With the previous approach we'd require at least an additional pointer
per item to store them in a list (15-18% increase in the overhead per
item). Instead we switch from handling collisions with overflow lists to
an open addressing scheme and store the actual table as variable-sized
indices pointing into an array of all inserted items in their original
order.
This can reduce the memory overhead even compared to the previous
implementation (especially for smaller tables), but because the array for
items is preallocated whenever the table is resized, it can be worse for
certain numbers of items. However, avoiding all the allocations required
by the previous design is actually a big advantage.
Depending on the usage pattern, the performance can improve quite a bit (in
particular when inserting many items). The raw lookup performance is a bit
slower as probing lengths increase with open addressing, but there are some
caching benefits due to the compact storage. So for general usage the
performance should be better. For instance, one test I did was counting the
occurrences of words in a list of 1'000'000 randomly selected words from a
dictionary of ~58'000 words (i.e. using a counter stored under each word as
key). The new implementation was ~8% faster on average while requiring
10% less memory.
Since we can't remove items from the array (would change the indices of all
items that follow it) we just mark them as removed and remove them once the
hash table is resized/rehashed (the cells in the hash table for these may
be reused). Due to this the latter may also happen if the number of stored
items does not increase e.g. after a series of remove/put operations (each
insertion requires storage in the array, no matter if items were removed).
So if the capacity is exhausted, the table is resized/rehashed (after lots
of removals the size may even be reduced) and all items marked as removed
are simply skipped.
Compared to the previous implementation the load factor/capacity is
lowered to reduce chances of collisions and to avoid primary clustering to
some degree. However, the latter in particular, but the open addressing
scheme in general, make this implementation completely unsuited for the
get_match() functionality (purposefully hashing to the same value and,
therefore, increasing the probing length and clustering). And keeping the
keys optionally sorted would complicate the code significantly. So we just
keep the existing hashlist_t implementation without adding code to maintain
the overall insertion order (we could add that feature optionally later, but
with the mentioned overhead for one or two pointers).
The maximum size is currently not changed. With the new implementation
this translates to a hard limit for the maximum number of items that can be
held in the table (=CAPACITY(MAX_SIZE)). Since this equals 715'827'882
items with the current settings, this shouldn't be a problem in practice,
the table alone would require 20 GiB in memory for that many items. The
hashlist_t implementation doesn't have that limitation due to the overflow
lists (it can store beyond it's capacity) but it itself would require over
29 GiB of memory to hold that many items.
This allows us to use it without having to initialize libcharon, which
was required for the logging (we probably could have included debug.h
instead of daemon.h to workaround that but this seems more correct).
This allows us to disable Unix sockets cleanly on Windows. Replaces some
read/write calls with recv/send counterparts, as Winsock does not like
read/writes.
This parser features several improvements over the existing one.
For instance, quoted strings (with escape sequences), unlimited includes,
relaxed newline handling (e.g. at the end of files or before/after { and }),
and the difference between empty and unset values (key = vs. key = "").
It also complains a lot more about invalid syntax. The current one accepts
pretty odd stuff (like settings or sections without name) without any
errors or warnings.
Currently we use the very versatile linked-list collection to store elements
with variable count. This is fine, but very inefficient: Due to the many
methods in the linked list, on 64-bit platforms an empty list alone is more
than 200 bytes. As we currently have about 50 lists per IKE_SA/CHILD_SA pair,
this takes up to 10KB just for managing the empty lists. This is about the
half of memory used by an IKE_SA/CHILD_SA pair, and obviously way too much.
The new array type is not an object, but a collection of functions on an
abstract type.
The following lists are per IKE_SA and should be considered for a replacement
with more efficient arrays (this uses load-testers on-demand created dynamic
configurations, other scenarios have different lists):
14 -> ike_sa_create() @ src/libcharon/sa/ike_sa.c:2198
10 -> auth_cfg_create() @ src/libstrongswan/credentials/auth_cfg.c:1088
6 -> task_manager_v2_create() @ src/libcharon/sa/ikev2/task_manager_v2.c:1505
6 -> proposal_create() @ src/libcharon/config/proposal.c:592
5 -> peer_cfg_create() @ src/libcharon/config/peer_cfg.c:657
4 -> child_sa_create() @ src/libcharon/sa/child_sa.c:1090
2 -> child_cfg_create() @ src/libcharon/config/child_cfg.c:536
1 -> ike_cfg_create() @ src/libcharon/config/ike_cfg.c:330
1 -> put_connected_peers() @ src/libcharon/sa/ike_sa_manager.c:854