[PATCH] hw_random: bcm2835: use hwrng_msleep() instead of cpu_relax()

[Jason A. Donenfeld]

On Mon, Oct 10, 2022 at 09:06:07AM -0600, Jason A. Donenfeld wrote:
> Rather than busy looping, yield back to the scheduler and sleep for a
> bit in the event that there's no data. This should hopefully prevent the
> stalls that Mark reported:
> 
> <6>[    3.362859] Freeing initrd memory: 16196K
> <3>[   23.160131] rcu: INFO: rcu_sched self-detected stall on CPU
> <3>[   23.166057] rcu:  0-....: (2099 ticks this GP) idle=03b4/1/0x40000002 softirq=28/28 fqs=1050
> <4>[   23.174895]       (t=2101 jiffies g=-1147 q=2353 ncpus=4)
> <4>[   23.180203] CPU: 0 PID: 49 Comm: hwrng Not tainted 6.0.0 #1
> <4>[   23.186125] Hardware name: BCM2835
> <4>[   23.189837] PC is at bcm2835_rng_read+0x30/0x6c
> <4>[   23.194709] LR is at hwrng_fillfn+0x71/0xf4
> <4>[   23.199218] pc : [<c07ccdc8>]    lr : [<c07cb841>]    psr: 40000033
> <4>[   23.205840] sp : f093df70  ip : 00000000  fp : 00000000
> <4>[   23.211404] r10: c3c7e800  r9 : 00000000  r8 : c17e6b20
> <4>[   23.216968] r7 : c17e6b64  r6 : c18b0a74  r5 : c07ccd99  r4 : c3f171c0
> <4>[   23.223855] r3 : 000fffff  r2 : 00000040  r1 : c3c7e800  r0 : c3f171c0
> <4>[   23.230743] Flags: nZcv  IRQs on  FIQs on  Mode SVC_32  ISA Thumb  Segment none
> <4>[   23.238426] Control: 50c5387d  Table: 0020406a  DAC: 00000051
> <4>[   23.244519] CPU: 0 PID: 49 Comm: hwrng Not tainted 6.0.0 #1
> 
> Link: https://lore.kernel.org/all/Y0QJLauamRnCDUef@sirena.org.uk/
> Signed-off-by: Jason A. Donenfeld <Jason at zx2c4.com>
> ---
> I haven't tested this. Somebody with access to that kernel CI infra that
> triggered this will need to test.

So I succeeded at testing this, sort of. I was able to reproduce the
hang on a CONFIG_PREEMPT=n kernel with this diff:

diff --git a/drivers/net/wireguard/main.c b/drivers/net/wireguard/main.c
index ee4da9ab8013..19e1186f0db0 100644
--- a/drivers/net/wireguard/main.c
+++ b/drivers/net/wireguard/main.c
@@ -15,12 +15,29 @@
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/genetlink.h>
+#include <linux/hw_random.h>
 #include <net/rtnetlink.h>

+static int derp_rng_read(struct hwrng *rng, void *buf, size_t max, bool wait)
+{
+	if (wait) {
+		for (;;)
+			cpu_relax();
+	}
+	return 0;
+}
+
+static struct hwrng derp_ops = {
+	.name = "flurpderp",
+	.read = derp_rng_read,
+};
+
 static int __init wg_mod_init(void)
 {
 	int ret;

+	hwrng_register(&derp_ops);
+
 	ret = wg_allowedips_slab_init();
 	if (ret < 0)
 		goto err_allowedips;


Next, I changed the cpu_relax() into hwrng_msleep(), as this patch does:

diff --git a/drivers/net/wireguard/main.c b/drivers/net/wireguard/main.c
index ee4da9ab8013..19e1186f0db0 100644
--- a/drivers/net/wireguard/main.c
+++ b/drivers/net/wireguard/main.c
@@ -15,12 +15,29 @@
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/genetlink.h>
+#include <linux/hw_random.h>
 #include <net/rtnetlink.h>

+static int derp_rng_read(struct hwrng *rng, void *buf, size_t max, bool wait)
+{
+	if (wait) {
+		for (;;)
+			hwrng_msleep(rng, 1000);
+	}
+	return 0;
+}
+
+static struct hwrng derp_ops = {
+	.name = "flurpderp",
+	.read = derp_rng_read,
+};
+
 static int __init wg_mod_init(void)
 {
 	int ret;

+	hwrng_register(&derp_ops);
+
 	ret = wg_allowedips_slab_init();
 	if (ret < 0)
 		goto err_allowedips;

And then the problem went away.

So I think this patch is a good one.

Jason