xref: /openbmc/linux/drivers/infiniband/hw/hfi1/sdma.h (revision d9fd5a71)
1 #ifndef _HFI1_SDMA_H
2 #define _HFI1_SDMA_H
3 /*
4  * Copyright(c) 2015 - 2018 Intel Corporation.
5  *
6  * This file is provided under a dual BSD/GPLv2 license.  When using or
7  * redistributing this file, you may do so under either license.
8  *
9  * GPL LICENSE SUMMARY
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  * Redistribution and use in source and binary forms, with or without
23  * modification, are permitted provided that the following conditions
24  * are met:
25  *
26  *  - Redistributions of source code must retain the above copyright
27  *    notice, this list of conditions and the following disclaimer.
28  *  - Redistributions in binary form must reproduce the above copyright
29  *    notice, this list of conditions and the following disclaimer in
30  *    the documentation and/or other materials provided with the
31  *    distribution.
32  *  - Neither the name of Intel Corporation nor the names of its
33  *    contributors may be used to endorse or promote products derived
34  *    from this software without specific prior written permission.
35  *
36  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
37  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
38  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
39  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
40  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
41  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
42  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
43  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
44  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
45  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
46  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
47  *
48  */
49 
50 #include <linux/types.h>
51 #include <linux/list.h>
52 #include <asm/byteorder.h>
53 #include <linux/workqueue.h>
54 #include <linux/rculist.h>
55 
56 #include "hfi.h"
57 #include "verbs.h"
58 #include "sdma_txreq.h"
59 
60 /* Hardware limit */
61 #define MAX_DESC 64
62 /* Hardware limit for SDMA packet size */
63 #define MAX_SDMA_PKT_SIZE ((16 * 1024) - 1)
64 
65 #define SDMA_MAP_NONE          0
66 #define SDMA_MAP_SINGLE        1
67 #define SDMA_MAP_PAGE          2
68 
69 #define SDMA_AHG_VALUE_MASK          0xffff
70 #define SDMA_AHG_VALUE_SHIFT         0
71 #define SDMA_AHG_INDEX_MASK          0xf
72 #define SDMA_AHG_INDEX_SHIFT         16
73 #define SDMA_AHG_FIELD_LEN_MASK      0xf
74 #define SDMA_AHG_FIELD_LEN_SHIFT     20
75 #define SDMA_AHG_FIELD_START_MASK    0x1f
76 #define SDMA_AHG_FIELD_START_SHIFT   24
77 #define SDMA_AHG_UPDATE_ENABLE_MASK  0x1
78 #define SDMA_AHG_UPDATE_ENABLE_SHIFT 31
79 
80 /* AHG modes */
81 
82 /*
83  * Be aware the ordering and values
84  * for SDMA_AHG_APPLY_UPDATE[123]
85  * are assumed in generating a skip
86  * count in submit_tx() in sdma.c
87  */
88 #define SDMA_AHG_NO_AHG              0
89 #define SDMA_AHG_COPY                1
90 #define SDMA_AHG_APPLY_UPDATE1       2
91 #define SDMA_AHG_APPLY_UPDATE2       3
92 #define SDMA_AHG_APPLY_UPDATE3       4
93 
94 /*
95  * Bits defined in the send DMA descriptor.
96  */
97 #define SDMA_DESC0_FIRST_DESC_FLAG      BIT_ULL(63)
98 #define SDMA_DESC0_LAST_DESC_FLAG       BIT_ULL(62)
99 #define SDMA_DESC0_BYTE_COUNT_SHIFT     48
100 #define SDMA_DESC0_BYTE_COUNT_WIDTH     14
101 #define SDMA_DESC0_BYTE_COUNT_MASK \
102 	((1ULL << SDMA_DESC0_BYTE_COUNT_WIDTH) - 1)
103 #define SDMA_DESC0_BYTE_COUNT_SMASK \
104 	(SDMA_DESC0_BYTE_COUNT_MASK << SDMA_DESC0_BYTE_COUNT_SHIFT)
105 #define SDMA_DESC0_PHY_ADDR_SHIFT       0
106 #define SDMA_DESC0_PHY_ADDR_WIDTH       48
107 #define SDMA_DESC0_PHY_ADDR_MASK \
108 	((1ULL << SDMA_DESC0_PHY_ADDR_WIDTH) - 1)
109 #define SDMA_DESC0_PHY_ADDR_SMASK \
110 	(SDMA_DESC0_PHY_ADDR_MASK << SDMA_DESC0_PHY_ADDR_SHIFT)
111 
112 #define SDMA_DESC1_HEADER_UPDATE1_SHIFT 32
113 #define SDMA_DESC1_HEADER_UPDATE1_WIDTH 32
114 #define SDMA_DESC1_HEADER_UPDATE1_MASK \
115 	((1ULL << SDMA_DESC1_HEADER_UPDATE1_WIDTH) - 1)
116 #define SDMA_DESC1_HEADER_UPDATE1_SMASK \
117 	(SDMA_DESC1_HEADER_UPDATE1_MASK << SDMA_DESC1_HEADER_UPDATE1_SHIFT)
118 #define SDMA_DESC1_HEADER_MODE_SHIFT    13
119 #define SDMA_DESC1_HEADER_MODE_WIDTH    3
