You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Eduard Zingerman says:
====================
bpf: propagate read/precision marks over state graph backedges
Current loop_entry-based states comparison logic does not handle the
following case:
.-> A --. Assume the states are visited in the order A, B, C.
| | | Assume that state B reaches a state equivalent to state A.
| v v At this point, state C is not processed yet, so state A
'-- B C has not received any read or precision marks from C.
As a result, these marks won't be propagated to B.
If B has incomplete marks, it is unsafe to use it in states_equal()
checks. This issue was first reported in [1].
This patch-set
--------------
Here is the gist of the algorithm implemented by this patch-set:
- Compute strongly connected components (SCCs) in the program CFG.
- When a verifier state enters an SCC, that state is recorded as the
SCC's entry point.
- When a verifier state is found to be equivalent to another
(e.g., B to A in the example above), it is recorded as a
states-graph backedge.
- Backedges are accumulated per SCC (*).
- When an SCC entry state reaches `branches == 0`, propagate read and
precision marks through the backedges until a fixed point is reached
(e.g., from A to B, from C to A, and then again from A to B).
(*) This is an oversimplification, see patch #8 for details.
Unfortunately, this means that commit [2] needs to be reverted,
as precision propagation requires access to jump history,
and backedges represent history not belonging to `env->cur_state`.
Details are provided in patch #8; a comment in `is_state_visited()`
explains most of the mechanics.
Patch #2 adds a `compute_scc()` function, which computes SCCs in the
program CFG. This function was tested using property-based testing in
[3], but it is not included in selftests.
Previous attempt
----------------
A previous attempt to fix this is described in [4]:
1. Within the states loop, `states_equal(... RANGE_WITHIN)` ignores
read and precision marks.
2. For states outside the loop, all registers for states within the
loop are marked as read and precise.
This approach led to an 86x regression on the `cond_break1` selftest.
In that test, one loop was followed by another, and a certain variable
was incremented in the second loop. This variable was marked as
precise due to rule (2), which hindered convergence in the first loop.
After some off-list discussion, it was decided that this might be a
typical case and such regressions are undesirable.
This patch-set avoids such eager precision markings.
Alternatives
------------
Another option is to associate a mask of read/written/precise stack
slots with each instruction. This mask can be populated during
verifier states exploration. Upon reaching an `EXIT` instruction or an
equivalent state, the accumulated masks can be used to propagate
read/written/precise bits across the program's control flow graph
using an analysis similar to use-def.
Unfortunately, a naive implementation of this approach [5] results in
a 10x regression in `veristat` for some `sched_ext` programs due to
the inability to express the must-write property. This issue requires
further investigation.
Changes in verification performance
-----------------------------------
There are some veristat regressions when comparing with master using
selftests and sched_ext BPF binaries. The comparison is done using
master from [6] and this patch-set from [7] where memory accounting
logic is added to veristat.
========= selftests: master vs patch-set =========
File Program Insns Peak memory (KiB)
--------------------- ----------------------------------- ----- ----- ---------------- ---- ----- ----------------
bpf_qdisc_fq.bpf.o bpf_fq_dequeue 1187 1645 +458 (+38.58%) 768 1240 +472 (+61.46%)
