1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005, 2006, 2007 Cisco Systems, Inc. All rights reserved.
4  * Copyright (c) 2005, 2006, 2007, 2008 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  */
35 
36 #include <linux/gfp.h>
37 #include <linux/export.h>
38 
39 #include <linux/mlx4/cmd.h>
40 #include <linux/mlx4/qp.h>
41 
42 #include "mlx4.h"
43 #include "icm.h"
44 
45 /* QP to support BF should have bits 6,7 cleared */
46 #define MLX4_BF_QP_SKIP_MASK	0xc0
47 #define MLX4_MAX_BF_QP_RANGE	0x40
48 
49 void mlx4_qp_event(struct mlx4_dev *dev, u32 qpn, int event_type)
50 {
51 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
52 	struct mlx4_qp *qp;
53 
54 	spin_lock(&qp_table->lock);
55 
56 	qp = __mlx4_qp_lookup(dev, qpn);
57 	if (qp)
58 		atomic_inc(&qp->refcount);
59 
60 	spin_unlock(&qp_table->lock);
61 
62 	if (!qp) {
63 		mlx4_dbg(dev, "Async event for none existent QP %08x\n", qpn);
64 		return;
65 	}
66 
67 	qp->event(qp, event_type);
68 
69 	if (atomic_dec_and_test(&qp->refcount))
70 		complete(&qp->free);
71 }
72 
73 /* used for INIT/CLOSE port logic */
74 static int is_master_qp0(struct mlx4_dev *dev, struct mlx4_qp *qp, int *real_qp0, int *proxy_qp0)
75 {
76 	/* this procedure is called after we already know we are on the master */
77 	/* qp0 is either the proxy qp0, or the real qp0 */
78 	u32 pf_proxy_offset = dev->phys_caps.base_proxy_sqpn + 8 * mlx4_master_func_num(dev);
79 	*proxy_qp0 = qp->qpn >= pf_proxy_offset && qp->qpn <= pf_proxy_offset + 1;
80 
81 	*real_qp0 = qp->qpn >= dev->phys_caps.base_sqpn &&
82 		qp->qpn <= dev->phys_caps.base_sqpn + 1;
83 
84 	return *real_qp0 || *proxy_qp0;
85 }
86 
87 static int __mlx4_qp_modify(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
88 		     enum mlx4_qp_state cur_state, enum mlx4_qp_state new_state,
89 		     struct mlx4_qp_context *context,
90 		     enum mlx4_qp_optpar optpar,
91 		     int sqd_event, struct mlx4_qp *qp, int native)
92 {
93 	static const u16 op[MLX4_QP_NUM_STATE][MLX4_QP_NUM_STATE] = {
94 		[MLX4_QP_STATE_RST] = {
95 			[MLX4_QP_STATE_RST]	= MLX4_CMD_2RST_QP,
96 			[MLX4_QP_STATE_ERR]	= MLX4_CMD_2ERR_QP,
97 			[MLX4_QP_STATE_INIT]	= MLX4_CMD_RST2INIT_QP,
98 		},
99 		[MLX4_QP_STATE_INIT]  = {
100 			[MLX4_QP_STATE_RST]	= MLX4_CMD_2RST_QP,
101 			[MLX4_QP_STATE_ERR]	= MLX4_CMD_2ERR_QP,
102 			[MLX4_QP_STATE_INIT]	= MLX4_CMD_INIT2INIT_QP,
103 			[MLX4_QP_STATE_RTR]	= MLX4_CMD_INIT2RTR_QP,
104 		},
105 		[MLX4_QP_STATE_RTR]   = {
106 			[MLX4_QP_STATE_RST]	= MLX4_CMD_2RST_QP,
107 			[MLX4_QP_STATE_ERR]	= MLX4_CMD_2ERR_QP,
108 			[MLX4_QP_STATE_RTS]	= MLX4_CMD_RTR2RTS_QP,
109 		},
110 		[MLX4_QP_STATE_RTS]   = {
111 			[MLX4_QP_STATE_RST]	= MLX4_CMD_2RST_QP,
112 			[MLX4_QP_STATE_ERR]	= MLX4_CMD_2ERR_QP,
113 			[MLX4_QP_STATE_RTS]	= MLX4_CMD_RTS2RTS_QP,
114 			[MLX4_QP_STATE_SQD]	= MLX4_CMD_RTS2SQD_QP,
115 		},
116 		[MLX4_QP_STATE_SQD] = {
117 			[MLX4_QP_STATE_RST]	= MLX4_CMD_2RST_QP,
118 			[MLX4_QP_STATE_ERR]	= MLX4_CMD_2ERR_QP,
119 			[MLX4_QP_STATE_RTS]	= MLX4_CMD_SQD2RTS_QP,
120 			[MLX4_QP_STATE_SQD]	= MLX4_CMD_SQD2SQD_QP,
121 		},
122 		[MLX4_QP_STATE_SQER] = {
123 			[MLX4_QP_STATE_RST]	= MLX4_CMD_2RST_QP,
124 			[MLX4_QP_STATE_ERR]	= MLX4_CMD_2ERR_QP,
125 			[MLX4_QP_STATE_RTS]	= MLX4_CMD_SQERR2RTS_QP,
126 		},
127 		[MLX4_QP_STATE_ERR] = {
128 			[MLX4_QP_STATE_RST]	= MLX4_CMD_2RST_QP,
129 			[MLX4_QP_STATE_ERR]	= MLX4_CMD_2ERR_QP,
130 		}
131 	};
132 
