
INDEX KEY PREFIX NODE
=============== ======= ====
13694892892489 13 C
13795289025897 137 D
13889dde88793 138 E
138bbb89003093 138 E
1394879524789 12 C
1458952489 1 B
9431809de993ba - A
b4542910809cd - A
e5284310def98 e F
e68428974237 e6 G
e7fffcbd443 e F
f3842239082 - A
To save memory, if a node can hold all the leaves in its portion of keyspace,
then the node will have all those leaves in it and will not have any metadata
pointers - even if some of those leaves would like to be in the same slot.
A node can contain a heterogeneous mix of leaves and metadata pointers.
Metadata pointers must be in the slots that match their subdivisions of key
space. The leaves can be in any slot not occupied by a metadata pointer. It
is guaranteed that none of the leaves in a node will match a slot occupied by a
metadata pointer. If the metadata pointer is there, any leaf whose key matches
the metadata key prefix must be in the subtree that the metadata pointer points
to.
In the above example list of index keys, node A will contain::
SLOT CONTENT INDEX KEY (PREFIX)
==== =============== ==================
1 PTR TO NODE B 1*
any LEAF 9431809de993ba
any LEAF b4542910809cd
e PTR TO NODE F e*
any LEAF f3842239082
and node B::
3 PTR TO NODE C 13*
any LEAF 1458952489
Shortcuts
---------
Shortcuts are metadata records that jump over a piece of keyspace. A shortcut
is a replacement for a series of single-occupancy nodes ascending through the
levels. Shortcuts exist to save memory and to speed up traversal.
It is possible for the root of the tree to be a shortcut - say, for example,
the tree contains at least 17 nodes all with key prefix ``1111``. The
insertion algorithm will insert a shortcut to skip over the ``1111`` keyspace
in a single bound and get to the fourth level where these actually become
different.
Splitting And Collapsing Nodes
------------------------------
Each node has a maximum capacity of 16 leaves and metadata pointers. If the
insertion algorithm finds that it is trying to insert a 17th object into a
node, that node will be split such that at least two leaves that have a common
key segment at that level end up in a separate node rooted on that slot for
that common key segment.
If the leaves in a full node and the leaf that is being inserted are
sufficiently similar, then a shortcut will be inserted into the tree.
When the number of objects in the subtree rooted at a node falls to 16 or
fewer, then the subtree will be collapsed down to a single node - and this will
ripple towards the root if possible.
Non-Recursive Iteration
-----------------------
Each node and shortcut contains a back pointer to its parent and the number of
slot in that parent that points to it. None-recursive iteration uses these to
proceed rootwards through the tree, going to the parent node, slot N + 1 to
make sure progress is made without the need for a stack.
The backpointers, however, make simultaneous alteration and iteration tricky.
Simultaneous Alteration And Iteration
-------------------------------------
There are a number of cases to consider:
1. Simple insert/replace. This involves simply replacing a NULL or old
matching leaf pointer with the pointer to the new leaf after a barrier.
The metadata blocks don't change otherwise. An old leaf won't be freed
until after the RCU grace period.
2. Simple delete. This involves just clearing an old matching leaf. The
metadata blocks don't change otherwise. The old leaf won't be freed until
after the RCU grace period.
3. Insertion replacing part of a subtree that we haven't yet entered. This
may involve replacement of part of that subtree - but that won't affect
the iteration as we won't have reached the pointer to it yet and the
ancestry blocks are not replaced (the layout of those does not change).
4. Insertion replacing nodes that we're actively processing. This isn't a
problem as we've passed the anchoring pointer and won't switch onto the
new layout until we follow the back pointers - at which point we've
already examined the leaves in the replaced node (we iterate over all the
leaves in a node before following any of its metadata pointers).
We might, however, re-see some leaves that have been split out into a new
branch that's in a slot further along than we were at.
5. Insertion replacing nodes that we're processing a dependent branch of.
This won't affect us until we follow the back pointers. Similar to (4).
6. Deletion collapsing a branch under us. This doesn't affect us because the
back pointers will get us back to the parent of the new node before we
could see the new node. The entire collapsed subtree is thrown away
unchanged - and will still be rooted on the same slot, so we shouldn't
process it a second time as we'll go back to slot + 1.
.. note::
Under some circumstances, we need to simultaneously change the parent
pointer and the parent slot pointer on a node (say, for example, we
inserted another node before it and moved it up a level). We cannot do
this without locking against a read - so we have to replace that node too.
However, when we're changing a shortcut into a node this isn't a problem
as shortcuts only have one slot and so the parent slot number isn't used
when traversing backwards over one. This means that it's okay to change
the slot number first - provided suitable barriers are used to make sure
the parent slot number is read after the back pointer.
