Wip strong claim

[mattbenjamin]

Proposed as a precursor to conditional XioMessenger changeseet.

[yuyuyu101]

I have a quick look at accelio source codes. It looks like you want to let memory buffer which is used by ibv_rq can be append to bufferlist and recycle by accelio?

[yuyuyu101]

Yeah, I'm not sure that I have got your idea. It looks like you can inherit "Raw" class which can add registered memory to bufferlist and a special destruction which can recycle memory when ptr->nref == 0.

[mattbenjamin]

Thanks for reviewing.

First, yes, the Xio changesets do derive a "raw" class with special destruction. We've had ones that just recycle the memory, and a stronger one does so via the related completion hook. You're right, that's the natural way to deal with zero-copy buffers in the current buffer API (yuyuyu101).

The issue turned out to be, not getting zero-copy to work, but ensuring that zero-copy buffers were returned in a timely manner. Because buffer::ptrs have been shared by default, we don't know when (or even if) every ref will be returned. So what strong claim is attempting to do is make sharing !default -for a specific class of buffers- (ones with the volatile property). I wanted a way to ensure that, unless a code path would assuredly return volatile buffers "in a timely manner" (something that seemed difficult to ensure "by default"), it would not get zero-copy buffers (an Accelio pool resource, whether or not they are registered).

Sage, will that have the same effect? (I'm genuinely not sure.)

[vuhuong]

liewegas,

Even though strong_claim_inplace() is called by default for claim() family and operator=; however, in buffer::ptr::strong_claim() we check if raw buffer volatile or not to deep copy. By the default the raw buffer is not volatile. We only set is_volatile() == true in case of Accelio's buffers.

[liewegas]

Okay, this makes sense. Perhaps renaming strong_claim_* to strong_claim_volatile_* would make this clearer.

Also, instead of a bool strong=true arg, what about a bool force_weak_volatile=false? That would make it clearer that the exceptional case specific to volatile buffers?