Re: [PATCH v2 09/30] cxlflash: Fix to stop interrupt processing on remove

[Matthew R. Ochs]

> On Sep 21, 2015, at 6:33 AM, Tomas Henzl  wrote:
> On 19.9.2015 01:26, Matthew R. Ochs wrote:
>>> On Sep 18, 2015, at 6:59 AM, Tomas Henzl  wrote:
>>> On 17.9.2015 19:16, Matthew R. Ochs wrote:
> On Sep 17, 2015, at 7:38 AM, Tomas Henzl  wrote:
> 
> On 16.9.2015 18:53, Matthew R. Ochs wrote:
>> Interrupt processing can run in parallel to a remove operation. This
>> can lead to a condition where the interrupt handler is processing with
>> memory that has been freed.
>> 
>> To avoid processing an interrupt while memory may be yanked, check for
>> removal while in the interrupt handler. Bail when removal is imminent.
>> 
>> Signed-off-by: Matthew R. Ochs 
>> Signed-off-by: Manoj N. Kumar 
>> ---
>> drivers/scsi/cxlflash/common.h |  2 ++
>> drivers/scsi/cxlflash/main.c   | 21 +++--
>> 2 files changed, 17 insertions(+), 6 deletions(-)
>> 
>> diff --git a/drivers/scsi/cxlflash/common.h 
>> b/drivers/scsi/cxlflash/common.h
>> index 1abe4e0..03d2cc6 100644
>> --- a/drivers/scsi/cxlflash/common.h
>> +++ b/drivers/scsi/cxlflash/common.h
>> @@ -103,6 +103,8 @@ struct cxlflash_cfg {
>>  enum cxlflash_lr_state lr_state;
>>  int lr_port;
>> 
>> +atomic_t remove_active;
>> +
>>  struct cxl_afu *cxl_afu;
>> 
>>  struct pci_pool *cxlflash_cmd_pool;
>> diff --git a/drivers/scsi/cxlflash/main.c b/drivers/scsi/cxlflash/main.c
>> index 6e85c77..89ee648 100644
>> --- a/drivers/scsi/cxlflash/main.c
>> +++ b/drivers/scsi/cxlflash/main.c
>> @@ -892,6 +892,7 @@ static void cxlflash_remove(struct pci_dev *pdev)
>>  spin_unlock_irqrestore(>tmf_waitq.lock, lock_flags);
>> 
>>  cfg->state = STATE_FAILTERM;
>> +atomic_inc(>remove_active);
> Hi Matthew,
> you could just call term_afu at this point, this way you don't
> need an additional check in all irq functions.
> Cheers,
> Tomas
 Hi Tomas,
 
 We actually do call term_afu() a few lines down from here. I don't follow
 how moving it here would help things.
>>> When you disable ints sooner (that is what term_afu does ?) you'll get no
>>> more ints later isn't this what you want?
>> Correct, that's what we want.
>> 
 The reason for the atomic was to provide something lightweight that we
 could check _inside_ the processing loop for the read-response queue
 handler. A check outside that loop doesn't really provide much in terms
 of closing or narrowing down the window of when freed memory can be
 accessed.
 
 As David Laight correctly pointed out, this approach does not completely
 close the window. We'd need something heavier to fully protect (e.g. a lock
 to wrap around the entire loop). I will look into adding this in a future 
 cycle
 when I can adequately test.
>>> term_afu calls free_irq and this function
>>> does not return until any executing interrupts for have completed.
>>> This is the sync mechanism you need, it's lightweight
>>> (does not add an additional check to your irq functions)
>>> and closes the race window completely.
>> Thanks for clarifying!
>> 
>> I looked at this closer and you are correct, free_irq() guarantees not
>> to return until the interrupt handler has completed. The current location
>> of term_afu() is appropriate as the memory that the handler touches is
>> not freed until the very end [by free_mem() and scsi_host_put()] of the
>> remove. Thus we can simply ignore this patch (I'll remove it in a v3).
> 
> OK. In some future patch please reorganize the remove function so,
> that it follows the template described in Documentation/PCI/pci.txt :
>   Disable the device from generating IRQs
>   Release the IRQ (free_irq())
>   Stop all DMA activity
>   Release DMA buffers (both streaming and coherent)
>   Unregister from other subsystems (e.g. scsi or netdev)
>   Release MMIO/IOP resources
>   Disable the device

Will do.

