[PATCH bpf v2 1/2] bpf: Skip detached progs in trampoline images that are still in use

["Florent Revest (Anthropic)" <florent.revest@linux.dev>]

bpf_tramp_image_put() makes sure a trampoline image is not freed while
a task may still be running in it, but nothing similar is done for the
progs called by that image. Detach drops the last prog reference right
away and the prog is freed after a grace period of its own RCU flavor,
on the basis that a task inside the prog holds rcu_read_lock() or
rcu_read_lock_trace(), and that a task still in the traced function
skips the fexit progs once the nop at ip_after_call is patched to a
jump.

That leaves out a task elsewhere in the image: sleeping in a sleepable
prog that runs before the detached one (with any preemption model), or
preempted in the few instructions between two progs. Neither holds
anything that delays the free of the prog it is about to call:

  CPU 0                               CPU 1
  in image I, sleeping in prog S
                                      detach P from I's trampoline
                                       -> new image, bpf_tramp_image_put(I)
                                      bpf_prog_put(P), last ref
                                      RCU / tasks trace GP, P freed
  back from S
  __bpf_prog_enter(P)
  call P->bpf_func

If S and P are fexit progs the task is already past the patched jump,
and fentry only images don't have one. On x86 this is an int3 in
poisoned bpf_prog_pack memory:

  Oops: int3: 0000 [#1] SMP NOPTI
  CPU: 18 UID: 0 PID: 94573 Comm: x169 Not tainted 6.18.44 #1 PREEMPT(lazy)
  RIP: 0010:0xffffffffc0601d8d
  Call Trace:
   <TASK>
   ? bpf_trampoline_6442515411+0x1a4/0x21b
   bpf_lsm_bprm_committed_creds+0x5/0x10
   security_bprm_committed_creds+0x5f/0x70
   begin_new_exec+0x2d6/0x410
   ...

We hit this in production on preemptible kernels when progs attached
through trampolines got detached while their hooks were busy, and it was
independently found with a fuzzer and KASAN.

Extend the ip_after_call mechanism to every prog: have the JITs emit a
patchable nop in front of each prog call sequence and record it in the
image, and when a prog is detached, patch its nop to a jump over the
call sequence in every image of the trampoline that is not freed yet.
The trampoline keeps a list of those images for that, and they hold a
reference on it until they are freed. Only the detached prog is skipped
so tasks in an old image keep running the progs that are still
attached. With the extra nops, BPF_MAX_TRAMP_LINKS progs no longer fit
in a page on arm64 and loongarch (and already didn't on powerpc), so
lower the limit there like s390 does.

Fixes: e21aa341785c ("bpf: Fix fexit trampoline.")
Reported-by: Sechang Lim <rhkrqnwk98@gmail.com>
Closes: https://lore.kernel.org/bpf/20260815071927.147049-1-zirajs7@gmail.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Assisted-by: Claude:unspecified
Signed-off-by: Florent Revest (Anthropic) <florent.revest@linux.dev>
---
 arch/arm64/net/bpf_jit_comp.c   | 29 ++++++++++++-------
 arch/loongarch/net/bpf_jit.c    | 35 ++++++++++++++--------
 arch/powerpc/net/bpf_jit_comp.c | 38 ++++++++++++++----------
 arch/riscv/net/bpf_jit_comp64.c | 34 ++++++++++++++--------
 arch/s390/net/bpf_jit_comp.c    | 35 ++++++++++++++++++----
 arch/x86/net/bpf_jit_comp.c     | 23 ++++++++++-----
 include/linux/bpf.h             | 40 ++++++++++++++++++++++++--
 kernel/bpf/trampoline.c         | 51 +++++++++++++++++++++++++++++++--
 8 files changed, 218 insertions(+), 67 deletions(-)

diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
index c18e005a41db..9c166bdfbc6a 100644
--- a/arch/arm64/net/bpf_jit_comp.c
+++ b/arch/arm64/net/bpf_jit_comp.c
@@ -2416,10 +2416,11 @@ bool bpf_jit_supports_subprog_tailcalls(void)
 	return true;
 }
 
-static void invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_node *node,
-			    int bargs_off, int retval_off, int run_ctx_off,
-			    bool save_ret)
+static void invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_image *im,
+			    struct bpf_tramp_node *node, int bargs_off,
+			    int retval_off, int run_ctx_off, bool save_ret)
 {
+	void *skip;
 	__le32 *branch;
 	u64 enter_prog;
 	u64 exit_prog;
@@ -2429,6 +2430,10 @@ static void invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_node *node,
 	enter_prog = (u64)bpf_trampoline_enter(p);
 	exit_prog = (u64)bpf_trampoline_exit(p);
 
+	/* nop, patched to skip this prog when it is detached */
+	skip = ctx->ro_image + ctx->idx;
+	emit(A64_NOP, ctx);
+
 	if (node->cookie == 0) {
 		/* if cookie is zero, one instruction is enough to store it */
 		emit(A64_STR64I(A64_ZR, A64_SP, run_ctx_off + cookie_off), ctx);
@@ -2481,11 +2486,13 @@ static void invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_node *node,
 	emit(A64_ADD_I(1, A64_R(2), A64_SP, run_ctx_off), ctx);
 
 	emit_call(exit_prog, ctx);
+
+	bpf_tramp_image_add_skip(im, p, skip, ctx->ro_image + ctx->idx);
 }
 
