[PATCH 0/3] LoongArch: Fix BPF trampoline related issues

[PATCH 0/3] LoongArch: Fix BPF trampoline related issues

[Hengqi Chen]

The following two selftest cases triggers oops on LoongArch:

    $ ./test_progs -a ns_bpf_qdisc -a tracing_struct

This small series tries to fix/workaround these issues.
See individual commit for details.

While at it, remove a duplicated flags check in __arch_prepare_bpf_trampoline().

Hengqi Chen (3):
  LoongArch: BPF: Remove duplicated flags check
  LoongArch: BPF: Sign extend struct ops return values properly
  LoongArch: BPF: No support of struct argument in trampoline programs

 arch/loongarch/net/bpf_jit.c | 54 +++++++++++++++++++++++++++++-------
 1 file changed, 44 insertions(+), 10 deletions(-)

--
2.43.5

[PATCH 1/3] LoongArch: BPF: Remove duplicated flags check

[Hengqi Chen]

The check for (BPF_TRAMP_F_ORIG_STACK | BPF_TRAMP_F_SHARE_IPMODIFY)
is duplicated in __arch_prepare_bpf_trampoline(). Remove it.

Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
---
 arch/loongarch/net/bpf_jit.c | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
index abfdb6bb5c38..b646c6b73014 100644
--- a/arch/loongarch/net/bpf_jit.c
+++ b/arch/loongarch/net/bpf_jit.c
@@ -1462,9 +1462,6 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
 	u32 **branches = NULL;
 
-	if (flags & (BPF_TRAMP_F_ORIG_STACK | BPF_TRAMP_F_SHARE_IPMODIFY))
-		return -ENOTSUPP;
-
 	/*
 	 * FP + 8       [ RA to parent func ] return address to parent
 	 *                    function
-- 
2.43.5

[PATCH 2/3] LoongArch: BPF: Sign extend struct ops return values properly

[Hengqi Chen]

The ns_bpf_qdisc selftest triggers a kernel panic:

    CPU 0 Unable to handle kernel paging request at virtual address 0000000000741d58, era == 90000000851b5ac0, ra == 90000000851b5aa4
    Oops[#1]:
    CPU: 0 UID: 0 PID: 449 Comm: test_progs Tainted: G           OE       6.16.0+ #3 PREEMPT(full)
    Tainted: [O]=OOT_MODULE, [E]=UNSIGNED_MODULE
    Hardware name: QEMU QEMU Virtual Machine, BIOS unknown 2/2/2022
    pc 90000000851b5ac0 ra 90000000851b5aa4 tp 90000001076b8000 sp 90000001076bb600
    a0 0000000000741ce8 a1 0000000000000001 a2 90000001076bb5c0 a3 0000000000000008
    a4 90000001004c4620 a5 9000000100741ce8 a6 0000000000000000 a7 0100000000000000
    t0 0000000000000010 t1 0000000000000000 t2 9000000104d24d30 t3 0000000000000001
    t4 4f2317da8a7e08c4 t5 fffffefffc002f00 t6 90000001004c4620 t7 ffffffffc61c5b3d
    t8 0000000000000000 u0 0000000000000001 s9 0000000000000050 s0 90000001075bc800
    s1 0000000000000040 s2 900000010597c400 s3 0000000000000008 s4 90000001075bc880
    s5 90000001075bc8f0 s6 0000000000000000 s7 0000000000741ce8 s8 0000000000000000
       ra: 90000000851b5aa4 __qdisc_run+0xac/0x8d8
      ERA: 90000000851b5ac0 __qdisc_run+0xc8/0x8d8
     CRMD: 000000b0 (PLV0 -IE -DA +PG DACF=CC DACM=CC -WE)
     PRMD: 00000004 (PPLV0 +PIE -PWE)
     EUEN: 00000007 (+FPE +SXE +ASXE -BTE)
     ECFG: 00071c1d (LIE=0,2-4,10-12 VS=7)
    ESTAT: 00010000 [PIL] (IS= ECode=1 EsubCode=0)
     BADV: 0000000000741d58
     PRID: 0014c010 (Loongson-64bit, Loongson-3A5000)
    Modules linked in: bpf_testmod(OE) [last unloaded: bpf_testmod(OE)]
    Process test_progs (pid: 449, threadinfo=000000009af02b3a, task=00000000e9ba4956)
    Stack : 0000000000000000 90000001075bc8ac 90000000869524a8 9000000100741ce8
            90000001075bc800 9000000100415300 90000001075bc8ac 0000000000000000
            900000010597c400 900000008694a000 0000000000000000 9000000105b59000
            90000001075bc800 9000000100741ce8 0000000000000050 900000008513000c
            9000000086936000 0000000100094d4c fffffff400676208 0000000000000000
            9000000105b59000 900000008694a000 9000000086bf0dc0 9000000105b59000
            9000000086bf0d68 9000000085147010 90000001075be788 0000000000000000
            9000000086bf0f98 0000000000000001 0000000000000010 9000000006015840
            0000000000000000 9000000086be6c40 0000000000000000 0000000000000000
            0000000000000000 4f2317da8a7e08c4 0000000000000101 4f2317da8a7e08c4
            ...
    Call Trace:
    [<90000000851b5ac0>] __qdisc_run+0xc8/0x8d8
    [<9000000085130008>] __dev_queue_xmit+0x578/0x10f0
    [<90000000853701c0>] ip6_finish_output2+0x2f0/0x950
    [<9000000085374bc8>] ip6_finish_output+0x2b8/0x448
    [<9000000085370b24>] ip6_xmit+0x304/0x858
    [<90000000853c4438>] inet6_csk_xmit+0x100/0x170
    [<90000000852b32f0>] __tcp_transmit_skb+0x490/0xdd0
    [<90000000852b47fc>] tcp_connect+0xbcc/0x1168
    [<90000000853b9088>] tcp_v6_connect+0x580/0x8a0
    [<90000000852e7738>] __inet_stream_connect+0x170/0x480
    [<90000000852e7a98>] inet_stream_connect+0x50/0x88
    [<90000000850f2814>] __sys_connect+0xe4/0x110
    [<90000000850f2858>] sys_connect+0x18/0x28
    [<9000000085520c94>] do_syscall+0x94/0x1a0
    [<9000000083df1fb8>] handle_syscall+0xb8/0x158

