Commit Graph

44 Commits

Author SHA1 Message Date
Matthew Parkinson
d588e8ecb1 Updated to match documentation. 2019-02-19 15:00:55 +00:00
Matthew Parkinson
aac2b9212e Merge pull request #21 from plietar/smart-next
Place the next pointer at a different place on every object.
2019-02-19 09:04:23 +00:00
Paul Liétar
ffe496dd18 Address review comments. 2019-02-18 21:29:32 +00:00
Paul Liétar
4ac3421487 Place the next pointer at a different place on every object.
This limits the collision in cache buckets, especially for larger objects.
2019-02-18 14:22:50 +00:00
David Chisnall
c47eed1922 Tweaks to end bounds checking.
Introduce a `OnePastEnd` option for the pointer immediately after the
end of the allocation.  This simplifies some of the logic in callers,
where they wants to say 'is base + length safe to use?'.

Also restructure some of the other logic somewhat.
2019-02-15 20:05:00 +01:00
Matthew Parkinson
5e149b083f One more constexpr 2019-02-15 17:43:27 +00:00
Matthew Parkinson
8698f87018 Added some constexpr to some ifs. 2019-02-15 16:12:38 +00:00
Matthew Parkinson
ac8e5e4e5b Fix comment to represent current PAL API. 2019-02-15 09:24:01 +00:00
Matthew Parkinson
adeecfb0f1 Added bug fix to alignment of using. 2019-02-14 20:08:04 +00:00
Paul Liétar
79b432ddab Fix the condition on when to allocate a new block.
Previously we would pretty much always allocate a new block for each
allocator, making them 16MiB aligned.
2019-02-12 12:47:22 +00:00
Matthew Parkinson
731757b29c Merge pull request #13 from Microsoft/matt-refactor
Simplification around remote allocation
2019-02-06 11:25:07 +00:00
Matthew Parkinson
f3897dd3e0 Code review feedback. 2019-02-05 16:57:09 +00:00
Matthew Parkinson
83c55fe5da Simplify remote allocator.
Don't store the sizeclass, as we can find it easily, and already are
going to touch though cache lines in the common case.
2019-02-05 13:27:19 +00:00
Matthew Parkinson
122a5de811 Refactor code to correctly handle low bits of allocator ID. 2019-02-05 13:11:55 +00:00
theodus
85cf5dd097 Avoid unnecessary allocation from realloc 2019-02-02 14:47:06 -05:00
Matthew Parkinson
ec59820970 Use placement new instead of init for metaslab. 2019-01-25 19:46:24 +00:00
Matthew Parkinson
06ff1ffe0a Renamed zero to init for metaslab. 2019-01-25 16:38:10 +00:00
Matthew Parkinson
e5c40c1231 Remove an auto as not correctly inferred. 2019-01-24 11:48:02 +00:00
Matthew Parkinson
d5fb7150a3 Clang format 2019-01-24 11:39:19 +00:00
Matthew Parkinson
294efe8e03 Hardening access to meta data
Introduce a wrapper that ensure various fields of meta data, represent
valid indexes into the the address space, and not beyond the current
slab/superslab.
2019-01-24 11:39:18 +00:00
Matthew Parkinson
2083b29c9b Remove a raw pointer usage
Metaslab can be handled as as ref, rather than a raw pointer.
2019-01-24 11:39:18 +00:00
Matthew Parkinson
e30c94cb2a Apply ModArray to Metaslab data
Prevent bugs accessing outside the Slabs meta data.
2019-01-24 11:39:18 +00:00
Matthew Parkinson
745a2a53d1 Clang format 2019-01-22 17:18:54 +00:00
Matthew Parkinson
cb70aa3f7f Restricted another array to bounds. 2019-01-22 11:19:57 +00:00
Matthew Parkinson
18dd15c2c0 Minor restructuring to move static_assert 2019-01-22 11:19:57 +00:00
Matthew Parkinson
bb5027b454 Harden sizeclass table
If a sizeclass in the metadata is corrupted, then this can be used to
force an index beyond the end of these tables. This extends the tables
to the next power of two, and uses a mask on the index, so they are
always either a valid piece of data, or zero.
2019-01-22 11:19:57 +00:00
Matthew Parkinson
94f8b886a0 Remove union in metaslab
The union in Metaslab provides no benefit in size, as the single byte
it effectively saves will be removed due to padding.

