Summary:
Previously: Each allocation happened in the first chain after any free.
Now: All allocation will happen in one chain until all buckets are full, this can reduce
fragmentation to some extent.
Reviewers: seoz, govi, shilpasingh, raster, cedric
Reviewed By: cedric
Subscribers: cedric, rajeshps
Differential Revision: https://phab.enlightenment.org/D2071
Signed-off-by: Cedric BAIL <cedric@osg.samsung.com>
When over-allocating (past "pool->max" items) a memory slice will be
allocated to the new item as a linked list using Eina_Inlist.
The original code was placing the Eina_Inlist structure (3 pointers)
at the beginning of the allocated memory. However the item must have
proper alignment based on "pool->item_size", otherwise a structure may
end with unaligned members. Take for example MIPS 32 bits, it uses 4
bytes pointers with 8 bytes double. A structure containing a double
could have it unaligned as 12 % 8 = 4 (12 is the size of Eina_Inlist,
that contains 3 pointers), and MIPS doesn't allow unaligned access.
Albin Tonnerre (Lutin) spotted this in his Debian MIPS test machine,
it was breaking at eet_data_get_double() that was storing an unaligned
double. This was being called from within edje test suite.
The current code will place the list node after the requested
"pool->item_size", of course guaranteeing the pointer inside the node
is aligned (otherwise a "char" or "short" would break its alignment).
Is there a title for the commiter with the longest revert of a revert list?
I screwed this one up. Had other local changes that sneaked in.
Resetting and doing it right now.
This reverts commit ee155b771d.
You can bet on reverting makes people speed up the process in fixing it.
Revert the revert here now that Cedric fixed it in eina.
This reverts commit 875e7cf74d.
This patch brings in constant E crash for me. It seems to be timing
related which somehow leads to a mem corruption. Revert it for now
to allow people using E while Cedric looks into it.
This reverts commit 12d34309c7.
To avoid people getting confused over this code, like myself, add a comment
to clarify that only the elements need to be set to zero while the payload
does not matter here.
Please check.
note1: Only lib and bin for now, but should be extended to other stuff
note2: distcheck does not work because eo_suite is failing.
SVN revision: 78758
eina headers used to be in their own directory, but now they live
inside src/lib/eina. The mempools were still including
include/lib/eina/ and with "sed" they became src/lib/eina/, which
already existed.
SVN revision: 77837
Now we always build the following memory pools statically:
- pass_through: calls malloc/free directly, useful to debug.
- chained_pool: default for ages.
- one_big: used by some embedded systems (should we remove?)
Removed:
- ememoa_fixed and ememoa_unknown: depends on a separate lib, not supported?
- buddy: nobody uses it?
NOTE: we do not need the src/modules/eina/mp/*/Makefile.am anymore
since they are statically built. But I'll keep these and the
references in src/modules/eina/mp/Makefile.am
SVN revision: 77792