120 #define SDMA_DESC1_HEADER_MODE_MASK \
121 	((1ULL << SDMA_DESC1_HEADER_MODE_WIDTH) - 1)
122 #define SDMA_DESC1_HEADER_MODE_SMASK \
123 	(SDMA_DESC1_HEADER_MODE_MASK << SDMA_DESC1_HEADER_MODE_SHIFT)
124 #define SDMA_DESC1_HEADER_INDEX_SHIFT   8
125 #define SDMA_DESC1_HEADER_INDEX_WIDTH   5
126 #define SDMA_DESC1_HEADER_INDEX_MASK \
127 	((1ULL << SDMA_DESC1_HEADER_INDEX_WIDTH) - 1)
128 #define SDMA_DESC1_HEADER_INDEX_SMASK \
129 	(SDMA_DESC1_HEADER_INDEX_MASK << SDMA_DESC1_HEADER_INDEX_SHIFT)
130 #define SDMA_DESC1_HEADER_DWS_SHIFT     4
131 #define SDMA_DESC1_HEADER_DWS_WIDTH     4
132 #define SDMA_DESC1_HEADER_DWS_MASK \
133 	((1ULL << SDMA_DESC1_HEADER_DWS_WIDTH) - 1)
134 #define SDMA_DESC1_HEADER_DWS_SMASK \
135 	(SDMA_DESC1_HEADER_DWS_MASK << SDMA_DESC1_HEADER_DWS_SHIFT)
136 #define SDMA_DESC1_GENERATION_SHIFT     2
137 #define SDMA_DESC1_GENERATION_WIDTH     2
138 #define SDMA_DESC1_GENERATION_MASK \
139 	((1ULL << SDMA_DESC1_GENERATION_WIDTH) - 1)
140 #define SDMA_DESC1_GENERATION_SMASK \
141 	(SDMA_DESC1_GENERATION_MASK << SDMA_DESC1_GENERATION_SHIFT)
142 #define SDMA_DESC1_INT_REQ_FLAG         BIT_ULL(1)
143 #define SDMA_DESC1_HEAD_TO_HOST_FLAG    BIT_ULL(0)
144 
145 enum sdma_states {
146 	sdma_state_s00_hw_down,
147 	sdma_state_s10_hw_start_up_halt_wait,
148 	sdma_state_s15_hw_start_up_clean_wait,
149 	sdma_state_s20_idle,
150 	sdma_state_s30_sw_clean_up_wait,
151 	sdma_state_s40_hw_clean_up_wait,
152 	sdma_state_s50_hw_halt_wait,
153 	sdma_state_s60_idle_halt_wait,
154 	sdma_state_s80_hw_freeze,
155 	sdma_state_s82_freeze_sw_clean,
156 	sdma_state_s99_running,
157 };
158 
159 enum sdma_events {
160 	sdma_event_e00_go_hw_down,
161 	sdma_event_e10_go_hw_start,
162 	sdma_event_e15_hw_halt_done,
163 	sdma_event_e25_hw_clean_up_done,
164 	sdma_event_e30_go_running,
165 	sdma_event_e40_sw_cleaned,
166 	sdma_event_e50_hw_cleaned,
167 	sdma_event_e60_hw_halted,
168 	sdma_event_e70_go_idle,
169 	sdma_event_e80_hw_freeze,
170 	sdma_event_e81_hw_frozen,
171 	sdma_event_e82_hw_unfreeze,
172 	sdma_event_e85_link_down,
173 	sdma_event_e90_sw_halted,
174 };
175 
176 struct sdma_set_state_action {
177 	unsigned op_enable:1;
178 	unsigned op_intenable:1;
179 	unsigned op_halt:1;
180 	unsigned op_cleanup:1;
181 	unsigned go_s99_running_tofalse:1;
182 	unsigned go_s99_running_totrue:1;
183 };
184 
185 struct sdma_state {
186 	struct kref          kref;
187 	struct completion    comp;
188 	enum sdma_states current_state;
189 	unsigned             current_op;
190 	unsigned             go_s99_running;
191 	/* debugging/development */
192 	enum sdma_states previous_state;
193 	unsigned             previous_op;
194 	enum sdma_events last_event;
195 };
196 
197 /**
198  * DOC: sdma exported routines
199  *
200  * These sdma routines fit into three categories:
201  * - The SDMA API for building and submitting packets
202  *   to the ring
203  *
204  * - Initialization and tear down routines to buildup
205  *   and tear down SDMA
206  *
207  * - ISR entrances to handle interrupts, state changes
208  *   and errors
209  */
210 
211 /**
212  * DOC: sdma PSM/verbs API
213  *
214  * The sdma API is designed to be used by both PSM
215  * and verbs to supply packets to the SDMA ring.
216  *
217  * The usage of the API is as follows:
218  *
219  * Embed a struct iowait in the QP or
220  * PQ.  The iowait should be initialized with a
221  * call to iowait_init().
222  *
223  * The user of the API should create an allocation method
224  * for their version of the txreq. slabs, pre-allocated lists,
225  * and dma pools can be used.  Once the user's overload of
226  * the sdma_txreq has been allocated, the sdma_txreq member
227  * must be initialized with sdma_txinit() or sdma_txinit_ahg().
228  *
229  * The txreq must be declared with the sdma_txreq first.
230  *
231  * The tx request, once initialized,  is manipulated with calls to
232  * sdma_txadd_daddr(), sdma_txadd_page(), or sdma_txadd_kvaddr()
233  * for each disjoint memory location.  It is the user's responsibility
234  * to understand the packet boundaries and page boundaries to do the
235  * appropriate number of sdma_txadd_* calls..  The user
236  * must be prepared to deal with failures from these routines due to
237  * either memory allocation or dma_mapping failures.
