BPF List
 help / color / mirror / Atom feed
From: Vineet Gupta <vineet.gupta@linux.dev>
To: bpf@vger.kernel.org
Cc: ast@kernel.org, daniel@iogearbox.net, andrii@kernel.org,
	eddyz87@gmail.com, memxor@gmail.com, martin.lau@linux.dev,
	song@kernel.org, yonghong.song@linux.dev, jolsa@kernel.org,
	emil@etsalapatis.com, ihor.solodrai@linux.dev,
	john.fastabend@gmail.com, shuah@kernel.org,
	linux-kernel@vger.kernel.org, linux-kselftest@vger.kernel.org,
	jose.marchesi@oracle.com, david.faust@oss.qualcomm.com,
	Vineet Gupta <vineet.gupta@linux.dev>
Subject: [PATCH bpf-next v2 7/8] selftests/bpf: write the mixed arena/scalar ALU tests in asm
Date: Wed, 28 Oct 2026 13:41:57 -0700	[thread overview]
Message-ID: <20261028204158.1021284-8-vineet.gupta@linux.dev> (raw)
In-Reply-To: <20261028204158.1021284-1-vineet.gupta@linux.dev>

They fail to build with BPF_GCC:

  progs/verifier_arena.c:566:5: error: optimization may eliminate reads and/or writes to register variables [-Werror=volatile-register-var]

for

  volatile register __u64 reg asm("r3");

and GCC means it.  With the warning silenced it drops both scalar
accesses, leaving nothing that uses r3 as a scalar:

  call	7
  if w0 == 0 goto .L136
  r9 = arena ll;r3 = 0;r3 = addr_space_cast(r3, 0, 1);
  .L136:
  r0 = 0
  exit

`reg = 1ULL << 32' and `reg += 1' are both gone.  The tests need one
instruction reached with r3 live as a scalar on one path and as an
arena pointer on the other; what is left gives the verifier no reason
to reject anything, so the __failure expectation would not hold.  A
local register variable is not guaranteed to survive a helper call
either, which GCC is also entitled to act on.

Nothing here needs a compiler to allocate registers, so write the
whole body in asm like mixed_arena_stack_atomic below, which tests the
same shape.  Both compilers then emit the same sequence, with the
branch landing on the shared ALU op:

  call 7
  r6=r0
  r3=1
  r3<<=32
  if w6==0 goto 4          <- lands on r3+=1
  r9=0 ll
  r3=0
  r3 = addr_space_cast(r3, 0, 1)
  r3+=1                    <- scalar on one path, arena pointer on the other
  r0=0
  exit

Take __load_if_JITed() from the sibling while here.  A program binding
an arena map is rejected in check_map_prog_compatibility() with "JIT is
required to use arena" during map resolution, before do_check() can
reach the shared instruction, so without a JIT these would fail on the
wrong message rather than skip.

Not exercised by a stock bpf-gcc, which compiles both tests out with the
rest of the arena block; the build failure above needs a compiler that
has arena support.  Under clang the rewrite is not a no-op either: the
pre-image spilled r3 to the stack and did the shared ALU op on r1
reloaded from it, so the test was not exercising one register live as
both a scalar and an arena pointer.

Signed-off-by: Vineet Gupta <vineet.gupta@linux.dev>
---
 .../selftests/bpf/progs/verifier_arena.c      | 85 ++++++++++---------
 1 file changed, 45 insertions(+), 40 deletions(-)

diff --git a/tools/testing/selftests/bpf/progs/verifier_arena.c b/tools/testing/selftests/bpf/progs/verifier_arena.c
index 21b66d51c61e..b84d8f1e0b39 100644
--- a/tools/testing/selftests/bpf/progs/verifier_arena.c
+++ b/tools/testing/selftests/bpf/progs/verifier_arena.c
@@ -539,53 +539,58 @@ int arena_ptr_add_arena_ptr(void *ctx)
 	return 0;
 }
 
+/*
+ * The r9 load is never read; it is there so libbpf records the arena map
+ * as used, which is what lets the verifier accept addr_space_cast at all.
+ */
 SEC("syscall")
 __failure __msg("same insn cannot be used with and without arena pointer")
-int mixed_arena_scalar_alu64_scalar_first(void *ctx)
-{
-	volatile register __u64 reg asm("r3");
-	__u32 pick_arena = bpf_get_prandom_u32();
-
-	reg = 1ULL << 32;
-
-	if (pick_arena) {
-		asm volatile (
-			"r9 = %[arena] ll;"
-			"%[reg] = 0;"
-			"%[reg] = addr_space_cast(%[reg], 0x0, 0x1);"
-			: [reg] "=r"(reg)
-			: __imm_addr(arena)
-			: "r9"
-		);
-	}
-
-	reg += 1;
-
-	return 0;
+__load_if_JITed()
+__naked void mixed_arena_scalar_alu64_scalar_first(void)
+{
+	asm volatile ("					\
+	call %[bpf_get_prandom_u32];			\
+	r6 = r0;					\
+	r3 = 1;						\
+	r3 <<= 32;					\
+	if w6 == 0 goto 1f;				\
+	r9 = %[arena] ll;				\
+	r3 = 0;						\
+	r3 = addr_space_cast(r3, 0, 1);			\
+1:	r3 += 1;					\
+	r0 = 0;						\
+	exit;						\
+"	:
+	: __imm_addr(arena),
+	  __imm(bpf_get_prandom_u32)
+	: __clobber_all);
 }
 