133 	struct mlx4_priv *priv = mlx4_priv(dev);
134 	struct mlx4_cmd_mailbox *mailbox;
135 	int ret = 0;
136 	int real_qp0 = 0;
137 	int proxy_qp0 = 0;
138 	u8 port;
139 
140 	if (cur_state >= MLX4_QP_NUM_STATE || new_state >= MLX4_QP_NUM_STATE ||
141 	    !op[cur_state][new_state])
142 		return -EINVAL;
143 
144 	if (op[cur_state][new_state] == MLX4_CMD_2RST_QP) {
145 		ret = mlx4_cmd(dev, 0, qp->qpn, 2,
146 			MLX4_CMD_2RST_QP, MLX4_CMD_TIME_CLASS_A, native);
147 		if (mlx4_is_master(dev) && cur_state != MLX4_QP_STATE_ERR &&
148 		    cur_state != MLX4_QP_STATE_RST &&
149 		    is_master_qp0(dev, qp, &real_qp0, &proxy_qp0)) {
150 			port = (qp->qpn & 1) + 1;
151 			if (proxy_qp0)
152 				priv->mfunc.master.qp0_state[port].proxy_qp0_active = 0;
153 			else
154 				priv->mfunc.master.qp0_state[port].qp0_active = 0;
155 		}
156 		return ret;
157 	}
158 
159 	mailbox = mlx4_alloc_cmd_mailbox(dev);
160 	if (IS_ERR(mailbox))
161 		return PTR_ERR(mailbox);
162 
163 	if (cur_state == MLX4_QP_STATE_RST && new_state == MLX4_QP_STATE_INIT) {
164 		u64 mtt_addr = mlx4_mtt_addr(dev, mtt);
165 		context->mtt_base_addr_h = mtt_addr >> 32;
166 		context->mtt_base_addr_l = cpu_to_be32(mtt_addr & 0xffffffff);
167 		context->log_page_size   = mtt->page_shift - MLX4_ICM_PAGE_SHIFT;
168 	}
169 
170 	*(__be32 *) mailbox->buf = cpu_to_be32(optpar);
171 	memcpy(mailbox->buf + 8, context, sizeof *context);
172 
173 	((struct mlx4_qp_context *) (mailbox->buf + 8))->local_qpn =
174 		cpu_to_be32(qp->qpn);
175 
176 	ret = mlx4_cmd(dev, mailbox->dma,
177 		       qp->qpn | (!!sqd_event << 31),
178 		       new_state == MLX4_QP_STATE_RST ? 2 : 0,
179 		       op[cur_state][new_state], MLX4_CMD_TIME_CLASS_C, native);
180 
181 	if (mlx4_is_master(dev) && is_master_qp0(dev, qp, &real_qp0, &proxy_qp0)) {
182 		port = (qp->qpn & 1) + 1;
183 		if (cur_state != MLX4_QP_STATE_ERR &&
184 		    cur_state != MLX4_QP_STATE_RST &&
185 		    new_state == MLX4_QP_STATE_ERR) {
186 			if (proxy_qp0)
187 				priv->mfunc.master.qp0_state[port].proxy_qp0_active = 0;
188 			else
189 				priv->mfunc.master.qp0_state[port].qp0_active = 0;
190 		} else if (new_state == MLX4_QP_STATE_RTR) {
191 			if (proxy_qp0)
192 				priv->mfunc.master.qp0_state[port].proxy_qp0_active = 1;
193 			else
194 				priv->mfunc.master.qp0_state[port].qp0_active = 1;
195 		}
196 	}
197 
198 	mlx4_free_cmd_mailbox(dev, mailbox);
199 	return ret;
200 }
201 
202 int mlx4_qp_modify(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
203 		   enum mlx4_qp_state cur_state, enum mlx4_qp_state new_state,
204 		   struct mlx4_qp_context *context,
205 		   enum mlx4_qp_optpar optpar,
206 		   int sqd_event, struct mlx4_qp *qp)
207 {
208 	return __mlx4_qp_modify(dev, mtt, cur_state, new_state, context,
209 				optpar, sqd_event, qp, 0);
210 }
211 EXPORT_SYMBOL_GPL(mlx4_qp_modify);
212 
213 int __mlx4_qp_reserve_range(struct mlx4_dev *dev, int cnt, int align,
214 			    int *base, u8 flags)
215 {
216 	u32 uid;
217 	int bf_qp = !!(flags & (u8)MLX4_RESERVE_ETH_BF_QP);
218 
219 	struct mlx4_priv *priv = mlx4_priv(dev);
220 	struct mlx4_qp_table *qp_table = &priv->qp_table;
221 
222 	if (cnt > MLX4_MAX_BF_QP_RANGE && bf_qp)
223 		return -ENOMEM;
224 
225 	uid = MLX4_QP_TABLE_ZONE_GENERAL;
226 	if (flags & (u8)MLX4_RESERVE_A0_QP) {
227 		if (bf_qp)
228 			uid = MLX4_QP_TABLE_ZONE_RAW_ETH;
229 		else
230 			uid = MLX4_QP_TABLE_ZONE_RSS;
231 	}
232 
