Re: [PATCH bpf-next 0/2] bpf, riscv: use BPF prog pack allocator in BPF JIT
From: Puranjay Mohan
Date: Mon Aug 14 2023 - 11:39:13 EST
Hi Björn,
On Mon, Aug 14, 2023 at 4:29 PM Björn Töpel <bjorn@xxxxxxxxxx> wrote:
>
> Puranjay Mohan <puranjay12@xxxxxxxxx> writes:
>
> > On Mon, Aug 14, 2023 at 12:40 PM Björn Töpel <bjorn@xxxxxxxxxx> wrote:
> >>
> >> Björn Töpel <bjorn@xxxxxxxxxx> writes:
> >>
> >> > Puranjay Mohan <puranjay12@xxxxxxxxx> writes:
> >> >
> >> >> BPF programs currently consume a page each on RISCV. For systems with many BPF
> >> >> programs, this adds significant pressure to instruction TLB. High iTLB pressure
> >> >> usually causes slow down for the whole system.
> >> >>
> >> >> Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue.
> >> >> It packs multiple BPF programs into a single huge page. It is currently only
> >> >> enabled for the x86_64 BPF JIT.
> >> >>
> >> >> I enabled this allocator on the ARM64 BPF JIT[2]. It is being reviewed now.
> >> >>
> >> >> This patch series enables the BPF prog pack allocator for the RISCV BPF JIT.
> >> >> This series needs a patch[3] from the ARM64 series to work.
> >> >>
> >> >> ======================================================
> >> >> Performance Analysis of prog pack allocator on RISCV64
> >> >> ======================================================
> >> >>
> >> >> Test setup:
> >> >> ===========
> >> >>
> >> >> Host machine: Debian GNU/Linux 11 (bullseye)
> >> >> Qemu Version: QEMU emulator version 8.0.3 (Debian 1:8.0.3+dfsg-1)
> >> >> u-boot-qemu Version: 2023.07+dfsg-1
> >> >> opensbi Version: 1.3-1
> >> >>
> >> >> To test the performance of the BPF prog pack allocator on RV, a stresser
> >> >> tool[4] linked below was built. This tool loads 8 BPF programs on the system and
> >> >> triggers 5 of them in an infinite loop by doing system calls.
> >> >>
> >> >> The runner script starts 20 instances of the above which loads 8*20=160 BPF
> >> >> programs on the system, 5*20=100 of which are being constantly triggered.
> >> >> The script is passed a command which would be run in the above environment.
> >> >>
> >> >> The script was run with following perf command:
> >> >> ./run.sh "perf stat -a \
> >> >> -e iTLB-load-misses \
> >> >> -e dTLB-load-misses \
> >> >> -e dTLB-store-misses \
> >> >> -e instructions \
> >> >> --timeout 60000"
> >> >>
> >> >> The output of the above command is discussed below before and after enabling the
> >> >> BPF prog pack allocator.
> >> >>
> >> >> The tests were run on qemu-system-riscv64 with 8 cpus, 16G memory. The rootfs
> >> >> was created using Bjorn's riscv-cross-builder[5] docker container linked below.
> >> >>
> >> >> Results
> >> >> =======
> >> >>
> >> >> Before enabling prog pack allocator:
> >> >> ------------------------------------
> >> >>
> >> >> Performance counter stats for 'system wide':
> >> >>
> >> >> 4939048 iTLB-load-misses
> >> >> 5468689 dTLB-load-misses
> >> >> 465234 dTLB-store-misses
> >> >> 1441082097998 instructions
> >> >>
> >> >> 60.045791200 seconds time elapsed
> >> >>
> >> >> After enabling prog pack allocator:
> >> >> -----------------------------------
> >> >>
> >> >> Performance counter stats for 'system wide':
> >> >>
> >> >> 3430035 iTLB-load-misses
> >> >> 5008745 dTLB-load-misses
> >> >> 409944 dTLB-store-misses
> >> >> 1441535637988 instructions
> >> >>
> >> >> 60.046296600 seconds time elapsed
> >> >>
> >> >> Improvements in metrics
> >> >> =======================
> >> >>
> >> >> It was expected that the iTLB-load-misses would decrease as now a single huge
> >> >> page is used to keep all the BPF programs compared to a single page for each
> >> >> program earlier.
> >> >>
> >> >> --------------------------------------------
> >> >> The improvement in iTLB-load-misses: -30.5 %
> >> >> --------------------------------------------
> >> >>
> >> >> I repeated this expriment more than 100 times in different setups and the
> >> >> improvement was always greater than 30%.
> >> >>
> >> >> This patch series is boot tested on the Starfive VisionFive 2 board[6].