+/*
+ * The r9 load is never read; it is there so libbpf records the arena map
+ * as used, which is what lets the verifier accept addr_space_cast at all.
+ */
 SEC("syscall")
 __failure __msg("same insn cannot be used with and without arena pointer")
-int mixed_arena_scalar_alu64_arena_first(void *ctx)
+__load_if_JITed()
+__naked void mixed_arena_scalar_alu64_arena_first(void)
 {
-	volatile register __u64 reg asm("r3");
-	__u32 pick_scalar = bpf_get_prandom_u32();
-
-	asm volatile (
-		"r9 = %[arena] ll;"
-		"%[reg] = 0;"
-		"%[reg] = addr_space_cast(%[reg], 0x0, 0x1);"
-		: [reg] "=r"(reg)
-		: __imm_addr(arena)
-		: "r9"
-	);
-
-	if (pick_scalar)
-		reg = 1ULL << 32;
-
-	reg += 1;
-
-	return 0;
+	asm volatile ("					\
+	call %[bpf_get_prandom_u32];			\
+	r6 = r0;					\
+	r9 = %[arena] ll;				\
+	r3 = 0;						\
+	r3 = addr_space_cast(r3, 0, 1);			\
+	if w6 == 0 goto 1f;				\
+	r3 = 1;						\
+	r3 <<= 32;					\
+1:	r3 += 1;					\
+	r0 = 0;						\
+	exit;						\
+"	:
+	: __imm_addr(arena),
+	  __imm(bpf_get_prandom_u32)
+	: __clobber_all);
 }
 
 SEC("syscall")
-- 
2.53.0-Meta


  parent reply	other threads:[~2026-10-28 20:42 UTC|newest]

Thread overview: 13+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-10-28 20:41 [PATCH bpf-next v2 0/8] selftests/bpf: cleanup bpf-gcc runs Vineet Gupta
2026-10-28 20:41 ` [PATCH bpf-next v2 1/8] selftests/bpf: Build callx tests with GCC that defines __BPF_FEATURE_CALLX Vineet Gupta
2026-10-28 21:22   ` bot+bpf-ci
2026-10-28 22:28     ` Vineet Gupta
2026-10-28 20:41 ` [PATCH bpf-next v2 2/8] selftests/bpf: don't pass clang-only warning flags to GCC Vineet Gupta
2026-10-28 20:41 ` [PATCH bpf-next v2 3/8] selftests/bpf: report SKIP for arena tests a compiler cannot build Vineet Gupta
2026-10-28 20:41 ` [PATCH bpf-next v2 4/8] selftests/bpf: make arena_base() opaque to GCC's -Warray-bounds Vineet Gupta
2026-10-28 20:41 ` [PATCH bpf-next v2 5/8] selftests/bpf: don't report arena tests as passing with an empty body Vineet Gupta
2026-10-28 20:41 ` [PATCH bpf-next v2 6/8] selftests/bpf: cast explicitly in the arena subprog and callback faults Vineet Gupta
2026-10-28 20:41 ` Vineet Gupta [this message]
2026-10-28 21:22   ` [PATCH bpf-next v2 7/8] selftests/bpf: write the mixed arena/scalar ALU tests in asm bot+bpf-ci
2026-10-28 23:14     ` Vineet Gupta
2026-10-28 20:41 ` [PATCH bpf-next v2 8/8] selftests/bpf: skip the struct_ops arena tests without compiler support Vineet Gupta

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20261028204158.1021284-8-vineet.gupta@linux.dev \
    --to=vineet.gupta@linux.dev \
    --cc=andrii@kernel.org \
    --cc=ast@kernel.org \
    --cc=bpf@vger.kernel.org \
    --cc=daniel@iogearbox.net \
    --cc=david.faust@oss.qualcomm.com \
    --cc=eddyz87@gmail.com \
    --cc=emil@etsalapatis.com \
    --cc=ihor.solodrai@linux.dev \
    --cc=john.fastabend@gmail.com \
    --cc=jolsa@kernel.org \
    --cc=jose.marchesi@oracle.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-kselftest@vger.kernel.org \
    --cc=martin.lau@linux.dev \
    --cc=memxor@gmail.com \
    --cc=shuah@kernel.org \
    --cc=song@kernel.org \
    --cc=yonghong.song@linux.dev \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox