1/*
2 * Intel MIC Platform Software Stack (MPSS)
3 *
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
6 *
7 * GPL LICENSE SUMMARY
8 *
9 * Copyright(c) 2014 Intel Corporation.
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of version 2 of the GNU General Public License as
13 * published by the Free Software Foundation.
14 *
15 * This program is distributed in the hope that it will be useful, but
16 * WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * BSD LICENSE
21 *
22 * Copyright(c) 2014 Intel Corporation.
23 *
24 * Redistribution and use in source and binary forms, with or without
25 * modification, are permitted provided that the following conditions
26 * are met:
27 *
28 * * Redistributions of source code must retain the above copyright
29 * notice, this list of conditions and the following disclaimer.
30 * * Redistributions in binary form must reproduce the above copyright
31 * notice, this list of conditions and the following disclaimer in
32 * the documentation and/or other materials provided with the
33 * distribution.
34 * * Neither the name of Intel Corporation nor the names of its
35 * contributors may be used to endorse or promote products derived
36 * from this software without specific prior written permission.
37 *
38 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
39 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
40 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
41 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
42 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
43 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
44 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
45 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
46 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
47 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
48 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
49 *
50 * Intel SCIF driver.
51 *
52 */
53#ifndef __SCIF_H__
54#define __SCIF_H__
55
56#include <linux/types.h>
57#include <linux/poll.h>
58#include <linux/device.h>
59#include <linux/scif_ioctl.h>
60
61#define SCIF_ACCEPT_SYNC 1
62#define SCIF_SEND_BLOCK 1
63#define SCIF_RECV_BLOCK 1
64
65enum {
66 SCIF_PROT_READ = (1 << 0),
67 SCIF_PROT_WRITE = (1 << 1)
68};
69
70enum {
71 SCIF_MAP_FIXED = 0x10,
72 SCIF_MAP_KERNEL = 0x20,
73};
74
75enum {
76 SCIF_FENCE_INIT_SELF = (1 << 0),
77 SCIF_FENCE_INIT_PEER = (1 << 1),
78 SCIF_SIGNAL_LOCAL = (1 << 4),
79 SCIF_SIGNAL_REMOTE = (1 << 5)
80};
81
82enum {
83 SCIF_RMA_USECPU = (1 << 0),
84 SCIF_RMA_USECACHE = (1 << 1),
85 SCIF_RMA_SYNC = (1 << 2),
86 SCIF_RMA_ORDERED = (1 << 3)
87};
88
89/* End of SCIF Admin Reserved Ports */
90#define SCIF_ADMIN_PORT_END 1024
91
92/* End of SCIF Reserved Ports */
93#define SCIF_PORT_RSVD 1088
94
95typedef struct scif_endpt *scif_epd_t;
96typedef struct scif_pinned_pages *scif_pinned_pages_t;
97
98/**
99 * struct scif_range - SCIF registered range used in kernel mode
100 * @cookie: cookie used internally by SCIF
101 * @nr_pages: number of pages of PAGE_SIZE
102 * @prot_flags: R/W protection
103 * @phys_addr: Array of bus addresses
104 * @va: Array of kernel virtual addresses backed by the pages in the phys_addr
105 * array. The va is populated only when called on the host for a remote
106 * SCIF connection on MIC. This is required to support the use case of DMA
107 * between MIC and another device which is not a SCIF node e.g., an IB or
108 * ethernet NIC.
109 */
110struct scif_range {
111 void *cookie;
112 int nr_pages;
113 int prot_flags;
114 dma_addr_t *phys_addr;
115 void __iomem **va;
116};
117
118/**
119 * struct scif_pollepd - SCIF endpoint to be monitored via scif_poll
120 * @epd: SCIF endpoint
121 * @events: requested events
122 * @revents: returned events
123 */
124struct scif_pollepd {
125 scif_epd_t epd;
126 __poll_t events;
127 __poll_t revents;
128};
129
130/**
131 * scif_peer_dev - representation of a peer SCIF device
132 *
133 * Peer devices show up as PCIe devices for the mgmt node but not the cards.
134 * The mgmt node discovers all the cards on the PCIe bus and informs the other
135 * cards about their peers. Upon notification of a peer a node adds a peer
136 * device to the peer bus to maintain symmetry in the way devices are
137 * discovered across all nodes in the SCIF network.
138 *
139 * @dev: underlying device
140 * @dnode - The destination node which this device will communicate with.
141 */
142struct scif_peer_dev {
143 struct device dev;
144 u8 dnode;
145};
146
147/**
148 * scif_client - representation of a SCIF client
149 * @name: client name
150 * @probe - client method called when a peer device is registered
151 * @remove - client method called when a peer device is unregistered
152 * @si - subsys_interface used internally for implementing SCIF clients
153 */
154struct scif_client {
155 const char *name;
156 void (*probe)(struct scif_peer_dev *spdev);
157 void (*remove)(struct scif_peer_dev *spdev);
158 struct subsys_interface si;
159};
160
161#define SCIF_OPEN_FAILED ((scif_epd_t)-1)
162#define SCIF_REGISTER_FAILED ((off_t)-1)
163#define SCIF_MMAP_FAILED ((void *)-1)
164
165/**
166 * scif_open() - Create an endpoint
167 *
168 * Return:
169 * Upon successful completion, scif_open() returns an endpoint descriptor to
170 * be used in subsequent SCIF functions calls to refer to that endpoint;
171 * otherwise in user mode SCIF_OPEN_FAILED (that is ((scif_epd_t)-1)) is
172 * returned and errno is set to indicate the error; in kernel mode a NULL
173 * scif_epd_t is returned.
174 *
175 * Errors:
176 * ENOMEM - Insufficient kernel memory was available
177 */
178scif_epd_t scif_open(void);
179
180/**
181 * scif_bind() - Bind an endpoint to a port
182 * @epd: endpoint descriptor
183 * @pn: port number
184 *
185 * scif_bind() binds endpoint epd to port pn, where pn is a port number on the
186 * local node. If pn is zero, a port number greater than or equal to
187 * SCIF_PORT_RSVD is assigned and returned. Each endpoint may be bound to
188 * exactly one local port. Ports less than 1024 when requested can only be bound
189 * by system (or root) processes or by processes executed by privileged users.
190 *
191 * Return:
192 * Upon successful completion, scif_bind() returns the port number to which epd
193 * is bound; otherwise in user mode -1 is returned and errno is set to
194 * indicate the error; in kernel mode the negative of one of the following
195 * errors is returned.
196 *
197 * Errors:
198 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
199 * EINVAL - the endpoint or the port is already bound
200 * EISCONN - The endpoint is already connected
201 * ENOSPC - No port number available for assignment
202 * EACCES - The port requested is protected and the user is not the superuser
203 */
204int scif_bind(scif_epd_t epd, u16 pn);
205
206/**
207 * scif_listen() - Listen for connections on an endpoint
208 * @epd: endpoint descriptor
209 * @backlog: maximum pending connection requests
210 *
211 * scif_listen() marks the endpoint epd as a listening endpoint - that is, as
212 * an endpoint that will be used to accept incoming connection requests. Once
213 * so marked, the endpoint is said to be in the listening state and may not be
214 * used as the endpoint of a connection.
215 *
216 * The endpoint, epd, must have been bound to a port.
217 *
218 * The backlog argument defines the maximum length to which the queue of
219 * pending connections for epd may grow. If a connection request arrives when
220 * the queue is full, the client may receive an error with an indication that
221 * the connection was refused.