___
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev

Re: [PATCH v2 09/30] cxlflash: Fix to stop interrupt processing on remove

[Tomas Henzl]

On 19.9.2015 01:26, Matthew R. Ochs wrote:
>> On Sep 18, 2015, at 6:59 AM, Tomas Henzl  wrote:
>> On 17.9.2015 19:16, Matthew R. Ochs wrote:
 On Sep 17, 2015, at 7:38 AM, Tomas Henzl  wrote:

 On 16.9.2015 18:53, Matthew R. Ochs wrote:
> Interrupt processing can run in parallel to a remove operation. This
> can lead to a condition where the interrupt handler is processing with
> memory that has been freed.
>
> To avoid processing an interrupt while memory may be yanked, check for
> removal while in the interrupt handler. Bail when removal is imminent.
>
> Signed-off-by: Matthew R. Ochs 
> Signed-off-by: Manoj N. Kumar 
> ---
> drivers/scsi/cxlflash/common.h |  2 ++
> drivers/scsi/cxlflash/main.c   | 21 +++--
> 2 files changed, 17 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/scsi/cxlflash/common.h 
> b/drivers/scsi/cxlflash/common.h
> index 1abe4e0..03d2cc6 100644
> --- a/drivers/scsi/cxlflash/common.h
> +++ b/drivers/scsi/cxlflash/common.h
> @@ -103,6 +103,8 @@ struct cxlflash_cfg {
>   enum cxlflash_lr_state lr_state;
>   int lr_port;
>
> + atomic_t remove_active;
> +
>   struct cxl_afu *cxl_afu;
>
>   struct pci_pool *cxlflash_cmd_pool;
> diff --git a/drivers/scsi/cxlflash/main.c b/drivers/scsi/cxlflash/main.c
> index 6e85c77..89ee648 100644
> --- a/drivers/scsi/cxlflash/main.c
> +++ b/drivers/scsi/cxlflash/main.c
> @@ -892,6 +892,7 @@ static void cxlflash_remove(struct pci_dev *pdev)
>   spin_unlock_irqrestore(>tmf_waitq.lock, lock_flags);
>
>   cfg->state = STATE_FAILTERM;
> + atomic_inc(>remove_active);
 Hi Matthew,
 you could just call term_afu at this point, this way you don't
 need an additional check in all irq functions.
 Cheers,
 Tomas
>>> Hi Tomas,
>>>
>>> We actually do call term_afu() a few lines down from here. I don't follow
>>> how moving it here would help things.
>> When you disable ints sooner (that is what term_afu does ?) you'll get no
>> more ints later isn't this what you want?
> Correct, that's what we want.
>
>>> The reason for the atomic was to provide something lightweight that we
>>> could check _inside_ the processing loop for the read-response queue
>>> handler. A check outside that loop doesn't really provide much in terms
>>> of closing or narrowing down the window of when freed memory can be
>>> accessed.
>>>
>>> As David Laight correctly pointed out, this approach does not completely
>>> close the window. We'd need something heavier to fully protect (e.g. a lock
>>> to wrap around the entire loop). I will look into adding this in a future 
>>> cycle
>>> when I can adequately test.
>> term_afu calls free_irq and this function
>> does not return until any executing interrupts for have completed.
>> This is the sync mechanism you need, it's lightweight
>> (does not add an additional check to your irq functions)
>> and closes the race window completely.
> Thanks for clarifying!
>
> I looked at this closer and you are correct, free_irq() guarantees not
> to return until the interrupt handler has completed. The current location
> of term_afu() is appropriate as the memory that the handler touches is
> not freed until the very end [by free_mem() and scsi_host_put()] of the
> remove. Thus we can simply ignore this patch (I'll remove it in a v3).

