1 /*
2  * Copyright (C) 2007 Michael Brown <mbrown@fensystems.co.uk>.
3  *
4  * Based in part upon the original driver by Mellanox Technologies
5  * Ltd.  Portions may be Copyright (c) Mellanox Technologies Ltd.
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License as
9  * published by the Free Software Foundation; either version 2 of the
10  * License, or any later version.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with this program; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
20  * 02110-1301, USA.
21  *
22  * You can also choose to distribute this program under the terms of
23  * the Unmodified Binary Distribution Licence (as given in the file
24  * COPYING.UBDL), provided that you have satisfied its requirements.
25  */
26 
27 FILE_LICENCE ( GPL2_OR_LATER_OR_UBDL );
28 
29 #include <stdint.h>
30 #include <stdlib.h>
31 #include <stdio.h>
32 #include <string.h>
33 #include <strings.h>
34 #include <unistd.h>
35 #include <errno.h>
36 #include <byteswap.h>
37 #include <ipxe/io.h>
38 #include <ipxe/pci.h>
39 #include <ipxe/pcibackup.h>
40 #include <ipxe/malloc.h>
41 #include <ipxe/umalloc.h>
42 #include <ipxe/iobuf.h>
43 #include <ipxe/netdevice.h>
44 #include <ipxe/infiniband.h>
45 #include <ipxe/ib_smc.h>
46 #include "arbel.h"
47 
48 /**
49  * @file
50  *
51  * Mellanox Arbel Infiniband HCA
52  *
53  */
54 
55 /***************************************************************************
56  *
57  * Queue number allocation
58  *
59  ***************************************************************************
60  */
61 
62 /**
63  * Allocate offset within usage bitmask
64  *
65  * @v bits		Usage bitmask
66  * @v bits_len		Length of usage bitmask
67  * @ret bit		First free bit within bitmask, or negative error
68  */
arbel_bitmask_alloc(arbel_bitmask_t * bits,unsigned int bits_len)69 static int arbel_bitmask_alloc ( arbel_bitmask_t *bits,
70 				 unsigned int bits_len ) {
71 	unsigned int bit = 0;
72 	arbel_bitmask_t mask = 1;
73 
74 	while ( bit < bits_len ) {
75 		if ( ( mask & *bits ) == 0 ) {
76 			*bits |= mask;
77 			return bit;
78 		}
79 		bit++;
80 		mask = ( mask << 1 ) | ( mask >> ( 8 * sizeof ( mask ) - 1 ) );
81 		if ( mask == 1 )
82 			bits++;
83 	}
84 	return -ENFILE;
85 }
86 
87 /**
88  * Free offset within usage bitmask
89  *
90  * @v bits		Usage bitmask
91  * @v bit		Bit within bitmask
92  */
arbel_bitmask_free(arbel_bitmask_t * bits,int bit)93 static void arbel_bitmask_free ( arbel_bitmask_t *bits, int bit ) {
94 	arbel_bitmask_t mask;
95 
96 	mask = ( 1 << ( bit % ( 8 * sizeof ( mask ) ) ) );
97 	bits += ( bit / ( 8 * sizeof ( mask ) ) );
98 	*bits &= ~mask;
99 }
100 
101 /***************************************************************************
102  *
103  * HCA commands
104  *
105  ***************************************************************************
106  */
107 
108 /**
109  * Wait for Arbel command completion
110  *
111  * @v arbel		Arbel device
112  * @ret rc		Return status code
113  */
arbel_cmd_wait(struct arbel * arbel,struct arbelprm_hca_command_register * hcr)114 static int arbel_cmd_wait ( struct arbel *arbel,
115 			    struct arbelprm_hca_command_register *hcr ) {
116 	unsigned int wait;
117 
118 	for ( wait = ARBEL_HCR_MAX_WAIT_MS ; wait ; wait-- ) {
119 		hcr->u.dwords[6] =
120 			readl ( arbel->config + ARBEL_HCR_REG ( 6 ) );
121 		if ( MLX_GET ( hcr, go ) == 0 )
122 			return 0;
123 		mdelay ( 1 );
124 	}
125 	return -EBUSY;
126 }
127 
128 /**
129  * Issue HCA command
130  *
131  * @v arbel		Arbel device
132  * @v command		Command opcode, flags and input/output lengths
133  * @v op_mod		Opcode modifier (0 if no modifier applicable)
134  * @v in		Input parameters
135  * @v in_mod		Input modifier (0 if no modifier applicable)
136  * @v out		Output parameters
137  * @ret rc		Return status code
138  */
arbel_cmd(struct arbel * arbel,unsigned long command,unsigned int op_mod,const void * in,unsigned int in_mod,void * out)139 static int arbel_cmd ( struct arbel *arbel, unsigned long command,
140 		       unsigned int op_mod, const void *in,
141 		       unsigned int in_mod, void *out ) {
142 	struct arbelprm_hca_command_register hcr;
143 	unsigned int opcode = ARBEL_HCR_OPCODE ( command );
144 	size_t in_len = ARBEL_HCR_IN_LEN ( command );
145 	size_t out_len = ARBEL_HCR_OUT_LEN ( command );
146 	void *in_buffer;
147 	void *out_buffer;
148 	unsigned int status;
149 	unsigned int i;
150 	int rc;
151 
152 	assert ( in_len <= ARBEL_MBOX_SIZE );
153 	assert ( out_len <= ARBEL_MBOX_SIZE );
154 
155 	DBGC2 ( arbel, "Arbel %p command %02x in %zx%s out %zx%s\n",
156 		arbel, opcode, in_len,
157 		( ( command & ARBEL_HCR_IN_MBOX ) ? "(mbox)" : "" ), out_len,
158 		( ( command & ARBEL_HCR_OUT_MBOX ) ? "(mbox)" : "" ) );
159 
160 	/* Check that HCR is free */
161 	if ( ( rc = arbel_cmd_wait ( arbel, &hcr ) ) != 0 ) {
162 		DBGC ( arbel, "Arbel %p command interface locked\n", arbel );
163 		return rc;
164 	}
165 
166 	/* Prepare HCR */
167 	memset ( &hcr, 0, sizeof ( hcr ) );
168 	in_buffer = &hcr.u.dwords[0];
169 	if ( in_len && ( command & ARBEL_HCR_IN_MBOX ) ) {
170 		in_buffer = arbel->mailbox_in;
171 		MLX_FILL_H ( &hcr, 0, in_param_h, virt_to_bus ( in_buffer ) );
172 		MLX_FILL_1 ( &hcr, 1, in_param_l, virt_to_bus ( in_buffer ) );
173 	}
174 	memcpy ( in_buffer, in, in_len );
175 	MLX_FILL_1 ( &hcr, 2, input_modifier, in_mod );
176 	out_buffer = &hcr.u.dwords[3];
177 	if ( out_len && ( command & ARBEL_HCR_OUT_MBOX ) ) {
178 		out_buffer = arbel->mailbox_out;
179 		MLX_FILL_H ( &hcr, 3, out_param_h,
180 			     virt_to_bus ( out_buffer ) );
181 		MLX_FILL_1 ( &hcr, 4, out_param_l,
182 			     virt_to_bus ( out_buffer ) );
183 	}
184 	MLX_FILL_3 ( &hcr, 6,
185 		     opcode, opcode,
186 		     opcode_modifier, op_mod,
187 		     go, 1 );
188 	DBGC ( arbel, "Arbel %p issuing command %04x\n", arbel, opcode );
189 	DBGC2_HDA ( arbel, virt_to_phys ( arbel->config + ARBEL_HCR_BASE ),
190 		    &hcr, sizeof ( hcr ) );
191 	if ( in_len && ( command & ARBEL_HCR_IN_MBOX ) ) {
192 		DBGC2 ( arbel, "Input mailbox:\n" );
193 		DBGC2_HDA ( arbel, virt_to_phys ( in_buffer ), in_buffer,
194 			    ( ( in_len < 512 ) ? in_len : 512 ) );
195 	}
196 
197 	/* Issue command */
198 	for ( i = 0 ; i < ( sizeof ( hcr ) / sizeof ( hcr.u.dwords[0] ) ) ;
199 	      i++ ) {
200 		writel ( hcr.u.dwords[i],
201 			 arbel->config + ARBEL_HCR_REG ( i ) );
202 		barrier();
203 	}
204 
205 	/* Wait for command completion */
206 	if ( ( rc = arbel_cmd_wait ( arbel, &hcr ) ) != 0 ) {
207 		DBGC ( arbel, "Arbel %p timed out waiting for command:\n",
208 		       arbel );
209 		DBGC_HD ( arbel, &hcr, sizeof ( hcr ) );
210 		return rc;
211 	}
212 
213 	/* Check command status */
214 	status = MLX_GET ( &hcr, status );
215 	if ( status != 0 ) {
216 		DBGC ( arbel, "Arbel %p command failed with status %02x:\n",
217 		       arbel, status );
218 		DBGC_HD ( arbel, &hcr, sizeof ( hcr ) );
219 		return -EIO;
220 	}
221 
222 	/* Read output parameters, if any */
223 	hcr.u.dwords[3] = readl ( arbel->config + ARBEL_HCR_REG ( 3 ) );
224 	hcr.u.dwords[4] = readl ( arbel->config + ARBEL_HCR_REG ( 4 ) );
225 	memcpy ( out, out_buffer, out_len );
226 	if ( out_len ) {
227 		DBGC2 ( arbel, "Output%s:\n",
228 			( command & ARBEL_HCR_OUT_MBOX ) ? " mailbox" : "" );
229 		DBGC2_HDA ( arbel, virt_to_phys ( out_buffer ), out_buffer,
230 			    ( ( out_len < 512 ) ? out_len : 512 ) );
231 	}
232 
233 	return 0;
234 }
235 
236 static inline int
arbel_cmd_query_dev_lim(struct arbel * arbel,struct arbelprm_query_dev_lim * dev_lim)237 arbel_cmd_query_dev_lim ( struct arbel *arbel,
238 			  struct arbelprm_query_dev_lim *dev_lim ) {
239 	return arbel_cmd ( arbel,
240 			   ARBEL_HCR_OUT_CMD ( ARBEL_HCR_QUERY_DEV_LIM,
241 					       1, sizeof ( *dev_lim ) ),
242 			   0, NULL, 0, dev_lim );
243 }
244 
245 static inline int
arbel_cmd_query_fw(struct arbel * arbel,struct arbelprm_query_fw * fw)246 arbel_cmd_query_fw ( struct arbel *arbel, struct arbelprm_query_fw *fw ) {
247 	return arbel_cmd ( arbel,
248 			   ARBEL_HCR_OUT_CMD ( ARBEL_HCR_QUERY_FW,
249 					       1, sizeof ( *fw ) ),
250 			   0, NULL, 0, fw );
251 }
252 
253 static inline int
arbel_cmd_init_hca(struct arbel * arbel,const struct arbelprm_init_hca * init_hca)254 arbel_cmd_init_hca ( struct arbel *arbel,
255 		     const struct arbelprm_init_hca *init_hca ) {
256 	return arbel_cmd ( arbel,
257 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_INIT_HCA,
258 					      1, sizeof ( *init_hca ) ),
259 			   0, init_hca, 0, NULL );
260 }
261 
262 static inline int
arbel_cmd_close_hca(struct arbel * arbel)263 arbel_cmd_close_hca ( struct arbel *arbel ) {
264 	return arbel_cmd ( arbel,
265 			   ARBEL_HCR_VOID_CMD ( ARBEL_HCR_CLOSE_HCA ),
266 			   0, NULL, 0, NULL );
267 }
268 
269 static inline int
arbel_cmd_init_ib(struct arbel * arbel,unsigned int port,const struct arbelprm_init_ib * init_ib)270 arbel_cmd_init_ib ( struct arbel *arbel, unsigned int port,
271 		    const struct arbelprm_init_ib *init_ib ) {
272 	return arbel_cmd ( arbel,
273 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_INIT_IB,
274 					      1, sizeof ( *init_ib ) ),
275 			   0, init_ib, port, NULL );
276 }
277 
278 static inline int
arbel_cmd_close_ib(struct arbel * arbel,unsigned int port)279 arbel_cmd_close_ib ( struct arbel *arbel, unsigned int port ) {
280 	return arbel_cmd ( arbel,
281 			   ARBEL_HCR_VOID_CMD ( ARBEL_HCR_CLOSE_IB ),
282 			   0, NULL, port, NULL );
283 }
284 
285 static inline int
arbel_cmd_sw2hw_mpt(struct arbel * arbel,unsigned int index,const struct arbelprm_mpt * mpt)286 arbel_cmd_sw2hw_mpt ( struct arbel *arbel, unsigned int index,
287 		      const struct arbelprm_mpt *mpt ) {
288 	return arbel_cmd ( arbel,
289 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_SW2HW_MPT,
290 					      1, sizeof ( *mpt ) ),
291 			   0, mpt, index, NULL );
292 }
293 
294 static inline int
arbel_cmd_map_eq(struct arbel * arbel,unsigned long index_map,const struct arbelprm_event_mask * mask)295 arbel_cmd_map_eq ( struct arbel *arbel, unsigned long index_map,
296 		   const struct arbelprm_event_mask *mask ) {
297 	return arbel_cmd ( arbel,
298 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_MAP_EQ,
299 					      0, sizeof ( *mask ) ),
300 			   0, mask, index_map, NULL );
301 }
302 
303 static inline int
arbel_cmd_sw2hw_eq(struct arbel * arbel,unsigned int index,const struct arbelprm_eqc * eqctx)304 arbel_cmd_sw2hw_eq ( struct arbel *arbel, unsigned int index,
305 		     const struct arbelprm_eqc *eqctx ) {
306 	return arbel_cmd ( arbel,
307 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_SW2HW_EQ,
308 					      1, sizeof ( *eqctx ) ),
309 			   0, eqctx, index, NULL );
310 }
311 
312 static inline int
arbel_cmd_hw2sw_eq(struct arbel * arbel,unsigned int index,struct arbelprm_eqc * eqctx)313 arbel_cmd_hw2sw_eq ( struct arbel *arbel, unsigned int index,
314 		     struct arbelprm_eqc *eqctx ) {
315 	return arbel_cmd ( arbel,
316 			   ARBEL_HCR_OUT_CMD ( ARBEL_HCR_HW2SW_EQ,
317 					       1, sizeof ( *eqctx ) ),
318 			   1, NULL, index, eqctx );
319 }
320 
321 static inline int
arbel_cmd_sw2hw_cq(struct arbel * arbel,unsigned long cqn,const struct arbelprm_completion_queue_context * cqctx)322 arbel_cmd_sw2hw_cq ( struct arbel *arbel, unsigned long cqn,
323 		     const struct arbelprm_completion_queue_context *cqctx ) {
324 	return arbel_cmd ( arbel,
325 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_SW2HW_CQ,
326 					      1, sizeof ( *cqctx ) ),
327 			   0, cqctx, cqn, NULL );
328 }
329 
330 static inline int
arbel_cmd_hw2sw_cq(struct arbel * arbel,unsigned long cqn,struct arbelprm_completion_queue_context * cqctx)331 arbel_cmd_hw2sw_cq ( struct arbel *arbel, unsigned long cqn,
332 		     struct arbelprm_completion_queue_context *cqctx) {
333 	return arbel_cmd ( arbel,
334 			   ARBEL_HCR_OUT_CMD ( ARBEL_HCR_HW2SW_CQ,
335 					       1, sizeof ( *cqctx ) ),
336 			   0, NULL, cqn, cqctx );
337 }
338 
339 static inline int
arbel_cmd_query_cq(struct arbel * arbel,unsigned long cqn,struct arbelprm_completion_queue_context * cqctx)340 arbel_cmd_query_cq ( struct arbel *arbel, unsigned long cqn,
341 		     struct arbelprm_completion_queue_context *cqctx ) {
342 	return arbel_cmd ( arbel,
343 			   ARBEL_HCR_OUT_CMD ( ARBEL_HCR_QUERY_CQ,
344 					       1, sizeof ( *cqctx ) ),
345 			   0, NULL, cqn, cqctx );
346 }
347 
348 static inline int
arbel_cmd_rst2init_qpee(struct arbel * arbel,unsigned long qpn,const struct arbelprm_qp_ee_state_transitions * ctx)349 arbel_cmd_rst2init_qpee ( struct arbel *arbel, unsigned long qpn,
350 			  const struct arbelprm_qp_ee_state_transitions *ctx ){
351 	return arbel_cmd ( arbel,
352 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_RST2INIT_QPEE,
353 					      1, sizeof ( *ctx ) ),
354 			   0, ctx, qpn, NULL );
355 }
356 
357 static inline int
arbel_cmd_init2rtr_qpee(struct arbel * arbel,unsigned long qpn,const struct arbelprm_qp_ee_state_transitions * ctx)358 arbel_cmd_init2rtr_qpee ( struct arbel *arbel, unsigned long qpn,
359 			  const struct arbelprm_qp_ee_state_transitions *ctx ){
360 	return arbel_cmd ( arbel,
361 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_INIT2RTR_QPEE,
362 					      1, sizeof ( *ctx ) ),
363 			   0, ctx, qpn, NULL );
364 }
365 
366 static inline int
arbel_cmd_rtr2rts_qpee(struct arbel * arbel,unsigned long qpn,const struct arbelprm_qp_ee_state_transitions * ctx)367 arbel_cmd_rtr2rts_qpee ( struct arbel *arbel, unsigned long qpn,
368 			 const struct arbelprm_qp_ee_state_transitions *ctx ) {
369 	return arbel_cmd ( arbel,
370 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_RTR2RTS_QPEE,
371 					      1, sizeof ( *ctx ) ),
372 			   0, ctx, qpn, NULL );
373 }
374 
375 static inline int
arbel_cmd_rts2rts_qpee(struct arbel * arbel,unsigned long qpn,const struct arbelprm_qp_ee_state_transitions * ctx)376 arbel_cmd_rts2rts_qpee ( struct arbel *arbel, unsigned long qpn,
377 			 const struct arbelprm_qp_ee_state_transitions *ctx ) {
378 	return arbel_cmd ( arbel,
379 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_RTS2RTS_QPEE,
380 					      1, sizeof ( *ctx ) ),
381 			   0, ctx, qpn, NULL );
382 }
383 
384 static inline int
arbel_cmd_2rst_qpee(struct arbel * arbel,unsigned long qpn)385 arbel_cmd_2rst_qpee ( struct arbel *arbel, unsigned long qpn ) {
386 	return arbel_cmd ( arbel,
387 			   ARBEL_HCR_VOID_CMD ( ARBEL_HCR_2RST_QPEE ),
388 			   0x03, NULL, qpn, NULL );
