RE: [PATCH 2/2] PCI: lock each enable/disable num_vfs operation in sysfs

[Tantilov, Emil S]

>-----Original Message-----
>From: Gavin Shan [mailto:gwshan@xxxxxxxxxxxxxxxxxx]
>Sent: Wednesday, January 04, 2017 3:12 PM
>To: Tantilov, Emil S <emil.s.tantilov@xxxxxxxxx>
>Cc: Gavin Shan <gwshan@xxxxxxxxxxxxxxxxxx>; linux-pci@xxxxxxxxxxxxxxx;
>intel-wired-lan@xxxxxxxxxxxxxxxx; Duyck, Alexander H
><alexander.h.duyck@xxxxxxxxx>; netdev@xxxxxxxxxxxxxxx; linux-
>kernel@xxxxxxxxxxxxxxx
>Subject: Re: [PATCH 2/2] PCI: lock each enable/disable num_vfs operation in
>sysfs
>
>On Wed, Jan 04, 2017 at 04:00:20PM +0000, Tantilov, Emil S wrote:
>>>On Tue, Jan 03, 2017 at 04:48:31PM -0800, Emil Tantilov wrote:
>>>>Enabling/disabling SRIOV via sysfs by echo-ing multiple values
>>>>simultaneously:
>>>>
>>>>echo 63 > /sys/class/net/ethX/device/sriov_numvfs&
>>>>echo 63 > /sys/class/net/ethX/device/sriov_numvfs
>>>>
>>>>sleep 5
>>>>
>>>>echo 0 > /sys/class/net/ethX/device/sriov_numvfs&
>>>>echo 0 > /sys/class/net/ethX/device/sriov_numvfs
>>>>
>>>>Results in the following bug:
>>>>
>>>>kernel BUG at drivers/pci/iov.c:495!
>>>>invalid opcode: 0000 [#1] SMP
>>>>CPU: 1 PID: 8050 Comm: bash Tainted: G   W   4.9.0-rc7-net-next #2092
>>>>RIP: 0010:[<ffffffff813b1647>]
>>>>	  [<ffffffff813b1647>] pci_iov_release+0x57/0x60
>>>>
>>>>Call Trace:
>>>> [<ffffffff81391726>] pci_release_dev+0x26/0x70
>>>> [<ffffffff8155be6e>] device_release+0x3e/0xb0
>>>> [<ffffffff81365ee7>] kobject_cleanup+0x67/0x180
>>>> [<ffffffff81365d9d>] kobject_put+0x2d/0x60
>>>> [<ffffffff8155bc27>] put_device+0x17/0x20
>>>> [<ffffffff8139c08a>] pci_dev_put+0x1a/0x20
>>>> [<ffffffff8139cb6b>] pci_get_dev_by_id+0x5b/0x90
>>>> [<ffffffff8139cca5>] pci_get_subsys+0x35/0x40
>>>> [<ffffffff8139ccc8>] pci_get_device+0x18/0x20
>>>> [<ffffffff8139ccfb>] pci_get_domain_bus_and_slot+0x2b/0x60
>>>> [<ffffffff813b09e7>] pci_iov_remove_virtfn+0x57/0x180
>>>> [<ffffffff813b0b95>] pci_disable_sriov+0x65/0x140
>>>> [<ffffffffa00a1af7>] ixgbe_disable_sriov+0xc7/0x1d0 [ixgbe]
>>>> [<ffffffffa00a1e9d>] ixgbe_pci_sriov_configure+0x3d/0x170 [ixgbe]
>>>> [<ffffffff8139d28c>] sriov_numvfs_store+0xdc/0x130
>>>>...
>>>>RIP  [<ffffffff813b1647>] pci_iov_release+0x57/0x60
>>>>
>>>>Use the existing mutex lock to protect each enable/disable operation.
>>>>
>>>>CC: Alexander Duyck <alexander.h.duyck@xxxxxxxxx>
>>>>Signed-off-by: Emil Tantilov <emil.s.tantilov@xxxxxxxxx>
>>>
>>>Emil, It's going to change semantics of pci_enable_sriov() and
>pci_disable_sriov().
>>>They can be invoked when writing to the sysfs entry, or loading PF's
>>>driver. With the change applied, the lock (pf->sriov->lock) isn't acquired and released
>>>in the PF's driver loading path.
>>
>>The enablement of SRIOV on driver load is done via deprecated module parameter.
>>Perhaps we can just remove it, although there are probably still people that use it
>>and may not be happy if we get rid of it.
>>
>
>Yeah, some drivers are still using the interface. So we cannot affect it
>until it can be droped.
>
>>>I think the reasonable way would be adding a flag in "struct sriov", to
>>>indicate someone is accessing the IOV capability through sysfs file. With this, the
>>>code returns with "-EBUSY" immediately for contenders. With it, nothing is going
>>>to be changed in PF's driver loading path.
>>
>>Flag is what I initially had in mind, but did not want to add extra locking if we
>>can make use of the existing.
>>
>
>The problem is sriov->lock wasn't introduced to protect the whole IOV capability.
>Instead, it protects the allocation of virtual bus (if needed). In your patch,
>it will be used to protect the whole IOV capability, ensure accessing the
>IOV capability exclusively. So the usage of this lock is changed.
>
>     code extracted from pci.h:
>
>     struct pci_sriov {
>            :
>            struct mutex lock;      /* lock for VF bus */
>            :
>     }
>
>The lock was introduced by commit d1b054da8 ("PCI: initialize and release
>SR-IOV capability"). If I'm correct enough, I don't think this lock is needed when
>pci_enable_sriov() or pci_disable_sriov() are called in driver because of
>module
>parameters. I don't see the usage case calling pci_disable_sriov() while
>previous pci_enable_sriov() isn't finished yet. Also, it's not needed in EEH
>subsystem.
>So I think the lock can be dropped, then it can be used to protect sysfs path.