OK. In some future patch please reorganize the remove function so,
that it follows the template described in Documentation/PCI/pci.txt :
Disable the device from generating IRQs
Release the IRQ (free_irq())
Stop all DMA activity
Release DMA buffers (both streaming and coherent)
Unregister from other subsystems (e.g. scsi or netdev)
Release MMIO/IOP resources
Disable the device

>
>
> -matt
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
> the body of a message to majord...@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

___
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev

Re: [PATCH v2 09/30] cxlflash: Fix to stop interrupt processing on remove

[Matthew R. Ochs]

> On Sep 18, 2015, at 6:59 AM, Tomas Henzl  wrote:
> On 17.9.2015 19:16, Matthew R. Ochs wrote:
>>> On Sep 17, 2015, at 7:38 AM, Tomas Henzl  wrote:
>>> 
>>> On 16.9.2015 18:53, Matthew R. Ochs wrote:
 Interrupt processing can run in parallel to a remove operation. This
 can lead to a condition where the interrupt handler is processing with
 memory that has been freed.
 
 To avoid processing an interrupt while memory may be yanked, check for
 removal while in the interrupt handler. Bail when removal is imminent.
 
 Signed-off-by: Matthew R. Ochs 
 Signed-off-by: Manoj N. Kumar 
 ---
 drivers/scsi/cxlflash/common.h |  2 ++
 drivers/scsi/cxlflash/main.c   | 21 +++--
 2 files changed, 17 insertions(+), 6 deletions(-)
 
 diff --git a/drivers/scsi/cxlflash/common.h 
 b/drivers/scsi/cxlflash/common.h
 index 1abe4e0..03d2cc6 100644
 --- a/drivers/scsi/cxlflash/common.h
 +++ b/drivers/scsi/cxlflash/common.h
 @@ -103,6 +103,8 @@ struct cxlflash_cfg {
enum cxlflash_lr_state lr_state;
int lr_port;
 
 +  atomic_t remove_active;
 +
struct cxl_afu *cxl_afu;
 
struct pci_pool *cxlflash_cmd_pool;
 diff --git a/drivers/scsi/cxlflash/main.c b/drivers/scsi/cxlflash/main.c
 index 6e85c77..89ee648 100644
 --- a/drivers/scsi/cxlflash/main.c
 +++ b/drivers/scsi/cxlflash/main.c
 @@ -892,6 +892,7 @@ static void cxlflash_remove(struct pci_dev *pdev)
spin_unlock_irqrestore(>tmf_waitq.lock, lock_flags);
 
cfg->state = STATE_FAILTERM;
 +  atomic_inc(>remove_active);
>>> Hi Matthew,
>>> you could just call term_afu at this point, this way you don't
>>> need an additional check in all irq functions.
>>> Cheers,
>>> Tomas
>> Hi Tomas,
>> 
>> We actually do call term_afu() a few lines down from here. I don't follow
>> how moving it here would help things.
> 
> When you disable ints sooner (that is what term_afu does ?) you'll get no
> more ints later isn't this what you want?

Correct, that's what we want.

>> The reason for the atomic was to provide something lightweight that we
>> could check _inside_ the processing loop for the read-response queue
>> handler. A check outside that loop doesn't really provide much in terms
>> of closing or narrowing down the window of when freed memory can be
>> accessed.
>> 
>> As David Laight correctly pointed out, this approach does not completely
>> close the window. We'd need something heavier to fully protect (e.g. a lock
>> to wrap around the entire loop). I will look into adding this in a future 
>> cycle
>> when I can adequately test.
> 
> term_afu calls free_irq and this function
> does not return until any executing interrupts for have completed.
> This is the sync mechanism you need, it's lightweight
> (does not add an additional check to your irq functions)
> and closes the race window completely.

Thanks for clarifying!

I looked at this closer and you are correct, free_irq() guarantees not
to return until the interrupt handler has completed. The current location
of term_afu() is appropriate as the memory that the handler touches is
not freed until the very end [by free_mem() and scsi_host_put()] of the
remove. Thus we can simply ignore this patch (I'll remove it in a v3).


-matt


___
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev

Re: [PATCH v2 09/30] cxlflash: Fix to stop interrupt processing on remove

[Tomas Henzl]

