The PCI specs say that peer-to-peer DMA should be supported between any two
devices that have a common upstream PCI-to-PCI bridge. But devices under
different Root Ports don't share a common upstream bridge, and PCIe r4.0,
sec 1.3.1, says routing peer-to-peer transactions between Root Ports in the
same Root Complex is optional.
Many Root Complexes, including AMD ZEN, *do* support peer-to-peer DMA even
between Root Ports. Add a whitelist and allow peer-to-peer DMA if both
participants are attached to a Root Complex known to support it.
Link: https://lore.kernel.org/linux-pci/20190418115859.2394-1-christian.koenig@amd.com
Signed-off-by: Christian König <christian.koenig@amd.com>
[bhelgaas: changelog]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Logan Gunthorpe <logang@deltatee.com>
At Maintainer Summit, Greg brought up a topic I proposed around
EXPORT_SYMBOL_GPL usage. The motivation was considerations for when
EXPORT_SYMBOL_GPL is warranted and the criteria for taking the exceptional
step of reclassifying an existing export. Specifically, I wanted to make
the case that although the line is fuzzy and hard to specify in abstract
terms, it is nonetheless clear that devm_memremap_pages() and HMM
(Heterogeneous Memory Management) have crossed it. The
devm_memremap_pages() facility should have been EXPORT_SYMBOL_GPL from the
beginning, and HMM as a derivative of that functionality should have
naturally picked up that designation as well.
Contrary to typical rules, the HMM infrastructure was merged upstream with
zero in-tree consumers. There was a promise at the time that those users
would be merged "soon", but it has been over a year with no drivers
arriving. While the Nouveau driver is about to belatedly make good on
that promise it is clear that HMM was targeted first and foremost at an
out-of-tree consumer.
HMM is derived from devm_memremap_pages(), a facility Christoph and I
spearheaded to support persistent memory. It combines a device lifetime
model with a dynamically created 'struct page' / memmap array for any
physical address range. It enables coordination and control of the many
code paths in the kernel built to interact with memory via 'struct page'
objects. With HMM the integration goes even deeper by allowing device
drivers to hook and manipulate page fault and page free events.
One interpretation of when EXPORT_SYMBOL is suitable is when it is
exporting stable and generic leaf functionality. The
devm_memremap_pages() facility continues to see expanding use cases,
peer-to-peer DMA being the most recent, with no clear end date when it
will stop attracting reworks and semantic changes. It is not suitable to
export devm_memremap_pages() as a stable 3rd party driver API due to the
fact that it is still changing and manipulates core behavior. Moreover,
it is not in the best interest of the long term development of the core
memory management subsystem to permit any external driver to effectively
define its own system-wide memory management policies with no
encouragement to engage with upstream.
I am also concerned that HMM was designed in a way to minimize further
engagement with the core-MM. That, with these hooks in place,
device-drivers are free to implement their own policies without much
consideration for whether and how the core-MM could grow to meet that
need. Going forward not only should HMM be EXPORT_SYMBOL_GPL, but the
core-MM should be allowed the opportunity and stimulus to change and
address these new use cases as first class functionality.
Original changelog:
hmm_devmem_add(), and hmm_devmem_add_resource() duplicated
devm_memremap_pages() and are now simple now wrappers around the core
facility to inject a dev_pagemap instance into the global pgmap_radix and
hook page-idle events. The devm_memremap_pages() interface is base
infrastructure for HMM. HMM has more and deeper ties into the kernel
memory management implementation than base ZONE_DEVICE which is itself a
EXPORT_SYMBOL_GPL facility.
Originally, the HMM page structure creation routines copied the
devm_memremap_pages() code and reused ZONE_DEVICE. A cleanup to unify the
implementations was discussed during the initial review:
http://lkml.iu.edu/hypermail/linux/kernel/1701.2/00812.html Recent work to
extend devm_memremap_pages() for the peer-to-peer-DMA facility enabled
this cleanup to move forward.
In addition to the integration with devm_memremap_pages() HMM depends on
other GPL-only symbols:
mmu_notifier_unregister_no_release
percpu_ref
region_intersects
__class_create
It goes further to consume / indirectly expose functionality that is not
exported to any other driver:
alloc_pages_vma
walk_page_range
HMM is derived from devm_memremap_pages(), and extends deep core-kernel
fundamentals. Similar to devm_memremap_pages(), mark its entry points
EXPORT_SYMBOL_GPL().
