Type-safe containers

Note

This section previously used the term list; it was changed to container to be more clear.

Common container interface

FRR includes a set of container implementations with abstracted common APIs. The purpose of this is easily allow swapping out one data structure for another while also making the code easier to read and write. There is one API for unsorted containers and a similar but not identical API for sorted containers - and heaps use a middle ground of both.

For unsorted containers, the following implementations exist:

  • single-linked list with tail pointer (e.g. STAILQ in BSD)

  • double-linked list

  • atomic single-linked list with tail pointer

Being partially sorted, the oddball structure:

  • an 8-ary heap

For sorted containers, these data structures are implemented:

  • single-linked list

  • atomic single-linked list

  • skiplist

  • red-black tree (based on OpenBSD RB_TREE)

  • hash table (note below)

Except for hash tables, each of the sorted data structures has a variant with unique and non-unique items. Hash tables always require unique items and mostly follow the “sorted” API but use the hash value as sorting key. Also, iterating while modifying does not work with hash tables. Conversely, the heap always has non-unique items, but iterating while modifying doesn’t work either.

The following sorted structures are likely to be implemented at some point in the future:

  • atomic skiplist

  • atomic hash table (note below)

The APIs are all designed to be as type-safe as possible. This means that there will be a compiler warning when an item doesn’t match the container, or the return value has a different type, or other similar situations. You should never use casts with these APIs. If a cast is necessary in relation to these APIs, there is probably something wrong with the overall design.

Only the following pieces use dynamically allocated memory:

  • the hash table itself is dynamically grown and shrunk

  • skiplists store up to 4 next pointers inline but will dynamically allocate memory to hold an item’s 5th up to 16th next pointer (if they exist)

  • the heap uses a dynamically grown and shrunk array of items

Cheat sheet

Available types:

DECLARE_LIST
DECLARE_ATOMLIST
DECLARE_DLIST

DECLARE_HEAP

DECLARE_SORTLIST_UNIQ
DECLARE_SORTLIST_NONUNIQ
DECLARE_ATOMLIST_UNIQ
DECLARE_ATOMLIST_NONUNIQ
DECLARE_SKIPLIST_UNIQ
DECLARE_SKIPLIST_NONUNIQ
DECLARE_RBTREE_UNIQ
DECLARE_RBTREE_NONUNIQ

DECLARE_HASH

Functions provided:

Function

LIST

HEAP

HASH

*_UNIQ

*_NONUNIQ

_init, _fini

yes

yes

yes

yes

yes

_first, _next, _next_safe,

_const_first, _const_next

yes

yes

yes

yes

yes

_last, _prev, _prev_safe,

_const_last, _const_prev

DLIST only

RB only

RB only

_swap_all

yes

yes

yes

yes

yes

_anywhere

yes

_add_head, _add_tail, _add_after

yes

_add

yes

yes

yes

yes

_member

yes

yes

yes

yes

yes

_del, _pop

yes

yes

yes

yes

yes

_find, _const_find

yes

yes

_find_lt, _find_gteq,

_const_find_lt, _const_find_gteq

yes

yes

use with frr_each() macros

yes

yes

yes

yes

yes

Datastructure type setup

Each of the data structures has a PREDECL_* and a DECLARE_* macro to set up an “instantiation” of the container. This works somewhat similar to C++ templating, though much simpler.

In all following text, the Z prefix is replaced with a name chosen for the instance of the datastructure.

The common setup pattern will look like this:

#include <typesafe.h>

PREDECL_XXX(Z);
struct item {
    int otherdata;
    struct Z_item mylistitem;
}

struct Z_head mylisthead;

/* unsorted: */
DECLARE_XXX(Z, struct item, mylistitem);

/* sorted, items that compare as equal cannot be added to list */
int compare_func(const struct item *a, const struct item *b);
DECLARE_XXX_UNIQ(Z, struct item, mylistitem, compare_func);

/* sorted, items that compare as equal can be added to list */
int compare_func(const struct item *a, const struct item *b);
DECLARE_XXX_NONUNIQ(Z, struct item, mylistitem, compare_func);

/* hash tables: */
int compare_func(const struct item *a, const struct item *b);
uint32_t hash_func(const struct item *a);
DECLARE_XXX(Z, struct item, mylistitem, compare_func, hash_func);

XXX is replaced with the name of the data structure, e.g. SKIPLIST or ATOMLIST. The DECLARE_XXX invocation can either occur in a .h file (if the container needs to be accessed from several C files) or it can be placed in a .c file (if the container is only accessed from that file.) The PREDECL_XXX invocation defines the struct Z_item and struct Z_head types and must therefore occur before these are used.