On 17.9.2015 19:16, Matthew R. Ochs wrote:
>> On Sep 17, 2015, at 7:38 AM, Tomas Henzl  wrote:
>>
>> On 16.9.2015 18:53, Matthew R. Ochs wrote:
>>> Interrupt processing can run in parallel to a remove operation. This
>>> can lead to a condition where the interrupt handler is processing with
>>> memory that has been freed.
>>>
>>> To avoid processing an interrupt while memory may be yanked, check for
>>> removal while in the interrupt handler. Bail when removal is imminent.
>>>
>>> Signed-off-by: Matthew R. Ochs 
>>> Signed-off-by: Manoj N. Kumar 
>>> ---
>>> drivers/scsi/cxlflash/common.h |  2 ++
>>> drivers/scsi/cxlflash/main.c   | 21 +++--
>>> 2 files changed, 17 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/drivers/scsi/cxlflash/common.h b/drivers/scsi/cxlflash/common.h
>>> index 1abe4e0..03d2cc6 100644
>>> --- a/drivers/scsi/cxlflash/common.h
>>> +++ b/drivers/scsi/cxlflash/common.h
>>> @@ -103,6 +103,8 @@ struct cxlflash_cfg {
>>> enum cxlflash_lr_state lr_state;
>>> int lr_port;
>>>
>>> +   atomic_t remove_active;
>>> +
>>> struct cxl_afu *cxl_afu;
>>>
>>> struct pci_pool *cxlflash_cmd_pool;
>>> diff --git a/drivers/scsi/cxlflash/main.c b/drivers/scsi/cxlflash/main.c
>>> index 6e85c77..89ee648 100644
>>> --- a/drivers/scsi/cxlflash/main.c
>>> +++ b/drivers/scsi/cxlflash/main.c
>>> @@ -892,6 +892,7 @@ static void cxlflash_remove(struct pci_dev *pdev)
>>> spin_unlock_irqrestore(>tmf_waitq.lock, lock_flags);
>>>
>>> cfg->state = STATE_FAILTERM;
>>> +   atomic_inc(>remove_active);
>> Hi Matthew,
>> you could just call term_afu at this point, this way you don't
>> need an additional check in all irq functions.
>> Cheers,
>> Tomas
> Hi Tomas,
>
> We actually do call term_afu() a few lines down from here. I don't follow
> how moving it here would help things.

When you disable ints sooner (that is what term_afu does ?) you'll get no
more ints later isn't this what you want?

>
> The reason for the atomic was to provide something lightweight that we
> could check _inside_ the processing loop for the read-response queue
> handler. A check outside that loop doesn't really provide much in terms
> of closing or narrowing down the window of when freed memory can be
> accessed.
>
> As David Laight correctly pointed out, this approach does not completely
> close the window. We'd need something heavier to fully protect (e.g. a lock
> to wrap around the entire loop). I will look into adding this in a future 
> cycle
> when I can adequately test.

term_afu calls free_irq and this function
does not return until any executing interrupts for have completed.
This is the sync mechanism you need, it's lightweight
(does not add an additional check to your irq functions)
and closes the race window completely.

--tm

>
>
> -matt
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
> the body of a message to majord...@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

___
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev

RE: [PATCH v2 09/30] cxlflash: Fix to stop interrupt processing on remove

[David Laight]

From: Linuxppc-dev Matthew R. Ochs
> Sent: 16 September 2015 22:28
> Interrupt processing can run in parallel to a remove operation. This
> can lead to a condition where the interrupt handler is processing with
> memory that has been freed.
> 
> To avoid processing an interrupt while memory may be yanked, check for
> removal while in the interrupt handler. Bail when removal is imminent.

On the face of it this just reduces the size of the window somewhat.

What happens if the interrupt routine reads the flag just before it is set
(so is processing the entry that is being removed) and is then (say)
interrupted by a higher priority interrupt that takes longer to execute than
the remove code?

You've still got an interrupt routine accessing freed memory.

David

___
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev

Re: [PATCH v2 09/30] cxlflash: Fix to stop interrupt processing on remove

[Tomas Henzl]