> >> >> The performance analysis was not done on the board because it doesn't
> >> >> expose iTLB-load-misses, etc. The stresser program was run on the board to test
> >> >> the loading and unloading of BPF programs
> >> >>
> >> >> [1] https://lore.kernel.org/bpf/20220204185742.271030-1-song@xxxxxxxxxx/
> >> >> [2] https://lore.kernel.org/all/20230626085811.3192402-1-puranjay12@xxxxxxxxx/
> >> >> [3] https://lore.kernel.org/all/20230626085811.3192402-2-puranjay12@xxxxxxxxx/
> >> >> [4] https://github.com/puranjaymohan/BPF-Allocator-Bench
> >> >> [5] https://github.com/bjoto/riscv-cross-builder
> >> >> [6] https://www.starfivetech.com/en/site/boards
> >> >>
> >> >> Puranjay Mohan (2):
> >> >> riscv: Extend patch_text_nosync() for multiple pages
> >> >> bpf, riscv: use prog pack allocator in the BPF JIT
> >> >
> >> > I get a hang for "test_tag", but it's not directly related to your
> >> > series, but rather "remote fence.i".
> >> >
> >> > | rcu: INFO: rcu_sched detected stalls on CPUs/tasks:
> >> > | rcu: 0-....: (1400 ticks this GP) idle=d5e4/1/0x4000000000000000 softirq=5542/5542 fqs=1862
> >> > | rcu: (detected by 1, t=5252 jiffies, g=10253, q=195 ncpus=4)
> >> > | Task dump for CPU 0:
> >> > | task:kworker/0:5 state:R running task stack:0 pid:319 ppid:2 flags:0x00000008
> >> > | Workqueue: events bpf_prog_free_deferred
> >> > | Call Trace:
> >> > | [<ffffffff80cbc444>] __schedule+0x2d0/0x940
> >> > | watchdog: BUG: soft lockup - CPU#0 stuck for 21s! [kworker/0:5:319]
> >> > | Modules linked in: nls_iso8859_1 drm fuse i2c_core drm_panel_orientation_quirks backlight dm_mod configfs ip_tables x_tables
> >> > | CPU: 0 PID: 319 Comm: kworker/0:5 Not tainted 6.5.0-rc5 #1
> >> > | Hardware name: riscv-virtio,qemu (DT)
> >> > | Workqueue: events bpf_prog_free_deferred
> >> > | epc : __sbi_rfence_v02_call.isra.0+0x74/0x11a
> >> > | ra : __sbi_rfence_v02+0xda/0x1a4
> >> > | epc : ffffffff8000ab4c ra : ffffffff8000accc sp : ff20000001c9bbd0
> >> > | gp : ffffffff82078c48 tp : ff600000888e6a40 t0 : ff20000001c9bd44
> >> > | t1 : 0000000000000000 t2 : 0000000000000040 s0 : ff20000001c9bbf0
> >> > | s1 : 0000000000000010 a0 : 0000000000000000 a1 : 0000000000000000
> >> > | a2 : 0000000000000000 a3 : 0000000000000000 a4 : 0000000000000000
> >> > | a5 : 0000000000000000 a6 : 0000000000000000 a7 : 0000000052464e43
> >> > | s2 : 000000000000ffff s3 : 00000000ffffffff s4 : ffffffff81667528
> >> > | s5 : 0000000000000000 s6 : 0000000000000000 s7 : 0000000000000000
> >> > | s8 : 0000000000000001 s9 : 0000000000000003 s10: 0000000000000040
> >> > | s11: ffffffff8207d240 t3 : 000000000000000f t4 : 000000000000002a
> >> > | t5 : ff600000872df140 t6 : ffffffff81e26828
> >> > | status: 0000000200000120 badaddr: 0000000000000000 cause: 8000000000000005
> >> > | [<ffffffff8000ab4c>] __sbi_rfence_v02_call.isra.0+0x74/0x11a
> >> > | [<ffffffff8000accc>] __sbi_rfence_v02+0xda/0x1a4
> >> > | [<ffffffff8000a886>] sbi_remote_fence_i+0x1e/0x26
> >> > | [<ffffffff8000cee2>] flush_icache_all+0x1a/0x48
> >> > | [<ffffffff80007736>] patch_text_nosync+0x6c/0x8c
> >> > | [<ffffffff8000f0f8>] bpf_arch_text_invalidate+0x62/0xac
> >> > | [<ffffffff8016c538>] bpf_prog_pack_free+0x9c/0x1b2
> >> > | [<ffffffff8016c84a>] bpf_jit_binary_pack_free+0x20/0x4a
> >> > | [<ffffffff8000f198>] bpf_jit_free+0x56/0x9e
> >> > | [<ffffffff8016b43a>] bpf_prog_free_deferred+0x15a/0x182
> >> > | [<ffffffff800576c4>] process_one_work+0x1b6/0x3d6
> >> > | [<ffffffff80057d52>] worker_thread+0x84/0x378
> >> > | [<ffffffff8005fc2c>] kthread+0xe8/0x108
> >> > | [<ffffffff80003ffa>] ret_from_fork+0xe/0x20
> >> >
> >> > I'm digging into that now, and I would appreciate if you could run the
> >> > test_tag on VF2 or similar (I'm missing that HW).
