xref: /linux/drivers/net/ethernet/mellanox/mlx4/qp.c (revision a4eb44a6435d6d8f9e642407a4a06f65eb90ca04)
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 		refcount_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 (refcount_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 	if ((cur_state == MLX4_QP_STATE_RTR) &&
171 	    (new_state == MLX4_QP_STATE_RTS) &&
172 	    dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2)
173 		context->roce_entropy =
174 			cpu_to_be16(mlx4_qp_roce_entropy(dev, qp->qpn));
175 
176 	*(__be32 *) mailbox->buf = cpu_to_be32(optpar);
177 	memcpy(mailbox->buf + 8, context, sizeof(*context));
178 
179 	((struct mlx4_qp_context *) (mailbox->buf + 8))->local_qpn =
180 		cpu_to_be32(qp->qpn);
181 
182 	ret = mlx4_cmd(dev, mailbox->dma,
183 		       qp->qpn | (!!sqd_event << 31),
184 		       new_state == MLX4_QP_STATE_RST ? 2 : 0,
185 		       op[cur_state][new_state], MLX4_CMD_TIME_CLASS_C, native);
186 
187 	if (mlx4_is_master(dev) && is_master_qp0(dev, qp, &real_qp0, &proxy_qp0)) {
188 		port = (qp->qpn & 1) + 1;
189 		if (cur_state != MLX4_QP_STATE_ERR &&
190 		    cur_state != MLX4_QP_STATE_RST &&
191 		    new_state == MLX4_QP_STATE_ERR) {
192 			if (proxy_qp0)
193 				priv->mfunc.master.qp0_state[port].proxy_qp0_active = 0;
194 			else
195 				priv->mfunc.master.qp0_state[port].qp0_active = 0;
196 		} else if (new_state == MLX4_QP_STATE_RTR) {
197 			if (proxy_qp0)
198 				priv->mfunc.master.qp0_state[port].proxy_qp0_active = 1;
199 			else
200 				priv->mfunc.master.qp0_state[port].qp0_active = 1;
201 		}
202 	}
203 
204 	mlx4_free_cmd_mailbox(dev, mailbox);
205 	return ret;
206 }
207 
208 int mlx4_qp_modify(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
209 		   enum mlx4_qp_state cur_state, enum mlx4_qp_state new_state,
210 		   struct mlx4_qp_context *context,
211 		   enum mlx4_qp_optpar optpar,
212 		   int sqd_event, struct mlx4_qp *qp)
213 {
214 	return __mlx4_qp_modify(dev, mtt, cur_state, new_state, context,
215 				optpar, sqd_event, qp, 0);
216 }
217 EXPORT_SYMBOL_GPL(mlx4_qp_modify);
218 
219 int __mlx4_qp_reserve_range(struct mlx4_dev *dev, int cnt, int align,
220 			    int *base, u8 flags)
221 {
222 	u32 uid;
223 	int bf_qp = !!(flags & (u8)MLX4_RESERVE_ETH_BF_QP);
224 
225 	struct mlx4_priv *priv = mlx4_priv(dev);
226 	struct mlx4_qp_table *qp_table = &priv->qp_table;
227 
228 	if (cnt > MLX4_MAX_BF_QP_RANGE && bf_qp)
229 		return -ENOMEM;
230 
231 	uid = MLX4_QP_TABLE_ZONE_GENERAL;
232 	if (flags & (u8)MLX4_RESERVE_A0_QP) {
233 		if (bf_qp)
234 			uid = MLX4_QP_TABLE_ZONE_RAW_ETH;
235 		else
236 			uid = MLX4_QP_TABLE_ZONE_RSS;
237 	}
238 
239 	*base = mlx4_zone_alloc_entries(qp_table->zones, uid, cnt, align,
240 					bf_qp ? MLX4_BF_QP_SKIP_MASK : 0, NULL);
241 	if (*base == -1)
242 		return -ENOMEM;
243 
244 	return 0;
245 }
246 
247 int mlx4_qp_reserve_range(struct mlx4_dev *dev, int cnt, int align,
248 			  int *base, u8 flags, u8 usage)
249 {
250 	u32 in_modifier = RES_QP | (((u32)usage & 3) << 30);
251 	u64 in_param = 0;
252 	u64 out_param;
253 	int err;
254 
255 	/* Turn off all unsupported QP allocation flags */
256 	flags &= dev->caps.alloc_res_qp_mask;
257 
258 	if (mlx4_is_mfunc(dev)) {
259 		set_param_l(&in_param, (((u32)flags) << 24) | (u32)cnt);
260 		set_param_h(&in_param, align);
261 		err = mlx4_cmd_imm(dev, in_param, &out_param,
262 				   in_modifier, RES_OP_RESERVE,
263 				   MLX4_CMD_ALLOC_RES,
264 				   MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
265 		if (err)
266 			return err;
267 
268 		*base = get_param_l(&out_param);
269 		return 0;
270 	}
271 	return __mlx4_qp_reserve_range(dev, cnt, align, base, flags);
272 }
273 EXPORT_SYMBOL_GPL(mlx4_qp_reserve_range);
274 
275 void __mlx4_qp_release_range(struct mlx4_dev *dev, int base_qpn, int cnt)
276 {
277 	struct mlx4_priv *priv = mlx4_priv(dev);
278 	struct mlx4_qp_table *qp_table = &priv->qp_table;
279 
280 	if (mlx4_is_qp_reserved(dev, (u32) base_qpn))
281 		return;
282 	mlx4_zone_free_entries_unique(qp_table->zones, base_qpn, cnt);
283 }
284 
285 void mlx4_qp_release_range(struct mlx4_dev *dev, int base_qpn, int cnt)
286 {
287 	u64 in_param = 0;
288 	int err;
289 
290 	if (!cnt)
291 		return;
292 
293 	if (mlx4_is_mfunc(dev)) {
294 		set_param_l(&in_param, base_qpn);
295 		set_param_h(&in_param, cnt);
296 		err = mlx4_cmd(dev, in_param, RES_QP, RES_OP_RESERVE,
297 			       MLX4_CMD_FREE_RES,
298 			       MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
299 		if (err) {
300 			mlx4_warn(dev, "Failed to release qp range base:%d cnt:%d\n",
301 				  base_qpn, cnt);
302 		}
303 	} else
304 		 __mlx4_qp_release_range(dev, base_qpn, cnt);
305 }
306 EXPORT_SYMBOL_GPL(mlx4_qp_release_range);
307 
308 int __mlx4_qp_alloc_icm(struct mlx4_dev *dev, int qpn)
309 {
310 	struct mlx4_priv *priv = mlx4_priv(dev);
311 	struct mlx4_qp_table *qp_table = &priv->qp_table;
312 	int err;
313 
314 	err = mlx4_table_get(dev, &qp_table->qp_table, qpn);
315 	if (err)
316 		goto err_out;
317 
318 	err = mlx4_table_get(dev, &qp_table->auxc_table, qpn);
319 	if (err)
320 		goto err_put_qp;
321 
322 	err = mlx4_table_get(dev, &qp_table->altc_table, qpn);
323 	if (err)
324 		goto err_put_auxc;
325 
326 	err = mlx4_table_get(dev, &qp_table->rdmarc_table, qpn);
327 	if (err)
328 		goto err_put_altc;
329 
330 	err = mlx4_table_get(dev, &qp_table->cmpt_table, qpn);
331 	if (err)
332 		goto err_put_rdmarc;
333 
334 	return 0;
335 
336 err_put_rdmarc:
337 	mlx4_table_put(dev, &qp_table->rdmarc_table, qpn);
