Skip to contentSkip to conversation

bpf: call get_random_u32() for random integers

PATCH In review Jason A. Donenfeld 3 participants 8 messages Dec 5, 2022 – Dec 6, 2022

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bpf: call get_random_u32() for random integers

Since BPF's bpf_user_rnd_u32() was introduced, there have been three significant developments in the RNG: 1) get_random_u32() returns the same types of bytes as /dev/urandom, eliminating the distinction between "kernel random bytes" and "userspace random bytes", 2) get_random_u32() operates mostly locklessly over percpu state, 3) get_random_u32() has become quite fast.

So rather than using the old clunky Tausworthe prandom code, just call get_random_u32(), which should fit BPF uses perfectly.

Cc Daniel Borkmann <[email protected]>
Cc Andrii Nakryiko <[email protected]>
Cc Martin KaFai Lau <[email protected]>
---
 include/linux/bpf.h   |  1 -
 kernel/bpf/core.c     | 17 +----------------
 kernel/bpf/verifier.c |  2 --
 net/core/filter.c     |  1 -
 4 files changed, 1 insertion(+), 20 deletions(-)
include/linux/bpf.h +0 -1
index 0566705c1d4e..aae89318789a 100644
@@ -2554,7 +2554,6 @@ const struct bpf_func_proto *tracing_prog_func_proto(
   enum bpf_func_id func_id, const struct bpf_prog *prog);
 
 /* Shared helpers among cBPF and eBPF. */
-void bpf_user_rnd_init_once(void);
 u64 bpf_user_rnd_u32(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
 u64 bpf_get_raw_cpu_id(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
 
kernel/bpf/core.c +1 -16
index 38159f39e2af..2cc28d63d761 100644
@@ -2579,14 +2579,6 @@ void bpf_prog_free(struct bpf_prog *fp)
 }
 EXPORT_SYMBOL_GPL(bpf_prog_free);
 
-/* RNG for unpriviledged user space with separated state from prandom_u32(). */
-static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
-
-void bpf_user_rnd_init_once(void)
-{
-	prandom_init_once(&bpf_user_rnd_state);
-}
-
 BPF_CALL_0(bpf_user_rnd_u32)
 {
 	/* Should someone ever have the rather unwise idea to use some
@@ -2595,14 +2587,7 @@ BPF_CALL_0(bpf_user_rnd_u32)
 	 * transformations. Register assignments from both sides are
 	 * different, f.e. classic always sets fn(ctx, A, X) here.
 	 */
-	struct rnd_state *state;
-	u32 res;
-
-	state = &get_cpu_var(bpf_user_rnd_state);
-	res = prandom_u32_state(state);
-	put_cpu_var(bpf_user_rnd_state);
-
-	return res;
+	return get_random_u32();
 }
 
 BPF_CALL_0(bpf_get_raw_cpu_id)
kernel/bpf/verifier.c +0 -2
index 225666307bba..75a1a6526165 100644
@@ -14045,8 +14045,6 @@ static int do_misc_fixups(struct bpf_verifier_env *env)
 
 		if (insn->imm == BPF_FUNC_get_route_realm)
 			prog->dst_needed = 1;
-		if (insn->imm == BPF_FUNC_get_prandom_u32)
-			bpf_user_rnd_init_once();
 		if (insn->imm == BPF_FUNC_override_return)
 			prog->kprobe_override = 1;
 		if (insn->imm == BPF_FUNC_tail_call) {
net/core/filter.c +0 -1
index bb0136e7a8e4..7a595ac0028d 100644
@@ -443,7 +443,6 @@ static bool convert_bpf_extensions(struct sock_filter *fp,
 			break;
 		case SKF_AD_OFF + SKF_AD_RANDOM:
 			*insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
-			bpf_user_rnd_init_once();
 			break;
 		}
 		break;
Signature · 1 line
2.38.1
maintainer
#1 lore raw
Quoted 6 lines · Jason A. Donenfeld

Since BPF's bpf_user_rnd_u32() was introduced, there have been three significant developments in the RNG: 1) get_random_u32() returns the same types of bytes as /dev/urandom, eliminating the distinction between "kernel random bytes" and "userspace random bytes", 2) get_random_u32() operates mostly locklessly over percpu state, 3) get_random_u32() has become quite fast.

Wrt "quite fast", do you have a comparison between the two? Asking as its often used in networking worst case on per packet basis (e.g. via XDP), would be useful to state concrete numbers for the two on a given machine.