389 }
390 
391 static inline int
arbel_cmd_query_qpee(struct arbel * arbel,unsigned long qpn,struct arbelprm_qp_ee_state_transitions * ctx)392 arbel_cmd_query_qpee ( struct arbel *arbel, unsigned long qpn,
393 		       struct arbelprm_qp_ee_state_transitions *ctx ) {
394 	return arbel_cmd ( arbel,
395 			   ARBEL_HCR_OUT_CMD ( ARBEL_HCR_QUERY_QPEE,
396 					       1, sizeof ( *ctx ) ),
397 			   0, NULL, qpn, ctx );
398 }
399 
400 static inline int
arbel_cmd_conf_special_qp(struct arbel * arbel,unsigned int qp_type,unsigned long base_qpn)401 arbel_cmd_conf_special_qp ( struct arbel *arbel, unsigned int qp_type,
402 			    unsigned long base_qpn ) {
403 	return arbel_cmd ( arbel,
404 			   ARBEL_HCR_VOID_CMD ( ARBEL_HCR_CONF_SPECIAL_QP ),
405 			   qp_type, NULL, base_qpn, NULL );
406 }
407 
408 static inline int
arbel_cmd_mad_ifc(struct arbel * arbel,unsigned int port,union arbelprm_mad * mad)409 arbel_cmd_mad_ifc ( struct arbel *arbel, unsigned int port,
410 		    union arbelprm_mad *mad ) {
411 	return arbel_cmd ( arbel,
412 			   ARBEL_HCR_INOUT_CMD ( ARBEL_HCR_MAD_IFC,
413 						 1, sizeof ( *mad ),
414 						 1, sizeof ( *mad ) ),
415 			   0x03, mad, port, mad );
416 }
417 
418 static inline int
arbel_cmd_read_mgm(struct arbel * arbel,unsigned int index,struct arbelprm_mgm_entry * mgm)419 arbel_cmd_read_mgm ( struct arbel *arbel, unsigned int index,
420 		     struct arbelprm_mgm_entry *mgm ) {
421 	return arbel_cmd ( arbel,
422 			   ARBEL_HCR_OUT_CMD ( ARBEL_HCR_READ_MGM,
423 					       1, sizeof ( *mgm ) ),
424 			   0, NULL, index, mgm );
425 }
426 
427 static inline int
arbel_cmd_write_mgm(struct arbel * arbel,unsigned int index,const struct arbelprm_mgm_entry * mgm)428 arbel_cmd_write_mgm ( struct arbel *arbel, unsigned int index,
429 		      const struct arbelprm_mgm_entry *mgm ) {
430 	return arbel_cmd ( arbel,
431 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_WRITE_MGM,
432 					      1, sizeof ( *mgm ) ),
433 			   0, mgm, index, NULL );
434 }
435 
436 static inline int
arbel_cmd_mgid_hash(struct arbel * arbel,const union ib_gid * gid,struct arbelprm_mgm_hash * hash)437 arbel_cmd_mgid_hash ( struct arbel *arbel, const union ib_gid *gid,
438 		      struct arbelprm_mgm_hash *hash ) {
439 	return arbel_cmd ( arbel,
440 			   ARBEL_HCR_INOUT_CMD ( ARBEL_HCR_MGID_HASH,
441 						 1, sizeof ( *gid ),
442 						 0, sizeof ( *hash ) ),
443 			   0, gid, 0, hash );
444 }
445 
446 static inline int
arbel_cmd_run_fw(struct arbel * arbel)447 arbel_cmd_run_fw ( struct arbel *arbel ) {
448 	return arbel_cmd ( arbel,
449 			   ARBEL_HCR_VOID_CMD ( ARBEL_HCR_RUN_FW ),
450 			   0, NULL, 0, NULL );
451 }
452 
453 static inline int
arbel_cmd_disable_lam(struct arbel * arbel)454 arbel_cmd_disable_lam ( struct arbel *arbel ) {
455 	return arbel_cmd ( arbel,
456 			   ARBEL_HCR_VOID_CMD ( ARBEL_HCR_DISABLE_LAM ),
457 			   0, NULL, 0, NULL );
458 }
459 
460 static inline int
arbel_cmd_enable_lam(struct arbel * arbel,struct arbelprm_access_lam * lam)461 arbel_cmd_enable_lam ( struct arbel *arbel, struct arbelprm_access_lam *lam ) {
462 	return arbel_cmd ( arbel,
463 			   ARBEL_HCR_OUT_CMD ( ARBEL_HCR_ENABLE_LAM,
464 					       1, sizeof ( *lam ) ),
465 			   1, NULL, 0, lam );
466 }
467 
468 static inline int
arbel_cmd_unmap_icm(struct arbel * arbel,unsigned int page_count,const struct arbelprm_scalar_parameter * offset)469 arbel_cmd_unmap_icm ( struct arbel *arbel, unsigned int page_count,
470 		      const struct arbelprm_scalar_parameter *offset ) {
471 	return arbel_cmd ( arbel,
472 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_UNMAP_ICM, 0,
473 					      sizeof ( *offset ) ),
474 			   0, offset, page_count, NULL );
475 }
476 
477 static inline int
arbel_cmd_map_icm(struct arbel * arbel,const struct arbelprm_virtual_physical_mapping * map)478 arbel_cmd_map_icm ( struct arbel *arbel,
479 		    const struct arbelprm_virtual_physical_mapping *map ) {
480 	return arbel_cmd ( arbel,
481 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_MAP_ICM,
482 					      1, sizeof ( *map ) ),
483 			   0, map, 1, NULL );
484 }
485 
486 static inline int
arbel_cmd_unmap_icm_aux(struct arbel * arbel)487 arbel_cmd_unmap_icm_aux ( struct arbel *arbel ) {
488 	return arbel_cmd ( arbel,
489 			   ARBEL_HCR_VOID_CMD ( ARBEL_HCR_UNMAP_ICM_AUX ),
490 			   0, NULL, 0, NULL );
491 }
492 
493 static inline int
arbel_cmd_map_icm_aux(struct arbel * arbel,const struct arbelprm_virtual_physical_mapping * map)494 arbel_cmd_map_icm_aux ( struct arbel *arbel,
495 			const struct arbelprm_virtual_physical_mapping *map ) {
496 	return arbel_cmd ( arbel,
497 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_MAP_ICM_AUX,
498 					      1, sizeof ( *map ) ),
499 			   0, map, 1, NULL );
500 }
501 
502 static inline int
arbel_cmd_set_icm_size(struct arbel * arbel,const struct arbelprm_scalar_parameter * icm_size,struct arbelprm_scalar_parameter * icm_aux_size)503 arbel_cmd_set_icm_size ( struct arbel *arbel,
504 			 const struct arbelprm_scalar_parameter *icm_size,
505 			 struct arbelprm_scalar_parameter *icm_aux_size ) {
506 	return arbel_cmd ( arbel,
507 			   ARBEL_HCR_INOUT_CMD ( ARBEL_HCR_SET_ICM_SIZE,
508 						 0, sizeof ( *icm_size ),
509 						 0, sizeof ( *icm_aux_size ) ),
510 			   0, icm_size, 0, icm_aux_size );
511 }
512 
513 static inline int
arbel_cmd_unmap_fa(struct arbel * arbel)514 arbel_cmd_unmap_fa ( struct arbel *arbel ) {
515 	return arbel_cmd ( arbel,
516 			   ARBEL_HCR_VOID_CMD ( ARBEL_HCR_UNMAP_FA ),
517 			   0, NULL, 0, NULL );
518 }
519 
520 static inline int
arbel_cmd_map_fa(struct arbel * arbel,const struct arbelprm_virtual_physical_mapping * map)521 arbel_cmd_map_fa ( struct arbel *arbel,
522 		   const struct arbelprm_virtual_physical_mapping *map ) {
523 	return arbel_cmd ( arbel,
524 			   ARBEL_HCR_IN_CMD ( ARBEL_HCR_MAP_FA,
525 					      1, sizeof ( *map ) ),
526 			   0, map, 1, NULL );
527 }
528 
529 /***************************************************************************
530  *
531  * MAD operations
532  *
533  ***************************************************************************
534  */
535 
536 /**
537  * Issue management datagram
538  *
539  * @v ibdev		Infiniband device
540  * @v mad		Management datagram
541  * @ret rc		Return status code
542  */
arbel_mad(struct ib_device * ibdev,union ib_mad * mad)543 static int arbel_mad ( struct ib_device *ibdev, union ib_mad *mad ) {
544 	struct arbel *arbel = ib_get_drvdata ( ibdev );
545 	union arbelprm_mad mad_ifc;
546 	int rc;
547 
548 	linker_assert ( sizeof ( *mad ) == sizeof ( mad_ifc.mad ),
549 			mad_size_mismatch );
550 
551 	/* Copy in request packet */
552 	memcpy ( &mad_ifc.mad, mad, sizeof ( mad_ifc.mad ) );
553 
554 	/* Issue MAD */
555 	if ( ( rc = arbel_cmd_mad_ifc ( arbel, ibdev->port,
556 					&mad_ifc ) ) != 0 ) {
557 		DBGC ( arbel, "Arbel %p port %d could not issue MAD IFC: %s\n",
558 		       arbel, ibdev->port, strerror ( rc ) );
559 		return rc;
560 	}
561 
562 	/* Copy out reply packet */
563 	memcpy ( mad, &mad_ifc.mad, sizeof ( *mad ) );
564 
565 	if ( mad->hdr.status != 0 ) {
566 		DBGC ( arbel, "Arbel %p port %d MAD IFC status %04x\n",
567 		       arbel, ibdev->port, ntohs ( mad->hdr.status ) );
568 		return -EIO;
569 	}
570 	return 0;
571 }
572 
573 /***************************************************************************
574  *
575  * Completion queue operations
576  *
577  ***************************************************************************
578  */
579 
580 /**
581  * Dump completion queue context (for debugging only)
582  *
583  * @v arbel		Arbel device
584  * @v cq		Completion queue
585  * @ret rc		Return status code
586  */
587 static __attribute__ (( unused )) int
arbel_dump_cqctx(struct arbel * arbel,struct ib_completion_queue * cq)588 arbel_dump_cqctx ( struct arbel *arbel, struct ib_completion_queue *cq ) {
589 	struct arbelprm_completion_queue_context cqctx;
590 	int rc;
591 
592 	memset ( &cqctx, 0, sizeof ( cqctx ) );
593 	if ( ( rc = arbel_cmd_query_cq ( arbel, cq->cqn, &cqctx ) ) != 0 ) {
594 		DBGC ( arbel, "Arbel %p CQN %#lx QUERY_CQ failed: %s\n",
595 		       arbel, cq->cqn, strerror ( rc ) );
596 		return rc;
597 	}
598 	DBGC ( arbel, "Arbel %p CQN %#lx context:\n", arbel, cq->cqn );
599 	DBGC_HDA ( arbel, 0, &cqctx, sizeof ( cqctx ) );
600 
601 	return 0;
602 }
603 
604 /**
605  * Create completion queue
606  *
607  * @v ibdev		Infiniband device
608  * @v cq		Completion queue
609  * @ret rc		Return status code
610  */
arbel_create_cq(struct ib_device * ibdev,struct ib_completion_queue * cq)611 static int arbel_create_cq ( struct ib_device *ibdev,
612 			     struct ib_completion_queue *cq ) {
613 	struct arbel *arbel = ib_get_drvdata ( ibdev );
614 	struct arbel_completion_queue *arbel_cq;
615 	struct arbelprm_completion_queue_context cqctx;
616 	struct arbelprm_cq_ci_db_record *ci_db_rec;
617 	struct arbelprm_cq_arm_db_record *arm_db_rec;
618 	int cqn_offset;
619 	unsigned int i;
620 	int rc;
621 
622 	/* Find a free completion queue number */
623 	cqn_offset = arbel_bitmask_alloc ( arbel->cq_inuse, ARBEL_MAX_CQS );
624 	if ( cqn_offset < 0 ) {
625 		DBGC ( arbel, "Arbel %p out of completion queues\n", arbel );
626 		rc = cqn_offset;
627 		goto err_cqn_offset;
628 	}
629 	cq->cqn = ( arbel->limits.reserved_cqs + cqn_offset );
630 
631 	/* Allocate control structures */
632 	arbel_cq = zalloc ( sizeof ( *arbel_cq ) );
633 	if ( ! arbel_cq ) {
634 		rc = -ENOMEM;
635 		goto err_arbel_cq;
636 	}
637 	arbel_cq->ci_doorbell_idx = arbel_cq_ci_doorbell_idx ( arbel, cq );
638 	arbel_cq->arm_doorbell_idx = arbel_cq_arm_doorbell_idx ( arbel, cq );
639 
640 	/* Allocate completion queue itself */
641 	arbel_cq->cqe_size = ( cq->num_cqes * sizeof ( arbel_cq->cqe[0] ) );
642 	arbel_cq->cqe = malloc_dma ( arbel_cq->cqe_size,
643 				     sizeof ( arbel_cq->cqe[0] ) );
644 	if ( ! arbel_cq->cqe ) {
645 		rc = -ENOMEM;
646 		goto err_cqe;
647 	}
648 	memset ( arbel_cq->cqe, 0, arbel_cq->cqe_size );
649 	for ( i = 0 ; i < cq->num_cqes ; i++ ) {
650 		MLX_FILL_1 ( &arbel_cq->cqe[i].normal, 7, owner, 1 );
651 	}
652 	barrier();
653 
654 	/* Initialise doorbell records */
655 	ci_db_rec = &arbel->db_rec[arbel_cq->ci_doorbell_idx].cq_ci;
656 	MLX_FILL_1 ( ci_db_rec, 0, counter, 0 );
657 	MLX_FILL_2 ( ci_db_rec, 1,
658 		     res, ARBEL_UAR_RES_CQ_CI,
659 		     cq_number, cq->cqn );
660 	arm_db_rec = &arbel->db_rec[arbel_cq->arm_doorbell_idx].cq_arm;
661 	MLX_FILL_1 ( arm_db_rec, 0, counter, 0 );
662 	MLX_FILL_2 ( arm_db_rec, 1,
663 		     res, ARBEL_UAR_RES_CQ_ARM,
664 		     cq_number, cq->cqn );
665 
666 	/* Hand queue over to hardware */
667 	memset ( &cqctx, 0, sizeof ( cqctx ) );
668 	MLX_FILL_1 ( &cqctx, 0, st, 0xa /* "Event fired" */ );
669 	MLX_FILL_H ( &cqctx, 1, start_address_h,
670 		     virt_to_bus ( arbel_cq->cqe ) );
671 	MLX_FILL_1 ( &cqctx, 2, start_address_l,
672 		     virt_to_bus ( arbel_cq->cqe ) );
673 	MLX_FILL_2 ( &cqctx, 3,
674 		     usr_page, arbel->limits.reserved_uars,
675 		     log_cq_size, fls ( cq->num_cqes - 1 ) );
676 	MLX_FILL_1 ( &cqctx, 5, c_eqn, arbel->eq.eqn );
677 	MLX_FILL_1 ( &cqctx, 6, pd, ARBEL_GLOBAL_PD );
678 	MLX_FILL_1 ( &cqctx, 7, l_key, arbel->lkey );
679 	MLX_FILL_1 ( &cqctx, 12, cqn, cq->cqn );
680 	MLX_FILL_1 ( &cqctx, 13,
681 		     cq_ci_db_record, arbel_cq->ci_doorbell_idx );
682 	MLX_FILL_1 ( &cqctx, 14,
683 		     cq_state_db_record, arbel_cq->arm_doorbell_idx );
684 	if ( ( rc = arbel_cmd_sw2hw_cq ( arbel, cq->cqn, &cqctx ) ) != 0 ) {
685 		DBGC ( arbel, "Arbel %p CQN %#lx SW2HW_CQ failed: %s\n",
686 		       arbel, cq->cqn, strerror ( rc ) );
687 		goto err_sw2hw_cq;
688 	}
689 
690 	DBGC ( arbel, "Arbel %p CQN %#lx ring [%08lx,%08lx), doorbell %08lx\n",
691 	       arbel, cq->cqn, virt_to_phys ( arbel_cq->cqe ),
692 	       ( virt_to_phys ( arbel_cq->cqe ) + arbel_cq->cqe_size ),
693 	       virt_to_phys ( ci_db_rec ) );
694 	ib_cq_set_drvdata ( cq, arbel_cq );
695 	return 0;
696 
697  err_sw2hw_cq:
698 	MLX_FILL_1 ( ci_db_rec, 1, res, ARBEL_UAR_RES_NONE );
699 	MLX_FILL_1 ( arm_db_rec, 1, res, ARBEL_UAR_RES_NONE );
700 	free_dma ( arbel_cq->cqe, arbel_cq->cqe_size );
701  err_cqe:
702 	free ( arbel_cq );
703  err_arbel_cq:
704 	arbel_bitmask_free ( arbel->cq_inuse, cqn_offset );
705  err_cqn_offset:
706 	return rc;
707 }
708 
709 /**
710  * Destroy completion queue
711  *
712  * @v ibdev		Infiniband device
713  * @v cq		Completion queue
714  */
arbel_destroy_cq(struct ib_device * ibdev,struct ib_completion_queue * cq)715 static void arbel_destroy_cq ( struct ib_device *ibdev,
716 			       struct ib_completion_queue *cq ) {
717 	struct arbel *arbel = ib_get_drvdata ( ibdev );
718 	struct arbel_completion_queue *arbel_cq = ib_cq_get_drvdata ( cq );
719 	struct arbelprm_completion_queue_context cqctx;
720 	struct arbelprm_cq_ci_db_record *ci_db_rec;
721 	struct arbelprm_cq_arm_db_record *arm_db_rec;
722 	int cqn_offset;
723 	int rc;
724 
725 	/* Take ownership back from hardware */
726 	if ( ( rc = arbel_cmd_hw2sw_cq ( arbel, cq->cqn, &cqctx ) ) != 0 ) {
727 		DBGC ( arbel, "Arbel %p CQN %#lx FATAL HW2SW_CQ failed: "
728 		       "%s\n", arbel, cq->cqn, strerror ( rc ) );
729 		/* Leak memory and return; at least we avoid corruption */
730 		return;
731 	}
732 
733 	/* Clear doorbell records */
734 	ci_db_rec = &arbel->db_rec[arbel_cq->ci_doorbell_idx].cq_ci;
735 	arm_db_rec = &arbel->db_rec[arbel_cq->arm_doorbell_idx].cq_arm;
736 	MLX_FILL_1 ( ci_db_rec, 1, res, ARBEL_UAR_RES_NONE );
737 	MLX_FILL_1 ( arm_db_rec, 1, res, ARBEL_UAR_RES_NONE );
738 
739 	/* Free memory */
740 	free_dma ( arbel_cq->cqe, arbel_cq->cqe_size );
741 	free ( arbel_cq );
742 
743 	/* Mark queue number as free */
744 	cqn_offset = ( cq->cqn - arbel->limits.reserved_cqs );
745 	arbel_bitmask_free ( arbel->cq_inuse, cqn_offset );
746 
747 	ib_cq_set_drvdata ( cq, NULL );
748 }
749 
750 /***************************************************************************
751  *
752  * Queue pair operations
753  *
754  ***************************************************************************
755  */
756 
757 /**
758  * Assign queue pair number
759  *
760  * @v ibdev		Infiniband device
761  * @v qp		Queue pair
762  * @ret rc		Return status code
763  */
arbel_alloc_qpn(struct ib_device * ibdev,struct ib_queue_pair * qp)764 static int arbel_alloc_qpn ( struct ib_device *ibdev,
765 			     struct ib_queue_pair *qp ) {
766 	struct arbel *arbel = ib_get_drvdata ( ibdev );
