Merge branch 'fix/sdmmc_wait_for_idle_busy_poll_starving_cpu_cores_v6.0' into 'release/v6.0'

fix(sdmmc): back off between CMD13 polls while waiting for card to be ready (v6.0)

See merge request espressif/esp-idf!52533

components/sdmmc/Kconfig

@@ -5,4 +5,39 @@ menu "SD Protocol Layer Configuration"
         help
             Enable SDIO support.
             Disabling this will skip SDIO-specific initialization steps
+
+    config SD_READY_POLL_PERIOD_START_US
+        int "Initial delay between card status polls (us)"
+        default 100
+        range 1 10000
+        help
+            While waiting for the card to leave its busy state (for example after a write),
+            the driver polls it with CMD13 (SEND_STATUS). This is the delay before the
+            first re-poll; the delay then doubles after every poll, up to one FreeRTOS
+            tick period.
+
+            Polling back-to-back without a delay turns every write into a storm of
+            hundreds of CMD13 commands, which occupies the host controller and slows down
+            unrelated work on both cores.
+
+            The exponential backoff means a card that becomes ready after 1.5 ms is
+            detected at about 1.5 ms rather than at the next FreeRTOS tick, using a
+            handful of commands instead of hundreds. Once the delay reaches one tick
+            period it stops growing, so a long busy period costs one poll per tick and
+            is detected within one tick of the card becoming ready.
+
+            Lower this value to detect very short busy periods sooner, at the cost of
+            more commands on the bus. Values larger than one FreeRTOS tick period are
+            clamped to it, so the largest useful setting depends on CONFIG_FREERTOS_HZ.
+
+            Delays shorter than one FreeRTOS tick period are spent busy-waiting, since
+            there is no shorter blocking sleep available; a delay of one tick blocks on
+            vTaskDelay() and lets other tasks run.
+
+            The busy-waited part of the ramp does not yield, so lower-priority tasks on
+            that core are held off for its duration. It lasts at most until the backoff
+            reaches one tick: with the default 100 us start, about 1.5 ms at a 1 kHz
+            tick (100+200+400+800 us) and about 12.7 ms at a 100 Hz tick. A card that
+            becomes ready sooner ends the wait sooner. No CMD13 is issued during the
+            delays, so the host controller and the bus stay idle throughout.
 endmenu

components/sdmmc/include/esp_private/sdmmc_common.h

@@ -17,6 +17,7 @@
 #pragma once
 
 #include <string.h>
+#include "sdkconfig.h"
 #include "esp_log.h"
 #include "esp_check.h"
 #include "esp_heap_caps.h"
@@ -38,6 +39,27 @@ extern "C" {
 #define SDMMC_INIT_WAIT_DATA_READY_TIMEOUT_US  (5000 * 1000)
 #define SDMMC_READY_FOR_DATA_TIMEOUT_US        (5000 * 1000)
 
+/* Delay between two consecutive card status (CMD13) polls while waiting for the
+ * card to become ready. Starts at SDMMC_READY_POLL_PERIOD_START_US and doubles
+ * after every poll, up to one FreeRTOS tick period.
+ *
+ * Polling without a delay turns every write into a storm of hundreds of CMD13
+ * commands, which occupies the host controller and slows down unrelated work.
+ *
+ * While the delay is below one tick period it is busy-waited and doubles, so a
+ * handful of commands covers a busy period of a few milliseconds and a card that
+ * becomes ready early is noticed without waiting for the next tick.
+ *
+ * Once the delay reaches one tick period it stops growing: from there on
+ * vTaskDelay() already yields and one command per tick is not a storm, while a
+ * longer delay would only delay noticing that the card is ready. Long busy
+ * periods therefore cost one poll per tick and are detected within one tick.
+ *
+ * One tick period is a hard cap: a configured start period larger than that is
+ * clamped to it as well.
+ */
+#define SDMMC_READY_POLL_PERIOD_START_US       CONFIG_SD_READY_POLL_PERIOD_START_US
+
 /* These delay values are mostly useful for cases when CD pin is not used, and
  * the card is removed. In this case, SDMMC peripheral may not always return
  * CMD_DONE / DATA_DONE interrupts after signaling the error. These timeouts work
@@ -184,6 +206,19 @@ static inline bool sdmmc_ready_for_data(uint32_t status)
     return (status & MMC_R1_READY_FOR_DATA) && (MMC_R1_CURRENT_STATE_STATUS(status) == MMC_R1_CURRENT_STATE_TRAN);
 }
 
+/**
+ * @brief Wait before the next card status poll, and back off for the poll after that
+ *
+ * Delays for *period_us, then doubles *period_us. Both the delay and the stored
+ * period are capped at one FreeRTOS tick period, so a *period_us larger than that
+ * on entry is clamped rather than used as-is. A delay shorter than one tick period
+ * is busy-waited, since no shorter blocking sleep is available; one tick period
+ * blocks on vTaskDelay() so that other tasks can run.
+ *
+ * @param[in,out] period_us  Delay to apply now; updated to the delay to apply next time
+ */
+void sdmmc_poll_delay_and_backoff(uint32_t* period_us);
+
 void sdmmc_flip_byte_order(uint32_t* response, size_t size);
 
 esp_err_t sdmmc_fix_host_flags(sdmmc_card_t* card);

components/sdmmc/sdmmc_common.c

@@ -18,6 +18,7 @@
 #include <inttypes.h>
 #include "esp_log.h"
 #include "freertos/FreeRTOS.h"
+#include "esp_rom_sys.h"
 #include "esp_timer.h"
 #include "esp_private/sdmmc_common.h"
 
