Am Mittwoch, dem 14.05.2025 um 12:23 +0200 schrieb Johannes Berg:
> + linux-wireless
>
> On Wed, 2025-05-14 at 09:32 +0000, Bert Karwatzki wrote:
>
> > Then I reapplied commit 76a853f86c97 hunk by hunk and found the one hunk that
> > causes the problem:
> >
> > diff --git a/net/mac80211/tx.c b/net/mac80211/tx.c
> > index 3e751dd3ae7b..63df21228029 100644
> > --- a/net/mac80211/tx.c
> > +++ b/net/mac80211/tx.c
> > @@ -4648,8 +4648,7 @@ static void ieee80211_8023_xmit(struct
> > ieee80211_sub_if_data *sdata,
> > memcpy(IEEE80211_SKB_CB(seg), info, sizeof(*info));
> > }
> >
> > - if (unlikely(skb->sk &&
> > - skb_shinfo(skb)->tx_flags & SKBTX_WIFI_STATUS)) {
> > + if (unlikely(skb->sk && sock_flag(skb->sk, SOCK_WIFI_STATUS))) {
> > info->status_data = ieee80211_store_ack_skb(local, skb,
> > &info->flags, NULL);
> > if (info->status_data)
>
> I think it crashed later on the status, but this inserts the skb into
> the IDR so the status can pick it up to return the status and afaict
> _that's_ where it crashed.
>
> Still I don't really know what could go wrong? The (copied) skb should
> still have been keeping the socket alive.
>
> > This is enough to cause a kernel panic when compiled with clang (clang-19.1.7
> > from debian sid). Compiling the same kernel with gcc (gcc-14.2.0 from debian
> > sid) shows no problem.
>
> Right, even stranger. But I can't even say you should look at this place
> (which inserts) or the other (which takes it out again and crashed) to
> compare the code :-/
>
>
> johannes
I've split off the problematic piece of code into an noinline function to simplify the disassembly:
diff --git a/net/mac80211/tx.c b/net/mac80211/tx.c
index 20de6e6b0929..075e012d9992 100644
--- a/net/mac80211/tx.c
+++ b/net/mac80211/tx.c
@@ -4582,7 +4582,19 @@ static bool ieee80211_tx_8023(struct ieee80211_sub_if_data *sdata,
return ret;
}
-static noinline void ieee80211_8023_xmit(struct ieee80211_sub_if_data *sdata,
+static noinline void ieee80211_8023_xmit_clang_debug_helper(struct sk_buff *skb,
+ struct ieee80211_local *local,
+ struct ieee80211_tx_info *info)
+{
+ if (unlikely(skb->sk && sock_flag(skb->sk, SOCK_WIFI_STATUS))) {
+ info->status_data = ieee80211_store_ack_skb(local, skb,
+ &info->flags, NULL);
+ if (info->status_data)
+ info->status_data_idr = 1;
+ }
+}
+
+static void ieee80211_8023_xmit(struct ieee80211_sub_if_data *sdata,
struct net_device *dev, struct sta_info *sta,
struct ieee80211_key *key, struct sk_buff *skb)
{
@@ -4648,12 +4660,7 @@ static noinline void ieee80211_8023_xmit(struct ieee80211_sub_if_data *sdata,
memcpy(IEEE80211_SKB_CB(seg), info, sizeof(*info));
}
- if (unlikely(skb->sk && sock_flag(skb->sk, SOCK_WIFI_STATUS))) {
- info->status_data = ieee80211_store_ack_skb(local, skb,
- &info->flags, NULL);
- if (info->status_data)
- info->status_data_idr = 1;
- }
+ ieee80211_8023_xmit_clang_debug_helper(skb, local, info);
dev_sw_netstats_tx_add(dev, skbs, len);
sta->deflink.tx_stats.packets[queue] += skbs;
This shows the the behaviour as the old code, i.e. kernel panic when compiled with clang(-19.1.7), no problem when
compiled with gcc(-14.2.0).