Obsolete blocks and leaves are freed up after an RCU grace period has passed,
so as long as anyone doing walking or iteration holds the RCU read lock, the
old superstructure should not go away on them.

View File

@ -1,36 +1,42 @@
Semantics and Behavior of Atomic and
Bitmask Operations
=======================================================
Semantics and Behavior of Atomic and Bitmask Operations
=======================================================
David S. Miller
:Author: David S. Miller
This document is intended to serve as a guide to Linux port
This document is intended to serve as a guide to Linux port
maintainers on how to implement atomic counter, bitops, and spinlock
interfaces properly.
The atomic_t type should be defined as a signed integer and
Atomic Type And Operations
==========================
The atomic_t type should be defined as a signed integer and
the atomic_long_t type as a signed long integer. Also, they should
be made opaque such that any kind of cast to a normal C integer type
will fail. Something like the following should suffice:
will fail. Something like the following should suffice::
typedef struct { int counter; } atomic_t;
typedef struct { long counter; } atomic_long_t;
Historically, counter has been declared volatile. This is now discouraged.
See Documentation/process/volatile-considered-harmful.rst for the complete rationale.
See :ref:`Documentation/process/volatile-considered-harmful.rst
<volatile_considered_harmful>` for the complete rationale.
local_t is very similar to atomic_t. If the counter is per CPU and only
updated by one CPU, local_t is probably more appropriate. Please see
Documentation/local_ops.txt for the semantics of local_t.
:ref:`Documentation/core-api/local_ops.rst <local_ops>` for the semantics of
local_t.
The first operations to implement for atomic_t's are the initializers and
plain reads.
plain reads. ::
#define ATOMIC_INIT(i) { (i) }
#define atomic_set(v, i) ((v)->counter = (i))
The first macro is used in definitions, such as:
The first macro is used in definitions, such as::
static atomic_t my_counter = ATOMIC_INIT(1);
static atomic_t my_counter = ATOMIC_INIT(1);
The initializer is atomic in that the return values of the atomic operations
are guaranteed to be correct reflecting the initialized value if the
@ -38,10 +44,10 @@ initializer is used before runtime. If the initializer is used at runtime, a
proper implicit or explicit read memory barrier is needed before reading the
value with atomic_read from another thread.
As with all of the atomic_ interfaces, replace the leading "atomic_"
with "atomic_long_" to operate on atomic_long_t.
As with all of the ``atomic_`` interfaces, replace the leading ``atomic_``
with ``atomic_long_`` to operate on atomic_long_t.
The second interface can be used at runtime, as in:
The second interface can be used at runtime, as in::
struct foo { atomic_t counter; };
...
@ -59,7 +65,7 @@ been set with this operation or set with another operation. A proper implicit
or explicit memory barrier is needed before the value set with the operation
is guaranteed to be readable with atomic_read from another thread.
Next, we have:
Next, we have::
#define atomic_read(v) ((v)->counter)
@ -73,20 +79,21 @@ initialization by any other thread is visible yet, so the user of the
interface must take care of that with a proper implicit or explicit memory
barrier.
*** WARNING: atomic_read() and atomic_set() DO NOT IMPLY BARRIERS! ***
.. warning::
Some architectures may choose to use the volatile keyword, barriers, or inline
assembly to guarantee some degree of immediacy for atomic_read() and
atomic_set(). This is not uniformly guaranteed, and may change in the future,
so all users of atomic_t should treat atomic_read() and atomic_set() as simple
C statements that may be reordered or optimized away entirely by the compiler
or processor, and explicitly invoke the appropriate compiler and/or memory
barrier for each use case. Failure to do so will result in code that may
suddenly break when used with different architectures or compiler
optimizations, or even changes in unrelated code which changes how the
compiler optimizes the section accessing atomic_t variables.
``atomic_read()`` and ``atomic_set()`` DO NOT IMPLY BARRIERS!
*** YOU HAVE BEEN WARNED! ***
Some architectures may choose to use the volatile keyword, barriers, or
inline assembly to guarantee some degree of immediacy for atomic_read()
and atomic_set(). This is not uniformly guaranteed, and may change in
the future, so all users of atomic_t should treat atomic_read() and
atomic_set() as simple C statements that may be reordered or optimized
away entirely by the compiler or processor, and explicitly invoke the
appropriate compiler and/or memory barrier for each use case. Failure
to do so will result in code that may suddenly break when used with
different architectures or compiler optimizations, or even changes in
unrelated code which changes how the compiler optimizes the section
accessing atomic_t variables.
Properly aligned pointers, longs, ints, and chars (and unsigned
equivalents) may be atomically loaded from and stored to in the same
@ -95,14 +102,14 @@ and WRITE_ONCE() macros should be used to prevent the compiler from using
optimizations that might otherwise optimize accesses out of existence on
the one hand, or that might create unsolicited accesses on the other.