> >> >
> >> > It seems like we're hitting a bug with this series, so let's try to
> >> > figure out where the problems is, prior merging it.
> >>
> >> Hmm, it looks like the bpf_arch_text_invalidate() implementation is a
> >> bit problematic:
> >>
> >> +int bpf_arch_text_invalidate(void *dst, size_t len)
> >> +{
> >> + __le32 *ptr;
> >> + int ret = 0;
> >> + u32 inval = 0;
> >> +
> >> + for (ptr = dst; ret == 0 && len >= sizeof(u32); len -= sizeof(u32)) {
> >> + mutex_lock(&text_mutex);
> >> + ret = patch_text_nosync(ptr++, &inval, sizeof(u32));
> >> + mutex_unlock(&text_mutex);
> >> + }
> >> +
> >> + return ret;
> >> +}
> >>
> >> Each patch_text_nosync() is a remote fence.i, and for a big "len", we'll
> >> be flooded with remote fences.
> >
> > I understand this now, thanks for debugging this.
> >
> > We are calling patch_text_nosync() for each word (u32) which calls
> > flush_icache_range() and therefore "fence.i" is inserted after every
> > word.
>
> But more importantly, it does a remote fence.i (which is an IPI to all
> cores).
>
> > I still don't fully understand how it causes this bug because I lack
> > the prerequisite
> > knowledge about test_tag and what the failing test is doing.
>
> The test_tag is part of kselftest/bpf:
> tools/testing/selftests/bpf/test_tag.c
>
> TL;DR: it generates a bunch of programs, where some have a length of,
> e.g, 41024. bpf_arch_text_invalidate() does ~10k of remote fences in
> that case.
>
> > But to solve this issue we would need a function like the x86
> > text_poke_set() that will only
> > insert a single "fence.i" after setting the whole memory area. This
> > can be done by
> > implementing a wrapper around patch_insn_write() which would set the memory area
> > and at the end call flush_icache_range().
> >
> > Something like:
> >
> > void *text_set_nosync(void *dst, int c, size_t len)
> > {
> > __le32 *ptr;
> > int ret = 0;
> >
> > for (ptr = dst; ret == 0 && len >= sizeof(u32); len -= sizeof(u32)) {
> > ret = patch_insn_write(ptr++, &c, sizeof(u32));
> > }
> > if(!ret)
> > flush_icache_range((uintptr_t) dst, (uintptr_t) dst + len);
> >
> > return ret;
> > }
> >
> > Let me know if this looks correct or we need more details here.
> > I will then send v2 with this implemented as a separate patch.
>
> Can't we do better here? Perhaps a similar pattern like the 2 page fill?
> Otherwise we'll have a bunch of fixmap updates as well.
I agree that we can make it more efficient by first copying the value to a
RW buffer using normal memcpy() and then copying that area to the RO area
using patch_insn_write(). Then it would solve both problems. Or we implement
a new function like patch_insn_write() that does the 2 page map and
set explicitly.
Which approach would you prefer?
1) Wrapper around patch_insn_write() that first memsets a RW buffer and then
copies the complete RW buffer to the RO area by calling
patch_insn_write() with len.
2) A new function like patch_insn_write() that takes dst, src, len and
maps the dst, 2 pages
at a time and sets it to *src in a loop.
>
> I'd keep the patch_ prefix in the name for consistency. Please measure
> the runtime of test_tag pre/after the change.
test_tag currently wouldn't even complete right? with the current
version of the patch?
>
> I don't know if your arm64 work has similar problems?
Thanks for bringing this up. I will revisit that and verify if
test_tag is working
there. There also the bpf_arch_text_invalidate() is calling
aarch64_insn_patch_text_nosync()
in a loop that in turn calls caches_clean_inval_pou(). So I might see
similar issues there.
I think https://github.com/kernel-patches doesn't run test_tag hence I
might have missed it.
>
>
> Björn
Thanks,
Puranjay