From: Thomas Otvald Jensen <thomas.otvald@gmail.com>
Subject: =?utf-8?b?bW1jX3Jlc2Nhbg==?= causes severe interrupt latency
Date: Thu, 10 Oct 2013 08:51:56 +0000 (UTC)
Message-ID: <loom.20131010T100432-754@post.gmane.org>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Sender: linux-mmc-owner@vger.kernel.org
To: linux-mmc@vger.kernel.org

Hi,

I am on an embedded system, and found that we had several ms interrupt delay
(causing UART overrun). The problem was identified as part of the mmc_rescan.

Our setup has enabled polling (MMC_CAP_NEEDS_POLL) to detect if an SD card
is present. This will cause a periodic rescheduling of mmc_rescan.

mmc_rescan -> mmc_rescan_try_freq -> mmc_power_up -> mmc_set_ios ->
host->ops->set_ios(host, ios);

We are using mmc/host/sdhci.c, so host->ops->set_ios(host, ios) -> sdhci_set_ios

And here is the problem:
The routine starts with 'spin_lock_irqsave'. This means I wont get any
interrupts until the lock is released. However this may take some time
(usually 3-5ms), but when looking at the code - in particular -
'sdhci_set_clock', then I see the comment: /* Wait max 20 ms */. We *still*
have interrupts disabled, so it could potentially add further to our delay.

(I know that e.g. sdhci_writew are platform specific, so FYI, I am on a
Freescale imx6 platform).

Q: Is this kind of interrupt delay expected (3-5ms)?
Q: Is it acceptable to have a loop with 20ms timeout with interrupts disabled?

Kind regards,
/Thomas