Re: [PATCH 02/16] PCI/P2PDMA: Avoid pci_get_slot() which sleeps

[Logan Gunthorpe]

On 2021-05-01 11:35 p.m., John Hubbard wrote:
> On 4/8/21 10:01 AM, Logan Gunthorpe wrote:
>> In order to use upstream_bridge_distance_warn() from a dma_map function,
>> it must not sleep. However, pci_get_slot() takes the pci_bus_sem so it
>> might sleep.
>>
>> In order to avoid this, try to get the host bridge's device from
>> bus->self, and if that is not set, just get the first element in the
>> device list. It should be impossible for the host bridge's device to
>> go away while references are held on child devices, so the first element
>> should not be able to change and, thus, this should be safe.
>>
>> Signed-off-by: Logan Gunthorpe <logang@xxxxxxxxxxxx>
>> ---
>>   drivers/pci/p2pdma.c | 14 ++++++++++++--
>>   1 file changed, 12 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
>> index bd89437faf06..473a08940fbc 100644
>> --- a/drivers/pci/p2pdma.c
>> +++ b/drivers/pci/p2pdma.c
>> @@ -311,16 +311,26 @@ static const struct pci_p2pdma_whitelist_entry {
>>   static bool __host_bridge_whitelist(struct pci_host_bridge *host,
>>   				    bool same_host_bridge)
>>   {
>> -	struct pci_dev *root = pci_get_slot(host->bus, PCI_DEVFN(0, 0));
>>   	const struct pci_p2pdma_whitelist_entry *entry;
>> +	struct pci_dev *root = host->bus->self;
>>   	unsigned short vendor, device;
>>   
>> +	/*
>> +	 * This makes the assumption that the first device on the bus is the
>> +	 * bridge itself and it has the devfn of 00.0. This assumption should
>> +	 * hold for the devices in the white list above, and if there are cases
>> +	 * where this isn't true they will have to be dealt with when such a
>> +	 * case is added to the whitelist.
> 
> Actually, it makes the assumption that the first device *in the list*
> (the host->bus-devices list) is 00.0.  The previous code made the
> assumption that you wrote.

The comment notes two assumptions (although the grammar is poor, which I
will fix). Yes, the previous code made the second assumption, the new
code makes both assumptions.

> By the way, pre-existing code comment: pci_p2pdma_whitelist[] seems
> really short. From a naive point of view, I'd expect that there must be
> a lot more CPUs/chipsets that can do pci p2p, what do you think? I
> wonder if we have to be so super strict, anyway. It just seems extremely
> limited, and I suspect there will be some additions to the list as soon
> as we start to use this.

Yes, well unfortunately we have no other way to determine what host
bridges can communicate with P2P. We settled on a whitelist when the
series was first patch. Nobody likes that situation, but nobody has
found anything better. We've been hoping standards bodies would give us
a flag but I haven't heard anything about that. At least AMD has been
able to guarantee us that all CPUs newer than Zen will support so that
covers a large swath. It would be nice if we could say something similar
for Intel.

> OK, yes this avoids taking the pci_bus_sem, but it's kind of cheating.
> Why is it OK to avoid taking any locks in order to retrieve the
> first entry from the list, but in order to retrieve any other entry, you
> have to aquire the pci_bus_sem, and get a reference as well? Something
> is inconsistent there.
> 
> The new version here also no longer takes a reference on the device,
> which is also cheating. But I'm guessing that the unstated assumption
> here is that there is always at least one entry in the list. But if
> that's true, then it's better to show clearly that assumption, instead
> of hiding it in an implicit call that skips both locking and reference
> counting.

Because we hold a reference to a child device of the bus. So the host
bus device can't go away until the child device has been released. An
earlier version of the P2PDMA patchset had a lot more extraneous get
device calls until someone else pointed this out.

> You could add a new function, which is a cut-down version of pci_get_slot(),
> like this, and call this from __host_bridge_whitelist():
> 
> /*
>   * A special purpose variant of pci_get_slot() that doesn't take the pci_bus_sem
>   * lock, and only looks for the 00.0 bus-device-function. Once the PCI bus is
>   * up, it is safe to call this, because there will always be a top-level PCI
>   * root device.
>   *
>   * Other assumptions: the root device is the first device in the list, and the
>   * root device is numbered 00.0.
>   */
> struct pci_dev *pci_get_root_slot(struct pci_bus *bus)
> {
> 	struct pci_dev *root;
> 	unsigned devfn = PCI_DEVFN(0, 0);
> 
> 	root = list_first_entry_or_null(&bus->devices, struct pci_dev,
> 					bus_list);
> 	if (root->devfn == devfn)
> 		goto out;
> 
> 	root = NULL;
>   out:
> 	pci_dev_get(root);
> 	return root;
> }
> EXPORT_SYMBOL(pci_get_root_slot);
> 
> ...I think that's a lot clearer to the reader, about what's going on here.

Per above, I think the reference count is unnecessary. But I could wrap
it in a static function for clarity. (There's no reason to export this
function).

> Note that I'm not really sure if it *is* safe, I would need to ask other
> PCIe subsystem developers with more experience. But I don't think anyone
> is trying to make p2pdma calls so early that PCIe buses are uninitialized.

Yeah, it's impossible to make a p2pdma call before the PCIe bus is
initialized. They have to have access to at least one PCI device before
they can even attempt it.

Logan