-static void invoke_bpf_mod_ret(struct jit_ctx *ctx, struct bpf_tramp_nodes *tn,
-			       int bargs_off, int retval_off, int run_ctx_off,
-			       __le32 **branches)
+static void invoke_bpf_mod_ret(struct jit_ctx *ctx, struct bpf_tramp_image *im,
+			       struct bpf_tramp_nodes *tn, int bargs_off,
+			       int retval_off, int run_ctx_off, __le32 **branches)
 {
 	int i;
 
@@ -2494,7 +2501,7 @@ static void invoke_bpf_mod_ret(struct jit_ctx *ctx, struct bpf_tramp_nodes *tn,
 	 */
 	emit(A64_STR64I(A64_ZR, A64_SP, retval_off), ctx);
 	for (i = 0; i < tn->nr_nodes; i++) {
-		invoke_bpf_prog(ctx, tn->nodes[i], bargs_off, retval_off,
+		invoke_bpf_prog(ctx, im, tn->nodes[i], bargs_off, retval_off,
 				run_ctx_off, true);
 		/* if (*(u64 *)(sp + retval_off) !=  0)
 		 *	goto do_fexit;
@@ -2880,7 +2887,7 @@ static int prepare_trampoline(struct jit_ctx *ctx, struct bpf_tramp_image *im,
 			store_func_meta(ctx, meta, func_meta_off);
 			cookie_bargs_off--;
 		}
-		invoke_bpf_prog(ctx, fentry->nodes[i], bargs_off,
+		invoke_bpf_prog(ctx, im, fentry->nodes[i], bargs_off,
 				retval_off, run_ctx_off,
 				flags & BPF_TRAMP_F_RET_FENTRY_RET);
 	}
@@ -2891,7 +2898,7 @@ static int prepare_trampoline(struct jit_ctx *ctx, struct bpf_tramp_image *im,
 		if (!branches)
 			return -ENOMEM;
 
-		invoke_bpf_mod_ret(ctx, fmod_ret, bargs_off, retval_off,
+		invoke_bpf_mod_ret(ctx, im, fmod_ret, bargs_off, retval_off,
 				   run_ctx_off, branches);
 	}
 
@@ -2928,7 +2935,7 @@ static int prepare_trampoline(struct jit_ctx *ctx, struct bpf_tramp_image *im,
 			store_func_meta(ctx, meta, func_meta_off);
 			cookie_bargs_off--;
 		}
-		invoke_bpf_prog(ctx, fexit->nodes[i], bargs_off, retval_off,
+		invoke_bpf_prog(ctx, im, fexit->nodes[i], bargs_off, retval_off,
 				run_ctx_off, false);
 	}
 
@@ -2992,7 +2999,7 @@ int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
 		.image = NULL,
 		.idx = 0,
 	};
-	struct bpf_tramp_image im;
+	struct bpf_tramp_image im = {};
 	struct arg_aux aaux;
 	int ret;
 
diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
index 4da278900938..0e6f1ad36c2d 100644
--- a/arch/loongarch/net/bpf_jit.c
+++ b/arch/loongarch/net/bpf_jit.c
@@ -1696,13 +1696,19 @@ static void restore_stk_args(struct jit_ctx *ctx, int nr_stk_args, int args_off,
 	}
 }
 
-static int invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_node *n,
-			   int args_off, int retval_off, int run_ctx_off, bool save_ret)
+static int invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_image *im,
+			   struct bpf_tramp_node *n, int args_off, int retval_off,
+			   int run_ctx_off, bool save_ret)
 {
-	int ret;
+	int i, ret;
 	u32 *branch;
 	struct bpf_prog *p = n->link->prog;
 	int cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
+	void *skip = ctx->ro_image + ctx->idx;
+
+	/* nops for move_imm+jirl, patched to skip this prog when it is detached */
+	for (i = 0; i < LOONGARCH_LONG_JUMP_NINSNS; i++)
+		emit_insn(ctx, nop);
 
 	if (n->cookie)
 		emit_store_stack_imm64(ctx, LOONGARCH_GPR_T1,
@@ -1755,13 +1761,17 @@ static int invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_node *n,
 	/* arg3: &run_ctx */
 	emit_insn(ctx, addid, LOONGARCH_GPR_A2, LOONGARCH_GPR_FP, -run_ctx_off);
 	ret = emit_call(ctx, (const u64)bpf_trampoline_exit(p));
+	if (ret)
+		return ret;
 
-	return ret;
+	bpf_tramp_image_add_skip(im, p, skip, ctx->ro_image + ctx->idx);
+	return 0;
 }
 
