[PATCH 1/2] hwrng: histb: wait for a batch of words

[David Yang]

The old figures were per-word, taken from a streaming average: ~16ms at
depth 255 is one word out of a batch every ~64ms, and ~1ms at depth 1 is
the 1ms poll granularity of the read path. The device fills the FIFO
with 4 words in one go, every ~200us per unit of post-process depth, so
at the depth of 144 this driver programs a batch lands every ~30ms and
an empty queue waits that long for a single word. The 30ms per-word
timeout falls inside that, so wait 500ms instead, which also covers the
restart after rewriting the control word.

Fixes: a2216e187471 ("crypto: hisilicon/trng - add support for HiSTB TRNG")
Signed-off-by: David Yang <mmyangfl@xxxxxxxxx>
---
 drivers/char/hw_random/histb-rng.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/drivers/char/hw_random/histb-rng.c b/drivers/char/hw_random/histb-rng.c
index 445b80beed62..d435d4ab9aa7 100644
--- a/drivers/char/hw_random/histb-rng.c
+++ b/drivers/char/hw_random/histb-rng.c
@@ -26,16 +26,21 @@ struct histb_rng_priv {
 };
 
 /*
- * Observed:
- * depth = 1 -> ~1ms
- * depth = 255 -> ~16ms
+ * Observed: the FIFO fills with 4 words at a time, roughly every 200us per
+ * unit of post-process depth, i.e. one batch per:
+ * depth = 144 -> ~30ms
+ * depth = 255 -> ~52ms
+ * Rewriting the control word delays the first batch by up to 300ms.
  */
+#define BATCH_TIMEOUT_US	(500 * USEC_PER_MSEC)
+
 static int histb_rng_wait(void __iomem *base)
 {
 	u32 val;
 
 	return readl_relaxed_poll_timeout(base + RNG_STAT, val,
-					  val & DATA_COUNT, 1000, 30 * 1000);
+					  val & DATA_COUNT, 1000,
+					  BATCH_TIMEOUT_US);
 }
 
 static void histb_rng_init(void __iomem *base, unsigned int depth)
-- 
2.53.0