[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <0dcc8f8e-4aab-4676-eb95-5f45b5cda1b2@nvidia.com>
Date: Tue, 22 Jun 2021 10:59:53 +0530
From: Vidya Sagar <vidyas@...dia.com>
To: Kishon Vijay Abraham I <kishon@...com>
CC: <linux-pci@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
"Lorenzo Pieralisi" <lorenzo.pieralisi@....com>,
Andrew Murray <amurray@...goodpenguin.co.uk>,
Bjorn Helgaas <bhelgaas@...gle.com>,
"Athani Nadeem Ladkhan" <nadeem@...ence.com>,
Tom Joseph <tjoseph@...ence.com>,
Manikanta Maddireddy <mmaddireddy@...dia.com>,
Om Prakash Singh <omp@...dia.com>,
Krishna Thota <kthota@...dia.com>
Subject: Re: [PATCH v2 2/5] PCI: endpoint: Replace spinlock with mutex
On 6/21/2021 7:07 PM, Kishon Vijay Abraham I wrote:
> External email: Use caution opening links or attachments
>
>
> Hi Vidya Sagar,
>
> On 21/06/21 3:08 pm, Vidya Sagar wrote:
>>
>>
>> On 6/21/2021 10:44 AM, Kishon Vijay Abraham I wrote:
>>> External email: Use caution opening links or attachments
>>>
>>>
>>> Hi Vidya Sagar,
>>>
>>> On 11/06/21 3:22 pm, Vidya Sagar wrote:
>>>> Hi Kishon,
>>>> Apologies for bringup it up this late.
>>>> I'm wondering if there was any issue which this patch tried to address?
>>>
>>> There was one function pci_epc_linkup() which was expected to be invoked
>>> in interrupt context (basically when the LINKUP interrupt is raised).
>>> But after it was moved to use atomic notifier, all the EPC core APIs
>>> were replaced to use mutex.
>>>> Actually, "The pci_epc_ops is not intended to be invoked from interrupt
>>>> context" isn't true in case of Tegra194. We do call
>>>> dw_pcie_ep_init_notify() API from threaded irq service routine and it
>>>> eventually calls mutext_lock() of pci_epc_get_features() which is
>>>> reusulting in the following warning log.
>>>> BUG: sleeping function called from invalid context at
>>>> kernel/locking/mutex.c:
>>>> Would like hear your comments on it.
>> After reviewing the logs and code again, I think it was my mistake to
>> come to early conclusion that it was because of calling mutex_lock() in
>> the atomic context. It is clear now.
>>
>> I would like to understand the reason behind putting locks in the epc
>> core driver before calling ops.
>
> There could be two different functions trying to configure endpoint
> controller (could be a multi-function endpoint) and the framework should
> guarantee the hardware is not accessed by both the functions simultaneously.
Not all ops functions need to be protected by the lock. for ex:-
get/set_msi(x)(), raise_irq(), get_features() don't need to be protected
by a global lock. So, transferring the synchronization responsibility to
the controller driver gives an efficient control on locking.
>
>> I believe the ops callers should implement lock if they are concurrently
>> accessing the ops instead of adding a global lock in the epc core.
> This can only protect within a function and not across multiple functions.
>> This would help in scenarios like the one below.
>>
>> We have a performance oriented endpoint function driver which calls
>> map, unmap & raise_irq ops from softirq context and because of
>> mutex_lock(), we can't do that now. epc core driver should not restrict
>> the function drivers to use only non-atomic functions.
>
> Not sure what exactly the function driver does but can't map/unmap be
> done for a big block once to optimize and operate on that buffer? I'd
> assume you are having a custom driver on the host side too?
We implemented a function driver that provides a virtual ethernet
interface. To complement this, we have a PCIe device driver on the host
that exposes a virtual ethernet interface in the host system.
start_xmit() in the function driver of the virtual ethernet interface is
a soft irq which maps/unmaps each skb buffer dynamically. So, a one time
static mapping is not possible. Similarly, because of the global lock,
start_xmit() can not raise_irq() to the host.
Thanks,
Vidya Sagar
>
> Thanks
> Kishon
>
>>
>> Thanks,
>> Vidya Sagar
>>>
>>> I don't think it is ideal to initialize EPC in interrupt context (unless
>>> there is a specific reason for it). EPC initialization can be moved to
>>> bottom half similar to how commands are handled after LINKUP.
>>
>>>
>>> Thanks
>>> Kishon
>>>
>>>>
>>>> Thanks,
>>>> Vidya Sagar
>>>>
>>>> On 2/12/2020 4:55 PM, Kishon Vijay Abraham I wrote:
>>>>> External email: Use caution opening links or attachments
>>>>>
>>>>>
>>>>> The pci_epc_ops is not intended to be invoked from interrupt context.
