diff mbox series

[1/2] scsi: ufs: Put SCSI host after remove it

Message ID 1576328616-30404-2-git-send-email-cang@codeaurora.org (mailing list archive)
State Changes Requested
Headers show
Series Modularize ufs-bsg | expand

Commit Message

Can Guo Dec. 14, 2019, 1:03 p.m. UTC
In ufshcd_remove(), after SCSI host is removed, put it once so that its
resources can be released.

Signed-off-by: Can Guo <cang@codeaurora.org>

Comments

Bart Van Assche Dec. 14, 2019, 6:32 p.m. UTC | #1
On 12/14/19 8:03 AM, Can Guo wrote:
> In ufshcd_remove(), after SCSI host is removed, put it once so that its
> resources can be released.
> 
> Signed-off-by: Can Guo <cang@codeaurora.org>
> 
> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> index b5966fa..a86b0fd 100644
> --- a/drivers/scsi/ufs/ufshcd.c
> +++ b/drivers/scsi/ufs/ufshcd.c
> @@ -8251,6 +8251,7 @@ void ufshcd_remove(struct ufs_hba *hba)
>   	ufs_bsg_remove(hba);
>   	ufs_sysfs_remove_nodes(hba->dev);
>   	scsi_remove_host(hba->host);
> +	scsi_host_put(hba->host);
>   	/* disable interrupts */
>   	ufshcd_disable_intr(hba, hba->intr_mask);
>   	ufshcd_hba_stop(hba, true);

Hi Can,

The UFS driver may queue work asynchronously and that asynchronous work 
may refer to the SCSI host, e.g. ufshcd_err_handler(). Is it guaranteed 
that all that asynchronous work has finished before scsi_host_put() is 
called?

Thanks,

Bart.
Can Guo Dec. 14, 2019, 10:24 p.m. UTC | #2
On 2019-12-15 02:32, Bart Van Assche wrote:
> On 12/14/19 8:03 AM, Can Guo wrote:
>> In ufshcd_remove(), after SCSI host is removed, put it once so that 
>> its
>> resources can be released.
>> 
>> Signed-off-by: Can Guo <cang@codeaurora.org>
>> 
>> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
>> index b5966fa..a86b0fd 100644
>> --- a/drivers/scsi/ufs/ufshcd.c
>> +++ b/drivers/scsi/ufs/ufshcd.c
>> @@ -8251,6 +8251,7 @@ void ufshcd_remove(struct ufs_hba *hba)
>>   	ufs_bsg_remove(hba);
>>   	ufs_sysfs_remove_nodes(hba->dev);
>>   	scsi_remove_host(hba->host);
>> +	scsi_host_put(hba->host);
>>   	/* disable interrupts */
>>   	ufshcd_disable_intr(hba, hba->intr_mask);
>>   	ufshcd_hba_stop(hba, true);
> 
> Hi Can,
> 
> The UFS driver may queue work asynchronously and that asynchronous
> work may refer to the SCSI host, e.g. ufshcd_err_handler(). Is it
> guaranteed that all that asynchronous work has finished before
> scsi_host_put() is called?
> 
> Thanks,
> 
> Bart.

Hi Bart,

Thanks for pointing it out. I noticed that you are changing this
path too in below 2 changes.
https://marc.info/?l=linux-scsi&m=157591520015924&w=2
https://marc.info/?l=linux-scsi&m=157591519915923&w=2

Actually the async works you pointed may also affect your change,
because you may tear down the hba->cmd_queue too early, as there can be
devm commands sent by clock gating, eeh_work and eh_work after that 
point,
meaning when blk_get_request is called in exec_dev_cmd(), hba->cmd_queue
may have been released already.

