Date:   Thu, 12 Dec 2019 16:53:51 +0000
From:   cang@...eaurora.org
To:     Avri Altman <Avri.Altman@....com>
Cc:     Bjorn Andersson <bjorn.andersson@...aro.org>,
        asutoshd@...eaurora.org, nguyenb@...eaurora.org,
        rnayak@...eaurora.org, linux-scsi@...r.kernel.org,
        kernel-team@...roid.com, saravanak@...gle.com, salyzyn@...gle.com,
        Alim Akhtar <alim.akhtar@...sung.com>,
        Pedro Sousa <pedrom.sousa@...opsys.com>,
        "James E.J. Bottomley" <jejb@...ux.ibm.com>,
        "Martin K. Petersen" <martin.petersen@...cle.com>,
        Evan Green <evgreen@...omium.org>,
        Kishon Vijay Abraham I <kishon@...com>,
        Stephen Boyd <swboyd@...omium.org>,
        Stanley Chu <stanley.chu@...iatek.com>,
        Vignesh Raghavendra <vigneshr@...com>,
        Bean Huo <beanhuo@...ron.com>,
        Bart Van Assche <bvanassche@....org>,
        Venkat Gopalakrishnan <venkatg@...eaurora.org>,
        Tomas Winkler <tomas.winkler@...el.com>,
        Arnd Bergmann <arnd@...db.de>,
        open list <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v2 2/3] scsi: ufs: Modulize ufs-bsg

On 2019-12-12 15:00, Avri Altman wrote:
>> 
>> 
>> On Wed 11 Dec 22:01 PST 2019, cang@...eaurora.org wrote:
>> 
>> > On 2019-12-12 12:53, Bjorn Andersson wrote:
>> > > On Wed 11 Dec 00:49 PST 2019, Can Guo wrote:
>> > >
>> > > > In order to improve the flexibility of ufs-bsg, modulizing it is a
>> > > > good choice. This change introduces tristate to ufs-bsg to allow
>> > > > users compile it as an external module.
>> > >
>> > > Can you please elaborate on what this "flexibility" is and why it's
>> > > a good thing?
>> > >
>> >
>> > ufs-bsg is a helpful gadget for debug/test purpose. But neither
>> > disabling it nor enabling it is the best way on a commercialized
>> > device. Disabling it means we cannot use it, while enabling it by
>> > default will expose all the DEVM/UIC/TM interfaces to user space,
>> > which is not "safe" on a commercialized device to let users play with it.
>> > Making it a module can resolve this, because only vendors can install
>> > it as they have the root permissions.
> Agree.
> We see that the public ufs-utils
> (https://github.com/westerndigitalcorporation/ufs-utils) that uses
> this infrastructure,
> is gaining momentum, and currently being used not only by chipset and
> flash vendors,
> but by end customers as well.
> This change will e.g. enable, field application engineers to debug
> issues in a safer mode.
> 

True, thank you for the comments.