233 	*base = mlx4_zone_alloc_entries(qp_table->zones, uid, cnt, align,
234 					bf_qp ? MLX4_BF_QP_SKIP_MASK : 0, NULL);
235 	if (*base == -1)
236 		return -ENOMEM;
237 
238 	return 0;
239 }
240 
241 int mlx4_qp_reserve_range(struct mlx4_dev *dev, int cnt, int align,
242 			  int *base, u8 flags)
243 {
244 	u64 in_param = 0;
245 	u64 out_param;
246 	int err;
247 
248 	/* Turn off all unsupported QP allocation flags */
249 	flags &= dev->caps.alloc_res_qp_mask;
250 
251 	if (mlx4_is_mfunc(dev)) {
252 		set_param_l(&in_param, (((u32)flags) << 24) | (u32)cnt);
253 		set_param_h(&in_param, align);
254 		err = mlx4_cmd_imm(dev, in_param, &out_param,
255 				   RES_QP, RES_OP_RESERVE,
256 				   MLX4_CMD_ALLOC_RES,
257 				   MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
258 		if (err)
259 			return err;
260 
261 		*base = get_param_l(&out_param);
262 		return 0;
263 	}
264 	return __mlx4_qp_reserve_range(dev, cnt, align, base, flags);
265 }
266 EXPORT_SYMBOL_GPL(mlx4_qp_reserve_range);
267 
268 void __mlx4_qp_release_range(struct mlx4_dev *dev, int base_qpn, int cnt)
269 {
270 	struct mlx4_priv *priv = mlx4_priv(dev);
271 	struct mlx4_qp_table *qp_table = &priv->qp_table;
272 
273 	if (mlx4_is_qp_reserved(dev, (u32) base_qpn))
274 		return;
275 	mlx4_zone_free_entries_unique(qp_table->zones, base_qpn, cnt);
276 }
277 
278 void mlx4_qp_release_range(struct mlx4_dev *dev, int base_qpn, int cnt)
279 {
280 	u64 in_param = 0;
281 	int err;
282 
283 	if (mlx4_is_mfunc(dev)) {
284 		set_param_l(&in_param, base_qpn);
285 		set_param_h(&in_param, cnt);
286 		err = mlx4_cmd(dev, in_param, RES_QP, RES_OP_RESERVE,
287 			       MLX4_CMD_FREE_RES,
288 			       MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
289 		if (err) {
290 			mlx4_warn(dev, "Failed to release qp range base:%d cnt:%d\n",
291 				  base_qpn, cnt);
292 		}
293 	} else
294 		 __mlx4_qp_release_range(dev, base_qpn, cnt);
295 }
296 EXPORT_SYMBOL_GPL(mlx4_qp_release_range);
297 
298 int __mlx4_qp_alloc_icm(struct mlx4_dev *dev, int qpn, gfp_t gfp)
299 {
300 	struct mlx4_priv *priv = mlx4_priv(dev);
301 	struct mlx4_qp_table *qp_table = &priv->qp_table;
302 	int err;
303 
304 	err = mlx4_table_get(dev, &qp_table->qp_table, qpn, gfp);
305 	if (err)
306 		goto err_out;
307 
308 	err = mlx4_table_get(dev, &qp_table->auxc_table, qpn, gfp);
309 	if (err)
310 		goto err_put_qp;
311 
312 	err = mlx4_table_get(dev, &qp_table->altc_table, qpn, gfp);
313 	if (err)
314 		goto err_put_auxc;
315 
316 	err = mlx4_table_get(dev, &qp_table->rdmarc_table, qpn, gfp);
317 	if (err)
318 		goto err_put_altc;
319 
320 	err = mlx4_table_get(dev, &qp_table->cmpt_table, qpn, gfp);
321 	if (err)
322 		goto err_put_rdmarc;
323 
324 	return 0;
325 
326 err_put_rdmarc:
327 	mlx4_table_put(dev, &qp_table->rdmarc_table, qpn);
328 
329 err_put_altc:
330 	mlx4_table_put(dev, &qp_table->altc_table, qpn);
331 
332 err_put_auxc:
333 	mlx4_table_put(dev, &qp_table->auxc_table, qpn);
334 
335 err_put_qp:
336 	mlx4_table_put(dev, &qp_table->qp_table, qpn);
337 
338 err_out:
339 	return err;
340 }
341 
342 static int mlx4_qp_alloc_icm(struct mlx4_dev *dev, int qpn, gfp_t gfp)
343 {
344 	u64 param = 0;
345 
346 	if (mlx4_is_mfunc(dev)) {
347 		set_param_l(&param, qpn);
348 		return mlx4_cmd_imm(dev, param, &param, RES_QP, RES_OP_MAP_ICM,
349 				    MLX4_CMD_ALLOC_RES, MLX4_CMD_TIME_CLASS_A,
350 				    MLX4_CMD_WRAPPED);
351 	}
