Stream tests fail with BPF_GCC, e.g.

  test_stream_arena_fault_address:FAIL:stream__open_and_load unexpected error: -13

They call stream__open_and_load(), which loads every program in the object,
and two of those do not verify:

  #481/11 stream_success/stream_arena_subprog_fault:FAIL
  #481/12 stream_success/stream_arena_callback_fault:FAIL

Both dereference an arena pointer and leave the cast to the compiler:

  int __arena *addr = (int __arena *)0xdeadbeef;
  *addr = 1;

Without arena support __arena is only btf_type_tag("arena"), so no
addr_space_cast is emitted and the store is to a scalar, which the
verifier rejects. GCC splits it into byte stores, 0xdeadbeef not
being aligned:

  r2=3735928559 ll
  *(u8*)(r2+0)=1            <- plain store, no cast

Their siblings in the same file do not rely on the compiler; they use
bpf_addr_space_cast(), which emits the instruction byte by byte and
works anywhere. Do the same here.

The address space has to come off the local as well. Left on, a
compiler that supports arenas casts a second time for the
dereference, on a pointer the explicit cast has already converted:

  r2 = addr_space_cast(r2, 0, 1)
  r3 = addr_space_cast(r2, 0, 1)
  *(u32*)(r3+0)=1

With it off, GCC with and without arena support and clang all emit the
same thing, which is what the test wants:

  r2 = addr_space_cast(r2, 0, 1)
  *(u32*)(r2+0)=1

What these two assert is the fault report, not how the pointer was
produced, so nothing they were testing is lost.

Measured on bpf-next with a stock bpf-gcc, test_progs before and
after. Every test that loads this object recovers, not just the arena
ones:

  stream_arena_fault_address                   FAIL -> OK
  stream_fd                                    FAIL -> OK
  stream_fd/blocking                           FAIL -> OK
  stream_fd/edge                               FAIL -> OK
  stream_fd/empty                              FAIL -> OK
  stream_fd/hup                                FAIL -> OK
  stream_fd/nmi_epoll                          FAIL -> OK
  stream_fd/nonblock                           FAIL -> OK
  stream_fd/open                               FAIL -> OK
  stream_oversize                              FAIL -> OK
  stream_partial_read                          FAIL -> OK
  stream_success                               FAIL -> OK
  stream_success/stream_arena_callback_fault   FAIL -> OK
  stream_success/stream_arena_subprog_fault    FAIL -> OK
  stream_syscall                               FAIL -> OK

and the subtests of the object that failed to load appear for the
first time:

  stream_arena_fault_address/read_fault             -> OK
  stream_arena_fault_address/write_fault            -> OK
  stream_arena_fault_address/xchg_fault             -> OK
  stream_arena_fault_address/cmpxchg_fault          -> OK
  stream_arena_fault_address/load_acquire_fault     -> OK

Signed-off-by: Vineet Gupta <vineet.gupta@linux.dev>
---
v3 -> v4: was 6/8. No longer drops #include <bpf_arena_common.h>,
which the arena_base() patch now needs in this file.

 tools/testing/selftests/bpf/progs/stream.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/tools/testing/selftests/bpf/progs/stream.c b/tools/testing/selftests/bpf/progs/stream.c
index b4d88d0e01ad..8c49a130150b 100644
--- a/tools/testing/selftests/bpf/progs/stream.c
+++ b/tools/testing/selftests/bpf/progs/stream.c
@@ -341,9 +341,10 @@ int stream_arena_cmpxchg_fault(void *ctx)
 
 static __noinline void subprog(void)
 {
-	int __arena *addr = (int __arena *)0xdeadbeef;
+	int *addr = (int *)0xdeadbeef;
 
 	arena_ptr = &arena;
+	bpf_addr_space_cast(addr, 0, 1);
 	*addr = 1;
 }
 
@@ -364,9 +365,10 @@ int stream_arena_subprog_fault(void *ctx)
 
 static __noinline int timer_cb(void *map, int *key, struct bpf_timer *timer)
 {
-	int __arena *addr = (int __arena *)0xdeadbeef;
+	int *addr = (int *)0xdeadbeef;
 
 	arena_ptr = &arena;
+	bpf_addr_space_cast(addr, 0, 1);
 	*addr = 1;
 	return 0;
 }
-- 
2.53.0-Meta