222 *
223 * Return:
224 * Upon successful completion, scif_listen() returns 0; otherwise in user mode
225 * -1 is returned and errno is set to indicate the error; in kernel mode the
226 * negative of one of the following errors is returned.
227 *
228 * Errors:
229 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
230 * EINVAL - the endpoint is not bound to a port
231 * EISCONN - The endpoint is already connected or listening
232 */
233int scif_listen(scif_epd_t epd, int backlog);
234
235/**
236 * scif_connect() - Initiate a connection on a port
237 * @epd: endpoint descriptor
238 * @dst: global id of port to which to connect
239 *
240 * The scif_connect() function requests the connection of endpoint epd to remote
241 * port dst. If the connection is successful, a peer endpoint, bound to dst, is
242 * created on node dst.node. On successful return, the connection is complete.
243 *
244 * If the endpoint epd has not already been bound to a port, scif_connect()
245 * will bind it to an unused local port.
246 *
247 * A connection is terminated when an endpoint of the connection is closed,
248 * either explicitly by scif_close(), or when a process that owns one of the
249 * endpoints of the connection is terminated.
250 *
251 * In user space, scif_connect() supports an asynchronous connection mode
252 * if the application has set the O_NONBLOCK flag on the endpoint via the
253 * fcntl() system call. Setting this flag will result in the calling process
254 * not to wait during scif_connect().
255 *
256 * Return:
257 * Upon successful completion, scif_connect() returns the port ID to which the
258 * endpoint, epd, is bound; otherwise in user mode -1 is returned and errno is
259 * set to indicate the error; in kernel mode the negative of one of the
260 * following errors is returned.
261 *
262 * Errors:
263 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
264 * ECONNREFUSED - The destination was not listening for connections or refused
265 * the connection request
266 * EINVAL - dst.port is not a valid port ID
267 * EISCONN - The endpoint is already connected
268 * ENOMEM - No buffer space is available
269 * ENODEV - The destination node does not exist, or the node is lost or existed,
270 * but is not currently in the network since it may have crashed
271 * ENOSPC - No port number available for assignment
272 * EOPNOTSUPP - The endpoint is listening and cannot be connected
273 */
274int scif_connect(scif_epd_t epd, struct scif_port_id *dst);
275
276/**
277 * scif_accept() - Accept a connection on an endpoint
278 * @epd: endpoint descriptor
279 * @peer: global id of port to which connected
280 * @newepd: new connected endpoint descriptor
281 * @flags: flags
282 *
283 * The scif_accept() call extracts the first connection request from the queue
284 * of pending connections for the port on which epd is listening. scif_accept()
285 * creates a new endpoint, bound to the same port as epd, and allocates a new
286 * SCIF endpoint descriptor, returned in newepd, for the endpoint. The new
287 * endpoint is connected to the endpoint through which the connection was
288 * requested. epd is unaffected by this call, and remains in the listening
289 * state.
290 *
291 * On successful return, peer holds the global port identifier (node id and
292 * local port number) of the port which requested the connection.
293 *
294 * A connection is terminated when an endpoint of the connection is closed,
295 * either explicitly by scif_close(), or when a process that owns one of the
296 * endpoints of the connection is terminated.
297 *
298 * The number of connections that can (subsequently) be accepted on epd is only
299 * limited by system resources (memory).
300 *
301 * The flags argument is formed by OR'ing together zero or more of the
302 * following values.
303 * SCIF_ACCEPT_SYNC - block until a connection request is presented. If
304 * SCIF_ACCEPT_SYNC is not in flags, and no pending
305 * connections are present on the queue, scif_accept()
306 * fails with an EAGAIN error
307 *
308 * In user mode, the select() and poll() functions can be used to determine
309 * when there is a connection request. In kernel mode, the scif_poll()
310 * function may be used for this purpose. A readable event will be delivered
311 * when a connection is requested.
312 *
313 * Return:
314 * Upon successful completion, scif_accept() returns 0; otherwise in user mode
315 * -1 is returned and errno is set to indicate the error; in kernel mode the
316 * negative of one of the following errors is returned.
317 *
318 * Errors:
319 * EAGAIN - SCIF_ACCEPT_SYNC is not set and no connections are present to be
320 * accepted or SCIF_ACCEPT_SYNC is not set and remote node failed to complete
321 * its connection request
322 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
323 * EINTR - Interrupted function
324 * EINVAL - epd is not a listening endpoint, or flags is invalid, or peer is
325 * NULL, or newepd is NULL
326 * ENODEV - The requesting node is lost or existed, but is not currently in the
327 * network since it may have crashed
328 * ENOMEM - Not enough space
329 * ENOENT - Secondary part of epd registration failed
330 */
331int scif_accept(scif_epd_t epd, struct scif_port_id *peer, scif_epd_t
332 *newepd, int flags);
333
334/**
335 * scif_close() - Close an endpoint
336 * @epd: endpoint descriptor
337 *
338 * scif_close() closes an endpoint and performs necessary teardown of
339 * facilities associated with that endpoint.
340 *
341 * If epd is a listening endpoint then it will no longer accept connection
342 * requests on the port to which it is bound. Any pending connection requests
343 * are rejected.
344 *
345 * If epd is a connected endpoint, then its peer endpoint is also closed. RMAs
346 * which are in-process through epd or its peer endpoint will complete before
347 * scif_close() returns. Registered windows of the local and peer endpoints are
348 * released as if scif_unregister() was called against each window.
349 *
350 * Closing a SCIF endpoint does not affect local registered memory mapped by
351 * a SCIF endpoint on a remote node. The local memory remains mapped by the peer
352 * SCIF endpoint explicitly removed by calling munmap(..) by the peer.
353 *
354 * If the peer endpoint's receive queue is not empty at the time that epd is
355 * closed, then the peer endpoint can be passed as the endpoint parameter to
356 * scif_recv() until the receive queue is empty.
357 *
358 * epd is freed and may no longer be accessed.
359 *
360 * Return:
361 * Upon successful completion, scif_close() returns 0; otherwise in user mode
362 * -1 is returned and errno is set to indicate the error; in kernel mode the
363 * negative of one of the following errors is returned.
364 *
365 * Errors:
366 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
367 */
368int scif_close(scif_epd_t epd);
369
370/**
371 * scif_send() - Send a message
372 * @epd: endpoint descriptor
373 * @msg: message buffer address
374 * @len: message length
375 * @flags: blocking mode flags
376 *
377 * scif_send() sends data to the peer of endpoint epd. Up to len bytes of data
378 * are copied from memory starting at address msg. On successful execution the
379 * return value of scif_send() is the number of bytes that were sent, and is
380 * zero if no bytes were sent because len was zero. scif_send() may be called
381 * only when the endpoint is in a connected state.
382 *
383 * If a scif_send() call is non-blocking, then it sends only those bytes which
384 * can be sent without waiting, up to a maximum of len bytes.
385 *
386 * If a scif_send() call is blocking, then it normally returns after sending
387 * all len bytes. If a blocking call is interrupted or the connection is
388 * reset, the call is considered successful if some bytes were sent or len is
389 * zero, otherwise the call is considered unsuccessful.
390 *
391 * In user mode, the select() and poll() functions can be used to determine
392 * when the send queue is not full. In kernel mode, the scif_poll() function
393 * may be used for this purpose.