352 	return __mlx4_qp_alloc_icm(dev, qpn, gfp);
353 }
354 
355 void __mlx4_qp_free_icm(struct mlx4_dev *dev, int qpn)
356 {
357 	struct mlx4_priv *priv = mlx4_priv(dev);
358 	struct mlx4_qp_table *qp_table = &priv->qp_table;
359 
360 	mlx4_table_put(dev, &qp_table->cmpt_table, qpn);
361 	mlx4_table_put(dev, &qp_table->rdmarc_table, qpn);
362 	mlx4_table_put(dev, &qp_table->altc_table, qpn);
363 	mlx4_table_put(dev, &qp_table->auxc_table, qpn);
364 	mlx4_table_put(dev, &qp_table->qp_table, qpn);
365 }
366 
367 static void mlx4_qp_free_icm(struct mlx4_dev *dev, int qpn)
368 {
369 	u64 in_param = 0;
370 
371 	if (mlx4_is_mfunc(dev)) {
372 		set_param_l(&in_param, qpn);
373 		if (mlx4_cmd(dev, in_param, RES_QP, RES_OP_MAP_ICM,
374 			     MLX4_CMD_FREE_RES, MLX4_CMD_TIME_CLASS_A,
375 			     MLX4_CMD_WRAPPED))
376 			mlx4_warn(dev, "Failed to free icm of qp:%d\n", qpn);
377 	} else
378 		__mlx4_qp_free_icm(dev, qpn);
379 }
380 
381 int mlx4_qp_alloc(struct mlx4_dev *dev, int qpn, struct mlx4_qp *qp, gfp_t gfp)
382 {
383 	struct mlx4_priv *priv = mlx4_priv(dev);
384 	struct mlx4_qp_table *qp_table = &priv->qp_table;
385 	int err;
386 
387 	if (!qpn)
388 		return -EINVAL;
389 
390 	qp->qpn = qpn;
391 
392 	err = mlx4_qp_alloc_icm(dev, qpn, gfp);
393 	if (err)
394 		return err;
395 
396 	spin_lock_irq(&qp_table->lock);
397 	err = radix_tree_insert(&dev->qp_table_tree, qp->qpn &
398 				(dev->caps.num_qps - 1), qp);
399 	spin_unlock_irq(&qp_table->lock);
400 	if (err)
401 		goto err_icm;
402 
403 	atomic_set(&qp->refcount, 1);
404 	init_completion(&qp->free);
405 
406 	return 0;
407 
408 err_icm:
409 	mlx4_qp_free_icm(dev, qpn);
410 	return err;
411 }
412 
413 EXPORT_SYMBOL_GPL(mlx4_qp_alloc);
414 
415 int mlx4_update_qp(struct mlx4_dev *dev, u32 qpn,
416 		   enum mlx4_update_qp_attr attr,
417 		   struct mlx4_update_qp_params *params)
418 {
419 	struct mlx4_cmd_mailbox *mailbox;
420 	struct mlx4_update_qp_context *cmd;
421 	u64 pri_addr_path_mask = 0;
422 	u64 qp_mask = 0;
423 	int err = 0;
424 
425 	mailbox = mlx4_alloc_cmd_mailbox(dev);
426 	if (IS_ERR(mailbox))
427 		return PTR_ERR(mailbox);
428 
429 	cmd = (struct mlx4_update_qp_context *)mailbox->buf;
430 
431 	if (!attr || (attr & ~MLX4_UPDATE_QP_SUPPORTED_ATTRS))
432 		return -EINVAL;
433 
434 	if (attr & MLX4_UPDATE_QP_SMAC) {
435 		pri_addr_path_mask |= 1ULL << MLX4_UPD_QP_PATH_MASK_MAC_INDEX;
436 		cmd->qp_context.pri_path.grh_mylmc = params->smac_index;
437 	}
438 
439 	if (attr & MLX4_UPDATE_QP_VSD) {
440 		qp_mask |= 1ULL << MLX4_UPD_QP_MASK_VSD;
441 		if (params->flags & MLX4_UPDATE_QP_PARAMS_FLAGS_VSD_ENABLE)
442 			cmd->qp_context.param3 |= cpu_to_be32(MLX4_STRIP_VLAN);
443 	}
444 
445 	if (attr & MLX4_UPDATE_QP_RATE_LIMIT) {
446 		qp_mask |= 1ULL << MLX4_UPD_QP_MASK_RATE_LIMIT;
447 		cmd->qp_context.rate_limit_params = cpu_to_be16((params->rate_unit << 14) | params->rate_val);
448 	}
449 
450 	if (attr & MLX4_UPDATE_QP_QOS_VPORT) {
451 		qp_mask |= 1ULL << MLX4_UPD_QP_MASK_QOS_VPP;
452 		cmd->qp_context.qos_vport = params->qos_vport;
453 	}
454 
455 	cmd->primary_addr_path_mask = cpu_to_be64(pri_addr_path_mask);
456 	cmd->qp_mask = cpu_to_be64(qp_mask);
457 
458 	err = mlx4_cmd(dev, mailbox->dma, qpn & 0xffffff, 0,
459 		       MLX4_CMD_UPDATE_QP, MLX4_CMD_TIME_CLASS_A,
460 		       MLX4_CMD_NATIVE);
461 
462 	mlx4_free_cmd_mailbox(dev, mailbox);
463 	return err;
464 }
465 EXPORT_SYMBOL_GPL(mlx4_update_qp);
466 