To switch between compatible data structures, only these two lines need to be changes. To switch to a data structure with a different API, some source changes are necessary.

Common iteration macros

The following iteration macros work across all data structures:

frr_each(Z, head, item)

Equivalent to:

for (item = Z_first(&head); item; item = Z_next(&head, item))

Note that this will fail if the container is modified while being iterated over.

frr_each_safe(Z, head, item)

Same as the previous, but the next element is pre-loaded into a “hidden” variable (named Z_safe.) Equivalent to:

for (item = Z_first(&head); item; item = next) {
    next = Z_next_safe(&head, item);
    ...
}

Warning

Iterating over hash tables while adding or removing items is not possible. The iteration position will be corrupted when the hash tables is resized while iterating. This will cause items to be skipped or iterated over twice.

frr_each_from(Z, head, item, from)

Iterates over the container, starting at item from. This variant is “safe” as in the previous macro. Equivalent to:

for (item = from; item; item = from) {
    from = Z_next_safe(&head, item);
    ...
}

Note

The from variable is written to. This is intentional - you can resume iteration after breaking out of the loop by keeping the from value persistent and reusing it for the next loop.

frr_rev_each(Z, head, item)
frr_rev_each_safe(Z, head, item)
frr_rev_each_from(Z, head, item, from)

Reverse direction variants of the above. Only supported on containers that implement _last and _prev (i.e. RBTREE and DLIST).

To iterate over const pointers, add _const to the name of the datastructure (Z above), e.g. frr_each (mylist, head, item) becomes frr_each (mylist_const, head, item).

Common API

The following documentation assumes that a container has been defined using Z as the name, and itemtype being the type of the items (e.g. struct item.)

void Z_init(struct Z_head*)

Initializes the container for use. For most implementations, this just sets some values. Hash tables are the only implementation that allocates memory in this call.

void Z_fini(struct Z_head*)

Reverse the effects of Z_init(). The container must be empty when this function is called.

Warning

This function may assert() if the container is not empty.

size_t Z_count(const struct Z_head*)

Returns the number of items in a structure. All structures store a counter in their Z_head so that calling this function completes in O(1).

Note

For atomic containers with concurrent access, the value will already be outdated by the time this function returns and can therefore only be used as an estimate.

bool Z_member(const struct Z_head*, const itemtype*)

Determines whether some item is a member of the given container. The item must either be valid on some container, or set to all zeroes.

On some containers, if no faster way to determine membership is possible, this is simply item == Z_find(head, item).

Not currently available for atomic containers.

const itemtype *Z_const_first(const struct Z_head*)
itemtype *Z_first(struct Z_head*)

Returns the first item in the structure, or NULL if the structure is empty. This is O(1) for all data structures except red-black trees where it is O(log n).

const itemtype *Z_const_last(const struct Z_head*)
itemtype *Z_last(struct Z_head*)

Last item in the structure, or NULL. Only available on containers that support reverse iteration (i.e. RBTREE and DLIST).

itemtype *Z_pop(struct Z_head*)

Remove and return the first item in the structure, or NULL if the structure is empty. Like Z_first(), this is O(1) for all data structures except red-black trees where it is O(log n) again.

This function can be used to build queues (with unsorted structures) or priority queues (with sorted structures.)

Another common pattern is deleting all container items:

while ((item = Z_pop(head)))
    item_free(item);

Note

This function can - and should - be used with hash tables. It is not affected by the “modification while iterating” problem. To remove all items from a hash table, use the loop demonstrated above.

const itemtype *Z_const_next(const struct Z_head*, const itemtype *prev)
itemtype *Z_next(struct Z_head*, itemtype *prev)

Return the item that follows after prev, or NULL if prev is the last item.

Warning

prev must not be NULL! Use Z_next_safe() if prev might be NULL.

itemtype *Z_next_safe(struct Z_head*, itemtype *prev)

Same as Z_next(), except that NULL is returned if prev is NULL.

const itemtype *Z_const_prev(const struct Z_head*, const itemtype *next)
itemtype *Z_prev(struct Z_head*, itemtype *next)
itemtype *Z_prev_safe(struct Z_head*, itemtype *next)

As above, but preceding item. Only available on structures that support reverse iteration (i.e. RBTREE and DLIST).

itemtype *Z_del(struct Z_head*, itemtype *item)

Remove item from the container and return it.