When compiled with clang the disassembly of the function is (from objdump -d)
000000000000a260 <ieee80211_8023_xmit_clang_debug_helper>:
a260: 48 8b 47 18 mov 0x18(%rdi),%rax
a264: 48 85 c0 test %rax,%rax
a267: 74 0c je a275 <ieee80211_8023_xmit_clang_debug_helper+0x15>
a269: 53 push %rbx
a26a: 48 f7 40 60 00 00 08 testq $0x80000,0x60(%rax)
a271: 00
a272: 75 07 jne a27b <ieee80211_8023_xmit_clang_debug_helper+0x1b>
a274: 5b pop %rbx
a275: 2e e9 00 00 00 00 cs jmp a27b <ieee80211_8023_xmit_clang_debug_helper+0x1b>
a27b: 48 89 f8 mov %rdi,%rax
a27e: 48 89 f7 mov %rsi,%rdi
a281: 48 89 c6 mov %rax,%rsi
a284: 48 89 d3 mov %rdx,%rbx
a287: 31 c9 xor %ecx,%ecx
a289: e8 02 ff ff ff call a190 <ieee80211_store_ack_skb>
a28e: 25 ff 1f 00 00 and $0x1fff,%eax
a293: 89 c2 mov %eax,%edx
a295: b9 0f 00 fe ff mov $0xfffe000f,%ecx
a29a: 23 4b 04 and 0x4(%rbx),%ecx
a29d: c1 e2 04 shl $0x4,%edx
a2a0: 09 d1 or %edx,%ecx
a2a2: 89 4b 04 mov %ecx,0x4(%rbx)
a2a5: 85 c0 test %eax,%eax
a2a7: 74 cb je a274 <ieee80211_8023_xmit_clang_debug_helper+0x14>
a2a9: 83 c9 08 or $0x8,%ecx
a2ac: 89 4b 04 mov %ecx,0x4(%rbx)
a2af: eb c3 jmp a274 <ieee80211_8023_xmit_clang_debug_helper+0x14>
a2b1: 66 66 66 66 66 66 2e data16 data16 data16 data16 data16 cs nopw 0x0(%rax,%rax,1)
a2b8: 0f 1f 84 00 00 00 00
a2bf: 00
When compiled with gcc the disassembly is
00000000000010e0 <ieee80211_8023_xmit_clang_debug_helper>:
10e0: 48 8b 4f 18 mov 0x18(%rdi),%rcx
10e4: 48 89 f8 mov %rdi,%rax
10e7: 48 85 c9 test %rcx,%rcx
10ea: 75 05 jne 10f1 <ieee80211_8023_xmit_clang_debug_helper+0x11>
10ec: e9 00 00 00 00 jmp 10f1 <ieee80211_8023_xmit_clang_debug_helper+0x11>
10f1: 48 8b 49 60 mov 0x60(%rcx),%rcx
10f5: f7 c1 00 00 08 00 test $0x80000,%ecx
10fb: 74 ef je 10ec <ieee80211_8023_xmit_clang_debug_helper+0xc>
10fd: 48 83 ec 08 sub $0x8,%rsp
1101: 48 89 f7 mov %rsi,%rdi
1104: 31 c9 xor %ecx,%ecx
1106: 48 89 c6 mov %rax,%rsi
1109: 48 89 14 24 mov %rdx,(%rsp)
110d: e8 ce f8 ff ff call 9e0 <ieee80211_store_ack_skb>
1112: 48 8b 14 24 mov (%rsp),%rdx
1116: 89 c1 mov %eax,%ecx
1118: 8b 42 04 mov 0x4(%rdx),%eax
111b: 81 e1 ff 1f 00 00 and $0x1fff,%ecx
1121: c1 e1 04 shl $0x4,%ecx
1124: 25 0f 00 fe ff and $0xfffe000f,%eax
1129: 09 c8 or %ecx,%eax
112b: 89 42 04 mov %eax,0x4(%rdx)
112e: a9 f0 ff 01 00 test $0x1fff0,%eax
1133: 74 04 je 1139 <ieee80211_8023_xmit_clang_debug_helper+0x59>
1135: 80 4a 04 08 orb $0x8,0x4(%rdx)
1139: 48 83 c4 08 add $0x8,%rsp
113d: e9 00 00 00 00 jmp 1142 <ieee80211_8023_xmit_clang_debug_helper+0x62>
1142: 66 66 2e 0f 1f 84 00 data16 cs nopw 0x0(%rax,%rax,1)
1149: 00 00 00 00
114d: 0f 1f 00 nopl (%rax)
1150: 90 nop
1151: 90 nop
1152: 90 nop
1153: 90 nop
1154: 90 nop
1155: 90 nop
1156: 90 nop
1157: 90 nop
1158: 90 nop
1159: 90 nop
115a: 90 nop
115b: 90 nop
115c: 90 nop
115d: 90 nop
115e: 90 nop
115f: 90 nop
I've not yet taken a closer look, but perhaps the error is obvious
for some one else.
Bert Karwatzki
