One of the ways you can get tripped up in kernel code is that sometimes something really major happens and it isn't in any way highlighted.

For instance - when reading from disk into a folio (whether file-backed or from swap), the folio is locked and will become unlocked once it is 'uptodate' (i.e. has the PG_uptodate flag set).

That renders all instances of trying to lock a folio a barrier that either waits for the read to happen or, as if often the case, incorporates some ability to tell the caller 'ok I need to wait on I/O try again later' as userland experiences with async I/O.

A good example of this is do_swap_page() [0], where we kick off either a synchronous or async read from disk (see block starting at line 3766), but later (line 3878) indicate SIGBUS if the page isn't uptodate.

The key here? Line 3826:

locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);

We try to lock the page, if I/O is done already we can, if it was synchronous we can, otherwise we either return VM_FAULT_RETRY to indicate the I/O is needed or we wait to acquire the lock.

After this line, we can safely assume the swap has been read from.

I kind of feel that a comment would be helpful here :>) [checking tip I see there isn't one, might add one myself], at the same time it's sort of assumed you'd know this was the case...

While I'm writing the book primarily to learn the mm in more depth, I hope pointing out stuff like this in it will help others over these kind of stumbling blocks.

[0]: https://elixir.bootlin.com/linux/v6.0/source/mm/memory.c#L3718