@@ -8263,6 +8232,7 @@ void ufshcd_remove(struct ufs_hba *hba)
  {
      ufs_bsg_remove(hba);
      ufs_sysfs_remove_nodes(hba->dev);
+    blk_cleanup_queue(hba->cmd_queue);
      scsi_remove_host(hba->host);
      /* disable interrupts */
      ufshcd_disable_intr(hba, hba->intr_mask);

How do you think if I replace my patch with below one?
In this way, you can also move blk_cleanup_queue() behind
cancel_work_sync(eh_work).

diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
index b5966fa..bd4ae75 100644
--- a/drivers/scsi/ufs/ufshcd.c
+++ b/drivers/scsi/ufs/ufshcd.c
@@ -8251,15 +8251,17 @@ void ufshcd_remove(struct ufs_hba *hba)
         ufs_bsg_remove(hba);
         ufs_sysfs_remove_nodes(hba->dev);
         scsi_remove_host(hba->host);
-       /* disable interrupts */
-       ufshcd_disable_intr(hba, hba->intr_mask);
-       ufshcd_hba_stop(hba, true);
-
         ufshcd_exit_clk_scaling(hba);
         ufshcd_exit_clk_gating(hba);
         if (ufshcd_is_clkscaling_supported(hba))
                 device_remove_file(hba->dev, 
&hba->clk_scaling.enable_attr);
+       cancel_work_sync(&hba->eeh_work);
+       cancel_work_sync(&hba->eh_work);
+       /* disable interrupts */
+       ufshcd_disable_intr(hba, hba->intr_mask);
+       ufshcd_hba_stop(hba, true);
         ufshcd_hba_exit(hba);
+       ufshcd_dealloc_host(hba);
  }
  EXPORT_SYMBOL_GPL(ufshcd_remove);
Bart Van Assche Dec. 15, 2019, 9:55 p.m. UTC | #3
On 2019-12-14 14:24, cang@codeaurora.org wrote:
> How do you think if I replace my patch with below one?
> In this way, you can also move blk_cleanup_queue() behind
> cancel_work_sync(eh_work).
> 
> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> index b5966fa..bd4ae75 100644
> --- a/drivers/scsi/ufs/ufshcd.c
> +++ b/drivers/scsi/ufs/ufshcd.c
> @@ -8251,15 +8251,17 @@ void ufshcd_remove(struct ufs_hba *hba)
>         ufs_bsg_remove(hba);
>         ufs_sysfs_remove_nodes(hba->dev);
>         scsi_remove_host(hba->host);
> -       /* disable interrupts */
> -       ufshcd_disable_intr(hba, hba->intr_mask);
> -       ufshcd_hba_stop(hba, true);
> -
>         ufshcd_exit_clk_scaling(hba);
>         ufshcd_exit_clk_gating(hba);
>         if (ufshcd_is_clkscaling_supported(hba))
>                 device_remove_file(hba->dev,
> &hba->clk_scaling.enable_attr);
> +       cancel_work_sync(&hba->eeh_work);
> +       cancel_work_sync(&hba->eh_work);
> +       /* disable interrupts */
> +       ufshcd_disable_intr(hba, hba->intr_mask);
> +       ufshcd_hba_stop(hba, true);
>         ufshcd_hba_exit(hba);
> +       ufshcd_dealloc_host(hba);
>  }
>  EXPORT_SYMBOL_GPL(ufshcd_remove);

Hi Can,

To which kernel tree does the above patch apply? I'm asking this because
I don't see the recently added blk_cleanup_queue() calls in the above
patch. Please start from Martin's latest scsi-queue branch when
preparing SCSI patches.

Additionally, is it on purpose that there is no scsi_host_put() call in
the above code? I'd like to keep that call because without that call a
memory leak will occur when unloading the ufshcd-core kernel driver.

