There should be an extra newline after variable declarations.

Can we get away without having both pindex and p?

Doesn't the increment of pindex belong in the outer loop?

This isn't really true in general. A hysteresis loop could prevent subsequent allocations until the amount of free memory has hit some threshold. I haven't spent much time in that part of UMA lately though and don't know what behaviour to expect offhand.

I'm also not really convinced that we should try to optimize this rarely executed path.

Yes. Careless haste on my part.

I'll take out the extra optimization loop. I would like to completely process sb in this iteration of the outer loop, though.

Shouldn't this be d_mask &= ~(1 << i);?

I would probably just initialize all entries in d to SWAPBLK_NONE and avoid having d_mask entirely.

As a matter of style, or of correctness? What's here is correct. Maybe if I flipped the bit with '^' to set it above, instead of using '|', then flipping it here would be stylistically okay?

If I drop d_mask, and initialize d[], then I have to walk the entire d array looking for values not equal to SWAPBLK_NONE, every time, where now I only have to set d_mask to zero and compare it to zero in all the normal, non-blocking iterations.

Perhaps I'm failing to understand what solution you are asking for.

Sorry, I misunderstood what the old code was doing. I do the find the use of | and & in that context easier to read.

I don't object to the use of a bitmask in general, and compared to this version find the old one easier to understand.

you can merge this w/ blk's def

what happens if SWAP_META_PAGES increases? there should be an assertion to guard against this

for ease of understanding, I would assert that d_mask == 0

keep in mind that this could hypothetically fail on iteration i and succeed on i + 1; however, I don't see this causing a problem

Ideally, d_mask would be of type pn_popmap_t, or an alias for that, but that type is hidden in subr_pctrie.c. It could be exposed in the header file.

As it happens, a static assert in subr_pctrie.c guarantees that int will be big enough for int.

So, do you want a static assert about sizeof(d_mask) or to expose pn_popmap_t?

As a followup, I think that it would be good to document here how this code handles the case where the destination already has a blk at a page index.

... and a comment here too.