394 *
395 * It is recommended that scif_send()/scif_recv() only be used for short
396 * control-type message communication between SCIF endpoints. The SCIF RMA
397 * APIs are expected to provide better performance for transfer sizes of
398 * 1024 bytes or longer for the current MIC hardware and software
399 * implementation.
400 *
401 * scif_send() will block until the entire message is sent if SCIF_SEND_BLOCK
402 * is passed as the flags argument.
403 *
404 * Return:
405 * Upon successful completion, scif_send() returns the number of bytes sent;
406 * otherwise in user mode -1 is returned and errno is set to indicate the
407 * error; in kernel mode the negative of one of the following errors is
408 * returned.
409 *
410 * Errors:
411 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
412 * ECONNRESET - Connection reset by peer
413 * EINVAL - flags is invalid, or len is negative
414 * ENODEV - The remote node is lost or existed, but is not currently in the
415 * network since it may have crashed
416 * ENOMEM - Not enough space
417 * ENOTCONN - The endpoint is not connected
418 */
419int scif_send(scif_epd_t epd, void *msg, int len, int flags);
420
421/**
422 * scif_recv() - Receive a message
423 * @epd: endpoint descriptor
424 * @msg: message buffer address
425 * @len: message buffer length
426 * @flags: blocking mode flags
427 *
428 * scif_recv() receives data from the peer of endpoint epd. Up to len bytes of
429 * data are copied to memory starting at address msg. On successful execution
430 * the return value of scif_recv() is the number of bytes that were received,
431 * and is zero if no bytes were received because len was zero. scif_recv() may
432 * be called only when the endpoint is in a connected state.
433 *
434 * If a scif_recv() call is non-blocking, then it receives only those bytes
435 * which can be received without waiting, up to a maximum of len bytes.
436 *
437 * If a scif_recv() call is blocking, then it normally returns after receiving
438 * all len bytes. If the blocking call was interrupted due to a disconnection,
439 * subsequent calls to scif_recv() will copy all bytes received upto the point
440 * of disconnection.
441 *
442 * In user mode, the select() and poll() functions can be used to determine
443 * when data is available to be received. In kernel mode, the scif_poll()
444 * function may be used for this purpose.
445 *
446 * It is recommended that scif_send()/scif_recv() only be used for short
447 * control-type message communication between SCIF endpoints. The SCIF RMA
448 * APIs are expected to provide better performance for transfer sizes of
449 * 1024 bytes or longer for the current MIC hardware and software
450 * implementation.
451 *
452 * scif_recv() will block until the entire message is received if
453 * SCIF_RECV_BLOCK is passed as the flags argument.
454 *
455 * Return:
456 * Upon successful completion, scif_recv() returns the number of bytes
457 * received; otherwise in user mode -1 is returned and errno is set to
458 * indicate the error; in kernel mode the negative of one of the following
459 * errors is returned.
460 *
461 * Errors:
462 * EAGAIN - The destination node is returning from a low power state
463 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
464 * ECONNRESET - Connection reset by peer
465 * EINVAL - flags is invalid, or len is negative
466 * ENODEV - The remote node is lost or existed, but is not currently in the
467 * network since it may have crashed
468 * ENOMEM - Not enough space
469 * ENOTCONN - The endpoint is not connected
470 */
471int scif_recv(scif_epd_t epd, void *msg, int len, int flags);
472
473/**
474 * scif_register() - Mark a memory region for remote access.
475 * @epd: endpoint descriptor
476 * @addr: starting virtual address
477 * @len: length of range
478 * @offset: offset of window
479 * @prot_flags: read/write protection flags
480 * @map_flags: mapping flags
481 *
482 * The scif_register() function opens a window, a range of whole pages of the
483 * registered address space of the endpoint epd, starting at offset po and
484 * continuing for len bytes. The value of po, further described below, is a
485 * function of the parameters offset and len, and the value of map_flags. Each
486 * page of the window represents the physical memory page which backs the
487 * corresponding page of the range of virtual address pages starting at addr
488 * and continuing for len bytes. addr and len are constrained to be multiples
489 * of the page size. A successful scif_register() call returns po.
490 *
491 * When SCIF_MAP_FIXED is set in the map_flags argument, po will be offset
492 * exactly, and offset is constrained to be a multiple of the page size. The
493 * mapping established by scif_register() will not replace any existing
494 * registration; an error is returned if any page within the range [offset,
495 * offset + len - 1] intersects an existing window.
496 *
497 * When SCIF_MAP_FIXED is not set, the implementation uses offset in an
498 * implementation-defined manner to arrive at po. The po value so chosen will
499 * be an area of the registered address space that the implementation deems
500 * suitable for a mapping of len bytes. An offset value of 0 is interpreted as
501 * granting the implementation complete freedom in selecting po, subject to
502 * constraints described below. A non-zero value of offset is taken to be a
503 * suggestion of an offset near which the mapping should be placed. When the
504 * implementation selects a value for po, it does not replace any extant
505 * window. In all cases, po will be a multiple of the page size.
506 *
507 * The physical pages which are so represented by a window are available for
508 * access in calls to mmap(), scif_readfrom(), scif_writeto(),
509 * scif_vreadfrom(), and scif_vwriteto(). While a window is registered, the
510 * physical pages represented by the window will not be reused by the memory
511 * subsystem for any other purpose. Note that the same physical page may be
512 * represented by multiple windows.
513 *
514 * Subsequent operations which change the memory pages to which virtual
515 * addresses are mapped (such as mmap(), munmap()) have no effect on
516 * existing window.
517 *
518 * If the process will fork(), it is recommended that the registered
519 * virtual address range be marked with MADV_DONTFORK. Doing so will prevent
520 * problems due to copy-on-write semantics.
521 *
522 * The prot_flags argument is formed by OR'ing together one or more of the
523 * following values.
524 * SCIF_PROT_READ - allow read operations from the window
525 * SCIF_PROT_WRITE - allow write operations to the window
526 *
527 * Return:
528 * Upon successful completion, scif_register() returns the offset at which the
529 * mapping was placed (po); otherwise in user mode SCIF_REGISTER_FAILED (that
530 * is (off_t *)-1) is returned and errno is set to indicate the error; in
531 * kernel mode the negative of one of the following errors is returned.
532 *
533 * Errors:
534 * EADDRINUSE - SCIF_MAP_FIXED is set in map_flags, and pages in the range
535 * [offset, offset + len -1] are already registered
536 * EAGAIN - The mapping could not be performed due to lack of resources
537 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
538 * ECONNRESET - Connection reset by peer
539 * EINVAL - map_flags is invalid, or prot_flags is invalid, or SCIF_MAP_FIXED is
540 * set in flags, and offset is not a multiple of the page size, or addr is not a
541 * multiple of the page size, or len is not a multiple of the page size, or is
542 * 0, or offset is negative
543 * ENODEV - The remote node is lost or existed, but is not currently in the
544 * network since it may have crashed
545 * ENOMEM - Not enough space
546 * ENOTCONN -The endpoint is not connected
547 */
548off_t scif_register(scif_epd_t epd, void *addr, size_t len, off_t offset,
549 int prot_flags, int map_flags);
550
551/**
552 * scif_unregister() - Mark a memory region for remote access.
553 * @epd: endpoint descriptor
554 * @offset: start of range to unregister
555 * @len: length of range to unregister
556 *
557 * The scif_unregister() function closes those previously registered windows
558 * which are entirely within the range [offset, offset + len - 1]. It is an
559 * error to specify a range which intersects only a subrange of a window.