Thanks,

Bart.
Can Guo Dec. 16, 2019, 1:34 a.m. UTC | #4
On 2019-12-16 05:55, Bart Van Assche wrote:
> On 2019-12-14 14:24, cang@codeaurora.org wrote:
>> How do you think if I replace my patch with below one?
>> In this way, you can also move blk_cleanup_queue() behind
>> cancel_work_sync(eh_work).
>> 
>> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
>> index b5966fa..bd4ae75 100644
>> --- a/drivers/scsi/ufs/ufshcd.c
>> +++ b/drivers/scsi/ufs/ufshcd.c
>> @@ -8251,15 +8251,17 @@ void ufshcd_remove(struct ufs_hba *hba)
>>         ufs_bsg_remove(hba);
>>         ufs_sysfs_remove_nodes(hba->dev);
>>         scsi_remove_host(hba->host);
>> -       /* disable interrupts */
>> -       ufshcd_disable_intr(hba, hba->intr_mask);
>> -       ufshcd_hba_stop(hba, true);
>> -
>>         ufshcd_exit_clk_scaling(hba);
>>         ufshcd_exit_clk_gating(hba);
>>         if (ufshcd_is_clkscaling_supported(hba))
>>                 device_remove_file(hba->dev,
>> &hba->clk_scaling.enable_attr);
>> +       cancel_work_sync(&hba->eeh_work);
>> +       cancel_work_sync(&hba->eh_work);
>> +       /* disable interrupts */
>> +       ufshcd_disable_intr(hba, hba->intr_mask);
>> +       ufshcd_hba_stop(hba, true);
>>         ufshcd_hba_exit(hba);
>> +       ufshcd_dealloc_host(hba);
>>  }
>>  EXPORT_SYMBOL_GPL(ufshcd_remove);
> 
> Hi Can,
> 
> To which kernel tree does the above patch apply? I'm asking this 
> because
> I don't see the recently added blk_cleanup_queue() calls in the above
> patch. Please start from Martin's latest scsi-queue branch when
> preparing SCSI patches.
> 
> Additionally, is it on purpose that there is no scsi_host_put() call in
> the above code? I'd like to keep that call because without that call a
> memory leak will occur when unloading the ufshcd-core kernel driver.
> 
> Thanks,
> 
> Bart.


Hi Bart,

This is applied to 5.5/scsi-queue. The two changes I patsed from you are
not merged yet, I am still doing code review to them, so there is no
blk_cleanup_queue() calls in my code base. I am just saying you may move
your blk_cleanup_queue() calls below cancel_work_sync(&hba->eh_work) if
my change applies. How do you think?

scsi_host_put() was there before but explicitly removed by
afa3dfd42d205b106787476647735aa1de1a5d02. I agree with you, without this
change, there is memory leak.

Thanks,

Can Guo.
Bart Van Assche Dec. 16, 2019, 2:39 a.m. UTC | #5
On 2019-12-15 17:34, cang@codeaurora.org wrote:
> This is applied to 5.5/scsi-queue. The two changes I patsed from you are
> not merged yet, I am still doing code review to them, so there is no
> blk_cleanup_queue() calls in my code base. I am just saying you may move
> your blk_cleanup_queue() calls below cancel_work_sync(&hba->eh_work) if
> my change applies. How do you think?
> 
> scsi_host_put() was there before but explicitly removed by
> afa3dfd42d205b106787476647735aa1de1a5d02. I agree with you, without this
> change, there is memory leak.

Hi Can,

Since your patch restores a call that was removed earlier, please
consider adding a Fixes: tag to your patch.

Please also have a look at
https://git.kernel.org/pub/scm/linux/kernel/git/mkp/scsi.git/log/?h=5.6/scsi-queue.
As one can see my patches that introduce blk_cleanup_queue() and
blk_mq_free_tag_set() calls have already been queued on Martin's
5.6/scsi-queue branch.

Bart.
Can Guo Dec. 16, 2019, 3:12 a.m. UTC | #6
On 2019-12-16 10:39, Bart Van Assche wrote:
> On 2019-12-15 17:34, cang@codeaurora.org wrote:
>> This is applied to 5.5/scsi-queue. The two changes I patsed from you 
>> are
>> not merged yet, I am still doing code review to them, so there is no
>> blk_cleanup_queue() calls in my code base. I am just saying you may 
>> move
>> your blk_cleanup_queue() calls below cancel_work_sync(&hba->eh_work) 
>> if
>> my change applies. How do you think?
>> 
>> scsi_host_put() was there before but explicitly removed by
>> afa3dfd42d205b106787476647735aa1de1a5d02. I agree with you, without 
>> this
>> change, there is memory leak.
> 
> Hi Can,
> 
> Since your patch restores a call that was removed earlier, please
> consider adding a Fixes: tag to your patch.
> 
> Please also have a look at
> https://git.kernel.org/pub/scm/linux/kernel/git/mkp/scsi.git/log/?h=5.6/scsi-queue.
> As one can see my patches that introduce blk_cleanup_queue() and
> blk_mq_free_tag_set() calls have already been queued on Martin's
> 5.6/scsi-queue branch.
> 
> Bart.