>>>>> Hence replace spin_lock_irqsave and spin_unlock_irqrestore with
>>>>> mutex_lock and mutex_unlock respectively.
>>>>>
>>>>> Signed-off-by: Kishon Vijay Abraham I <kishon@...com>
>>>>> ---
>>>>> drivers/pci/endpoint/pci-epc-core.c | 82
>>>>> +++++++++++------------------
>>>>> include/linux/pci-epc.h | 6 +--
>>>>> 2 files changed, 34 insertions(+), 54 deletions(-)
>>>>>
>>>>> diff --git a/drivers/pci/endpoint/pci-epc-core.c
>>>>> b/drivers/pci/endpoint/pci-epc-core.c
>>>>> index 2f6436599fcb..e51a12ed85bb 100644
>>>>> --- a/drivers/pci/endpoint/pci-epc-core.c
>>>>> +++ b/drivers/pci/endpoint/pci-epc-core.c
>>>>> @@ -120,7 +120,6 @@ const struct pci_epc_features
>>>>> *pci_epc_get_features(struct pci_epc *epc,
>>>>> u8 func_no)
>>>>> {
>>>>> const struct pci_epc_features *epc_features;
>>>>> - unsigned long flags;
>>>>>
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions)
>>>>> return NULL;
>>>>> @@ -128,9 +127,9 @@ const struct pci_epc_features
>>>>> *pci_epc_get_features(struct pci_epc *epc,
>>>>> if (!epc->ops->get_features)
>>>>> return NULL;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> epc_features = epc->ops->get_features(epc, func_no);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> return epc_features;
>>>>> }
>>>>> @@ -144,14 +143,12 @@ EXPORT_SYMBOL_GPL(pci_epc_get_features);
>>>>> */
>>>>> void pci_epc_stop(struct pci_epc *epc)
>>>>> {
>>>>> - unsigned long flags;
>>>>> -
>>>>> if (IS_ERR(epc) || !epc->ops->stop)
>>>>> return;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> epc->ops->stop(epc);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>> }
>>>>> EXPORT_SYMBOL_GPL(pci_epc_stop);
>>>>>
>>>>> @@ -164,7 +161,6 @@ EXPORT_SYMBOL_GPL(pci_epc_stop);
>>>>> int pci_epc_start(struct pci_epc *epc)
>>>>> {
>>>>> int ret;
>>>>> - unsigned long flags;
>>>>>
>>>>> if (IS_ERR(epc))
>>>>> return -EINVAL;
>>>>> @@ -172,9 +168,9 @@ int pci_epc_start(struct pci_epc *epc)
>>>>> if (!epc->ops->start)
>>>>> return 0;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> ret = epc->ops->start(epc);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> return ret;
>>>>> }
>>>>> @@ -193,7 +189,6 @@ int pci_epc_raise_irq(struct pci_epc *epc, u8
>>>>> func_no,
>>>>> enum pci_epc_irq_type type, u16 interrupt_num)
>>>>> {
>>>>> int ret;
>>>>> - unsigned long flags;
>>>>>
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions)
>>>>> return -EINVAL;
>>>>> @@ -201,9 +196,9 @@ int pci_epc_raise_irq(struct pci_epc *epc, u8
>>>>> func_no,
>>>>> if (!epc->ops->raise_irq)
>>>>> return 0;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> ret = epc->ops->raise_irq(epc, func_no, type, interrupt_num);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> return ret;
>>>>> }
>>>>> @@ -219,7 +214,6 @@ EXPORT_SYMBOL_GPL(pci_epc_raise_irq);
>>>>> int pci_epc_get_msi(struct pci_epc *epc, u8 func_no)
>>>>> {
>>>>> int interrupt;
>>>>> - unsigned long flags;
>>>>>
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions)
>>>>> return 0;
>>>>> @@ -227,9 +221,9 @@ int pci_epc_get_msi(struct pci_epc *epc, u8
>>>>> func_no)
>>>>> if (!epc->ops->get_msi)
>>>>> return 0;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> interrupt = epc->ops->get_msi(epc, func_no);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> if (interrupt < 0)
>>>>> return 0;
>>>>> @@ -252,7 +246,6 @@ int pci_epc_set_msi(struct pci_epc *epc, u8
>>>>> func_no, u8 interrupts)
>>>>> {
>>>>> int ret;
>>>>> u8 encode_int;
>>>>> - unsigned long flags;
>>>>>
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions ||
>>>>> interrupts > 32)
