Re: [PATCH v7 01/13] PCI/P2PDMA: Support peer-to-peer memory

[Bart Van Assche]

On Tue, 2018-09-25 at 10:22 -0600, Logan Gunthorpe wrote:
+AD4 +AFs ... +AF0

Hi Logan,

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?

+AD4 +-static void pci+AF8-p2pdma+AF8-percpu+AF8-kill(void +ACo-data)
+AD4 +-+AHs
+AD4 +-	struct percpu+AF8-ref +ACo-ref +AD0 data+ADs
+AD4 +-
+AD4 +-	if (percpu+AF8-ref+AF8-is+AF8-dying(ref))
+AD4 +-		return+ADs
+AD4 +-
+AD4 +-	percpu+AF8-ref+AF8-kill(ref)+ADs
+AD4 +-+AH0

The percpu+AF8-ref+AF8-is+AF8-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+AF8-p2pdma+AF8-percpu+AF8-kill()? If so,
which mechanism serializes these multiple calls?

+AD4 +-static void pci+AF8-p2pdma+AF8-release(void +ACo-data)
+AD4 +-+AHs
+AD4 +-	struct pci+AF8-dev +ACo-pdev +AD0 data+ADs
+AD4 +-
+AD4 +-	if (+ACE-pdev-+AD4-p2pdma)
+AD4 +-		return+ADs
+AD4 +-
+AD4 +-	wait+AF8-for+AF8-completion(+ACY-pdev-+AD4-p2pdma-+AD4-devmap+AF8-ref+AF8-done)+ADs
+AD4 +-	percpu+AF8-ref+AF8-exit(+ACY-pdev-+AD4-p2pdma-+AD4-devmap+AF8-ref)+ADs
+AD4 +-
+AD4 +-	gen+AF8-pool+AF8-destroy(pdev-+AD4-p2pdma-+AD4-pool)+ADs
+AD4 +-	pdev-+AD4-p2pdma +AD0 NULL+ADs
+AD4 +-+AH0

Which code frees the memory pdev-+AD4-p2pdma points at? Other functions similar to
pci+AF8-p2pdma+AF8-release() call devm+AF8-remove+AF8-action(), e.g. hmm+AF8-devmem+AF8-ref+AF8-exit().

+AD4 +-static int pci+AF8-p2pdma+AF8-setup(struct pci+AF8-dev +ACo-pdev)
+AD4 +-+AHs
+AD4 +-	int error +AD0 -ENOMEM+ADs
+AD4 +-	struct pci+AF8-p2pdma +ACo-p2p+ADs
+AD4 +-
+AD4 +-	p2p +AD0 devm+AF8-kzalloc(+ACY-pdev-+AD4-dev, sizeof(+ACo-p2p), GFP+AF8-KERNEL)+ADs
+AD4 +-	if (+ACE-p2p)
+AD4 +-		return -ENOMEM+ADs
+AD4 +-
+AD4 +-	p2p-+AD4-pool +AD0 gen+AF8-pool+AF8-create(PAGE+AF8-SHIFT, dev+AF8-to+AF8-node(+ACY-pdev-+AD4-dev))+ADs
+AD4 +-	if (+ACE-p2p-+AD4-pool)
+AD4 +-		goto out+ADs
+AD4 +-
+AD4 +-	init+AF8-completion(+ACY-p2p-+AD4-devmap+AF8-ref+AF8-done)+ADs
+AD4 +-	error +AD0 percpu+AF8-ref+AF8-init(+ACY-p2p-+AD4-devmap+AF8-ref,
+AD4 +-			pci+AF8-p2pdma+AF8-percpu+AF8-release, 0, GFP+AF8-KERNEL)+ADs
+AD4 +-	if (error)
+AD4 +-		goto out+AF8-pool+AF8-destroy+ADs
+AD4 +-
+AD4 +-	percpu+AF8-ref+AF8-switch+AF8-to+AF8-atomic+AF8-sync(+ACY-p2p-+AD4-devmap+AF8-ref)+ADs

Why are percpu+AF8-ref+AF8-init() and percpu+AF8-ref+AF8-switch+AF8-to+AF8-atomic+AF8-sync() called
separately instead of passing PERCPU+AF8-REF+AF8-INIT+AF8-ATOMIC to percpu+AF8-ref+AF8-init()?
Would using PERCPU+AF8-REF+AF8-INIT+AF8-ATOMIC eliminate a call+AF8-rcu+AF8-sched() call and
hence make this function faster?

+AD4 +-static struct pci+AF8-dev +ACo-find+AF8-parent+AF8-pci+AF8-dev(struct device +ACo-dev)
+AD4 +-+AHs
+AD4 +-	struct device +ACo-parent+ADs
+AD4 +-
+AD4 +-	dev +AD0 get+AF8-device(dev)+ADs
+AD4 +-
+AD4 +-	while (dev) +AHs
+AD4 +-		if (dev+AF8-is+AF8-pci(dev))
+AD4 +-			return to+AF8-pci+AF8-dev(dev)+ADs
+AD4 +-
+AD4 +-		parent +AD0 get+AF8-device(dev-+AD4-parent)+ADs
+AD4 +-		put+AF8-device(dev)+ADs
+AD4 +-		dev +AD0 parent+ADs
+AD4 +-	+AH0
+AD4 +-
+AD4 +-	return NULL+ADs
+AD4 +-+AH0

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.

+AD4 +-static void seq+AF8-buf+AF8-print+AF8-bus+AF8-devfn(struct seq+AF8-buf +ACo-buf, struct pci+AF8-dev +ACo-pdev)
+AD4 +-+AHs
+AD4 +-	if (+ACE-buf)
+AD4 +-		return+ADs
+AD4 +-
+AD4 +-	seq+AF8-buf+AF8-printf(buf, +ACIAJQ-s+ADsAIg, pci+AF8-name(pdev))+ADs
+AD4 +-+AH0

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+AF8-buf+AF8-print+AF8-bus+AF8-devfn()?

+AD4 +-struct pci+AF8-p2pdma+AF8-client +AHs
+AD4 +-	struct list+AF8-head list+ADs
+AD4 +-	struct pci+AF8-dev +ACo-client+ADs
+AD4 +-	struct pci+AF8-dev +ACo-provider+ADs
+AD4 +-+AH0AOw

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.

+AD4 +-/+ACoAKg
+AD4 +- +ACo pci+AF8-free+AF8-p2pmem - allocate peer-to-peer DMA memory
+AD4 +- +ACo +AEA-pdev: the device the memory was allocated from
+AD4 +- +ACo +AEA-addr: address of the memory that was allocated
+AD4 +- +ACo +AEA-size: number of bytes that was allocated
+AD4 +- +ACo-/
+AD4 +-void pci+AF8-free+AF8-p2pmem(struct pci+AF8-dev +ACo-pdev, void +ACo-addr, size+AF8-t size)
+AD4 +-+AHs
+AD4 +-	gen+AF8-pool+AF8-free(pdev-+AD4-p2pdma-+AD4-pool, (uintptr+AF8-t)addr, size)+ADs
+AD4 +-	percpu+AF8-ref+AF8-put(+ACY-pdev-+AD4-p2pdma-+AD4-devmap+AF8-ref)+ADs
+AD4 +-+AH0
+AD4 +-EXPORT+AF8-SYMBOL+AF8-GPL(pci+AF8-free+AF8-p2pmem)+ADs

Please fix the header of this function - there is a copy-paste error in the
function header.

Thanks,

Bart.