338 
339 err_put_altc:
340 	mlx4_table_put(dev, &qp_table->altc_table, qpn);
341 
342 err_put_auxc:
343 	mlx4_table_put(dev, &qp_table->auxc_table, qpn);
344 
345 err_put_qp:
346 	mlx4_table_put(dev, &qp_table->qp_table, qpn);
347 
348 err_out:
349 	return err;
350 }
351 
352 static int mlx4_qp_alloc_icm(struct mlx4_dev *dev, int qpn)
353 {
354 	u64 param = 0;
355 
356 	if (mlx4_is_mfunc(dev)) {
357 		set_param_l(&param, qpn);
358 		return mlx4_cmd_imm(dev, param, &param, RES_QP, RES_OP_MAP_ICM,
359 				    MLX4_CMD_ALLOC_RES, MLX4_CMD_TIME_CLASS_A,
360 				    MLX4_CMD_WRAPPED);
361 	}
362 	return __mlx4_qp_alloc_icm(dev, qpn);
363 }
364 
365 void __mlx4_qp_free_icm(struct mlx4_dev *dev, int qpn)
366 {
367 	struct mlx4_priv *priv = mlx4_priv(dev);
368 	struct mlx4_qp_table *qp_table = &priv->qp_table;
369 
370 	mlx4_table_put(dev, &qp_table->cmpt_table, qpn);
371 	mlx4_table_put(dev, &qp_table->rdmarc_table, qpn);
372 	mlx4_table_put(dev, &qp_table->altc_table, qpn);
373 	mlx4_table_put(dev, &qp_table->auxc_table, qpn);
374 	mlx4_table_put(dev, &qp_table->qp_table, qpn);
375 }
376 
377 static void mlx4_qp_free_icm(struct mlx4_dev *dev, int qpn)
378 {
379 	u64 in_param = 0;
380 
381 	if (mlx4_is_mfunc(dev)) {
382 		set_param_l(&in_param, qpn);
383 		if (mlx4_cmd(dev, in_param, RES_QP, RES_OP_MAP_ICM,
384 			     MLX4_CMD_FREE_RES, MLX4_CMD_TIME_CLASS_A,
385 			     MLX4_CMD_WRAPPED))
386 			mlx4_warn(dev, "Failed to free icm of qp:%d\n", qpn);
387 	} else
388 		__mlx4_qp_free_icm(dev, qpn);
389 }
390 
391 struct mlx4_qp *mlx4_qp_lookup(struct mlx4_dev *dev, u32 qpn)
392 {
393 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
394 	struct mlx4_qp *qp;
395 
396 	spin_lock_irq(&qp_table->lock);
397 
398 	qp = __mlx4_qp_lookup(dev, qpn);
399 
400 	spin_unlock_irq(&qp_table->lock);
401 	return qp;
402 }
403 
404 int mlx4_qp_alloc(struct mlx4_dev *dev, int qpn, struct mlx4_qp *qp)
405 {
406 	struct mlx4_priv *priv = mlx4_priv(dev);
407 	struct mlx4_qp_table *qp_table = &priv->qp_table;
408 	int err;
409 
410 	if (!qpn)
411 		return -EINVAL;
412 
413 	qp->qpn = qpn;
414 
415 	err = mlx4_qp_alloc_icm(dev, qpn);
416 	if (err)
417 		return err;
418 
419 	spin_lock_irq(&qp_table->lock);
420 	err = radix_tree_insert(&dev->qp_table_tree, qp->qpn &
421 				(dev->caps.num_qps - 1), qp);
422 	spin_unlock_irq(&qp_table->lock);
423 	if (err)
424 		goto err_icm;
425 
426 	refcount_set(&qp->refcount, 1);
427 	init_completion(&qp->free);
428 
429 	return 0;
430 
431 err_icm:
432 	mlx4_qp_free_icm(dev, qpn);
433 	return err;
434 }
435 
436 EXPORT_SYMBOL_GPL(mlx4_qp_alloc);
437 
438 int mlx4_update_qp(struct mlx4_dev *dev, u32 qpn,
439 		   enum mlx4_update_qp_attr attr,
440 		   struct mlx4_update_qp_params *params)