-static int invoke_bpf(struct jit_ctx *ctx, struct bpf_tramp_nodes *tn,
-		      int args_off, int retval_off, int run_ctx_off,
-		      int func_meta_off, bool save_ret, u64 func_meta, int cookie_off)
+static int invoke_bpf(struct jit_ctx *ctx, struct bpf_tramp_image *im,
+		      struct bpf_tramp_nodes *tn, int args_off, int retval_off,
+		      int run_ctx_off, int func_meta_off, bool save_ret,
+		      u64 func_meta, int cookie_off)
 {
 	int i, cur_cookie = (cookie_off - args_off) / 8;
 
@@ -1774,7 +1784,8 @@ static int invoke_bpf(struct jit_ctx *ctx, struct bpf_tramp_nodes *tn,
 			emit_store_stack_imm64(ctx, LOONGARCH_GPR_T1, -func_meta_off, meta);
 			cur_cookie--;
 		}
-		err = invoke_bpf_prog(ctx, tn->nodes[i], args_off, retval_off, run_ctx_off, save_ret);
+		err = invoke_bpf_prog(ctx, im, tn->nodes[i], args_off, retval_off,
+				      run_ctx_off, save_ret);
 		if (err)
 			return err;
 	}
@@ -2017,7 +2028,7 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
 	}
 
 	if (fentry->nr_nodes) {
-		ret = invoke_bpf(ctx, fentry, args_off, retval_off, run_ctx_off, func_meta_off,
+		ret = invoke_bpf(ctx, im, fentry, args_off, retval_off, run_ctx_off, func_meta_off,
 				 flags & BPF_TRAMP_F_RET_FENTRY_RET, func_meta, cookie_off);
 		if (ret)
 			return ret;
@@ -2029,7 +2040,7 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
 
 		emit_insn(ctx, std, LOONGARCH_GPR_ZERO, LOONGARCH_GPR_FP, -retval_off);
 		for (i = 0; i < fmod_ret->nr_nodes; i++) {
-			ret = invoke_bpf_prog(ctx, fmod_ret->nodes[i],
+			ret = invoke_bpf_prog(ctx, im, fmod_ret->nodes[i],
 					      args_off, retval_off, run_ctx_off, true);
 			if (ret)
 				goto out;
@@ -2068,7 +2079,7 @@ static int __arch_prepare_bpf_trampoline(struct jit_ctx *ctx, struct bpf_tramp_i
 		emit_store_stack_imm64(ctx, LOONGARCH_GPR_T1, -func_meta_off, func_meta);
 
 	if (fexit->nr_nodes) {
-		ret = invoke_bpf(ctx, fexit, args_off, retval_off, run_ctx_off,
+		ret = invoke_bpf(ctx, im, fexit, args_off, retval_off, run_ctx_off,
 				 func_meta_off, false, func_meta, cookie_off);
 		if (ret)
 			goto out;
@@ -2178,7 +2189,7 @@ int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
 {
 	int ret;
 	struct jit_ctx ctx;
-	struct bpf_tramp_image im;
+	struct bpf_tramp_image im = {};
 
 	ctx.image = NULL;
 	ctx.idx = 0;
diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
index 7b07b43575f1..95ea27148c91 100644
--- a/arch/powerpc/net/bpf_jit_comp.c
+++ b/arch/powerpc/net/bpf_jit_comp.c
@@ -602,14 +602,18 @@ int arch_protect_bpf_trampoline(void *image, unsigned int size)
 }
 
 static int invoke_bpf_prog(u32 *image, u32 *ro_image, struct codegen_context *ctx,
-			   struct bpf_tramp_node *n, int regs_off, int retval_off,
-			   int run_ctx_off, bool save_ret)
+			   struct bpf_tramp_image *im, struct bpf_tramp_node *n,
+			   int regs_off, int retval_off, int run_ctx_off, bool save_ret)
 {
 	struct bpf_prog *p = n->link->prog;
 	ppc_inst_t branch_insn;
-	u32 jmp_idx;
+	u32 jmp_idx, skip_idx;
 	int ret = 0;
 
+	/* nop, patched to skip this prog when it is detached */
+	skip_idx = ctx->idx;
+	EMIT(PPC_RAW_NOP());
+
 	/* Save cookie */
 	if (IS_ENABLED(CONFIG_PPC64)) {
 		PPC_LI64(_R3, n->cookie);
@@ -679,13 +683,17 @@ static int invoke_bpf_prog(u32 *image, u32 *ro_image, struct codegen_context *ct
 	EMIT(PPC_RAW_ADDI(_R5, _R1, run_ctx_off));
 	ret = bpf_jit_emit_func_call_rel(image, ro_image, ctx,
 					 (unsigned long)bpf_trampoline_exit(p));
+	if (ret)
+		return ret;
 
-	return ret;
+	if (ro_image) /* image is NULL for dummy pass */
+		bpf_tramp_image_add_skip(im, p, &ro_image[skip_idx], &ro_image[ctx->idx]);
+	return 0;
 }
 
 static int invoke_bpf_mod_ret(u32 *image, u32 *ro_image, struct codegen_context *ctx,
-			      struct bpf_tramp_nodes *tn, int regs_off, int retval_off,
-			      int run_ctx_off, u32 *branches)
+			      struct bpf_tramp_image *im, struct bpf_tramp_nodes *tn,
+			      int regs_off, int retval_off, int run_ctx_off, u32 *branches)
 {
 	int i;
 
