Restrict supported alignment for malloc_domainset_aligned(9) to PAGE_SIZE.

[kib]

Summary

UMA page_alloc() does not take an alignment, so UMA can only handle alignment less then page size.

Noted by: alc

[markj]

Sorry for not catching this.

sys/kern/kern_malloc.c

1176

Shouldn't we also formally request page alignment for zones with size > PAGE_SIZE?

[kib]

sys/kern/kern_malloc.c

1176

I am not sure that this usage makes much sense, since at least busdma code wouldn't do it. Anyway, I changed it.

[kib]

Align larger zone on PAGE_SIZE.

[markj]

sys/kern/kern_malloc.c

780

This also needs change to size = roundup2(align), I think.

[kib]

Also use the right zone for allocation, by bumping size to the multiple of align.

Reported by: melifaro

[kib]

Unconditionally roundup.

[jrtc27]

sys/kern/kern_malloc.c

1176

MIN from param.h?

[alc]

sys/kern/kern_malloc.c

1176

I believe that an allocation request for 384 bytes with 128 byte alignment will still go to the malloc-384 zone, yes? But, we aren't requiring better than UMA_ALIGN_PTR alignment.

[kib]

sys/kern/kern_malloc.c

1176

bde@ did not liked the macro AFAIR.

[kib]

roundup alloc size to next power of two

[alc]

sys/kern/kern_malloc.c

1176

That's also my recollection. He preferred the inline functions from libkern.h.

[jrtc27]

sys/kern/kern_malloc.c

781

Isn't this going to be slow on architectures without that instruction? What are the actual requirements here? This could at least do with a comment; flsl is a rather obscure function that's non-obvious unless you've seen it before.

[alc]

sys/kern/kern_malloc.c

781

A "count leading zeroes" instruction is available on every widely used architecture, including all of the architectures supported by FreeBSD. That said, the last time that I looked at the various cpufunc.h implementations, only 32-bit and 64-bit arm and x86 used the builtins that provide access to these instructions.

The objective is to round up to the next power of 2, because we can only guarantee alignment for power-of-2-sized allocations. I agree that a comment to this effect would help.

[kib]

Add comment.
Bump size up to align (restore this part).

[kib]

Restore -1 for zone align calculation.

Reported by: pho

[jrtc27]

sys/kern/kern_malloc.c

781

Ok, which then confirms my suspicion that this is wrong: if size is a power of two it doubles the allocation size. https://godbolt.org/z/a6dhEq (using the arm64 implementation since it's a simple one in terms of the builtin).

[jrtc27]

sys/kern/kern_malloc.c

781

(you want to subtract 1 before passing to flsl, and be careful about 0 - 1 if that's possible)

[pho]

I ran tests with D28219.82616.diff on amd64 for two hours. No problems seen.

[kib]

Do not unnecessary double alloc size.

[pho]

I ran tests with D28219.82645.diff for 6 hours. No problems seen.

[alc]

sys/kern/kern_malloc.c

783

I think that this sentence would be clearer without "at size".

784

785

786

[kib]

Edits for the comment.

[alc]

sys/kern/kern_malloc.c

786

Drop the first "the" at the beginning of the line and add "their" at the end.

[alc]

sys/kern/kern_malloc.c

789–791

Isn't the following one line sufficient?

asize = size < align ? align : 1UL << flsl(size - 1);

[kib]

Simplify calculation. Edit comment.