    Code: 4001ad80  2400873f  2400832d <240073cc> 001137ff  001133ff  6407b41f  001503cc  0280041d

    ---[ end trace 0000000000000000 ]---

The bpf_fifo_dequeue prog returns a skb which is a pointer.
The pointer is treated as a 32bit value and sign extend to
64bit in epilogue. This behavior is right for most bpf prog
types but wrong for struct ops which requires LoongArch ABI.

So let's sign extend struct ops return values according to
the return value spec in function model.

Fixes: 6abf17d690d8 ("LoongArch: BPF: Add struct ops support for trampoline")
Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
---
 arch/loongarch/net/bpf_jit.c | 45 ++++++++++++++++++++++++++++++------
 1 file changed, 38 insertions(+), 7 deletions(-)

diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
index b646c6b73014..6754e5231ece 100644
--- a/arch/loongarch/net/bpf_jit.c
+++ b/arch/loongarch/net/bpf_jit.c
@@ -1361,7 +1361,8 @@ static void restore_args(struct jit_ctx *ctx, int nargs, int args_off)
 }
 
 static int invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_link *l,
-			   int args_off, int retval_off, int run_ctx_off, bool save_ret)
+			   const struct btf_func_model *m, int args_off,
+			   int retval_off, int run_ctx_off, bool save_ret)
 {
 	int ret;
 	u32 *branch;
@@ -1425,13 +1426,14 @@ static int invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_link *l,
 }
 
 static void invoke_bpf_mod_ret(struct jit_ctx *ctx, struct bpf_tramp_links *tl,
-			       int args_off, int retval_off, int run_ctx_off, u32 **branches)
+			       const struct btf_func_model *m, int args_off,
+			       int retval_off, int run_ctx_off, u32 **branches)
 {
 	int i;
 
 	emit_insn(ctx, std, LOONGARCH_GPR_ZERO, LOONGARCH_GPR_FP, -retval_off);
 	for (i = 0; i < tl->nr_links; i++) {
-		invoke_bpf_prog(ctx, tl->links[i], args_off, retval_off, run_ctx_off, true);
+		invoke_bpf_prog(ctx, tl->links[i], m, args_off, retval_off, run_ctx_off, true);
 		emit_insn(ctx, ldd, LOONGARCH_GPR_T1, LOONGARCH_GPR_FP, -retval_off);
 		branches[i] = (u32 *)ctx->image + ctx->idx;
 		emit_insn(ctx, nop);
@@ -1448,6 +1450,28 @@ void arch_free_bpf_trampoline(void *image, unsigned int size)
 	bpf_prog_pack_free(image, size);
 }
 
