linux/Documentation/ntb.txt
Linus Torvalds 486088bc46 This series converts a number of top-level documents to the RST format
without incorporating them into the Sphinx tree.  The hope is to bring some
 uniformity to kernel documentation and, perhaps more importantly, have our
 existing docs serve as an example of the desired formatting for those that
 will be added later.
 -----BEGIN PGP SIGNATURE-----
 
 iQJDBAABCAAtFiEEUOvtSCFqLxY/7px3jc41VjAi5XoFAllpIuQPHGNvcmJldEBs
 d24ubmV0AAoJEI3ONVYwIuV66VIP/1GNACrp606m7HAvfskGyt26RB7hwM3+5ct2
 4LDwBW8l1IZ5efgMFsEur9QjMPexiDCkzYqNj4B8vjQLk2dVl5w+eMJ4KY5hRGHe
 dfZtyTrbzh9gQtPOufNXkTiFJBQXlbrcoLb9JK0S8OdC9LM6yN5wNwTHddmoieiC
 +cmSt5iPf2/olNqZjtjAK/ZS+AHMb3C+Um/tMamiV5uMXmCWA9Xsz1HWZKbB3WVe
 X7zLiyt5BVAOLXKir8fCUNMh0qTkq4K+mDVC7TPnHae+fAudjA+sZycuJr2ebNtR
 JK15wWArUaQj7G90eB8VM1ZCN8eyYaUxXpHGw5+2A5p7gXvytfiF2VahJfJLLSjK
 gM7ZZcwynYe7/tzqf2Bmz7uqnAqm9cSMdxplGPdoqq0RUCE+nFDoslbcGw+0YS+c
 OM8FxCMNHYdjtX18pKQZ4H2KJ2qy5FqA0tyRnlWDj6+pNPQeAQoQ2zUEq0w/qSk6
 ArQAxaNhuhPjhaqL8qhyd3MnTAncWAIbuW6FBd5RbyYmGz/jJ0srxRwnrJEZzFc9
 LaDUSDPZ8frtLd8nVNHgblD2yBWAwRMcFbwt5oZS2yIYPg6VBdHyCosZBDTsmZP0
 7IIeW5trsagK2ktp2KooPZ5cOgeXzLmUnEDPEoMoJ6yPSsHvDDzpItJLot8HZrH2
 ahaOgM3e
 =HvQS
 -----END PGP SIGNATURE-----

Merge tag 'standardize-docs' of git://git.lwn.net/linux

Pull documentation format standardization from Jonathan Corbet:
 "This series converts a number of top-level documents to the RST format
  without incorporating them into the Sphinx tree. The hope is to bring
  some uniformity to kernel documentation and, perhaps more importantly,
  have our existing docs serve as an example of the desired formatting
  for those that will be added later.

  Mauro has gone through and fixed up a lot of top-level documentation
  files to make them conform to the RST format, but without moving or
  renaming them in any way. This will help when we incorporate the ones
  we want to keep into the Sphinx doctree, but the real purpose is to
  bring a bit of uniformity to our documentation and let the top-level
  docs serve as examples for those writing new ones"

* tag 'standardize-docs' of git://git.lwn.net/linux: (84 commits)
  docs: kprobes.txt: Fix whitespacing
  tee.txt: standardize document format
  cgroup-v2.txt: standardize document format
  dell_rbu.txt: standardize document format
  zorro.txt: standardize document format
  xz.txt: standardize document format
  xillybus.txt: standardize document format
  vfio.txt: standardize document format
  vfio-mediated-device.txt: standardize document format
  unaligned-memory-access.txt: standardize document format
  this_cpu_ops.txt: standardize document format
  svga.txt: standardize document format
  static-keys.txt: standardize document format
  smsc_ece1099.txt: standardize document format
  SM501.txt: standardize document format
  siphash.txt: standardize document format
  sgi-ioc4.txt: standardize document format
  SAK.txt: standardize document format
  rpmsg.txt: standardize document format
  robust-futexes.txt: standardize document format
  ...
2017-07-15 12:58:58 -07:00