@@ -696,8 +704,8 @@ static int invoke_bpf_mod_ret(u32 *image, u32 *ro_image, struct codegen_context
 	EMIT(PPC_RAW_LI(_R3, 0));
 	EMIT(PPC_RAW_STL(_R3, _R1, retval_off));
 	for (i = 0; i < tn->nr_nodes; i++) {
-		if (invoke_bpf_prog(image, ro_image, ctx, tn->nodes[i], regs_off, retval_off,
-				    run_ctx_off, true))
+		if (invoke_bpf_prog(image, ro_image, ctx, im, tn->nodes[i], regs_off,
+				    retval_off, run_ctx_off, true))
 			return -EINVAL;
 
 		/*
@@ -1043,8 +1051,8 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_im
 			cookie_ctx_off--;
 		}
 
-		if (invoke_bpf_prog(image, ro_image, ctx, fentry->nodes[i], regs_off, retval_off,
-				    run_ctx_off, flags & BPF_TRAMP_F_RET_FENTRY_RET))
+		if (invoke_bpf_prog(image, ro_image, ctx, im, fentry->nodes[i], regs_off,
+				    retval_off, run_ctx_off, flags & BPF_TRAMP_F_RET_FENTRY_RET))
 			return -EINVAL;
 	}
 
@@ -1053,7 +1061,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_im
 		if (!branches)
 			return -ENOMEM;
 
-		if (invoke_bpf_mod_ret(image, ro_image, ctx, fmod_ret, regs_off, retval_off,
+		if (invoke_bpf_mod_ret(image, ro_image, ctx, im, fmod_ret, regs_off, retval_off,
 				       run_ctx_off, branches)) {
 			ret = -EINVAL;
 			goto cleanup;
@@ -1123,8 +1131,8 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_im
 			cookie_ctx_off--;
 		}
 
-		if (invoke_bpf_prog(image, ro_image, ctx, fexit->nodes[i], regs_off, retval_off,
-				    run_ctx_off, false)) {
+		if (invoke_bpf_prog(image, ro_image, ctx, im, fexit->nodes[i], regs_off,
+				    retval_off, run_ctx_off, false)) {
 			ret = -EINVAL;
 			goto cleanup;
 		}
@@ -1192,7 +1200,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_im
 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
 			     struct bpf_tramp_nodes *tnodes, void *func_addr)
 {
-	struct bpf_tramp_image im;
+	struct bpf_tramp_image im = {};
 	int ret;
 
 	ret = __arch_prepare_bpf_trampoline(&im, NULL, NULL, NULL, m, flags, tnodes, func_addr);
@@ -1320,7 +1328,7 @@ int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
 
 	/*
 	 * If we are not poking at bpf prog entry, then we are simply patching in/out
-	 * an unconditional branch instruction at im->ip_after_call
+	 * an unconditional branch instruction in a trampoline image
 	 */
 	if (offset) {
 		if (old_t == BPF_MOD_CALL || new_t == BPF_MOD_CALL) {
diff --git a/arch/riscv/net/bpf_jit_comp64.c b/arch/riscv/net/bpf_jit_comp64.c
index 151031e97a24..21f8ce2ec89a 100644
--- a/arch/riscv/net/bpf_jit_comp64.c
+++ b/arch/riscv/net/bpf_jit_comp64.c
@@ -904,12 +904,18 @@ static void emit_store_stack_imm64(u8 reg, int stack_off, u64 imm64,
 	emit_sd(RV_REG_FP, stack_off, reg, ctx);
 }
 
-static int invoke_bpf_prog(struct bpf_tramp_node *node, int args_off, int retval_off,
-			   int run_ctx_off, bool save_ret, struct rv_jit_context *ctx)
+static int invoke_bpf_prog(struct bpf_tramp_image *im, struct bpf_tramp_node *node,
+			   int args_off, int retval_off, int run_ctx_off, bool save_ret,
+			   struct rv_jit_context *ctx)
 {
 	int ret, branch_off;
 	struct bpf_prog *p = node->link->prog;
 	int cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
+	void *skip = ctx->ro_insns + ctx->ninsns;
+
+	/* 2 nops for auipc+jalr, patched to skip this prog when it is detached */
+	emit(rv_nop(), ctx);
+	emit(rv_nop(), ctx);
 
 	if (node->cookie)
 		emit_store_stack_imm64(RV_REG_T1, -run_ctx_off + cookie_off, node->cookie, ctx);
@@ -962,13 +968,17 @@ static int invoke_bpf_prog(struct bpf_tramp_node *node, int args_off, int retval
 	/* arg3: &run_ctx */
 	emit_addi(RV_REG_A2, RV_REG_FP, -run_ctx_off, ctx);
 	ret = emit_call((const u64)bpf_trampoline_exit(p), true, ctx);
+	if (ret)
+		return ret;
 
-	return ret;
+	bpf_tramp_image_add_skip(im, p, skip, ctx->ro_insns + ctx->ninsns);
+	return 0;
 }
 