238  *
239  * The mapping specifics for each memory location are recorded
240  * in the tx. Memory locations added with sdma_txadd_page()
241  * and sdma_txadd_kvaddr() are automatically mapped when added
242  * to the tx and nmapped as part of the progress processing in the
243  * SDMA interrupt handling.
244  *
245  * sdma_txadd_daddr() is used to add an dma_addr_t memory to the
246  * tx.   An example of a use case would be a pre-allocated
247  * set of headers allocated via dma_pool_alloc() or
248  * dma_alloc_coherent().  For these memory locations, it
249  * is the responsibility of the user to handle that unmapping.
250  * (This would usually be at an unload or job termination.)
251  *
252  * The routine sdma_send_txreq() is used to submit
253  * a tx to the ring after the appropriate number of
254  * sdma_txadd_* have been done.
255  *
256  * If it is desired to send a burst of sdma_txreqs, sdma_send_txlist()
257  * can be used to submit a list of packets.
258  *
259  * The user is free to use the link overhead in the struct sdma_txreq as
260  * long as the tx isn't in flight.
261  *
262  * The extreme degenerate case of the number of descriptors
263  * exceeding the ring size is automatically handled as
264  * memory locations are added.  An overflow of the descriptor
265  * array that is part of the sdma_txreq is also automatically
266  * handled.
267  *
268  */
269 
270 /**
271  * DOC: Infrastructure calls
272  *
273  * sdma_init() is used to initialize data structures and
274  * CSRs for the desired number of SDMA engines.
275  *
276  * sdma_start() is used to kick the SDMA engines initialized
277  * with sdma_init().   Interrupts must be enabled at this
278  * point since aspects of the state machine are interrupt
279  * driven.
280  *
281  * sdma_engine_error() and sdma_engine_interrupt() are
282  * entrances for interrupts.
283  *
284  * sdma_map_init() is for the management of the mapping
285  * table when the number of vls is changed.
286  *
287  */
288 
289 /*
290  * struct hw_sdma_desc - raw 128 bit SDMA descriptor
291  *
292  * This is the raw descriptor in the SDMA ring
293  */
294 struct hw_sdma_desc {
295 	/* private:  don't use directly */
296 	__le64 qw[2];
297 };
298 
299 /**
300  * struct sdma_engine - Data pertaining to each SDMA engine.
301  * @dd: a back-pointer to the device data
302  * @ppd: per port back-pointer
303  * @imask: mask for irq manipulation
304  * @idle_mask: mask for determining if an interrupt is due to sdma_idle
305  *
306  * This structure has the state for each sdma_engine.
307  *
308  * Accessing to non public fields are not supported
309  * since the private members are subject to change.
310  */
311 struct sdma_engine {
312 	/* read mostly */
313 	struct hfi1_devdata *dd;
314 	struct hfi1_pportdata *ppd;
315 	/* private: */
316 	void __iomem *tail_csr;
317 	u64 imask;			/* clear interrupt mask */
318 	u64 idle_mask;
319 	u64 progress_mask;
320 	u64 int_mask;
321 	/* private: */
322 	volatile __le64      *head_dma; /* DMA'ed by chip */
323 	/* private: */
324 	dma_addr_t            head_phys;
325 	/* private: */
326 	struct hw_sdma_desc *descq;
327 	/* private: */
328 	unsigned descq_full_count;
329 	struct sdma_txreq **tx_ring;
330 	/* private: */
331 	dma_addr_t            descq_phys;
332 	/* private */
333 	u32 sdma_mask;
334 	/* private */
335 	struct sdma_state state;
336 	/* private */
337 	int cpu;
338 	/* private: */
339 	u8 sdma_shift;
340 	/* private: */
341 	u8 this_idx; /* zero relative engine */
342 	/* protect changes to senddmactrl shadow */
343 	spinlock_t senddmactrl_lock;
344 	/* private: */
345 	u64 p_senddmactrl;		/* shadow per-engine SendDmaCtrl */
346 
347 	/* read/write using tail_lock */
348 	spinlock_t            tail_lock ____cacheline_aligned_in_smp;
349 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
350 	/* private: */
351 	u64                   tail_sn;
352 #endif
353 	/* private: */
354 	u32                   descq_tail;
355 	/* private: */
356 	unsigned long         ahg_bits;
357 	/* private: */
358 	u16                   desc_avail;
359 	/* private: */
360 	u16                   tx_tail;
361 	/* private: */
362 	u16 descq_cnt;
363 
364 	/* read/write using head_lock */
365 	/* private: */
366 	seqlock_t            head_lock ____cacheline_aligned_in_smp;
367 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
368 	/* private: */
369 	u64                   head_sn;
370 #endif
371 	/* private: */
372 	u32                   descq_head;
373 	/* private: */
374 	u16                   tx_head;
375 	/* private: */
376 	u64                   last_status;
377 	/* private */
378 	u64                     err_cnt;
379 	/* private */
380 	u64                     sdma_int_cnt;
381 	u64                     idle_int_cnt;
382 	u64                     progress_int_cnt;
383 
384 	/* private: */
385 	seqlock_t            waitlock;
386 	struct list_head      dmawait;
387 
388 	/* CONFIG SDMA for now, just blindly duplicate */
389 	/* private: */
390 	struct tasklet_struct sdma_hw_clean_up_task
391 		____cacheline_aligned_in_smp;
392 
393 	/* private: */
394 	struct tasklet_struct sdma_sw_clean_up_task
395 		____cacheline_aligned_in_smp;
396 	/* private: */
397 	struct work_struct err_halt_worker;
398 	/* private */
399 	struct timer_list     err_progress_check_timer;
400 	u32                   progress_check_head;
401 	/* private: */
402 	struct work_struct flush_worker;
403 	/* protect flush list */
404 	spinlock_t flushlist_lock;
405 	/* private: */
406 	struct list_head flushlist;
407 	struct cpumask cpu_mask;
408 	struct kobject kobj;
409 	u32 msix_intr;
410 };
411 
412 int sdma_init(struct hfi1_devdata *dd, u8 port);
413 void sdma_start(struct hfi1_devdata *dd);
414 void sdma_exit(struct hfi1_devdata *dd);
415 void sdma_clean(struct hfi1_devdata *dd, size_t num_engines);
416 void sdma_all_running(struct hfi1_devdata *dd);
417 void sdma_all_idle(struct hfi1_devdata *dd);
418 void sdma_freeze_notify(struct hfi1_devdata *dd, int go_idle);
419 void sdma_freeze(struct hfi1_devdata *dd);
420 void sdma_unfreeze(struct hfi1_devdata *dd);
421 void sdma_wait(struct hfi1_devdata *dd);
422 
423 /**
424  * sdma_empty() - idle engine test
425  * @engine: sdma engine
426  *
427  * Currently used by verbs as a latency optimization.
