Date: Fri, 21 Jun 2024 17:12:15 +0800
From: Jiwei Sun <sunjw10@...look.com>
To: Paul M Stillwell Jr <paul.m.stillwell.jr@...el.com>,
 Bjorn Helgaas <helgaas@...nel.org>
Cc: nirmal.patel@...ux.intel.com, jonathan.derrick@...ux.dev,
 lpieralisi@...nel.org, kw@...ux.com, robh@...nel.org, bhelgaas@...gle.com,
 linux-pci@...r.kernel.org, linux-kernel@...r.kernel.org, sunjw10@...ovo.com,
 ahuang12@...ovo.com, Thomas Gleixner <tglx@...utronix.de>,
 Keith Busch <kbusch@...nel.org>
Subject: Re: [PATCH] PCI: vmd: Use raw spinlock for cfg_lock

On 6/20/24 23:10, Paul M Stillwell Jr wrote:
> On 6/20/2024 1:57 AM, Jiwei Sun wrote:
>>
>>
>> On 6/20/24 04:00, Bjorn Helgaas wrote:
>>> [+cc Thomas in case he has msi_lock comment, Keith in case he has
>>> cfg_lock comment]
>>>
>>> On Wed, Jun 19, 2024 at 07:27:59PM +0800, Jiwei Sun wrote:
>>>> From: Jiwei Sun <sunjw10@...ovo.com>
>>>>
>>>> If the kernel is built with the following configurations and booting
>>>>    CONFIG_VMD=y
>>>>    CONFIG_DEBUG_LOCKDEP=y
>>>>    CONFIG_DEBUG_SPINLOCK=y
>>>>    CONFIG_PROVE_LOCKING=y
>>>>    CONFIG_PROVE_RAW_LOCK_NESTING=y
>>>>
>>>> The following log appears,
>>>>
>>>> =============================
>>>> [ BUG: Invalid wait context ]
>>>> 6.10.0-rc4 #80 Not tainted
>>>> -----------------------------
>>>> kworker/18:2/633 is trying to lock:
>>>> ffff888c474e5648 (&vmd->cfg_lock){....}-{3:3}, at: vmd_pci_write+0x185/0x2a0
>>>> other info that might help us debug this:
>>>> context-{5:5}
>>>> 4 locks held by kworker/18:2/633:
>>>>   #0: ffff888100108958 ((wq_completion)events){+.+.}-{0:0}, at: process_one_work+0xf78/0x1920
>>>>   #1: ffffc9000ae1fd90 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_one_work+0x7fe/0x1920
>>>>   #2: ffff888c483508a8 (&md->mutex){+.+.}-{4:4}, at: __pci_enable_msi_range+0x208/0x800
>>>>   #3: ffff888c48329bd8 (&dev->msi_lock){....}-{2:2}, at: pci_msi_update_mask+0x91/0x170
>>>> stack backtrace:
>>>> CPU: 18 PID: 633 Comm: kworker/18:2 Not tainted 6.10.0-rc4 #80 7c0f2526417bfbb7579e3c3442683c5961773c75
>>>> Hardware name: Lenovo ThinkSystem SR630/-[7X01RCZ000]-, BIOS IVEL60O-2.71 09/28/2020
>>>> Workqueue: events work_for_cpu_fn
>>>> Call Trace:
>>>>   <TASK>
>>>>   dump_stack_lvl+0x7c/0xc0
>>>>   __lock_acquire+0x9e5/0x1ed0
>>>>   lock_acquire+0x194/0x490
>>>>   _raw_spin_lock_irqsave+0x42/0x90
>>>>   vmd_pci_write+0x185/0x2a0
>>>>   pci_msi_update_mask+0x10c/0x170
>>>>   __pci_enable_msi_range+0x291/0x800
>>>>   pci_alloc_irq_vectors_affinity+0x13e/0x1d0
>>>>   pcie_portdrv_probe+0x570/0xe60
>>>>   local_pci_probe+0xdc/0x190
>>>>   work_for_cpu_fn+0x4e/0xa0
