net: bpf: arm64: address randomize and write protect JIT code
authorDaniel Borkmann <dborkman@redhat.com>
Tue, 16 Sep 2014 07:48:50 +0000 (08:48 +0100)
committerCatalin Marinas <catalin.marinas@arm.com>
Mon, 20 Oct 2014 16:47:03 +0000 (17:47 +0100)
This is the ARM64 variant for 314beb9bcab ("x86: bpf_jit_comp: secure bpf
jit against spraying attacks").

Thanks to commit 11d91a770f1f ("arm64: Add CONFIG_DEBUG_SET_MODULE_RONX
support") which added necessary infrastructure, we can now implement
RO marking of eBPF generated JIT image pages and randomize start offset
for the JIT code, so that it does not reside directly on a page boundary
anymore. Likewise, the holes are filled with illegal instructions: here
we use BRK #0x100 (opcode 0xd4202000) to trigger a fault in the kernel
(unallocated BRKs would trigger a fault through do_debug_exception). This
seems more reliable as we don't have a guaranteed undefined instruction
space on ARM64.

This is basically the ARM64 variant of what we already have in ARM via
commit 55309dd3d4cd ("net: bpf: arm: address randomize and write protect
JIT code"). Moreover, this commit also presents a merge resolution due to
conflicts with commit 60a3b2253c41 ("net: bpf: make eBPF interpreter images
read-only") as we don't use kfree() in bpf_jit_free() anymore to release
the locked bpf_prog structure, but instead bpf_prog_unlock_free() through
a different allocator.

JIT tested on aarch64 with BPF test suite.

Reference: http://mainisusuallyafunction.blogspot.com/2012/11/attacking-hardened-linux-systems-with.html
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Reviewed-by: Zi Shen Lim <zlim.lnx@gmail.com>
Acked-by: Will Deacon <will.deacon@arm.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
arch/arm64/net/bpf_jit_comp.c

index 7ae33545535b96fd363d08ede9634d8ff0a30ad3..71088952ed270eacd1249273fe1798c02993b332 100644 (file)
 #define pr_fmt(fmt) "bpf_jit: " fmt
 
 #include <linux/filter.h>
-#include <linux/moduleloader.h>
 #include <linux/printk.h>
 #include <linux/skbuff.h>
 #include <linux/slab.h>
+
 #include <asm/byteorder.h>
 #include <asm/cacheflush.h>
+#include <asm/debug-monitors.h>
 
 #include "bpf_jit.h"
 
@@ -119,6 +120,14 @@ static inline int bpf2a64_offset(int bpf_to, int bpf_from,
        return to - from;
 }
 
+static void jit_fill_hole(void *area, unsigned int size)
+{
+       u32 *ptr;
+       /* We are guaranteed to have aligned memory. */
+       for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
+               *ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT);
+}
+
 static inline int epilogue_offset(const struct jit_ctx *ctx)
 {
        int to = ctx->offset[ctx->prog->len - 1];
@@ -613,8 +622,10 @@ void bpf_jit_compile(struct bpf_prog *prog)
 
 void bpf_int_jit_compile(struct bpf_prog *prog)
 {
+       struct bpf_binary_header *header;
        struct jit_ctx ctx;
        int image_size;
+       u8 *image_ptr;
 
        if (!bpf_jit_enable)
                return;
@@ -636,23 +647,25 @@ void bpf_int_jit_compile(struct bpf_prog *prog)
                goto out;
 
        build_prologue(&ctx);
-
        build_epilogue(&ctx);
 
        /* Now we know the actual image size. */
        image_size = sizeof(u32) * ctx.idx;
-       ctx.image = module_alloc(image_size);
-       if (unlikely(ctx.image == NULL))
+       header = bpf_jit_binary_alloc(image_size, &image_ptr,
+                                     sizeof(u32), jit_fill_hole);
+       if (header == NULL)
                goto out;
 
        /* 2. Now, the actual pass. */
 
+       ctx.image = (u32 *)image_ptr;
        ctx.idx = 0;
+
        build_prologue(&ctx);
 
        ctx.body_offset = ctx.idx;
        if (build_body(&ctx)) {
-               module_free(NULL, ctx.image);
+               bpf_jit_binary_free(header);
                goto out;
        }
 
@@ -663,17 +676,25 @@ void bpf_int_jit_compile(struct bpf_prog *prog)
                bpf_jit_dump(prog->len, image_size, 2, ctx.image);
 
        bpf_flush_icache(ctx.image, ctx.image + ctx.idx);
+
+       set_memory_ro((unsigned long)header, header->pages);
        prog->bpf_func = (void *)ctx.image;
        prog->jited = 1;
-
 out:
        kfree(ctx.offset);
 }
 
 void bpf_jit_free(struct bpf_prog *prog)
 {
-       if (prog->jited)
-               module_free(NULL, prog->bpf_func);
+       unsigned long addr = (unsigned long)prog->bpf_func & PAGE_MASK;
+       struct bpf_binary_header *header = (void *)addr;
+
+       if (!prog->jited)
+               goto free_filter;
+
+       set_memory_rw(addr, header->pages);
+       bpf_jit_binary_free(header);
 
-       kfree(prog);
+free_filter:
+       bpf_prog_unlock_free(prog);
 }