diff mbox series

[9/9] accel/habanalabs: gradual sleep in polling memory macro

Message ID 20240619063425.1377327-9-obitton@habana.ai (mailing list archive)
State New, archived
Headers show
Series [1/9] accel/habanalbs/gaudi2: reduce interrupt count to 128 | expand

Commit Message

Ofir Bitton June 19, 2024, 6:34 a.m. UTC
From: Didi Freiman <dfreiman@habana.ai>

It’s better to avoid long sleeps right from the beginning of the polling
since the data may be available much sooner than the sleep period.
Because polling host memory is inexpensive, this change gradually
increases the sleep time up to the user-requested period.

Signed-off-by: Didi Freiman <dfreiman@habana.ai>
Reviewed-by: Ofir Bitton <obitton@habana.ai>
---
 drivers/accel/habanalabs/common/habanalabs.h | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
diff mbox series

Patch

diff --git a/drivers/accel/habanalabs/common/habanalabs.h b/drivers/accel/habanalabs/common/habanalabs.h
index a06e5a966f45..6f27ce4fa01b 100644
--- a/drivers/accel/habanalabs/common/habanalabs.h
+++ b/drivers/accel/habanalabs/common/habanalabs.h
@@ -2729,11 +2729,16 @@  void hl_wreg(struct hl_device *hdev, u32 reg, u32 val);
  * updated directly by the device. If false, the host memory being polled will
  * be updated by host CPU. Required so host knows whether or not the memory
  * might need to be byte-swapped before returning value to caller.
+ *
+ * On the first 4 polling iterations the macro goes to sleep for short period of
+ * time that gradually increases and reaches sleep_us on the fifth iteration.
  */
 #define hl_poll_timeout_memory(hdev, addr, val, cond, sleep_us, timeout_us, \
 				mem_written_by_device) \
 ({ \
+	u64 __sleep_step_us; \
 	ktime_t __timeout; \
+	u8 __step = 8; \
 	\
 	__timeout = ktime_add_us(ktime_get(), timeout_us); \
 	might_sleep_if(sleep_us); \
@@ -2751,8 +2756,10 @@  void hl_wreg(struct hl_device *hdev, u32 reg, u32 val);
 				(val) = le32_to_cpu(*(__le32 *) &(val)); \
 			break; \
 		} \
-		if (sleep_us) \
-			usleep_range((sleep_us >> 2) + 1, sleep_us); \
+		__sleep_step_us = sleep_us >> __step; \
+		if (__sleep_step_us) \
+			usleep_range((__sleep_step_us >> 2) + 1, __sleep_step_us); \
+		__step >>= 1; \
 	} \
 	(cond) ? 0 : -ETIMEDOUT; \
 })