231 lines
11 KiB
Plaintext

===========
NTB Drivers
===========
NTB (Non-Transparent Bridge) is a type of PCI-Express bridge chip that connects
the separate memory systems of two or more computers to the same PCI-Express
fabric. Existing NTB hardware supports a common feature set: doorbell
registers and memory translation windows, as well as non common features like
scratchpad and message registers. Scratchpad registers are read-and-writable
registers that are accessible from either side of the device, so that peers can
exchange a small amount of information at a fixed address. Message registers can
be utilized for the same purpose. Additionally they are provided with with
special status bits to make sure the information isn't rewritten by another
peer. Doorbell registers provide a way for peers to send interrupt events.
Memory windows allow translated read and write access to the peer memory.
NTB Core Driver (ntb)
=====================
The NTB core driver defines an api wrapping the common feature set, and allows
clients interested in NTB features to discover NTB the devices supported by
hardware drivers. The term "client" is used here to mean an upper layer
component making use of the NTB api. The term "driver," or "hardware driver,"
is used here to mean a driver for a specific vendor and model of NTB hardware.
NTB Client Drivers
==================
NTB client drivers should register with the NTB core driver. After
registering, the client probe and remove functions will be called appropriately
as ntb hardware, or hardware drivers, are inserted and removed. The
registration uses the Linux Device framework, so it should feel familiar to
anyone who has written a pci driver.
NTB Typical client driver implementation
----------------------------------------
Primary purpose of NTB is to share some peace of memory between at least two
systems. So the NTB device features like Scratchpad/Message registers are
mainly used to perform the proper memory window initialization. Typically
there are two types of memory window interfaces supported by the NTB API:
inbound translation configured on the local ntb port and outbound translation
configured by the peer, on the peer ntb port. The first type is
depicted on the next figure
Inbound translation:
Memory: Local NTB Port: Peer NTB Port: Peer MMIO:
____________
| dma-mapped |-ntb_mw_set_trans(addr) |
| memory | _v____________ | ______________
| (addr) |<======| MW xlat addr |<====| MW base addr |<== memory-mapped IO
|------------| |--------------| | |--------------|
So typical scenario of the first type memory window initialization looks:
1) allocate a memory region, 2) put translated address to NTB config,
3) somehow notify a peer device of performed initialization, 4) peer device
maps corresponding outbound memory window so to have access to the shared
memory region.
The second type of interface, that implies the shared windows being
initialized by a peer device, is depicted on the figure:
Outbound translation:
Memory: Local NTB Port: Peer NTB Port: Peer MMIO:
____________ ______________
| dma-mapped | | | MW base addr |<== memory-mapped IO
| memory | | |--------------|
| (addr) |<===================| MW xlat addr |<-ntb_peer_mw_set_trans(addr)
|------------| | |--------------|
Typical scenario of the second type interface initialization would be:
1) allocate a memory region, 2) somehow deliver a translated address to a peer
device, 3) peer puts the translated address to NTB config, 4) peer device maps
outbound memory window so to have access to the shared memory region.
As one can see the described scenarios can be combined in one portable
algorithm.
Local device:
1) Allocate memory for a shared window
2) Initialize memory window by translated address of the allocated region
(it may fail if local memory window initialization is unsupported)
3) Send the translated address and memory window index to a peer device
Peer device:
1) Initialize memory window with retrieved address of the allocated
by another device memory region (it may fail if peer memory window
initialization is unsupported)
2) Map outbound memory window
In accordance with this scenario, the NTB Memory Window API can be used as
follows:
Local device:
1) ntb_mw_count(pidx) - retrieve number of memory ranges, which can
be allocated for memory windows between local device and peer device
of port with specified index.
2) ntb_get_align(pidx, midx) - retrieve parameters restricting the
shared memory region alignment and size. Then memory can be properly
allocated.
3) Allocate physically contiguous memory region in compliance with
restrictions retrieved in 2).
4) ntb_mw_set_trans(pidx, midx) - try to set translation address of
the memory window with specified index for the defined peer device
(it may fail if local translated address setting is not supported)
5) Send translated base address (usually together with memory window
number) to the peer device using, for instance, scratchpad or message
registers.
Peer device:
1) ntb_peer_mw_set_trans(pidx, midx) - try to set received from other
device (related to pidx) translated address for specified memory
window. It may fail if retrieved address, for instance, exceeds
maximum possible address or isn't properly aligned.
2) ntb_peer_mw_get_addr(widx) - retrieve MMIO address to map the memory
