[TS-1742] Freelists to use 64bit version w/ Double Word Compare and Swap

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.3.2
• Assignee: Brian Geffon
• Reporter: Brian Geffon
• Created: 08/Mar/13 05:03
• Updated: 05/Jun/13 22:42
• Resolved: 11/Mar/13 23:17

Description

So to those of you familiar with the freelists you know that it works this way the head pointer uses the upper 16 bits for a version to prevent the ABA problem. The big drawback to this is that it requires the following macros to get at the pointer or the version:

#define FREELIST_POINTER(_x) ((void*)(((((intptr_t)(_x).data)<<16)>>16) | \
 (((~((((intptr_t)(_x).data)<<16>>63)-1))>>48)<<48)))  // sign extend
#define FREELIST_VERSION(_x) (((intptr_t)(_x).data)>>48)
#define SET_FREELIST_POINTER_VERSION(_x,_p,_v) \
  (_x).data = ((((intptr_t)(_p))&0x0000FFFFFFFFFFFFULL) | (((_v)&0xFFFFULL) << 48))

Additionally, since this only leaves 16 bits it limits the number of versions you can have, well more and more x86_64 processors support DCAS (double word compare and swap / 128bit CAS). This means that we can use 64bits for a version which basically makes the versions unlimited but more importantly it takes those macros above and simplifies them to:

#define FREELIST_POINTER(_x) (_x).s.pointer
#define FREELIST_VERSION(_x) (_x).s.version
#define SET_FREELIST_POINTER_VERSION(_x,_p,_v) \
(_x).s.pointer = _p; (_x).s.version = _v

As you can imagine this will have a performance improvement, in my simple tests I measured a performance improvement of around 6%. Unfortunately, I'm not an expert with this stuff and I would really appreciate more community feedback before I commit this patch.

Note: this only applies if you're not using a reclaimable freelist.

Issue Links

• relates to: TS-1746 Crash report: UnixNetVConnection::reenable > ink_atomiclist_push > ink_atomic_cas