Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751705AbdCSNRn (ORCPT ); Sun, 19 Mar 2017 09:17:43 -0400 Received: from mail-it0-f66.google.com ([209.85.214.66]:34296 "EHLO mail-it0-f66.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751262AbdCSNRk (ORCPT ); Sun, 19 Mar 2017 09:17:40 -0400 MIME-Version: 1.0 In-Reply-To: References: <20170312133250.GA7772@singhal-Inspiron-5558> <6e55c61d-7587-4191-1fc5-de43e26986d7@metafoo.de> From: Gargi Sharma Date: Sun, 19 Mar 2017 18:46:14 +0530 Message-ID: Subject: Re: [Outreachy kernel] Re: [PATCH] staging: iio: ade7753: replace mlock with driver private lock To: Jonathan Cameron Cc: Lars-Peter Clausen , simran singhal , Michael Hennerich , Hartmut Knaack , Pete Meerwald-Stadler , Greg Kroah-Hartman , linux-iio@vger.kernel.org, devel@driverdev.osuosl.org, linux-kernel@vger.kernel.org, outreachy-kernel@googlegroups.com Content-Type: text/plain; charset=UTF-8 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 7477 Lines: 226 On Sun, Mar 19, 2017 at 4:01 PM, Jonathan Cameron wrote: > On 17/03/17 09:32, Gargi Sharma wrote: >> On Mon, Mar 13, 2017 at 5:30 PM, Lars-Peter Clausen wrote: >>> >>> On 03/12/2017 02:32 PM, simran singhal wrote: >>>> The IIO subsystem is redefining iio_dev->mlock to be used by >>>> the IIO core only for protecting device operating mode changes. >>>> ie. Changes between INDIO_DIRECT_MODE, INDIO_BUFFER_* modes. >>>> >>>> In this driver, mlock was being used to protect hardware state >>>> changes. Replace it with a lock in the devices global data. >>>> >>>> Fix some coding style issues related to white space also. >>>> >>>> Signed-off-by: simran singhal >>>> --- >>>> drivers/staging/iio/meter/ade7753.c | 14 ++++++++------ >>>> 1 file changed, 8 insertions(+), 6 deletions(-) >>>> >>>> diff --git a/drivers/staging/iio/meter/ade7753.c b/drivers/staging/iio/meter/ade7753.c >>>> index dfd8b71..ca99d82 100644 >>>> --- a/drivers/staging/iio/meter/ade7753.c >>>> +++ b/drivers/staging/iio/meter/ade7753.c >>>> @@ -81,12 +81,14 @@ >>>> * @tx: transmit buffer >>>> * @rx: receive buffer >>>> * @buf_lock: mutex to protect tx and rx >>>> + * @lock: protect sensor state >>> >>> It might make sense to reuse the existing lock which currently protects the >>> read/write functions. You can do this by introducing a variant of >>> ade7753_spi_{read,write}_reg_16() that does not take a lock and use these to >>> implement the read-modify-write cycle in a protected section. >> >> There are other read/write functions for example, >> ade7753_spi_{read/write}_reg_8 that use the mutex as well. Should a >> variant of these functions be introduced as well? Also, how does one >> go about implementing RMW inside a protected section. > Hmm. Simran has also been progressing with patches for this. > I was trying to work through a patch for ade7754. So ran into the same problem :) > You raise a good question. There are other read/modify/write sequences in > the driver. They don't have the same issue with potentially deadlocking > against the buf lock as they are all using the spi subsystems provisions > for small write/read cycles where buffer protection is handled internally. > > So let us address the cases in turn: > > static int ade7753_reset(struct device *dev) > { > u16 val; > int ret; > > ret = ade7753_spi_read_reg_16(dev, ADE7753_MODE, &val); > if (ret) > return ret; > > val |= BIT(6); /* Software Chip Reset */ > > return ade7753_spi_write_reg_16(dev, ADE7753_MODE, val); > } > This is only called in the device initialization. At that point > we should be fine in assuming no parallel calls. Crucial point > is it is before the call to iio_device_register which exposes > the userspace interfaces. > > static int ade7753_set_irq(struct