467 void mlx4_qp_remove(struct mlx4_dev *dev, struct mlx4_qp *qp)
468 {
469 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
470 	unsigned long flags;
471 
472 	spin_lock_irqsave(&qp_table->lock, flags);
473 	radix_tree_delete(&dev->qp_table_tree, qp->qpn & (dev->caps.num_qps - 1));
474 	spin_unlock_irqrestore(&qp_table->lock, flags);
475 }
476 EXPORT_SYMBOL_GPL(mlx4_qp_remove);
477 
478 void mlx4_qp_free(struct mlx4_dev *dev, struct mlx4_qp *qp)
479 {
480 	if (atomic_dec_and_test(&qp->refcount))
481 		complete(&qp->free);
482 	wait_for_completion(&qp->free);
483 
484 	mlx4_qp_free_icm(dev, qp->qpn);
485 }
486 EXPORT_SYMBOL_GPL(mlx4_qp_free);
487 
488 static int mlx4_CONF_SPECIAL_QP(struct mlx4_dev *dev, u32 base_qpn)
489 {
490 	return mlx4_cmd(dev, 0, base_qpn, 0, MLX4_CMD_CONF_SPECIAL_QP,
491 			MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
492 }
493 
494 #define MLX4_QP_TABLE_RSS_ETH_PRIORITY 2
495 #define MLX4_QP_TABLE_RAW_ETH_PRIORITY 1
496 #define MLX4_QP_TABLE_RAW_ETH_SIZE     256
497 
498 static int mlx4_create_zones(struct mlx4_dev *dev,
499 			     u32 reserved_bottom_general,
500 			     u32 reserved_top_general,
501 			     u32 reserved_bottom_rss,
502 			     u32 start_offset_rss,
503 			     u32 max_table_offset)
504 {
505 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
506 	struct mlx4_bitmap (*bitmap)[MLX4_QP_TABLE_ZONE_NUM] = NULL;
507 	int bitmap_initialized = 0;
508 	u32 last_offset;
509 	int k;
510 	int err;
511 
512 	qp_table->zones = mlx4_zone_allocator_create(MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP);
513 
514 	if (NULL == qp_table->zones)
515 		return -ENOMEM;
516 
517 	bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
518 
519 	if (NULL == bitmap) {
520 		err = -ENOMEM;
521 		goto free_zone;
522 	}
523 
524 	err = mlx4_bitmap_init(*bitmap + MLX4_QP_TABLE_ZONE_GENERAL, dev->caps.num_qps,
525 			       (1 << 23) - 1, reserved_bottom_general,
526 			       reserved_top_general);
527 
528 	if (err)
529 		goto free_bitmap;
530 
531 	++bitmap_initialized;
532 
533 	err = mlx4_zone_add_one(qp_table->zones, *bitmap + MLX4_QP_TABLE_ZONE_GENERAL,
534 				MLX4_ZONE_FALLBACK_TO_HIGHER_PRIO |
535 				MLX4_ZONE_USE_RR, 0,
536 				0, qp_table->zones_uids + MLX4_QP_TABLE_ZONE_GENERAL);
537 
538 	if (err)
539 		goto free_bitmap;
540 
541 	err = mlx4_bitmap_init(*bitmap + MLX4_QP_TABLE_ZONE_RSS,
542 			       reserved_bottom_rss,
543 			       reserved_bottom_rss - 1,
544 			       dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW],
545 			       reserved_bottom_rss - start_offset_rss);
546 
547 	if (err)
548 		goto free_bitmap;
549 
550 	++bitmap_initialized;
551 
552 	err = mlx4_zone_add_one(qp_table->zones, *bitmap + MLX4_QP_TABLE_ZONE_RSS,
553 				MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO |
554 				MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO |
555 				MLX4_ZONE_USE_RR, MLX4_QP_TABLE_RSS_ETH_PRIORITY,
556 				0, qp_table->zones_uids + MLX4_QP_TABLE_ZONE_RSS);
557 
558 	if (err)
559 		goto free_bitmap;
560 
561 	last_offset = dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW];
562 	/*  We have a single zone for the A0 steering QPs area of the FW. This area
563 	 *  needs to be split into subareas. One set of subareas is for RSS QPs
564 	 *  (in which qp number bits 6 and/or 7 are set); the other set of subareas
565 	 *  is for RAW_ETH QPs, which require that both bits 6 and 7 are zero.