For example consider the following code:
For example consider the following code::
while (a > 0)
do_something();
If the compiler can prove that do_something() does not store to the
variable a, then the compiler is within its rights transforming this to
the following:
the following::
tmp = a;
if (a > 0)
@ -110,14 +117,14 @@ the following:
do_something();
If you don't want the compiler to do this (and you probably don't), then
you should use something like the following:
you should use something like the following::
while (READ_ONCE(a) < 0)
do_something();
Alternatively, you could place a barrier() call in the loop.
For another example, consider the following code:
For another example, consider the following code::
tmp_a = a;
do_something_with(tmp_a);
@ -125,7 +132,7 @@ For another example, consider the following code:
If the compiler can prove that do_something_with() does not store to the
variable a, then the compiler is within its rights to manufacture an
additional load as follows:
additional load as follows::
tmp_a = a;
do_something_with(tmp_a);
@ -139,7 +146,7 @@ The compiler would be likely to manufacture this additional load if
do_something_with() was an inline function that made very heavy use
of registers: reloading from variable a could save a flush to the
stack and later reload. To prevent the compiler from attacking your
code in this manner, write the following:
code in this manner, write the following::
tmp_a = READ_ONCE(a);
do_something_with(tmp_a);
@ -147,7 +154,7 @@ code in this manner, write the following:
For a final example, consider the following code, assuming that the
variable a is set at boot time before the second CPU is brought online
and never changed later, so that memory barriers are not needed:
and never changed later, so that memory barriers are not needed::
if (a)
b = 9;
@ -155,7 +162,7 @@ and never changed later, so that memory barriers are not needed:
b = 42;
The compiler is within its rights to manufacture an additional store
by transforming the above code into the following:
by transforming the above code into the following::
b = 42;
if (a)
@ -163,7 +170,7 @@ by transforming the above code into the following:
This could come as a fatal surprise to other code running concurrently
that expected b to never have the value 42 if a was zero. To prevent
the compiler from doing this, write something like:
the compiler from doing this, write something like::
if (a)
WRITE_ONCE(b, 9);
@ -173,10 +180,12 @@ the compiler from doing this, write something like:
Don't even -think- about doing this without proper use of memory barriers,
locks, or atomic operations if variable a can change at runtime!
*** WARNING: READ_ONCE() OR WRITE_ONCE() DO NOT IMPLY A BARRIER! ***
.. warning::
``READ_ONCE()`` OR ``WRITE_ONCE()`` DO NOT IMPLY A BARRIER!
Now, we move onto the atomic operation interfaces typically implemented with
the help of assembly code.
the help of assembly code. ::
void atomic_add(int i, atomic_t *v);
void atomic_sub(int i, atomic_t *v);
@ -192,7 +201,7 @@ One very important aspect of these two routines is that they DO NOT
require any explicit memory barriers. They need only perform the
atomic_t counter update in an SMP safe manner.
Next, we have:
Next, we have::
int atomic_inc_return(atomic_t *v);
int atomic_dec_return(atomic_t *v);
@ -214,7 +223,7 @@ If the atomic instructions used in an implementation provide explicit
memory barrier semantics which satisfy the above requirements, that is
fine as well.
Let's move on:
Let's move on::
int atomic_add_return(int i, atomic_t *v);
int atomic_sub_return(int i, atomic_t *v);
@ -224,7 +233,7 @@ explicit counter adjustment is given instead of the implicit "1".
This means that like atomic_{inc,dec}_return(), the memory barrier
semantics are required.
Next:
Next::
int atomic_inc_and_test(atomic_t *v);
int atomic_dec_and_test(atomic_t *v);
@ -234,13 +243,13 @@ given atomic counter. They return a boolean indicating whether the
resulting counter value was zero or not.
Again, these primitives provide explicit memory barrier semantics around
the atomic operation.
the atomic operation::
int atomic_sub_and_test(int i, atomic_t *v);
This is identical to atomic_dec_and_test() except that an explicit
decrement is given instead of the implicit "1". This primitive must
provide explicit memory barrier semantics around the operation.
provide explicit memory barrier semantics around the operation::
int atomic_add_negative(int i, atomic_t *v);
@ -249,7 +258,7 @@ is return which indicates whether the resulting counter value is negative.
This primitive must provide explicit memory barrier semantics around
the operation.
