RE: [PATCH 2/6] x86: hv: hv_init.c: Replace alloc_page() with kmem_cache_alloc()
From: Vitaly Kuznetsov
Date: Fri May 10 2019 - 09:22:57 EST
Michael Kelley <mikelley@xxxxxxxxxxxxx> writes:
> From: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> Sent: Wednesday, May 8, 2019 7:55 AM
>> >> Sorry, my bad: I meant to say "not cache-like" (these allocations are
>> >> not 'cache') but the typo made it completely incomprehensible.
>> > No worries! Thank you for sharing your thoughts with me, Vitaly.
>> > Do you know of any alternatives to kmem_cache that can allocate memory
>> > in a specified size (different than a guest page size) with alignment?
>> > Memory allocated by alloc_page() is aligned but limited to the guest
>> > page size, and kmalloc() can allocate a particular size but it seems
>> > that it does not guarantee alignment. I am asking this while considering
>> > the changes for architecture independent code.
>> I think we can consider these allocations being DMA-like (because
>> Hypervisor accesses this memory too) so you can probably take a look at
>> dma_pool_create()/dma_pool_alloc() and friends.
> I've taken a look at dma_pool_create(), and it takes a "struct device"
> argument with which the DMA pool will be associated. That probably
> makes DMA pools a bad choice for this usage. Pages need to be allocated
> pretty early during boot for Hyper-V communication, and even if the
> device subsystem is initialized early enough to create a fake device,
> such a dependency seems rather dubious.
We can probably use dma_pool_create()/dma_pool_alloc() from vmbus module
but these 'early' allocations may not have a device to bind to indeed.
> kmem_cache_create/alloc() seems like the only choice to get
> guaranteed alignment. Do you see any actual problem with
> using kmem_cache_*, other than the naming? It seems like these
> kmem_cache_* functions really just act as a sub-allocator for
> known-size allocations, and "cache" is a common usage
> pattern, but not necessarily the only usage pattern.
Yes, it's basically the name - it makes it harder to read the code and
some future refactoring of kmem_cache_* may not take our use-case into
account (as we're misusing the API). We can try renaming it to something
generic of course and see what -mm people have to say :-)