-static int invoke_bpf(struct bpf_tramp_nodes *tn, int args_off, int retval_off,
-		      int run_ctx_off, int func_meta_off, bool save_ret, u64 func_meta,
-		      int cookie_off, struct rv_jit_context *ctx)
+static int invoke_bpf(struct bpf_tramp_image *im, struct bpf_tramp_nodes *tn,
+		      int args_off, int retval_off, int run_ctx_off, int func_meta_off,
+		      bool save_ret, u64 func_meta, int cookie_off,
+		      struct rv_jit_context *ctx)
 {
 	int i, cur_cookie = (cookie_off - args_off) / 8;
 
@@ -981,8 +991,8 @@ static int invoke_bpf(struct bpf_tramp_nodes *tn, int args_off, int retval_off,
 			emit_store_stack_imm64(RV_REG_T1, -func_meta_off, meta, ctx);
 			cur_cookie--;
 		}
-		err = invoke_bpf_prog(tn->nodes[i], args_off, retval_off, run_ctx_off,
-				      save_ret, ctx);
+		err = invoke_bpf_prog(im, tn->nodes[i], args_off, retval_off,
+				      run_ctx_off, save_ret, ctx);
 		if (err)
 			return err;
 	}
@@ -1170,7 +1180,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
 	}
 
 	if (fentry->nr_nodes) {
-		ret = invoke_bpf(fentry, args_off, retval_off, run_ctx_off, func_meta_off,
+		ret = invoke_bpf(im, fentry, args_off, retval_off, run_ctx_off, func_meta_off,
 				 flags & BPF_TRAMP_F_RET_FENTRY_RET, func_meta, cookie_off, ctx);
 		if (ret)
 			return ret;
@@ -1184,7 +1194,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
 		/* cleanup to avoid garbage return value confusion */
 		emit_sd(RV_REG_FP, -retval_off, RV_REG_ZERO, ctx);
 		for (i = 0; i < fmod_ret->nr_nodes; i++) {
-			ret = invoke_bpf_prog(fmod_ret->nodes[i], args_off, retval_off,
+			ret = invoke_bpf_prog(im, fmod_ret->nodes[i], args_off, retval_off,
 					      run_ctx_off, true, ctx);
 			if (ret)
 				goto out;
@@ -1230,7 +1240,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
 		emit_store_stack_imm64(RV_REG_T1, -func_meta_off, func_meta, ctx);
 
 	if (fexit->nr_nodes) {
-		ret = invoke_bpf(fexit, args_off, retval_off, run_ctx_off, func_meta_off,
+		ret = invoke_bpf(im, fexit, args_off, retval_off, run_ctx_off, func_meta_off,
 				 false, func_meta, cookie_off, ctx);
 		if (ret)
 			goto out;
@@ -1299,7 +1309,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
 			     struct bpf_tramp_nodes *tnodes, void *func_addr)
 {
-	struct bpf_tramp_image im;
+	struct bpf_tramp_image im = {};
 	struct rv_jit_context ctx;
 	int ret;
 
diff --git a/arch/s390/net/bpf_jit_comp.c b/arch/s390/net/bpf_jit_comp.c
index c4b47070bb59..286a7c9215cb 100644
--- a/arch/s390/net/bpf_jit_comp.c
+++ b/arch/s390/net/bpf_jit_comp.c
@@ -2566,6 +2566,8 @@ struct bpf_tramp_jit {
 	int r14_off;		/* Offset of saved %r14, has to be at the
 				 * bottom */
 	int do_fexit;		/* do_fexit: label */
+	int skip[BPF_MAX_TRAMP_LINKS];	/* skip: labels after each prog */
+	int nr_progs;
 };
 
 static void load_imm64(struct bpf_jit *jit, int dst_reg, u64 val)
@@ -2584,6 +2586,7 @@ static void emit_store_stack_imm64(struct bpf_jit *jit, int tmp_reg, int stack_o
 }
 
 static int invoke_bpf_prog(struct bpf_tramp_jit *tjit,
+			   struct bpf_tramp_image *im,
 			   const struct btf_func_model *m,
 			   struct bpf_tramp_node *node, bool save_ret)
 {
@@ -2591,8 +2594,20 @@ static int invoke_bpf_prog(struct bpf_tramp_jit *tjit,
 	int cookie_off = tjit->run_ctx_off +
 			 offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
 	struct bpf_prog *p = node->link->prog;
+	void *skip = jit->prg_buf + jit->prg;
+	int idx = tjit->nr_progs++;
 	int patch;
 
+	if (idx >= ARRAY_SIZE(tjit->skip))
+		return -E2BIG;
+
+	/*
+	 * nop, patched to skip this prog when it is detached
+	 */
+
+	/* brcl 0,skip */
+	EMIT6_PCREL_RILC(0xc0040000, 0, tjit->skip[idx]);
+
 	/*
 	 * run_ctx.cookie = node->cookie;
 	 */
@@ -2652,10 +2667,15 @@ static int invoke_bpf_prog(struct bpf_tramp_jit *tjit,
 	/* brasl %r14,__bpf_prog_exit */
 	EMIT6_PCREL_RILB_PTR(0xc0050000, REG_14, bpf_trampoline_exit(p));
 