+/*
+ * Sign-extend the register if necessary
+ */
+static int sign_extend(struct jit_ctx *ctx, int r, u8 size)
+{
+	switch (size) {
+	case 1:
+		emit_insn(ctx, extwb, r, r);
+		return 0;
+	case 2:
+		emit_insn(ctx, extwh, r, r);
+		return 0;
+	case 4:
+		emit_insn(ctx, addiw, r, r, 0);
+		return 0;
+	case 8:
+		return 0;
+	default:
+		return -1;
+	}
+}
+
 static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_image *im,
 					 const struct btf_func_model *m, struct bpf_tramp_links *tlinks,
 					 void *func_addr, u32 flags)
@@ -1599,8 +1623,8 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
 	}
 
 	for (i = 0; i < fentry->nr_links; i++) {
-		ret = invoke_bpf_prog(ctx, fentry->links[i], args_off, retval_off,
-				      run_ctx_off, flags & BPF_TRAMP_F_RET_FENTRY_RET);
+		ret = invoke_bpf_prog(ctx, fentry->links[i], m, args_off, retval_off,
+			      run_ctx_off, flags & BPF_TRAMP_F_RET_FENTRY_RET);
 		if (ret)
 			return ret;
 	}
@@ -1609,7 +1633,7 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
 		if (!branches)
 			return -ENOMEM;
 
-		invoke_bpf_mod_ret(ctx, fmod_ret, args_off, retval_off, run_ctx_off, branches);
+		invoke_bpf_mod_ret(ctx, fmod_ret, m, args_off, retval_off, run_ctx_off, branches);
 	}
 
 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
@@ -1635,7 +1659,8 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
 	}
 
 	for (i = 0; i < fexit->nr_links; i++) {
-		ret = invoke_bpf_prog(ctx, fexit->links[i], args_off, retval_off, run_ctx_off, false);
+		ret = invoke_bpf_prog(ctx, fexit->links[i], m, args_off,
+				      retval_off, run_ctx_off, false);
 		if (ret)
 			goto out;
 	}
@@ -1654,6 +1679,12 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
 	if (save_ret) {
 		emit_insn(ctx, ldd, LOONGARCH_GPR_A0, LOONGARCH_GPR_FP, -retval_off);
 		emit_insn(ctx, ldd, regmap[BPF_REG_0], LOONGARCH_GPR_FP, -(retval_off - 8));
+		if (is_struct_ops) {
+			move_reg(ctx, LOONGARCH_GPR_A0, regmap[BPF_REG_0]);
+			ret = sign_extend(ctx, LOONGARCH_GPR_A0, m->ret_size);
+			if (ret)
+				goto out;
+		}
 	}
 
 	emit_insn(ctx, ldd, LOONGARCH_GPR_S1, LOONGARCH_GPR_FP, -sreg_off);
-- 
2.43.5

[PATCH 3/3] LoongArch: BPF: No support of struct argument in trampoline programs

[Hengqi Chen]

The current implementation does not support struct argument.
This cause a oops when running bpf selftest:

