lkml.org 
[lkml]   [2019]   [Dec]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v2 2/3] scsi: ufs: Modulize ufs-bsg
On 2019-12-12 14:37, Bjorn Andersson wrote:
> On Wed 11 Dec 22:01 PST 2019, cang@codeaurora.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.
>>
>> > >
>> > > Signed-off-by: Can Guo <cang@codeaurora.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...
>
> [..]
>> > > 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

Thanks for the suggestion! I didn't even think about this before. I
will go with the platform_device_register_data() way, it will be much
easier. After I get my new patchset tested I will upload it for review.

Best Regards,
Can Guo.

\
 
 \ /
  Last update: 2019-12-12 17:46    [W:0.204 / U:0.508 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site