Quoted 2 lines

So rather than using the old clunky Tausworthe prandom code, just call get_random_u32(), which should fit BPF uses perfectly.

#2 lore raw
Quoted 11 lines · Daniel Borkmann
Quoted 6 lines · Jason A. Donenfeld

Since BPF's bpf_user_rnd_u32() was introduced, there have been three significant developments in the RNG: 1) get_random_u32() returns the same types of bytes as /dev/urandom, eliminating the distinction between "kernel random bytes" and "userspace random bytes", 2) get_random_u32() operates mostly locklessly over percpu state, 3) get_random_u32() has become quite fast.

Wrt "quite fast", do you have a comparison between the two? Asking as its often used in networking worst case on per packet basis (e.g. via XDP), would be useful to state concrete numbers for the two on a given machine.

Median of 25 cycles vs median of 38, on my Tiger Lake machine. So a little slower, but too small of a difference to matter.

#3 lore raw
Quoted 9 lines · Jason A. Donenfeld
Quoted 8 lines · Daniel Borkmann
Quoted 6 lines · Jason A. Donenfeld

Since BPF's bpf_user_rnd_u32() was introduced, there have been three significant developments in the RNG: 1) get_random_u32() returns the same types of bytes as /dev/urandom, eliminating the distinction between "kernel random bytes" and "userspace random bytes", 2) get_random_u32() operates mostly locklessly over percpu state, 3) get_random_u32() has become quite fast.

Quoted 6 lines
Quoted 3 lines

Wrt "quite fast", do you have a comparison between the two? Asking as its often used in networking worst case on per packet basis (e.g. via XDP), would be useful to state concrete numbers for the two on a given machine.

Median of 25 cycles vs median of 38, on my Tiger Lake machine. So a little slower, but too small of a difference to matter.

Assuming a 3Ghz CPU clock (so 3 cycles per nanosecond), that's an additional overhead of ~4.3 ns. When processing 10 Gbps at line rate with small packets, the per-packet processing budget is 67.2 ns, so those extra 4.3 ns will eat up ~6.4% of the budget.

So in other words, "too small a difference to matter" is definitely not true in general. It really depends on the use case; if someone is using this to, say, draw per-packet random numbers to compute a drop frequency on ingress, that extra processing time will most likely result in a quite measurable drop in performance.

-Toke

#4 lore raw

Hi Toke,

Quoted 24 lines · Toke Høiland-Jørgensen
Quoted 9 lines · Jason A. Donenfeld
Quoted 8 lines · Daniel Borkmann
Quoted 6 lines · Jason A. Donenfeld

Since BPF's bpf_user_rnd_u32() was introduced, there have been three significant developments in the RNG: 1) get_random_u32() returns the same types of bytes as /dev/urandom, eliminating the distinction between "kernel random bytes" and "userspace random bytes", 2) get_random_u32() operates mostly locklessly over percpu state, 3) get_random_u32() has become quite fast.

Quoted 6 lines
Quoted 3 lines

Wrt "quite fast", do you have a comparison between the two? Asking as its often used in networking worst case on per packet basis (e.g. via XDP), would be useful to state concrete numbers for the two on a given machine.

Median of 25 cycles vs median of 38, on my Tiger Lake machine. So a little slower, but too small of a difference to matter.

Assuming a 3Ghz CPU clock (so 3 cycles per nanosecond), that's an additional overhead of ~4.3 ns. When processing 10 Gbps at line rate with small packets, the per-packet processing budget is 67.2 ns, so those extra 4.3 ns will eat up ~6.4% of the budget.

Quoted 5 lines

So in other words, "too small a difference to matter" is definitely not true in general. It really depends on the use case; if someone is using this to, say, draw per-packet random numbers to compute a drop frequency on ingress, that extra processing time will most likely result in a quite measurable drop in performance.

Huh, neat calculation, I'll keep that method in mind.

Alright, sorry for the noise here. I'll check back in if I ever manage to eliminate that performance gap.

Jason

#5 lore raw
Quoted 27 lines · Jason A. Donenfeld

Hi Toke,

Quoted 24 lines · Toke Høiland-Jørgensen
Quoted 9 lines · Jason A. Donenfeld
Quoted 8 lines · Daniel Borkmann
Quoted 6 lines · Jason A. Donenfeld

Since BPF's bpf_user_rnd_u32() was introduced, there have been three significant developments in the RNG: 1) get_random_u32() returns the same types of bytes as /dev/urandom, eliminating the distinction between "kernel random bytes" and "userspace random bytes", 2) get_random_u32() operates mostly locklessly over percpu state, 3) get_random_u32() has become quite fast.