>> >
>> > > >
>> > > > Signed-off-by: Can Guo <cang@...eaurora.org>
>> > > > ---
>> > > >  drivers/scsi/ufs/Kconfig   |  3 ++-
>> > > >  drivers/scsi/ufs/Makefile  |  2 +-  drivers/scsi/ufs/ufs_bsg.c |
>> > > > 49
>> > > > +++++++++++++++++++++++++++++++++++++++++++---
>> > > >  drivers/scsi/ufs/ufs_bsg.h |  8 --------
>> > > > drivers/scsi/ufs/ufshcd.c  | 36 ++++++++++++++++++++++++++++++----
>> > > >  drivers/scsi/ufs/ufshcd.h  |  7 ++++++-
>> > > >  6 files changed, 87 insertions(+), 18 deletions(-)
>> > > >
>> > > > diff --git a/drivers/scsi/ufs/Kconfig b/drivers/scsi/ufs/Kconfig
>> > > > index d14c224..72620ce 100644
>> > > > --- a/drivers/scsi/ufs/Kconfig
>> > > > +++ b/drivers/scsi/ufs/Kconfig
>> > > > @@ -38,6 +38,7 @@ config SCSI_UFSHCD
>> > > >   select PM_DEVFREQ
>> > > >   select DEVFREQ_GOV_SIMPLE_ONDEMAND
>> > > >   select NLS
>> > > > + select BLK_DEV_BSGLIB
>> > >
>> > > Why is this needed?
>> > >
>> >
>> > Because ufshcd.c needs to call some funcs defined in bsg lib.
>> >
>> > > >   ---help---
>> > > >   This selects the support for UFS devices in Linux, say Y and make
>> > > >     sure that you know the name of your UFS host adapter (the card
>> > > > @@ -143,7 +144,7 @@ config SCSI_UFS_TI_J721E
>> > > >     If unsure, say N.
>> > > >
>> > > >  config SCSI_UFS_BSG
>> > > > - bool "Universal Flash Storage BSG device node"
>> > > > + tristate "Universal Flash Storage BSG device node"
>> > > >   depends on SCSI_UFSHCD
>> > > >   select BLK_DEV_BSGLIB
>> > > >   help
>> > > > diff --git a/drivers/scsi/ufs/Makefile b/drivers/scsi/ufs/Makefile
>> > > > index 94c6c5d..904eff1 100644
>> > > > --- a/drivers/scsi/ufs/Makefile
>> > > > +++ b/drivers/scsi/ufs/Makefile
>> > > > @@ -6,7 +6,7 @@ obj-$(CONFIG_SCSI_UFS_CDNS_PLATFORM) +=
>> > > > cdns-pltfrm.o
>> > > >  obj-$(CONFIG_SCSI_UFS_QCOM) += ufs-qcom.o
>> > > >  obj-$(CONFIG_SCSI_UFSHCD) += ufshcd-core.o
>> > > >  ufshcd-core-y                            += ufshcd.o ufs-sysfs.o
>> > > > -ufshcd-core-$(CONFIG_SCSI_UFS_BSG)       += ufs_bsg.o
>> > > > +obj-$(CONFIG_SCSI_UFS_BSG)       += ufs_bsg.o
>> > > >  obj-$(CONFIG_SCSI_UFSHCD_PCI) += ufshcd-pci.o
>> > > >  obj-$(CONFIG_SCSI_UFSHCD_PLATFORM) += ufshcd-pltfrm.o
>> > > >  obj-$(CONFIG_SCSI_UFS_HISI) += ufs-hisi.o diff --git
>> > > > a/drivers/scsi/ufs/ufs_bsg.c b/drivers/scsi/ufs/ufs_bsg.c index
>> > > > 3a2e68f..302222f 100644
>> > > > --- a/drivers/scsi/ufs/ufs_bsg.c
>> > > > +++ b/drivers/scsi/ufs/ufs_bsg.c
>> > > > @@ -164,13 +164,15 @@ static int ufs_bsg_request(struct bsg_job *job)
>> > > >   */
>> > > >  void ufs_bsg_remove(struct ufs_hba *hba)  {
>> > > > - struct device *bsg_dev = &hba->bsg_dev;
>> > > > + struct device *bsg_dev = hba->bsg_dev;
>> > > >
>> > > >   if (!hba->bsg_queue)
>> > > >           return;
>> > > >
>> > > >   bsg_remove_queue(hba->bsg_queue);
>> > > >
>> > > > + hba->bsg_dev = NULL;
>> > > > + hba->bsg_queue = NULL;
>> > > >   device_del(bsg_dev);
>> > > >   put_device(bsg_dev);
>> > > >  }
>> > > > @@ -178,6 +180,7 @@ void ufs_bsg_remove(struct ufs_hba *hba)
>> > > >  static inline void ufs_bsg_node_release(struct device *dev)
>> > > >  {
>> > > >   put_device(dev->parent);
>> > > > + kfree(dev);
>> > > >  }
>> > > >
>> > > >  /**
>> > > > @@ -186,14 +189,19 @@ static inline void ufs_bsg_node_release(struct
>> > > > device *dev)
>> > > >   *
>> > > >   * Called during initial loading of the driver, and before
>> > > > scsi_scan_host.
>> > > >   */
>> > > > -int ufs_bsg_probe(struct ufs_hba *hba)
>> > > > +static int ufs_bsg_probe(struct ufs_hba *hba)
>> > > >  {
>> > > > - struct device *bsg_dev = &hba->bsg_dev;
>> > > > + struct device *bsg_dev;
>> > > >   struct Scsi_Host *shost = hba->host;
>> > > >   struct device *parent = &shost->shost_gendev;
>> > > >   struct request_queue *q;
>> > > >   int ret;
>> > > >
>> > > > + bsg_dev = kzalloc(sizeof(*bsg_dev), GFP_KERNEL);
>> > > > + if (!bsg_dev)
>> > > > +         return -ENOMEM;
>> > > > +
>> > > > + hba->bsg_dev = bsg_dev;
>> > > >   device_initialize(bsg_dev);
>> > > >
>> > > >   bsg_dev->parent = get_device(parent);
>> > > > @@ -217,6 +225,41 @@ int ufs_bsg_probe(struct ufs_hba *hba)
>> > > >
>> > > >  out:
>> > > >   dev_err(bsg_dev, "fail to initialize a bsg dev %d\n",
>> > > > shost->host_no);
>> > > > + hba->bsg_dev = NULL;
>> > > >   put_device(bsg_dev);
>> > > >   return ret;
>> > > >  }
>> > > > +
>> > > > +static int __init ufs_bsg_init(void)
>> > > > +{
>> > > > + struct list_head *hba_list = NULL;
>> > > > + struct ufs_hba *hba;
>> > > > + int ret = 0;
>> > > > +
>> > > > + ufshcd_get_hba_list_lock(&hba_list);
>> > > > + list_for_each_entry(hba, hba_list, list) {
>> > > > +         ret = ufs_bsg_probe(hba);
>> > > > +         if (ret)
>> > > > +                 break;
>> > > > + }
>> > >
>> > > So what happens if I go CONFIG_SCSI_UFS_BSG=y and
>> > > CONFIG_SCSI_UFS_QCOM=y?
>> > >
>> > > Wouldn't that mean that ufs_bsg_init() is called before ufshcd_init()
>> > > has added the controller to the list? And even in the even that they are
>> > > both =m, what happens if they are invoked in the "wrong" order?
>> > >
>> >
>> > In the case that CONFIG_SCSI_UFS_BSG=y and CONFIG_SCSI_UFS_QCOM=y,
>> > I give late_initcall_sync(ufs_bsg_init) to make sure ufs_bsg_init
>> > is invoked only after platform driver is probed. I tested this combination.
>> >
>> > In the case that both of them are "m", installing ufs-bsg before ufs-qcom
>> > is installed would have no effect as ufs_hba_list is empty, which is
>> > expected.
>> 
>> Why is it the expected behavior that bsg may or may not probe 
>> depending
>> on the driver load order and potentially timing of the initialization.
>> 
>> > And in real cases, as the UFS is the boot device, UFS driver will always
>> > be probed during bootup.
>> >
>> 
>> The UFS driver will load and probe because it's mentioned in the
>> devicetree, but if either the ufs drivers or any of its dependencies
>> (phy, resets, clocks, etc) are built as modules it might very well
>> finish probing after lateinitcall.
>> 
>> So in the even that the bsg is =y and any of these drivers are =m, or 
>> if
>> you're having bad luck with your timing, the list will be empty.
>> 
>> As described below, if bsg=m, then there's nothing that will load the
>> module and the bsg will not probe...
> Right.
> bsg=y and ufshcd=m is a bad idea, and should be avoided.
> 

