Messages in this thread | | | From | Tianyu Lan <> | Subject | Re: [Resend RFC PATCH V4 09/13] x86/Swiotlb/HV: Add Swiotlb bounce buffer remap function for HV IVM | Date | Wed, 21 Jul 2021 18:28:48 +0800 |
| |
Thanks for review.
On 7/20/2021 9:54 PM, Christoph Hellwig wrote: > > Please split the swiotlb changes into a separate patch from the > consumer.
OK. Will update.
> >> } >> + >> +/* >> + * hv_map_memory - map memory to extra space in the AMD SEV-SNP Isolation VM. >> + */ >> +unsigned long hv_map_memory(unsigned long addr, unsigned long size) >> +{ >> + unsigned long *pfns = kcalloc(size / HV_HYP_PAGE_SIZE, >> + sizeof(unsigned long), >> + GFP_KERNEL); >> + unsigned long vaddr; >> + int i; >> + >> + if (!pfns) >> + return (unsigned long)NULL; >> + >> + for (i = 0; i < size / HV_HYP_PAGE_SIZE; i++) >> + pfns[i] = virt_to_hvpfn((void *)addr + i * HV_HYP_PAGE_SIZE) + >> + (ms_hyperv.shared_gpa_boundary >> HV_HYP_PAGE_SHIFT); >> + >> + vaddr = (unsigned long)vmap_pfn(pfns, size / HV_HYP_PAGE_SIZE, >> + PAGE_KERNEL_IO); >> + kfree(pfns); >> + >> + return vaddr; > > This seems to miss a 'select VMAP_PFN'.
VMAP_PFN has been selected in the previous patch "RFC PATCH V4 08/13] HV/Vmbus: Initialize VMbus ring buffer for Isolation VM"
> But more importantly I don't > think this actually works. Various DMA APIs do expect a struct page > backing, so how is this going to work with say dma_mmap_attrs or > dma_get_sgtable_attrs?
dma_mmap_attrs() and dma_get_sgtable_attrs() get input virtual address belonging to backing memory with struct page and returns bounce buffer dma physical address which is below shared_gpa_boundary(vTOM) and passed to Hyper-V via vmbus protocol.
The new map virtual address is only to access bounce buffer in swiotlb code and will not be used other places. It's stored in the mem->vstart. So the new API set_memory_decrypted_map() in this series is only called in the swiotlb code. Other platforms may replace set_memory_decrypted() with set_memory_decrypted_map() as requested.
> >> +static unsigned long __map_memory(unsigned long addr, unsigned long size) >> +{ >> + if (hv_is_isolation_supported()) >> + return hv_map_memory(addr, size); >> + >> + return addr; >> +} >> + >> +static void __unmap_memory(unsigned long addr) >> +{ >> + if (hv_is_isolation_supported()) >> + hv_unmap_memory(addr); >> +} >> + >> +unsigned long set_memory_decrypted_map(unsigned long addr, unsigned long size) >> +{ >> + if (__set_memory_enc_dec(addr, size / PAGE_SIZE, false)) >> + return (unsigned long)NULL; >> + >> + return __map_memory(addr, size); >> +} >> + >> +int set_memory_encrypted_unmap(unsigned long addr, unsigned long size) >> +{ >> + __unmap_memory(addr); >> + return __set_memory_enc_dec(addr, size / PAGE_SIZE, true); >> +} > > Why this obsfucation into all kinds of strange helpers? Also I think > we want an ops vectors (or alternative calls) instead of the random > if checks here.
Yes, agree and will add ops for different platforms to map/unmap memory.
> >> + * @vstart: The virtual start address of the swiotlb memory pool. The swiotlb >> + * memory pool may be remapped in the memory encrypted case and store > > Normall we'd call this vaddr or cpu_addr.
OK. Will update.
> >> - set_memory_decrypted((unsigned long)vaddr, bytes >> PAGE_SHIFT); >> - memset(vaddr, 0, bytes); >> + mem->vstart = (void *)set_memory_decrypted_map((unsigned long)vaddr, bytes); > > Please always pass kernel virtual addresses as pointers. > > And I think these APIs might need better names, e.g. > > arch_dma_map_decrypted and arch_dma_unmap_decrypted. > > Also these will need fallback versions for non-x86 architectures that > currently use memory encryption.
Sure. Will update in the next version.
| |