    $ ./test_progs -a tracing_struct
    CPU -1 Unable to handle kernel paging request at virtual address 0000000000000018, era == 90000000845659f4, ra == 90000000845659e8
    Oops[#1]:
    CPU -1 Unable to handle kernel paging request at virtual address 0000000000000018, era == 9000000085bef268, ra == 90000000844f3938
    rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
    rcu:     1-...0: (19 ticks this GP) idle=1094/1/0x4000000000000000 softirq=1380/1382 fqs=801
    rcu:     (detected by 0, t=5252 jiffies, g=1197, q=52 ncpus=4)
    Sending NMI from CPU 0 to CPUs 1:
    rcu: rcu_preempt kthread starved for 2495 jiffies! g1197 f0x0 RCU_GP_DOING_FQS(6) ->state=0x0 ->cpu=2
    rcu:     Unless rcu_preempt kthread gets sufficient CPU time, OOM is now expected behavior.
    rcu: RCU grace-period kthread stack dump:
    task:rcu_preempt     state:I stack:0     pid:15    tgid:15    ppid:2      task_flags:0x208040 flags:0x00000800
    Stack : 9000000100423e80 0000000000000402 0000000000000010 90000001003b0680
            9000000085d88000 0000000000000000 0000000000000040 9000000087159350
            9000000085c2b9b0 0000000000000001 900000008704a000 0000000000000005
            00000000ffff355b 00000000ffff355b 0000000000000000 0000000000000004
            9000000085d90510 0000000000000000 0000000000000002 7b5d998f8281e86e
            00000000ffff355c 7b5d998f8281e86e 000000000000003f 9000000087159350
            900000008715bf98 0000000000000005 9000000087036000 900000008704a000
            9000000100407c98 90000001003aff80 900000008715c4c0 9000000085c2b9b0
            00000000ffff355b 9000000085c33d3c 00000000000000b4 0000000000000000
            9000000007002150 00000000ffff355b 9000000084615480 0000000007000002
            ...
    Call Trace:
    [<9000000085c2a868>] __schedule+0x410/0x1520
    [<9000000085c2b9ac>] schedule+0x34/0x190
    [<9000000085c33d38>] schedule_timeout+0x98/0x140
    [<90000000845e9120>] rcu_gp_fqs_loop+0x5f8/0x868
    [<90000000845ed538>] rcu_gp_kthread+0x260/0x2e0
    [<900000008454e8a4>] kthread+0x144/0x238
    [<9000000085c26b60>] ret_from_kernel_thread+0x28/0xc8
    [<90000000844f20e4>] ret_from_kernel_thread_asm+0xc/0x88

    rcu: Stack dump where RCU GP kthread last ran:
    Sending NMI from CPU 0 to CPUs 2:
    NMI backtrace for cpu 2 skipped: idling at idle_exit+0x0/0x4

Reject it for now.

Fixes: f9b6b41f0cf3 ("LoongArch: BPF: Add basic bpf trampoline support")
Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
---
 arch/loongarch/net/bpf_jit.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
index 6754e5231ece..a87f51f5b708 100644
--- a/arch/loongarch/net/bpf_jit.c
+++ b/arch/loongarch/net/bpf_jit.c
@@ -1516,6 +1516,12 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
 	if (m->nr_args > LOONGARCH_MAX_REG_ARGS)
 		return -ENOTSUPP;
 
+	/* don't support struct argument */
+	for (i = 0; i < m->nr_args; i++) {
+		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
+			return -ENOTSUPP;
+	}
+
 	if (flags & (BPF_TRAMP_F_ORIG_STACK | BPF_TRAMP_F_SHARE_IPMODIFY))
 		return -ENOTSUPP;
 
-- 
2.43.5

Re: [PATCH 1/3] LoongArch: BPF: Remove duplicated flags check

[Tiezhu Yang]

On 2025/8/21 下午5:10, Hengqi Chen wrote:
> The check for (BPF_TRAMP_F_ORIG_STACK | BPF_TRAMP_F_SHARE_IPMODIFY)
> is duplicated in __arch_prepare_bpf_trampoline(). Remove it.
> 
> Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
> ---
>   arch/loongarch/net/bpf_jit.c | 3 ---
>   1 file changed, 3 deletions(-)
> 
> diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
> index abfdb6bb5c38..b646c6b73014 100644
> --- a/arch/loongarch/net/bpf_jit.c
> +++ b/arch/loongarch/net/bpf_jit.c
> @@ -1462,9 +1462,6 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
>   	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
>   	u32 **branches = NULL;
>   
> -	if (flags & (BPF_TRAMP_F_ORIG_STACK | BPF_TRAMP_F_SHARE_IPMODIFY))
> -		return -ENOTSUPP;
> -
>   	/*
>   	 * FP + 8       [ RA to parent func ] return address to parent
>   	 *                    function
> 

Acked-by: Tiezhu Yang <yangtiezhu@loongson.cn>

Maybe "stack_size = 0;" in this function is redundant too, if yes,
you can remove it together.

Thanks,
Tiezhu

Re: [PATCH 2/3] LoongArch: BPF: Sign extend struct ops return values properly

[Tiezhu Yang]

On 2025/8/21 下午5:10, Hengqi Chen wrote:
> The ns_bpf_qdisc selftest triggers a kernel panic:
...