Yeah, I will get it addressed in the next patchset.

Thanks,
Can Guo.

>> 
>> [..]
>> > > > diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
>> [..]
>> > > >  void ufshcd_remove(struct ufs_hba *hba)
>> > > >  {
>> > > > - ufs_bsg_remove(hba);
>> > > > + struct device *bsg_dev = hba->bsg_dev;
>> > > > +
>> > > > + mutex_lock(&ufs_hba_list_lock);
>> > > > + list_del(&hba->list);
>> > > > + if (hba->bsg_queue) {
>> > > > +         bsg_remove_queue(hba->bsg_queue);
>> > > > +         device_del(bsg_dev);
>> > >
>> > > Am I reading this correct in that you probe the bsg_dev form initcall
>> > > and you delete it as the ufshcd instance is removed? That's not okay.
>> > >
>> > > Regards,
>> > > Bjorn
>> > >
>> >
>> > If ufshcd is removed, its ufs-bsg, if exists, should also be removed.
>> > Could you please enlighten me a better way to do this? Thanks.
>> >
>> 
>> It's the asymmetry that I don't like.
>> 
>> Perhaps if you instead make ufshcd platform_device_register_data() the
>> bsg device you would solve the probe ordering, the remove will be
>> symmetric and module autoloading will work as well (although then you
>> need a MODULE_ALIAS of platform:device-name).
>> 
>> Regards,
>> Bjorn

Powered by blists - more mailing lists