>>>>   process_one_work+0x86d/0x1920
>>>>   process_scheduled_works+0xd7/0x140
>>>>   worker_thread+0x3e9/0xb90
>>>>   kthread+0x2e9/0x3d0
>>>>   ret_from_fork+0x2d/0x60
>>>>   ret_from_fork_asm+0x1a/0x30
>>>>   </TASK>
>>>>
>>>> The root cause is that the dev->msi_lock is a raw spinlock, but
>>>> vmd->cfg_lock is a spinlock.
>>>
>>> Can you expand this a little bit?  This isn't enough unless one
>>> already knows the difference between raw_spinlock_t and spinlock_t,
>>> which I didn't.
>>>
>>> Documentation/locking/locktypes.rst says they are the same except when
>>> CONFIG_PREEMPT_RT is set (might be worth mentioning with the config
>>> list above?), but that with CONFIG_PREEMPT_RT, spinlock_t is based on
>>> rt_mutex.
>>>
>>> And I guess there's a rule that you can't acquire rt_mutex while
>>> holding a raw_spinlock.
>>
>> Thanks for your review and comments. Sorry for not explaining this clearly.
>> Yes, you are right, if CONFIG_PREEMPT_RT is not set, the spinlock_t is
>> based on raw_spinlock, there is no any question in the above call trace.
>>
>> But as you mentioned, if CONFIG_PREEMPT_RT is set, the spinlock_t is based
>> on rt_mutex, a task will be scheduled when waiting for rt_mutex. For example,
>> there are two threads are trying to hold a rt_mutex lock, if A hold the
>> lock firstly, and B will be scheduled in rtlock_slowlock_locked() waiting
>> for A to release the lock. The raw_spinlock is a real spinning lock, which
>> is not allowed the task of the raw_spinlock owner is scheduled in its
>> critical region. In other words, we should not try to acquire rt_mutex lock
>> in the critical region of the raw_spinlock when CONFIG_PREEMPT_RT is set.
>>
>> CONFIG_PROVE_LOCKING and CONFIG_PROVE_RAW_LOCK_NESTING options are
>> used to detect the invalid lock nesting (the raw_spinlock vs. spinlock
>> nesting checks) [1]. Here is the call path:
>>
>>    pci_msi_update_mask  ---> hold raw_spinlock dev->msi_lock
>>      pci_write_config_dword
>>       pci_bus_write_config_dword
>>         vmd_pci_write   ---> hold spinlock_t vmd->cfg_lock
>>
>> The above call path is the invalid lock nesting becuase the vmd driver
>> tries to acquire the vmd->cfg_lock spinlock within the raw_spinlock
>> region (dev->msi_lock). That's why the message "BUG: Invalid wait contex"
>> is shown.
>>
> 
> It looks like this only happens when CONFIG_PREEMPT_RT is set so I would mention that in the commit message (as Bjorn mentioned). I also think thsi level of detail is helpful and should be in the commit message as well since it's not obvious to the casual observer :)