441 {
442 	struct mlx4_cmd_mailbox *mailbox;
443 	struct mlx4_update_qp_context *cmd;
444 	u64 pri_addr_path_mask = 0;
445 	u64 qp_mask = 0;
446 	int err = 0;
447 
448 	if (!attr || (attr & ~MLX4_UPDATE_QP_SUPPORTED_ATTRS))
449 		return -EINVAL;
450 
451 	mailbox = mlx4_alloc_cmd_mailbox(dev);
452 	if (IS_ERR(mailbox))
453 		return PTR_ERR(mailbox);
454 
455 	cmd = (struct mlx4_update_qp_context *)mailbox->buf;
456 
457 	if (attr & MLX4_UPDATE_QP_SMAC) {
458 		pri_addr_path_mask |= 1ULL << MLX4_UPD_QP_PATH_MASK_MAC_INDEX;
459 		cmd->qp_context.pri_path.grh_mylmc = params->smac_index;
460 	}
461 
462 	if (attr & MLX4_UPDATE_QP_ETH_SRC_CHECK_MC_LB) {
463 		if (!(dev->caps.flags2
464 		      & MLX4_DEV_CAP_FLAG2_UPDATE_QP_SRC_CHECK_LB)) {
465 			mlx4_warn(dev,
466 				  "Trying to set src check LB, but it isn't supported\n");
467 			err = -EOPNOTSUPP;
468 			goto out;
469 		}
470 		pri_addr_path_mask |=
471 			1ULL << MLX4_UPD_QP_PATH_MASK_ETH_SRC_CHECK_MC_LB;
472 		if (params->flags &
473 		    MLX4_UPDATE_QP_PARAMS_FLAGS_ETH_CHECK_MC_LB) {
474 			cmd->qp_context.pri_path.fl |=
475 				MLX4_FL_ETH_SRC_CHECK_MC_LB;
476 		}
477 	}
478 
479 	if (attr & MLX4_UPDATE_QP_VSD) {
480 		qp_mask |= 1ULL << MLX4_UPD_QP_MASK_VSD;
481 		if (params->flags & MLX4_UPDATE_QP_PARAMS_FLAGS_VSD_ENABLE)
482 			cmd->qp_context.param3 |= cpu_to_be32(MLX4_STRIP_VLAN);
483 	}
484 
485 	if (attr & MLX4_UPDATE_QP_RATE_LIMIT) {
486 		qp_mask |= 1ULL << MLX4_UPD_QP_MASK_RATE_LIMIT;
487 		cmd->qp_context.rate_limit_params = cpu_to_be16((params->rate_unit << 14) | params->rate_val);
488 	}
489 
490 	if (attr & MLX4_UPDATE_QP_QOS_VPORT) {
491 		if (!(dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_QOS_VPP)) {
492 			mlx4_warn(dev, "Granular QoS per VF is not enabled\n");
493 			err = -EOPNOTSUPP;
494 			goto out;
495 		}
496 
497 		qp_mask |= 1ULL << MLX4_UPD_QP_MASK_QOS_VPP;
498 		cmd->qp_context.qos_vport = params->qos_vport;
499 	}
500 
501 	cmd->primary_addr_path_mask = cpu_to_be64(pri_addr_path_mask);
502 	cmd->qp_mask = cpu_to_be64(qp_mask);
503 
504 	err = mlx4_cmd(dev, mailbox->dma, qpn & 0xffffff, 0,
505 		       MLX4_CMD_UPDATE_QP, MLX4_CMD_TIME_CLASS_A,
506 		       MLX4_CMD_NATIVE);
507 out:
508 	mlx4_free_cmd_mailbox(dev, mailbox);
509 	return err;
510 }
511 EXPORT_SYMBOL_GPL(mlx4_update_qp);
512 
513 void mlx4_qp_remove(struct mlx4_dev *dev, struct mlx4_qp *qp)
514 {
515 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
516 	unsigned long flags;
517 
518 	spin_lock_irqsave(&qp_table->lock, flags);
519 	radix_tree_delete(&dev->qp_table_tree, qp->qpn & (dev->caps.num_qps - 1));
520 	spin_unlock_irqrestore(&qp_table->lock, flags);
521 }
522 EXPORT_SYMBOL_GPL(mlx4_qp_remove);
523 
