[PATCH] mmc: jz4740: Remove the host before resetting the controller

[Olivier Louvignes]

jz4740_mmc_remove() masks every interrupt and resets the controller
before mmc_remove_host(), which removes the card and lets the card's
driver send its last commands. Those commands then run on a controller
that was just reset, and can wait out the driver's 5 s request timeout.

With an SDIO card that cost 5 s on every unbind: on a HiBy R1 (Ingenic
X1600, a BCM43430A1 driven by brcmfmac), the CMD52 that
sdio_release_irq() sends from brcmfmac's remove waited 5.1 s. The
command had finished (END_CMD_RES set in MSC_STAT), but MSC_IMASK read
0xffffffff although the request had unmasked END_CMD_RES. MSC_STAT can
still show IS_RESETTING after jz4740_mmc_reset() returns, so the reset
most likely completed after that write and undid it.

Remove the host first, then quiesce the controller, as other host
drivers do. On the R1, unbinding the Wi-Fi's MSC now takes 0.65 s.

Fixes: 61bfbdb85687 ("MMC: Add support for the controller on JZ4740 SoCs.")
Signed-off-by: Olivier Louvignes <olivier@xxxxxxxxx>
---
 drivers/mmc/host/jz4740_mmc.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/mmc/host/jz4740_mmc.c b/drivers/mmc/host/jz4740_mmc.c
index 6a3c26b7c82d..923cf1450de5 100644
--- a/drivers/mmc/host/jz4740_mmc.c
+++ b/drivers/mmc/host/jz4740_mmc.c
@@ -1151,12 +1151,13 @@ static void jz4740_mmc_remove(struct platform_device *pdev)
 {
 	struct jz4740_mmc_host *host = platform_get_drvdata(pdev);
 
+	/* The card's driver may still send commands as it goes */
+	mmc_remove_host(host->mmc);
+
 	timer_delete_sync(&host->timeout_timer);
 	jz4740_mmc_set_irq_enabled(host, 0xff, false);
 	jz4740_mmc_reset(host);
 
-	mmc_remove_host(host->mmc);
-
 	free_irq(host->irq, host);
 
 	if (host->use_dma)

base-commit: fd9af27f6c2319713bb23832066585f8a59a1771
-- 
2.39.5