566 	 *  Currently, the values returned by the FW (A0 steering area starting qp number
567 	 *  and A0 steering area size) are such that there are only two subareas -- one
568 	 *  for RSS and one for RAW_ETH.
569 	 */
570 	for (k = MLX4_QP_TABLE_ZONE_RSS + 1; k < sizeof(*bitmap)/sizeof((*bitmap)[0]);
571 	     k++) {
572 		int size;
573 		u32 offset = start_offset_rss;
574 		u32 bf_mask;
575 		u32 requested_size;
576 
577 		/* Assuming MLX4_BF_QP_SKIP_MASK is consecutive ones, this calculates
578 		 * a mask of all LSB bits set until (and not including) the first
579 		 * set bit of  MLX4_BF_QP_SKIP_MASK. For example, if MLX4_BF_QP_SKIP_MASK
580 		 * is 0xc0, bf_mask will be 0x3f.
581 		 */
582 		bf_mask = (MLX4_BF_QP_SKIP_MASK & ~(MLX4_BF_QP_SKIP_MASK - 1)) - 1;
583 		requested_size = min((u32)MLX4_QP_TABLE_RAW_ETH_SIZE, bf_mask + 1);
584 
585 		if (((last_offset & MLX4_BF_QP_SKIP_MASK) &&
586 		     ((int)(max_table_offset - last_offset)) >=
587 		     roundup_pow_of_two(MLX4_BF_QP_SKIP_MASK)) ||
588 		    (!(last_offset & MLX4_BF_QP_SKIP_MASK) &&
589 		     !((last_offset + requested_size - 1) &
590 		       MLX4_BF_QP_SKIP_MASK)))
591 			size = requested_size;
592 		else {
593 			u32 candidate_offset =
594 				(last_offset | MLX4_BF_QP_SKIP_MASK | bf_mask) + 1;
595 
596 			if (last_offset & MLX4_BF_QP_SKIP_MASK)
597 				last_offset = candidate_offset;
598 
599 			/* From this point, the BF bits are 0 */
600 
601 			if (last_offset > max_table_offset) {
602 				/* need to skip */
603 				size = -1;
604 			} else {
605 				size = min3(max_table_offset - last_offset,
606 					    bf_mask - (last_offset & bf_mask),
607 					    requested_size);
608 				if (size < requested_size) {
609 					int candidate_size;
610 
611 					candidate_size = min3(
612 						max_table_offset - candidate_offset,
613 						bf_mask - (last_offset & bf_mask),
614 						requested_size);
615 
616 					/*  We will not take this path if last_offset was
617 					 *  already set above to candidate_offset
618 					 */
619 					if (candidate_size > size) {
620 						last_offset = candidate_offset;
621 						size = candidate_size;
622 					}
623 				}
624 			}
625 		}
626 
627 		if (size > 0) {
628 			/* mlx4_bitmap_alloc_range will find a contiguous range of "size"
629 			 * QPs in which both bits 6 and 7 are zero, because we pass it the
630 			 * MLX4_BF_SKIP_MASK).
631 			 */
632 			offset = mlx4_bitmap_alloc_range(
633 					*bitmap + MLX4_QP_TABLE_ZONE_RSS,
634 					size, 1,
635 					MLX4_BF_QP_SKIP_MASK);
636 
637 			if (offset == (u32)-1) {
638 				err = -ENOMEM;
639 				break;
640 			}
641 
642 			last_offset = offset + size;
643 
644 			err = mlx4_bitmap_init(*bitmap + k, roundup_pow_of_two(size),
645 					       roundup_pow_of_two(size) - 1, 0,
646 					       roundup_pow_of_two(size) - size);
647 		} else {
648 			/* Add an empty bitmap, we'll allocate from different zones (since
649 			 * at least one is reserved)
650 			 */
651 			err = mlx4_bitmap_init(*bitmap + k, 1,
652 					       MLX4_QP_TABLE_RAW_ETH_SIZE - 1, 0,
653 					       0);
654 			mlx4_bitmap_alloc_range(*bitmap + k, 1, 1, 0);
655 		}
656 
657 		if (err)
658 			break;
659 
660 		++bitmap_initialized;
661 
662 		err = mlx4_zone_add_one(qp_table->zones, *bitmap + k,
663 					MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO |
664 					MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO |
665 					MLX4_ZONE_USE_RR, MLX4_QP_TABLE_RAW_ETH_PRIORITY,
666 					offset, qp_table->zones_uids + k);