> The bpf_fifo_dequeue prog returns a skb which is a pointer.
> The pointer is treated as a 32bit value and sign extend to
> 64bit in epilogue. This behavior is right for most bpf prog
> types but wrong for struct ops which requires LoongArch ABI.
> 
> So let's sign extend struct ops return values according to
> the return value spec in function model.
> 
> Fixes: 6abf17d690d8 ("LoongArch: BPF: Add struct ops support for trampoline")
> Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>

The following related testcases passed on LoongArch:

   sudo modprobe test_bpf
   sudo ./test_progs -a ns_bpf_qdisc
   sudo ./test_progs -t struct_ops -d struct_ops_multi_pages

Tested-by: Tiezhu Yang <yangtiezhu@loongson.cn>

Thanks,
Tiezhu

Re: [PATCH 3/3] LoongArch: BPF: No support of struct argument in trampoline programs

[Tiezhu Yang]

On 2025/8/21 下午5:10, Hengqi Chen wrote:
> The current implementation does not support struct argument.

...

> Reject it for now.
> 
> Fixes: f9b6b41f0cf3 ("LoongArch: BPF: Add basic bpf trampoline support")
> Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
> ---
>   arch/loongarch/net/bpf_jit.c | 6 ++++++
>   1 file changed, 6 insertions(+)
> 
> diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
> index 6754e5231ece..a87f51f5b708 100644
> --- a/arch/loongarch/net/bpf_jit.c
> +++ b/arch/loongarch/net/bpf_jit.c
> @@ -1516,6 +1516,12 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
>   	if (m->nr_args > LOONGARCH_MAX_REG_ARGS)
>   		return -ENOTSUPP;
>   
> +	/* don't support struct argument */
> +	for (i = 0; i < m->nr_args; i++) {
> +		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
> +			return -ENOTSUPP;
> +	}
> +

Acked-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Tested-by: Tiezhu Yang <yangtiezhu@loongson.cn>

By the way, I have 2 patches locally, one patch is to implement this 
feature, but there exist some other problems now, maybe it is related 
with the following failed testcase:

   sudo ./test_progs -t module_attach

Anyway, I will send a RFC series later for your review.

Thanks,
Tiezhu

Re: [PATCH 0/3] LoongArch: Fix BPF trampoline related issues

[Vincent Li]

On Thu, Aug 21, 2025 at 5:16 AM Hengqi Chen <hengqi.chen@gmail.com> wrote:
>
> The following two selftest cases triggers oops on LoongArch:
>
>     $ ./test_progs -a ns_bpf_qdisc -a tracing_struct
>

I tested this patch series by running bpf selftests as below to bypass
the tests that would hang up the loongarch kernel, I also denied the
tracing_struct that I think is related to module_attach and
subprogs_extable, these three tests I think is  related to
fentry/fexit on kernel module functions and that would hang up
loongarch kernel. The rest of bpf selftests failed report for the
loongarch kernel  is here [0].

./test_progs \
--deny=module_attach \
--deny=struct_ops \
--deny=subprogs_extable \
--deny=tracing_struct \
--deny=timer_lockup \
--json-summary="./ns_bpf_qdisc.json"

...
#213/1   ns_bpf_qdisc/fifo:OK
#213/2   ns_bpf_qdisc/fq:OK
#213/3   ns_bpf_qdisc/attach to mq:OK
#213/4   ns_bpf_qdisc/attach to non root:OK
#213/5   ns_bpf_qdisc/incompl_ops:OK
#213     ns_bpf_qdisc:OK
...

[0]: https://www.bpfire.net/download/ns_bpf_qdisc.json.txt

Tested-by: Vincent Li <vincent.mc.li@gmail.com>

> This small series tries to fix/workaround these issues.
> See individual commit for details.
>
> While at it, remove a duplicated flags check in __arch_prepare_bpf_trampoline().
>
> Hengqi Chen (3):
>   LoongArch: BPF: Remove duplicated flags check
>   LoongArch: BPF: Sign extend struct ops return values properly
>   LoongArch: BPF: No support of struct argument in trampoline programs
>
>  arch/loongarch/net/bpf_jit.c | 54 +++++++++++++++++++++++++++++-------
>  1 file changed, 44 insertions(+), 10 deletions(-)
>
> --
> 2.43.5