767 	unsigned int port_offset;
768 	int qpn_offset;
769 
770 	/* Calculate queue pair number */
771 	port_offset = ( ibdev->port - ARBEL_PORT_BASE );
772 
773 	switch ( qp->type ) {
774 	case IB_QPT_SMI:
775 		qp->qpn = ( arbel->special_qpn_base + port_offset );
776 		return 0;
777 	case IB_QPT_GSI:
778 		qp->qpn = ( arbel->special_qpn_base + 2 + port_offset );
779 		return 0;
780 	case IB_QPT_UD:
781 	case IB_QPT_RC:
782 		/* Find a free queue pair number */
783 		qpn_offset = arbel_bitmask_alloc ( arbel->qp_inuse,
784 						   ARBEL_MAX_QPS );
785 		if ( qpn_offset < 0 ) {
786 			DBGC ( arbel, "Arbel %p out of queue pairs\n",
787 			       arbel );
788 			return qpn_offset;
789 		}
790 		qp->qpn = ( ( random() & ARBEL_QPN_RANDOM_MASK ) |
791 			    ( arbel->qpn_base + qpn_offset ) );
792 		return 0;
793 	default:
794 		DBGC ( arbel, "Arbel %p unsupported QP type %d\n",
795 		       arbel, qp->type );
796 		return -ENOTSUP;
797 	}
798 }
799 
800 /**
801  * Free queue pair number
802  *
803  * @v ibdev		Infiniband device
804  * @v qp		Queue pair
805  */
arbel_free_qpn(struct ib_device * ibdev,struct ib_queue_pair * qp)806 static void arbel_free_qpn ( struct ib_device *ibdev,
807 			     struct ib_queue_pair *qp ) {
808 	struct arbel *arbel = ib_get_drvdata ( ibdev );
809 	int qpn_offset;
810 
811 	qpn_offset = ( ( qp->qpn & ~ARBEL_QPN_RANDOM_MASK ) - arbel->qpn_base );
812 	if ( qpn_offset >= 0 )
813 		arbel_bitmask_free ( arbel->qp_inuse, qpn_offset );
814 }
815 
816 /**
817  * Calculate transmission rate
818  *
819  * @v av		Address vector
820  * @ret arbel_rate	Arbel rate
821  */
arbel_rate(struct ib_address_vector * av)822 static unsigned int arbel_rate ( struct ib_address_vector *av ) {
823 	return ( ( ( av->rate >= IB_RATE_2_5 ) && ( av->rate <= IB_RATE_120 ) )
824 		 ? ( av->rate + 5 ) : 0 );
825 }
826 
827 /** Queue pair transport service type map */
828 static uint8_t arbel_qp_st[] = {
829 	[IB_QPT_SMI] = ARBEL_ST_MLX,
830 	[IB_QPT_GSI] = ARBEL_ST_MLX,
831 	[IB_QPT_UD] = ARBEL_ST_UD,
832 	[IB_QPT_RC] = ARBEL_ST_RC,
833 };
834 
835 /**
836  * Dump queue pair context (for debugging only)
837  *
838  * @v arbel		Arbel device
839  * @v qp		Queue pair
840  * @ret rc		Return status code
841  */
842 static __attribute__ (( unused )) int
arbel_dump_qpctx(struct arbel * arbel,struct ib_queue_pair * qp)843 arbel_dump_qpctx ( struct arbel *arbel, struct ib_queue_pair *qp ) {
844 	struct arbelprm_qp_ee_state_transitions qpctx;
845 	int rc;
846 
847 	memset ( &qpctx, 0, sizeof ( qpctx ) );
848 	if ( ( rc = arbel_cmd_query_qpee ( arbel, qp->qpn, &qpctx ) ) != 0 ) {
849 		DBGC ( arbel, "Arbel %p QPN %#lx QUERY_QPEE failed: %s\n",
850 		       arbel, qp->qpn, strerror ( rc ) );
851 		return rc;
852 	}
853 	DBGC ( arbel, "Arbel %p QPN %#lx context:\n", arbel, qp->qpn );
854 	DBGC_HDA ( arbel, 0, &qpctx.u.dwords[2], ( sizeof ( qpctx ) - 8 ) );
855 
856 	return 0;
857 }
858 
859 /**
860  * Create send work queue
861  *
862  * @v arbel_send_wq	Send work queue
863  * @v num_wqes		Number of work queue entries
864  * @ret rc		Return status code
865  */
arbel_create_send_wq(struct arbel_send_work_queue * arbel_send_wq,unsigned int num_wqes)866 static int arbel_create_send_wq ( struct arbel_send_work_queue *arbel_send_wq,
867 				  unsigned int num_wqes ) {
868 	union arbel_send_wqe *wqe;
869 	union arbel_send_wqe *next_wqe;
870 	unsigned int wqe_idx_mask;
871 	unsigned int i;
872 
873 	/* Allocate work queue */
874 	arbel_send_wq->wqe_size = ( num_wqes *
875 				    sizeof ( arbel_send_wq->wqe[0] ) );
876 	arbel_send_wq->wqe = malloc_dma ( arbel_send_wq->wqe_size,
877 					  sizeof ( arbel_send_wq->wqe[0] ) );
878 	if ( ! arbel_send_wq->wqe )
879 		return -ENOMEM;
880 	memset ( arbel_send_wq->wqe, 0, arbel_send_wq->wqe_size );
881 
882 	/* Link work queue entries */
883 	wqe_idx_mask = ( num_wqes - 1 );
884 	for ( i = 0 ; i < num_wqes ; i++ ) {
885 		wqe = &arbel_send_wq->wqe[i];
886 		next_wqe = &arbel_send_wq->wqe[ ( i + 1 ) & wqe_idx_mask ];
887 		MLX_FILL_1 ( &wqe->next, 0, nda_31_6,
888 			     ( virt_to_bus ( next_wqe ) >> 6 ) );
889 		MLX_FILL_1 ( &wqe->next, 1, always1, 1 );
890 	}
891 
892 	return 0;
893 }
894 
895 /**
896  * Create receive work queue
897  *
898  * @v arbel_recv_wq	Receive work queue
899  * @v num_wqes		Number of work queue entries
900  * @v type		Queue pair type
901  * @ret rc		Return status code
902  */
arbel_create_recv_wq(struct arbel_recv_work_queue * arbel_recv_wq,unsigned int num_wqes,enum ib_queue_pair_type type)903 static int arbel_create_recv_wq ( struct arbel_recv_work_queue *arbel_recv_wq,
904 				  unsigned int num_wqes,
905 				  enum ib_queue_pair_type type ) {
906 	struct arbelprm_recv_wqe *wqe;
907 	struct arbelprm_recv_wqe *next_wqe;
908 	unsigned int wqe_idx_mask;
909 	size_t nds;
910 	unsigned int i;
911 	unsigned int j;
912 	int rc;
913 
914 	/* Allocate work queue */
915 	arbel_recv_wq->wqe_size = ( num_wqes *
916 				    sizeof ( arbel_recv_wq->wqe[0] ) );
917 	arbel_recv_wq->wqe = malloc_dma ( arbel_recv_wq->wqe_size,
918 					  sizeof ( arbel_recv_wq->wqe[0] ) );
919 	if ( ! arbel_recv_wq->wqe ) {
920 		rc = -ENOMEM;
921 		goto err_alloc_wqe;
922 	}
923 	memset ( arbel_recv_wq->wqe, 0, arbel_recv_wq->wqe_size );
924 
925 	/* Allocate GRH entries, if needed */
926 	if ( ( type == IB_QPT_SMI ) || ( type == IB_QPT_GSI ) ||
927 	     ( type == IB_QPT_UD ) ) {
928 		arbel_recv_wq->grh_size = ( num_wqes *
929 					    sizeof ( arbel_recv_wq->grh[0] ) );
930 		arbel_recv_wq->grh = malloc_dma ( arbel_recv_wq->grh_size,
931 						  sizeof ( void * ) );
932 		if ( ! arbel_recv_wq->grh ) {
933 			rc = -ENOMEM;
934 			goto err_alloc_grh;
935 		}
936 	}
937 
938 	/* Link work queue entries */
939 	wqe_idx_mask = ( num_wqes - 1 );
940 	nds = ( ( offsetof ( typeof ( *wqe ), data ) +
941 		  sizeof ( wqe->data[0] ) ) >> 4 );
942 	for ( i = 0 ; i < num_wqes ; i++ ) {
943 		wqe = &arbel_recv_wq->wqe[i].recv;
944 		next_wqe = &arbel_recv_wq->wqe[( i + 1 ) & wqe_idx_mask].recv;
945 		MLX_FILL_1 ( &wqe->next, 0, nda_31_6,
946 			     ( virt_to_bus ( next_wqe ) >> 6 ) );
947 		MLX_FILL_1 ( &wqe->next, 1, nds, nds );
948 		for ( j = 0 ; ( ( ( void * ) &wqe->data[j] ) <
949 				( ( void * ) ( wqe + 1 ) ) ) ; j++ ) {
950 			MLX_FILL_1 ( &wqe->data[j], 1,
951 				     l_key, ARBEL_INVALID_LKEY );
952 		}
953 	}
954 
955 	return 0;
956 
957 	free_dma ( arbel_recv_wq->grh, arbel_recv_wq->grh_size );
958  err_alloc_grh:
959 	free_dma ( arbel_recv_wq->wqe, arbel_recv_wq->wqe_size );
960  err_alloc_wqe:
961 	return rc;
962 }
963 
964 /**
965  * Create queue pair
966  *
967  * @v ibdev		Infiniband device
968  * @v qp		Queue pair
969  * @ret rc		Return status code
970  */
arbel_create_qp(struct ib_device * ibdev,struct ib_queue_pair * qp)971 static int arbel_create_qp ( struct ib_device *ibdev,
972 			     struct ib_queue_pair *qp ) {
973 	struct arbel *arbel = ib_get_drvdata ( ibdev );
974 	struct arbel_queue_pair *arbel_qp;
975 	struct arbelprm_qp_ee_state_transitions qpctx;
976 	struct arbelprm_qp_db_record *send_db_rec;
977 	struct arbelprm_qp_db_record *recv_db_rec;
978 	physaddr_t send_wqe_base_adr;
979 	physaddr_t recv_wqe_base_adr;
980 	physaddr_t wqe_base_adr;
981 	int rc;
982 
983 	/* Warn about dysfunctional code
984 	 *
985 	 * Arbel seems to crash the system as soon as the first send
986 	 * WQE completes on an RC queue pair.  (NOPs complete
987 	 * successfully, so this is a problem specific to the work
988 	 * queue rather than the completion queue.)  The cause of this
989 	 * problem has remained unknown for over a year.  Patches to
990 	 * fix this are welcome.
991 	 */
992 	if ( qp->type == IB_QPT_RC )
993 		DBG ( "*** WARNING: Arbel RC support is non-functional ***\n" );
994 
995 	/* Calculate queue pair number */
996 	if ( ( rc = arbel_alloc_qpn ( ibdev, qp ) ) != 0 )
997 		goto err_alloc_qpn;
998 
999 	/* Allocate control structures */
1000 	arbel_qp = zalloc ( sizeof ( *arbel_qp ) );
1001 	if ( ! arbel_qp ) {
1002 		rc = -ENOMEM;
1003 		goto err_arbel_qp;
1004 	}
1005 	arbel_qp->send.doorbell_idx = arbel_send_doorbell_idx ( arbel, qp );
1006 	arbel_qp->recv.doorbell_idx = arbel_recv_doorbell_idx ( arbel, qp );
1007 
1008 	/* Create send and receive work queues */
1009 	if ( ( rc = arbel_create_send_wq ( &arbel_qp->send,
1010 					   qp->send.num_wqes ) ) != 0 )
1011 		goto err_create_send_wq;
1012 	if ( ( rc = arbel_create_recv_wq ( &arbel_qp->recv, qp->recv.num_wqes,
1013 					   qp->type ) ) != 0 )
1014 		goto err_create_recv_wq;
1015 
1016 	/* Send and receive work queue entries must be within the same 4GB */
1017 	send_wqe_base_adr = virt_to_bus ( arbel_qp->send.wqe );
1018 	recv_wqe_base_adr = virt_to_bus ( arbel_qp->recv.wqe );
1019 	if ( ( sizeof ( physaddr_t ) > sizeof ( uint32_t ) ) &&
1020 	     ( ( ( ( uint64_t ) send_wqe_base_adr ) >> 32 ) !=
1021 	       ( ( ( uint64_t ) recv_wqe_base_adr ) >> 32 ) ) ) {
1022 		DBGC ( arbel, "Arbel %p QPN %#lx cannot support send %08lx "
1023 		       "recv %08lx\n", arbel, qp->qpn,
1024 		       send_wqe_base_adr, recv_wqe_base_adr );
1025 		rc = -ENOTSUP;
1026 		goto err_unsupported_address_split;
1027 	}
1028 	wqe_base_adr = send_wqe_base_adr;
1029 
1030 	/* Initialise doorbell records */
1031 	send_db_rec = &arbel->db_rec[arbel_qp->send.doorbell_idx].qp;
1032 	MLX_FILL_1 ( send_db_rec, 0, counter, 0 );
1033 	MLX_FILL_2 ( send_db_rec, 1,
1034 		     res, ARBEL_UAR_RES_SQ,
1035 		     qp_number, qp->qpn );
1036 	recv_db_rec = &arbel->db_rec[arbel_qp->recv.doorbell_idx].qp;
1037 	MLX_FILL_1 ( recv_db_rec, 0, counter, 0 );
1038 	MLX_FILL_2 ( recv_db_rec, 1,
1039 		     res, ARBEL_UAR_RES_RQ,
1040 		     qp_number, qp->qpn );
1041 
1042 	/* Transition queue to INIT state */
1043 	memset ( &qpctx, 0, sizeof ( qpctx ) );
1044 	MLX_FILL_3 ( &qpctx, 2,
1045 		     qpc_eec_data.de, 1,
1046 		     qpc_eec_data.pm_state, ARBEL_PM_STATE_MIGRATED,
1047 		     qpc_eec_data.st, arbel_qp_st[qp->type] );
1048 	MLX_FILL_4 ( &qpctx, 4,
1049 		     qpc_eec_data.log_rq_size, fls ( qp->recv.num_wqes - 1 ),
1050 		     qpc_eec_data.log_rq_stride,
1051 		     ( fls ( sizeof ( arbel_qp->recv.wqe[0] ) - 1 ) - 4 ),
1052 		     qpc_eec_data.log_sq_size, fls ( qp->send.num_wqes - 1 ),
1053 		     qpc_eec_data.log_sq_stride,
1054 		     ( fls ( sizeof ( arbel_qp->send.wqe[0] ) - 1 ) - 4 ) );
1055 	MLX_FILL_1 ( &qpctx, 5,
1056 		     qpc_eec_data.usr_page, arbel->limits.reserved_uars );
1057 	MLX_FILL_1 ( &qpctx, 10, qpc_eec_data.primary_address_path.port_number,
1058 		     ibdev->port );
1059 	MLX_FILL_1 ( &qpctx, 27, qpc_eec_data.pd, ARBEL_GLOBAL_PD );
1060 	MLX_FILL_H ( &qpctx, 28, qpc_eec_data.wqe_base_adr_h, wqe_base_adr );
1061 	MLX_FILL_1 ( &qpctx, 29, qpc_eec_data.wqe_lkey, arbel->lkey );
1062 	MLX_FILL_1 ( &qpctx, 30, qpc_eec_data.ssc, 1 );
1063 	MLX_FILL_1 ( &qpctx, 33, qpc_eec_data.cqn_snd, qp->send.cq->cqn );
1064 	MLX_FILL_1 ( &qpctx, 34, qpc_eec_data.snd_wqe_base_adr_l,
1065 		     ( send_wqe_base_adr >> 6 ) );
1066 	MLX_FILL_1 ( &qpctx, 35, qpc_eec_data.snd_db_record_index,
1067 		     arbel_qp->send.doorbell_idx );
1068 	MLX_FILL_4 ( &qpctx, 38,
1069 		     qpc_eec_data.rre, 1,
1070 		     qpc_eec_data.rwe, 1,
1071 		     qpc_eec_data.rae, 1,
1072 		     qpc_eec_data.rsc, 1 );
1073 	MLX_FILL_1 ( &qpctx, 41, qpc_eec_data.cqn_rcv, qp->recv.cq->cqn );
1074 	MLX_FILL_1 ( &qpctx, 42, qpc_eec_data.rcv_wqe_base_adr_l,
1075 		     ( recv_wqe_base_adr >> 6 ) );
1076 	MLX_FILL_1 ( &qpctx, 43, qpc_eec_data.rcv_db_record_index,
1077 		     arbel_qp->recv.doorbell_idx );
1078 	if ( ( rc = arbel_cmd_rst2init_qpee ( arbel, qp->qpn, &qpctx )) != 0 ){
1079 		DBGC ( arbel, "Arbel %p QPN %#lx RST2INIT_QPEE failed: %s\n",
1080 		       arbel, qp->qpn, strerror ( rc ) );
1081 		goto err_rst2init_qpee;
1082 	}
1083 	arbel_qp->state = ARBEL_QP_ST_INIT;
1084 
1085 	DBGC ( arbel, "Arbel %p QPN %#lx send ring [%08lx,%08lx), doorbell "
1086 	       "%08lx\n", arbel, qp->qpn, virt_to_phys ( arbel_qp->send.wqe ),
1087 	       ( virt_to_phys ( arbel_qp->send.wqe ) +
1088 		 arbel_qp->send.wqe_size ),
1089 	       virt_to_phys ( send_db_rec ) );
1090 	DBGC ( arbel, "Arbel %p QPN %#lx receive ring [%08lx,%08lx), doorbell "
1091 	       "%08lx\n", arbel, qp->qpn, virt_to_phys ( arbel_qp->recv.wqe ),
1092 	       ( virt_to_phys ( arbel_qp->recv.wqe ) +
1093 		 arbel_qp->recv.wqe_size ),
1094 	       virt_to_phys ( recv_db_rec ) );
1095 	DBGC ( arbel, "Arbel %p QPN %#lx send CQN %#lx receive CQN %#lx\n",
1096 	       arbel, qp->qpn, qp->send.cq->cqn, qp->recv.cq->cqn );
1097 	ib_qp_set_drvdata ( qp, arbel_qp );
1098 	return 0;
1099 
1100 	arbel_cmd_2rst_qpee ( arbel, qp->qpn );
1101  err_rst2init_qpee:
1102 	MLX_FILL_1 ( send_db_rec, 1, res, ARBEL_UAR_RES_NONE );
1103 	MLX_FILL_1 ( recv_db_rec, 1, res, ARBEL_UAR_RES_NONE );
1104  err_unsupported_address_split:
1105 	free_dma ( arbel_qp->recv.grh, arbel_qp->recv.grh_size );
1106 	free_dma ( arbel_qp->recv.wqe, arbel_qp->recv.wqe_size );
1107  err_create_recv_wq:
1108 	free_dma ( arbel_qp->send.wqe, arbel_qp->send.wqe_size );
1109  err_create_send_wq:
1110 	free ( arbel_qp );
1111  err_arbel_qp:
1112 	arbel_free_qpn ( ibdev, qp );
1113  err_alloc_qpn:
1114 	return rc;
1115 }
1116 
1117 /**
1118  * Modify queue pair
1119  *
1120  * @v ibdev		Infiniband device
1121  * @v qp		Queue pair
1122  * @ret rc		Return status code
1123  */
arbel_modify_qp(struct ib_device * ibdev,struct ib_queue_pair * qp)1124 static int arbel_modify_qp ( struct ib_device *ibdev,
1125 			     struct ib_queue_pair *qp ) {
1126 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1127 	struct arbel_queue_pair *arbel_qp = ib_qp_get_drvdata ( qp );
1128 	struct arbelprm_qp_ee_state_transitions qpctx;
1129 	int rc;
1130 
1131 	/* Transition queue to RTR state, if applicable */
1132 	if ( arbel_qp->state < ARBEL_QP_ST_RTR ) {
1133 		memset ( &qpctx, 0, sizeof ( qpctx ) );
1134 		MLX_FILL_2 ( &qpctx, 4,
1135 			     qpc_eec_data.mtu, ARBEL_MTU_2048,
1136 			     qpc_eec_data.msg_max, 31 );
1137 		MLX_FILL_1 ( &qpctx, 7,
1138 			     qpc_eec_data.remote_qpn_een, qp->av.qpn );
1139 		MLX_FILL_2 ( &qpctx, 11,
1140 			     qpc_eec_data.primary_address_path.rnr_retry,
1141 			     ARBEL_RETRY_MAX,
1142 			     qpc_eec_data.primary_address_path.rlid,
1143 			     qp->av.lid );
1144 		MLX_FILL_2 ( &qpctx, 12,
1145 			     qpc_eec_data.primary_address_path.ack_timeout,
1146 			     14 /* 4.096us * 2^(14) = 67ms */,
1147 			     qpc_eec_data.primary_address_path.max_stat_rate,