428  *
429  * Return:
430  * 1 - empty, 0 - non-empty
431  */
432 static inline int sdma_empty(struct sdma_engine *sde)
433 {
434 	return sde->descq_tail == sde->descq_head;
435 }
436 
437 static inline u16 sdma_descq_freecnt(struct sdma_engine *sde)
438 {
439 	return sde->descq_cnt -
440 		(sde->descq_tail -
441 		 READ_ONCE(sde->descq_head)) - 1;
442 }
443 
444 static inline u16 sdma_descq_inprocess(struct sdma_engine *sde)
445 {
446 	return sde->descq_cnt - sdma_descq_freecnt(sde);
447 }
448 
449 /*
450  * Either head_lock or tail lock required to see
451  * a steady state.
452  */
453 static inline int __sdma_running(struct sdma_engine *engine)
454 {
455 	return engine->state.current_state == sdma_state_s99_running;
456 }
457 
458 /**
459  * sdma_running() - state suitability test
460  * @engine: sdma engine
461  *
462  * sdma_running probes the internal state to determine if it is suitable
463  * for submitting packets.
464  *
465  * Return:
466  * 1 - ok to submit, 0 - not ok to submit
467  *
468  */
469 static inline int sdma_running(struct sdma_engine *engine)
470 {
471 	unsigned long flags;
472 	int ret;
473 
474 	spin_lock_irqsave(&engine->tail_lock, flags);
475 	ret = __sdma_running(engine);
476 	spin_unlock_irqrestore(&engine->tail_lock, flags);
477 	return ret;
478 }
479 
480 void _sdma_txreq_ahgadd(
481 	struct sdma_txreq *tx,
482 	u8 num_ahg,
483 	u8 ahg_entry,
484 	u32 *ahg,
485 	u8 ahg_hlen);
486 
487 /**
488  * sdma_txinit_ahg() - initialize an sdma_txreq struct with AHG
489  * @tx: tx request to initialize
490  * @flags: flags to key last descriptor additions
491  * @tlen: total packet length (pbc + headers + data)
492  * @ahg_entry: ahg entry to use  (0 - 31)
493  * @num_ahg: ahg descriptor for first descriptor (0 - 9)
494  * @ahg: array of AHG descriptors (up to 9 entries)
495  * @ahg_hlen: number of bytes from ASIC entry to use
496  * @cb: callback
497  *
498  * The allocation of the sdma_txreq and it enclosing structure is user
499  * dependent.  This routine must be called to initialize the user independent
500  * fields.
501  *
502  * The currently supported flags are SDMA_TXREQ_F_URGENT,
503  * SDMA_TXREQ_F_AHG_COPY, and SDMA_TXREQ_F_USE_AHG.
504  *
505  * SDMA_TXREQ_F_URGENT is used for latency sensitive situations where the
506  * completion is desired as soon as possible.
507  *
508  * SDMA_TXREQ_F_AHG_COPY causes the header in the first descriptor to be
509  * copied to chip entry. SDMA_TXREQ_F_USE_AHG causes the code to add in
510  * the AHG descriptors into the first 1 to 3 descriptors.
511  *
512  * Completions of submitted requests can be gotten on selected
513  * txreqs by giving a completion routine callback to sdma_txinit() or
514  * sdma_txinit_ahg().  The environment in which the callback runs
515  * can be from an ISR, a tasklet, or a thread, so no sleeping
516  * kernel routines can be used.   Aspects of the sdma ring may
517  * be locked so care should be taken with locking.
518  *
519  * The callback pointer can be NULL to avoid any callback for the packet
520  * being submitted. The callback will be provided this tx, a status, and a flag.
521  *
522  * The status will be one of SDMA_TXREQ_S_OK, SDMA_TXREQ_S_SENDERROR,
523  * SDMA_TXREQ_S_ABORTED, or SDMA_TXREQ_S_SHUTDOWN.
524  *
525  * The flag, if the is the iowait had been used, indicates the iowait
526  * sdma_busy count has reached zero.
527  *
528  * user data portion of tlen should be precise.   The sdma_txadd_* entrances
529  * will pad with a descriptor references 1 - 3 bytes when the number of bytes
530  * specified in tlen have been supplied to the sdma_txreq.