dynptr_success.bpf.o test_copy_from_user_str_dynptr 208 279 +71 (+34.13%) 512 1024 +512 (+100.00%)
dynptr_success.bpf.o test_copy_from_user_task_str_dynptr 205 263 +58 (+28.29%) 512 1024 +512 (+100.00%)
dynptr_success.bpf.o test_probe_read_kernel_str_dynptr 686 857 +171 (+24.93%) 992 1724 +732 (+73.79%)
dynptr_success.bpf.o test_probe_read_user_str_dynptr 689 860 +171 (+24.82%) 1016 1744 +728 (+71.65%)
iters.bpf.o checkpoint_states_deletion 1211 1216 +5 (+0.41%) 512 1280 +768 (+150.00%)
pyperf600_iter.bpf.o on_event 2591 5929 +3338 (+128.83%) 4744 11176 +6432 (+135.58%)
verifier_gotol.bpf.o gotol_large_imm 40004 40004 +0 (+0.00%) 1024 1536 +512 (+50.00%)
Total progs: 3725
Old success: 2157
New success: 2157
total_insns diff min: 0.00%
total_insns diff max: 128.83%
0 -> value: 0
value -> 0: 0
total_insns abs max old: 837,487
total_insns abs max new: 837,487
0 .. 5 %: 3710
5 .. 15 %: 6
20 .. 30 %: 6
30 .. 40 %: 2
125 .. 130 %: 1
mem_peak diff min: -27.78%
mem_peak diff max: 198.44%
mem_peak abs max old: 269,312 KiB
mem_peak abs max new: 269,312 KiB
-30 .. -20 %: 1
-5 .. 0 %: 18
0 .. 5 %: 3568
5 .. 15 %: 4
15 .. 25 %: 3
45 .. 55 %: 4
60 .. 70 %: 1
70 .. 80 %: 2
100 .. 110 %: 3
135 .. 145 %: 1
150 .. 160 %: 1
195 .. 200 %: 1
========= scx: master vs patch-set =========
Program Insns Peak memory (KiB)
------------------------ ----- ----- --------------- ----- ----- -----------------
arena_topology_node_init 2133 2395 +262 (+12.28%) 768 768 +0 (+0.00%)
chaos_dispatch 2835 2868 +33 (+1.16%) 1972 1720 -252 (-12.78%)
chaos_init 4324 5210 +886 (+20.49%) 2528 3028 +500 (+19.78%)
lavd_cpu_offline 5107 5726 +619 (+12.12%) 4188 6304 +2116 (+50.53%)
lavd_cpu_online 5107 5726 +619 (+12.12%) 4188 6304 +2116 (+50.53%)
lavd_dispatch 41775 47601 +5826 (+13.95%) 6196 29192 +22996 (+371.14%)
lavd_enqueue 20238 24188 +3950 (+19.52%) 22084 42156 +20072 (+90.89%)
lavd_init 6974 7685 +711 (+10.20%) 5428 6928 +1500 (+27.63%)
lavd_select_cpu 22138 26088 +3950 (+17.84%) 24448 43688 +19240 (+78.70%)
layered_dispatch 17847 26581 +8734 (+48.94%) 11728 28740 +17012 (+145.05%)
layered_dump 1891 2098 +207 (+10.95%) 2036 3048 +1012 (+49.71%)
layered_runnable 2606 2634 +28 (+1.07%) 748 1240 +492 (+65.78%)
p2dq_init 3691 4554 +863 (+23.38%) 2016 2528 +512 (+25.40%)
rusty_enqueue 28853 28853 +0 (+0.00%) 2072 1824 -248 (-11.97%)
rusty_init_task 31128 31128 +0 (+0.00%) 2176 2560 +384 (+17.65%)
Total progs: 148
Old success: 135
New success: 135
total_insns diff min: 0.00%
total_insns diff max: 48.94%
0 -> value: 0
value -> 0: 0
total_insns abs max old: 41,775
total_insns abs max new: 47,601
0 .. 5 %: 133
5 .. 15 %: 7
15 .. 25 %: 4
35 .. 45 %: 3
45 .. 50 %: 1
mem_peak diff min: -12.78%
mem_peak diff max: 371.14%
mem_peak abs max old: 24,448 KiB
mem_peak abs max new: 43,688 KiB
-15 .. -5 %: 2
-5 .. 0 %: 2
0 .. 5 %: 129
5 .. 15 %: 1
15 .. 25 %: 2
25 .. 35 %: 2
45 .. 55 %: 3
65 .. 75 %: 1
75 .. 85 %: 1
90 .. 100 %: 1
145 .. 155 %: 1
195 .. 205 %: 1
370 .. 375 %: 1
Changelog
---------
v1: https://lore.kernel.org/bpf/[email protected]/
v1 -> v2:
- Rebase
- added mem_peak statistics (Alexei)
- selftests: fixed comments and removed useless r7 assignments (Yonghong)
v2: https://lore.kernel.org/bpf/[email protected]/
v2 -> v3:
- Rebase
Links
-----
[1] https://lore.kernel.org/bpf/[email protected]/
[2] commit 96a30e4 ("bpf: use common instruction history across all states")
[3] https://github.com/eddyz87/scc-test
[4] https://lore.kernel.org/bpf/[email protected]/
[5] https://github.com/eddyz87/bpf/tree/propagate-read-and-precision-in-cfg
[6] https://github.com/eddyz87/bpf/tree/veristat-memory-accounting
[7] https://github.com/eddyz87/bpf/tree/scc-accumulate-backedges
====================
Link: https://patch.msgid.link/[email protected]
Signed-off-by: Alexei Starovoitov <[email protected]>
0 commit comments