560 *
561 * On a successful return, pages within the window may no longer be specified
562 * in calls to mmap(), scif_readfrom(), scif_writeto(), scif_vreadfrom(),
563 * scif_vwriteto(), scif_get_pages, and scif_fence_signal(). The window,
564 * however, continues to exist until all previous references against it are
565 * removed. A window is referenced if there is a mapping to it created by
566 * mmap(), or if scif_get_pages() was called against the window
567 * (and the pages have not been returned via scif_put_pages()). A window is
568 * also referenced while an RMA, in which some range of the window is a source
569 * or destination, is in progress. Finally a window is referenced while some
570 * offset in that window was specified to scif_fence_signal(), and the RMAs
571 * marked by that call to scif_fence_signal() have not completed. While a
572 * window is in this state, its registered address space pages are not
573 * available for use in a new registered window.
574 *
575 * When all such references to the window have been removed, its references to
576 * all the physical pages which it represents are removed. Similarly, the
577 * registered address space pages of the window become available for
578 * registration in a new window.
579 *
580 * Return:
581 * Upon successful completion, scif_unregister() returns 0; otherwise in user
582 * mode -1 is returned and errno is set to indicate the error; in kernel mode
583 * the negative of one of the following errors is returned. In the event of an
584 * error, no windows are unregistered.
585 *
586 * Errors:
587 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
588 * ECONNRESET - Connection reset by peer
589 * EINVAL - the range [offset, offset + len - 1] intersects a subrange of a
590 * window, or offset is negative
591 * ENODEV - The remote node is lost or existed, but is not currently in the
592 * network since it may have crashed
593 * ENOTCONN - The endpoint is not connected
594 * ENXIO - Offsets in the range [offset, offset + len - 1] are invalid for the
595 * registered address space of epd
596 */
597int scif_unregister(scif_epd_t epd, off_t offset, size_t len);
598
599/**
600 * scif_readfrom() - Copy from a remote address space
601 * @epd: endpoint descriptor
602 * @loffset: offset in local registered address space to
603 * which to copy
604 * @len: length of range to copy
605 * @roffset: offset in remote registered address space
606 * from which to copy
607 * @rma_flags: transfer mode flags
608 *
609 * scif_readfrom() copies len bytes from the remote registered address space of
610 * the peer of endpoint epd, starting at the offset roffset to the local
611 * registered address space of epd, starting at the offset loffset.
612 *
613 * Each of the specified ranges [loffset, loffset + len - 1] and [roffset,
614 * roffset + len - 1] must be within some registered window or windows of the
615 * local and remote nodes. A range may intersect multiple registered windows,
616 * but only if those windows are contiguous in the registered address space.
617 *
618 * If rma_flags includes SCIF_RMA_USECPU, then the data is copied using
619 * programmed read/writes. Otherwise the data is copied using DMA. If rma_-
620 * flags includes SCIF_RMA_SYNC, then scif_readfrom() will return after the
621 * transfer is complete. Otherwise, the transfer may be performed asynchron-
622 * ously. The order in which any two asynchronous RMA operations complete
623 * is non-deterministic. The synchronization functions, scif_fence_mark()/
624 * scif_fence_wait() and scif_fence_signal(), can be used to synchronize to
625 * the completion of asynchronous RMA operations on the same endpoint.
626 *
627 * The DMA transfer of individual bytes is not guaranteed to complete in
628 * address order. If rma_flags includes SCIF_RMA_ORDERED, then the last
629 * cacheline or partial cacheline of the source range will become visible on
630 * the destination node after all other transferred data in the source
631 * range has become visible on the destination node.
632 *
633 * The optimal DMA performance will likely be realized if both
634 * loffset and roffset are cacheline aligned (are a multiple of 64). Lower
635 * performance will likely be realized if loffset and roffset are not
636 * cacheline aligned but are separated by some multiple of 64. The lowest level
637 * of performance is likely if loffset and roffset are not separated by a
638 * multiple of 64.
639 *
640 * The rma_flags argument is formed by ORing together zero or more of the
641 * following values.
642 * SCIF_RMA_USECPU - perform the transfer using the CPU, otherwise use the DMA
643 * engine.
644 * SCIF_RMA_SYNC - perform the transfer synchronously, returning after the
645 * transfer has completed. Passing this flag results in the
646 * current implementation busy waiting and consuming CPU cycles
647 * while the DMA transfer is in progress for best performance by
648 * avoiding the interrupt latency.
649 * SCIF_RMA_ORDERED - ensure that the last cacheline or partial cacheline of
650 * the source range becomes visible on the destination node
651 * after all other transferred data in the source range has
652 * become visible on the destination
653 *
654 * Return:
655 * Upon successful completion, scif_readfrom() returns 0; otherwise in user
656 * mode -1 is returned and errno is set to indicate the error; in kernel mode
657 * the negative of one of the following errors is returned.
658 *
659 * Errors:
660 * EACCESS - Attempt to write to a read-only range
661 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
662 * ECONNRESET - Connection reset by peer
663 * EINVAL - rma_flags is invalid
664 * ENODEV - The remote node is lost or existed, but is not currently in the
665 * network since it may have crashed
666 * ENOTCONN - The endpoint is not connected
667 * ENXIO - The range [loffset, loffset + len - 1] is invalid for the registered
668 * address space of epd, or, The range [roffset, roffset + len - 1] is invalid
669 * for the registered address space of the peer of epd, or loffset or roffset
670 * is negative
671 */
672int scif_readfrom(scif_epd_t epd, off_t loffset, size_t len, off_t
673 roffset, int rma_flags);
674
675/**
676 * scif_writeto() - Copy to a remote address space
677 * @epd: endpoint descriptor
678 * @loffset: offset in local registered address space
679 * from which to copy
680 * @len: length of range to copy
681 * @roffset: offset in remote registered address space to
682 * which to copy
683 * @rma_flags: transfer mode flags
684 *
685 * scif_writeto() copies len bytes from the local registered address space of
686 * epd, starting at the offset loffset to the remote registered address space
687 * of the peer of endpoint epd, starting at the offset roffset.
688 *
689 * Each of the specified ranges [loffset, loffset + len - 1] and [roffset,
690 * roffset + len - 1] must be within some registered window or windows of the
691 * local and remote nodes. A range may intersect multiple registered windows,
692 * but only if those windows are contiguous in the registered address space.
693 *
694 * If rma_flags includes SCIF_RMA_USECPU, then the data is copied using
695 * programmed read/writes. Otherwise the data is copied using DMA. If rma_-
696 * flags includes SCIF_RMA_SYNC, then scif_writeto() will return after the
697 * transfer is complete. Otherwise, the transfer may be performed asynchron-
698 * ously. The order in which any two asynchronous RMA operations complete
699 * is non-deterministic. The synchronization functions, scif_fence_mark()/
700 * scif_fence_wait() and scif_fence_signal(), can be used to synchronize to
701 * the completion of asynchronous RMA operations on the same endpoint.
702 *
703 * The DMA transfer of individual bytes is not guaranteed to complete in
704 * address order. If rma_flags includes SCIF_RMA_ORDERED, then the last
705 * cacheline or partial cacheline of the source range will become visible on
706 * the destination node after all other transferred data in the source
707 * range has become visible on the destination node.
708 *
709 * The optimal DMA performance will likely be realized if both
710 * loffset and roffset are cacheline aligned (are a multiple of 64). Lower
711 * performance will likely be realized if loffset and roffset are not cacheline
712 * aligned but are separated by some multiple of 64. The lowest level of
713 * performance is likely if loffset and roffset are not separated by a multiple
714 * of 64.
