1
0
Fork 0

powerpc/bpf: Fix tail call implementation

We have seen many crashes on powerpc hosts while loading bpf programs.

The problem here is that bpf_int_jit_compile() does a first pass
to compute the program length.

Then it allocates memory to store the generated program and
calls bpf_jit_build_body() a second time (and a third time
later)

What I have observed is that the second bpf_jit_build_body()
could end up using few more words than expected.

If bpf_jit_binary_alloc() put the space for the program
at the end of the allocated page, we then write on
a non mapped memory.

It appears that bpf_jit_emit_tail_call() calls
bpf_jit_emit_common_epilogue() while ctx->seen might not
be stable.

Only after the second pass we can be sure ctx->seen wont be changed.

Trying to avoid a second pass seems quite complex and probably
not worth it.

Fixes: ce0761419f ("powerpc/bpf: Implement support for tail calls")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Naveen N. Rao <naveen.n.rao@linux.ibm.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Martin KaFai Lau <kafai@fb.com>
Cc: Song Liu <songliubraving@fb.com>
Cc: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20191101033444.143741-1-edumazet@google.com
alistair/sunxi64-5.4-dsi
Eric Dumazet 2019-10-31 20:34:44 -07:00 committed by Daniel Borkmann
parent ff1c08e1f7
commit 7de0869093
1 changed files with 13 additions and 0 deletions

View File

@ -1141,6 +1141,19 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
goto out_addrs;
}
/*
* If we have seen a tail call, we need a second pass.
* This is because bpf_jit_emit_common_epilogue() is called
* from bpf_jit_emit_tail_call() with a not yet stable ctx->seen.
*/
if (cgctx.seen & SEEN_TAILCALL) {
cgctx.idx = 0;
if (bpf_jit_build_body(fp, 0, &cgctx, addrs, false)) {
fp = org_fp;
goto out_addrs;
}
}
/*
* Pretend to build prologue, given the features we've seen. This will
* update ctgtx.idx as it pretends to output instructions, then we can