LinuxKPI: make __kmalloc() play by the rules

[bz]

Summary

According to Documentation/core-api/dma-api.rst kmalloc() is supposd
to provide physically contiguous memory. [1]

In order to guarantee that allocations are contiguous even if using
larger than PAGE_SIZE check the size and use contigmalloc if needed.
This makes use of 9e6544dd6e02 (and following) allowing free(9) to
also work for contigmalloced memory.

Sponsored by: The FreeBSD Foundation
Pointed out by: jhb [1]
MFC after: 3 days

[emaste]

Looks reasonable to me. Aside, it seems we don't document boundary=0 in contigmalloc(9).

[bz]

In D46656#1063325, @emaste wrote:

Looks reasonable to me. Aside, it seems we don't document boundary=0 in contigmalloc(9).

Yes, though most drivers in sys/dev do not set it; vm_page_alloc_contig documents it as

"If the boundary
parameter is non-zero, the pages constituting the run will not cross a
physical address that is a multiple of the parameter value, which must be
a power of two."

[emaste]

I followed a very long call chain to arrive at vm_addr_bound_ok, which confirmed -- and the nice thing is boundary == 0 doesn't even need to be a special case:

/*
 * Do the first and last addresses of a range match in all bits except the ones
 * in -boundary (a power-of-two)?  For boundary == 0, all addresses match.
 */
static inline bool
vm_addr_bound_ok(vm_paddr_t pa, vm_paddr_t size, vm_paddr_t boundary)
{                               
        KASSERT(powerof2(boundary), ("%s: boundary is not a power of 2: %#jx",
            __func__, (uintmax_t)boundary));
        return (((pa ^ (pa + size - 1)) & -boundary) == 0);
}

If I have a chance I'll propose a contigmalloc(9) addition.

sys/compat/linuxkpi/common/src/linux_slab.c

221

is this needed? (either for contigmalloc, or to match Linux KPI?)

Looking, kmem_alloc_contig_domain does asize = round_page(size); so unless we need it to be explicit about Linux behaviour perhaps just passing _s through is sensible?

[bz]

sys/compat/linuxkpi/common/src/linux_slab.c

221

[ and so does contigmalloc itself: malloc_type_allocated(type, round_page(size)); ]

I think this is no longer needed; I'll remove it.

[bz]

simplify as @emaste pointed out; we no longer need to round up to page size.

[emaste]

LGTM, we can wait for @jhb as well though

[bz]

@jhb any comments?