On 16.9.2015 18:53, Matthew R. Ochs wrote:
> Interrupt processing can run in parallel to a remove operation. This
> can lead to a condition where the interrupt handler is processing with
> memory that has been freed.
> 
> To avoid processing an interrupt while memory may be yanked, check for
> removal while in the interrupt handler. Bail when removal is imminent.
>
> Signed-off-by: Matthew R. Ochs 
> Signed-off-by: Manoj N. Kumar 
> ---
>  drivers/scsi/cxlflash/common.h |  2 ++
>  drivers/scsi/cxlflash/main.c   | 21 +++--
>  2 files changed, 17 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/scsi/cxlflash/common.h b/drivers/scsi/cxlflash/common.h
> index 1abe4e0..03d2cc6 100644
> --- a/drivers/scsi/cxlflash/common.h
> +++ b/drivers/scsi/cxlflash/common.h
> @@ -103,6 +103,8 @@ struct cxlflash_cfg {
>   enum cxlflash_lr_state lr_state;
>   int lr_port;
>  
> + atomic_t remove_active;
> +
>   struct cxl_afu *cxl_afu;
>  
>   struct pci_pool *cxlflash_cmd_pool;
> diff --git a/drivers/scsi/cxlflash/main.c b/drivers/scsi/cxlflash/main.c
> index 6e85c77..89ee648 100644
> --- a/drivers/scsi/cxlflash/main.c
> +++ b/drivers/scsi/cxlflash/main.c
> @@ -892,6 +892,7 @@ static void cxlflash_remove(struct pci_dev *pdev)
>   spin_unlock_irqrestore(>tmf_waitq.lock, lock_flags);
>  
>   cfg->state = STATE_FAILTERM;
> + atomic_inc(>remove_active);

Hi Matthew,
you could just call term_afu at this point, this way you don't
need an additional check in all irq functions.
Cheers,
Tomas

>   cxlflash_stop_term_user_contexts(cfg);
>  
>   switch (cfg->init_state) {
> @@ -1380,16 +1381,20 @@ static void afu_err_intr_init(struct afu *afu)
>  static irqreturn_t cxlflash_sync_err_irq(int irq, void *data)
>  {
>   struct afu *afu = (struct afu *)data;
> + struct cxlflash_cfg *cfg = afu->parent;
>   u64 reg;
>   u64 reg_unmasked;
>  
> + if (atomic_read(>remove_active))
> + goto out;
> +
>   reg = readq_be(>host_map->intr_status);
>   reg_unmasked = (reg & SISL_ISTATUS_UNMASK);
>  
>   if (reg_unmasked == 0UL) {
>   pr_err("%s: %llX: spurious interrupt, intr_status %016llX\n",
>  __func__, (u64)afu, reg);
> - goto cxlflash_sync_err_irq_exit;
> + goto out;
>   }
>  
>   pr_err("%s: %llX: unexpected interrupt, intr_status %016llX\n",
> @@ -1397,7 +1402,7 @@ static irqreturn_t cxlflash_sync_err_irq(int irq, void 
> *data)
>  
>   writeq_be(reg_unmasked, >host_map->intr_clear);
>  
> -cxlflash_sync_err_irq_exit:
> +out:
>   pr_debug("%s: returning rc=%d\n", __func__, IRQ_HANDLED);
>   return IRQ_HANDLED;
>  }
> @@ -1412,6 +1417,7 @@ cxlflash_sync_err_irq_exit:
>  static irqreturn_t cxlflash_rrq_irq(int irq, void *data)
>  {
>   struct afu *afu = (struct afu *)data;
> + struct cxlflash_cfg *cfg = afu->parent;
>   struct afu_cmd *cmd;
>   bool toggle = afu->toggle;
>   u64 entry,
> @@ -1421,8 +1427,10 @@ static irqreturn_t cxlflash_rrq_irq(int irq, void 
> *data)
>  
>   /* Process however many RRQ entries that are ready */
>   while (true) {
> - entry = *hrrq_curr;
> + if (atomic_read(>remove_active))
> + goto out;
>  
> + entry = *hrrq_curr;
>   if ((entry & SISL_RESP_HANDLE_T_BIT) != toggle)
>   break;
>  
> @@ -1440,7 +1448,7 @@ static irqreturn_t cxlflash_rrq_irq(int irq, void *data)
>  
>   afu->hrrq_curr = hrrq_curr;
>   afu->toggle = toggle;
> -
> +out:
>   return IRQ_HANDLED;
>  }
>  
> @@ -1454,7 +1462,7 @@ static irqreturn_t cxlflash_rrq_irq(int irq, void *data)
>  static irqreturn_t cxlflash_async_err_irq(int irq, void *data)
>  {
>   struct afu *afu = (struct afu *)data;
> - struct cxlflash_cfg *cfg;
> + struct cxlflash_cfg *cfg = afu->parent;
>   u64 reg_unmasked;
>   const struct asyc_intr_info *info;
>   struct sisl_global_map *global = >afu_map->global;
> @@ -1462,7 +1470,8 @@ static irqreturn_t cxlflash_async_err_irq(int irq, void 
> *data)
>   u8 port;
>   int i;
>  
> - cfg = afu->parent;
> + if (atomic_read(>remove_active))
> + goto out;
>  
>   reg = readq_be(>regs.aintr_status);
>   reg_unmasked = (reg & SISL_ASTATUS_UNMASK);