1148 			     arbel_rate ( &qp->av ) );
1149 		memcpy ( &qpctx.u.dwords[14], &qp->av.gid,
1150 			 sizeof ( qp->av.gid ) );
1151 		MLX_FILL_1 ( &qpctx, 30,
1152 			     qpc_eec_data.retry_count, ARBEL_RETRY_MAX );
1153 		MLX_FILL_1 ( &qpctx, 39,
1154 			     qpc_eec_data.next_rcv_psn, qp->recv.psn );
1155 		MLX_FILL_1 ( &qpctx, 40,
1156 			     qpc_eec_data.ra_buff_indx,
1157 			     ( arbel->limits.reserved_rdbs +
1158 			       ( ( qp->qpn & ~ARBEL_QPN_RANDOM_MASK ) -
1159 				 arbel->special_qpn_base ) ) );
1160 		if ( ( rc = arbel_cmd_init2rtr_qpee ( arbel, qp->qpn,
1161 						      &qpctx ) ) != 0 ) {
1162 			DBGC ( arbel, "Arbel %p QPN %#lx INIT2RTR_QPEE failed:"
1163 			       " %s\n", arbel, qp->qpn, strerror ( rc ) );
1164 			return rc;
1165 		}
1166 		arbel_qp->state = ARBEL_QP_ST_RTR;
1167 	}
1168 
1169 	/* Transition queue to RTS state, if applicable */
1170 	if ( arbel_qp->state < ARBEL_QP_ST_RTS ) {
1171 		memset ( &qpctx, 0, sizeof ( qpctx ) );
1172 		MLX_FILL_1 ( &qpctx, 11,
1173 			     qpc_eec_data.primary_address_path.rnr_retry,
1174 			     ARBEL_RETRY_MAX );
1175 		MLX_FILL_1 ( &qpctx, 12,
1176 			     qpc_eec_data.primary_address_path.ack_timeout,
1177 			     14 /* 4.096us * 2^(14) = 67ms */ );
1178 		MLX_FILL_2 ( &qpctx, 30,
1179 			     qpc_eec_data.retry_count, ARBEL_RETRY_MAX,
1180 			     qpc_eec_data.sic, 1 );
1181 		MLX_FILL_1 ( &qpctx, 32,
1182 			     qpc_eec_data.next_send_psn, qp->send.psn );
1183 		if ( ( rc = arbel_cmd_rtr2rts_qpee ( arbel, qp->qpn,
1184 						     &qpctx ) ) != 0 ) {
1185 			DBGC ( arbel, "Arbel %p QPN %#lx RTR2RTS_QPEE failed: "
1186 			       "%s\n", arbel, qp->qpn, strerror ( rc ) );
1187 			return rc;
1188 		}
1189 		arbel_qp->state = ARBEL_QP_ST_RTS;
1190 	}
1191 
1192 	/* Update parameters in RTS state */
1193 	memset ( &qpctx, 0, sizeof ( qpctx ) );
1194 	MLX_FILL_1 ( &qpctx, 0, opt_param_mask, ARBEL_QPEE_OPT_PARAM_QKEY );
1195 	MLX_FILL_1 ( &qpctx, 44, qpc_eec_data.q_key, qp->qkey );
1196 	if ( ( rc = arbel_cmd_rts2rts_qpee ( arbel, qp->qpn, &qpctx ) ) != 0 ){
1197 		DBGC ( arbel, "Arbel %p QPN %#lx RTS2RTS_QPEE failed: %s\n",
1198 		       arbel, qp->qpn, strerror ( rc ) );
1199 		return rc;
1200 	}
1201 
1202 	return 0;
1203 }
1204 
1205 /**
1206  * Destroy queue pair
1207  *
1208  * @v ibdev		Infiniband device
1209  * @v qp		Queue pair
1210  */
arbel_destroy_qp(struct ib_device * ibdev,struct ib_queue_pair * qp)1211 static void arbel_destroy_qp ( struct ib_device *ibdev,
1212 			       struct ib_queue_pair *qp ) {
1213 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1214 	struct arbel_queue_pair *arbel_qp = ib_qp_get_drvdata ( qp );
1215 	struct arbelprm_qp_db_record *send_db_rec;
1216 	struct arbelprm_qp_db_record *recv_db_rec;
1217 	int rc;
1218 
1219 	/* Take ownership back from hardware */
1220 	if ( ( rc = arbel_cmd_2rst_qpee ( arbel, qp->qpn ) ) != 0 ) {
1221 		DBGC ( arbel, "Arbel %p QPN %#lx FATAL 2RST_QPEE failed: "
1222 		       "%s\n", arbel, qp->qpn, strerror ( rc ) );
1223 		/* Leak memory and return; at least we avoid corruption */
1224 		return;
1225 	}
1226 
1227 	/* Clear doorbell records */
1228 	send_db_rec = &arbel->db_rec[arbel_qp->send.doorbell_idx].qp;
1229 	recv_db_rec = &arbel->db_rec[arbel_qp->recv.doorbell_idx].qp;
1230 	MLX_FILL_1 ( send_db_rec, 1, res, ARBEL_UAR_RES_NONE );
1231 	MLX_FILL_1 ( recv_db_rec, 1, res, ARBEL_UAR_RES_NONE );
1232 
1233 	/* Free memory */
1234 	free_dma ( arbel_qp->recv.grh, arbel_qp->recv.grh_size );
1235 	free_dma ( arbel_qp->recv.wqe, arbel_qp->recv.wqe_size );
1236 	free_dma ( arbel_qp->send.wqe, arbel_qp->send.wqe_size );
1237 	free ( arbel_qp );
1238 
1239 	/* Mark queue number as free */
1240 	arbel_free_qpn ( ibdev, qp );
1241 
1242 	ib_qp_set_drvdata ( qp, NULL );
1243 }
1244 
1245 /***************************************************************************
1246  *
1247  * Work request operations
1248  *
1249  ***************************************************************************
1250  */
1251 
1252 /**
1253  * Ring doorbell register in UAR
1254  *
1255  * @v arbel		Arbel device
1256  * @v db_reg		Doorbell register structure
1257  * @v offset		Address of doorbell
1258  */
arbel_ring_doorbell(struct arbel * arbel,union arbelprm_doorbell_register * db_reg,unsigned int offset)1259 static void arbel_ring_doorbell ( struct arbel *arbel,
1260 				  union arbelprm_doorbell_register *db_reg,
1261 				  unsigned int offset ) {
1262 
1263 	DBGC2 ( arbel, "Arbel %p ringing doorbell %08x:%08x at %lx\n",
1264 		arbel, ntohl ( db_reg->dword[0] ), ntohl ( db_reg->dword[1] ),
1265 		virt_to_phys ( arbel->uar + offset ) );
1266 
1267 	barrier();
1268 	writel ( db_reg->dword[0], ( arbel->uar + offset + 0 ) );
1269 	barrier();
1270 	writel ( db_reg->dword[1], ( arbel->uar + offset + 4 ) );
1271 }
1272 
1273 /** GID used for GID-less send work queue entries */
1274 static const union ib_gid arbel_no_gid = {
1275 	.bytes = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0 },
1276 };
1277 
1278 /**
1279  * Construct UD send work queue entry
1280  *
1281  * @v ibdev		Infiniband device
1282  * @v qp		Queue pair
1283  * @v dest		Destination address vector
1284  * @v iobuf		I/O buffer
1285  * @v wqe		Send work queue entry
1286  * @ret nds		Work queue entry size
1287  */
arbel_fill_ud_send_wqe(struct ib_device * ibdev,struct ib_queue_pair * qp __unused,struct ib_address_vector * dest,struct io_buffer * iobuf,union arbel_send_wqe * wqe)1288 static size_t arbel_fill_ud_send_wqe ( struct ib_device *ibdev,
1289 				       struct ib_queue_pair *qp __unused,
1290 				       struct ib_address_vector *dest,
1291 				       struct io_buffer *iobuf,
1292 				       union arbel_send_wqe *wqe ) {
1293 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1294 	const union ib_gid *gid;
1295 
1296 	/* Construct this work queue entry */
1297 	MLX_FILL_1 ( &wqe->ud.ctrl, 0, always1, 1 );
1298 	MLX_FILL_2 ( &wqe->ud.ud, 0,
1299 		     ud_address_vector.pd, ARBEL_GLOBAL_PD,
1300 		     ud_address_vector.port_number, ibdev->port );
1301 	MLX_FILL_2 ( &wqe->ud.ud, 1,
1302 		     ud_address_vector.rlid, dest->lid,
1303 		     ud_address_vector.g, dest->gid_present );
1304 	MLX_FILL_2 ( &wqe->ud.ud, 2,
1305 		     ud_address_vector.max_stat_rate, arbel_rate ( dest ),
1306 		     ud_address_vector.msg, 3 );
1307 	MLX_FILL_1 ( &wqe->ud.ud, 3, ud_address_vector.sl, dest->sl );
1308 	gid = ( dest->gid_present ? &dest->gid : &arbel_no_gid );
1309 	memcpy ( &wqe->ud.ud.u.dwords[4], gid, sizeof ( *gid ) );
1310 	MLX_FILL_1 ( &wqe->ud.ud, 8, destination_qp, dest->qpn );
1311 	MLX_FILL_1 ( &wqe->ud.ud, 9, q_key, dest->qkey );
1312 	MLX_FILL_1 ( &wqe->ud.data[0], 0, byte_count, iob_len ( iobuf ) );
1313 	MLX_FILL_1 ( &wqe->ud.data[0], 1, l_key, arbel->lkey );
1314 	MLX_FILL_H ( &wqe->ud.data[0], 2,
1315 		     local_address_h, virt_to_bus ( iobuf->data ) );
1316 	MLX_FILL_1 ( &wqe->ud.data[0], 3,
1317 		     local_address_l, virt_to_bus ( iobuf->data ) );
1318 
1319 	return ( offsetof ( typeof ( wqe->ud ), data[1] ) >> 4 );
1320 }
1321 
1322 /**
1323  * Construct MLX send work queue entry
1324  *
1325  * @v ibdev		Infiniband device
1326  * @v qp		Queue pair
1327  * @v dest		Destination address vector
1328  * @v iobuf		I/O buffer
1329  * @v wqe		Send work queue entry
1330  * @ret nds		Work queue entry size
1331  */
arbel_fill_mlx_send_wqe(struct ib_device * ibdev,struct ib_queue_pair * qp,struct ib_address_vector * dest,struct io_buffer * iobuf,union arbel_send_wqe * wqe)1332 static size_t arbel_fill_mlx_send_wqe ( struct ib_device *ibdev,
1333 					struct ib_queue_pair *qp,
1334 					struct ib_address_vector *dest,
1335 					struct io_buffer *iobuf,
1336 					union arbel_send_wqe *wqe ) {
1337 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1338 	struct io_buffer headers;
1339 
1340 	/* Construct IB headers */
1341 	iob_populate ( &headers, &wqe->mlx.headers, 0,
1342 		       sizeof ( wqe->mlx.headers ) );
1343 	iob_reserve ( &headers, sizeof ( wqe->mlx.headers ) );
1344 	ib_push ( ibdev, &headers, qp, iob_len ( iobuf ), dest );
1345 
1346 	/* Construct this work queue entry */
1347 	MLX_FILL_5 ( &wqe->mlx.ctrl, 0,
1348 		     c, 1 /* generate completion */,
1349 		     icrc, 0 /* generate ICRC */,
1350 		     max_statrate, arbel_rate ( dest ),
1351 		     slr, 0,
1352 		     v15, ( ( qp->ext_qpn == IB_QPN_SMI ) ? 1 : 0 ) );
1353 	MLX_FILL_1 ( &wqe->mlx.ctrl, 1, rlid, dest->lid );
1354 	MLX_FILL_1 ( &wqe->mlx.data[0], 0,
1355 		     byte_count, iob_len ( &headers ) );
1356 	MLX_FILL_1 ( &wqe->mlx.data[0], 1, l_key, arbel->lkey );
1357 	MLX_FILL_H ( &wqe->mlx.data[0], 2,
1358 		     local_address_h, virt_to_bus ( headers.data ) );
1359 	MLX_FILL_1 ( &wqe->mlx.data[0], 3,
1360 		     local_address_l, virt_to_bus ( headers.data ) );
1361 	MLX_FILL_1 ( &wqe->mlx.data[1], 0,
1362 		     byte_count, ( iob_len ( iobuf ) + 4 /* ICRC */ ) );
1363 	MLX_FILL_1 ( &wqe->mlx.data[1], 1, l_key, arbel->lkey );
1364 	MLX_FILL_H ( &wqe->mlx.data[1], 2,
1365 		     local_address_h, virt_to_bus ( iobuf->data ) );
1366 	MLX_FILL_1 ( &wqe->mlx.data[1], 3,
1367 		     local_address_l, virt_to_bus ( iobuf->data ) );
1368 
1369 	return ( offsetof ( typeof ( wqe->mlx ), data[2] ) >> 4 );
1370 }
1371 
1372 /**
1373  * Construct RC send work queue entry
1374  *
1375  * @v ibdev		Infiniband device
1376  * @v qp		Queue pair
1377  * @v dest		Destination address vector
1378  * @v iobuf		I/O buffer
1379  * @v wqe		Send work queue entry
1380  * @ret nds		Work queue entry size
1381  */
arbel_fill_rc_send_wqe(struct ib_device * ibdev,struct ib_queue_pair * qp __unused,struct ib_address_vector * dest __unused,struct io_buffer * iobuf,union arbel_send_wqe * wqe)1382 static size_t arbel_fill_rc_send_wqe ( struct ib_device *ibdev,
1383 				       struct ib_queue_pair *qp __unused,
1384 				       struct ib_address_vector *dest __unused,
1385 				       struct io_buffer *iobuf,
1386 				       union arbel_send_wqe *wqe ) {
1387 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1388 
1389 	/* Construct this work queue entry */
1390 	MLX_FILL_1 ( &wqe->rc.ctrl, 0, always1, 1 );
1391 	MLX_FILL_1 ( &wqe->rc.data[0], 0, byte_count, iob_len ( iobuf ) );
1392 	MLX_FILL_1 ( &wqe->rc.data[0], 1, l_key, arbel->lkey );
1393 	MLX_FILL_H ( &wqe->rc.data[0], 2,
1394 		     local_address_h, virt_to_bus ( iobuf->data ) );
1395 	MLX_FILL_1 ( &wqe->rc.data[0], 3,
1396 		     local_address_l, virt_to_bus ( iobuf->data ) );
1397 
1398 	return ( offsetof ( typeof ( wqe->rc ), data[1] ) >> 4 );
1399 }
1400 
1401 /** Work queue entry constructors */
1402 static size_t
1403 ( * arbel_fill_send_wqe[] ) ( struct ib_device *ibdev,
1404 			      struct ib_queue_pair *qp,
1405 			      struct ib_address_vector *dest,
1406 			      struct io_buffer *iobuf,
1407 			      union arbel_send_wqe *wqe ) = {
1408 	[IB_QPT_SMI] = arbel_fill_mlx_send_wqe,
1409 	[IB_QPT_GSI] = arbel_fill_mlx_send_wqe,
1410 	[IB_QPT_UD] = arbel_fill_ud_send_wqe,
1411 	[IB_QPT_RC] = arbel_fill_rc_send_wqe,
1412 };
1413 
1414 /**
1415  * Post send work queue entry
1416  *
1417  * @v ibdev		Infiniband device
1418  * @v qp		Queue pair
1419  * @v dest		Destination address vector
1420  * @v iobuf		I/O buffer
1421  * @ret rc		Return status code
1422  */
arbel_post_send(struct ib_device * ibdev,struct ib_queue_pair * qp,struct ib_address_vector * dest,struct io_buffer * iobuf)1423 static int arbel_post_send ( struct ib_device *ibdev,
1424 			     struct ib_queue_pair *qp,
1425 			     struct ib_address_vector *dest,
1426 			     struct io_buffer *iobuf ) {
1427 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1428 	struct arbel_queue_pair *arbel_qp = ib_qp_get_drvdata ( qp );
1429 	struct ib_work_queue *wq = &qp->send;
1430 	struct arbel_send_work_queue *arbel_send_wq = &arbel_qp->send;
1431 	union arbel_send_wqe *prev_wqe;
1432 	union arbel_send_wqe *wqe;
1433 	struct arbelprm_qp_db_record *qp_db_rec;
1434 	union arbelprm_doorbell_register db_reg;
1435 	unsigned long wqe_idx_mask;
1436 	size_t nds;
1437 
1438 	/* Allocate work queue entry */
1439 	wqe_idx_mask = ( wq->num_wqes - 1 );
1440 	if ( wq->iobufs[wq->next_idx & wqe_idx_mask] ) {
1441 		DBGC ( arbel, "Arbel %p QPN %#lx send queue full",
1442 		       arbel, qp->qpn );
1443 		return -ENOBUFS;
1444 	}
1445 	wq->iobufs[wq->next_idx & wqe_idx_mask] = iobuf;
1446 	prev_wqe = &arbel_send_wq->wqe[(wq->next_idx - 1) & wqe_idx_mask];
1447 	wqe = &arbel_send_wq->wqe[wq->next_idx & wqe_idx_mask];
1448 
1449 	/* Construct work queue entry */
1450 	memset ( ( ( ( void * ) wqe ) + sizeof ( wqe->next ) ), 0,
1451 		 ( sizeof ( *wqe ) - sizeof ( wqe->next ) ) );
1452 	assert ( qp->type < ( sizeof ( arbel_fill_send_wqe ) /
1453 			      sizeof ( arbel_fill_send_wqe[0] ) ) );
1454 	assert ( arbel_fill_send_wqe[qp->type] != NULL );
1455 	nds = arbel_fill_send_wqe[qp->type] ( ibdev, qp, dest, iobuf, wqe );
1456 	DBGCP ( arbel, "Arbel %p QPN %#lx posting send WQE %#lx:\n",
1457 		arbel, qp->qpn, ( wq->next_idx & wqe_idx_mask ) );
1458 	DBGCP_HDA ( arbel, virt_to_phys ( wqe ), wqe, sizeof ( *wqe ) );
1459 
1460 	/* Update previous work queue entry's "next" field */
1461 	MLX_SET ( &prev_wqe->next, nopcode, ARBEL_OPCODE_SEND );
1462 	MLX_FILL_3 ( &prev_wqe->next, 1,
1463 		     nds, nds,
1464 		     f, 0,
1465 		     always1, 1 );
1466 
1467 	/* Update doorbell record */
1468 	barrier();
1469 	qp_db_rec = &arbel->db_rec[arbel_send_wq->doorbell_idx].qp;
1470 	MLX_FILL_1 ( qp_db_rec, 0,
1471 		     counter, ( ( wq->next_idx + 1 ) & 0xffff ) );
1472 
1473 	/* Ring doorbell register */
1474 	MLX_FILL_4 ( &db_reg.send, 0,
1475 		     nopcode, ARBEL_OPCODE_SEND,
1476 		     f, 0,
1477 		     wqe_counter, ( wq->next_idx & 0xffff ),
1478 		     wqe_cnt, 1 );
1479 	MLX_FILL_2 ( &db_reg.send, 1,
1480 		     nds, nds,
1481 		     qpn, qp->qpn );
1482 	arbel_ring_doorbell ( arbel, &db_reg, ARBEL_DB_POST_SND_OFFSET );
1483 
1484 	/* Update work queue's index */
1485 	wq->next_idx++;
1486 
1487 	return 0;
1488 }
1489 
1490 /**
1491  * Post receive work queue entry
1492  *
1493  * @v ibdev		Infiniband device
1494  * @v qp		Queue pair
1495  * @v iobuf		I/O buffer
1496  * @ret rc		Return status code
1497  */
arbel_post_recv(struct ib_device * ibdev,struct ib_queue_pair * qp,struct io_buffer * iobuf)1498 static int arbel_post_recv ( struct ib_device *ibdev,
1499 			     struct ib_queue_pair *qp,
1500 			     struct io_buffer *iobuf ) {