667 
668 		if (err)
669 			break;
670 	}
671 
672 	if (err)
673 		goto free_bitmap;
674 
675 	qp_table->bitmap_gen = *bitmap;
676 
677 	return err;
678 
679 free_bitmap:
680 	for (k = 0; k < bitmap_initialized; k++)
681 		mlx4_bitmap_cleanup(*bitmap + k);
682 	kfree(bitmap);
683 free_zone:
684 	mlx4_zone_allocator_destroy(qp_table->zones);
685 	return err;
686 }
687 
688 static void mlx4_cleanup_qp_zones(struct mlx4_dev *dev)
689 {
690 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
691 
692 	if (qp_table->zones) {
693 		int i;
694 
695 		for (i = 0;
696 		     i < sizeof(qp_table->zones_uids)/sizeof(qp_table->zones_uids[0]);
697 		     i++) {
698 			struct mlx4_bitmap *bitmap =
699 				mlx4_zone_get_bitmap(qp_table->zones,
700 						     qp_table->zones_uids[i]);
701 
702 			mlx4_zone_remove_one(qp_table->zones, qp_table->zones_uids[i]);
703 			if (NULL == bitmap)
704 				continue;
705 
706 			mlx4_bitmap_cleanup(bitmap);
707 		}
708 		mlx4_zone_allocator_destroy(qp_table->zones);
709 		kfree(qp_table->bitmap_gen);
710 		qp_table->bitmap_gen = NULL;
711 		qp_table->zones = NULL;
712 	}
713 }
714 
715 int mlx4_init_qp_table(struct mlx4_dev *dev)
716 {
717 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
718 	int err;
719 	int reserved_from_top = 0;
720 	int reserved_from_bot;
721 	int k;
722 	int fixed_reserved_from_bot_rv = 0;
723 	int bottom_reserved_for_rss_bitmap;
724 	u32 max_table_offset = dev->caps.dmfs_high_rate_qpn_base +
725 			dev->caps.dmfs_high_rate_qpn_range;
726 
727 	spin_lock_init(&qp_table->lock);
728 	INIT_RADIX_TREE(&dev->qp_table_tree, GFP_ATOMIC);
729 	if (mlx4_is_slave(dev))
730 		return 0;
731 
732 	/* We reserve 2 extra QPs per port for the special QPs.  The
733 	 * block of special QPs must be aligned to a multiple of 8, so
734 	 * round up.
735 	 *
736 	 * We also reserve the MSB of the 24-bit QP number to indicate
737 	 * that a QP is an XRC QP.
738 	 */
739 	for (k = 0; k <= MLX4_QP_REGION_BOTTOM; k++)
740 		fixed_reserved_from_bot_rv += dev->caps.reserved_qps_cnt[k];
741 
742 	if (fixed_reserved_from_bot_rv < max_table_offset)
743 		fixed_reserved_from_bot_rv = max_table_offset;
744 
745 	/* We reserve at least 1 extra for bitmaps that we don't have enough space for*/
746 	bottom_reserved_for_rss_bitmap =
747 		roundup_pow_of_two(fixed_reserved_from_bot_rv + 1);
748 	dev->phys_caps.base_sqpn = ALIGN(bottom_reserved_for_rss_bitmap, 8);
749 
750 	{
751 		int sort[MLX4_NUM_QP_REGION];
752 		int i, j, tmp;
753 		int last_base = dev->caps.num_qps;
754 
755 		for (i = 1; i < MLX4_NUM_QP_REGION; ++i)
756 			sort[i] = i;
757 
758 		for (i = MLX4_NUM_QP_REGION; i > MLX4_QP_REGION_BOTTOM; --i) {
759 			for (j = MLX4_QP_REGION_BOTTOM + 2; j < i; ++j) {
760 				if (dev->caps.reserved_qps_cnt[sort[j]] >
761 				    dev->caps.reserved_qps_cnt[sort[j - 1]]) {
762 					tmp             = sort[j];
763 					sort[j]         = sort[j - 1];
764 					sort[j - 1]     = tmp;
765 				}
766 			}
767 		}
768 
769 		for (i = MLX4_QP_REGION_BOTTOM + 1; i < MLX4_NUM_QP_REGION; ++i) {
770 			last_base -= dev->caps.reserved_qps_cnt[sort[i]];
771 			dev->caps.reserved_qps_base[sort[i]] = last_base;
772 			reserved_from_top +=
773 				dev->caps.reserved_qps_cnt[sort[i]];
774 		}
775 	}
776 
777        /* Reserve 8 real SQPs in both native and SRIOV modes.
778 	* In addition, in SRIOV mode, reserve 8 proxy SQPs per function
779 	* (for all PFs and VFs), and 8 corresponding tunnel QPs.
780 	* Each proxy SQP works opposite its own tunnel QP.
