quoted

I still don't understand how a page in the middle of a contiguous range ends
up coming from the slab while others don't.

I haven't investigate the origin of the IO
yet. I suspect the first 2 pages are the superblocks of the raid
(mdp_superblock_1 and bitmap_super_s) and the rest of the IO is the bitmap.

Well, if these indeed are different origins and just *happen* to be a 
mixture
of slab originated pages and non-slab pages combined together in a 
single bio of a bvec entry,
I'd suspect that it would be more beneficial to split the bvec 
(essentially not allow bio_add_page
to append the page to tail bvec depending on a queue limit (similar to 
how we handle sg gaps).

quoted

Ofir, can you please check which condition in sendpage_ok actually fails?

It failed because the page has slab, page count is 1. Sorry for not
clarifying this.

"skbuff: !sendpage_ok - page: 0x54f9f140 (pfn: 120757). is_slab: 1, page_count: 1"
                                                                  ^
The print I used:
pr_info(
     "!sendpage_ok - page: 0x%p (pfn: %lx). is_slab: %u, page_count: %u\n",
     (void *)page,
     page_to_pfn(page),
     page_address(page),
     !!PageSlab(page),
     page_count(page)
);