715 *
716 * The rma_flags argument is formed by ORing together zero or more of the
717 * following values.
718 * SCIF_RMA_USECPU - perform the transfer using the CPU, otherwise use the DMA
719 * engine.
720 * SCIF_RMA_SYNC - perform the transfer synchronously, returning after the
721 * transfer has completed. Passing this flag results in the
722 * current implementation busy waiting and consuming CPU cycles
723 * while the DMA transfer is in progress for best performance by
724 * avoiding the interrupt latency.
725 * SCIF_RMA_ORDERED - ensure that the last cacheline or partial cacheline of
726 * the source range becomes visible on the destination node
727 * after all other transferred data in the source range has
728 * become visible on the destination
729 *
730 * Return:
731 * Upon successful completion, scif_readfrom() returns 0; otherwise in user
732 * mode -1 is returned and errno is set to indicate the error; in kernel mode
733 * the negative of one of the following errors is returned.
734 *
735 * Errors:
736 * EACCESS - Attempt to write to a read-only range
737 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
738 * ECONNRESET - Connection reset by peer
739 * EINVAL - rma_flags is invalid
740 * ENODEV - The remote node is lost or existed, but is not currently in the
741 * network since it may have crashed
742 * ENOTCONN - The endpoint is not connected
743 * ENXIO - The range [loffset, loffset + len - 1] is invalid for the registered
744 * address space of epd, or, The range [roffset , roffset + len -1] is invalid
745 * for the registered address space of the peer of epd, or loffset or roffset
746 * is negative
747 */
748int scif_writeto(scif_epd_t epd, off_t loffset, size_t len, off_t
749 roffset, int rma_flags);
750
751/**
752 * scif_vreadfrom() - Copy from a remote address space
753 * @epd: endpoint descriptor
754 * @addr: address to which to copy
755 * @len: length of range to copy
756 * @roffset: offset in remote registered address space
757 * from which to copy
758 * @rma_flags: transfer mode flags
759 *
760 * scif_vreadfrom() copies len bytes from the remote registered address
761 * space of the peer of endpoint epd, starting at the offset roffset, to local
762 * memory, starting at addr.
763 *
764 * The specified range [roffset, roffset + len - 1] must be within some
765 * registered window or windows of the remote nodes. The range may
766 * intersect multiple registered windows, but only if those windows are
767 * contiguous in the registered address space.
768 *
769 * If rma_flags includes SCIF_RMA_USECPU, then the data is copied using
770 * programmed read/writes. Otherwise the data is copied using DMA. If rma_-
771 * flags includes SCIF_RMA_SYNC, then scif_vreadfrom() will return after the
772 * transfer is complete. Otherwise, the transfer may be performed asynchron-
773 * ously. The order in which any two asynchronous RMA operations complete
774 * is non-deterministic. The synchronization functions, scif_fence_mark()/
775 * scif_fence_wait() and scif_fence_signal(), can be used to synchronize to
776 * the completion of asynchronous RMA operations on the same endpoint.
777 *
778 * The DMA transfer of individual bytes is not guaranteed to complete in
779 * address order. If rma_flags includes SCIF_RMA_ORDERED, then the last
780 * cacheline or partial cacheline of the source range will become visible on
781 * the destination node after all other transferred data in the source
782 * range has become visible on the destination node.
783 *
784 * If rma_flags includes SCIF_RMA_USECACHE, then the physical pages which back
785 * the specified local memory range may be remain in a pinned state even after
786 * the specified transfer completes. This may reduce overhead if some or all of
787 * the same virtual address range is referenced in a subsequent call of
788 * scif_vreadfrom() or scif_vwriteto().
789 *
790 * The optimal DMA performance will likely be realized if both
791 * addr and roffset are cacheline aligned (are a multiple of 64). Lower
792 * performance will likely be realized if addr and roffset are not
793 * cacheline aligned but are separated by some multiple of 64. The lowest level
794 * of performance is likely if addr and roffset are not separated by a
795 * multiple of 64.
796 *
797 * The rma_flags argument is formed by ORing together zero or more of the
798 * following values.
799 * SCIF_RMA_USECPU - perform the transfer using the CPU, otherwise use the DMA
800 * engine.
801 * SCIF_RMA_USECACHE - enable registration caching
802 * SCIF_RMA_SYNC - perform the transfer synchronously, returning after the
803 * transfer has completed. Passing this flag results in the
804 * current implementation busy waiting and consuming CPU cycles
805 * while the DMA transfer is in progress for best performance by
806 * avoiding the interrupt latency.
807 * SCIF_RMA_ORDERED - ensure that the last cacheline or partial cacheline of
808 * the source range becomes visible on the destination node
809 * after all other transferred data in the source range has
810 * become visible on the destination
811 *
812 * Return:
813 * Upon successful completion, scif_vreadfrom() returns 0; otherwise in user
814 * mode -1 is returned and errno is set to indicate the error; in kernel mode
815 * the negative of one of the following errors is returned.
816 *
817 * Errors:
818 * EACCESS - Attempt to write to a read-only range
819 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
820 * ECONNRESET - Connection reset by peer
821 * EINVAL - rma_flags is invalid
822 * ENODEV - The remote node is lost or existed, but is not currently in the
823 * network since it may have crashed
824 * ENOTCONN - The endpoint is not connected
825 * ENXIO - Offsets in the range [roffset, roffset + len - 1] are invalid for the
826 * registered address space of epd
827 */
828int scif_vreadfrom(scif_epd_t epd, void *addr, size_t len, off_t roffset,
829 int rma_flags);
830
831/**
832 * scif_vwriteto() - Copy to a remote address space
833 * @epd: endpoint descriptor
834 * @addr: address from which to copy
835 * @len: length of range to copy
836 * @roffset: offset in remote registered address space to
837 * which to copy
838 * @rma_flags: transfer mode flags
839 *
840 * scif_vwriteto() copies len bytes from the local memory, starting at addr, to
841 * the remote registered address space of the peer of endpoint epd, starting at
842 * the offset roffset.
843 *
844 * The specified range [roffset, roffset + len - 1] must be within some
845 * registered window or windows of the remote nodes. The range may intersect
846 * multiple registered windows, but only if those windows are contiguous in the
847 * registered address space.
848 *
849 * If rma_flags includes SCIF_RMA_USECPU, then the data is copied using
850 * programmed read/writes. Otherwise the data is copied using DMA. If rma_-
851 * flags includes SCIF_RMA_SYNC, then scif_vwriteto() will return after the
852 * transfer is complete. Otherwise, the transfer may be performed asynchron-
853 * ously. The order in which any two asynchronous RMA operations complete
854 * is non-deterministic. The synchronization functions, scif_fence_mark()/
855 * scif_fence_wait() and scif_fence_signal(), can be used to synchronize to
856 * the completion of asynchronous RMA operations on the same endpoint.
857 *
858 * The DMA transfer of individual bytes is not guaranteed to complete in
859 * address order. If rma_flags includes SCIF_RMA_ORDERED, then the last
860 * cacheline or partial cacheline of the source range will become visible on
861 * the destination node after all other transferred data in the source
862 * range has become visible on the destination node.
863 *
864 * If rma_flags includes SCIF_RMA_USECACHE, then the physical pages which back
865 * the specified local memory range may be remain in a pinned state even after
866 * the specified transfer completes. This may reduce overhead if some or all of
867 * the same virtual address range is referenced in a subsequent call of
868 * scif_vreadfrom() or scif_vwriteto().
