Hi Sean, On 09/04/2018 05:32 PM, Sean Christopherson wrote: ...
/* Use the static page for the first CPUs, allocate otherwise */- if (cpu < HVC_BOOT_ARRAY_SIZE) + if (cpu < HVC_BOOT_ARRAY_SIZE) { p = &hv_clock_boot[cpu]; - else - p = kzalloc(sizeof(*p), GFP_KERNEL); + } else { + int rc; + unsigned int sz = sizeof(*p); + + if (sev_active()) + sz = PAGE_ALIGN(sz);Hmm, again we're wasting a fairly sizable amount of memory since each CPU is doing a separate 4k allocation. What if we defined an auxilary array in __decrypted to be used for cpus > HVC_BOOT_ARRAY_SIZE when SEV is active? struct pvclock_vsyscall_time_info is 32 bytes so we could handle the max of 8192 CPUs with 256kb of data (252kb if you subtract the pre-existing 4k page), i.e. the SEV case wouldn't need additional memory beyond the 2mb page that's reserved for __decrypted. The non-SEV case could do free_kernel_image_pages() on the unused array (which would need to be page sized) so it wouldn't waste memory.
Sure, we could use unused memory range from .data..decrypted section for cpu > HVC_BOOT_ARRAY_SIZE when SEV is active. I will rework the patch. -Brijesh