Quoted 6 lines
Quoted 3 lines

Wrt "quite fast", do you have a comparison between the two? Asking as its often used in networking worst case on per packet basis (e.g. via XDP), would be useful to state concrete numbers for the two on a given machine.

Median of 25 cycles vs median of 38, on my Tiger Lake machine. So a little slower, but too small of a difference to matter.

Assuming a 3Ghz CPU clock (so 3 cycles per nanosecond), that's an additional overhead of ~4.3 ns. When processing 10 Gbps at line rate with small packets, the per-packet processing budget is 67.2 ns, so those extra 4.3 ns will eat up ~6.4% of the budget.

Quoted 7 lines
Quoted 5 lines

So in other words, "too small a difference to matter" is definitely not true in general. It really depends on the use case; if someone is using this to, say, draw per-packet random numbers to compute a drop frequency on ingress, that extra processing time will most likely result in a quite measurable drop in performance.

Huh, neat calculation, I'll keep that method in mind.

Yeah, I find that thinking in "time budget per packet" helps a lot! For completeness, the 67.2 ns comes from 10 Gbps / 84 bytes (that's 64-byte packets + 20 bytes of preamble and inter-packet gap), which is 14.8 Mpps, or 67.2 ns/pkt.

Quoted 2 lines

Alright, sorry for the noise here. I'll check back in if I ever manage to eliminate that performance gap.

One approach that we use generally for XDP (and which may or may not help for this case) is that we try to amortise as much fixed cost as we can over a batch of packets. Because XDP processing always happens in the NAPI receive poll loop, we have a natural batching mechanism, and we know that for that whole loop we'll keep running on the same CPU.

So for instance, if there's a large fixed component of the overhead of get_random_u32(), we could have bpf_user_rnd_u32() populate a larger per-CPU buffer and then just emit u32 chunks of that as long as we're still in the same NAPI loop as the first call. Or something to that effect. Not sure if this makes sense for this use case, but figured I'd throw the idea out there :)

-Toke

#6 lore raw

Hi Toke,

Quoted 6 lines · Toke Høiland-Jørgensen

So for instance, if there's a large fixed component of the overhead of get_random_u32(), we could have bpf_user_rnd_u32() populate a larger per-CPU buffer and then just emit u32 chunks of that as long as we're still in the same NAPI loop as the first call. Or something to that effect. Not sure if this makes sense for this use case, but figured I'd throw the idea out there :)

Actually, this already is how get_random_u32() works! It buffers a bunch of u32s in percpu batches, and doles them out as requested. However, this API currently works in all contexts, including in interrupts. So each call results in disabling irqs and reenabling them. If I bifurcated batches into irq batches and non-irq batches, so that we only needed to disable preemption for the non-irq batches, that'd probably improve things quite a bit, since then the overhead really would reduce to just a memcpy for the majority of calls. But I don't know if adding that duplication of all code paths is really worth the huge hassle.

Jason

#7 lore raw
Quoted 10 lines · Jason A. Donenfeld

Hi Toke,

Quoted 6 lines · Toke Høiland-Jørgensen

So for instance, if there's a large fixed component of the overhead of get_random_u32(), we could have bpf_user_rnd_u32() populate a larger per-CPU buffer and then just emit u32 chunks of that as long as we're still in the same NAPI loop as the first call. Or something to that effect. Not sure if this makes sense for this use case, but figured I'd throw the idea out there :)

Quoted 2 lines

Actually, this already is how get_random_u32() works! It buffers a bunch of u32s in percpu batches, and doles them out as requested.

Ah, right. Not terribly surprised you already did this!

Quoted 8 lines

However, this API currently works in all contexts, including in interrupts. So each call results in disabling irqs and reenabling them. If I bifurcated batches into irq batches and non-irq batches, so that we only needed to disable preemption for the non-irq batches, that'd probably improve things quite a bit, since then the overhead really would reduce to just a memcpy for the majority of calls. But I don't know if adding that duplication of all code paths is really worth the huge hassle.

Right, makes sense; happy to leave that decision entirely up to you :)

-Toke