bpf, x86, arm64: Enable jit by default when not built as always-on

After Spectre 2 fix via 290af86 ("bpf: introduce BPF_JIT_ALWAYS_ON
config") most major distros use BPF_JIT_ALWAYS_ON configuration these days
which compiles out the BPF interpreter entirely and always enables the
JIT. Also given recent fix in e1608f3 ("bpf: Avoid setting bpf insns
pages read-only when prog is jited"), we additionally avoid fragmenting
the direct map for the BPF insns pages sitting in the general data heap
since they are not used during execution. Latter is only needed when run
through the interpreter.

Since both x86 and arm64 JITs have seen a lot of exposure over the years,
are generally most up to date and maintained, there is more downside in
!BPF_JIT_ALWAYS_ON configurations to have the interpreter enabled by default
rather than the JIT. Add a ARCH_WANT_DEFAULT_BPF_JIT config which archs can
use to set the bpf_jit_{enable,kallsyms} to 1. Back in the days the
bpf_jit_kallsyms knob was set to 0 by default since major distros still
had /proc/kallsyms addresses exposed to unprivileged user space which is
not the case anymore. Hence both knobs are set via BPF_JIT_DEFAULT_ON which
is set to 'y' in case of BPF_JIT_ALWAYS_ON or ARCH_WANT_DEFAULT_BPF_JIT.

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Will Deacon <will@kernel.org>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Link: https://lore.kernel.org/bpf/f78ad24795c2966efcc2ee19025fa3459f622185.1575903816.git.daniel@iogearbox.net

arch/arm64/Kconfig

@@ -69,6 +69,7 @@ config ARM64
 	select ARCH_SUPPORTS_INT128 if CC_HAS_INT128 && (GCC_VERSION >= 50000 || CC_IS_CLANG)
 	select ARCH_SUPPORTS_NUMA_BALANCING
 	select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
+	select ARCH_WANT_DEFAULT_BPF_JIT
 	select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
 	select ARCH_WANT_FRAME_POINTERS
 	select ARCH_WANT_HUGE_PMD_SHARE if ARM64_4K_PAGES || (ARM64_16K_PAGES && !ARM64_VA_BITS_36)

arch/x86/Kconfig

@@ -93,6 +93,7 @@ config X86
 	select ARCH_USE_QUEUED_RWLOCKS
 	select ARCH_USE_QUEUED_SPINLOCKS
 	select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
+	select ARCH_WANT_DEFAULT_BPF_JIT	if X86_64
 	select ARCH_WANTS_DYNAMIC_TASK_STRUCT
 	select ARCH_WANT_HUGE_PMD_SHARE
 	select ARCH_WANTS_THP_SWAP		if X86_64

init/Kconfig

@@ -1604,13 +1604,20 @@ config BPF_SYSCALL
 	  Enable the bpf() system call that allows to manipulate eBPF
 	  programs and maps via file descriptors.
 
+config ARCH_WANT_DEFAULT_BPF_JIT
+	bool
+
 config BPF_JIT_ALWAYS_ON
 	bool "Permanently enable BPF JIT and remove BPF interpreter"
 	depends on BPF_SYSCALL && HAVE_EBPF_JIT && BPF_JIT
 	help
 	  Enables BPF JIT and removes BPF interpreter to avoid
 	  speculative execution of BPF instructions by the interpreter
 
+config BPF_JIT_DEFAULT_ON
+	def_bool ARCH_WANT_DEFAULT_BPF_JIT || BPF_JIT_ALWAYS_ON
+	depends on HAVE_EBPF_JIT && BPF_JIT
+
 config USERFAULTFD
 	bool "Enable userfaultfd() system call"
 	depends on MMU

kernel/bpf/core.c

@@ -520,9 +520,9 @@ void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
 
 #ifdef CONFIG_BPF_JIT
 /* All BPF JIT sysctl knobs here. */
-int bpf_jit_enable   __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_ALWAYS_ON);
+int bpf_jit_enable   __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
+int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
 int bpf_jit_harden   __read_mostly;
-int bpf_jit_kallsyms __read_mostly;
 long bpf_jit_limit   __read_mostly;
 
 static __always_inline void