On Sun, Nov 18, 2018 at 4:15 AM Michael Niewöhner [off-list ref] wrote:

Hi Louis,

On Wed, 2018-09-26 at 11:24 +0800, Louis Collard wrote:

quoted

Some systems, for example embedded systems, do not generate
enough entropy on boot through interrupts, and boot may be blocked for
several minutes waiting for a call to getrandom to complete.

Currently, random data is read from a hwrng when it is registered,
and is loaded into primary_crng. This data is treated in the same
way as data that is device-specific but otherwise unchanging, and
so primary_crng cannot become initialized with the data from the
hwrng.

This change causes the data initially read from the hwrng to be
treated the same as subsequent data that is read from the hwrng if
it's quality score is non-zero.

The implications of this are:

The data read from hwrng can cause primary_crng to become
initialized, therefore avoiding problems of getrandom blocking
on boot.

Calls to getrandom (with GRND_RANDOM) may be using entropy
exclusively (or in practise, almost exclusively) from the hwrng.

Regarding the latter point; this behavior is the same as if a
user specified a quality score of 1 (bit of entropy per 1024 bits)
so hopefully this is not too scary a change to make.

This change is the result of the discussion here:
https://patchwork.kernel.org/patch/10453893/

Signed-off-by: Louis Collard <redacted>
Acked-by: Jarkko Sakkinen <redacted>
---
 drivers/char/hw_random/core.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/drivers/char/hw_random/core.c b/drivers/char/hw_random/core.c
index aaf9e5afaad4..47f358aa0c3d 100644
--- a/drivers/char/hw_random/core.c
+++ b/drivers/char/hw_random/core.c

@@ -24,6 +24,7 @@
 #include <linux/sched.h>
 #include <linux/slab.h>
 #include <linux/uaccess.h>
+#include <crypto/chacha20.h>

 #define RNG_MODULE_NAME              "hw_random"

@@ -64,13 +65,17 @@ static size_t rng_buffer_size(void)
 static void add_early_randomness(struct hwrng *rng)
 {
      int bytes_read;
-     size_t size = min_t(size_t, 16, rng_buffer_size());
+     /* Read enough to initialize crng. */
+     size_t size = 2*CHACHA20_KEY_SIZE;

      mutex_lock(&reading_mutex);
      bytes_read = rng_get_data(rng, rng_buffer, size, 1);
      mutex_unlock(&reading_mutex);
      if (bytes_read > 0)
-             add_device_randomness(rng_buffer, bytes_read);
+             /* Allow crng to become initialized, but do not add
+              * entropy to the pool.
+              */
+             add_hwgenerator_randomness(rng_buffer, bytes_read, 0);
 }

 static inline void cleanup_rng(struct kref *kref)

I found your patch by chance, searching for a solution for crng init delay on my
headless machine. Unfortunately it hardly makes any difference for me. With the
patch the system hangs for about 80s instead of 120s until the "crng init done"
message.In contrast, doing a `cat /dev/hwrng >/dev/random` or running rngd
initializes the crng instantly.

Isn't that delay the problem this patch tries to fix? Any idea what is wrong
here?

Thanks!

Best regards
Michael

Yes that is the problem this is trying to address. My guess would be
rng_get_data() is not returning as much data as requested, so the
delay is reduced but not eliminated. Looking at implementation of
rng_get_data() it appears this could be caused by device support for
read() vs data_read(). I don't have a good feel for whether looping to
retrieve more data here would be acceptable, it is certainly a bigger
change than currently proposed.

Thanks,
Louis