531  *
532  * ahg_hlen is used to determine the number of on-chip entry bytes to
533  * use as the header.   This is for cases where the stored header is
534  * larger than the header to be used in a packet.  This is typical
535  * for verbs where an RDMA_WRITE_FIRST is larger than the packet in
536  * and RDMA_WRITE_MIDDLE.
537  *
538  */
539 static inline int sdma_txinit_ahg(
540 	struct sdma_txreq *tx,
541 	u16 flags,
542 	u16 tlen,
543 	u8 ahg_entry,
544 	u8 num_ahg,
545 	u32 *ahg,
546 	u8 ahg_hlen,
547 	void (*cb)(struct sdma_txreq *, int))
548 {
549 	if (tlen == 0)
550 		return -ENODATA;
551 	if (tlen > MAX_SDMA_PKT_SIZE)
552 		return -EMSGSIZE;
553 	tx->desc_limit = ARRAY_SIZE(tx->descs);
554 	tx->descp = &tx->descs[0];
555 	INIT_LIST_HEAD(&tx->list);
556 	tx->num_desc = 0;
557 	tx->flags = flags;
558 	tx->complete = cb;
559 	tx->coalesce_buf = NULL;
560 	tx->wait = NULL;
561 	tx->packet_len = tlen;
562 	tx->tlen = tx->packet_len;
563 	tx->descs[0].qw[0] = SDMA_DESC0_FIRST_DESC_FLAG;
564 	tx->descs[0].qw[1] = 0;
565 	if (flags & SDMA_TXREQ_F_AHG_COPY)
566 		tx->descs[0].qw[1] |=
567 			(((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK)
568 				<< SDMA_DESC1_HEADER_INDEX_SHIFT) |
569 			(((u64)SDMA_AHG_COPY & SDMA_DESC1_HEADER_MODE_MASK)
570 				<< SDMA_DESC1_HEADER_MODE_SHIFT);
571 	else if (flags & SDMA_TXREQ_F_USE_AHG && num_ahg)
572 		_sdma_txreq_ahgadd(tx, num_ahg, ahg_entry, ahg, ahg_hlen);
573 	return 0;
574 }
575 
576 /**
577  * sdma_txinit() - initialize an sdma_txreq struct (no AHG)
578  * @tx: tx request to initialize
579  * @flags: flags to key last descriptor additions
580  * @tlen: total packet length (pbc + headers + data)
581  * @cb: callback pointer
582  *
583  * The allocation of the sdma_txreq and it enclosing structure is user
584  * dependent.  This routine must be called to initialize the user
585  * independent fields.
586  *
587  * The currently supported flags is SDMA_TXREQ_F_URGENT.
588  *
589  * SDMA_TXREQ_F_URGENT is used for latency sensitive situations where the
590  * completion is desired as soon as possible.
591  *
592  * Completions of submitted requests can be gotten on selected
593  * txreqs by giving a completion routine callback to sdma_txinit() or
594  * sdma_txinit_ahg().  The environment in which the callback runs
595  * can be from an ISR, a tasklet, or a thread, so no sleeping
596  * kernel routines can be used.   The head size of the sdma ring may
597  * be locked so care should be taken with locking.
598  *
599  * The callback pointer can be NULL to avoid any callback for the packet
600  * being submitted.
601  *
602  * The callback, if non-NULL,  will be provided this tx and a status.  The
603  * status will be one of SDMA_TXREQ_S_OK, SDMA_TXREQ_S_SENDERROR,
604  * SDMA_TXREQ_S_ABORTED, or SDMA_TXREQ_S_SHUTDOWN.
605  *
606  */
607 static inline int sdma_txinit(
608 	struct sdma_txreq *tx,
609 	u16 flags,
610 	u16 tlen,
611 	void (*cb)(struct sdma_txreq *, int))
612 {
613 	return sdma_txinit_ahg(tx, flags, tlen, 0, 0, NULL, 0, cb);
614 }
615 
616 /* helpers - don't use */
617 static inline int sdma_mapping_type(struct sdma_desc *d)
618 {
619 	return (d->qw[1] & SDMA_DESC1_GENERATION_SMASK)
620 		>> SDMA_DESC1_GENERATION_SHIFT;
621 }
622 
623 static inline size_t sdma_mapping_len(struct sdma_desc *d)
624 {
625 	return (d->qw[0] & SDMA_DESC0_BYTE_COUNT_SMASK)
626 		>> SDMA_DESC0_BYTE_COUNT_SHIFT;
627 }
628 
629 static inline dma_addr_t sdma_mapping_addr(struct sdma_desc *d)
630 {
631 	return (d->qw[0] & SDMA_DESC0_PHY_ADDR_SMASK)
632 		>> SDMA_DESC0_PHY_ADDR_SHIFT;
633 }
634 
635 static inline void make_tx_sdma_desc(
636 	struct sdma_txreq *tx,
637 	int type,
638 	dma_addr_t addr,
639 	size_t len)
640 {
641 	struct sdma_desc *desc = &tx->descp[tx->num_desc];
642 
643 	if (!tx->num_desc) {
644 		/* qw[0] zero; qw[1] first, ahg mode already in from init */
645 		desc->qw[1] |= ((u64)type & SDMA_DESC1_GENERATION_MASK)
646 				<< SDMA_DESC1_GENERATION_SHIFT;
647 	} else {
648 		desc->qw[0] = 0;
649 		desc->qw[1] = ((u64)type & SDMA_DESC1_GENERATION_MASK)
650 				<< SDMA_DESC1_GENERATION_SHIFT;
651 	}
652 	desc->qw[0] |= (((u64)addr & SDMA_DESC0_PHY_ADDR_MASK)
653 				<< SDMA_DESC0_PHY_ADDR_SHIFT) |