1501 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1502 	struct arbel_queue_pair *arbel_qp = ib_qp_get_drvdata ( qp );
1503 	struct ib_work_queue *wq = &qp->recv;
1504 	struct arbel_recv_work_queue *arbel_recv_wq = &arbel_qp->recv;
1505 	struct arbelprm_recv_wqe *wqe;
1506 	struct arbelprm_wqe_segment_data_ptr *data;
1507 	struct ib_global_route_header *grh;
1508 	union arbelprm_doorbell_record *db_rec;
1509 	unsigned int wqe_idx_mask;
1510 
1511 	/* Allocate work queue entry */
1512 	wqe_idx_mask = ( wq->num_wqes - 1 );
1513 	if ( wq->iobufs[wq->next_idx & wqe_idx_mask] ) {
1514 		DBGC ( arbel, "Arbel %p QPN %#lx receive queue full\n",
1515 		       arbel, qp->qpn );
1516 		return -ENOBUFS;
1517 	}
1518 	wq->iobufs[wq->next_idx & wqe_idx_mask] = iobuf;
1519 	wqe = &arbel_recv_wq->wqe[wq->next_idx & wqe_idx_mask].recv;
1520 
1521 	/* Construct work queue entry */
1522 	data = &wqe->data[0];
1523 	if ( arbel_recv_wq->grh ) {
1524 		grh = &arbel_recv_wq->grh[wq->next_idx & wqe_idx_mask];
1525 		MLX_FILL_1 ( data, 0, byte_count, sizeof ( *grh ) );
1526 		MLX_FILL_1 ( data, 1, l_key, arbel->lkey );
1527 		MLX_FILL_H ( data, 2, local_address_h, virt_to_bus ( grh ) );
1528 		MLX_FILL_1 ( data, 3, local_address_l, virt_to_bus ( grh ) );
1529 		data++;
1530 	}
1531 	MLX_FILL_1 ( data, 0, byte_count, iob_tailroom ( iobuf ) );
1532 	MLX_FILL_1 ( data, 1, l_key, arbel->lkey );
1533 	MLX_FILL_H ( data, 2, local_address_h, virt_to_bus ( iobuf->data ) );
1534 	MLX_FILL_1 ( data, 3, local_address_l, virt_to_bus ( iobuf->data ) );
1535 
1536 	/* Update doorbell record */
1537 	barrier();
1538 	db_rec = &arbel->db_rec[arbel_recv_wq->doorbell_idx];
1539 	MLX_FILL_1 ( &db_rec->qp, 0,
1540 		     counter, ( ( wq->next_idx + 1 ) & 0xffff ) );
1541 
1542 	/* Update work queue's index */
1543 	wq->next_idx++;
1544 
1545 	return 0;
1546 }
1547 
1548 /**
1549  * Handle completion
1550  *
1551  * @v ibdev		Infiniband device
1552  * @v cq		Completion queue
1553  * @v cqe		Hardware completion queue entry
1554  * @ret rc		Return status code
1555  */
arbel_complete(struct ib_device * ibdev,struct ib_completion_queue * cq,union arbelprm_completion_entry * cqe)1556 static int arbel_complete ( struct ib_device *ibdev,
1557 			    struct ib_completion_queue *cq,
1558 			    union arbelprm_completion_entry *cqe ) {
1559 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1560 	struct ib_work_queue *wq;
1561 	struct ib_queue_pair *qp;
1562 	struct arbel_queue_pair *arbel_qp;
1563 	struct arbel_send_work_queue *arbel_send_wq;
1564 	struct arbel_recv_work_queue *arbel_recv_wq;
1565 	struct arbelprm_recv_wqe *recv_wqe;
1566 	struct io_buffer *iobuf;
1567 	struct ib_address_vector recv_dest;
1568 	struct ib_address_vector recv_source;
1569 	struct ib_global_route_header *grh;
1570 	struct ib_address_vector *source;
1571 	unsigned int opcode;
1572 	unsigned long qpn;
1573 	int is_send;
1574 	unsigned long wqe_adr;
1575 	unsigned long wqe_idx;
1576 	size_t len;
1577 	int rc = 0;
1578 
1579 	/* Parse completion */
1580 	qpn = MLX_GET ( &cqe->normal, my_qpn );
1581 	is_send = MLX_GET ( &cqe->normal, s );
1582 	wqe_adr = ( MLX_GET ( &cqe->normal, wqe_adr ) << 6 );
1583 	opcode = MLX_GET ( &cqe->normal, opcode );
1584 	if ( opcode >= ARBEL_OPCODE_RECV_ERROR ) {
1585 		/* "s" field is not valid for error opcodes */
1586 		is_send = ( opcode == ARBEL_OPCODE_SEND_ERROR );
1587 		DBGC ( arbel, "Arbel %p CQN %#lx %s QPN %#lx syndrome %#x "
1588 		       "vendor %#x\n", arbel, cq->cqn,
1589 		       ( is_send ? "send" : "recv" ), qpn,
1590 		       MLX_GET ( &cqe->error, syndrome ),
1591 		       MLX_GET ( &cqe->error, vendor_code ) );
1592 		DBGC_HDA ( arbel, virt_to_phys ( cqe ), cqe, sizeof ( *cqe ) );
1593 		rc = -EIO;
1594 		/* Don't return immediately; propagate error to completer */
1595 	}
1596 
1597 	/* Identify work queue */
1598 	wq = ib_find_wq ( cq, qpn, is_send );
1599 	if ( ! wq ) {
1600 		DBGC ( arbel, "Arbel %p CQN %#lx unknown %s QPN %#lx\n",
1601 		       arbel, cq->cqn, ( is_send ? "send" : "recv" ), qpn );
1602 		return -EIO;
1603 	}
1604 	qp = wq->qp;
1605 	arbel_qp = ib_qp_get_drvdata ( qp );
1606 	arbel_send_wq = &arbel_qp->send;
1607 	arbel_recv_wq = &arbel_qp->recv;
1608 
1609 	/* Identify work queue entry index */
1610 	if ( is_send ) {
1611 		wqe_idx = ( ( wqe_adr - virt_to_bus ( arbel_send_wq->wqe ) ) /
1612 			    sizeof ( arbel_send_wq->wqe[0] ) );
1613 		assert ( wqe_idx < qp->send.num_wqes );
1614 	} else {
1615 		wqe_idx = ( ( wqe_adr - virt_to_bus ( arbel_recv_wq->wqe ) ) /
1616 			    sizeof ( arbel_recv_wq->wqe[0] ) );
1617 		assert ( wqe_idx < qp->recv.num_wqes );
1618 	}
1619 
1620 	DBGCP ( arbel, "Arbel %p CQN %#lx QPN %#lx %s WQE %#lx completed:\n",
1621 		arbel, cq->cqn, qp->qpn, ( is_send ? "send" : "recv" ),
1622 		wqe_idx );
1623 	DBGCP_HDA ( arbel, virt_to_phys ( cqe ), cqe, sizeof ( *cqe ) );
1624 
1625 	/* Identify I/O buffer */
1626 	iobuf = wq->iobufs[wqe_idx];
1627 	if ( ! iobuf ) {
1628 		DBGC ( arbel, "Arbel %p CQN %#lx QPN %#lx empty %s WQE %#lx\n",
1629 		       arbel, cq->cqn, qp->qpn, ( is_send ? "send" : "recv" ),
1630 		       wqe_idx );
1631 		return -EIO;
1632 	}
1633 	wq->iobufs[wqe_idx] = NULL;
1634 
1635 	if ( is_send ) {
1636 		/* Hand off to completion handler */
1637 		ib_complete_send ( ibdev, qp, iobuf, rc );
1638 	} else {
1639 		/* Set received length */
1640 		len = MLX_GET ( &cqe->normal, byte_cnt );
1641 		recv_wqe = &arbel_recv_wq->wqe[wqe_idx].recv;
1642 		assert ( MLX_GET ( &recv_wqe->data[0], local_address_l ) ==
1643 			 virt_to_bus ( iobuf->data ) );
1644 		assert ( MLX_GET ( &recv_wqe->data[0], byte_count ) ==
1645 			 iob_tailroom ( iobuf ) );
1646 		MLX_FILL_1 ( &recv_wqe->data[0], 0, byte_count, 0 );
1647 		MLX_FILL_1 ( &recv_wqe->data[0], 1,
1648 			     l_key, ARBEL_INVALID_LKEY );
1649 		memset ( &recv_dest, 0, sizeof ( recv_dest ) );
1650 		recv_dest.qpn = qpn;
1651 		switch ( qp->type ) {
1652 		case IB_QPT_SMI:
1653 		case IB_QPT_GSI:
1654 		case IB_QPT_UD:
1655 			/* Locate corresponding GRH */
1656 			assert ( arbel_recv_wq->grh != NULL );
1657 			grh = &arbel_recv_wq->grh[wqe_idx];
1658 			len -= sizeof ( *grh );
1659 			/* Construct address vector */
1660 			source = &recv_source;
1661 			memset ( source, 0, sizeof ( *source ) );
1662 			source->qpn = MLX_GET ( &cqe->normal, rqpn );
1663 			source->lid = MLX_GET ( &cqe->normal, rlid );
1664 			source->sl = MLX_GET ( &cqe->normal, sl );
1665 			recv_dest.gid_present = source->gid_present =
1666 				MLX_GET ( &cqe->normal, g );
1667 			memcpy ( &recv_dest.gid, &grh->dgid,
1668 				 sizeof ( recv_dest.gid ) );
1669 			memcpy ( &source->gid, &grh->sgid,
1670 				 sizeof ( source->gid ) );
1671 			break;
1672 		case IB_QPT_RC:
1673 			source = &qp->av;
1674 			break;
1675 		default:
1676 			assert ( 0 );
1677 			return -EINVAL;
1678 		}
1679 		assert ( len <= iob_tailroom ( iobuf ) );
1680 		iob_put ( iobuf, len );
1681 		/* Hand off to completion handler */
1682 		ib_complete_recv ( ibdev, qp, &recv_dest, source, iobuf, rc );
1683 	}
1684 
1685 	return rc;
1686 }
1687 
1688 /**
1689  * Poll completion queue
1690  *
1691  * @v ibdev		Infiniband device
1692  * @v cq		Completion queue
1693  */
arbel_poll_cq(struct ib_device * ibdev,struct ib_completion_queue * cq)1694 static void arbel_poll_cq ( struct ib_device *ibdev,
1695 			    struct ib_completion_queue *cq ) {
1696 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1697 	struct arbel_completion_queue *arbel_cq = ib_cq_get_drvdata ( cq );
1698 	struct arbelprm_cq_ci_db_record *ci_db_rec;
1699 	union arbelprm_completion_entry *cqe;
1700 	unsigned int cqe_idx_mask;
1701 	int rc;
1702 
1703 	while ( 1 ) {
1704 		/* Look for completion entry */
1705 		cqe_idx_mask = ( cq->num_cqes - 1 );
1706 		cqe = &arbel_cq->cqe[cq->next_idx & cqe_idx_mask];
1707 		if ( MLX_GET ( &cqe->normal, owner ) != 0 ) {
1708 			/* Entry still owned by hardware; end of poll */
1709 			break;
1710 		}
1711 
1712 		/* Handle completion */
1713 		if ( ( rc = arbel_complete ( ibdev, cq, cqe ) ) != 0 ) {
1714 			DBGC ( arbel, "Arbel %p CQN %#lx failed to complete: "
1715 			       "%s\n", arbel, cq->cqn, strerror ( rc ) );
1716 			DBGC_HD ( arbel, cqe, sizeof ( *cqe ) );
1717 		}
1718 
1719 		/* Return ownership to hardware */
1720 		MLX_FILL_1 ( &cqe->normal, 7, owner, 1 );
1721 		barrier();
1722 		/* Update completion queue's index */
1723 		cq->next_idx++;
1724 		/* Update doorbell record */
1725 		ci_db_rec = &arbel->db_rec[arbel_cq->ci_doorbell_idx].cq_ci;
1726 		MLX_FILL_1 ( ci_db_rec, 0,
1727 			     counter, ( cq->next_idx & 0xffffffffUL ) );
1728 	}
1729 }
1730 
1731 /***************************************************************************
1732  *
1733  * Event queues
1734  *
1735  ***************************************************************************
1736  */
1737 
1738 /**
1739  * Create event queue
1740  *
1741  * @v arbel		Arbel device
1742  * @ret rc		Return status code
1743  */
arbel_create_eq(struct arbel * arbel)1744 static int arbel_create_eq ( struct arbel *arbel ) {
1745 	struct arbel_event_queue *arbel_eq = &arbel->eq;
1746 	struct arbelprm_eqc eqctx;
1747 	struct arbelprm_event_mask mask;
1748 	unsigned int i;
1749 	int rc;
1750 
1751 	/* Select event queue number */
1752 	arbel_eq->eqn = arbel->limits.reserved_eqs;
1753 
1754 	/* Calculate doorbell address */
1755 	arbel_eq->doorbell = ( arbel->eq_ci_doorbells +
1756 			       ARBEL_DB_EQ_OFFSET ( arbel_eq->eqn ) );
1757 
1758 	/* Allocate event queue itself */
1759 	arbel_eq->eqe_size =
1760 		( ARBEL_NUM_EQES * sizeof ( arbel_eq->eqe[0] ) );
1761 	arbel_eq->eqe = malloc_dma ( arbel_eq->eqe_size,
1762 				     sizeof ( arbel_eq->eqe[0] ) );
1763 	if ( ! arbel_eq->eqe ) {
1764 		rc = -ENOMEM;
1765 		goto err_eqe;
1766 	}
1767 	memset ( arbel_eq->eqe, 0, arbel_eq->eqe_size );
1768 	for ( i = 0 ; i < ARBEL_NUM_EQES ; i++ ) {
1769 		MLX_FILL_1 ( &arbel_eq->eqe[i].generic, 7, owner, 1 );
1770 	}
1771 	barrier();
1772 
1773 	/* Hand queue over to hardware */
1774 	memset ( &eqctx, 0, sizeof ( eqctx ) );
1775 	MLX_FILL_1 ( &eqctx, 0, st, 0xa /* "Fired" */ );
1776 	MLX_FILL_H ( &eqctx, 1,
1777 		     start_address_h, virt_to_phys ( arbel_eq->eqe ) );
1778 	MLX_FILL_1 ( &eqctx, 2,
1779 		     start_address_l, virt_to_phys ( arbel_eq->eqe ) );
1780 	MLX_FILL_1 ( &eqctx, 3, log_eq_size, fls ( ARBEL_NUM_EQES - 1 ) );
1781 	MLX_FILL_1 ( &eqctx, 6, pd, ARBEL_GLOBAL_PD );
1782 	MLX_FILL_1 ( &eqctx, 7, lkey, arbel->lkey );
1783 	if ( ( rc = arbel_cmd_sw2hw_eq ( arbel, arbel_eq->eqn,
1784 					 &eqctx ) ) != 0 ) {
1785 		DBGC ( arbel, "Arbel %p EQN %#lx SW2HW_EQ failed: %s\n",
1786 		       arbel, arbel_eq->eqn, strerror ( rc ) );
1787 		goto err_sw2hw_eq;
1788 	}
1789 
1790 	/* Map events to this event queue */
1791 	memset ( &mask, 0xff, sizeof ( mask ) );
1792 	if ( ( rc = arbel_cmd_map_eq ( arbel,
1793 				       ( ARBEL_MAP_EQ | arbel_eq->eqn ),
1794 				       &mask ) ) != 0 ) {
1795 		DBGC ( arbel, "Arbel %p EQN %#lx MAP_EQ failed: %s\n",
1796 		       arbel, arbel_eq->eqn, strerror ( rc )  );
1797 		goto err_map_eq;
1798 	}
1799 
1800 	DBGC ( arbel, "Arbel %p EQN %#lx ring [%08lx,%08lx), doorbell %08lx\n",
1801 	       arbel, arbel_eq->eqn, virt_to_phys ( arbel_eq->eqe ),
1802 	       ( virt_to_phys ( arbel_eq->eqe ) + arbel_eq->eqe_size ),
1803 	       virt_to_phys ( arbel_eq->doorbell ) );
1804 	return 0;
1805 
1806  err_map_eq:
1807 	arbel_cmd_hw2sw_eq ( arbel, arbel_eq->eqn, &eqctx );
1808  err_sw2hw_eq:
1809 	free_dma ( arbel_eq->eqe, arbel_eq->eqe_size );
1810  err_eqe:
1811 	memset ( arbel_eq, 0, sizeof ( *arbel_eq ) );
1812 	return rc;
1813 }
1814 
1815 /**
1816  * Destroy event queue
1817  *
1818  * @v arbel		Arbel device
1819  */
arbel_destroy_eq(struct arbel * arbel)1820 static void arbel_destroy_eq ( struct arbel *arbel ) {
1821 	struct arbel_event_queue *arbel_eq = &arbel->eq;
1822 	struct arbelprm_eqc eqctx;
1823 	struct arbelprm_event_mask mask;
1824 	int rc;
1825 
1826 	/* Unmap events from event queue */
1827 	memset ( &mask, 0, sizeof ( mask ) );
1828 	MLX_FILL_1 ( &mask, 1, port_state_change, 1 );
1829 	if ( ( rc = arbel_cmd_map_eq ( arbel,
1830 				       ( ARBEL_UNMAP_EQ | arbel_eq->eqn ),
1831 				       &mask ) ) != 0 ) {
1832 		DBGC ( arbel, "Arbel %p EQN %#lx FATAL MAP_EQ failed to "
1833 		       "unmap: %s\n", arbel, arbel_eq->eqn, strerror ( rc ) );
1834 		/* Continue; HCA may die but system should survive */
1835 	}
1836 
1837 	/* Take ownership back from hardware */
1838 	if ( ( rc = arbel_cmd_hw2sw_eq ( arbel, arbel_eq->eqn,
1839 					 &eqctx ) ) != 0 ) {
1840 		DBGC ( arbel, "Arbel %p EQN %#lx FATAL HW2SW_EQ failed: %s\n",
1841 		       arbel, arbel_eq->eqn, strerror ( rc ) );
1842 		/* Leak memory and return; at least we avoid corruption */
1843 		return;
1844 	}
1845 
1846 	/* Free memory */
1847 	free_dma ( arbel_eq->eqe, arbel_eq->eqe_size );
1848 	memset ( arbel_eq, 0, sizeof ( *arbel_eq ) );
1849 }
1850 
1851 /**
1852  * Handle port state event
1853  *
1854  * @v arbel		Arbel device
1855  * @v eqe		Port state change event queue entry
1856  */
arbel_event_port_state_change(struct arbel * arbel,union arbelprm_event_entry * eqe)1857 static void arbel_event_port_state_change ( struct arbel *arbel,
1858 					    union arbelprm_event_entry *eqe){
1859 	unsigned int port;
1860 	int link_up;
1861 
1862 	/* Get port and link status */
1863 	port = ( MLX_GET ( &eqe->port_state_change, data.p ) - 1 );
1864 	link_up = ( MLX_GET ( &eqe->generic, event_sub_type ) & 0x04 );
1865 	DBGC ( arbel, "Arbel %p port %d link %s\n", arbel, ( port + 1 ),
1866 	       ( link_up ? "up" : "down" ) );
1867 
1868 	/* Sanity check */
1869 	if ( port >= ARBEL_NUM_PORTS ) {
1870 		DBGC ( arbel, "Arbel %p port %d does not exist!\n",
1871 		       arbel, ( port + 1 ) );
1872 		return;
1873 	}
1874 
1875 	/* Update MAD parameters */
1876 	ib_smc_update ( arbel->ibdev[port], arbel_mad );
1877 }
1878 
1879 /**
1880  * Poll event queue
1881  *
1882  * @v ibdev		Infiniband device
1883  */
arbel_poll_eq(struct ib_device * ibdev)1884 static void arbel_poll_eq ( struct ib_device *ibdev ) {
1885 	struct arbel *arbel = ib_get_drvdata ( ibdev );
1886 	struct arbel_event_queue *arbel_eq = &arbel->eq;
1887 	union arbelprm_event_entry *eqe;
1888 	union arbelprm_eq_doorbell_register db_reg;
1889 	unsigned int eqe_idx_mask;
1890 	unsigned int event_type;
1891 
1892 	/* No event is generated upon reaching INIT, so we must poll
1893 	 * separately for link state changes while we remain DOWN.