781 	*
782 	* The QPs are arranged as follows:
783 	* a. 8 real SQPs
784 	* b. All the proxy SQPs (8 per function)
785 	* c. All the tunnel QPs (8 per function)
786 	*/
787 	reserved_from_bot = mlx4_num_reserved_sqps(dev);
788 	if (reserved_from_bot + reserved_from_top > dev->caps.num_qps) {
789 		mlx4_err(dev, "Number of reserved QPs is higher than number of QPs\n");
790 		return -EINVAL;
791 	}
792 
793 	err = mlx4_create_zones(dev, reserved_from_bot, reserved_from_bot,
794 				bottom_reserved_for_rss_bitmap,
795 				fixed_reserved_from_bot_rv,
796 				max_table_offset);
797 
798 	if (err)
799 		return err;
800 
801 	if (mlx4_is_mfunc(dev)) {
802 		/* for PPF use */
803 		dev->phys_caps.base_proxy_sqpn = dev->phys_caps.base_sqpn + 8;
804 		dev->phys_caps.base_tunnel_sqpn = dev->phys_caps.base_sqpn + 8 + 8 * MLX4_MFUNC_MAX;
805 
806 		/* In mfunc, calculate proxy and tunnel qp offsets for the PF here,
807 		 * since the PF does not call mlx4_slave_caps */
808 		dev->caps.qp0_tunnel = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
809 		dev->caps.qp0_proxy = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
810 		dev->caps.qp1_tunnel = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
811 		dev->caps.qp1_proxy = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
812 
813 		if (!dev->caps.qp0_tunnel || !dev->caps.qp0_proxy ||
814 		    !dev->caps.qp1_tunnel || !dev->caps.qp1_proxy) {
815 			err = -ENOMEM;
816 			goto err_mem;
817 		}
818 
819 		for (k = 0; k < dev->caps.num_ports; k++) {
820 			dev->caps.qp0_proxy[k] = dev->phys_caps.base_proxy_sqpn +
821 				8 * mlx4_master_func_num(dev) + k;
822 			dev->caps.qp0_tunnel[k] = dev->caps.qp0_proxy[k] + 8 * MLX4_MFUNC_MAX;
823 			dev->caps.qp1_proxy[k] = dev->phys_caps.base_proxy_sqpn +
824 				8 * mlx4_master_func_num(dev) + MLX4_MAX_PORTS + k;
825 			dev->caps.qp1_tunnel[k] = dev->caps.qp1_proxy[k] + 8 * MLX4_MFUNC_MAX;
826 		}
827 	}
828 
829 
830 	err = mlx4_CONF_SPECIAL_QP(dev, dev->phys_caps.base_sqpn);
831 	if (err)
832 		goto err_mem;
833 
834 	return err;
835 
836 err_mem:
837 	kfree(dev->caps.qp0_tunnel);
838 	kfree(dev->caps.qp0_proxy);
839 	kfree(dev->caps.qp1_tunnel);
840 	kfree(dev->caps.qp1_proxy);
841 	dev->caps.qp0_tunnel = dev->caps.qp0_proxy =
842 		dev->caps.qp1_tunnel = dev->caps.qp1_proxy = NULL;
843 	mlx4_cleanup_qp_zones(dev);
844 	return err;
845 }
846 
847 void mlx4_cleanup_qp_table(struct mlx4_dev *dev)
848 {
849 	if (mlx4_is_slave(dev))
850 		return;
851 
852 	mlx4_CONF_SPECIAL_QP(dev, 0);
853 
854 	mlx4_cleanup_qp_zones(dev);
855 }
856 
857 int mlx4_qp_query(struct mlx4_dev *dev, struct mlx4_qp *qp,
858 		  struct mlx4_qp_context *context)
859 {
860 	struct mlx4_cmd_mailbox *mailbox;
861 	int err;
862 
863 	mailbox = mlx4_alloc_cmd_mailbox(dev);
864 	if (IS_ERR(mailbox))
865 		return PTR_ERR(mailbox);
866 
867 	err = mlx4_cmd_box(dev, 0, mailbox->dma, qp->qpn, 0,
868 			   MLX4_CMD_QUERY_QP, MLX4_CMD_TIME_CLASS_A,
869 			   MLX4_CMD_WRAPPED);
870 	if (!err)
871 		memcpy(context, mailbox->buf + 8, sizeof *context);
872 
873 	mlx4_free_cmd_mailbox(dev, mailbox);
874 	return err;
875 }
876 EXPORT_SYMBOL_GPL(mlx4_qp_query);
877 
878 int mlx4_qp_to_ready(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
879 		     struct mlx4_qp_context *context,
880 		     struct mlx4_qp *qp, enum mlx4_qp_state *qp_state)
881 {
882 	int err;
883 	int i;
884 	enum mlx4_qp_state states[] = {
885 		MLX4_QP_STATE_RST,
886 		MLX4_QP_STATE_INIT,
887 		MLX4_QP_STATE_RTR,
888 		MLX4_QP_STATE_RTS
889 	};
890 
891 	for (i = 0; i < ARRAY_SIZE(states) - 1; i++) {
892 		context->flags &= cpu_to_be32(~(0xf << 28));
893 		context->flags |= cpu_to_be32(states[i + 1] << 28);
894 		if (states[i + 1] != MLX4_QP_STATE_RTR)
895 			context->params2 &= ~MLX4_QP_BIT_FPP;
896 		err = mlx4_qp_modify(dev, mtt, states[i], states[i + 1],
897 				     context, 0, 0, qp);
898 		if (err) {
899 			mlx4_err(dev, "Failed to bring QP to state: %d with error: %d\n",
900 				 states[i + 1], err);
901 			return err;
902 		}
903 
904 		*qp_state = states[i + 1];
905 	}
906 
907 	return 0;
908 }
909 EXPORT_SYMBOL_GPL(mlx4_qp_to_ready);
910