On 8/28/24 4:44 PM, Alexei Starovoitov wrote:
On Wed, Aug 28, 2024 at 3:47 PM Yonghong Song <yonghong.song@xxxxxxxxx> wrote:
It can be je/je too, no?
Yes. It is possible.
so 128 - 4 instead of 128 - 3 ?
You probably mean "127 - 4 instead of 127 - 3" since
the maximum value is 127.
Yes, of course :)
I checked 127 - 4 = 0x7c and indeed we should. See below examples:
20e: 48 85 ff test rdi,rdi
211: XX XX je 0x291
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28d: XX XX XX XX XX XX je 0x212
293: bf 03 00 00 00 mov edi,0x3
=>
20e: 48 85 ff test rdi,rdi
211: XX XX XX XX XX XX je 0x297 (0x293 - 0x213)
217: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
291: XX XX je 0x217 (0x217 - 0x293)
293: bf 03 00 00 00 mov edi,0x3
=>
20e: 48 85 ff test rdi,rdi
211: XX XX je 0x28f (0x293 - 0x217)
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28d: XX XX je 0x213 (0x213 - 0x293) // -0x80 allowed
293: bf 03 00 00 00 mov edi,0x3
=>
20e: 48 85 ff test rdi,rdi
211: XX XX XX XX XX XX je 0x28f (0x293 - 0x213)
217: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
291: XX XX je 0x217 (0x217 - 0x293)
293: bf 03 00 00 00 mov edi,0x3
=>
...
Here 0x293 - 0x217 = 0x7c
How did you craft such a test?
Can we add it as a selftest somehow?
This is not from a complete test. I assumed some state during convergence
and from there to further derive states. But I will try to see whether
I can construct actual test cases or not.
+static bool is_imm8_cond_offset(int value)
+{
+ return value <= 124 && value >= -128;
the other side needs the same treatment, no ?
good question. From my understanding, the non-convergence in the
above needs both forward and backport conditions. The solution we
are using is based on putting a limitation on forward conditions
w.r.t. jit code gen.
Another solution is actually to put a limitation on backward
conditions. For example, let us say the above is_imm8_cond_offset()
has
return value <= 127 && value > -124
See below example:
20e: 48 85 ff test rdi,rdi
211: XX XX je 0x291
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28d: XX XX XX XX XX XX je 0x212
293: bf 03 00 00 00 mov edi,0x3
=>
20e: 48 85 ff test rdi,rdi
211: XX XX XX XX XX XX je 0x297 (0x293 - 0x213)
217: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
291: XX XX XX XX XX XX je 0x21b (0x217 - 0x293)
297: bf 03 00 00 00 mov edi,0x3
=>
20e: 48 85 ff test rdi,rdi
211: XX XX XX XX XX XX je 0x297 (0x297 - 0x217)
217: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
291: XX XX XX XX XX XX je 0x217 (0x217 - 0x297)
297: bf 03 00 00 00 mov edi,0x3
converged here.
So I think we do not need to limit both sides. One side should be enough.
I see and agree when both sides are je/je.
What if the earlier one is a jmp ?
Then we can hit:
if (nops != 0 && nops != 3) {
pr_err("unexpected jump padding: %d bytes\n",
nops);
?
So one side of "jmp_cond padding" and the same side in "jump padding"
needs to do it?
I did some further experiments with pattern like
jmp <-> je
and
jmp <-> jmp
The below is the illustration (not from a complete test):
================
211: XX XX jmp 0x291
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28d: XX XX XX XX XX XX je 0x212
293: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX XX XX XX jmp (0x293 - 0x213)
216: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
291: XX XX je (0x216 - 0x293)
293: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX jmp (0x293 - 0x216 = 0x7d)
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28d: XX XX je (0x213 - 0x293)
293: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX XX XX XX jmp (0x293 - 0x213)
216: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
291: XX XX je (0x216 - 0x293)
293: bf 03 00 00 00 mov edi,0x3
=>
...
not converged!
================
211: XX XX jmp 0x291
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28c: XX XX XX XX XX XX je 0x212
292: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX jmp (0x292 - 0x213)
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28c: XX XX je (0x213 - 0x292)
28e: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX jmp (0x28e - 0x213 = 0x7b)
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28c: XX XX je (0x213 - 0x28e)
28e: bf 03 00 00 00 mov edi,0x3
converged!
=================
211: XX XX jmp 0x291
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28e: XX XX XX XX XX jmp 0x212
293: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX XX XX XX jmp (0x293 - 0x213)
216: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
292: XX XX jmp (0x216 - 0x293)
294: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX jmp (0x294 - 0x216 = 0x7e)
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28e: XX XX XX XX XX jmp (0x213 - 0x294)
294: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX jmp (0x294 - 0x216 = 0x7e)
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28e: XX XX XX XX XX jmp (0x213 - 0x294)
294: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX XX XX XX jmp (0x294 - 0x213)
216: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
292: XX XX jmp (0x216 - 0x294)
294: bf 03 00 00 00 mov edi,0x3
=>
...
no converged!
===================================
211: XX XX jmp 0x291
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28d: XX XX XX XX XX jmp 0x212
292: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX jmp (0x292 - 0x213)
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28d: XX XX jmp (0x213 - 0x292)
290: bf 03 00 00 00 mov edi,0x3
=>
211: XX XX jmp (0x290 - 0x213 = 0x7d)
213: 48 8b 77 00 mov rsi,QWORD PTR [rdi+0x0]
...
28d: XX XX jmp (0x213 - 0x290)
290: bf 03 00 00 00 mov edi,0x3
converged!
So I emulated je <-> je, je <-> jmp, jmp <-> je and jmp <-> jmp.
So we need to apply the same checking is_imm8_cond_offset() to jmp insn.
This should cover all cases.
Hitting the following
if (nops != 0 && nops != 3) {
pr_err("unexpected jump padding: %d bytes\n",
nops);
is not due to the above illustration with 'jmp' insn as indeed
its insn length changes with 0 or 3. But it is due to some jmp/cond_jmp
insn inside je/jmp <-> je/jmp.