+	/* skip: */
+	tjit->skip[idx] = jit->prg;
+	bpf_tramp_image_add_skip(im, p, skip, jit->prg_buf + jit->prg);
+
 	return 0;
 }
 
 static int invoke_bpf(struct bpf_tramp_jit *tjit,
+		      struct bpf_tramp_image *im,
 		      const struct btf_func_model *m,
 		      struct bpf_tramp_nodes *tn, bool save_ret,
 		      u64 func_meta, int cookie_off)
@@ -2670,7 +2690,7 @@ static int invoke_bpf(struct bpf_tramp_jit *tjit,
 			emit_store_stack_imm64(jit, REG_0, tjit->func_meta_off, meta);
 			cur_cookie--;
 		}
-		if (invoke_bpf_prog(tjit, m, tn->nodes[i], save_ret))
+		if (invoke_bpf_prog(tjit, im, m, tn->nodes[i], save_ret))
 			return -EINVAL;
 	}
 
@@ -2712,6 +2732,11 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
 	u64 func_meta;
 	int i, j;
 
+	/* The skip labels are taken from the previous pass. */
+	tjit->nr_progs = 0;
+	if (im)
+		im->nr_skips = 0;
+
 	/* Support as many stack arguments as "mvc" instruction can handle. */
 	nr_reg_args = min_t(int, m->nr_args, MAX_NR_REG_ARGS);
 	nr_stack_args = m->nr_args - nr_reg_args;
@@ -2875,7 +2900,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
 		emit_store_stack_imm64(jit, REG_0, tjit->retval_off, 0);
 	}
 
-	if (invoke_bpf(tjit, m, fentry, flags & BPF_TRAMP_F_RET_FENTRY_RET,
+	if (invoke_bpf(tjit, im, m, fentry, flags & BPF_TRAMP_F_RET_FENTRY_RET,
 		       func_meta, cookie_off))
 		return -EINVAL;
 
@@ -2889,7 +2914,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
 		       0xf000 | tjit->retval_off);
 
 		for (i = 0; i < fmod_ret->nr_nodes; i++) {
-			if (invoke_bpf_prog(tjit, m, fmod_ret->nodes[i], true))
+			if (invoke_bpf_prog(tjit, im, m, fmod_ret->nodes[i], true))
 				return -EINVAL;
 
 			/*
@@ -2962,7 +2987,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
 
 	/* do_fexit: */
 	tjit->do_fexit = jit->prg;
-	if (invoke_bpf(tjit, m, fexit, false, func_meta, cookie_off))
+	if (invoke_bpf(tjit, im, m, fexit, false, func_meta, cookie_off))
 		return -EINVAL;
 
 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
@@ -3016,7 +3041,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
 			     struct bpf_tramp_nodes *tnodes, void *orig_call)
 {
-	struct bpf_tramp_image im;
+	struct bpf_tramp_image im = {};
 	struct bpf_tramp_jit tjit;
 	int ret;
 
diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
index 2853e87797a7..0d974da4f509 100644
--- a/arch/x86/net/bpf_jit_comp.c
+++ b/arch/x86/net/bpf_jit_comp.c
@@ -3217,16 +3217,21 @@ static void restore_regs(const struct btf_func_model *m, u8 **prog,
 }
 
 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
+			   struct bpf_tramp_image *im,
 			   struct bpf_tramp_node *node, int stack_size,
 			   int run_ctx_off, bool save_ret,
 			   void *image, void *rw_image)
 {
 	u8 *prog = *pprog;
-	u8 *jmp_insn;
+	u8 *jmp_insn, *skip;
 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
 	struct bpf_prog *p = node->link->prog;
 	u64 cookie = node->cookie;
 
+	/* nop, patched to skip this prog when it is detached */
+	skip = image + (prog - (u8 *)rw_image);
+	emit_nops(&prog, X86_PATCH_SIZE);
+
 	/* mov rdi, cookie */
 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
 
@@ -3301,6 +3306,8 @@ static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
 	if (emit_rsb_call(&prog, bpf_trampoline_exit(p), image + (prog - (u8 *)rw_image)))
 		return -EINVAL;
 
+	bpf_tramp_image_add_skip(im, p, skip, image + (prog - (u8 *)rw_image));
+
 	*pprog = prog;
 	return 0;
 }
@@ -3332,6 +3339,7 @@ static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
 }
 
 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
+		      struct bpf_tramp_image *im,
 		      struct bpf_tramp_nodes *tl, int stack_size,
 		      int run_ctx_off, int func_meta_off, bool save_ret,
 		      void *image, void *rw_image, u64 func_meta,
@@ -3346,7 +3354,7 @@ static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
 				func_meta | (cur_cookie << BPF_TRAMP_COOKIE_INDEX_SHIFT));
 			cur_cookie--;
 		}
-		if (invoke_bpf_prog(m, &prog, tl->nodes[i], stack_size,
+		if (invoke_bpf_prog(m, &prog, im, tl->nodes[i], stack_size,
 				    run_ctx_off, save_ret, image, rw_image))
 			return -EINVAL;
 	}
@@ -3355,6 +3363,7 @@ static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
 }
 