device *dev, bool enable) > { > int ret; > u8 irqen; > > ret = ade7753_spi_read_reg_8(dev, ADE7753_IRQEN, &irqen); > if (ret) > goto error_ret; > > if (enable) > irqen |= BIT(3); /* Enables an interrupt when a data is > * present in the waveform register > */ > else > irqen &= ~BIT(3); > > ret = ade7753_spi_write_reg_8(dev, ADE7753_IRQEN, irqen); > > error_ret: > return ret; > } > > This one is actually safe because it is the only function that > modifies that particular register. > > /* Power down the device */ > static int ade7753_stop_device(struct device *dev) > { > u16 val; > int ret; > > ret = ade7753_spi_read_reg_16(dev, ADE7753_MODE, &val); > if (ret) > return ret; > > val |= BIT(4); /* AD converters can be turned off */ > > return ade7753_spi_write_reg_16(dev, ADE7753_MODE, val); > } > > Only called in remove (after userspace interfaces have been > removed by the iio_device_unregister call so also should not > be running concurrently with much else. > The only nested lock here is ade7754_spi_write_reg_16, so as long as that is refactored, it'll be fine. > So I think all the other cases are safe. Perhaps it would have > been better to have had a lock around them, purely to make > the code more resilient against future changes though. > Probably a job to do as part of a larger scale pile of work > on that driver rather than as a one off patch. Another question that I have is why are we writing inside a read function(ade7754_spi_read_reg_24)? static int ade7754_spi_read_reg_24(struct device *dev, u8 reg_address, u32 *val) { struct iio_dev *indio_dev = dev_to_iio_dev(dev); struct ade7754_state *st = iio_priv(indio_dev); int ret; struct spi_transfer xfers[] = { { .tx_buf = st->tx, .rx_buf = st->rx, .bits_per_word = 8, .len = 4, }, }; mutex_lock(&st->buf_lock); st->tx[0] = ADE7754_READ_REG(reg_address); st->tx[1] = 0; st->tx[2] = 0; st->tx[3] = 0; ret = spi_sync_transfer(st->us, xfers, ARRAY_SIZE(xfers)); if (ret) { dev_err(&st->us->dev, "problem when reading 24 bit register 0x%02X", reg_address); goto error_ret; } *val = (st->rx[1] << 16) | (st->rx[2] << 8) | st->rx[3]; error_ret: mutex_unlock(&st->buf_lock); return ret; } Thanks! Gargi > > Jonathan > > > > > > > >> >> >>> >>> Looking through the driver there seem to be other places as well that do >>> read-modify-write that should be protected by a lock, but currently are not. >>> This might be a good task. >> >> Am I right in understanding that we want to introduce mutex lock for >> writes in other drivers as well? >> >> Thanks, >> Gargi >>> >>>> **/ >>>> struct ade7753_state { >>>> - struct spi_device *us; >>>> - struct mutex buf_lock; >>>> - u8 tx[ADE7753_MAX_TX] ____cacheline_aligned; >>>> - u8 rx[ADE7753_MAX_RX]; >>>> + struct spi_device *us; >>>> + struct mutex buf_lock; >>>> + struct mutex lock; /* protect sensor state */ >>>> + u8 tx[ADE7753_MAX_TX] ____cacheline_aligned; >>>> + u8 rx[ADE7753_MAX_RX]; >>>> }; >>>> >>>> static int ade7753_spi_write_reg_8(struct device *dev, >>>> @@ -484,7 +486,7 @@ static ssize_t ade7753_write_frequency(struct device *dev, >>>> if (!val) >>>> return -EINVAL; >>>> >>>> - mutex_lock(&indio_dev->mlock); >>>> + mutex_lock(&st->lock); >>>> >>>> t = 27900 / val; >>>> if (t > 0) >>>> @@ -505,7 +507,7 @@ static ssize_t ade7753_write_frequency(struct device *dev, >>>> ret = ade7753_spi_write_reg_16(dev, ADE7753_MODE, reg); >>>> >>>> out: >>>> - mutex_unlock(&indio_dev->mlock); >>>> + mutex_unlock(&st->lock); >>>> >>>> return ret ? ret : len; >>>> } >>>> >>> >>> -- >> -- >> To unsubscribe from this list: send the line "unsubscribe linux-iio" in >> the body of a message to majordomo@vger.kernel.org >> More majordomo info at http://vger.kernel.org/majordomo-info.html >> >