On 10/9/2024 10:33 PM, Dave Hansen wrote: > On 10/9/24 09:31, Neeraj Upadhyay wrote: >>> Second, this looks to be allocating a potentially large physically >>> contiguous chunk of memory, then handing it out 4k at a time. The loop is: >>> >>> buf = alloc(NR_CPUS * PAGE_SIZE); >>> for (i = 0; i < NR_CPUS; i++) >>> foo[i] = buf + i * PAGE_SIZE; >>> >>> but could be: >>> >>> for (i = 0; i < NR_CPUS; i++) >>> foo[i] = alloc(PAGE_SIZE); >>> >>> right? >> >> Single contiguous allocation is done here to avoid TLB impact due to backing page >> accesses (e.g. sending ipi requires writing to target CPU's backing page). >> I can change it to allocation in chunks of size 2M instead of one big allocation. >> Is that fine? Also, as described in commit message, reserving entire 2M chunk >> for backing pages also prevents splitting of NPT entries into individual 4K entries. >> This can happen if part of a 2M page is not allocated for backing pages by guest >> and page state change (from private to shared) is done for that part. > > Ick. > > First, this needs to be thoroughly commented, not in the changelogs. > Ok. > Second, this is premature optimization at its finest. Just imagine if > _every_ site that needed 16k or 32k of shared memory decided to allocate > a 2M chunk for this _and_ used it sparsely. What's the average number > of vCPUs in a guest. 4? 8? > Got it. > The absolute minimum that we can do here is some stupid infrastructure > that you call for allocating shared pages, or for things that _will_ be > converted to shared so they get packed. > > But hacking uncommented 2M allocations into every site seems like > insanity to me. > > IMNHO, you can either invest the time to put the infrastructure in place > and get 2M pages, or you can live with the suboptimal performance of 4k. I will start with 4K. For later, I will get the performance numbers to propose a change in allocation scheme - for ex, allocating a bigger contiguous batch from the total allocation required for backing pages (num_possible_cpus() * 4K) without doing 2M reservation. - Neeraj
