mirror of
https://review.haiku-os.org/haiku
synced 2025-02-23 14:08:15 +01:00
0b69689a18
although it put the freed space in the free list. Alternating
invocations of alloc_...() and free_...() would thus increase the
header space unboundedly.
* Unified the way a data_header and data_node access the header space.
Originally data_header::data_space and data_node::header_space had to
be kept in sync, which some functions failed to do. Introduced a
header_space structure, which is located in data_header and referenced
by data_node, so that accessing it either way does always keep both
structures in sync.
* Removed the special handling for the data node created with a buffer.
Since remove_{header,trailer}() and trim() could remove it without
knowing, the last reference to the first data header would be freed
prematurely, causing operations on freed memory, and in the end a
second free which screwed the object cache's free list. This crashed
Haiku e.g. when running OpenSSH's "forwarding" test. Now the first
created node is just a node like any other. It's allocated in the data
header, too.
* Changed the mechanism how data nodes are allocated. Now they will
always be allocated on a header associated with the buffer for which
they are created. This fixes a race condition when freeing them. They
would otherwise modify the free list of a header which might be
accessed by another thread at the same time (added a TODO explaining
how the old code could possibly be fixed). Also squashed several
TODOs related to running out of header space when allocating a node.
If the buffer runs out of header space, it will simply allocate a new
header, now.
* Dealt with some TODOs regarding reverting the buffer to its previous
state when running out of memory after allocating a few nodes.
* Added several TODOs.
git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@25256 a95241bf-73f2-0310-859d-f6bbb57e9c96