 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
+			      struct bpf_tramp_image *im,
 			      struct bpf_tramp_nodes *tl, int stack_size,
 			      int run_ctx_off, u8 **branches,
 			      void *image, void *rw_image)
@@ -3368,7 +3377,7 @@ static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
 	for (i = 0; i < tl->nr_nodes; i++) {
-		if (invoke_bpf_prog(m, &prog, tl->nodes[i], stack_size, run_ctx_off, true,
+		if (invoke_bpf_prog(m, &prog, im, tl->nodes[i], stack_size, run_ctx_off, true,
 				    image, rw_image))
 			return -EINVAL;
 
@@ -3640,7 +3649,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_im
 	}
 
 	if (fentry->nr_nodes) {
-		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off, func_meta_off,
+		if (invoke_bpf(m, &prog, im, fentry, regs_off, run_ctx_off, func_meta_off,
 			       flags & BPF_TRAMP_F_RET_FENTRY_RET, image, rw_image,
 			       func_meta, cookie_off))
 			return -EINVAL;
@@ -3652,7 +3661,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_im
 		if (!branches)
 			return -ENOMEM;
 
-		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
+		if (invoke_bpf_mod_ret(m, &prog, im, fmod_ret, regs_off,
 				       run_ctx_off, branches, image, rw_image)) {
 			ret = -EINVAL;
 			goto cleanup;
@@ -3708,7 +3717,7 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_im
 		emit_store_stack_imm64(&prog, BPF_REG_0, -func_meta_off, func_meta);
 
 	if (fexit->nr_nodes) {
-		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, func_meta_off,
+		if (invoke_bpf(m, &prog, im, fexit, regs_off, run_ctx_off, func_meta_off,
 			       false, image, rw_image, func_meta, cookie_off)) {
 			ret = -EINVAL;
 			goto cleanup;
@@ -3811,7 +3820,7 @@ int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *i
 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
 			     struct bpf_tramp_nodes *tnodes, void *func_addr)
 {
-	struct bpf_tramp_image im;
+	struct bpf_tramp_image im = {};
 	void *image;
 	int ret;
 
diff --git a/include/linux/bpf.h b/include/linux/bpf.h
index e57af902560c..9dcd3bc395eb 100644
--- a/include/linux/bpf.h
+++ b/include/linux/bpf.h
@@ -1258,11 +1258,15 @@ struct btf_func_model {
 #define BPF_TRAMP_F_INDIRECT		BIT(8)
 
 /* Each call __bpf_prog_enter + call bpf_func + call __bpf_prog_exit is ~50
- * bytes on x86.
+ * bytes on x86. The trampoline image has to fit in PAGE_SIZE.
  */
 enum {
-#if defined(__s390x__)
+#if defined(__s390x__) || defined(__powerpc64__)
 	BPF_MAX_TRAMP_LINKS = 27,
+#elif defined(__loongarch__)
+	BPF_MAX_TRAMP_LINKS = 33,
+#elif defined(__aarch64__)
+	BPF_MAX_TRAMP_LINKS = 37,
 #else
 	BPF_MAX_TRAMP_LINKS = 38,
 #endif
@@ -1363,6 +1367,17 @@ enum bpf_tramp_prog_type {
 	BPF_TRAMP_FSESSION,
 };
 
+/*
+ * Each prog call in a trampoline image is preceded by a nop. When the prog is
+ * detached, the nop is patched to a jump to target, right after the call, so
+ * that tasks still running in the image skip the prog.
+ */
+struct bpf_tramp_skip {
+	struct bpf_prog *prog;
+	void *nop;
+	void *target;
+};
+
 struct bpf_tramp_image {
 	void *image;
 	int size;
@@ -1374,8 +1389,27 @@ struct bpf_tramp_image {
 		struct rcu_head rcu;
 		struct work_struct work;
 	};
+	/* entry in tr->images, the image holds a reference on tr */
+	struct bpf_trampoline *tr;
+	struct list_head list;
+	struct bpf_tramp_skip *skips;
+	int nr_skips;
 };
 
