rp23xx: add TRNG driver (hardware /dev/random and /dev/urandom)

[ricardgb]

Summary

The RP2350 (rp23xx) embeds an Arm CryptoCell-style true random number
generator — a sampled ring oscillator conditioned by von-Neumann, CRNGT and
autocorrelation health tests, presenting 192 bits in the six EHR_DATA words —
but there is no driver for it, so /dev/urandom on rp23xx is the software
xorshift PRNG.

This adds arch/arm/src/rp23xx/rp23xx_rng.c, a polling driver that serves the
conditioned hardware bits as /dev/random and /dev/urandom. Entropy is drawn
rarely (an mbedtls CTR_DRBG seeds once and reseeds occasionally), so a blocking
poll is used rather than interrupt plumbing; health-test failures discard the
block and re-arm the source, so only conditioned entropy is ever returned.

New CONFIG_RP23XX_TRNG selects ARCH_HAVE_RNG, which switches /dev/urandom
from the software PRNG to the hardware source (backing getrandom() and thus any
entropy consumer, e.g. mbedtls).

Impact

• rp23xx gains a real hardware entropy source; opt-in via CONFIG_RP23XX_TRNG.
• No change when the option is off (default n).

Testing

Tested on a Raspberry Pi Pico 2 W:

• /dev/random and /dev/urandom register from the TRNG;
• a 32-byte read returns at real ring-oscillator sampling latency (~10 ms);
• two independent 32-byte samples differ (live, non-stuck source);
• the CryptoCell CRNGT continuous test — which would stall a stuck source —
  passes, and mbedtls seeds its CTR_DRBG from /dev/urandom successfully.

---

Disclosure: this change was prepared by an AI agent (Claude Code) at the
direction of the author, who reviewed and tested it on hardware before
submission.

🤖 Generated with Claude Code

[github-actions]

MemBrowse Memory Report

No memory changes detected for:

• arduino-mega2560
• esp32-devkitc
• hifive1-revb
• mirtoo
• qemu-armv8a
• qemu-intel64
• rx65n-rsk2mb
• s698pm-dkit
• stm32-nucleo-f103rb

[ricardgb]

Thanks for the review @xiaoxiang781216 — all four points addressed in d2531b2:

• dropped the local MIN fallback and used the sys/param.h macro directly (added the include)
• n = MIN(len - got, RP23XX_TRNG_EHR_BYTES); and return got; casts removed as suggested
• sample[][] in the startup health check is no longer static (192 bytes on the stack)

Also fixed the check CI failures from the previous push (nxstyle long lines and codespell hits on "EHR"); tools/checkpatch.sh -c -u -m -g now passes locally on the commit.

(As before: this change was made with the help of an AI agent and human-reviewed before pushing.)

[linguini1]

Please include the logs from your testing.

Also, as per the contribution guide, please mark your commit with the Assisted-by field. Co-authored commits are not permitted from AI tools, as only a human can certify the commit.

[ricardgb]

@linguini1 commit trailer updated to Assisted-by: per the contribution guide (and rebased onto the latest master while rewriting). Test logs from the Pico 2 W validation runs will follow in a separate comment.

[ricardgb]

@linguini1 on-hardware test logs from a Raspberry Pi Pico 2 W (RP2350), NuttX 13.0.0-RC0, with CONFIG_RP23XX_TRNG=y (DEV_RANDOM=y, DEV_URANDOM_ARCH=y).

Both devices register:

nsh> ls /dev
/dev:
 console  null  random  telnet  ttyACM0  urandom  zero

/dev/urandom reads at the hardware ring-oscillator latency (~10 ms / block, not the software PRNG):

nsh> dd if=/dev/urandom of=/dev/null bs=64 count=1
64 bytes (1 blocks) copied, 10000 usec, 6 KB/s

Two independent /dev/urandom samples differ (high entropy, no repeats / constants / runs):

sample A: 62 5e 5f 88 c4 f3 e3 68 12 91 68 4c 03 03 38 c0  ed 18 71 fe a9 2d 67 be ...
sample B: de f9 c7 27 df 42 89 91 dc 94 c1 9e 69 fd 7b 65  c8 aa 49 6f d7 cb ad a3 ...

/dev/random (blocking hardware source) returns real entropy — slower, as it draws fresh conditioned bits through the VN/CRNGT/autocorrelation health tests:

nsh> dd if=/dev/random of=/dev/null bs=32 count=1
32 bytes (1 blocks) copied, 500000 usec, 0 KB/s
nsh> dd if=/dev/random of=/dev/null bs=64 count=1
64 bytes (1 blocks) copied, 370000 usec, 0 KB/s

The boot self-test gates the source (a failed source is marked unhealthy and every read returns -EIO); all reads above succeed, so the source is healthy.

Disclosure: these tests were run and captured with the help of an AI agent (Claude Code, Anthropic); results are from real hardware and human-reviewed.