Hi Bart,

Sure, I will add the Fixes tag and rebase my changes. How about the 
logic
part of this change? Does it look good to you?

Sorry I was not aware of that your changes have been applied to 
5.6/scsi-queue.
I am still trying to get it tested on my setups...
Anyways, aside of hba->cmd_queue, tearing down hba->tmf_queue before
scsi_remove_host() may be problem too. Requests can still be
sent before and during scsi_remove_host(). If a request timed out,
task abort will be invoked to abort the request, during which
hba->tmf_queue is expected to be present. Please correct me if I am 
wrong.

Thanks,

Can Guo.
Can Guo Dec. 16, 2019, 5:46 a.m. UTC | #7
On 2019-12-16 11:12, cang@codeaurora.org wrote:
> On 2019-12-16 10:39, Bart Van Assche wrote:
>> On 2019-12-15 17:34, cang@codeaurora.org wrote:
>>> This is applied to 5.5/scsi-queue. The two changes I patsed from you 
>>> are
>>> not merged yet, I am still doing code review to them, so there is no
>>> blk_cleanup_queue() calls in my code base. I am just saying you may 
>>> move
>>> your blk_cleanup_queue() calls below cancel_work_sync(&hba->eh_work) 
>>> if
>>> my change applies. How do you think?
>>> 
>>> scsi_host_put() was there before but explicitly removed by
>>> afa3dfd42d205b106787476647735aa1de1a5d02. I agree with you, without 
>>> this
>>> change, there is memory leak.
>> 
>> Hi Can,
>> 
>> Since your patch restores a call that was removed earlier, please
>> consider adding a Fixes: tag to your patch.
>> 
>> Please also have a look at
>> https://git.kernel.org/pub/scm/linux/kernel/git/mkp/scsi.git/log/?h=5.6/scsi-queue.
>> As one can see my patches that introduce blk_cleanup_queue() and
>> blk_mq_free_tag_set() calls have already been queued on Martin's
>> 5.6/scsi-queue branch.
>> 
>> Bart.
> 
> Hi Bart,
> 
> Sure, I will add the Fixes tag and rebase my changes. How about the 
> logic
> part of this change? Does it look good to you?
> 
> Sorry I was not aware of that your changes have been applied to 
> 5.6/scsi-queue.
> I am still trying to get it tested on my setups...
> Anyways, aside of hba->cmd_queue, tearing down hba->tmf_queue before
> scsi_remove_host() may be problem too. Requests can still be
> sent before and during scsi_remove_host(). If a request timed out,
> task abort will be invoked to abort the request, during which
> hba->tmf_queue is expected to be present. Please correct me if I am 
> wrong.
> 
> Thanks,
> 
> Can Guo.

Hi Bart

Just found that I should also remove the ufshcd_dealloc_host() called
in ufshcd_pci_remove() to make sure the deallocation is only handled by
ufshcd_remove().

Thanks,

Can Guo.
Can Guo Dec. 16, 2019, 2:31 p.m. UTC | #8
On 2019-12-15 02:32, Bart Van Assche wrote:
> On 12/14/19 8:03 AM, Can Guo wrote:
>> In ufshcd_remove(), after SCSI host is removed, put it once so that 
>> its
>> resources can be released.
>> 
>> Signed-off-by: Can Guo <cang@codeaurora.org>
>> 
>> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
>> index b5966fa..a86b0fd 100644
>> --- a/drivers/scsi/ufs/ufshcd.c
>> +++ b/drivers/scsi/ufs/ufshcd.c
>> @@ -8251,6 +8251,7 @@ void ufshcd_remove(struct ufs_hba *hba)
>>   	ufs_bsg_remove(hba);
>>   	ufs_sysfs_remove_nodes(hba->dev);
>>   	scsi_remove_host(hba->host);
>> +	scsi_host_put(hba->host);
>>   	/* disable interrupts */
>>   	ufshcd_disable_intr(hba, hba->intr_mask);
>>   	ufshcd_hba_stop(hba, true);
> 
> Hi Can,
> 
> The UFS driver may queue work asynchronously and that asynchronous
> work may refer to the SCSI host, e.g. ufshcd_err_handler(). Is it
> guaranteed that all that asynchronous work has finished before
> scsi_host_put() is called?
> 
> Thanks,
> 
> Bart.