654 			(((u64)len & SDMA_DESC0_BYTE_COUNT_MASK)
655 				<< SDMA_DESC0_BYTE_COUNT_SHIFT);
656 }
657 
658 /* helper to extend txreq */
659 int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx,
660 			   int type, void *kvaddr, struct page *page,
661 			   unsigned long offset, u16 len);
662 int _pad_sdma_tx_descs(struct hfi1_devdata *, struct sdma_txreq *);
663 void __sdma_txclean(struct hfi1_devdata *, struct sdma_txreq *);
664 
665 static inline void sdma_txclean(struct hfi1_devdata *dd, struct sdma_txreq *tx)
666 {
667 	if (tx->num_desc)
668 		__sdma_txclean(dd, tx);
669 }
670 
671 /* helpers used by public routines */
672 static inline void _sdma_close_tx(struct hfi1_devdata *dd,
673 				  struct sdma_txreq *tx)
674 {
675 	tx->descp[tx->num_desc].qw[0] |=
676 		SDMA_DESC0_LAST_DESC_FLAG;
677 	tx->descp[tx->num_desc].qw[1] |=
678 		dd->default_desc1;
679 	if (tx->flags & SDMA_TXREQ_F_URGENT)
680 		tx->descp[tx->num_desc].qw[1] |=
681 			(SDMA_DESC1_HEAD_TO_HOST_FLAG |
682 			 SDMA_DESC1_INT_REQ_FLAG);
683 }
684 
685 static inline int _sdma_txadd_daddr(
686 	struct hfi1_devdata *dd,
687 	int type,
688 	struct sdma_txreq *tx,
689 	dma_addr_t addr,
690 	u16 len)
691 {
692 	int rval = 0;
693 
694 	make_tx_sdma_desc(
695 		tx,
696 		type,
697 		addr, len);
698 	WARN_ON(len > tx->tlen);
699 	tx->tlen -= len;
700 	/* special cases for last */
701 	if (!tx->tlen) {
702 		if (tx->packet_len & (sizeof(u32) - 1)) {
703 			rval = _pad_sdma_tx_descs(dd, tx);
704 			if (rval)
705 				return rval;
706 		} else {
707 			_sdma_close_tx(dd, tx);
708 		}
709 	}
710 	tx->num_desc++;
711 	return rval;
712 }
713 
714 /**
715  * sdma_txadd_page() - add a page to the sdma_txreq
716  * @dd: the device to use for mapping
717  * @tx: tx request to which the page is added
718  * @page: page to map
719  * @offset: offset within the page
720  * @len: length in bytes
721  *
722  * This is used to add a page/offset/length descriptor.
723  *
724  * The mapping/unmapping of the page/offset/len is automatically handled.
725  *
726  * Return:
727  * 0 - success, -ENOSPC - mapping fail, -ENOMEM - couldn't
728  * extend/coalesce descriptor array
729  */
730 static inline int sdma_txadd_page(
731 	struct hfi1_devdata *dd,
732 	struct sdma_txreq *tx,
733 	struct page *page,
734 	unsigned long offset,
735 	u16 len)
736 {
737 	dma_addr_t addr;
738 	int rval;
739 
740 	if ((unlikely(tx->num_desc == tx->desc_limit))) {
741 		rval = ext_coal_sdma_tx_descs(dd, tx, SDMA_MAP_PAGE,
742 					      NULL, page, offset, len);
743 		if (rval <= 0)
744 			return rval;
745 	}
746 
747 	addr = dma_map_page(
748 		       &dd->pcidev->dev,
749 		       page,
750 		       offset,
751 		       len,
752 		       DMA_TO_DEVICE);
753 
754 	if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) {
755 		__sdma_txclean(dd, tx);
756 		return -ENOSPC;
757 	}
758 
759 	return _sdma_txadd_daddr(
760 			dd, SDMA_MAP_PAGE, tx, addr, len);
761 }
762 
763 /**
764  * sdma_txadd_daddr() - add a dma address to the sdma_txreq
765  * @dd: the device to use for mapping
766  * @tx: sdma_txreq to which the page is added
767  * @addr: dma address mapped by caller
768  * @len: length in bytes
769  *
770  * This is used to add a descriptor for memory that is already dma mapped.
771  *
772  * In this case, there is no unmapping as part of the progress processing for
773  * this memory location.
774  *
775  * Return:
776  * 0 - success, -ENOMEM - couldn't extend descriptor array
777  */
778 
779 static inline int sdma_txadd_daddr(
780 	struct hfi1_devdata *dd,
781 	struct sdma_txreq *tx,
782 	dma_addr_t addr,
783 	u16 len)
784 {
785 	int rval;
786 
787 	if ((unlikely(tx->num_desc == tx->desc_limit))) {
788 		rval = ext_coal_sdma_tx_descs(dd, tx, SDMA_MAP_NONE,
789 					      NULL, NULL, 0, 0);
790 		if (rval <= 0)
791 			return rval;
792 	}
793 
794 	return _sdma_txadd_daddr(dd, SDMA_MAP_NONE, tx, addr, len);
795 }
796 
797 /**
798  * sdma_txadd_kvaddr() - add a kernel virtual address to sdma_txreq
799  * @dd: the device to use for mapping
800  * @tx: sdma_txreq to which the page is added
801  * @kvaddr: the kernel virtual address
802  * @len: length in bytes
803  *
804  * This is used to add a descriptor referenced by the indicated kvaddr and
805  * len.