+static inline void bpf_tramp_image_add_skip(struct bpf_tramp_image *im, struct bpf_prog *prog,
+					    void *nop, void *target)
+{
+	struct bpf_tramp_skip *skip;
+
+	/* struct_ops trampolines and arch_bpf_trampoline_size() have no image */
+	if (!im || !im->skips)
+		return;
+	skip = &im->skips[im->nr_skips++];
+	skip->prog = prog;
+	skip->nop = nop;
+	skip->target = target;
+}
+
 struct bpf_trampoline {
 	/* hlist for trampoline_key_table */
 	struct hlist_node hlist_key;
@@ -1402,6 +1436,8 @@ struct bpf_trampoline {
 	int progs_cnt[BPF_TRAMP_MAX];
 	/* Executable image of trampoline */
 	struct bpf_tramp_image *cur_image;
+	/* Images not freed yet, cur_image and older ones still in use */
+	struct list_head images;
 	/* Used as temporary old image storage for multi_attach */
 	struct {
 		struct bpf_tramp_image *old_image;
diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
index 90b70ea0d370..b82d8592319a 100644
--- a/kernel/bpf/trampoline.c
+++ b/kernel/bpf/trampoline.c
@@ -403,6 +403,7 @@ static struct bpf_trampoline *bpf_trampoline_lookup(u64 key, unsigned long ip)
 	refcount_set(&tr->refcnt, 1);
 	for (i = 0; i < BPF_TRAMP_MAX; i++)
 		INIT_HLIST_HEAD(&tr->progs_hlist[i]);
+	INIT_LIST_HEAD(&tr->images);
 out:
 	mutex_unlock(&trampoline_mutex);
 	return tr;
@@ -565,14 +566,21 @@ static void bpf_tramp_image_free(struct bpf_tramp_image *im)
 	arch_free_bpf_trampoline(im->image, im->size);
 	bpf_jit_uncharge_modmem(im->size);
 	percpu_ref_exit(&im->pcref);
+	kfree(im->skips);
 	kfree_rcu(im, rcu);
 }
 
 static void __bpf_tramp_image_put_deferred(struct work_struct *work)
 {
+	struct bpf_trampoline *tr;
 	struct bpf_tramp_image *im;
 
 	im = container_of(work, struct bpf_tramp_image, work);
+	tr = im->tr;
+	trampoline_lock(tr);
+	list_del(&im->list);
+	trampoline_unlock(tr);
+	bpf_trampoline_put(tr);
 	bpf_tramp_image_free(im);
 }
 
@@ -658,7 +666,7 @@ static void bpf_tramp_image_put(struct bpf_tramp_image *im)
 	call_rcu_tasks_trace(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
 }
 
-static struct bpf_tramp_image *bpf_tramp_image_alloc(u64 key, int size)
+static struct bpf_tramp_image *bpf_tramp_image_alloc(u64 key, int size, int nr_progs)
 {
 	struct bpf_tramp_image *im;
 	struct bpf_ksym *ksym;
@@ -669,6 +677,10 @@ static struct bpf_tramp_image *bpf_tramp_image_alloc(u64 key, int size)
 	if (!im)
 		goto out;
 
+	im->skips = kzalloc_objs(*im->skips, nr_progs);
+	if (!im->skips)
+		goto out_free_im;
+
 	err = bpf_jit_charge_modmem(size);
 	if (err)
 		goto out_free_im;
@@ -695,11 +707,37 @@ static struct bpf_tramp_image *bpf_tramp_image_alloc(u64 key, int size)
 out_uncharge:
 	bpf_jit_uncharge_modmem(size);
 out_free_im:
+	kfree(im->skips);
 	kfree(im);
 out:
 	return ERR_PTR(err);
 }
 
+/*
+ * prog was detached and can be freed, but tasks may still be running in images
+ * that call it, sleeping in an earlier prog for example. Patch these images to
+ * jump over prog.
+ */
+static void bpf_trampoline_skip_prog(struct bpf_trampoline *tr, struct bpf_prog *prog)
+{
+	struct bpf_tramp_image *im;
+	int i, err;
+
+	list_for_each_entry(im, &tr->images, list) {
+		for (i = 0; i < im->nr_skips; i++) {
+			struct bpf_tramp_skip *skip = &im->skips[i];
+
+			if (skip->prog != prog)
+				continue;
+			err = bpf_arch_text_poke(skip->nop, BPF_MOD_NOP, BPF_MOD_JUMP,
+						 NULL, skip->target);
+			WARN_ON_ONCE(err);
+			/* not a nop anymore, and prog's address can be reused */
+			skip->prog = NULL;
+		}
+	}
+}
+
 void bpf_trampoline_set_flags(struct bpf_trampoline *tr, u32 flags)
 {
 	trampoline_lock(tr);
@@ -771,7 +809,7 @@ static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mut
 		goto out;
 	}
 
-	im = bpf_tramp_image_alloc(tr->key, size);
+	im = bpf_tramp_image_alloc(tr->key, size, total);
 	if (IS_ERR(im)) {
 		err = PTR_ERR(im);
 		goto out;
@@ -806,8 +844,14 @@ static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mut
 #endif
 
 out_free:
-	if (err)
+	if (err) {
 		bpf_tramp_image_free(im);
+	} else {
+		/* track the image until it is freed, for bpf_trampoline_skip_prog() */
+		refcount_inc(&tr->refcnt);
+		im->tr = tr;
+		list_add(&im->list, &tr->images);
+	}
 out:
 	/* If any error happens, restore previous flags */
 	if (err)
@@ -937,6 +981,7 @@ static void bpf_trampoline_remove_prog(struct bpf_trampoline *tr,
 	}
 	hlist_del_init(&node->tramp_hlist);
 	tr->progs_cnt[kind]--;
+	bpf_trampoline_skip_prog(tr, node->link->prog);
 }
 
 static int __bpf_trampoline_link_prog(struct bpf_tramp_node *node,
-- 
2.55.0