Note

This function’s behaviour is undefined if item is not actually on the container. Some structures return NULL in this case while others return item. The function may also call assert() (but most don’t.)

itemtype *Z_swap_all(struct Z_head*, struct Z_head*)

Swap the contents of 2 containers (of identical type). This exchanges the contents of the two head structures and updates pointers if necessary for the particular data structure. Fast for all structures.

(Not currently available on atomic containers.)

Todo

Z_del_after() / Z_del_hint()?

API for unsorted structures

Since the insertion position is not pre-defined for unsorted data, there are several functions exposed to insert data:

Note

item must not be NULL for any of the following functions.

DECLARE_XXX(Z, type, field)
Parameters:
  • XXX (listtype) – LIST, DLIST or ATOMLIST to select a data structure implementation.

  • Z (token) – Gives the name prefix that is used for the functions created for this instantiation. DECLARE_XXX(foo, ...) gives struct foo_item, foo_add_head(), foo_count(), etc. Note that this must match the value given in PREDECL_XXX(foo).

  • type (typename) – Specifies the data type of the list items, e.g. struct item. Note that struct must be added here, it is not automatically added.

  • field (token) – References a struct member of type that must be typed as struct foo_item. This struct member is used to store “next” pointers or other data structure specific data.

void Z_add_head(struct Z_head*, itemtype *item)

Insert an item at the beginning of the structure, before the first item. This is an O(1) operation for non-atomic lists.

void Z_add_tail(struct Z_head*, itemtype *item)

Insert an item at the end of the structure, after the last item. This is also an O(1) operation for non-atomic lists.

void Z_add_after(struct Z_head*, itemtype *after, itemtype *item)

Insert item behind after. If after is NULL, the item is inserted at the beginning of the list as with Z_add_head(). This is also an O(1) operation for non-atomic lists.

A common pattern is to keep a “previous” pointer around while iterating:

itemtype *prev = NULL, *item;

frr_each_safe(Z, head, item) {
    if (something) {
        Z_add_after(head, prev, item);
        break;
    }
    prev = item;
}

Todo

maybe flip the order of item & after? Z_add_after(head, item, after)

bool Z_anywhere(const itemtype*)

Returns whether an item is a member of any container of this type. The item must either be valid on some container, or set to all zeroes.

Guaranteed to be fast (pointer compare or similar.)

Not currently available for sorted and atomic containers. Might be added for sorted containers at some point (when needed.)

API for sorted structures

Sorted data structures do not need to have an insertion position specified, therefore the insertion calls are different from unsorted containers. Also, sorted containers can be searched for a value.

DECLARE_XXX_UNIQ(Z, type, field, compare_func)
Parameters:
  • XXX (listtype) – One of the following: SORTLIST (single-linked sorted list), SKIPLIST (skiplist), RBTREE (RB-tree) or ATOMSORT (atomic single-linked list).

  • Z (token) – Gives the name prefix that is used for the functions created for this instantiation. DECLARE_XXX(foo, ...) gives struct foo_item, foo_add(), foo_count(), etc. Note that this must match the value given in PREDECL_XXX(foo).

  • type (typename) – Specifies the data type of the items, e.g. struct item. Note that struct must be added here, it is not automatically added.

  • field (token) – References a struct member of type that must be typed as struct foo_item. This struct member is used to store “next” pointers or other data structure specific data.

  • compare_func (funcptr) – Item comparison function, must have the following function signature: int function(const itemtype *, const itemtype*). This function may be static if the container is only used in one file.

DECLARE_XXX_NONUNIQ(Z, type, field, compare_func)

Same as above, but allow adding multiple items to the container that compare as equal in compare_func. Ordering between these items is undefined and depends on the container implementation.

itemtype *Z_add(struct Z_head*, itemtype *item)

Insert an item at the appropriate sorted position. If another item exists in the container that compares as equal (compare_func() == 0), item is not inserted and the already-existing item in the container is returned. Otherwise, on successful insertion, NULL is returned.

For _NONUNIQ containers, this function always returns NULL since item can always be successfully added to the container.

const itemtype *Z_const_find(const struct Z_head*, const itemtype *ref)
itemtype *Z_find(struct Z_head*, const itemtype *ref)

Search the container for an item that compares equal to ref. If no equal item is found, return NULL.