806  *
807  * The mapping/unmapping of the kvaddr and len is automatically handled.
808  *
809  * Return:
810  * 0 - success, -ENOSPC - mapping fail, -ENOMEM - couldn't extend/coalesce
811  * descriptor array
812  */
813 static inline int sdma_txadd_kvaddr(
814 	struct hfi1_devdata *dd,
815 	struct sdma_txreq *tx,
816 	void *kvaddr,
817 	u16 len)
818 {
819 	dma_addr_t addr;
820 	int rval;
821 
822 	if ((unlikely(tx->num_desc == tx->desc_limit))) {
823 		rval = ext_coal_sdma_tx_descs(dd, tx, SDMA_MAP_SINGLE,
824 					      kvaddr, NULL, 0, len);
825 		if (rval <= 0)
826 			return rval;
827 	}
828 
829 	addr = dma_map_single(
830 		       &dd->pcidev->dev,
831 		       kvaddr,
832 		       len,
833 		       DMA_TO_DEVICE);
834 
835 	if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) {
836 		__sdma_txclean(dd, tx);
837 		return -ENOSPC;
838 	}
839 
840 	return _sdma_txadd_daddr(
841 			dd, SDMA_MAP_SINGLE, tx, addr, len);
842 }
843 
844 struct iowait_work;
845 
846 int sdma_send_txreq(struct sdma_engine *sde,
847 		    struct iowait_work *wait,
848 		    struct sdma_txreq *tx,
849 		    bool pkts_sent);
850 int sdma_send_txlist(struct sdma_engine *sde,
851 		     struct iowait_work *wait,
852 		     struct list_head *tx_list,
853 		     u16 *count_out);
854 
855 int sdma_ahg_alloc(struct sdma_engine *sde);
856 void sdma_ahg_free(struct sdma_engine *sde, int ahg_index);
857 
858 /**
859  * sdma_build_ahg - build ahg descriptor
860  * @data
861  * @dwindex
862  * @startbit
863  * @bits
864  *
865  * Build and return a 32 bit descriptor.
866  */
867 static inline u32 sdma_build_ahg_descriptor(
868 	u16 data,
869 	u8 dwindex,
870 	u8 startbit,
871 	u8 bits)
872 {
873 	return (u32)(1UL << SDMA_AHG_UPDATE_ENABLE_SHIFT |
874 		((startbit & SDMA_AHG_FIELD_START_MASK) <<
875 		SDMA_AHG_FIELD_START_SHIFT) |
876 		((bits & SDMA_AHG_FIELD_LEN_MASK) <<
877 		SDMA_AHG_FIELD_LEN_SHIFT) |
878 		((dwindex & SDMA_AHG_INDEX_MASK) <<
879 		SDMA_AHG_INDEX_SHIFT) |
880 		((data & SDMA_AHG_VALUE_MASK) <<
881 		SDMA_AHG_VALUE_SHIFT));
882 }
883 
884 /**
885  * sdma_progress - use seq number of detect head progress
886  * @sde: sdma_engine to check
887  * @seq: base seq count
888  * @tx: txreq for which we need to check descriptor availability
889  *
890  * This is used in the appropriate spot in the sleep routine
891  * to check for potential ring progress.  This routine gets the
892  * seqcount before queuing the iowait structure for progress.
893  *
894  * If the seqcount indicates that progress needs to be checked,
895  * re-submission is detected by checking whether the descriptor
896  * queue has enough descriptor for the txreq.
897  */
898 static inline unsigned sdma_progress(struct sdma_engine *sde, unsigned seq,
899 				     struct sdma_txreq *tx)
900 {
901 	if (read_seqretry(&sde->head_lock, seq)) {
902 		sde->desc_avail = sdma_descq_freecnt(sde);
903 		if (tx->num_desc > sde->desc_avail)
904 			return 0;
905 		return 1;
906 	}
907 	return 0;
908 }
909 
910 /**
911  * sdma_iowait_schedule() - initialize wait structure
912  * @sde: sdma_engine to schedule
913  * @wait: wait struct to schedule
914  *
915  * This function initializes the iowait
916  * structure embedded in the QP or PQ.
917  *
918  */
919 static inline void sdma_iowait_schedule(
920 	struct sdma_engine *sde,
921 	struct iowait *wait)
922 {
923 	struct hfi1_pportdata *ppd = sde->dd->pport;
924 
925 	iowait_schedule(wait, ppd->hfi1_wq, sde->cpu);
926 }
927 
928 /* for use by interrupt handling */
929 void sdma_engine_error(struct sdma_engine *sde, u64 status);
930 void sdma_engine_interrupt(struct sdma_engine *sde, u64 status);
931 
932 /*
933  *
934  * The diagram below details the relationship of the mapping structures
935  *
936  * Since the mapping now allows for non-uniform engines per vl, the
937  * number of engines for a vl is either the vl_engines[vl] or
938  * a computation based on num_sdma/num_vls:
939  *
940  * For example:
941  * nactual = vl_engines ? vl_engines[vl] : num_sdma/num_vls
942  *
943  * n = roundup to next highest power of 2 using nactual
944  *
945  * In the case where there are num_sdma/num_vls doesn't divide
946  * evenly, the extras are added from the last vl downward.