1894 	 */
1895 	if ( ib_is_open ( ibdev ) &&
1896 	     ( ibdev->port_state == IB_PORT_STATE_DOWN ) ) {
1897 		ib_smc_update ( ibdev, arbel_mad );
1898 	}
1899 
1900 	/* Poll event queue */
1901 	while ( 1 ) {
1902 		/* Look for event entry */
1903 		eqe_idx_mask = ( ARBEL_NUM_EQES - 1 );
1904 		eqe = &arbel_eq->eqe[arbel_eq->next_idx & eqe_idx_mask];
1905 		if ( MLX_GET ( &eqe->generic, owner ) != 0 ) {
1906 			/* Entry still owned by hardware; end of poll */
1907 			break;
1908 		}
1909 		DBGCP ( arbel, "Arbel %p EQN %#lx event:\n",
1910 			arbel, arbel_eq->eqn );
1911 		DBGCP_HDA ( arbel, virt_to_phys ( eqe ),
1912 			    eqe, sizeof ( *eqe ) );
1913 
1914 		/* Handle event */
1915 		event_type = MLX_GET ( &eqe->generic, event_type );
1916 		switch ( event_type ) {
1917 		case ARBEL_EV_PORT_STATE_CHANGE:
1918 			arbel_event_port_state_change ( arbel, eqe );
1919 			break;
1920 		default:
1921 			DBGC ( arbel, "Arbel %p EQN %#lx unrecognised event "
1922 			       "type %#x:\n",
1923 			       arbel, arbel_eq->eqn, event_type );
1924 			DBGC_HDA ( arbel, virt_to_phys ( eqe ),
1925 				   eqe, sizeof ( *eqe ) );
1926 			break;
1927 		}
1928 
1929 		/* Return ownership to hardware */
1930 		MLX_FILL_1 ( &eqe->generic, 7, owner, 1 );
1931 		barrier();
1932 
1933 		/* Update event queue's index */
1934 		arbel_eq->next_idx++;
1935 
1936 		/* Ring doorbell */
1937 		MLX_FILL_1 ( &db_reg.ci, 0, ci, arbel_eq->next_idx );
1938 		writel ( db_reg.dword[0], arbel_eq->doorbell );
1939 	}
1940 }
1941 
1942 /***************************************************************************
1943  *
1944  * Firmware control
1945  *
1946  ***************************************************************************
1947  */
1948 
1949 /**
1950  * Map virtual to physical address for firmware usage
1951  *
1952  * @v arbel		Arbel device
1953  * @v map		Mapping function
1954  * @v va		Virtual address
1955  * @v pa		Physical address
1956  * @v len		Length of region
1957  * @ret rc		Return status code
1958  */
arbel_map_vpm(struct arbel * arbel,int (* map)(struct arbel * arbel,const struct arbelprm_virtual_physical_mapping *),uint64_t va,physaddr_t pa,size_t len)1959 static int arbel_map_vpm ( struct arbel *arbel,
1960 			   int ( *map ) ( struct arbel *arbel,
1961 			     const struct arbelprm_virtual_physical_mapping* ),
1962 			   uint64_t va, physaddr_t pa, size_t len ) {
1963 	struct arbelprm_virtual_physical_mapping mapping;
1964 	physaddr_t start;
1965 	physaddr_t low;
1966 	physaddr_t high;
1967 	physaddr_t end;
1968 	size_t size;
1969 	int rc;
1970 
1971 	/* Sanity checks */
1972 	assert ( ( va & ( ARBEL_PAGE_SIZE - 1 ) ) == 0 );
1973 	assert ( ( pa & ( ARBEL_PAGE_SIZE - 1 ) ) == 0 );
1974 	assert ( ( len & ( ARBEL_PAGE_SIZE - 1 ) ) == 0 );
1975 	assert ( len != 0 );
1976 
1977 	/* Calculate starting points */
1978 	start = pa;
1979 	end = ( start + len );
1980 	size = ( 1UL << ( fls ( start ^ end ) - 1 ) );
1981 	low = high = ( end & ~( size - 1 ) );
1982 	assert ( start < low );
1983 	assert ( high <= end );
1984 
1985 	/* These mappings tend to generate huge volumes of
1986 	 * uninteresting debug data, which basically makes it
1987 	 * impossible to use debugging otherwise.
1988 	 */
1989 	DBG_DISABLE ( DBGLVL_LOG | DBGLVL_EXTRA );
1990 
1991 	/* Map blocks in descending order of size */
1992 	while ( size >= ARBEL_PAGE_SIZE ) {
1993 
1994 		/* Find the next candidate block */
1995 		if ( ( low - size ) >= start ) {
1996 			low -= size;
1997 			pa = low;
1998 		} else if ( high <= ( end - size ) ) {
1999 			pa = high;
2000 			high += size;
2001 		} else {
2002 			size >>= 1;
2003 			continue;
2004 		}
2005 		assert ( ( va & ( size - 1 ) ) == 0 );
2006 		assert ( ( pa & ( size - 1 ) ) == 0 );
2007 
2008 		/* Map this block */
2009 		memset ( &mapping, 0, sizeof ( mapping ) );
2010 		MLX_FILL_1 ( &mapping, 0, va_h, ( va >> 32 ) );
2011 		MLX_FILL_1 ( &mapping, 1, va_l, ( va >> 12 ) );
2012 		MLX_FILL_H ( &mapping, 2, pa_h, pa );
2013 		MLX_FILL_2 ( &mapping, 3,
2014 			     log2size, ( ( fls ( size ) - 1 ) - 12 ),
2015 			     pa_l, ( pa >> 12 ) );
2016 		if ( ( rc = map ( arbel, &mapping ) ) != 0 ) {
2017 			DBG_ENABLE ( DBGLVL_LOG | DBGLVL_EXTRA );
2018 			DBGC ( arbel, "Arbel %p could not map %08llx+%zx to "
2019 			       "%08lx: %s\n",
2020 			       arbel, va, size, pa, strerror ( rc ) );
2021 			return rc;
2022 		}
2023 		va += size;
2024 	}
2025 	assert ( low == start );
2026 	assert ( high == end );
2027 
2028 	DBG_ENABLE ( DBGLVL_LOG | DBGLVL_EXTRA );
2029 	return 0;
2030 }
2031 
2032 /**
2033  * Start firmware running
2034  *
2035  * @v arbel		Arbel device
2036  * @ret rc		Return status code
2037  */
arbel_start_firmware(struct arbel * arbel)2038 static int arbel_start_firmware ( struct arbel *arbel ) {
2039 	struct arbelprm_query_fw fw;
2040 	struct arbelprm_access_lam lam;
2041 	unsigned int fw_pages;
2042 	size_t fw_len;
2043 	physaddr_t fw_base;
2044 	uint64_t eq_set_ci_base_addr;
2045 	int rc;
2046 
2047 	/* Get firmware parameters */
2048 	if ( ( rc = arbel_cmd_query_fw ( arbel, &fw ) ) != 0 ) {
2049 		DBGC ( arbel, "Arbel %p could not query firmware: %s\n",
2050 		       arbel, strerror ( rc ) );
2051 		goto err_query_fw;
2052 	}
2053 	DBGC ( arbel, "Arbel %p firmware version %d.%d.%d\n", arbel,
2054 	       MLX_GET ( &fw, fw_rev_major ), MLX_GET ( &fw, fw_rev_minor ),
2055 	       MLX_GET ( &fw, fw_rev_subminor ) );
2056 	fw_pages = MLX_GET ( &fw, fw_pages );
2057 	DBGC ( arbel, "Arbel %p requires %d kB for firmware\n",
2058 	       arbel, ( fw_pages * 4 ) );
2059 	eq_set_ci_base_addr =
2060 		( ( (uint64_t) MLX_GET ( &fw, eq_set_ci_base_addr_h ) << 32 ) |
2061 		  ( (uint64_t) MLX_GET ( &fw, eq_set_ci_base_addr_l ) ) );
2062 	arbel->eq_ci_doorbells = ioremap ( eq_set_ci_base_addr, 0x200 );
2063 
2064 	/* Enable locally-attached memory.  Ignore failure; there may
2065 	 * be no attached memory.
2066 	 */
2067 	arbel_cmd_enable_lam ( arbel, &lam );
2068 
2069 	/* Allocate firmware pages and map firmware area */
2070 	fw_len = ( fw_pages * ARBEL_PAGE_SIZE );
2071 	if ( ! arbel->firmware_area ) {
2072 		arbel->firmware_len = fw_len;
2073 		arbel->firmware_area = umalloc ( arbel->firmware_len );
2074 		if ( ! arbel->firmware_area ) {
2075 			rc = -ENOMEM;
2076 			goto err_alloc_fa;
2077 		}
2078 	} else {
2079 		assert ( arbel->firmware_len == fw_len );
2080 	}
2081 	fw_base = user_to_phys ( arbel->firmware_area, 0 );
2082 	DBGC ( arbel, "Arbel %p firmware area at [%08lx,%08lx)\n",
2083 	       arbel, fw_base, ( fw_base + fw_len ) );
2084 	if ( ( rc = arbel_map_vpm ( arbel, arbel_cmd_map_fa,
2085 				    0, fw_base, fw_len ) ) != 0 ) {
2086 		DBGC ( arbel, "Arbel %p could not map firmware: %s\n",
2087 		       arbel, strerror ( rc ) );
2088 		goto err_map_fa;
2089 	}
2090 
2091 	/* Start firmware */
2092 	if ( ( rc = arbel_cmd_run_fw ( arbel ) ) != 0 ) {
2093 		DBGC ( arbel, "Arbel %p could not run firmware: %s\n",
2094 		       arbel, strerror ( rc ) );
2095 		goto err_run_fw;
2096 	}
2097 
2098 	DBGC ( arbel, "Arbel %p firmware started\n", arbel );
2099 	return 0;
2100 
2101  err_run_fw:
2102 	arbel_cmd_unmap_fa ( arbel );
2103  err_map_fa:
2104  err_alloc_fa:
2105  err_query_fw:
2106 	return rc;
2107 }
2108 
2109 /**
2110  * Stop firmware running
2111  *
2112  * @v arbel		Arbel device
2113  */
arbel_stop_firmware(struct arbel * arbel)2114 static void arbel_stop_firmware ( struct arbel *arbel ) {
2115 	int rc;
2116 
2117 	if ( ( rc = arbel_cmd_unmap_fa ( arbel ) ) != 0 ) {
2118 		DBGC ( arbel, "Arbel %p FATAL could not stop firmware: %s\n",
2119 		       arbel, strerror ( rc ) );
2120 		/* Leak memory and return; at least we avoid corruption */
2121 		arbel->firmware_area = UNULL;
2122 		return;
2123 	}
2124 }
2125 
2126 /***************************************************************************
2127  *
2128  * Infinihost Context Memory management
2129  *
2130  ***************************************************************************
2131  */
2132 
2133 /**
2134  * Get device limits
2135  *
2136  * @v arbel		Arbel device
2137  * @ret rc		Return status code
2138  */
arbel_get_limits(struct arbel * arbel)2139 static int arbel_get_limits ( struct arbel *arbel ) {
2140 	struct arbelprm_query_dev_lim dev_lim;
2141 	int rc;
2142 
2143 	if ( ( rc = arbel_cmd_query_dev_lim ( arbel, &dev_lim ) ) != 0 ) {
2144 		DBGC ( arbel, "Arbel %p could not get device limits: %s\n",
2145 		       arbel, strerror ( rc ) );
2146 		return rc;
2147 	}
2148 
2149 	arbel->limits.reserved_qps =
2150 		( 1 << MLX_GET ( &dev_lim, log2_rsvd_qps ) );
2151 	arbel->limits.qpc_entry_size = MLX_GET ( &dev_lim, qpc_entry_sz );
2152 	arbel->limits.eqpc_entry_size = MLX_GET ( &dev_lim, eqpc_entry_sz );
2153 	arbel->limits.reserved_srqs =
2154 		( 1 << MLX_GET ( &dev_lim, log2_rsvd_srqs ) );
2155 	arbel->limits.srqc_entry_size = MLX_GET ( &dev_lim, srq_entry_sz );
2156 	arbel->limits.reserved_ees =
2157 		( 1 << MLX_GET ( &dev_lim, log2_rsvd_ees ) );
2158 	arbel->limits.eec_entry_size = MLX_GET ( &dev_lim, eec_entry_sz );
2159 	arbel->limits.eeec_entry_size = MLX_GET ( &dev_lim, eeec_entry_sz );
2160 	arbel->limits.reserved_cqs =
2161 		( 1 << MLX_GET ( &dev_lim, log2_rsvd_cqs ) );
2162 	arbel->limits.cqc_entry_size = MLX_GET ( &dev_lim, cqc_entry_sz );
2163 	arbel->limits.reserved_mtts =
2164 		( 1 << MLX_GET ( &dev_lim, log2_rsvd_mtts ) );
2165 	arbel->limits.mtt_entry_size = MLX_GET ( &dev_lim, mtt_entry_sz );
2166 	arbel->limits.reserved_mrws =
2167 		( 1 << MLX_GET ( &dev_lim, log2_rsvd_mrws ) );
2168 	arbel->limits.mpt_entry_size = MLX_GET ( &dev_lim, mpt_entry_sz );
2169 	arbel->limits.reserved_rdbs =
2170 		( 1 << MLX_GET ( &dev_lim, log2_rsvd_rdbs ) );
2171 	arbel->limits.reserved_eqs = MLX_GET ( &dev_lim, num_rsvd_eqs );
2172 	arbel->limits.eqc_entry_size = MLX_GET ( &dev_lim, eqc_entry_sz );
2173 	arbel->limits.reserved_uars = MLX_GET ( &dev_lim, num_rsvd_uars );
2174 	arbel->limits.uar_scratch_entry_size =
2175 		MLX_GET ( &dev_lim, uar_scratch_entry_sz );
2176 
2177 	DBGC ( arbel, "Arbel %p reserves %d x %#zx QPC, %d x %#zx EQPC, "
2178 	       "%d x %#zx SRQC\n", arbel,
2179 	       arbel->limits.reserved_qps, arbel->limits.qpc_entry_size,
2180 	       arbel->limits.reserved_qps, arbel->limits.eqpc_entry_size,
2181 	       arbel->limits.reserved_srqs, arbel->limits.srqc_entry_size );
2182 	DBGC ( arbel, "Arbel %p reserves %d x %#zx EEC, %d x %#zx EEEC, "
2183 	       "%d x %#zx CQC\n", arbel,
2184 	       arbel->limits.reserved_ees, arbel->limits.eec_entry_size,
2185 	       arbel->limits.reserved_ees, arbel->limits.eeec_entry_size,
2186 	       arbel->limits.reserved_cqs, arbel->limits.cqc_entry_size );
2187 	DBGC ( arbel, "Arbel %p reserves %d x %#zx EQC, %d x %#zx MTT, "
2188 	       "%d x %#zx MPT\n", arbel,
2189 	       arbel->limits.reserved_eqs, arbel->limits.eqc_entry_size,
2190 	       arbel->limits.reserved_mtts, arbel->limits.mtt_entry_size,
2191 	       arbel->limits.reserved_mrws, arbel->limits.mpt_entry_size );
2192 	DBGC ( arbel, "Arbel %p reserves %d x %#zx RDB, %d x %#zx UAR, "
2193 	       "%d x %#zx UAR scratchpad\n", arbel,
2194 	       arbel->limits.reserved_rdbs, ARBEL_RDB_ENTRY_SIZE,
2195 	       arbel->limits.reserved_uars, ARBEL_PAGE_SIZE,
2196 	       arbel->limits.reserved_uars,
2197 	       arbel->limits.uar_scratch_entry_size );
2198 
2199 	return 0;
2200 }
2201 
2202 /**
2203  * Align ICM table
2204  *
2205  * @v icm_offset	Current ICM offset
2206  * @v len		ICM table length
2207  * @ret icm_offset	ICM offset
2208  */
icm_align(size_t icm_offset,size_t len)2209 static size_t icm_align ( size_t icm_offset, size_t len ) {
2210 
2211 	/* Round up to a multiple of the table size */
2212 	assert ( len == ( 1UL << ( fls ( len ) - 1 ) ) );
2213 	return ( ( icm_offset + len - 1 ) & ~( len - 1 ) );
2214 }
2215 
2216 /**
2217  * Allocate ICM
2218  *
2219  * @v arbel		Arbel device
2220  * @v init_hca		INIT_HCA structure to fill in
2221  * @ret rc		Return status code
2222  */
arbel_alloc_icm(struct arbel * arbel,struct arbelprm_init_hca * init_hca)2223 static int arbel_alloc_icm ( struct arbel *arbel,
2224 			     struct arbelprm_init_hca *init_hca ) {
2225 	struct arbelprm_scalar_parameter icm_size;
2226 	struct arbelprm_scalar_parameter icm_aux_size;
2227 	struct arbelprm_scalar_parameter unmap_icm;
2228 	union arbelprm_doorbell_record *db_rec;
2229 	size_t icm_offset = 0;
2230 	unsigned int log_num_uars, log_num_qps, log_num_srqs, log_num_ees;
2231 	unsigned int log_num_cqs, log_num_mtts, log_num_mpts, log_num_rdbs;
2232 	unsigned int log_num_eqs, log_num_mcs;
2233 	size_t icm_len, icm_aux_len;
2234 	size_t len;
2235 	physaddr_t icm_phys;
2236 	int rc;
2237 
2238 	/* Calculate number of each object type within ICM */
2239 	log_num_qps = fls ( arbel->limits.reserved_qps +
2240 			    ARBEL_RSVD_SPECIAL_QPS + ARBEL_MAX_QPS - 1 );
2241 	log_num_srqs = fls ( arbel->limits.reserved_srqs - 1 );
2242 	log_num_ees = fls ( arbel->limits.reserved_ees - 1 );
2243 	log_num_cqs = fls ( arbel->limits.reserved_cqs + ARBEL_MAX_CQS - 1 );
2244 	log_num_eqs = fls ( arbel->limits.reserved_eqs + ARBEL_MAX_EQS - 1 );
2245 	log_num_mtts = fls ( arbel->limits.reserved_mtts - 1 );
2246 	log_num_mpts = fls ( arbel->limits.reserved_mrws + 1 - 1 );
2247 	log_num_rdbs = fls ( arbel->limits.reserved_rdbs +
2248 			     ARBEL_RSVD_SPECIAL_QPS + ARBEL_MAX_QPS - 1 );
2249 	log_num_uars = fls ( arbel->limits.reserved_uars +
2250 			     1 /* single UAR used */ - 1 );
2251 	log_num_mcs = ARBEL_LOG_MULTICAST_HASH_SIZE;
2252 
2253 	/* Queue pair contexts */
2254 	len = ( ( 1 << log_num_qps ) * arbel->limits.qpc_entry_size );
2255 	icm_offset = icm_align ( icm_offset, len );
2256 	MLX_FILL_2 ( init_hca, 13,
2257 		     qpc_eec_cqc_eqc_rdb_parameters.qpc_base_addr_l,
2258 		     ( icm_offset >> 7 ),
2259 		     qpc_eec_cqc_eqc_rdb_parameters.log_num_of_qp,
2260 		     log_num_qps );
2261 	DBGC ( arbel, "Arbel %p ICM QPC is %d x %#zx at [%zx,%zx)\n",
2262 	       arbel, ( 1 << log_num_qps ), arbel->limits.qpc_entry_size,
2263 	       icm_offset, ( icm_offset + len ) );
2264 	icm_offset += len;
2265 
2266 	/* Extended queue pair contexts */
2267 	len = ( ( 1 << log_num_qps ) * arbel->limits.eqpc_entry_size );
2268 	icm_offset = icm_align ( icm_offset, len );
2269 	MLX_FILL_1 ( init_hca, 25,
2270 		     qpc_eec_cqc_eqc_rdb_parameters.eqpc_base_addr_l,
2271 		     icm_offset );
2272 	DBGC ( arbel, "Arbel %p ICM EQPC is %d x %#zx at [%zx,%zx)\n",
2273 	       arbel, ( 1 << log_num_qps ), arbel->limits.eqpc_entry_size,
2274 	       icm_offset, ( icm_offset + len ) );
2275 	icm_offset += len;
2276 
2277 	/* Completion queue contexts */
2278 	len = ( ( 1 << log_num_cqs ) * arbel->limits.cqc_entry_size );
2279 	icm_offset = icm_align ( icm_offset, len );
2280 	MLX_FILL_2 ( init_hca, 21,