Thanks for your suggestions and comments, I totally agree with you. 
I will add those key information into V2 patch commit message.

Thanks,
Regards,
Jiwei

> 
> Paul
> 
>> [1] https://lore.kernel.org/lkml/YBBA81osV7cHN2fb@hirez.programming.kicks-ass.net/
>>
>> Thanks,
>> Regards,
>> Jiwei
>>
>>>
>>> The dev->msi_lock was added by 77e89afc25f3 ("PCI/MSI: Protect
>>> msi_desc::masked for multi-MSI") and only used in
>>> pci_msi_update_mask():
>>>
>>>    raw_spin_lock_irqsave(lock, flags);
>>>    desc->pci.msi_mask &= ~clear;
>>>    desc->pci.msi_mask |= set;
>>>    pci_write_config_dword(msi_desc_to_pci_dev(desc), desc->pci.mask_pos,
>>>              desc->pci.msi_mask);
>>>    raw_spin_unlock_irqrestore(lock, flags);
>>>
>>> The vmd->cfg_lock was added by 185a383ada2e ("x86/PCI: Add driver for
>>> Intel Volume Management Device (VMD)") and is only used around VMD
>>> config accesses, e.g.,
>>>
>>>    * CPU may deadlock if config space is not serialized on some versions of this
>>>    * hardware, so all config space access is done under a spinlock.
>>>
>>>    static int vmd_pci_read(...)
>>>    {
>>>      spin_lock_irqsave(&vmd->cfg_lock, flags);
>>>      switch (len) {
>>>      case 1:
>>>         *value = readb(addr);
>>>         break;
>>>      case 2:
>>>         *value = readw(addr);
>>>         break;
>>>      case 4:
>>>         *value = readl(addr);
>>>         break;
>>>      default:
>>>         ret = -EINVAL;
>>>         break;
>>>      }
>>>      spin_unlock_irqrestore(&vmd->cfg_lock, flags);
>>>    }
>>>
>>> IIUC those reads turn into single PCIe MMIO reads, so I wouldn't
>>> expect any concurrency issues there that need locking.
>>>
>>> But apparently there's something weird that can deadlock the CPU.
>>>
>>>> Signed-off-by: Jiwei Sun<sunjw10@...ovo.com>
>>>> Suggested-by: Adrian Huang <ahuang12@...ovo.com>
>>>> ---
>>>>   drivers/pci/controller/vmd.c | 12 ++++++------
>>>>   1 file changed, 6 insertions(+), 6 deletions(-)
>>>>
>>>> diff --git a/drivers/pci/controller/vmd.c b/drivers/pci/controller/vmd.c
>>>> index 87b7856f375a..45d0ebf96adc 100644
>>>> --- a/drivers/pci/controller/vmd.c
>>>> +++ b/drivers/pci/controller/vmd.c
>>>> @@ -125,7 +125,7 @@ struct vmd_irq_list {
>>>>   struct vmd_dev {
>>>>       struct pci_dev        *dev;
>>>>   -    spinlock_t        cfg_lock;
>>>> +    raw_spinlock_t        cfg_lock;
>>>>       void __iomem        *cfgbar;
>>>>         int msix_count;
>>>> @@ -402,7 +402,7 @@ static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
>>>>       if (!addr)
>>>>           return -EFAULT;
>>>>   -    spin_lock_irqsave(&vmd->cfg_lock, flags);
>>>> +    raw_spin_lock_irqsave(&vmd->cfg_lock, flags);
>>>>       switch (len) {
>>>>       case 1:
>>>>           *value = readb(addr);
>>>> @@ -417,7 +417,7 @@ static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
>>>>           ret = -EINVAL;
>>>>           break;
>>>>       }
>>>> -    spin_unlock_irqrestore(&vmd->cfg_lock, flags);
>>>> +    raw_spin_unlock_irqrestore(&vmd->cfg_lock, flags);
>>>>       return ret;
>>>>   }
>>>>   @@ -437,7 +437,7 @@ static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
>>>>       if (!addr)
>>>>           return -EFAULT;
>>>>   -    spin_lock_irqsave(&vmd->cfg_lock, flags);
>>>> +    raw_spin_lock_irqsave(&vmd->cfg_lock, flags);
>>>>       switch (len) {
>>>>       case 1:
>>>>           writeb(value, addr);
>>>> @@ -455,7 +455,7 @@ static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
>>>>           ret = -EINVAL;
>>>>           break;
>>>>       }
>>>> -    spin_unlock_irqrestore(&vmd->cfg_lock, flags);
>>>> +    raw_spin_unlock_irqrestore(&vmd->cfg_lock, flags);
>>>>       return ret;
>>>>   }
>>>>   @@ -1015,7 +1015,7 @@ static int vmd_probe(struct pci_dev *dev, const struct pci_device_id *id)
>>>>       if (features & VMD_FEAT_OFFSET_FIRST_VECTOR)
>>>>           vmd->first_vec = 1;
>>>>   -    spin_lock_init(&vmd->cfg_lock);
>>>> +    raw_spin_lock_init(&vmd->cfg_lock);
>>>>       pci_set_drvdata(dev, vmd);
>>>>       err = vmd_enable_domain(vmd, features);
>>>>       if (err)
>>>> -- 
>>>> 2.27.0
>>>>
>>
> 

Powered by blists - more mailing lists