947  *
948  * For the case where n > nactual, the engines are assigned
949  * in a round robin fashion wrapping back to the first engine
950  * for a particular vl.
951  *
952  *               dd->sdma_map
953  *                    |                                   sdma_map_elem[0]
954  *                    |                                +--------------------+
955  *                    v                                |       mask         |
956  *               sdma_vl_map                           |--------------------|
957  *      +--------------------------+                   | sde[0] -> eng 1    |
958  *      |    list (RCU)            |                   |--------------------|
959  *      |--------------------------|                 ->| sde[1] -> eng 2    |
960  *      |    mask                  |              --/  |--------------------|
961  *      |--------------------------|            -/     |        *           |
962  *      |    actual_vls (max 8)    |          -/       |--------------------|
963  *      |--------------------------|       --/         | sde[n-1] -> eng n  |
964  *      |    vls (max 8)           |     -/            +--------------------+
965  *      |--------------------------|  --/
966  *      |    map[0]                |-/
967  *      |--------------------------|                   +---------------------+
968  *      |    map[1]                |---                |       mask          |
969  *      |--------------------------|   \----           |---------------------|
970  *      |           *              |        \--        | sde[0] -> eng 1+n   |
971  *      |           *              |           \----   |---------------------|
972  *      |           *              |                \->| sde[1] -> eng 2+n   |
973  *      |--------------------------|                   |---------------------|
974  *      |   map[vls - 1]           |-                  |         *           |
975  *      +--------------------------+ \-                |---------------------|
976  *                                     \-              | sde[m-1] -> eng m+n |
977  *                                       \             +---------------------+
978  *                                        \-
979  *                                          \
980  *                                           \-        +----------------------+
981  *                                             \-      |       mask           |
982  *                                               \     |----------------------|
983  *                                                \-   | sde[0] -> eng 1+m+n  |
984  *                                                  \- |----------------------|
985  *                                                    >| sde[1] -> eng 2+m+n  |
986  *                                                     |----------------------|
987  *                                                     |         *            |
988  *                                                     |----------------------|
989  *                                                     | sde[o-1] -> eng o+m+n|
990  *                                                     +----------------------+
991  *
992  */
993 
994 /**
995  * struct sdma_map_elem - mapping for a vl
996  * @mask - selector mask
997  * @sde - array of engines for this vl
998  *
999  * The mask is used to "mod" the selector
1000  * to produce index into the trailing
1001  * array of sdes.
1002  */
1003 struct sdma_map_elem {
1004 	u32 mask;
1005 	struct sdma_engine *sde[];
1006 };
1007 
1008 /**
1009  * struct sdma_map_el - mapping for a vl
1010  * @engine_to_vl - map of an engine to a vl
1011  * @list - rcu head for free callback
1012  * @mask - vl mask to "mod" the vl to produce an index to map array
1013  * @actual_vls - number of vls
1014  * @vls - number of vls rounded to next power of 2
1015  * @map - array of sdma_map_elem entries
1016  *
1017  * This is the parent mapping structure.  The trailing
1018  * members of the struct point to sdma_map_elem entries, which
1019  * in turn point to an array of sde's for that vl.
1020  */
1021 struct sdma_vl_map {
1022 	s8 engine_to_vl[TXE_NUM_SDMA_ENGINES];
1023 	struct rcu_head list;
1024 	u32 mask;
1025 	u8 actual_vls;
1026 	u8 vls;
1027 	struct sdma_map_elem *map[];
1028 };
1029 
1030 int sdma_map_init(
1031 	struct hfi1_devdata *dd,
1032 	u8 port,
1033 	u8 num_vls,
1034 	u8 *vl_engines);
1035 
1036 /* slow path */
1037 void _sdma_engine_progress_schedule(struct sdma_engine *sde);
1038 
1039 /**
1040  * sdma_engine_progress_schedule() - schedule progress on engine
1041  * @sde: sdma_engine to schedule progress
1042  *
1043  * This is the fast path.
1044  *
1045  */
1046 static inline void sdma_engine_progress_schedule(
1047 	struct sdma_engine *sde)
1048 {
1049 	if (!sde || sdma_descq_inprocess(sde) < (sde->descq_cnt / 8))
1050 		return;
1051 	_sdma_engine_progress_schedule(sde);
1052 }
1053 
1054 struct sdma_engine *sdma_select_engine_sc(
1055 	struct hfi1_devdata *dd,
1056 	u32 selector,
1057 	u8 sc5);
1058 
1059 struct sdma_engine *sdma_select_engine_vl(
1060 	struct hfi1_devdata *dd,
1061 	u32 selector,
1062 	u8 vl);
1063 
1064 struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd,
1065 					    u32 selector, u8 vl);
1066 ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf);
1067 ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf,
1068 				size_t count);
1069 int sdma_engine_get_vl(struct sdma_engine *sde);
1070 void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *);
1071 void sdma_seqfile_dump_cpu_list(struct seq_file *s, struct hfi1_devdata *dd,
1072 				unsigned long cpuid);
1073 
1074 #ifdef CONFIG_SDMA_VERBOSITY
1075 void sdma_dumpstate(struct sdma_engine *);
1076 #endif
1077 static inline char *slashstrip(char *s)
1078 {
1079 	char *r = s;
1080 
1081 	while (*s)
1082 		if (*s++ == '/')
1083 			r = s;
1084 	return r;
1085 }
1086 
1087 u16 sdma_get_descq_cnt(void);
1088 
1089 extern uint mod_num_sdma;
1090 
1091 void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid);
1092 
1093 #endif
1094