>>>>> @@ -263,9 +256,9 @@ int pci_epc_set_msi(struct pci_epc *epc, u8
>>>>> func_no, u8 interrupts)
>>>>>
>>>>> encode_int = order_base_2(interrupts);
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> ret = epc->ops->set_msi(epc, func_no, encode_int);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> return ret;
>>>>> }
>>>>> @@ -281,7 +274,6 @@ EXPORT_SYMBOL_GPL(pci_epc_set_msi);
>>>>> int pci_epc_get_msix(struct pci_epc *epc, u8 func_no)
>>>>> {
>>>>> int interrupt;
>>>>> - unsigned long flags;
>>>>>
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions)
>>>>> return 0;
>>>>> @@ -289,9 +281,9 @@ int pci_epc_get_msix(struct pci_epc *epc, u8
>>>>> func_no)
>>>>> if (!epc->ops->get_msix)
>>>>> return 0;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> interrupt = epc->ops->get_msix(epc, func_no);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> if (interrupt < 0)
>>>>> return 0;
>>>>> @@ -311,7 +303,6 @@ EXPORT_SYMBOL_GPL(pci_epc_get_msix);
>>>>> int pci_epc_set_msix(struct pci_epc *epc, u8 func_no, u16
>>>>> interrupts)
>>>>> {
>>>>> int ret;
>>>>> - unsigned long flags;
>>>>>
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions ||
>>>>> interrupts < 1 || interrupts > 2048)
>>>>> @@ -320,9 +311,9 @@ int pci_epc_set_msix(struct pci_epc *epc, u8
>>>>> func_no, u16 interrupts)
>>>>> if (!epc->ops->set_msix)
>>>>> return 0;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> ret = epc->ops->set_msix(epc, func_no, interrupts - 1);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> return ret;
>>>>> }
>>>>> @@ -339,17 +330,15 @@ EXPORT_SYMBOL_GPL(pci_epc_set_msix);
>>>>> void pci_epc_unmap_addr(struct pci_epc *epc, u8 func_no,
>>>>> phys_addr_t phys_addr)
>>>>> {
>>>>> - unsigned long flags;
>>>>> -
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions)
>>>>> return;
>>>>>
>>>>> if (!epc->ops->unmap_addr)
>>>>> return;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> epc->ops->unmap_addr(epc, func_no, phys_addr);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>> }
>>>>> EXPORT_SYMBOL_GPL(pci_epc_unmap_addr);
>>>>>
>>>>> @@ -367,7 +356,6 @@ int pci_epc_map_addr(struct pci_epc *epc, u8
>>>>> func_no,
>>>>> phys_addr_t phys_addr, u64 pci_addr, size_t
>>>>> size)
>>>>> {
>>>>> int ret;
>>>>> - unsigned long flags;
>>>>>
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions)
>>>>> return -EINVAL;
>>>>> @@ -375,9 +363,9 @@ int pci_epc_map_addr(struct pci_epc *epc, u8
>>>>> func_no,
>>>>> if (!epc->ops->map_addr)
>>>>> return 0;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> ret = epc->ops->map_addr(epc, func_no, phys_addr, pci_addr,
>>>>> size);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> return ret;
>>>>> }
>>>>> @@ -394,8 +382,6 @@ EXPORT_SYMBOL_GPL(pci_epc_map_addr);
>>>>> void pci_epc_clear_bar(struct pci_epc *epc, u8 func_no,
>>>>> struct pci_epf_bar *epf_bar)
>>>>> {
>>>>> - unsigned long flags;
>>>>> -
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions ||
>>>>> (epf_bar->barno == BAR_5 &&
>>>>> epf_bar->flags & PCI_BASE_ADDRESS_MEM_TYPE_64))
>>>>> @@ -404,9 +390,9 @@ void pci_epc_clear_bar(struct pci_epc *epc, u8
>>>>> func_no,
>>>>> if (!epc->ops->clear_bar)
>>>>> return;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> epc->ops->clear_bar(epc, func_no, epf_bar);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>> }
>>>>> EXPORT_SYMBOL_GPL(pci_epc_clear_bar);
>>>>>
>>>>> @@ -422,7 +408,6 @@ int pci_epc_set_bar(struct pci_epc *epc, u8
>>>>> func_no,
>>>>> struct pci_epf_bar *epf_bar)
>>>>> {
>>>>> int ret;
>>>>> - unsigned long irq_flags;
>>>>> int flags = epf_bar->flags;
>>>>>
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions ||
>>>>> @@ -437,9 +422,9 @@ int pci_epc_set_bar(struct pci_epc *epc, u8
>>>>> func_no,
>>>>> if (!epc->ops->set_bar)
>>>>> return 0;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, irq_flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> ret = epc->ops->set_bar(epc, func_no, epf_bar);
>>>>> - spin_unlock_irqrestore(&epc->lock, irq_flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> return ret;
>>>>> }
>>>>> @@ -460,7 +445,6 @@ int pci_epc_write_header(struct pci_epc *epc, u8
>>>>> func_no,
>>>>> struct pci_epf_header *header)
>>>>> {
>>>>> int ret;
>>>>> - unsigned long flags;
>>>>>
>>>>> if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions)
>>>>> return -EINVAL;
>>>>> @@ -468,9 +452,9 @@ int pci_epc_write_header(struct pci_epc *epc, u8
>>>>> func_no,
>>>>> if (!epc->ops->write_header)
>>>>> return 0;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> ret = epc->ops->write_header(epc, func_no, header);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> return ret;
>>>>> }
>>>>> @@ -487,8 +471,6 @@ EXPORT_SYMBOL_GPL(pci_epc_write_header);
>>>>> */
>>>>> int pci_epc_add_epf(struct pci_epc *epc, struct pci_epf *epf)
>>>>> {
>>>>> - unsigned long flags;
>>>>> -
>>>>> if (epf->epc)
>>>>> return -EBUSY;
>>>>>
>>>>> @@ -500,9 +482,9 @@ int pci_epc_add_epf(struct pci_epc *epc, struct
>>>>> pci_epf *epf)
>>>>>
>>>>> epf->epc = epc;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> list_add_tail(&epf->list, &epc->pci_epf);
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>>
>>>>> return 0;
>>>>> }
>>>>> @@ -517,15 +499,13 @@ EXPORT_SYMBOL_GPL(pci_epc_add_epf);
>>>>> */
>>>>> void pci_epc_remove_epf(struct pci_epc *epc, struct pci_epf *epf)
>>>>> {
>>>>> - unsigned long flags;
>>>>> -
>>>>> if (!epc || IS_ERR(epc) || !epf)
>>>>> return;
>>>>>
>>>>> - spin_lock_irqsave(&epc->lock, flags);
>>>>> + mutex_lock(&epc->lock);
>>>>> list_del(&epf->list);
>>>>> epf->epc = NULL;
>>>>> - spin_unlock_irqrestore(&epc->lock, flags);
>>>>> + mutex_unlock(&epc->lock);
>>>>> }
>>>>> EXPORT_SYMBOL_GPL(pci_epc_remove_epf);
>>>>>
>>>>> @@ -604,7 +584,7 @@ __pci_epc_create(struct device *dev, const struct
>>>>> pci_epc_ops *ops,
>>>>> goto err_ret;
>>>>> }
>>>>>
>>>>> - spin_lock_init(&epc->lock);
>>>>> + mutex_init(&epc->lock);
>>>>> INIT_LIST_HEAD(&epc->pci_epf);
>>>>> ATOMIC_INIT_NOTIFIER_HEAD(&epc->notifier);
>>>>>
>>>>> diff --git a/include/linux/pci-epc.h b/include/linux/pci-epc.h
>>>>> index 36644ccd32ac..9dd60f2e9705 100644
>>>>> --- a/include/linux/pci-epc.h
>>>>> +++ b/include/linux/pci-epc.h
>>>>> @@ -88,7 +88,7 @@ struct pci_epc_mem {
>>>>> * @mem: address space of the endpoint controller
>>>>> * @max_functions: max number of functions that can be configured in
>>>>> this EPC
>>>>> * @group: configfs group representing the PCI EPC device
>>>>> - * @lock: spinlock to protect pci_epc ops
>>>>> + * @lock: mutex to protect pci_epc ops
>>>>> * @notifier: used to notify EPF of any EPC events (like linkup)
>>>>> */
>>>>> struct pci_epc {
>>>>> @@ -98,8 +98,8 @@ struct pci_epc {
>>>>> struct pci_epc_mem *mem;
>>>>> u8 max_functions;
>>>>> struct config_group *group;
>>>>> - /* spinlock to protect against concurrent access of EP
>>>>> controller */
>>>>> - spinlock_t lock;
>>>>> + /* mutex to protect against concurrent access of EP
>>>>> controller */
>>>>> + struct mutex lock;
>>>>> struct atomic_notifier_head notifier;
>>>>> };
>>>>>
>>>>> --
>>>>> 2.17.1
>>>>>
Powered by blists - more mailing lists