[logang@deltatee.com: PCI/P2PDMA: match interface changes to devm_memremap_pages()]
Link: http://lkml.kernel.org/r/20181130225911.2900-1-logang@deltatee.com
Link: http://lkml.kernel.org/r/154275560565.76910.15919297436557795278.stgit@dwillia2-desk3.amr.corp.intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Logan Gunthorpe <logang@deltatee.com>
Cc: "Jérôme Glisse" <jglisse@redhat.com>
Cc: Balbir Singh <bsingharora@gmail.com>,
Cc: Michal Hocko <mhocko@suse.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Fix typos, spellos, and grammar in p2pdma.rst and p2pdma.c.
Fix return value(s) in function pci_p2pmem_alloc_sgl().
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Acked-by: Logan Gunthorpe <logang@deltatee.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Users of the P2PDMA infrastructure will typically need a way for the user
to tell the kernel to use P2P resources. Typically this will be a simple
on/off boolean operation but sometimes it may be desirable for the user to
specify the exact device to use for the P2P operation.
Add new helpers for attributes which take a boolean or a PCI device. Any
boolean as accepted by strtobool() turn P2P on or off (such as 'y', 'n',
'1', '0', etc). Specifying a full PCI device name/BDF will select the
specific device.
Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
The DMA address used when mapping PCI P2P memory must be the PCI bus
address. Thus, introduce pci_p2pmem_map_sg() to map the correct addresses
when using P2P memory.
Memory mapped in this way does not need to be unmapped and thus if we
provided pci_p2pmem_unmap_sg() it would be empty. This breaks the expected
balance between map/unmap but was left out as an empty function doesn't
really provide any benefit. In the future, if this call becomes necessary
it can be added without much difficulty.
For this, we assume that an SGL passed to these functions contain all P2P
memory or no P2P memory.
Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Add a sysfs group to display statistics about P2P memory that is registered
in each PCI device.
Attributes in the group display the total amount of P2P memory, the amount
available and whether it is published or not.
Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Some PCI devices may have memory mapped in a BAR space that's intended for
use in peer-to-peer transactions. To enable such transactions the memory
must be registered with ZONE_DEVICE pages so it can be used by DMA
interfaces in existing drivers.
Add an interface for other subsystems to find and allocate chunks of P2P
memory as necessary to facilitate transfers between two PCI peers:
struct pci_dev *pci_p2pmem_find[_many]();
int pci_p2pdma_distance[_many]();
void *pci_alloc_p2pmem();
The new interface requires a driver to collect a list of client devices
involved in the transaction then call pci_p2pmem_find() to obtain any
suitable P2P memory. Alternatively, if the caller knows a device which
provides P2P memory, they can use pci_p2pdma_distance() to determine if it
is usable. With a suitable p2pmem device, memory can then be allocated
with pci_alloc_p2pmem() for use in DMA transactions.
Depending on hardware, using peer-to-peer memory may reduce the bandwidth
of the transfer but can significantly reduce pressure on system memory.
This may be desirable in many cases: for example a system could be designed
with a small CPU connected to a PCIe switch by a small number of lanes
which would maximize the number of lanes available to connect to NVMe
devices.
The code is designed to only utilize the p2pmem device if all the devices
involved in a transfer are behind the same PCI bridge. This is because we
have no way of knowing whether peer-to-peer routing between PCIe Root Ports
is supported (PCIe r4.0, sec 1.3.1). Additionally, the benefits of P2P
transfers that go through the RC is limited to only reducing DRAM usage
and, in some cases, coding convenience. The PCI-SIG may be exploring
adding a new capability bit to advertise whether this is possible for
future hardware.
This commit includes significant rework and feedback from Christoph
Hellwig.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
[bhelgaas: fold in fix from Keith Busch <keith.busch@intel.com>:
https://lore.kernel.org/linux-pci/20181012155920.15418-1-keith.busch@intel.com,
to address comment from Dan Carpenter <dan.carpenter@oracle.com>, fold in
https://lore.kernel.org/linux-pci/20181017160510.17926-1-logang@deltatee.com]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Christian König