___
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev

Re: [PATCH v2 09/30] cxlflash: Fix to stop interrupt processing on remove

[Matthew R. Ochs]

> On Sep 17, 2015, at 6:58 AM, David Laight  wrote:
> 
> From: Linuxppc-dev Matthew R. Ochs
>> Sent: 16 September 2015 22:28
>> Interrupt processing can run in parallel to a remove operation. This
>> can lead to a condition where the interrupt handler is processing with
>> memory that has been freed.
>> 
>> To avoid processing an interrupt while memory may be yanked, check for
>> removal while in the interrupt handler. Bail when removal is imminent.
> 
> On the face of it this just reduces the size of the window somewhat.

Agreed.

> 
> What happens if the interrupt routine reads the flag just before it is set
> (so is processing the entry that is being removed) and is then (say)
> interrupted by a higher priority interrupt that takes longer to execute than
> the remove code?

Understood. To completely close we'd need to either introduce a lock or a
reciprocal flag/count such that the remove doesn't make forward progress
until after interrupt processing has completed. I can look at introducing such
a mechanism in a later patch to fully remove the exposure.


-matt


___
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev

Re: [PATCH v2 09/30] cxlflash: Fix to stop interrupt processing on remove

[Matthew R. Ochs]

> On Sep 17, 2015, at 7:38 AM, Tomas Henzl  wrote:
> 
> On 16.9.2015 18:53, Matthew R. Ochs wrote:
>> Interrupt processing can run in parallel to a remove operation. This
>> can lead to a condition where the interrupt handler is processing with
>> memory that has been freed.
>> 
>> To avoid processing an interrupt while memory may be yanked, check for
>> removal while in the interrupt handler. Bail when removal is imminent.
>> 
>> Signed-off-by: Matthew R. Ochs 
>> Signed-off-by: Manoj N. Kumar 
>> ---
>> drivers/scsi/cxlflash/common.h |  2 ++
>> drivers/scsi/cxlflash/main.c   | 21 +++--
>> 2 files changed, 17 insertions(+), 6 deletions(-)
>> 
>> diff --git a/drivers/scsi/cxlflash/common.h b/drivers/scsi/cxlflash/common.h
>> index 1abe4e0..03d2cc6 100644
>> --- a/drivers/scsi/cxlflash/common.h
>> +++ b/drivers/scsi/cxlflash/common.h
>> @@ -103,6 +103,8 @@ struct cxlflash_cfg {
>>  enum cxlflash_lr_state lr_state;
>>  int lr_port;
>> 
>> +atomic_t remove_active;
>> +
>>  struct cxl_afu *cxl_afu;
>> 
>>  struct pci_pool *cxlflash_cmd_pool;
>> diff --git a/drivers/scsi/cxlflash/main.c b/drivers/scsi/cxlflash/main.c
>> index 6e85c77..89ee648 100644
>> --- a/drivers/scsi/cxlflash/main.c
>> +++ b/drivers/scsi/cxlflash/main.c
>> @@ -892,6 +892,7 @@ static void cxlflash_remove(struct pci_dev *pdev)
>>  spin_unlock_irqrestore(>tmf_waitq.lock, lock_flags);
>> 
>>  cfg->state = STATE_FAILTERM;
>> +atomic_inc(>remove_active);
> 
> Hi Matthew,
> you could just call term_afu at this point, this way you don't
> need an additional check in all irq functions.
> Cheers,
> Tomas

Hi Tomas,

We actually do call term_afu() a few lines down from here. I don't follow
how moving it here would help things.

The reason for the atomic was to provide something lightweight that we
could check _inside_ the processing loop for the read-response queue
handler. A check outside that loop doesn't really provide much in terms
of closing or narrowing down the window of when freed memory can be
accessed.

As David Laight correctly pointed out, this approach does not completely
close the window. We'd need something heavier to fully protect (e.g. a lock
to wrap around the entire loop). I will look into adding this in a future cycle
when I can adequately test.


-matt

___
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev