Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
bpf: support precision propagation in the presence of subprogs
Add support precision backtracking in the presence of subprogram frames in
jump history.
This means supporting a few different kinds of subprogram invocation
situations, all requiring a slightly different handling in precision
backtracking handling logic:
- static subprogram calls;
- global subprogram calls;
- callback-calling helpers/kfuncs.
For each of those we need to handle a few precision propagation cases:
- what to do with precision of subprog returns (r0);
- what to do with precision of input arguments;
- for all of them callee-saved registers in caller function should be
propagated ignoring subprog/callback part of jump history.
N.B. Async callback-calling helpers (currently only
bpf_timer_set_callback()) are transparent to all this because they set
a separate async callback environment and thus callback's history is not
shared with main program's history. So as far as all the changes in this
commit goes, such helper is just a regular helper.
Let's look at all these situation in more details. Let's start with
static subprogram being called, using an exxerpt of a simple main
program and its static subprog, indenting subprog's frame slightly to
make everything clear.
frame 0 frame 1 precision set
======= ======= =============
9: r6 = 456;
10: r1 = 123; r6
11: call pc+10; r1, r6
22: r0 = r1; r1
23: exit r0
12: r1 = <map_pointer> r0, r6
13: r1 += r0; r0, r6
14: r1 += r6; r6;
15: exit
As can be seen above main function is passing 123 as single argument to
an identity (`return x;`) subprog. Returned value is used to adjust map
pointer offset, which forces r0 to be marked as precise. Then
instruction #14 does the same for callee-saved r6, which will have to be
backtracked all the way to instruction #9. For brevity, precision sets
for instruction #13 and #14 are combined in the diagram above.
First, for subprog calls, r0 returned from subprog (in frame 0) has to
go into subprog's frame 1, and should be cleared from frame 0. So we go
back into subprog's frame knowing we need to mark r0 precise. We then
see that insn #22 sets r0 from r1, so now we care about marking r1
precise. When we pop up from subprog's frame back into caller at
insn #11 we keep r1, as it's an argument-passing register, so we eventually
find `10: r1 = 123;` and satify precision propagation chain for insn #13.
This example demonstrates two sets of rules:
- r0 returned after subprog call has to be moved into subprog's r0 set;
- *static* subprog arguments (r1-r5) are moved back to caller precision set.
Let's look at what happens with callee-saved precision propagation. Insn #14
mark r6 as precise. When we get into subprog's frame, we keep r6 in
frame 0's precision set *only*. Subprog itself has its own set of
independent r6-r10 registers and is not affected. When we eventually
made our way out of subprog frame we keep r6 in precision set until we
reach `9: r6 = 456;`, satisfying propagation. r6-r10 propagation is
perhaps the simplest aspect, it always stays in its original frame.
That's pretty much all we have to do to support precision propagation
across *static subprog* invocation.
Let's look at what happens when we have global subprog invocation.
frame 0 frame 1 precision set
======= ======= =============
9: r6 = 456;
10: r1 = 123; r6
11: call pc+10; # global subprog r6
12: r1 = <map_pointer> r0, r6
13: r1 += r0; r0, r6
14: r1 += r6; r6;
15: exit
Starting from insn #13, r0 has to be precise. We backtrack all the way
to insn #11 (call pc+10) and see that subprog is global, so was already
validated in isolation. As opposed to static subprog, global subprog
always returns unknown scalar r0, so that satisfies precision
propagation and we drop r0 from precision set. We are done for insns #13.
Now for insn #14. r6 is in precision set, we backtrack to `call pc+10;`.
Here we need to recognize that this is effectively both exit and entry
to global subprog, which means we stay in caller's frame. So we carry on
with r6 still in precision set, until we satisfy it at insn #9. The only
hard part with global subprogs is just knowing when it's a global func.
Lastly, callback-calling helpers and kfuncs do simulate subprog calls,
so jump history will have subprog instructions in between caller
program's instructions, but the rules of propagating r0 and r1-r5
differ, because we don't actually directly call callback. We actually
call helper/kfunc, which at runtime will call subprog, so the only
difference between normal helper/kfunc handling is that we need to make
sure to skip callback simulatinog part of jump history.
Let's look at an example to make this clearer.
frame 0 frame 1 precision set
======= ======= =============
8: r6 = 456;
9: r1 = 123; r6
10: r2 = &callback; r6
11: call bpf_loop; r6
22: r0 = r1;
23: exit
12: r1 = <map_pointer> r0, r6
13: r1 += r0; r0, r6
14: r1 += r6; r6;
15: exit
Again, insn #13 forces r0 to be precise. As soon as we get to `23: exit`
we see that this isn't actually a static subprog call (it's `call
bpf_loop;` helper call instead). So we clear r0 from precision set.
For callee-saved register, there is no difference: it stays in frame 0's
precision set, we go through insn #22 and #23, ignoring them until we
get back to caller frame 0, eventually satisfying precision backtrack
logic at insn #8 (`r6 = 456;`).
Assuming callback needed to set r0 as precise at insn #23, we'd
backtrack to insn #22, switching from r0 to r1, and then at the point
when we pop back to frame 0 at insn #11, we'll clear r1-r5 from
precision set, as we don't really do a subprog call directly, so there
is no input argument precision propagation.
That's pretty much it. With these changes, it seems like the only still
unsupported situation for precision backpropagation is the case when
program is accessing stack through registers other than r10. This is
still left as unsupported (though rare) case for now.
As for results. For selftests, few positive changes for bigger programs,
cls_redirect in dynptr variant benefitting the most:
[vmuser@archvm bpf]$ ./veristat -C ~/subprog-precise-before-results.csv ~/subprog-precise-after-results.csv -f @veristat.cfg -e file,prog,insns -f 'insns_diff!=0'
File Program Insns (A) Insns (B) Insns (DIFF)
---------------------------------------- ------------- --------- --------- ----------------
pyperf600_bpf_loop.bpf.linked1.o on_event 2060 2002 -58 (-2.82%)
test_cls_redirect_dynptr.bpf.linked1.o cls_redirect 15660 2914 -12746 (-81.39%)
test_cls_redirect_subprogs.bpf.linked1.o cls_redirect 61620 59088 -2532 (-4.11%)
xdp_synproxy_kern.bpf.linked1.o syncookie_tc 109980 86278 -23702 (-21.55%)
xdp_synproxy_kern.bpf.linked1.o syncookie_xdp 97716 85147 -12569 (-12.86%)
Cilium progress don't really regress. They don't use subprogs and are
mostly unaffected, but some other fixes and improvements could have
changed something. This doesn't appear to be the case:
[vmuser@archvm bpf]$ ./veristat -C ~/subprog-precise-before-results-cilium.csv ~/subprog-precise-after-results-cilium.csv -e file,prog,insns -f 'insns_diff!=0'
File Program Insns (A) Insns (B) Insns (DIFF)
------------- ------------------------------ --------- --------- ------------
bpf_host.o tail_nodeport_nat_ingress_ipv6 4983 5003 +20 (+0.40%)
bpf_lxc.o tail_nodeport_nat_ingress_ipv6 4983 5003 +20 (+0.40%)
bpf_overlay.o tail_nodeport_nat_ingress_ipv6 4983 5003 +20 (+0.40%)
bpf_xdp.o tail_handle_nat_fwd_ipv6 12475 12504 +29 (+0.23%)
bpf_xdp.o tail_nodeport_nat_ingress_ipv6 6363 6371 +8 (+0.13%)
Looking at (somewhat anonymized) Meta production programs, we see mostly
insignificant variation in number of instructions, with one program
(syar_bind6_protect6) benefitting the most at -17%.
[vmuser@archvm bpf]$ ./veristat -C ~/subprog-precise-before-results-fbcode.csv ~/subprog-precise-after-results-fbcode.csv -e prog,insns -f 'insns_diff!=0'
Program Insns (A) Insns (B) Insns (DIFF)
------------------------ --------- --------- ----------------
on_request_context_event 597 585 -12 (-2.01%)
read_async_py_stack 43789 43657 -132 (-0.30%)
read_sync_py_stack 35041 37599 +2558 (+7.30%)
rrm_usdt 946 940 -6 (-0.63%)
sysarmor_inet6_bind 28863 28249 -614 (-2.13%)
sysarmor_inet_bind 28845 28240 -605 (-2.10%)
syar_bind4_protect4 154145 147640 -6505 (-4.22%)
syar_bind6_protect6 165242 137088 -28154 (-17.04%)
syar_task_exit_setgid 21289 19720 -1569 (-7.37%)
syar_task_exit_setuid 21290 19721 -1569 (-7.37%)
do_uprobe 19967 19413 -554 (-2.77%)
tw_twfw_ingress 215877 204833 -11044 (-5.12%)
tw_twfw_tc_in 215877 204833 -11044 (-5.12%)
But checking duration (wall clock) differences, that is the actual time taken
by verifier to validate programs, we see a sometimes dramatic improvements, all
the way to about 16x improvements:
[vmuser@archvm bpf]$ ./veristat -C ~/subprog-precise-before-results-meta.csv ~/subprog-precise-after-results-meta.csv -e prog,duration -s duration_diff^ | head -n20
Program Duration (us) (A) Duration (us) (B) Duration (us) (DIFF)
---------------------------------------- ----------------- ----------------- --------------------
tw_twfw_ingress 4488374 272836 -4215538 (-93.92%)
tw_twfw_tc_in 4339111 268175 -4070936 (-93.82%)
tw_twfw_egress 3521816 270751 -3251065 (-92.31%)
tw_twfw_tc_eg 3472878 284294 -3188584 (-91.81%)
balancer_ingress 343119 291391 -51728 (-15.08%)
syar_bind6_protect6 78992 64782 -14210 (-17.99%)
ttls_tc_ingress 11739 8176 -3563 (-30.35%)
kprobe__security_inode_link 13864 11341 -2523 (-18.20%)
read_sync_py_stack 21927 19442 -2485 (-11.33%)
read_async_py_stack 30444 28136 -2308 (-7.58%)
syar_task_exit_setuid 10256 8440 -1816 (-17.71%)
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>- Loading branch information
