Le 08/11/2022 à 19:41, Song Liu a écrit : > On Tue, Nov 8, 2022 at 3:27 AM Mike Rapoport <rppt@xxxxxxxxxx> wrote: >> >> Hi Song, >> >> On Mon, Nov 07, 2022 at 02:39:16PM -0800, Song Liu wrote: >>> This patchset tries to address the following issues: >>> >>> 1. Direct map fragmentation >>> >>> On x86, STRICT_*_RWX requires the direct map of any RO+X memory to be also >>> RO+X. These set_memory_* calls cause 1GB page table entries to be split >>> into 2MB and 4kB ones. This fragmentation in direct map results in bigger >>> and slower page table, and pressure for both instruction and data TLB. >>> >>> Our previous work in bpf_prog_pack tries to address this issue from BPF >>> program side. Based on the experiments by Aaron Lu [4], bpf_prog_pack has >>> greatly reduced direct map fragmentation from BPF programs. >> >> Usage of set_memory_* APIs with memory allocated from vmalloc/modules >> virtual range does not change the direct map, but only updates the >> permissions in vmalloc range. The direct map splits occur in >> vm_remove_mappings() when the memory is *freed*. >> >> That said, both bpf_prog_pack and these patches do reduce the >> fragmentation, but this happens because the memory is freed to the system >> in 2M chunks and there are no splits of 2M pages. Besides, since the same >> 2M page used for many BPF programs there should be way less vfree() calls. >> >>> 2. iTLB pressure from BPF program >>> >>> Dynamic kernel text such as modules and BPF programs (even with current >>> bpf_prog_pack) use 4kB pages on x86, when the total size of modules and >>> BPF program is big, we can see visible performance drop caused by high >>> iTLB miss rate. >> >> Like Luis mentioned several times already, it would be nice to see numbers. >> >>> 3. TLB shootdown for short-living BPF programs >>> >>> Before bpf_prog_pack loading and unloading BPF programs requires global >>> TLB shootdown. This patchset (and bpf_prog_pack) replaces it with a local >>> TLB flush. >>> >>> 4. Reduce memory usage by BPF programs (in some cases) >>> >>> Most BPF programs and various trampolines are small, and they often >>> occupies a whole page. From a random server in our fleet, 50% of the >>> loaded BPF programs are less than 500 byte in size, and 75% of them are >>> less than 2kB in size. Allowing these BPF programs to share 2MB pages >>> would yield some memory saving for systems with many BPF programs. For >>> systems with only small number of BPF programs, this patch may waste a >>> little memory by allocating one 2MB page, but using only part of it. >> >> I'm not convinced there are memory savings here. Unless you have hundreds >> of BPF programs, most of 2M page will be wasted, won't it? >> So for systems that have moderate use of BPF most of the 2M page will be >> unused, right? > > There will be some memory waste in such cases. But it will get better with: > 1) With 4/5 and 5/5, BPF programs will share this 2MB page with kernel .text > section (_stext to _etext); > 2) modules, ftrace, kprobe will also share this 2MB page; > 3) There are bigger BPF programs in many use cases. And what I love with this series (for powerpc/32) is that we will likely now be able to have bpf, ftrace, kprobe without the performance cost of CONFIG_MODULES. Today, CONFIG_MODULES means page mapping, which means handling of kernel page in ITLB miss handlers. By using some of the space between end of rodata and start of inittext, we are able to use ROX linear memory which is mapped by blocks. It means there is no need to handle kernel text in ITLB handler (You can look at https://elixir.bootlin.com/linux/v6.1-rc3/source/arch/powerpc/kernel/head_8xx.S#L191 to better understand what I'm talking about). Thanks Christophe
