Maya Nakamura <[email protected]> writes: > On Fri, Apr 05, 2019 at 01:31:02PM +0200, Vitaly Kuznetsov wrote: >> Maya Nakamura <[email protected]> writes: >> >> > @@ -98,18 +99,20 @@ EXPORT_SYMBOL_GPL(hyperv_pcpu_input_arg); >> > u32 hv_max_vp_index; >> > EXPORT_SYMBOL_GPL(hv_max_vp_index); >> > >> > +struct kmem_cache *cachep; >> > +EXPORT_SYMBOL_GPL(cachep); >> > + >> > static int hv_cpu_init(unsigned int cpu) >> > { >> > u64 msr_vp_index; >> > struct hv_vp_assist_page **hvp = &hv_vp_assist_page[smp_processor_id()]; >> > void **input_arg; >> > - struct page *pg; >> > >> > input_arg = (void **)this_cpu_ptr(hyperv_pcpu_input_arg); >> > - pg = alloc_page(GFP_KERNEL); >> > - if (unlikely(!pg)) >> > + *input_arg = kmem_cache_alloc(cachep, GFP_KERNEL); >> >> I'm not sure use of kmem_cache is justified here: pages we allocate are >> not cache-line and all these allocations are supposed to persist for the >> lifetime of the guest. In case you think that even on x86 it will be >> possible to see PAGE_SIZE != HV_HYP_PAGE_SIZE you can use alloc_pages() >> instead. >> > Thank you for your feedback, Vitaly! > > Will you please tell me how cache-line relates to kmem_cache? > > I understand that alloc_pages() would work when PAGE_SIZE <= > HV_HYP_PAGE_SIZE, but I think that it would not work if PAGE_SIZE > > HV_HYP_PAGE_SIZE.
Sorry, my bad: I meant to say "not cache-like" (these allocations are not 'cache') but the typo made it completely incomprehensible. > >> Also, in case the idea is to generalize stuff, what will happen if >> PAGE_SIZE > HV_HYP_PAGE_SIZE? Who will guarantee proper alignment? >> >> I think we can leave hypercall arguments, vp_assist and similar pages >> alone for now: the code is not going to be shared among architectures >> anyways. >> > About the alignment, kmem_cache_create() aligns memory with its third > parameter, offset. Yes, I know, I was trying to think about a (hypothetical) situation when page sizes differ: what would be the memory alignment requirements from the hypervisor for e.g. hypercall arguments? In case it's always HV_HYP_PAGE_SIZE we're good but could it be PAGE_SIZE (for e.g. TLB flush hypercall)? I don't know. For x86 this discussion probably makes no sense. I'm, however, struggling to understand what benefit we will get from the change. Maybe just leave it as-is for now and fix arch-independent code only? And later, if we decide to generalize this code, make another approach? (Not insisting, just a suggestion) > >> > @@ -338,7 +349,10 @@ void __init hyperv_init(void) >> > guest_id = generate_guest_id(0, LINUX_VERSION_CODE, 0); >> > wrmsrl(HV_X64_MSR_GUEST_OS_ID, guest_id); >> > >> > - hv_hypercall_pg = __vmalloc(PAGE_SIZE, GFP_KERNEL, PAGE_KERNEL_RX); >> > + hv_hypercall_pg = kmem_cache_alloc(cachep, GFP_KERNEL); >> > + if (hv_hypercall_pg) >> > + set_memory_x((unsigned long)hv_hypercall_pg, 1); >> >> _RX is not writeable, right? >> > Yes, you are correct. I should use set_memory_ro() in addition to > set_memory_x(). > >> > @@ -416,6 +431,7 @@ void hyperv_cleanup(void) >> > * let hypercall operations fail safely rather than >> > * panic the kernel for using invalid hypercall page >> > */ >> > + kmem_cache_free(cachep, hv_hypercall_pg); >> >> Please don't do that: hyperv_cleanup() is called on kexec/kdump and >> we're trying to do the bare minimum to allow next kernel to boot. Doing >> excessive work here will likely lead to consequent problems (we're >> already crashing the case it's kdump!). >> > Thank you for the explanation! I will remove that. > -- Vitaly