Hi Bart,

As SCSI host is allocated in ufshcd_platform_init() during platform
drive probe, it is much more appropriate if platform driver calls
ufshcd_dealloc_host() in their own drv->remove() path. How do you
think if I change it as below? If it is OK to you, please ignore my
previous mails.

diff --git a/drivers/scsi/ufs/ufs-qcom.c b/drivers/scsi/ufs/ufs-qcom.c
index 3d4582e..ea45756 100644
--- a/drivers/scsi/ufs/ufs-qcom.c
+++ b/drivers/scsi/ufs/ufs-qcom.c
@@ -3239,6 +3239,7 @@ static int ufs_qcom_remove(struct platform_device 
*pdev)

         pm_runtime_get_sync(&(pdev)->dev);
         ufshcd_remove(hba);
+       ufshcd_dealloc_host(hba);
         return 0;
  }

Thanks,

Can Guo.
Bart Van Assche Dec. 16, 2019, 5:39 p.m. UTC | #9
On 12/16/19 6:31 AM, cang@codeaurora.org wrote:
> As SCSI host is allocated in ufshcd_platform_init() during platform
> drive probe, it is much more appropriate if platform driver calls
> ufshcd_dealloc_host() in their own drv->remove() path. How do you
> think if I change it as below? If it is OK to you, please ignore my
> previous mails.
> 
> diff --git a/drivers/scsi/ufs/ufs-qcom.c b/drivers/scsi/ufs/ufs-qcom.c
> index 3d4582e..ea45756 100644
> --- a/drivers/scsi/ufs/ufs-qcom.c
> +++ b/drivers/scsi/ufs/ufs-qcom.c
> @@ -3239,6 +3239,7 @@ static int ufs_qcom_remove(struct platform_device 
> *pdev)
> 
>          pm_runtime_get_sync(&(pdev)->dev);
>          ufshcd_remove(hba);
> +       ufshcd_dealloc_host(hba);
>          return 0;
>   }

Hi Can,

Apparently some UFS drivers call ufshcd_remove() only and others (PCIe) 
call both ufshcd_remove() and ufshcd_dealloc_host(). I think that the 
above change will cause trouble for the PCIe driver unless the 
ufshcd_dealloc_host() call is removed from ufshcd_pci_remove().

Thanks,

Bart.
Bart Van Assche Dec. 16, 2019, 5:44 p.m. UTC | #10
On 12/15/19 7:12 PM, cang@codeaurora.org wrote:
> Sure, I will add the Fixes tag and rebase my changes. How about the logic
> part of this change? Does it look good to you?

Hi Can,

You may want to ask someone who is more familiar with the UFS driver 
than I to have a look. I'm not a UFS expert ...

> Sorry I was not aware of that your changes have been applied to 
> 5.6/scsi-queue.
> I am still trying to get it tested on my setups...
> Anyways, aside of hba->cmd_queue, tearing down hba->tmf_queue before
> scsi_remove_host() may be problem too. Requests can still be
> sent before and during scsi_remove_host(). If a request timed out,
> task abort will be invoked to abort the request, during which
> hba->tmf_queue is expected to be present. Please correct me if I am wrong.

I agree that the code I added in ufshcd_remove() probably needs to be 
moved somewhere below the scsi_remove_host() call.