@@ -441,12 +442,33 @@ uint32_t sdmmc_get_erase_timeout_ms(const sdmmc_card_t* card, int arg, size_t er
     }
 }
 
+void sdmmc_poll_delay_and_backoff(uint32_t* period_us)
+{
+    const uint32_t us_per_tick = portTICK_PERIOD_MS * 1000;
+    /* Clamp on entry as well: a configured start period longer than a tick would
+     * otherwise be used as-is and never brought back down to the cap. */
+    uint32_t delay_us = MIN(*period_us, us_per_tick);
+
+    if (delay_us < us_per_tick) {
+        /* No blocking sleep with sub-tick resolution is available, busy-wait instead.
+         * The point of the delay is to keep CMD13 off the bus, which this still does. */
+        esp_rom_delay_us(delay_us);
+    } else {
+        vTaskDelay(1);
+    }
+
+    /* Stop growing once the delay reaches one tick period. At that point vTaskDelay()
+     * already yields and one command per tick is not a storm, so a longer delay would
+     * only add overshoot to the time the card is detected as ready. */
+    *period_us = MIN(delay_us * 2, us_per_tick);
+}
+
 esp_err_t sdmmc_wait_for_idle(sdmmc_card_t* card, uint32_t status)
 {
     assert(!host_is_spi(card));
     esp_err_t err = ESP_OK;
     size_t count = 0;
-    int64_t yield_delay_us = 100 * 1000; // initially 100ms
+    uint32_t poll_period_us = SDMMC_READY_POLL_PERIOD_START_US;
     int64_t t0 = esp_timer_get_time();
     int64_t t1 = 0;
     /* SD mode: wait for the card to become idle based on R1 status */
@@ -455,10 +477,7 @@ esp_err_t sdmmc_wait_for_idle(sdmmc_card_t* card, uint32_t status)
         if (t1 - t0 > SDMMC_READY_FOR_DATA_TIMEOUT_US) {
             return ESP_ERR_TIMEOUT;
         }
-        if (t1 - t0 > yield_delay_us) {
+        sdmmc_poll_delay_and_backoff(&poll_period_us);
-            yield_delay_us *= 2;
-            vTaskDelay(1);
-        }
         err = sdmmc_send_cmd_send_status(card, &status);
         if (err != ESP_OK) {
             ESP_LOGE(TAG, "%s: sdmmc_send_cmd_send_status returned 0x%x", __func__, err);

components/sdmmc/sdmmc_sd.c

@@ -150,7 +150,7 @@ esp_err_t sdmmc_init_sd_wait_data_ready(sdmmc_card_t* card)
     /* Wait for the card to be ready for data transfers */
     uint32_t status = 0;
     uint32_t count = 0;
-    int64_t yield_delay_us = 100 * 1000; // initially 100ms
+    uint32_t poll_period_us = SDMMC_READY_POLL_PERIOD_START_US;
     int64_t t0 = esp_timer_get_time();
     int64_t t1 = 0;
     while (!host_is_spi(card) && !(status & MMC_R1_READY_FOR_DATA)) {
@@ -159,10 +159,7 @@ esp_err_t sdmmc_init_sd_wait_data_ready(sdmmc_card_t* card)
             ESP_LOGE(TAG, "init wait data ready - timeout");
             return ESP_ERR_TIMEOUT;
         }
-        if (t1 - t0 > yield_delay_us) {
+        sdmmc_poll_delay_and_backoff(&poll_period_us);
-            yield_delay_us *= 2;
-            vTaskDelay(1);
-        }
         esp_err_t err = sdmmc_send_cmd_send_status(card, &status);
         if (err != ESP_OK) {
             return err;
@@ -396,7 +393,7 @@ static esp_err_t read_tuning_block(sdmmc_card_t *card)
 
     uint32_t status = 0;
     size_t count = 0;
-    int64_t yield_delay_us = 100 * 1000; // initially 100ms
+    uint32_t poll_period_us = SDMMC_READY_POLL_PERIOD_START_US;
     int64_t t0 = esp_timer_get_time();
     int64_t t1 = 0;
     while (!host_is_spi(card) && !(status & MMC_R1_READY_FOR_DATA)) {
@@ -405,10 +402,7 @@ static esp_err_t read_tuning_block(sdmmc_card_t *card)
             ESP_LOGW(TAG, "read sectors dma - timeout");
             return ESP_ERR_TIMEOUT;
         }
-        if (t1 - t0 > yield_delay_us) {
+        sdmmc_poll_delay_and_backoff(&poll_period_us);
-            yield_delay_us *= 2;
-            vTaskDelay(1);
-        }
         ret = sdmmc_send_cmd_send_status(card, &status);
         if (ret != ESP_OK) {
             ESP_LOGW(TAG, "%s: sdmmc_send_cmd_send_status returned 0x%x", __func__, ret);