869 *
870 * The optimal DMA performance will likely be realized if both
871 * addr and offset are cacheline aligned (are a multiple of 64). Lower
872 * performance will likely be realized if addr and offset are not cacheline
873 * aligned but are separated by some multiple of 64. The lowest level of
874 * performance is likely if addr and offset are not separated by a multiple of
875 * 64.
876 *
877 * The rma_flags argument is formed by ORing together zero or more of the
878 * following values.
879 * SCIF_RMA_USECPU - perform the transfer using the CPU, otherwise use the DMA
880 * engine.
881 * SCIF_RMA_USECACHE - allow registration caching
882 * SCIF_RMA_SYNC - perform the transfer synchronously, returning after the
883 * transfer has completed. Passing this flag results in the
884 * current implementation busy waiting and consuming CPU cycles
885 * while the DMA transfer is in progress for best performance by
886 * avoiding the interrupt latency.
887 * SCIF_RMA_ORDERED - ensure that the last cacheline or partial cacheline of
888 * the source range becomes visible on the destination node
889 * after all other transferred data in the source range has
890 * become visible on the destination
891 *
892 * Return:
893 * Upon successful completion, scif_vwriteto() returns 0; otherwise in user
894 * mode -1 is returned and errno is set to indicate the error; in kernel mode
895 * the negative of one of the following errors is returned.
896 *
897 * Errors:
898 * EACCESS - Attempt to write to a read-only range
899 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
900 * ECONNRESET - Connection reset by peer
901 * EINVAL - rma_flags is invalid
902 * ENODEV - The remote node is lost or existed, but is not currently in the
903 * network since it may have crashed
904 * ENOTCONN - The endpoint is not connected
905 * ENXIO - Offsets in the range [roffset, roffset + len - 1] are invalid for the
906 * registered address space of epd
907 */
908int scif_vwriteto(scif_epd_t epd, void *addr, size_t len, off_t roffset,
909 int rma_flags);
910
911/**
912 * scif_fence_mark() - Mark previously issued RMAs
913 * @epd: endpoint descriptor
914 * @flags: control flags
915 * @mark: marked value returned as output.
916 *
917 * scif_fence_mark() returns after marking the current set of all uncompleted
918 * RMAs initiated through the endpoint epd or the current set of all
919 * uncompleted RMAs initiated through the peer of endpoint epd. The RMAs are
920 * marked with a value returned at mark. The application may subsequently call
921 * scif_fence_wait(), passing the value returned at mark, to await completion
922 * of all RMAs so marked.
923 *
924 * The flags argument has exactly one of the following values.
925 * SCIF_FENCE_INIT_SELF - RMA operations initiated through endpoint
926 * epd are marked
927 * SCIF_FENCE_INIT_PEER - RMA operations initiated through the peer
928 * of endpoint epd are marked
929 *
930 * Return:
931 * Upon successful completion, scif_fence_mark() returns 0; otherwise in user
932 * mode -1 is returned and errno is set to indicate the error; in kernel mode
933 * the negative of one of the following errors is returned.
934 *
935 * Errors:
936 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
937 * ECONNRESET - Connection reset by peer
938 * EINVAL - flags is invalid
939 * ENODEV - The remote node is lost or existed, but is not currently in the
940 * network since it may have crashed
941 * ENOTCONN - The endpoint is not connected
942 * ENOMEM - Insufficient kernel memory was available
943 */
944int scif_fence_mark(scif_epd_t epd, int flags, int *mark);
945
946/**
947 * scif_fence_wait() - Wait for completion of marked RMAs
948 * @epd: endpoint descriptor
949 * @mark: mark request
950 *
951 * scif_fence_wait() returns after all RMAs marked with mark have completed.
952 * The value passed in mark must have been obtained in a previous call to
953 * scif_fence_mark().
954 *
955 * Return:
956 * Upon successful completion, scif_fence_wait() returns 0; otherwise in user
957 * mode -1 is returned and errno is set to indicate the error; in kernel mode
958 * the negative of one of the following errors is returned.
959 *
960 * Errors:
961 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
962 * ECONNRESET - Connection reset by peer
963 * ENODEV - The remote node is lost or existed, but is not currently in the
964 * network since it may have crashed
965 * ENOTCONN - The endpoint is not connected
966 * ENOMEM - Insufficient kernel memory was available
967 */
968int scif_fence_wait(scif_epd_t epd, int mark);
969
970/**
971 * scif_fence_signal() - Request a memory update on completion of RMAs
972 * @epd: endpoint descriptor
973 * @loff: local offset
974 * @lval: local value to write to loffset
975 * @roff: remote offset
976 * @rval: remote value to write to roffset
977 * @flags: flags
978 *
979 * scif_fence_signal() returns after marking the current set of all uncompleted
980 * RMAs initiated through the endpoint epd or marking the current set of all
981 * uncompleted RMAs initiated through the peer of endpoint epd.
982 *
983 * If flags includes SCIF_SIGNAL_LOCAL, then on completion of the RMAs in the
984 * marked set, lval is written to memory at the address corresponding to offset
985 * loff in the local registered address space of epd. loff must be within a
986 * registered window. If flags includes SCIF_SIGNAL_REMOTE, then on completion
987 * of the RMAs in the marked set, rval is written to memory at the address
988 * corresponding to offset roff in the remote registered address space of epd.
989 * roff must be within a remote registered window of the peer of epd. Note
990 * that any specified offset must be DWORD (4 byte / 32 bit) aligned.
991 *
992 * The flags argument is formed by OR'ing together the following.
993 * Exactly one of the following values.
994 * SCIF_FENCE_INIT_SELF - RMA operations initiated through endpoint
995 * epd are marked
996 * SCIF_FENCE_INIT_PEER - RMA operations initiated through the peer
997 * of endpoint epd are marked
998 * One or more of the following values.
999 * SCIF_SIGNAL_LOCAL - On completion of the marked set of RMAs, write lval to
1000 * memory at the address corresponding to offset loff in the local
1001 * registered address space of epd.
1002 * SCIF_SIGNAL_REMOTE - On completion of the marked set of RMAs, write rval to
1003 * memory at the address corresponding to offset roff in the remote
1004 * registered address space of epd.
1005 *
1006 * Return:
1007 * Upon successful completion, scif_fence_signal() returns 0; otherwise in
1008 * user mode -1 is returned and errno is set to indicate the error; in kernel
1009 * mode the negative of one of the following errors is returned.
1010 *
1011 * Errors:
1012 * EBADF, ENOTTY - epd is not a valid endpoint descriptor
1013 * ECONNRESET - Connection reset by peer
1014 * EINVAL - flags is invalid, or loff or roff are not DWORD aligned
1015 * ENODEV - The remote node is lost or existed, but is not currently in the
1016 * network since it may have crashed
1017 * ENOTCONN - The endpoint is not connected
1018 * ENXIO - loff is invalid for the registered address of epd, or roff is invalid
1019 * for the registered address space, of the peer of epd
1020 */
1021int scif_fence_signal(scif_epd_t epd, off_t loff, u64 lval, off_t roff,
1022 u64 rval, int flags);
1023
1024/**
1025 * scif_get_node_ids() - Return information about online nodes
1026 * @nodes: array in which to return online node IDs
1027 * @len: number of entries in the nodes array
1028 * @self: address to place the node ID of the local node
1029 *
1030 * scif_get_node_ids() fills in the nodes array with up to len node IDs of the
1031 * nodes in the SCIF network. If there is not enough space in nodes, as
1032 * indicated by the len parameter, only len node IDs are returned in nodes. The
1033 * return value of scif_get_node_ids() is the total number of nodes currently in
1034 * the SCIF network. By checking the return value against the len parameter,
1035 * the user may determine if enough space for nodes was allocated.