Thanks,

Bart.
Greg Kroah-Hartman Dec. 16, 2019, 6:05 p.m. UTC | #11
On Mon, Dec 16, 2019 at 10:31:29PM +0800, cang@codeaurora.org wrote:
> On 2019-12-15 02:32, Bart Van Assche wrote:
> > On 12/14/19 8:03 AM, Can Guo wrote:
> > > In ufshcd_remove(), after SCSI host is removed, put it once so that
> > > its
> > > resources can be released.
> > > 
> > > Signed-off-by: Can Guo <cang@codeaurora.org>
> > > 
> > > diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> > > index b5966fa..a86b0fd 100644
> > > --- a/drivers/scsi/ufs/ufshcd.c
> > > +++ b/drivers/scsi/ufs/ufshcd.c
> > > @@ -8251,6 +8251,7 @@ void ufshcd_remove(struct ufs_hba *hba)
> > >   	ufs_bsg_remove(hba);
> > >   	ufs_sysfs_remove_nodes(hba->dev);
> > >   	scsi_remove_host(hba->host);
> > > +	scsi_host_put(hba->host);
> > >   	/* disable interrupts */
> > >   	ufshcd_disable_intr(hba, hba->intr_mask);
> > >   	ufshcd_hba_stop(hba, true);
> > 
> > Hi Can,
> > 
> > The UFS driver may queue work asynchronously and that asynchronous
> > work may refer to the SCSI host, e.g. ufshcd_err_handler(). Is it
> > guaranteed that all that asynchronous work has finished before
> > scsi_host_put() is called?
> > 
> > Thanks,
> > 
> > Bart.
> 
> Hi Bart,
> 
> As SCSI host is allocated in ufshcd_platform_init() during platform
> drive probe, it is much more appropriate if platform driver calls
> ufshcd_dealloc_host() in their own drv->remove() path. How do you
> think if I change it as below? If it is OK to you, please ignore my
> previous mails.
> 
> diff --git a/drivers/scsi/ufs/ufs-qcom.c b/drivers/scsi/ufs/ufs-qcom.c
> index 3d4582e..ea45756 100644
> --- a/drivers/scsi/ufs/ufs-qcom.c
> +++ b/drivers/scsi/ufs/ufs-qcom.c
> @@ -3239,6 +3239,7 @@ static int ufs_qcom_remove(struct platform_device
> *pdev)
> 
>         pm_runtime_get_sync(&(pdev)->dev);
>         ufshcd_remove(hba);
> +       ufshcd_dealloc_host(hba);
>         return 0;
>  }

Wait, why is this a platform device?  Don't you hang off of a pci
device?  Or am I missing something earlier in this patchset?

thanks,

greg k-h
Can Guo Dec. 17, 2019, 12:46 a.m. UTC | #12
On 2019-12-17 01:39, Bart Van Assche wrote:
> On 12/16/19 6:31 AM, cang@codeaurora.org wrote:
>> As SCSI host is allocated in ufshcd_platform_init() during platform
>> drive probe, it is much more appropriate if platform driver calls
>> ufshcd_dealloc_host() in their own drv->remove() path. How do you
>> think if I change it as below? If it is OK to you, please ignore my
>> previous mails.
>> 
>> diff --git a/drivers/scsi/ufs/ufs-qcom.c b/drivers/scsi/ufs/ufs-qcom.c
>> index 3d4582e..ea45756 100644
>> --- a/drivers/scsi/ufs/ufs-qcom.c
>> +++ b/drivers/scsi/ufs/ufs-qcom.c
>> @@ -3239,6 +3239,7 @@ static int ufs_qcom_remove(struct 
>> platform_device *pdev)
>> 
>>          pm_runtime_get_sync(&(pdev)->dev);
>>          ufshcd_remove(hba);
>> +       ufshcd_dealloc_host(hba);
>>          return 0;
>>   }
> 
> Hi Can,
> 
> Apparently some UFS drivers call ufshcd_remove() only and others
> (PCIe) call both ufshcd_remove() and ufshcd_dealloc_host(). I think
> that the above change will cause trouble for the PCIe driver unless
> the ufshcd_dealloc_host() call is removed from ufshcd_pci_remove().
> 
> Thanks,
> 
> Bart.

