Remove Pinned

[cagatay-y]

Pinned was used to keep track of if a raw pointer to the object was handed out via the use of a boolean and free or not free accordingly. The initial commit replaces it with Rc to keep track of the sharing in a standard and safer way. The second commit eliminates the need to keep track of shares completely by removing the sharing of TransferTokens and adapting the only instance of the use of the returned TransferToken.

[cagatay-y]

I think it is possible that I introduced a memory leak. Can we profile the memory consumption of main and this branch to check?

[mkroening]

We'll sure, you could accumulate the sizes of all allocs minus all deallocs somewhere and compare.
I did something similar for debugging purposes, hacked into the allocator once using an AtomicUsize for #899.
You could also start VMs with minimal memory and create/destroy connections a lot and see if we run out of memory.

[cagatay-y]

I've tested the branch with siege using 10 concurrent connections for 115 seconds (just under the xtask timeout). It handled 34460 requests in total without error. The QEMU process also seems to have a stable memory usage during the test and returns the memory at the end of the test.

[mkroening]

Thanks for testing! How much memory did the VM get?

[cagatay-y]

I am not on my computer right now, so I cannot check to be sure but if I recall correctly, ~95 MB during the test and ~80 MB at idle. The VM limit was the xtask default (128 MB (?)).

[mkroening]

Thanks, this looks great! :)

Sorry for taking so long with the review.

Just a few remarks regarding the introduction of additional unsafe blocks.

[cagatay-y]

Looking closer at the uses of unsafe, I think they are fundamentally not thread-safe. It is possible to use synchronization to make it safe, but I realized that the only uses of shared Transfers are inside the module itself to check for Transfer status and they can be avoided. The latest commit does that.

If it is preferable to keep the possibility to check status through the Transfer for future use, we can also switch to using synchronization. I kept the initial commit (the one with reference counting) to make it more convenient to review the changes to the original PR and to switch back to the Rc approach if necessary. The two commits can be squashed if the approach described above is acceptable.

[mkroening]

Thanks a lot! This looks great.

Just one question about MaybeUninit inline.

Another question: Is the PR description still up to date?

You mention one-bit refcounts. What do you mean with that? I mean a refcounter that can only count to one is exactly the same as just having a single owner.

The description also talks about replacing Pinned<T> with Rc<T>, but to me, the code reads more like replacing it with Box<T> and T, which is good, I was just wondering about the description. :)

[cagatay-y]

I've updated the PR description and implemented the change from MaybeUninit to Option.

[mkroening]

Thanks a lot! :)