2281 		     qpc_eec_cqc_eqc_rdb_parameters.cqc_base_addr_l,
2282 		     ( icm_offset >> 6 ),
2283 		     qpc_eec_cqc_eqc_rdb_parameters.log_num_of_cq,
2284 		     log_num_cqs );
2285 	DBGC ( arbel, "Arbel %p ICM CQC is %d x %#zx at [%zx,%zx)\n",
2286 	       arbel, ( 1 << log_num_cqs ), arbel->limits.cqc_entry_size,
2287 	       icm_offset, ( icm_offset + len ) );
2288 	icm_offset += len;
2289 
2290 	/* Event queue contexts */
2291 	len = ( ( 1 << log_num_eqs ) * arbel->limits.eqc_entry_size );
2292 	icm_offset = icm_align ( icm_offset, len );
2293 	MLX_FILL_2 ( init_hca, 33,
2294 		     qpc_eec_cqc_eqc_rdb_parameters.eqc_base_addr_l,
2295 		     ( icm_offset >> 6 ),
2296 		     qpc_eec_cqc_eqc_rdb_parameters.log_num_eq,
2297 		     log_num_eqs );
2298 	DBGC ( arbel, "Arbel %p ICM EQC is %d x %#zx at [%zx,%zx)\n",
2299 	       arbel, ( 1 << log_num_eqs ), arbel->limits.eqc_entry_size,
2300 	       icm_offset, ( icm_offset + len ) );
2301 	icm_offset += len;
2302 
2303 	/* End-to-end contexts */
2304 	len = ( ( 1 << log_num_ees ) * arbel->limits.eec_entry_size );
2305 	icm_offset = icm_align ( icm_offset, len );
2306 	MLX_FILL_2 ( init_hca, 17,
2307 		     qpc_eec_cqc_eqc_rdb_parameters.eec_base_addr_l,
2308 		     ( icm_offset >> 7 ),
2309 		     qpc_eec_cqc_eqc_rdb_parameters.log_num_of_ee,
2310 		     log_num_ees );
2311 	DBGC ( arbel, "Arbel %p ICM EEC is %d x %#zx at [%zx,%zx)\n",
2312 	       arbel, ( 1 << log_num_ees ), arbel->limits.eec_entry_size,
2313 	       icm_offset, ( icm_offset + len ) );
2314 	icm_offset += len;
2315 
2316 	/* Shared receive queue contexts */
2317 	len = ( ( 1 << log_num_srqs ) * arbel->limits.srqc_entry_size );
2318 	icm_offset = icm_align ( icm_offset, len );
2319 	MLX_FILL_2 ( init_hca, 19,
2320 		     qpc_eec_cqc_eqc_rdb_parameters.srqc_base_addr_l,
2321 		     ( icm_offset >> 5 ),
2322 		     qpc_eec_cqc_eqc_rdb_parameters.log_num_of_srq,
2323 		     log_num_srqs );
2324 	DBGC ( arbel, "Arbel %p ICM SRQC is %d x %#zx at [%zx,%zx)\n",
2325 	       arbel, ( 1 << log_num_srqs ), arbel->limits.srqc_entry_size,
2326 	       icm_offset, ( icm_offset + len ) );
2327 	icm_offset += len;
2328 
2329 	/* Memory protection table */
2330 	len = ( ( 1 << log_num_mpts ) * arbel->limits.mpt_entry_size );
2331 	icm_offset = icm_align ( icm_offset, len );
2332 	MLX_FILL_1 ( init_hca, 61,
2333 		     tpt_parameters.mpt_base_adr_l, icm_offset );
2334 	MLX_FILL_1 ( init_hca, 62,
2335 		     tpt_parameters.log_mpt_sz, log_num_mpts );
2336 	DBGC ( arbel, "Arbel %p ICM MPT is %d x %#zx at [%zx,%zx)\n",
2337 	       arbel, ( 1 << log_num_mpts ), arbel->limits.mpt_entry_size,
2338 	       icm_offset, ( icm_offset + len ) );
2339 	icm_offset += len;
2340 
2341 	/* Remote read data base table */
2342 	len = ( ( 1 << log_num_rdbs ) * ARBEL_RDB_ENTRY_SIZE );
2343 	icm_offset = icm_align ( icm_offset, len );
2344 	MLX_FILL_1 ( init_hca, 37,
2345 		     qpc_eec_cqc_eqc_rdb_parameters.rdb_base_addr_l,
2346 		     icm_offset );
2347 	DBGC ( arbel, "Arbel %p ICM RDB is %d x %#zx at [%zx,%zx)\n",
2348 	       arbel, ( 1 << log_num_rdbs ), ARBEL_RDB_ENTRY_SIZE,
2349 	       icm_offset, ( icm_offset + len ) );
2350 	icm_offset += len;
2351 
2352 	/* Extended end-to-end contexts */
2353 	len = ( ( 1 << log_num_ees ) * arbel->limits.eeec_entry_size );
2354 	icm_offset = icm_align ( icm_offset, len );
2355 	MLX_FILL_1 ( init_hca, 29,
2356 		     qpc_eec_cqc_eqc_rdb_parameters.eeec_base_addr_l,
2357 		     icm_offset );
2358 	DBGC ( arbel, "Arbel %p ICM EEEC is %d x %#zx at [%zx,%zx)\n",
2359 	       arbel, ( 1 << log_num_ees ), arbel->limits.eeec_entry_size,
2360 	       icm_offset, ( icm_offset + len ) );
2361 	icm_offset += len;
2362 
2363 	/* Multicast table */
2364 	len = ( ( 1 << log_num_mcs ) * sizeof ( struct arbelprm_mgm_entry ) );
2365 	icm_offset = icm_align ( icm_offset, len );
2366 	MLX_FILL_1 ( init_hca, 49,
2367 		     multicast_parameters.mc_base_addr_l, icm_offset );
2368 	MLX_FILL_1 ( init_hca, 52,
2369 		     multicast_parameters.log_mc_table_entry_sz,
2370 		     fls ( sizeof ( struct arbelprm_mgm_entry ) - 1 ) );
2371 	MLX_FILL_1 ( init_hca, 53,
2372 		     multicast_parameters.mc_table_hash_sz,
2373 		     ( 1 << log_num_mcs ) );
2374 	MLX_FILL_1 ( init_hca, 54,
2375 		     multicast_parameters.log_mc_table_sz,
2376 		     log_num_mcs /* Only one entry per hash */ );
2377 	DBGC ( arbel, "Arbel %p ICM MC is %d x %#zx at [%zx,%zx)\n", arbel,
2378 	       ( 1 << log_num_mcs ), sizeof ( struct arbelprm_mgm_entry ),
2379 	       icm_offset, ( icm_offset + len ) );
2380 	icm_offset += len;
2381 
2382 	/* Memory translation table */
2383 	len = ( ( 1 << log_num_mtts ) * arbel->limits.mtt_entry_size );
2384 	icm_offset = icm_align ( icm_offset, len );
2385 	MLX_FILL_1 ( init_hca, 65,
2386 		     tpt_parameters.mtt_base_addr_l, icm_offset );
2387 	DBGC ( arbel, "Arbel %p ICM MTT is %d x %#zx at [%zx,%zx)\n",
2388 	       arbel, ( 1 << log_num_mtts ), arbel->limits.mtt_entry_size,
2389 	       icm_offset, ( icm_offset + len ) );
2390 	icm_offset += len;
2391 
2392 	/* User access region scratchpads */
2393 	len = ( ( 1 << log_num_uars ) * arbel->limits.uar_scratch_entry_size );
2394 	icm_offset = icm_align ( icm_offset, len );
2395 	MLX_FILL_1 ( init_hca, 77,
2396 		     uar_parameters.uar_scratch_base_addr_l, icm_offset );
2397 	DBGC ( arbel, "Arbel %p UAR scratchpad is %d x %#zx at [%zx,%zx)\n",
2398 	       arbel, ( 1 << log_num_uars ),
2399 	       arbel->limits.uar_scratch_entry_size,
2400 	       icm_offset, ( icm_offset + len ) );
2401 	icm_offset += len;
2402 
2403 	/* Record amount of ICM to be allocated */
2404 	icm_offset = icm_align ( icm_offset, ARBEL_PAGE_SIZE );
2405 	icm_len = icm_offset;
2406 
2407 	/* User access region contexts
2408 	 *
2409 	 * The reserved UAR(s) do not need to be backed by physical
2410 	 * memory, and our UAR is allocated separately; neither are
2411 	 * part of the umalloc()ed ICM block, but both contribute to
2412 	 * the total length of ICM virtual address space.
2413 	 */
2414 	len = ( ( 1 << log_num_uars ) * ARBEL_PAGE_SIZE );
2415 	icm_offset = icm_align ( icm_offset, len );
2416 	MLX_FILL_1 ( init_hca, 74, uar_parameters.log_max_uars, log_num_uars );
2417 	MLX_FILL_1 ( init_hca, 79,
2418 		     uar_parameters.uar_context_base_addr_l, icm_offset );
2419 	arbel->db_rec_offset =
2420 		( icm_offset +
2421 		  ( arbel->limits.reserved_uars * ARBEL_PAGE_SIZE ) );
2422 	DBGC ( arbel, "Arbel %p UAR is %d x %#zx at [%zx,%zx), doorbells "
2423 	       "[%zx,%zx)\n", arbel, ( 1 << log_num_uars ), ARBEL_PAGE_SIZE,
2424 	       icm_offset, ( icm_offset + len ), arbel->db_rec_offset,
2425 	       ( arbel->db_rec_offset + ARBEL_PAGE_SIZE ) );
2426 	icm_offset += len;
2427 
2428 	/* Get ICM auxiliary area size */
2429 	memset ( &icm_size, 0, sizeof ( icm_size ) );
2430 	MLX_FILL_1 ( &icm_size, 1, value, icm_len );
2431 	if ( ( rc = arbel_cmd_set_icm_size ( arbel, &icm_size,
2432 					     &icm_aux_size ) ) != 0 ) {
2433 		DBGC ( arbel, "Arbel %p could not set ICM size: %s\n",
2434 		       arbel, strerror ( rc ) );
2435 		goto err_set_icm_size;
2436 	}
2437 	icm_aux_len = ( MLX_GET ( &icm_aux_size, value ) * ARBEL_PAGE_SIZE );
2438 
2439 	/* Allocate ICM data and auxiliary area */
2440 	DBGC ( arbel, "Arbel %p requires %zd kB ICM and %zd kB AUX ICM\n",
2441 	       arbel, ( icm_len / 1024 ), ( icm_aux_len / 1024 ) );
2442 	if ( ! arbel->icm ) {
2443 		arbel->icm_len = icm_len;
2444 		arbel->icm_aux_len = icm_aux_len;
2445 		arbel->icm = umalloc ( arbel->icm_len + arbel->icm_aux_len );
2446 		if ( ! arbel->icm ) {
2447 			rc = -ENOMEM;
2448 			goto err_alloc_icm;
2449 		}
2450 	} else {
2451 		assert ( arbel->icm_len == icm_len );
2452 		assert ( arbel->icm_aux_len == icm_aux_len );
2453 	}
2454 	icm_phys = user_to_phys ( arbel->icm, 0 );
2455 
2456 	/* Allocate doorbell UAR */
2457 	arbel->db_rec = malloc_dma ( ARBEL_PAGE_SIZE, ARBEL_PAGE_SIZE );
2458 	if ( ! arbel->db_rec ) {
2459 		rc = -ENOMEM;
2460 		goto err_alloc_doorbell;
2461 	}
2462 
2463 	/* Map ICM auxiliary area */
2464 	DBGC ( arbel, "Arbel %p ICM AUX at [%08lx,%08lx)\n",
2465 	       arbel, icm_phys, ( icm_phys + arbel->icm_aux_len ) );
2466 	if ( ( rc = arbel_map_vpm ( arbel, arbel_cmd_map_icm_aux,
2467 				    0, icm_phys, arbel->icm_aux_len ) ) != 0 ){
2468 		DBGC ( arbel, "Arbel %p could not map AUX ICM: %s\n",
2469 		       arbel, strerror ( rc ) );
2470 		goto err_map_icm_aux;
2471 	}
2472 	icm_phys += arbel->icm_aux_len;
2473 
2474 	/* Map ICM area */
2475 	DBGC ( arbel, "Arbel %p ICM at [%08lx,%08lx)\n",
2476 	       arbel, icm_phys, ( icm_phys + arbel->icm_len ) );
2477 	if ( ( rc = arbel_map_vpm ( arbel, arbel_cmd_map_icm,
2478 				    0, icm_phys, arbel->icm_len ) ) != 0 ) {
2479 		DBGC ( arbel, "Arbel %p could not map ICM: %s\n",
2480 		       arbel, strerror ( rc ) );
2481 		goto err_map_icm;
2482 	}
2483 	icm_phys += arbel->icm_len;
2484 
2485 	/* Map doorbell UAR */
2486 	DBGC ( arbel, "Arbel %p UAR at [%08lx,%08lx)\n",
2487 	       arbel, virt_to_phys ( arbel->db_rec ),
2488 	       ( virt_to_phys ( arbel->db_rec ) + ARBEL_PAGE_SIZE ) );
2489 	if ( ( rc = arbel_map_vpm ( arbel, arbel_cmd_map_icm,
2490 				    arbel->db_rec_offset,
2491 				    virt_to_phys ( arbel->db_rec ),
2492 				    ARBEL_PAGE_SIZE ) ) != 0 ) {
2493 		DBGC ( arbel, "Arbel %p could not map doorbell UAR: %s\n",
2494 		       arbel, strerror ( rc ) );
2495 		goto err_map_doorbell;
2496 	}
2497 
2498 	/* Initialise doorbell records */
2499 	memset ( arbel->db_rec, 0, ARBEL_PAGE_SIZE );
2500 	db_rec = &arbel->db_rec[ARBEL_GROUP_SEPARATOR_DOORBELL];
2501 	MLX_FILL_1 ( &db_rec->qp, 1, res, ARBEL_UAR_RES_GROUP_SEP );
2502 
2503 	return 0;
2504 
2505 	memset ( &unmap_icm, 0, sizeof ( unmap_icm ) );
2506 	MLX_FILL_1 ( &unmap_icm, 1, value, arbel->db_rec_offset );
2507 	arbel_cmd_unmap_icm ( arbel, 1, &unmap_icm );
2508  err_map_doorbell:
2509 	memset ( &unmap_icm, 0, sizeof ( unmap_icm ) );
2510 	arbel_cmd_unmap_icm ( arbel, ( arbel->icm_len / ARBEL_PAGE_SIZE ),
2511 			      &unmap_icm );
2512  err_map_icm:
2513 	arbel_cmd_unmap_icm_aux ( arbel );
2514  err_map_icm_aux:
2515 	free_dma ( arbel->db_rec, ARBEL_PAGE_SIZE );
2516 	arbel->db_rec= NULL;
2517  err_alloc_doorbell:
2518  err_alloc_icm:
2519  err_set_icm_size:
2520 	return rc;
2521 }
2522 
2523 /**
2524  * Free ICM
2525  *
2526  * @v arbel		Arbel device
2527  */
arbel_free_icm(struct arbel * arbel)2528 static void arbel_free_icm ( struct arbel *arbel ) {
2529 	struct arbelprm_scalar_parameter unmap_icm;
2530 
2531 	memset ( &unmap_icm, 0, sizeof ( unmap_icm ) );
2532 	MLX_FILL_1 ( &unmap_icm, 1, value, arbel->db_rec_offset );
2533 	arbel_cmd_unmap_icm ( arbel, 1, &unmap_icm );
2534 	memset ( &unmap_icm, 0, sizeof ( unmap_icm ) );
2535 	arbel_cmd_unmap_icm ( arbel, ( arbel->icm_len / ARBEL_PAGE_SIZE ),
2536 			      &unmap_icm );
2537 	arbel_cmd_unmap_icm_aux ( arbel );
2538 	free_dma ( arbel->db_rec, ARBEL_PAGE_SIZE );
2539 	arbel->db_rec = NULL;
2540 }
2541 
2542 /***************************************************************************
2543  *
2544  * Initialisation and teardown
2545  *
2546  ***************************************************************************
2547  */
2548 
2549 /**
2550  * Reset device
2551  *
2552  * @v arbel		Arbel device
2553  */
arbel_reset(struct arbel * arbel)2554 static void arbel_reset ( struct arbel *arbel ) {
2555 	struct pci_device *pci = arbel->pci;
2556 	struct pci_config_backup backup;
2557 	static const uint8_t backup_exclude[] =
2558 		PCI_CONFIG_BACKUP_EXCLUDE ( 0x58, 0x5c );
2559 	uint16_t vendor;
2560 	unsigned int i;
2561 
2562 	/* Perform device reset and preserve PCI configuration */
2563 	pci_backup ( pci, &backup, backup_exclude );
2564 	writel ( ARBEL_RESET_MAGIC,
2565 		 ( arbel->config + ARBEL_RESET_OFFSET ) );
2566 	for ( i = 0 ; i < ARBEL_RESET_WAIT_TIME_MS ; i++ ) {
2567 		mdelay ( 1 );
2568 		pci_read_config_word ( pci, PCI_VENDOR_ID, &vendor );
2569 		if ( vendor != 0xffff )
2570 			break;
2571 	}
2572 	pci_restore ( pci, &backup, backup_exclude );
2573 }
2574 
2575 /**
2576  * Set up memory protection table
2577  *
2578  * @v arbel		Arbel device
2579  * @ret rc		Return status code
2580  */
arbel_setup_mpt(struct arbel * arbel)2581 static int arbel_setup_mpt ( struct arbel *arbel ) {
2582 	struct arbelprm_mpt mpt;
2583 	uint32_t key;
2584 	int rc;
2585 
2586 	/* Derive key */
2587 	key = ( arbel->limits.reserved_mrws | ARBEL_MKEY_PREFIX );
2588 	arbel->lkey = ( ( key << 8 ) | ( key >> 24 ) );
2589 
2590 	/* Initialise memory protection table */
2591 	memset ( &mpt, 0, sizeof ( mpt ) );
2592 	MLX_FILL_7 ( &mpt, 0,
2593 		     a, 1,
2594 		     rw, 1,
2595 		     rr, 1,
2596 		     lw, 1,
2597 		     lr, 1,
2598 		     pa, 1,
2599 		     r_w, 1 );
2600 	MLX_FILL_1 ( &mpt, 2, mem_key, key );
2601 	MLX_FILL_2 ( &mpt, 3,
2602 		     pd, ARBEL_GLOBAL_PD,
2603 		     rae, 1 );
2604 	MLX_FILL_1 ( &mpt, 6, reg_wnd_len_h, 0xffffffffUL );
2605 	MLX_FILL_1 ( &mpt, 7, reg_wnd_len_l, 0xffffffffUL );
2606 	if ( ( rc = arbel_cmd_sw2hw_mpt ( arbel, arbel->limits.reserved_mrws,
2607 					  &mpt ) ) != 0 ) {
2608 		DBGC ( arbel, "Arbel %p could not set up MPT: %s\n",
2609 		       arbel, strerror ( rc ) );
2610 		return rc;
2611 	}
2612 
2613 	return 0;
2614 }
2615 
2616 /**
2617  * Configure special queue pairs
2618  *
2619  * @v arbel		Arbel device
2620  * @ret rc		Return status code
2621  */
arbel_configure_special_qps(struct arbel * arbel)2622 static int arbel_configure_special_qps ( struct arbel *arbel ) {
2623 	unsigned int smi_qpn_base;
2624 	unsigned int gsi_qpn_base;
2625 	int rc;
2626 
2627 	/* Special QP block must be aligned on an even number */
2628 	arbel->special_qpn_base = ( ( arbel->limits.reserved_qps + 1 ) & ~1 );
2629 	arbel->qpn_base = ( arbel->special_qpn_base +
2630 			    ARBEL_NUM_SPECIAL_QPS );
2631 	DBGC ( arbel, "Arbel %p special QPs at [%lx,%lx]\n", arbel,
2632 	       arbel->special_qpn_base, ( arbel->qpn_base - 1 ) );
2633 	smi_qpn_base = arbel->special_qpn_base;
2634 	gsi_qpn_base = ( smi_qpn_base + 2 );
2635 
2636 	/* Issue commands to configure special QPs */
2637 	if ( ( rc = arbel_cmd_conf_special_qp ( arbel, 0,
2638 						smi_qpn_base ) ) != 0 ) {
2639 		DBGC ( arbel, "Arbel %p could not configure SMI QPs: %s\n",
2640 		       arbel, strerror ( rc ) );
2641 		return rc;
2642 	}
2643 	if ( ( rc = arbel_cmd_conf_special_qp ( arbel, 1,
2644 						gsi_qpn_base ) ) != 0 ) {
2645 		DBGC ( arbel, "Arbel %p could not configure GSI QPs: %s\n",
2646 		       arbel, strerror ( rc ) );
2647 		return rc;
2648 	}
2649 
2650 	return 0;
2651 }
2652 
2653 /**
2654  * Start Arbel device
2655  *
2656  * @v arbel		Arbel device
2657  * @v running		Firmware is already running
2658  * @ret rc		Return status code
2659  */
