Re: [PATCH 3/3] VMCI: Fix some error handling paths in vmci_guest_probe_device()

[Vishnu Dasa <vdasa@vmware.com>]

> On Feb 11, 2022, at 12:06 PM, Christophe JAILLET <christophe.jaillet@wanadoo.fr> wrote:
> 
> Le 11/02/2022 à 15:09, Dan Carpenter a écrit :
>> On Thu, Feb 10, 2022 at 11:27:34PM +0100, Christophe JAILLET wrote:
>>> The 'err_remove_vmci_dev_g' error label is not at the right place.
>>> This could lead to un-released resource.
>>> 
>>> There is also a missing label. If pci_alloc_irq_vectors() fails, the
>>> previous vmci_event_subscribe() call must be undone.
>>> 
>>> Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
>>> ---
>>> Review with GREAT care.
>>> 
>>> This patch is a recent rebase of an old patch that has never been
>>> submitted.
>>> This function is huge and modifying its error handling path is error
>>> prone (at least for me).
>>> 
>>> The patch is compile-tested only.
>> There is still one bug.  Sorry if the line numbers are off.
> 
> Thanks for the review Dan.
> Much appreciated.

Thanks, Christophe and Dan!

> 
>> drivers/misc/vmw_vmci/vmci_guest.c
>>    705                if (capabilities & VMCI_CAPS_NOTIFICATIONS) {
>>    706                        vmci_dev->notification_bitmap = dma_alloc_coherent(
>>                                                                   ^^^^^
>> Alloc
>>    707                                &pdev->dev, PAGE_SIZE, &vmci_dev->notification_base,
>>    708                                GFP_KERNEL);
>>    709                        if (!vmci_dev->notification_bitmap) {
>>    710                                dev_warn(&pdev->dev,
>>    711                                         "Unable to allocate notification bitmap\n");
>>    712                        } else {
>>    713                                memset(vmci_dev->notification_bitmap, 0, PAGE_SIZE);
>>    714                                caps_in_use |= VMCI_CAPS_NOTIFICATIONS;
>>    715                        }
>>    716                }
>>    717
>>    718                if (mmio_base != NULL) {
>>    719                        if (capabilities & VMCI_CAPS_DMA_DATAGRAM) {
>>    720                                caps_in_use |= VMCI_CAPS_DMA_DATAGRAM;
>>    721                        } else {
>>    722                                dev_err(&pdev->dev,
>>    723                                        "Missing capability: VMCI_CAPS_DMA_DATAGRAM\n");
>>    724                                error = -ENXIO;
>>    725                                goto err_free_data_buffers;
>> This should be goto err_free_notification_bitmap;
> 
> Agreed.
> The error handling path still looked odd to me because 2 things were undone without a label between the 2 steps.
> That was it. An err_free_notification_bitmap should be added and used.
> I missed it.

Good catch.  This fixes recent code, so a separate patch would be good.
"[PATCH v3 3/8] VMCI: dma dg: detect DMA datagram capability"

> 
>>    726                        }
>>    727                }
>> On of the rules for error handling is that the unwind code should mirror
>> the allocation code but instead of that this code will have:
>> Alloc:
>> 	if (capabilities & VMCI_CAPS_NOTIFICATIONS)
>> Free:
>> 	if (vmci_dev->notification_bitmap)
>> It's the same if statement but you wouldn't really know it from just
>> looking at it so it's confusing.
> 
> This one is fine I think. If the allocation of notification_bitmap fails, it is not an error. So it looks fine to test it the way it is done.
> Or we should have both 'if'.
> 

Right.  And we would need to check 'capabilities & VMCI_CAPS_NOTIFICATIONS',
'caps_in_use & VMCI_CAPS_NOTIFICATIONS' and then
'vmci_dev->notification_bitmap' if we go that route.  I think we can leave it as is.

> 
>> Whatever...  But where this really
>> hurts is with:
>> Alloc:
>> 	if (vmci_dev->exclusive_vectors) {
>> 		error = request_irq(pci_irq_vector(pdev, 1), ...
>> Free:
>> 	free_irq(pci_irq_vector(pdev, 1), vmci_dev);
>> No if statement.  It works because it's the last allocation but it's
>> confusing and fragile.
> 
> Agreed.

Sorry, I hope I'm not missing something obvious or misunderstanding the point.
But I don't think the problem implied here exists?

If 'request_irq(pci_irq_vector(pdev, 0), ...' fails we goto err_disable_msi and there
is no free_irq in this path.  If 'request_irq(pci_irq_vector(pdev, 1), ...' fails then we
goto err_free_irq and we do 'free_irq(pci_irq_vector(pdev, 0), vmci_dev)'.  Note that
this is for the previous one without the check for vmci_dev->exclusive_vectors.

> 
>> The other question I had was:
>>    882        err_remove_bitmap:
>>    883                if (vmci_dev->notification_bitmap) {
>>    884                        vmci_write_reg(vmci_dev, VMCI_CONTROL_RESET, VMCI_CONTROL_ADDR);
>>                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
>> This doesn't mirror anything in the allocation code so who knows if its
>> done in the correct place/order.
> 
> Agreed. It puzzled me as well.
> 
> vmci_guest_remove_device() also has this kind of code, but it is not done the same way in this function. It is unconditional and not close to the dma_free_coherent() call.
> Odd.
> 
> I won't touch it by myself :)
> 
>>    885                        dma_free_coherent(&pdev->dev, PAGE_SIZE,
>>    886                                          vmci_dev->notification_bitmap,
>>    887                                          vmci_dev->notification_base);
>>    888                }
>> regards,
>> dan carpenter
> 
> All your comments are unrelated to my patch and looks like additional fixes.
> 
> Until recently, this file was mostly untouched.
> So, let see if a maintainer looks interested in these patches and if he prefers a patch that fixes everything or several patches, maybe easier to review.
> 
> Once again, big thanks.
> 
> CJ