Then:
Then::
int atomic_xchg(atomic_t *v, int new);
@ -257,14 +266,14 @@ This performs an atomic exchange operation on the atomic variable v, setting
the given new value. It returns the old value that the atomic variable v had
just before the operation.
atomic_xchg must provide explicit memory barriers around the operation.
atomic_xchg must provide explicit memory barriers around the operation. ::
int atomic_cmpxchg(atomic_t *v, int old, int new);
This performs an atomic compare exchange operation on the atomic value v,
with the given old and new values. Like all atomic_xxx operations,
atomic_cmpxchg will only satisfy its atomicity semantics as long as all
other accesses of *v are performed through atomic_xxx operations.
other accesses of \*v are performed through atomic_xxx operations.
atomic_cmpxchg must provide explicit memory barriers around the operation,
although if the comparison fails then no memory ordering guarantees are
@ -273,7 +282,7 @@ required.
The semantics for atomic_cmpxchg are the same as those defined for 'cas'
below.
Finally:
Finally::
int atomic_add_unless(atomic_t *v, int a, int u);
@ -289,12 +298,12 @@ atomic_inc_not_zero, equivalent to atomic_add_unless(v, 1, 0)
If a caller requires memory barrier semantics around an atomic_t
operation which does not return a value, a set of interfaces are
defined which accomplish this:
defined which accomplish this::
void smp_mb__before_atomic(void);
void smp_mb__after_atomic(void);
For example, smp_mb__before_atomic() can be used like so:
For example, smp_mb__before_atomic() can be used like so::
obj->dead = 1;
smp_mb__before_atomic();
@ -315,67 +324,69 @@ atomic_t implementation above can have disastrous results. Here is
an example, which follows a pattern occurring frequently in the Linux
kernel. It is the use of atomic counters to implement reference
counting, and it works such that once the counter falls to zero it can
be guaranteed that no other entity can be accessing the object:
be guaranteed that no other entity can be accessing the object::
static void obj_list_add(struct obj *obj, struct list_head *head)
{
obj->active = 1;
list_add(&obj->list, head);
}
static void obj_list_add(struct obj *obj, struct list_head *head)
{
obj->active = 1;
list_add(&obj->list, head);
}
static void obj_list_del(struct obj *obj)
{
list_del(&obj->list);
obj->active = 0;
}
static void obj_list_del(struct obj *obj)
{
list_del(&obj->list);
obj->active = 0;
}
static void obj_destroy(struct obj *obj)
{
BUG_ON(obj->active);
kfree(obj);
}
static void obj_destroy(struct obj *obj)
{
BUG_ON(obj->active);
kfree(obj);
}
struct obj *obj_list_peek(struct list_head *head)
{
if (!list_empty(head)) {
struct obj *obj_list_peek(struct list_head *head)
{
if (!list_empty(head)) {
struct obj *obj;
obj = list_entry(head->next, struct obj, list);
atomic_inc(&obj->refcnt);
return obj;
}
return NULL;
}
void obj_poke(void)
{
struct obj *obj;
obj = list_entry(head->next, struct obj, list);
atomic_inc(&obj->refcnt);
return obj;
spin_lock(&global_list_lock);
obj = obj_list_peek(&global_list);
spin_unlock(&global_list_lock);
if (obj) {
obj->ops->poke(obj);
if (atomic_dec_and_test(&obj->refcnt))
obj_destroy(obj);
}
}
return NULL;
}
void obj_poke(void)
{
struct obj *obj;
void obj_timeout(struct obj *obj)
{
spin_lock(&global_list_lock);
obj_list_del(obj);
spin_unlock(&global_list_lock);
spin_lock(&global_list_lock);
obj = obj_list_peek(&global_list);
spin_unlock(&global_list_lock);
if (obj) {
obj->ops->poke(obj);
if (atomic_dec_and_test(&obj->refcnt))
obj_destroy(obj);
}
}
void obj_timeout(struct obj *obj)
{
spin_lock(&global_list_lock);
obj_list_del(obj);
spin_unlock(&global_list_lock);
.. note::
if (atomic_dec_and_test(&obj->refcnt))
obj_destroy(obj);
}
(This is a simplification of the ARP queue management in the
generic neighbour discover code of the networking. Olaf Kirch
found a bug wrt. memory barriers in kfree_skb() that exposed
the atomic_t memory barrier requirements quite clearly.)
This is a simplification of the ARP queue management in the generic
neighbour discover code of the networking. Olaf Kirch found a bug wrt.
memory barriers in kfree_skb() that exposed the atomic_t memory barrier
requirements quite clearly.
Given the above scheme, it must be the case that the obj->active
update done by the obj list deletion be visible to other processors
@ -383,7 +394,7 @@ before the atomic counter decrement is performed.
Otherwise, the counter could fall to zero, yet obj->active would still
be set, thus triggering the assertion in obj_destroy(). The error
sequence looks like this:
sequence looks like this::
cpu 0 cpu 1
obj_poke() obj_timeout()
@ -420,6 +431,10 @@ same scheme.
Another note is that the atomic_t operations returning values are
extremely slow on an old 386.
Atomic Bitmask
==============
We will now cover the atomic bitmask operations. You will find that
their SMP and memory barrier semantics are similar in shape and scope
to the atomic_t ops above.