arbel_start(struct arbel * arbel,int running)2660 static int arbel_start ( struct arbel *arbel, int running ) {
2661 	struct arbelprm_init_hca init_hca;
2662 	unsigned int i;
2663 	int rc;
2664 
2665 	/* Start firmware if not already running */
2666 	if ( ! running ) {
2667 		if ( ( rc = arbel_start_firmware ( arbel ) ) != 0 )
2668 			goto err_start_firmware;
2669 	}
2670 
2671 	/* Allocate ICM */
2672 	memset ( &init_hca, 0, sizeof ( init_hca ) );
2673 	if ( ( rc = arbel_alloc_icm ( arbel, &init_hca ) ) != 0 )
2674 		goto err_alloc_icm;
2675 
2676 	/* Initialise HCA */
2677 	if ( ( rc = arbel_cmd_init_hca ( arbel, &init_hca ) ) != 0 ) {
2678 		DBGC ( arbel, "Arbel %p could not initialise HCA: %s\n",
2679 		       arbel, strerror ( rc ) );
2680 		goto err_init_hca;
2681 	}
2682 
2683 	/* Set up memory protection */
2684 	if ( ( rc = arbel_setup_mpt ( arbel ) ) != 0 )
2685 		goto err_setup_mpt;
2686 	for ( i = 0 ; i < ARBEL_NUM_PORTS ; i++ )
2687 		arbel->ibdev[i]->rdma_key = arbel->lkey;
2688 
2689 	/* Set up event queue */
2690 	if ( ( rc = arbel_create_eq ( arbel ) ) != 0 )
2691 		goto err_create_eq;
2692 
2693 	/* Configure special QPs */
2694 	if ( ( rc = arbel_configure_special_qps ( arbel ) ) != 0 )
2695 		goto err_conf_special_qps;
2696 
2697 	return 0;
2698 
2699  err_conf_special_qps:
2700 	arbel_destroy_eq ( arbel );
2701  err_create_eq:
2702  err_setup_mpt:
2703 	arbel_cmd_close_hca ( arbel );
2704  err_init_hca:
2705 	arbel_free_icm ( arbel );
2706  err_alloc_icm:
2707 	arbel_stop_firmware ( arbel );
2708  err_start_firmware:
2709 	return rc;
2710 }
2711 
2712 /**
2713  * Stop Arbel device
2714  *
2715  * @v arbel		Arbel device
2716  */
arbel_stop(struct arbel * arbel)2717 static void arbel_stop ( struct arbel *arbel ) {
2718 	arbel_destroy_eq ( arbel );
2719 	arbel_cmd_close_hca ( arbel );
2720 	arbel_free_icm ( arbel );
2721 	arbel_stop_firmware ( arbel );
2722 	arbel_reset ( arbel );
2723 }
2724 
2725 /**
2726  * Open Arbel device
2727  *
2728  * @v arbel		Arbel device
2729  * @ret rc		Return status code
2730  */
arbel_open(struct arbel * arbel)2731 static int arbel_open ( struct arbel *arbel ) {
2732 	int rc;
2733 
2734 	/* Start device if applicable */
2735 	if ( arbel->open_count == 0 ) {
2736 		if ( ( rc = arbel_start ( arbel, 0 ) ) != 0 )
2737 			return rc;
2738 	}
2739 
2740 	/* Increment open counter */
2741 	arbel->open_count++;
2742 
2743 	return 0;
2744 }
2745 
2746 /**
2747  * Close Arbel device
2748  *
2749  * @v arbel		Arbel device
2750  */
arbel_close(struct arbel * arbel)2751 static void arbel_close ( struct arbel *arbel ) {
2752 
2753 	/* Decrement open counter */
2754 	assert ( arbel->open_count != 0 );
2755 	arbel->open_count--;
2756 
2757 	/* Stop device if applicable */
2758 	if ( arbel->open_count == 0 )
2759 		arbel_stop ( arbel );
2760 }
2761 
2762 /***************************************************************************
2763  *
2764  * Infiniband link-layer operations
2765  *
2766  ***************************************************************************
2767  */
2768 
2769 /**
2770  * Initialise Infiniband link
2771  *
2772  * @v ibdev		Infiniband device
2773  * @ret rc		Return status code
2774  */
arbel_ib_open(struct ib_device * ibdev)2775 static int arbel_ib_open ( struct ib_device *ibdev ) {
2776 	struct arbel *arbel = ib_get_drvdata ( ibdev );
2777 	struct arbelprm_init_ib init_ib;
2778 	int rc;
2779 
2780 	/* Open hardware */
2781 	if ( ( rc = arbel_open ( arbel ) ) != 0 )
2782 		goto err_open;
2783 
2784 	/* Initialise IB */
2785 	memset ( &init_ib, 0, sizeof ( init_ib ) );
2786 	MLX_FILL_3 ( &init_ib, 0,
2787 		     mtu_cap, ARBEL_MTU_2048,
2788 		     port_width_cap, 3,
2789 		     vl_cap, 1 );
2790 	MLX_FILL_1 ( &init_ib, 1, max_gid, 1 );
2791 	MLX_FILL_1 ( &init_ib, 2, max_pkey, 64 );
2792 	if ( ( rc = arbel_cmd_init_ib ( arbel, ibdev->port,
2793 					&init_ib ) ) != 0 ) {
2794 		DBGC ( arbel, "Arbel %p port %d could not intialise IB: %s\n",
2795 		       arbel, ibdev->port, strerror ( rc ) );
2796 		goto err_init_ib;
2797 	}
2798 
2799 	/* Update MAD parameters */
2800 	ib_smc_update ( ibdev, arbel_mad );
2801 
2802 	return 0;
2803 
2804  err_init_ib:
2805 	arbel_close ( arbel );
2806  err_open:
2807 	return rc;
2808 }
2809 
2810 /**
2811  * Close Infiniband link
2812  *
2813  * @v ibdev		Infiniband device
2814  */
arbel_ib_close(struct ib_device * ibdev)2815 static void arbel_ib_close ( struct ib_device *ibdev ) {
2816 	struct arbel *arbel = ib_get_drvdata ( ibdev );
2817 	int rc;
2818 
2819 	/* Close IB */
2820 	if ( ( rc = arbel_cmd_close_ib ( arbel, ibdev->port ) ) != 0 ) {
2821 		DBGC ( arbel, "Arbel %p port %d could not close IB: %s\n",
2822 		       arbel, ibdev->port, strerror ( rc ) );
2823 		/* Nothing we can do about this */
2824 	}
2825 
2826 	/* Close hardware */
2827 	arbel_close ( arbel );
2828 }
2829 
2830 /**
2831  * Inform embedded subnet management agent of a received MAD
2832  *
2833  * @v ibdev		Infiniband device
2834  * @v mad		MAD
2835  * @ret rc		Return status code
2836  */
arbel_inform_sma(struct ib_device * ibdev,union ib_mad * mad)2837 static int arbel_inform_sma ( struct ib_device *ibdev, union ib_mad *mad ) {
2838 	int rc;
2839 
2840 	/* Send the MAD to the embedded SMA */
2841 	if ( ( rc = arbel_mad ( ibdev, mad ) ) != 0 )
2842 		return rc;
2843 
2844 	/* Update parameters held in software */
2845 	ib_smc_update ( ibdev, arbel_mad );
2846 
2847 	return 0;
2848 }
2849 
2850 /***************************************************************************
2851  *
2852  * Multicast group operations
2853  *
2854  ***************************************************************************
2855  */
2856 
2857 /**
2858  * Attach to multicast group
2859  *
2860  * @v ibdev		Infiniband device
2861  * @v qp		Queue pair
2862  * @v gid		Multicast GID
2863  * @ret rc		Return status code
2864  */
arbel_mcast_attach(struct ib_device * ibdev,struct ib_queue_pair * qp,union ib_gid * gid)2865 static int arbel_mcast_attach ( struct ib_device *ibdev,
2866 				struct ib_queue_pair *qp,
2867 				union ib_gid *gid ) {
2868 	struct arbel *arbel = ib_get_drvdata ( ibdev );
2869 	struct arbelprm_mgm_hash hash;
2870 	struct arbelprm_mgm_entry mgm;
2871 	unsigned int index;
2872 	int rc;
2873 
2874 	/* Generate hash table index */
2875 	if ( ( rc = arbel_cmd_mgid_hash ( arbel, gid, &hash ) ) != 0 ) {
2876 		DBGC ( arbel, "Arbel %p could not hash GID: %s\n",
2877 		       arbel, strerror ( rc ) );
2878 		return rc;
2879 	}
2880 	index = MLX_GET ( &hash, hash );
2881 
2882 	/* Check for existing hash table entry */
2883 	if ( ( rc = arbel_cmd_read_mgm ( arbel, index, &mgm ) ) != 0 ) {
2884 		DBGC ( arbel, "Arbel %p could not read MGM %#x: %s\n",
2885 		       arbel, index, strerror ( rc ) );
2886 		return rc;
2887 	}
2888 	if ( MLX_GET ( &mgm, mgmqp_0.qi ) != 0 ) {
2889 		/* FIXME: this implementation allows only a single QP
2890 		 * per multicast group, and doesn't handle hash
2891 		 * collisions.  Sufficient for IPoIB but may need to
2892 		 * be extended in future.
2893 		 */
2894 		DBGC ( arbel, "Arbel %p MGID index %#x already in use\n",
2895 		       arbel, index );
2896 		return -EBUSY;
2897 	}
2898 
2899 	/* Update hash table entry */
2900 	MLX_FILL_2 ( &mgm, 8,
2901 		     mgmqp_0.qpn_i, qp->qpn,
2902 		     mgmqp_0.qi, 1 );
2903 	memcpy ( &mgm.u.dwords[4], gid, sizeof ( *gid ) );
2904 	if ( ( rc = arbel_cmd_write_mgm ( arbel, index, &mgm ) ) != 0 ) {
2905 		DBGC ( arbel, "Arbel %p could not write MGM %#x: %s\n",
2906 		       arbel, index, strerror ( rc ) );
2907 		return rc;
2908 	}
2909 
2910 	return 0;
2911 }
2912 
2913 /**
2914  * Detach from multicast group
2915  *
2916  * @v ibdev		Infiniband device
2917  * @v qp		Queue pair
2918  * @v gid		Multicast GID
2919  */
arbel_mcast_detach(struct ib_device * ibdev,struct ib_queue_pair * qp __unused,union ib_gid * gid)2920 static void arbel_mcast_detach ( struct ib_device *ibdev,
2921 				 struct ib_queue_pair *qp __unused,
2922 				 union ib_gid *gid ) {
2923 	struct arbel *arbel = ib_get_drvdata ( ibdev );
2924 	struct arbelprm_mgm_hash hash;
2925 	struct arbelprm_mgm_entry mgm;
2926 	unsigned int index;
2927 	int rc;
2928 
2929 	/* Generate hash table index */
2930 	if ( ( rc = arbel_cmd_mgid_hash ( arbel, gid, &hash ) ) != 0 ) {
2931 		DBGC ( arbel, "Arbel %p could not hash GID: %s\n",
2932 		       arbel, strerror ( rc ) );
2933 		return;
2934 	}
2935 	index = MLX_GET ( &hash, hash );
2936 
2937 	/* Clear hash table entry */
2938 	memset ( &mgm, 0, sizeof ( mgm ) );
2939 	if ( ( rc = arbel_cmd_write_mgm ( arbel, index, &mgm ) ) != 0 ) {
2940 		DBGC ( arbel, "Arbel %p could not write MGM %#x: %s\n",
2941 		       arbel, index, strerror ( rc ) );
2942 		return;
2943 	}
2944 }
2945 
2946 /** Arbel Infiniband operations */
2947 static struct ib_device_operations arbel_ib_operations = {
2948 	.create_cq	= arbel_create_cq,
2949 	.destroy_cq	= arbel_destroy_cq,
2950 	.create_qp	= arbel_create_qp,
2951 	.modify_qp	= arbel_modify_qp,
2952 	.destroy_qp	= arbel_destroy_qp,
2953 	.post_send	= arbel_post_send,
2954 	.post_recv	= arbel_post_recv,
2955 	.poll_cq	= arbel_poll_cq,
2956 	.poll_eq	= arbel_poll_eq,
2957 	.open		= arbel_ib_open,
2958 	.close		= arbel_ib_close,
2959 	.mcast_attach	= arbel_mcast_attach,
2960 	.mcast_detach	= arbel_mcast_detach,
2961 	.set_port_info	= arbel_inform_sma,
2962 	.set_pkey_table	= arbel_inform_sma,
2963 };
2964 
2965 /***************************************************************************
2966  *
2967  * PCI interface
2968  *
2969  ***************************************************************************
2970  */
2971 
2972 /**
2973  * Allocate Arbel device
2974  *
2975  * @ret arbel		Arbel device
2976  */
arbel_alloc(void)2977 static struct arbel * arbel_alloc ( void ) {
2978 	struct arbel *arbel;
2979 
2980 	/* Allocate Arbel device */
2981 	arbel = zalloc ( sizeof ( *arbel ) );
2982 	if ( ! arbel )
2983 		goto err_arbel;
2984 
2985 	/* Allocate space for mailboxes */
2986 	arbel->mailbox_in = malloc_dma ( ARBEL_MBOX_SIZE, ARBEL_MBOX_ALIGN );
2987 	if ( ! arbel->mailbox_in )
2988 		goto err_mailbox_in;
2989 	arbel->mailbox_out = malloc_dma ( ARBEL_MBOX_SIZE, ARBEL_MBOX_ALIGN );
2990 	if ( ! arbel->mailbox_out )
2991 		goto err_mailbox_out;
2992 
2993 	return arbel;
2994 
2995 	free_dma ( arbel->mailbox_out, ARBEL_MBOX_SIZE );
2996  err_mailbox_out:
2997 	free_dma ( arbel->mailbox_in, ARBEL_MBOX_SIZE );
2998  err_mailbox_in:
2999 	free ( arbel );
3000  err_arbel:
3001 	return NULL;
3002 }
3003 
3004 /**
3005  * Free Arbel device
3006  *
3007  * @v arbel		Arbel device
3008  */
arbel_free(struct arbel * arbel)3009 static void arbel_free ( struct arbel *arbel ) {
3010 
3011 	ufree ( arbel->icm );
3012 	ufree ( arbel->firmware_area );
3013 	free_dma ( arbel->mailbox_out, ARBEL_MBOX_SIZE );
3014 	free_dma ( arbel->mailbox_in, ARBEL_MBOX_SIZE );
3015 	free ( arbel );
3016 }
3017 
3018 /**
3019  * Probe PCI device
3020  *
3021  * @v pci		PCI device
3022  * @v id		PCI ID
3023  * @ret rc		Return status code
3024  */
arbel_probe(struct pci_device * pci)3025 static int arbel_probe ( struct pci_device *pci ) {
3026 	struct arbel *arbel;
3027 	struct ib_device *ibdev;
3028 	int i;
3029 	int rc;
3030 
3031 	/* Allocate Arbel device */
3032 	arbel = arbel_alloc();
3033 	if ( ! arbel ) {
3034 		rc = -ENOMEM;
3035 		goto err_alloc;
3036 	}
3037 	pci_set_drvdata ( pci, arbel );
3038 	arbel->pci = pci;
3039 
3040 	/* Fix up PCI device */
3041 	adjust_pci_device ( pci );
3042 
3043 	/* Map PCI BARs */
3044 	arbel->config = ioremap ( pci_bar_start ( pci, ARBEL_PCI_CONFIG_BAR ),
3045 				  ARBEL_PCI_CONFIG_BAR_SIZE );
3046 	arbel->uar = ioremap ( ( pci_bar_start ( pci, ARBEL_PCI_UAR_BAR ) +
3047 				 ARBEL_PCI_UAR_IDX * ARBEL_PCI_UAR_SIZE ),
3048 			       ARBEL_PCI_UAR_SIZE );
3049 
3050 	/* Allocate Infiniband devices */
3051 	for ( i = 0 ; i < ARBEL_NUM_PORTS ; i++ ) {
3052 		ibdev = alloc_ibdev ( 0 );
3053 		if ( ! ibdev ) {
3054 			rc = -ENOMEM;
3055 			goto err_alloc_ibdev;
3056 		}
3057 		arbel->ibdev[i] = ibdev;
3058 		ibdev->op = &arbel_ib_operations;
3059 		ibdev->dev = &pci->dev;
3060 		ibdev->port = ( ARBEL_PORT_BASE + i );
3061 		ib_set_drvdata ( ibdev, arbel );
3062 	}
3063 
3064 	/* Reset device */
3065 	arbel_reset ( arbel );
3066 
3067 	/* Start firmware */
3068 	if ( ( rc = arbel_start_firmware ( arbel ) ) != 0 )
3069 		goto err_start_firmware;
3070 
3071 	/* Get device limits */
3072 	if ( ( rc = arbel_get_limits ( arbel ) ) != 0 )
3073 		goto err_get_limits;
3074 
3075 	/* Start device */
3076 	if ( ( rc = arbel_start ( arbel, 1 ) ) != 0 )
3077 		goto err_start;
3078 
3079 	/* Initialise parameters using SMC */
3080 	for ( i = 0 ; i < ARBEL_NUM_PORTS ; i++ )
3081 		ib_smc_init ( arbel->ibdev[i], arbel_mad );
3082 
3083 	/* Register Infiniband devices */
3084 	for ( i = 0 ; i < ARBEL_NUM_PORTS ; i++ ) {
3085 		if ( ( rc = register_ibdev ( arbel->ibdev[i] ) ) != 0 ) {
3086 			DBGC ( arbel, "Arbel %p port %d could not register IB "
3087 			       "device: %s\n", arbel,
3088 			       arbel->ibdev[i]->port, strerror ( rc ) );
3089 			goto err_register_ibdev;
3090 		}
3091 	}
3092 
3093 	/* Leave device quiescent until opened */
3094 	if ( arbel->open_count == 0 )
3095 		arbel_stop ( arbel );
3096 
3097 	return 0;
3098 
3099 	i = ARBEL_NUM_PORTS;
3100  err_register_ibdev:
3101 	for ( i-- ; i >= 0 ; i-- )
3102 		unregister_ibdev ( arbel->ibdev[i] );
3103 	arbel_stop ( arbel );
3104  err_start:
3105  err_get_limits:
3106 	arbel_stop_firmware ( arbel );
3107  err_start_firmware:
3108 	i = ARBEL_NUM_PORTS;
3109  err_alloc_ibdev:
3110 	for ( i-- ; i >= 0 ; i-- )
3111 		ibdev_put ( arbel->ibdev[i] );
3112 	iounmap ( arbel->uar );
3113 	iounmap ( arbel->config );
3114 	arbel_free ( arbel );
3115  err_alloc:
3116 	return rc;
3117 }
3118 
3119 /**
3120  * Remove PCI device
3121  *
3122  * @v pci		PCI device
3123  */
arbel_remove(struct pci_device * pci)3124 static void arbel_remove ( struct pci_device *pci ) {
3125 	struct arbel *arbel = pci_get_drvdata ( pci );
3126 	int i;
3127 
3128 	for ( i = ( ARBEL_NUM_PORTS - 1 ) ; i >= 0 ; i-- )
3129 		unregister_ibdev ( arbel->ibdev[i] );
3130 	for ( i = ( ARBEL_NUM_PORTS - 1 ) ; i >= 0 ; i-- )
3131 		ibdev_put ( arbel->ibdev[i] );
3132 	iounmap ( arbel->uar );
3133 	iounmap ( arbel->config );
3134 	arbel_free ( arbel );
3135 }
3136 
3137 static struct pci_device_id arbel_nics[] = {
3138 	PCI_ROM ( 0x15b3, 0x6282, "mt25218", "MT25218 HCA driver", 0 ),
3139 	PCI_ROM ( 0x15b3, 0x6274, "mt25204", "MT25204 HCA driver", 0 ),
3140 };
3141 
3142 struct pci_driver arbel_driver __pci_driver = {
3143 	.ids = arbel_nics,
3144 	.id_count = ( sizeof ( arbel_nics ) / sizeof ( arbel_nics[0] ) ),
3145 	.probe = arbel_probe,
3146 	.remove = arbel_remove,
3147 };
3148