Hi Bart,

You may get me wrong. I mean we should do like what ufshcd-pci.c does.
As driver probe routine allocates SCSI host, then driver remove() should
de-allocate it. Meaning ufs_qcom_remove() should call both 
ufshcd_remove()
and ufshcd_dealloc_host().

diff --git a/drivers/scsi/ufs/ufs-qcom.c b/drivers/scsi/ufs/ufs-qcom.c
index 3d4582e..ea45756 100644
--- a/drivers/scsi/ufs/ufs-qcom.c
+++ b/drivers/scsi/ufs/ufs-qcom.c
@@ -3239,6 +3239,7 @@ static int ufs_qcom_remove(struct platform_device 
*pdev)

           pm_runtime_get_sync(&(pdev)->dev);
           ufshcd_remove(hba);
  +       ufshcd_dealloc_host(hba);
           return 0;
    }

Thanks,

Can Guo.
Can Guo Dec. 17, 2019, 12:50 a.m. UTC | #13
On 2019-12-17 02:05, Greg KH wrote:
> On Mon, Dec 16, 2019 at 10:31:29PM +0800, cang@codeaurora.org wrote:
>> On 2019-12-15 02:32, Bart Van Assche wrote:
>> > On 12/14/19 8:03 AM, Can Guo wrote:
>> > > In ufshcd_remove(), after SCSI host is removed, put it once so that
>> > > its
>> > > resources can be released.
>> > >
>> > > Signed-off-by: Can Guo <cang@codeaurora.org>
>> > >
>> > > diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
>> > > index b5966fa..a86b0fd 100644
>> > > --- a/drivers/scsi/ufs/ufshcd.c
>> > > +++ b/drivers/scsi/ufs/ufshcd.c
>> > > @@ -8251,6 +8251,7 @@ void ufshcd_remove(struct ufs_hba *hba)
>> > >   	ufs_bsg_remove(hba);
>> > >   	ufs_sysfs_remove_nodes(hba->dev);
>> > >   	scsi_remove_host(hba->host);
>> > > +	scsi_host_put(hba->host);
>> > >   	/* disable interrupts */
>> > >   	ufshcd_disable_intr(hba, hba->intr_mask);
>> > >   	ufshcd_hba_stop(hba, true);
>> >
>> > Hi Can,
>> >
>> > The UFS driver may queue work asynchronously and that asynchronous
>> > work may refer to the SCSI host, e.g. ufshcd_err_handler(). Is it
>> > guaranteed that all that asynchronous work has finished before
>> > scsi_host_put() is called?
>> >
>> > Thanks,
>> >
>> > Bart.
>> 
>> Hi Bart,
>> 
>> As SCSI host is allocated in ufshcd_platform_init() during platform
>> drive probe, it is much more appropriate if platform driver calls
>> ufshcd_dealloc_host() in their own drv->remove() path. How do you
>> think if I change it as below? If it is OK to you, please ignore my
>> previous mails.
>> 
>> diff --git a/drivers/scsi/ufs/ufs-qcom.c b/drivers/scsi/ufs/ufs-qcom.c
>> index 3d4582e..ea45756 100644
>> --- a/drivers/scsi/ufs/ufs-qcom.c
>> +++ b/drivers/scsi/ufs/ufs-qcom.c
>> @@ -3239,6 +3239,7 @@ static int ufs_qcom_remove(struct 
>> platform_device
>> *pdev)
>> 
>>         pm_runtime_get_sync(&(pdev)->dev);
>>         ufshcd_remove(hba);
>> +       ufshcd_dealloc_host(hba);
>>         return 0;
>>  }
> 
> Wait, why is this a platform device?  Don't you hang off of a pci
> device?  Or am I missing something earlier in this patchset?
> 
> thanks,
> 
> greg k-h

Hi Greg,

I am not saying someone is a platform device here. My point is
whoever allocates the SCSI host in its drv->probe(), should
de-allocate it in its own drv->remove(), just like what ufshcd-pci.c
does.