@ -427,7 +442,7 @@ to the atomic_t ops above.
Native atomic bit operations are defined to operate on objects aligned
to the size of an "unsigned long" C data type, and are least of that
size. The endianness of the bits within each "unsigned long" are the
native endianness of the cpu.
native endianness of the cpu. ::
void set_bit(unsigned long nr, volatile unsigned long *addr);
void clear_bit(unsigned long nr, volatile unsigned long *addr);
@ -437,7 +452,7 @@ These routines set, clear, and change, respectively, the bit number
indicated by "nr" on the bit mask pointed to by "ADDR".
They must execute atomically, yet there are no implicit memory barrier
semantics required of these interfaces.
semantics required of these interfaces. ::
int test_and_set_bit(unsigned long nr, volatile unsigned long *addr);
int test_and_clear_bit(unsigned long nr, volatile unsigned long *addr);
@ -466,7 +481,7 @@ must provide explicit memory barrier semantics around their execution.
All memory operations before the atomic bit operation call must be
made visible globally before the atomic bit operation is made visible.
Likewise, the atomic bit operation must be visible globally before any
subsequent memory operation is made visible. For example:
subsequent memory operation is made visible. For example::
obj->dead = 1;
if (test_and_set_bit(0, &obj->flags))
@ -479,7 +494,7 @@ done by test_and_set_bit() becomes visible. Likewise, the atomic
memory operation done by test_and_set_bit() must become visible before
"obj->killed = 1;" is visible.
Finally there is the basic operation:
Finally there is the basic operation::
int test_bit(unsigned long nr, __const__ volatile unsigned long *addr);
@ -488,13 +503,13 @@ pointed to by "addr".
If explicit memory barriers are required around {set,clear}_bit() (which do
not return a value, and thus does not need to provide memory barrier
semantics), two interfaces are provided:
semantics), two interfaces are provided::
void smp_mb__before_atomic(void);
void smp_mb__after_atomic(void);
They are used as follows, and are akin to their atomic_t operation
brothers:
brothers::
/* All memory operations before this call will
* be globally visible before the clear_bit().
@ -511,7 +526,7 @@ There are two special bitops with lock barrier semantics (acquire/release,
same as spinlocks). These operate in the same way as their non-_lock/unlock
postfixed variants, except that they are to provide acquire/release semantics,
respectively. This means they can be used for bit_spin_trylock and
bit_spin_unlock type operations without specifying any more barriers.
bit_spin_unlock type operations without specifying any more barriers. ::
int test_and_set_bit_lock(unsigned long nr, unsigned long *addr);
void clear_bit_unlock(unsigned long nr, unsigned long *addr);
@ -526,7 +541,7 @@ provided. They are used in contexts where some other higher-level SMP
locking scheme is being used to protect the bitmask, and thus less
expensive non-atomic operations may be used in the implementation.
They have names similar to the above bitmask operation interfaces,
except that two underscores are prefixed to the interface name.
except that two underscores are prefixed to the interface name. ::
void __set_bit(unsigned long nr, volatile unsigned long *addr);
void __clear_bit(unsigned long nr, volatile unsigned long *addr);
@ -542,9 +557,11 @@ The routines xchg() and cmpxchg() must provide the same exact
memory-barrier semantics as the atomic and bit operations returning
values.
Note: If someone wants to use xchg(), cmpxchg() and their variants,
linux/atomic.h should be included rather than asm/cmpxchg.h, unless
the code is in arch/* and can take care of itself.
.. note::
If someone wants to use xchg(), cmpxchg() and their variants,
linux/atomic.h should be included rather than asm/cmpxchg.h, unless the
code is in arch/* and can take care of itself.
Spinlocks and rwlocks have memory barrier expectations as well.
The rule to follow is simple:
@ -558,7 +575,7 @@ The rule to follow is simple:
Which finally brings us to _atomic_dec_and_lock(). There is an
architecture-neutral version implemented in lib/dec_and_lock.c,
but most platforms will wish to optimize this in assembler.
but most platforms will wish to optimize this in assembler. ::
int _atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock);
@ -573,7 +590,7 @@ sure the spinlock operation is globally visible before any
subsequent memory operation.
We can demonstrate this operation more clearly if we define
an abstract atomic operation:
an abstract atomic operation::
long cas(long *mem, long old, long new);
@ -584,48 +601,48 @@ an abstract atomic operation:
3) Regardless, the current value at "mem" is returned.