1036 *
1037 * The node ID of the local node is returned at self.
1038 *
1039 * Return:
1040 * Upon successful completion, scif_get_node_ids() returns the actual number of
1041 * online nodes in the SCIF network including 'self'; otherwise in user mode
1042 * -1 is returned and errno is set to indicate the error; in kernel mode no
1043 * errors are returned.
1044 */
1045int scif_get_node_ids(u16 *nodes, int len, u16 *self);
1046
1047/**
1048 * scif_pin_pages() - Pin a set of pages
1049 * @addr: Virtual address of range to pin
1050 * @len: Length of range to pin
1051 * @prot_flags: Page protection flags
1052 * @map_flags: Page classification flags
1053 * @pinned_pages: Handle to pinned pages
1054 *
1055 * scif_pin_pages() pins (locks in physical memory) the physical pages which
1056 * back the range of virtual address pages starting at addr and continuing for
1057 * len bytes. addr and len are constrained to be multiples of the page size. A
1058 * successful scif_pin_pages() call returns a handle to pinned_pages which may
1059 * be used in subsequent calls to scif_register_pinned_pages().
1060 *
1061 * The pages will remain pinned as long as there is a reference against the
1062 * scif_pinned_pages_t value returned by scif_pin_pages() and until
1063 * scif_unpin_pages() is called, passing the scif_pinned_pages_t value. A
1064 * reference is added to a scif_pinned_pages_t value each time a window is
1065 * created by calling scif_register_pinned_pages() and passing the
1066 * scif_pinned_pages_t value. A reference is removed from a
1067 * scif_pinned_pages_t value each time such a window is deleted.
1068 *
1069 * Subsequent operations which change the memory pages to which virtual
1070 * addresses are mapped (such as mmap(), munmap()) have no effect on the
1071 * scif_pinned_pages_t value or windows created against it.
1072 *
1073 * If the process will fork(), it is recommended that the registered
1074 * virtual address range be marked with MADV_DONTFORK. Doing so will prevent
1075 * problems due to copy-on-write semantics.
1076 *
1077 * The prot_flags argument is formed by OR'ing together one or more of the
1078 * following values.
1079 * SCIF_PROT_READ - allow read operations against the pages
1080 * SCIF_PROT_WRITE - allow write operations against the pages
1081 * The map_flags argument can be set as SCIF_MAP_KERNEL to interpret addr as a
1082 * kernel space address. By default, addr is interpreted as a user space
1083 * address.
1084 *
1085 * Return:
1086 * Upon successful completion, scif_pin_pages() returns 0; otherwise the
1087 * negative of one of the following errors is returned.
1088 *
1089 * Errors:
1090 * EINVAL - prot_flags is invalid, map_flags is invalid, or offset is negative
1091 * ENOMEM - Not enough space
1092 */
1093int scif_pin_pages(void *addr, size_t len, int prot_flags, int map_flags,
1094 scif_pinned_pages_t *pinned_pages);
1095
1096/**
1097 * scif_unpin_pages() - Unpin a set of pages
1098 * @pinned_pages: Handle to pinned pages to be unpinned
1099 *
1100 * scif_unpin_pages() prevents scif_register_pinned_pages() from registering new
1101 * windows against pinned_pages. The physical pages represented by pinned_pages
1102 * will remain pinned until all windows previously registered against
1103 * pinned_pages are deleted (the window is scif_unregister()'d and all
1104 * references to the window are removed (see scif_unregister()).
1105 *
1106 * pinned_pages must have been obtain from a previous call to scif_pin_pages().
1107 * After calling scif_unpin_pages(), it is an error to pass pinned_pages to
1108 * scif_register_pinned_pages().
1109 *
1110 * Return:
1111 * Upon successful completion, scif_unpin_pages() returns 0; otherwise the
1112 * negative of one of the following errors is returned.
1113 *
1114 * Errors:
1115 * EINVAL - pinned_pages is not valid
1116 */
1117int scif_unpin_pages(scif_pinned_pages_t pinned_pages);
1118
1119/**
1120 * scif_register_pinned_pages() - Mark a memory region for remote access.
1121 * @epd: endpoint descriptor
1122 * @pinned_pages: Handle to pinned pages
1123 * @offset: Registered address space offset
1124 * @map_flags: Flags which control where pages are mapped
1125 *
1126 * The scif_register_pinned_pages() function opens a window, a range of whole
1127 * pages of the registered address space of the endpoint epd, starting at
1128 * offset po. The value of po, further described below, is a function of the
1129 * parameters offset and pinned_pages, and the value of map_flags. Each page of
1130 * the window represents a corresponding physical memory page of the range
1131 * represented by pinned_pages; the length of the window is the same as the
1132 * length of range represented by pinned_pages. A successful
1133 * scif_register_pinned_pages() call returns po as the return value.
1134 *
1135 * When SCIF_MAP_FIXED is set in the map_flags argument, po will be offset
1136 * exactly, and offset is constrained to be a multiple of the page size. The
1137 * mapping established by scif_register_pinned_pages() will not replace any
1138 * existing registration; an error is returned if any page of the new window
1139 * would intersect an existing window.
1140 *
1141 * When SCIF_MAP_FIXED is not set, the implementation uses offset in an
1142 * implementation-defined manner to arrive at po. The po so chosen will be an
1143 * area of the registered address space that the implementation deems suitable
1144 * for a mapping of the required size. An offset value of 0 is interpreted as
1145 * granting the implementation complete freedom in selecting po, subject to
1146 * constraints described below. A non-zero value of offset is taken to be a
1147 * suggestion of an offset near which the mapping should be placed. When the
1148 * implementation selects a value for po, it does not replace any extant
1149 * window. In all cases, po will be a multiple of the page size.
1150 *
1151 * The physical pages which are so represented by a window are available for
1152 * access in calls to scif_get_pages(), scif_readfrom(), scif_writeto(),
1153 * scif_vreadfrom(), and scif_vwriteto(). While a window is registered, the
1154 * physical pages represented by the window will not be reused by the memory
1155 * subsystem for any other purpose. Note that the same physical page may be
1156 * represented by multiple windows.
1157 *
1158 * Windows created by scif_register_pinned_pages() are unregistered by
1159 * scif_unregister().
1160 *
1161 * The map_flags argument can be set to SCIF_MAP_FIXED which interprets a
1162 * fixed offset.
1163 *
1164 * Return:
1165 * Upon successful completion, scif_register_pinned_pages() returns the offset
1166 * at which the mapping was placed (po); otherwise the negative of one of the
1167 * following errors is returned.