By removing the union, we get stronger properties over sizeclass, and
remove an out-of-bounds access.
2019-01-18 21:11:01 +00:00
David Chisnall
20b84b9810 Merge pull request #7 from Microsoft/mjp41/refactor
Fixes remote and pagemap
2019-01-18 17:11:25 +00:00
Matthew Parkinson
c1e23c497f Made alignment a template parameter. 2019-01-18 17:00:10 +00:00
Matthew Parkinson
07b19d43cb Merge pull request #4 from Microsoft/fbsdlibc
Merge changes required for using snmalloc in FreeBSD libc
2019-01-18 16:08:51 +00:00
David Chisnall
c3d46a9d8e Address review comments.
Move slow_allocator into a separate header in the snmalloc namespace and
rename it for consistency with the rest of the codebase.  Delete its
copy and move constructors / assignment operators.
2019-01-18 15:57:28 +00:00
Matthew Parkinson
a62e77f930 Align pagemap entries to OS_PAGE_SIZE
Guarantee the page map is page aligned.  Fix public API.
2019-01-18 15:22:08 +00:00
Matthew Parkinson
62ad70b17e Fixes to remote deallocation
The encoding in the top bits for the size class did not respect kernel
pointers.  Using an intptr_t means, we can use a signed shift to
maintain the kernel pointers.
2019-01-18 14:30:23 +00:00
Matthew Parkinson
806f7e47cc Rename TypeAlloc to Pool
Minor refactor to clarify purpose of the pool.
2019-01-18 11:12:09 +00:00
Matthew Parkinson
4ad12f33cf Fix spacing from a bad clang format. 2019-01-17 16:47:55 +00:00
Matthew Parkinson
ee9d899aa3 Remove GRANULARITY check
This check was not doing anything.  It was from an earlier dependency
ordering issue that has now been fixed.
2019-01-17 14:38:14 +00:00
Matthew Parkinson
164df40372 Add configuration option for small address spaces
This adds a configuration option for 1MiB superslabs with 16KiB slabs.
2019-01-17 14:09:50 +00:00
Matthew Parkinson
b097b47aa2 Handle 2^16 slabs
This code assumes that 16bit uint wrapping can be used, rather than
masking by SLAB_SIZE-1.  Add explicit mask, so we can change the
SLAB_SIZE to be smaller.
2019-01-17 14:09:50 +00:00
Matthew Parkinson
ab57c86e3a Addressing CR feedback 2019-01-16 17:03:15 +00:00
Matthew Parkinson
233be30731 Implement a FlatPagemap
The current pagemap assumes there is at least one level of indexing.
This commit introduces a new pagemap that is completely flat.  This is
useful for 32bit, where there is no need to introduce the indexing
structure.

This pagemap is also considerably faster for 64bit platforms, but
does require a global allocation of 16MiB of the flat page map.
2019-01-16 14:38:11 +00:00
Matthew Parkinson
82595dc9cd Made Remote deallocation work on 32bit
The design of Remote used the top bits of the allocator id to encode
the sizeclass of the deallocation.  On 32bit, or on a platform that uses all the bits
we cannot use these bits for a sizeclass.

This commit uses the bottom bit of the allocator id (which is
guaranteed to be 0), to indicate if the object is the minimum
allocation size.  If it is not the minimum allocation size the
subsequent byte is used to encode the sizeclass.

The code uses constexpr to decide which strategy to use.
2019-01-16 14:38:10 +00:00
Matthew Parkinson
74018a1c91 Correct PageMap public API 2019-01-16 14:38:10 +00:00
theodus
e2190f376f Fix typo in threadalloc.h 2019-01-16 09:22:01 +00:00
Matthew Parkinson
4f9d991449 Initial commit of snmalloc
History squashed from internal development.

Internal history has commit hash:
  e27a0e485c44a5003a802de2661ce3b21e120316
2019-01-15 14:17:55 +00:00