As an example usage, here is what an atomic counter update
might look like:
might look like::
void example_atomic_inc(long *counter)
{
long old, new, ret;
void example_atomic_inc(long *counter)
{
long old, new, ret;
while (1) {
old = *counter;
new = old + 1;
while (1) {
old = *counter;
new = old + 1;
ret = cas(counter, old, new);
if (ret == old)
break;
}
}
Let's use cas() in order to build a pseudo-C atomic_dec_and_lock():
int _atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock)
{
long old, new, ret;
int went_to_zero;
went_to_zero = 0;
while (1) {
old = atomic_read(atomic);
new = old - 1;
if (new == 0) {
went_to_zero = 1;
spin_lock(lock);
}
ret = cas(atomic, old, new);
if (ret == old)
break;
if (went_to_zero) {
spin_unlock(lock);
went_to_zero = 0;
ret = cas(counter, old, new);
if (ret == old)
break;
}
}
return went_to_zero;
}
Let's use cas() in order to build a pseudo-C atomic_dec_and_lock()::
int _atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock)
{
long old, new, ret;
int went_to_zero;
went_to_zero = 0;
while (1) {
old = atomic_read(atomic);
new = old - 1;
if (new == 0) {
went_to_zero = 1;
spin_lock(lock);
}
ret = cas(atomic, old, new);
if (ret == old)
break;
if (went_to_zero) {
spin_unlock(lock);
went_to_zero = 0;
}
}
return went_to_zero;
}
Now, as far as memory barriers go, as long as spin_lock()
strictly orders all subsequent memory operations (including
@ -635,6 +652,7 @@ Said another way, _atomic_dec_and_lock() must guarantee that
a counter dropping to zero is never made visible before the
spinlock being acquired.
Note that this also means that for the case where the counter
is not dropping to zero, there are no memory ordering
requirements.
.. note::
Note that this also means that for the case where the counter is not
dropping to zero, there are no memory ordering requirements.

View File

@ -11,6 +11,9 @@ Core utilities
.. toctree::
:maxdepth: 1
assoc_array
atomic_ops
local_ops
workqueue
Interfaces for kernel debugging

View File

@ -0,0 +1,206 @@
.. _local_ops:
=================================================
Semantics and Behavior of Local Atomic Operations
=================================================
:Author: Mathieu Desnoyers
This document explains the purpose of the local atomic operations, how
to implement them for any given architecture and shows how they can be used
properly. It also stresses on the precautions that must be taken when reading
those local variables across CPUs when the order of memory writes matters.
.. note::
Note that ``local_t`` based operations are not recommended for general
kernel use. Please use the ``this_cpu`` operations instead unless there is
really a special purpose. Most uses of ``local_t`` in the kernel have been
replaced by ``this_cpu`` operations. ``this_cpu`` operations combine the
relocation with the ``local_t`` like semantics in a single instruction and
yield more compact and faster executing code.
Purpose of local atomic operations
==================================
Local atomic operations are meant to provide fast and highly reentrant per CPU
counters. They minimize the performance cost of standard atomic operations by
removing the LOCK prefix and memory barriers normally required to synchronize
across CPUs.
Having fast per CPU atomic counters is interesting in many cases: it does not
require disabling interrupts to protect from interrupt handlers and it permits
coherent counters in NMI handlers. It is especially useful for tracing purposes
and for various performance monitoring counters.
Local atomic operations only guarantee variable modification atomicity wrt the
CPU which owns the data. Therefore, care must taken to make sure that only one
CPU writes to the ``local_t`` data. This is done by using per cpu data and
making sure that we modify it from within a preemption safe context. It is
however permitted to read ``local_t`` data from any CPU: it will then appear to
be written out of order wrt other memory writes by the owner CPU.
Implementation for a given architecture
=======================================
It can be done by slightly modifying the standard atomic operations: only
their UP variant must be kept. It typically means removing LOCK prefix (on
i386 and x86_64) and any SMP synchronization barrier. If the architecture does
not have a different behavior between SMP and UP, including
``asm-generic/local.h`` in your architecture's ``local.h`` is sufficient.
The ``local_t`` type is defined as an opaque ``signed long`` by embedding an
``atomic_long_t`` inside a structure. This is made so a cast from this type to
a ``long`` fails. The definition looks like::
typedef struct { atomic_long_t a; } local_t;
Rules to follow when using local atomic operations
==================================================
* Variables touched by local ops must be per cpu variables.
* *Only* the CPU owner of these variables must write to them.
* This CPU can use local ops from any context (process, irq, softirq, nmi, ...)
to update its ``local_t`` variables.
* Preemption (or interrupts) must be disabled when using local ops in
process context to make sure the process won't be migrated to a
different CPU between getting the per-cpu variable and doing the
actual local op.
* When using local ops in interrupt context, no special care must be
taken on a mainline kernel, since they will run on the local CPU with
preemption already disabled. I suggest, however, to explicitly
disable preemption anyway to make sure it will still work correctly on
-rt kernels.
* Reading the local cpu variable will provide the current copy of the
variable.