That's pretty much what this patch does, except I kept the locking for EEH since 
it is the only driver that calls pci_iov_add/remove_virtfn() directly.

I'll write it up and run some tests, although I have no way to test EEH.
 
>>>Also, there are some minor comments as below and I guess most of them won't
>>>be applied if you take my suggestion eventually. However, I'm trying to make
>>>the comments complete.
>>
>>Thanks a lot for reviewing!
>>
>>>
>>>>---
>>>> drivers/pci/pci-sysfs.c |   24 +++++++++++++++++-------
>>>> 1 file changed, 17 insertions(+), 7 deletions(-)
>>>>
>>>>diff --git a/drivers/pci/pci-sysfs.c b/drivers/pci/pci-sysfs.c
>>>>index 0666287..5b54cf5 100644
>>>>--- a/drivers/pci/pci-sysfs.c
>>>>+++ b/drivers/pci/pci-sysfs.c
>>>>@@ -472,7 +472,9 @@ static ssize_t sriov_numvfs_store(struct device
>*dev,
>>>> 				  const char *buf, size_t count)
>>>> {
>>>> 	struct pci_dev *pdev = to_pci_dev(dev);
>>>>+	struct pci_sriov *iov = pdev->sriov;
>>>> 	int ret;
>>>>+
>>>
>>>Unnecessary change.
>>>
>>>> 	u16 num_vfs;
>>>>
>>>> 	ret = kstrtou16(buf, 0, &num_vfs);
>>>>@@ -482,38 +484,46 @@ static ssize_t sriov_numvfs_store(struct device
>>>*dev,
>>>> 	if (num_vfs > pci_sriov_get_totalvfs(pdev))
>>>> 		return -ERANGE;
>>>>
>>>>+	mutex_lock(&iov->dev->sriov->lock);
>>>>+
>>>> 	if (num_vfs == pdev->sriov->num_VFs)
>>>>-		return count;		/* no change */
>>>>+		goto exit;
>>>>
>>>> 	/* is PF driver loaded w/callback */
>>>> 	if (!pdev->driver || !pdev->driver->sriov_configure) {
>>>> 		dev_info(&pdev->dev, "Driver doesn't support SRIOV
>>>configuration via sysfs\n");
>>>>-		return -ENOSYS;
>>>>+		ret = -EINVAL;
>>>>+		goto exit;
>>>
>>>Why we need change the error code here?
>>
>>checkpatch was complaining about the use of the ENOSYS error code being specific
>>and even though it was not my patch introducing it I had to change it to shut it up.
>>
>
>Right, it's reserved for attempt to call nonexisting syscall, but I think
>ENOENT might be more indicative than EINVAL in this specific case?

ENOENT is for a missing file, but if we got this far in the code then there must've been 
a sysfs file. This is pretty straightforward "not supported" error, which is why I picked
EINVAL.

Thanks,
Emil