1168 *
1169 * Errors:
1170 * EADDRINUSE - SCIF_MAP_FIXED is set in map_flags and pages in the new window
1171 * would intersect an existing window
1172 * EAGAIN - The mapping could not be performed due to lack of resources
1173 * ECONNRESET - Connection reset by peer
1174 * EINVAL - map_flags is invalid, or SCIF_MAP_FIXED is set in map_flags, and
1175 * offset is not a multiple of the page size, or offset is negative
1176 * ENODEV - The remote node is lost or existed, but is not currently in the
1177 * network since it may have crashed
1178 * ENOMEM - Not enough space
1179 * ENOTCONN - The endpoint is not connected
1180 */
1181off_t scif_register_pinned_pages(scif_epd_t epd,
1182 scif_pinned_pages_t pinned_pages,
1183 off_t offset, int map_flags);
1184
1185/**
1186 * scif_get_pages() - Add references to remote registered pages
1187 * @epd: endpoint descriptor
1188 * @offset: remote registered offset
1189 * @len: length of range of pages
1190 * @pages: returned scif_range structure
1191 *
1192 * scif_get_pages() returns the addresses of the physical pages represented by
1193 * those pages of the registered address space of the peer of epd, starting at
1194 * offset and continuing for len bytes. offset and len are constrained to be
1195 * multiples of the page size.
1196 *
1197 * All of the pages in the specified range [offset, offset + len - 1] must be
1198 * within a single window of the registered address space of the peer of epd.
1199 *
1200 * The addresses are returned as a virtually contiguous array pointed to by the
1201 * phys_addr component of the scif_range structure whose address is returned in
1202 * pages. The nr_pages component of scif_range is the length of the array. The
1203 * prot_flags component of scif_range holds the protection flag value passed
1204 * when the pages were registered.
1205 *
1206 * Each physical page whose address is returned by scif_get_pages() remains
1207 * available and will not be released for reuse until the scif_range structure
1208 * is returned in a call to scif_put_pages(). The scif_range structure returned
1209 * by scif_get_pages() must be unmodified.
1210 *
1211 * It is an error to call scif_close() on an endpoint on which a scif_range
1212 * structure of that endpoint has not been returned to scif_put_pages().
1213 *
1214 * Return:
1215 * Upon successful completion, scif_get_pages() returns 0; otherwise the
1216 * negative of one of the following errors is returned.
1217 * Errors:
1218 * ECONNRESET - Connection reset by peer.
1219 * EINVAL - offset is not a multiple of the page size, or offset is negative, or
1220 * len is not a multiple of the page size
1221 * ENODEV - The remote node is lost or existed, but is not currently in the
1222 * network since it may have crashed
1223 * ENOTCONN - The endpoint is not connected
1224 * ENXIO - Offsets in the range [offset, offset + len - 1] are invalid
1225 * for the registered address space of the peer epd
1226 */
1227int scif_get_pages(scif_epd_t epd, off_t offset, size_t len,
1228 struct scif_range **pages);
1229
1230/**
1231 * scif_put_pages() - Remove references from remote registered pages
1232 * @pages: pages to be returned
1233 *
1234 * scif_put_pages() releases a scif_range structure previously obtained by
1235 * calling scif_get_pages(). The physical pages represented by pages may
1236 * be reused when the window which represented those pages is unregistered.
1237 * Therefore, those pages must not be accessed after calling scif_put_pages().
1238 *
1239 * Return:
1240 * Upon successful completion, scif_put_pages() returns 0; otherwise the
1241 * negative of one of the following errors is returned.
1242 * Errors:
1243 * EINVAL - pages does not point to a valid scif_range structure, or
1244 * the scif_range structure pointed to by pages was already returned
1245 * ENODEV - The remote node is lost or existed, but is not currently in the
1246 * network since it may have crashed
1247 * ENOTCONN - The endpoint is not connected
1248 */
1249int scif_put_pages(struct scif_range *pages);
1250
1251/**
1252 * scif_poll() - Wait for some event on an endpoint
1253 * @epds: Array of endpoint descriptors
1254 * @nepds: Length of epds
1255 * @timeout: Upper limit on time for which scif_poll() will block
1256 *
1257 * scif_poll() waits for one of a set of endpoints to become ready to perform
1258 * an I/O operation.
1259 *
1260 * The epds argument specifies the endpoint descriptors to be examined and the
1261 * events of interest for each endpoint descriptor. epds is a pointer to an
1262 * array with one member for each open endpoint descriptor of interest.
1263 *
1264 * The number of items in the epds array is specified in nepds. The epd field
1265 * of scif_pollepd is an endpoint descriptor of an open endpoint. The field
1266 * events is a bitmask specifying the events which the application is
1267 * interested in. The field revents is an output parameter, filled by the
1268 * kernel with the events that actually occurred. The bits returned in revents
1269 * can include any of those specified in events, or one of the values EPOLLERR,
1270 * EPOLLHUP, or EPOLLNVAL. (These three bits are meaningless in the events
1271 * field, and will be set in the revents field whenever the corresponding
1272 * condition is true.)
1273 *
1274 * If none of the events requested (and no error) has occurred for any of the
1275 * endpoint descriptors, then scif_poll() blocks until one of the events occurs.
1276 *
1277 * The timeout argument specifies an upper limit on the time for which
1278 * scif_poll() will block, in milliseconds. Specifying a negative value in
1279 * timeout means an infinite timeout.
1280 *
1281 * The following bits may be set in events and returned in revents.
1282 * EPOLLIN - Data may be received without blocking. For a connected
1283 * endpoint, this means that scif_recv() may be called without blocking. For a
1284 * listening endpoint, this means that scif_accept() may be called without
1285 * blocking.
1286 * EPOLLOUT - Data may be sent without blocking. For a connected endpoint, this
1287 * means that scif_send() may be called without blocking. EPOLLOUT may also be
1288 * used to block waiting for a non-blocking connect to complete. This bit value
1289 * has no meaning for a listening endpoint and is ignored if specified.
1290 *
1291 * The following bits are only returned in revents, and are ignored if set in
1292 * events.
1293 * EPOLLERR - An error occurred on the endpoint
1294 * EPOLLHUP - The connection to the peer endpoint was disconnected
1295 * EPOLLNVAL - The specified endpoint descriptor is invalid.
1296 *
1297 * Return:
1298 * Upon successful completion, scif_poll() returns a non-negative value. A
1299 * positive value indicates the total number of endpoint descriptors that have
1300 * been selected (that is, endpoint descriptors for which the revents member is
1301 * non-zero). A value of 0 indicates that the call timed out and no endpoint
1302 * descriptors have been selected. Otherwise in user mode -1 is returned and
1303 * errno is set to indicate the error; in kernel mode the negative of one of
1304 * the following errors is returned.
1305 *
1306 * Errors:
1307 * EINTR - A signal occurred before any requested event
1308 * EINVAL - The nepds argument is greater than {OPEN_MAX}
1309 * ENOMEM - There was no space to allocate file descriptor tables
1310 */
1311int scif_poll(struct scif_pollepd *epds, unsigned int nepds, long timeout);
1312
1313/**
1314 * scif_client_register() - Register a SCIF client
1315 * @client: client to be registered
1316 *
1317 * scif_client_register() registers a SCIF client. The probe() method
1318 * of the client is called when SCIF peer devices come online and the
1319 * remove() method is called when the peer devices disappear.
1320 *
1321 * Return:
1322 * Upon successful completion, scif_client_register() returns a non-negative
1323 * value. Otherwise the return value is the same as subsys_interface_register()
1324 * in the kernel.
1325 */
1326int scif_client_register(struct scif_client *client);
1327
1328/**
1329 * scif_client_unregister() - Unregister a SCIF client
1330 * @client: client to be unregistered
1331 *
1332 * scif_client_unregister() unregisters a SCIF client.
1333 *
1334 * Return:
1335 * None
1336 */
1337void scif_client_unregister(struct scif_client *client);
1338
1339#endif /* __SCIF_H__ */
1340