* Reads of these variables can be done from any CPU, because updates to
"``long``", aligned, variables are always atomic. Since no memory
synchronization is done by the writer CPU, an outdated copy of the
variable can be read when reading some *other* cpu's variables.
How to use local atomic operations
==================================
::
#include <linux/percpu.h>
#include <asm/local.h>
static DEFINE_PER_CPU(local_t, counters) = LOCAL_INIT(0);
Counting
========
Counting is done on all the bits of a signed long.
In preemptible context, use ``get_cpu_var()`` and ``put_cpu_var()`` around
local atomic operations: it makes sure that preemption is disabled around write
access to the per cpu variable. For instance::
local_inc(&get_cpu_var(counters));
put_cpu_var(counters);
If you are already in a preemption-safe context, you can use
``this_cpu_ptr()`` instead::
local_inc(this_cpu_ptr(&counters));
Reading the counters
====================
Those local counters can be read from foreign CPUs to sum the count. Note that
the data seen by local_read across CPUs must be considered to be out of order
relatively to other memory writes happening on the CPU that owns the data::
long sum = 0;
for_each_online_cpu(cpu)
sum += local_read(&per_cpu(counters, cpu));
If you want to use a remote local_read to synchronize access to a resource
between CPUs, explicit ``smp_wmb()`` and ``smp_rmb()`` memory barriers must be used
respectively on the writer and the reader CPUs. It would be the case if you use
the ``local_t`` variable as a counter of bytes written in a buffer: there should
be a ``smp_wmb()`` between the buffer write and the counter increment and also a
``smp_rmb()`` between the counter read and the buffer read.
Here is a sample module which implements a basic per cpu counter using
``local.h``::
/* test-local.c
*
* Sample module for local.h usage.
*/
#include <asm/local.h>
#include <linux/module.h>
#include <linux/timer.h>
static DEFINE_PER_CPU(local_t, counters) = LOCAL_INIT(0);
static struct timer_list test_timer;
/* IPI called on each CPU. */
static void test_each(void *info)
{
/* Increment the counter from a non preemptible context */
printk("Increment on cpu %d\n", smp_processor_id());
local_inc(this_cpu_ptr(&counters));
/* This is what incrementing the variable would look like within a
* preemptible context (it disables preemption) :
*
* local_inc(&get_cpu_var(counters));
* put_cpu_var(counters);
*/
}
static void do_test_timer(unsigned long data)
{
int cpu;
/* Increment the counters */
on_each_cpu(test_each, NULL, 1);
/* Read all the counters */
printk("Counters read from CPU %d\n", smp_processor_id());
for_each_online_cpu(cpu) {
printk("Read : CPU %d, count %ld\n", cpu,
local_read(&per_cpu(counters, cpu)));
}
del_timer(&test_timer);
test_timer.expires = jiffies + 1000;
add_timer(&test_timer);
}
static int __init test_init(void)
{
/* initialize the timer that will increment the counter */
init_timer(&test_timer);
test_timer.function = do_test_timer;
test_timer.expires = jiffies + 1;
add_timer(&test_timer);
return 0;
}
static void __exit test_exit(void)
{
del_timer_sync(&test_timer);
}
module_init(test_init);
module_exit(test_exit);
MODULE_LICENSE("GPL");
MODULE_AUTHOR("Mathieu Desnoyers");
MODULE_DESCRIPTION("Local Atomic Ops");

View File

@ -1,191 +0,0 @@
Semantics and Behavior of Local Atomic Operations
Mathieu Desnoyers
This document explains the purpose of the local atomic operations, how
to implement them for any given architecture and shows how they can be used
properly. It also stresses on the precautions that must be taken when reading
those local variables across CPUs when the order of memory writes matters.
Note that local_t based operations are not recommended for general kernel use.
Please use the this_cpu operations instead unless there is really a special purpose.
Most uses of local_t in the kernel have been replaced by this_cpu operations.
this_cpu operations combine the relocation with the local_t like semantics in
a single instruction and yield more compact and faster executing code.
* Purpose of local atomic operations
Local atomic operations are meant to provide fast and highly reentrant per CPU
counters. They minimize the performance cost of standard atomic operations by
removing the LOCK prefix and memory barriers normally required to synchronize
across CPUs.
Having fast per CPU atomic counters is interesting in many cases : it does not
require disabling interrupts to protect from interrupt handlers and it permits
coherent counters in NMI handlers. It is especially useful for tracing purposes
and for various performance monitoring counters.
Local atomic operations only guarantee variable modification atomicity wrt the
CPU which owns the data. Therefore, care must taken to make sure that only one
CPU writes to the local_t data. This is done by using per cpu data and making
sure that we modify it from within a preemption safe context. It is however
permitted to read local_t data from any CPU : it will then appear to be written
out of order wrt other memory writes by the owner CPU.