window so to have an access to the shared memory.
Also it is worth to note, that method ntb_mw_count(pidx) should return the
same value as ntb_peer_mw_count() on the peer with port index - pidx.
NTB Transport Client (ntb\_transport) and NTB Netdev (ntb\_netdev)
------------------------------------------------------------------
The primary client for NTB is the Transport client, used in tandem with NTB
Netdev. These drivers function together to create a logical link to the peer,
across the ntb, to exchange packets of network data. The Transport client
establishes a logical link to the peer, and creates queue pairs to exchange
messages and data. The NTB Netdev then creates an ethernet device using a
Transport queue pair. Network data is copied between socket buffers and the
Transport queue pair buffer. The Transport client may be used for other things
besides Netdev, however no other applications have yet been written.
NTB Ping Pong Test Client (ntb\_pingpong)
-----------------------------------------
The Ping Pong test client serves as a demonstration to exercise the doorbell
and scratchpad registers of NTB hardware, and as an example simple NTB client.
Ping Pong enables the link when started, waits for the NTB link to come up, and
then proceeds to read and write the doorbell scratchpad registers of the NTB.
The peers interrupt each other using a bit mask of doorbell bits, which is
shifted by one in each round, to test the behavior of multiple doorbell bits
and interrupt vectors. The Ping Pong driver also reads the first local
scratchpad, and writes the value plus one to the first peer scratchpad, each
round before writing the peer doorbell register.
Module Parameters:
* unsafe - Some hardware has known issues with scratchpad and doorbell
registers. By default, Ping Pong will not attempt to exercise such
hardware. You may override this behavior at your own risk by setting
unsafe=1.
* delay\_ms - Specify the delay between receiving a doorbell
interrupt event and setting the peer doorbell register for the next
round.
* init\_db - Specify the doorbell bits to start new series of rounds. A new
series begins once all the doorbell bits have been shifted out of
range.
* dyndbg - It is suggested to specify dyndbg=+p when loading this module, and
then to observe debugging output on the console.
NTB Tool Test Client (ntb\_tool)
--------------------------------
The Tool test client serves for debugging, primarily, ntb hardware and drivers.
The Tool provides access through debugfs for reading, setting, and clearing the
NTB doorbell, and reading and writing scratchpads.
The Tool does not currently have any module parameters.
Debugfs Files:
* *debugfs*/ntb\_tool/*hw*/
A directory in debugfs will be created for each
NTB device probed by the tool. This directory is shortened to *hw*
below.
* *hw*/db
This file is used to read, set, and clear the local doorbell. Not
all operations may be supported by all hardware. To read the doorbell,
read the file. To set the doorbell, write `s` followed by the bits to
set (eg: `echo 's 0x0101' > db`). To clear the doorbell, write `c`
followed by the bits to clear.
* *hw*/mask
This file is used to read, set, and clear the local doorbell mask.
See *db* for details.
* *hw*/peer\_db
This file is used to read, set, and clear the peer doorbell.
See *db* for details.
* *hw*/peer\_mask
This file is used to read, set, and clear the peer doorbell
mask. See *db* for details.
* *hw*/spad
This file is used to read and write local scratchpads. To read
the values of all scratchpads, read the file. To write values, write a
series of pairs of scratchpad number and value
(eg: `echo '4 0x123 7 0xabc' > spad`
# to set scratchpads `4` and `7` to `0x123` and `0xabc`, respectively).
* *hw*/peer\_spad
This file is used to read and write peer scratchpads. See
*spad* for details.
NTB Hardware Drivers
====================
NTB hardware drivers should register devices with the NTB core driver. After
registering, clients probe and remove functions will be called.
NTB Intel Hardware Driver (ntb\_hw\_intel)
------------------------------------------
The Intel hardware driver supports NTB on Xeon and Atom CPUs.
Module Parameters:
* b2b\_mw\_idx
If the peer ntb is to be accessed via a memory window, then use
this memory window to access the peer ntb. A value of zero or positive
starts from the first mw idx, and a negative value starts from the last
mw idx. Both sides MUST set the same value here! The default value is
`-1`.
* b2b\_mw\_share
If the peer ntb is to be accessed via a memory window, and if
the memory window is large enough, still allow the client to use the
second half of the memory window for address translation to the peer.
* xeon\_b2b\_usd\_bar2\_addr64
If using B2B topology on Xeon hardware, use
this 64 bit address on the bus between the NTB devices for the window
at BAR2, on the upstream side of the link.
* xeon\_b2b\_usd\_bar4\_addr64 - See *xeon\_b2b\_bar2\_addr64*.
* xeon\_b2b\_usd\_bar4\_addr32 - See *xeon\_b2b\_bar2\_addr64*.
* xeon\_b2b\_usd\_bar5\_addr32 - See *xeon\_b2b\_bar2\_addr64*.
* xeon\_b2b\_dsd\_bar2\_addr64 - See *xeon\_b2b\_bar2\_addr64*.
* xeon\_b2b\_dsd\_bar4\_addr64 - See *xeon\_b2b\_bar2\_addr64*.
* xeon\_b2b\_dsd\_bar4\_addr32 - See *xeon\_b2b\_bar2\_addr64*.
* xeon\_b2b\_dsd\_bar5\_addr32 - See *xeon\_b2b\_bar2\_addr64*.