This function is likely used with a temporary stack-allocated value for ref like so:

itemtype searchfor = { .foo = 123 };

itemtype *item = Z_find(head, &searchfor);

Note

The Z_find() function is only available for containers that contain unique items (i.e. DECLARE_XXX_UNIQ.) This is because on a container with non-unique items, more than one item may compare as equal to the item that is searched for.

const itemtype *Z_const_find_gteq(const struct Z_head*, const itemtype *ref)
itemtype *Z_find_gteq(struct Z_head*, const itemtype *ref)

Search the container for an item that compares greater or equal to ref. See Z_find() above.

const itemtype *Z_const_find_lt(const struct Z_head*, const itemtype *ref)
itemtype *Z_find_lt(struct Z_head*, const itemtype *ref)

Search the container for an item that compares less than ref. See Z_find() above.

API for hash tables

DECLARE_HASH(Z, type, field, compare_func, hash_func)
Parameters:
  • HASH (listtype) – Only HASH is currently available.

  • Z (token) – Gives the name prefix that is used for the functions created for this instantiation. DECLARE_XXX(foo, ...) gives struct foo_item, foo_add(), foo_count(), etc. Note that this must match the value given in PREDECL_XXX(foo).

  • type (typename) – Specifies the data type of the items, e.g. struct item. Note that struct must be added here, it is not automatically added.

  • field (token) – References a struct member of type that must be typed as struct foo_item. This struct member is used to store “next” pointers or other data structure specific data.

  • compare_func (funcptr) – Item comparison function, must have the following function signature: int function(const itemtype *, const itemtype*). This function may be static if the container is only used in one file. For hash tables, this function is only used to check for equality, the ordering is ignored.

  • hash_func (funcptr) – Hash calculation function, must have the following function signature: uint32_t function(const itemtype *). The hash value for items stored in a hash table is cached in each item, so this value need not be cached by the user code.

Warning

Items that compare as equal cannot be inserted. Refer to the notes about sorted structures in the previous section.

void Z_init_size(struct Z_head*, size_t size)

Same as Z_init() but preset the minimum hash table to size.

Hash tables also support Z_add() and Z_find() with the same semantics as noted above. Z_find_gteq() and Z_find_lt() are not provided for hash tables.

Hash table invariants

There are several ways to injure yourself using the hash table API.

First, note that there are two functions related to computing uniqueness of objects inserted into the hash table. There is a hash function and a comparison function. The hash function computes the hash of the object. Our hash table implementation uses chaining. This means that your hash function does not have to be perfect; multiple objects having the same computed hash will be placed into a linked list corresponding to that key. The closer to perfect the hash function, the better performance, as items will be more evenly distributed and the chain length will not be long on any given lookup, minimizing the number of list operations required to find the correct item. However, the comparison function must be perfect, in the sense that any two unique items inserted into the hash table must compare not equal. At insertion time, if you try to insert an item that compares equal to an existing item the insertion will not happen and hash_get() will return the existing item. However, this invariant must be maintained while the object is in the hash table. Suppose you insert items A and B into the hash table which both hash to the same value 1234 but do not compare equal. They will be placed in a chain like so:

1234 : A -> B

Now suppose you do something like this elsewhere in the code:

*A = *B

I.e. you copy all fields of B into A, such that the comparison function now says that they are equal based on their contents. At this point when you look up B in the hash table, hash_get() will search the chain for the first item that compares equal to B, which will be A. This leads to insidious bugs.

Warning

Never modify the values looked at by the comparison or hash functions after inserting an item into a hash table.

A similar situation can occur with the hash allocation function. hash_get() accepts a function pointer that it will call to get the item that should be inserted into the list if the provided item is not already present. There is a builtin function, hash_alloc_intern, that will simply return the item you provided; if you always want to store the value you pass to hash_get you should use this one. If you choose to provide a different one, that function must return a new item that hashes and compares equal to the one you provided to hash_get(). If it does not the behavior of the hash table is undefined.

Warning

Always make sure your hash allocation function returns a value that hashes and compares equal to the item you provided to hash_get().

Finally, if you maintain pointers to items you have inserted into a hash table, then before deallocating them you must release them from the hash table. This is basic memory management but worth repeating as bugs have arisen from failure to do this.