* Implementation for a given architecture
It can be done by slightly modifying the standard atomic operations : only
their UP variant must be kept. It typically means removing LOCK prefix (on
i386 and x86_64) and any SMP synchronization barrier. If the architecture does
not have a different behavior between SMP and UP, including asm-generic/local.h
in your architecture's local.h is sufficient.
The local_t type is defined as an opaque signed long by embedding an
atomic_long_t inside a structure. This is made so a cast from this type to a
long fails. The definition looks like :
typedef struct { atomic_long_t a; } local_t;
* Rules to follow when using local atomic operations
- Variables touched by local ops must be per cpu variables.
- _Only_ the CPU owner of these variables must write to them.
- This CPU can use local ops from any context (process, irq, softirq, nmi, ...)
to update its local_t variables.
- Preemption (or interrupts) must be disabled when using local ops in
process context to make sure the process won't be migrated to a
different CPU between getting the per-cpu variable and doing the
actual local op.
- When using local ops in interrupt context, no special care must be
taken on a mainline kernel, since they will run on the local CPU with
preemption already disabled. I suggest, however, to explicitly
disable preemption anyway to make sure it will still work correctly on
-rt kernels.
- Reading the local cpu variable will provide the current copy of the
variable.
- Reads of these variables can be done from any CPU, because updates to
"long", aligned, variables are always atomic. Since no memory
synchronization is done by the writer CPU, an outdated copy of the
variable can be read when reading some _other_ cpu's variables.
* How to use local atomic operations
#include <linux/percpu.h>
#include <asm/local.h>
static DEFINE_PER_CPU(local_t, counters) = LOCAL_INIT(0);
* Counting
Counting is done on all the bits of a signed long.
In preemptible context, use get_cpu_var() and put_cpu_var() around local atomic
operations : it makes sure that preemption is disabled around write access to
the per cpu variable. For instance :
local_inc(&get_cpu_var(counters));
put_cpu_var(counters);
If you are already in a preemption-safe context, you can use
this_cpu_ptr() instead.
local_inc(this_cpu_ptr(&counters));
* Reading the counters
Those local counters can be read from foreign CPUs to sum the count. Note that
the data seen by local_read across CPUs must be considered to be out of order
relatively to other memory writes happening on the CPU that owns the data.
long sum = 0;
for_each_online_cpu(cpu)
sum += local_read(&per_cpu(counters, cpu));
If you want to use a remote local_read to synchronize access to a resource
between CPUs, explicit smp_wmb() and smp_rmb() memory barriers must be used
respectively on the writer and the reader CPUs. It would be the case if you use
the local_t variable as a counter of bytes written in a buffer : there should
be a smp_wmb() between the buffer write and the counter increment and also a
smp_rmb() between the counter read and the buffer read.
Here is a sample module which implements a basic per cpu counter using local.h.
--- BEGIN ---
/* test-local.c
*
* Sample module for local.h usage.
*/
#include <asm/local.h>
#include <linux/module.h>
#include <linux/timer.h>
static DEFINE_PER_CPU(local_t, counters) = LOCAL_INIT(0);
static struct timer_list test_timer;
/* IPI called on each CPU. */
static void test_each(void *info)
{
/* Increment the counter from a non preemptible context */
printk("Increment on cpu %d\n", smp_processor_id());
local_inc(this_cpu_ptr(&counters));
/* This is what incrementing the variable would look like within a
* preemptible context (it disables preemption) :
*
* local_inc(&get_cpu_var(counters));
* put_cpu_var(counters);
*/
}
static void do_test_timer(unsigned long data)
{
int cpu;
/* Increment the counters */
on_each_cpu(test_each, NULL, 1);
/* Read all the counters */
printk("Counters read from CPU %d\n", smp_processor_id());
for_each_online_cpu(cpu) {
printk("Read : CPU %d, count %ld\n", cpu,
local_read(&per_cpu(counters, cpu)));
}
del_timer(&test_timer);
test_timer.expires = jiffies + 1000;
add_timer(&test_timer);
}
static int __init test_init(void)
{
/* initialize the timer that will increment the counter */
init_timer(&test_timer);
test_timer.function = do_test_timer;
test_timer.expires = jiffies + 1;
add_timer(&test_timer);
return 0;
}
static void __exit test_exit(void)
{
del_timer_sync(&test_timer);
}
module_init(test_init);
module_exit(test_exit);
MODULE_LICENSE("GPL");
MODULE_AUTHOR("Mathieu Desnoyers");
MODULE_DESCRIPTION("Local Atomic Ops");
--- END ---

View File

@ -1,3 +1,6 @@
.. _volatile_considered_harmful:
Why the "volatile" type class should not be used
------------------------------------------------