Thanks,

Can Guo.
Bart Van Assche Dec. 17, 2019, 1:15 a.m. UTC | #14
On 12/16/19 4:46 PM, cang@codeaurora.org wrote:
> On 2019-12-17 01:39, Bart Van Assche wrote:
>> Apparently some UFS drivers call ufshcd_remove() only and others
>> (PCIe) call both ufshcd_remove() and ufshcd_dealloc_host(). I think
>> that the above change will cause trouble for the PCIe driver unless
>> the ufshcd_dealloc_host() call is removed from ufshcd_pci_remove().
>
> You may get me wrong. I mean we should do like what ufshcd-pci.c does.
> As driver probe routine allocates SCSI host, then driver remove() should
> de-allocate it. Meaning ufs_qcom_remove() should call both ufshcd_remove()
> and ufshcd_dealloc_host().
> 
> diff --git a/drivers/scsi/ufs/ufs-qcom.c b/drivers/scsi/ufs/ufs-qcom.c
> index 3d4582e..ea45756 100644
> --- a/drivers/scsi/ufs/ufs-qcom.c
> +++ b/drivers/scsi/ufs/ufs-qcom.c
> @@ -3239,6 +3239,7 @@ static int ufs_qcom_remove(struct platform_device 
> *pdev)
> 
>            pm_runtime_get_sync(&(pdev)->dev);
>            ufshcd_remove(hba);
>   +       ufshcd_dealloc_host(hba);
>            return 0;
>     }

Hi Can,

If it is possible to move the ufshcd_dealloc_host() into ufshcd_remove() 
then I would prefer to do that. If all UFS transport drivers need that 
call then I think that call should happen from the UFS core instead of 
from the transport drivers.

Thanks,

Bart.
Can Guo Dec. 17, 2019, 1:31 a.m. UTC | #15
On 2019-12-17 09:15, Bart Van Assche wrote:
> On 12/16/19 4:46 PM, cang@codeaurora.org wrote:
>> On 2019-12-17 01:39, Bart Van Assche wrote:
>>> Apparently some UFS drivers call ufshcd_remove() only and others
>>> (PCIe) call both ufshcd_remove() and ufshcd_dealloc_host(). I think
>>> that the above change will cause trouble for the PCIe driver unless
>>> the ufshcd_dealloc_host() call is removed from ufshcd_pci_remove().
>> 
>> You may get me wrong. I mean we should do like what ufshcd-pci.c does.
>> As driver probe routine allocates SCSI host, then driver remove() 
>> should
>> de-allocate it. Meaning ufs_qcom_remove() should call both 
>> ufshcd_remove()
>> and ufshcd_dealloc_host().
>> 
>> diff --git a/drivers/scsi/ufs/ufs-qcom.c b/drivers/scsi/ufs/ufs-qcom.c
>> index 3d4582e..ea45756 100644
>> --- a/drivers/scsi/ufs/ufs-qcom.c
>> +++ b/drivers/scsi/ufs/ufs-qcom.c
>> @@ -3239,6 +3239,7 @@ static int ufs_qcom_remove(struct 
>> platform_device *pdev)
>> 
>>            pm_runtime_get_sync(&(pdev)->dev);
>>            ufshcd_remove(hba);
>>   +       ufshcd_dealloc_host(hba);
>>            return 0;
>>     }
> 
> Hi Can,
> 
> If it is possible to move the ufshcd_dealloc_host() into
> ufshcd_remove() then I would prefer to do that. If all UFS transport
> drivers need that call then I think that call should happen from the
> UFS core instead of from the transport drivers.
> 
> Thanks,
> 
> Bart.

Yeah, that is an once for all solution, but I not sure if PCI folks are
OK if I remove the ufshcd_dealloc_host() call from their driver.
In next version, I will try to make such change and see.

Thanks,
Can Guo.
diff mbox series

Patch

diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
index b5966fa..a86b0fd 100644
--- a/drivers/scsi/ufs/ufshcd.c
+++ b/drivers/scsi/ufs/ufshcd.c
@@ -8251,6 +8251,7 @@  void ufshcd_remove(struct ufs_hba *hba)
 	ufs_bsg_remove(hba);
 	ufs_sysfs_remove_nodes(hba->dev);
 	scsi_remove_host(hba->host);
+	scsi_host_put(hba->host);
 	/* disable interrupts */
 	ufshcd_disable_intr(hba, hba->intr_mask);
 	ufshcd_hba_stop(hba, true);