API for heaps

Heaps provide the same API as the sorted data structures, except:

  • none of the find functions (Z_find(), Z_find_gteq() or Z_find_lt()) are available.

  • iterating over the heap yields the items in semi-random order, only the first item is guaranteed to be in order and actually the “lowest” item on the heap. Being a heap, only the rebalancing performed on removing the first item (either through Z_pop() or Z_del()) causes the new lowest item to bubble up to the front.

  • all heap modifications are O(log n). However, cacheline efficiency and latency is likely quite a bit better than with other data structures.

Atomic lists

atomlist.h provides an unsorted and a sorted atomic single-linked list. Since atomic memory accesses can be considerably slower than plain memory accessses (depending on the CPU type), these lists should only be used where necessary.

The following guarantees are provided regarding concurrent access:

  • the operations are lock-free but not wait-free.

    Lock-free means that it is impossible for all threads to be blocked. Some thread will always make progress, regardless of what other threads do. (This even includes a random thread being stopped by a debugger in a random location.)

    Wait-free implies that the time any single thread might spend in one of the calls is bounded. This is not provided here since it is not normally relevant to practical operations. What this means is that if some thread is hammering a particular list with requests, it is possible that another thread is blocked for an extended time. The lock-free guarantee still applies since the hammering thread is making progress.

  • without a RCU mechanism in place, the point of contention for atomic lists is memory deallocation. As it is, a rwlock is required for correct operation. The read lock must be held for all accesses, including reading the list, adding items to the list, and removing items from the list. The write lock must be acquired and released before deallocating any list element. If this is not followed, an use-after-free can occur as a MT race condition when an element gets deallocated while another thread is accessing the list.

    Note

    The write lock does not need to be held for deleting items from the list, and there should not be any instructions between the pthread_rwlock_wrlock and pthread_rwlock_unlock. The write lock is used as a sequence point, not as an exclusion mechanism.

  • insertion operations are always safe to do with the read lock held. Added items are immediately visible after the insertion call returns and should not be touched anymore.

  • when removing a particular (pre-determined) item, the caller must ensure that no other thread is attempting to remove that same item. If this cannot be guaranteed by architecture, a separate lock might need to be added.

  • concurrent pop calls are always safe to do with only the read lock held. This does not fall under the previous rule since the pop call will select the next item if the first is already being removed by another thread.

    Deallocation locking still applies. Assume another thread starts reading the list, but gets task-switched by the kernel while reading the first item. pop will happily remove and return that item. If it is deallocated without acquiring and releasing the write lock, the other thread will later resume execution and try to access the now-deleted element.

  • the list count should be considered an estimate. Since there might be concurrent insertions or removals in progress, it might already be outdated by the time the call returns. No attempt is made to have it be correct even for a nanosecond.

Overall, atomic lists are well-suited for MT queues; concurrent insertion, iteration and removal operations will work with the read lock held.

Code snippets

Iteration:

struct item *i;

pthread_rwlock_rdlock(&itemhead_rwlock);
frr_each(itemlist, &itemhead, i) {
  /* lock must remain held while iterating */
  ...
}
pthread_rwlock_unlock(&itemhead_rwlock);

Head removal (pop) and deallocation:

struct item *i;

pthread_rwlock_rdlock(&itemhead_rwlock);
i = itemlist_pop(&itemhead);
pthread_rwlock_unlock(&itemhead_rwlock);

/* i might still be visible for another thread doing an
 * frr_each() (but won't be returned by another pop()) */
...

pthread_rwlock_wrlock(&itemhead_rwlock);
pthread_rwlock_unlock(&itemhead_rwlock);
/* i now guaranteed to be gone from the list.
 * note nothing between wrlock() and unlock() */
XFREE(MTYPE_ITEM, i);

FAQ

What are the semantics of const in the container APIs?

const pointers to list heads and/or items are interpreted to mean that both the container itself as well as the data items are read-only.

Why is it PREDECL + DECLARE instead of DECLARE + DEFINE?

The rule is that a DEFINE must be in a .c file, and linked exactly once because it defines some kind of global symbol. This is not the case for the data structure macros; they only define static symbols and it is perfectly fine to include both PREDECL and DECLARE in a header file. It is also perfectly fine to have the same DECLARE statement in 2 .c files, but only if the macro arguments are identical. Maybe don’t do that unless you really need it.

FRR lists

Todo

document

BSD lists

Todo

refer to external docs