Re: [PATCH v7 01/13] PCI/P2PDMA: Support peer-to-peer memory
From: Logan Gunthorpe
Date: Tue Sep 25 2018 - 14:09:56 EST
On 2018-09-25 11:25 a.m., Bart Van Assche wrote:
> It's great to see this patch series making progress. Unfortunately I didn't
> have the time earlier to have a closer look at this patch series. I hope that
> you don't mind that I ask a few questions about the implementation?
Thanks for the review Bart!
>> +static void pci_p2pdma_percpu_kill(void *data)
>> +{
>> + struct percpu_ref *ref = data;
>> +
>> + if (percpu_ref_is_dying(ref))
>> + return;
>> +
>> + percpu_ref_kill(ref);
>> +}
>
> The percpu_ref_is_dying() test should either be removed or a comment should be
> added above it that explains why it is necessary. Is the purpose of that call
> perhaps to protect against multiple calls of pci_p2pdma_percpu_kill()? If so,
> which mechanism serializes these multiple calls?
Hmm, yes, this was copied from device DAX, but I see it has been removed
from there since then. I'll remove it for v8.
>> +static void pci_p2pdma_release(void *data)
>> +{
>> + struct pci_dev *pdev = data;
>> +
>> + if (!pdev->p2pdma)
>> + return;
>> +
>> + wait_for_completion(&pdev->p2pdma->devmap_ref_done);
>> + percpu_ref_exit(&pdev->p2pdma->devmap_ref);
>> +
>> + gen_pool_destroy(pdev->p2pdma->pool);
>> + pdev->p2pdma = NULL;
>> +}
>
> Which code frees the memory pdev->p2pdma points at? Other functions similar to
> pci_p2pdma_release() call devm_remove_action(), e.g. hmm_devmem_ref_exit().
pdev->p2pdma is allocated using devm so it will be freed when the PCI
driver is being unwound. pci_p2pdma_release() is a devm action itself
which is registered right after the devm_kzalloc() call. Therefore the
memory will be freed in the next devm action. I don't exactly know what
hmm is doing there but we don't have similar actions to remove.
>> +static int pci_p2pdma_setup(struct pci_dev *pdev)
>> +{
>> + int error = -ENOMEM;
>> + struct pci_p2pdma *p2p;
>> +
>> + p2p = devm_kzalloc(&pdev->dev, sizeof(*p2p), GFP_KERNEL);
>> + if (!p2p)
>> + return -ENOMEM;
>> +
>> + p2p->pool = gen_pool_create(PAGE_SHIFT, dev_to_node(&pdev->dev));
>> + if (!p2p->pool)
>> + goto out;
>> +
>> + init_completion(&p2p->devmap_ref_done);
>> + error = percpu_ref_init(&p2p->devmap_ref,
>> + pci_p2pdma_percpu_release, 0, GFP_KERNEL);
>> + if (error)
>> + goto out_pool_destroy;
>> +
>> + percpu_ref_switch_to_atomic_sync(&p2p->devmap_ref);
>
> Why are percpu_ref_init() and percpu_ref_switch_to_atomic_sync() called
> separately instead of passing PERCPU_REF_INIT_ATOMIC to percpu_ref_init()?
> Would using PERCPU_REF_INIT_ATOMIC eliminate a call_rcu_sched() call and
> hence make this function faster?
I can't even remember why we are switching to atomic at all. It probably
shouldn't be there. I'll remove it for v8.
>> +static struct pci_dev *find_parent_pci_dev(struct device *dev)
>
> The above function increases the reference count of the device it returns a
> pointer to. It is a good habit to explain such behavior above the function
> definition.
Will do.
>> +static void seq_buf_print_bus_devfn(struct seq_buf *buf, struct pci_dev *pdev)
>> +{
>> + if (!buf)
>> + return;
>> +
>> + seq_buf_printf(buf, "%s;", pci_name(pdev));
>> +}
>
> NULL checks in functions that print to a seq buffer are unusual. Is it
> possible that a NULL pointer gets passed as the first argument to
> seq_buf_print_bus_devfn()?
Yes. There are two paths here one that's verbose and one that's not. In
the non-verbose case, we pass NULL instead of the seq_buf, so both calls
need to ensure the seq_buf is not NULL before trying to print to it.
>> +struct pci_p2pdma_client {
>> + struct list_head list;
>> + struct pci_dev *client;
>> + struct pci_dev *provider;
>> +};
>
> Is there a reason that the peer-to-peer client and server code exist in the
> same source file? If not, have you considered to split the p2pdma.c file into
> two files - one with the code for devices that provide p2p functionality and
> another file with the code that supports p2p users? I think that would make it
> easier to follow the code.
I see what you're saying but generally I get push back against adding
extra files. I'm going to leave it the way it is unless other people
voice their opinions in favour of the change.
>> +/**
>> + * pci_free_p2pmem - allocate peer-to-peer DMA memory
>> + * @pdev: the device the memory was allocated from
>> + * @addr: address of the memory that was allocated
>> + * @size: number of bytes that was allocated
>> + */
>> +void pci_free_p2pmem(struct pci_dev *pdev, void *addr, size_t size)
>> +{
>> + gen_pool_free(pdev->p2pdma->pool, (uintptr_t)addr, size);
>> + percpu_ref_put(&pdev->p2pdma->devmap_ref);
>> +}
>> +EXPORT_SYMBOL_GPL(pci_free_p2pmem);
>
> Please fix the header of this function - there is a copy-paste error in the
> function header.
Will do.
Logan