524 void mlx4_qp_free(struct mlx4_dev *dev, struct mlx4_qp *qp)
525 {
526 	if (refcount_dec_and_test(&qp->refcount))
527 		complete(&qp->free);
528 	wait_for_completion(&qp->free);
529 
530 	mlx4_qp_free_icm(dev, qp->qpn);
531 }
532 EXPORT_SYMBOL_GPL(mlx4_qp_free);
533 
534 static int mlx4_CONF_SPECIAL_QP(struct mlx4_dev *dev, u32 base_qpn)
535 {
536 	return mlx4_cmd(dev, 0, base_qpn, 0, MLX4_CMD_CONF_SPECIAL_QP,
537 			MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
538 }
539 
540 #define MLX4_QP_TABLE_RSS_ETH_PRIORITY 2
541 #define MLX4_QP_TABLE_RAW_ETH_PRIORITY 1
542 #define MLX4_QP_TABLE_RAW_ETH_SIZE     256
543 
544 static int mlx4_create_zones(struct mlx4_dev *dev,
545 			     u32 reserved_bottom_general,
546 			     u32 reserved_top_general,
547 			     u32 reserved_bottom_rss,
548 			     u32 start_offset_rss,
549 			     u32 max_table_offset)
550 {
551 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
552 	struct mlx4_bitmap (*bitmap)[MLX4_QP_TABLE_ZONE_NUM] = NULL;
553 	int bitmap_initialized = 0;
554 	u32 last_offset;
555 	int k;
556 	int err;
557 
558 	qp_table->zones = mlx4_zone_allocator_create(MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP);
559 
560 	if (NULL == qp_table->zones)
561 		return -ENOMEM;
562 
563 	bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
564 
565 	if (NULL == bitmap) {
566 		err = -ENOMEM;
567 		goto free_zone;
568 	}
569 
570 	err = mlx4_bitmap_init(*bitmap + MLX4_QP_TABLE_ZONE_GENERAL, dev->caps.num_qps,
571 			       (1 << 23) - 1, reserved_bottom_general,
572 			       reserved_top_general);
573 
574 	if (err)
575 		goto free_bitmap;
576 
577 	++bitmap_initialized;
578 
579 	err = mlx4_zone_add_one(qp_table->zones, *bitmap + MLX4_QP_TABLE_ZONE_GENERAL,
580 				MLX4_ZONE_FALLBACK_TO_HIGHER_PRIO |
581 				MLX4_ZONE_USE_RR, 0,
582 				0, qp_table->zones_uids + MLX4_QP_TABLE_ZONE_GENERAL);
583 
584 	if (err)
585 		goto free_bitmap;
586 
587 	err = mlx4_bitmap_init(*bitmap + MLX4_QP_TABLE_ZONE_RSS,
588 			       reserved_bottom_rss,
589 			       reserved_bottom_rss - 1,
590 			       dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW],
591 			       reserved_bottom_rss - start_offset_rss);
592 
593 	if (err)
594 		goto free_bitmap;
595 
596 	++bitmap_initialized;
597 
598 	err = mlx4_zone_add_one(qp_table->zones, *bitmap + MLX4_QP_TABLE_ZONE_RSS,
599 				MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO |
600 				MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO |
601 				MLX4_ZONE_USE_RR, MLX4_QP_TABLE_RSS_ETH_PRIORITY,
602 				0, qp_table->zones_uids + MLX4_QP_TABLE_ZONE_RSS);
603 
604 	if (err)
605 		goto free_bitmap;
606 
607 	last_offset = dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW];
608 	/*  We have a single zone for the A0 steering QPs area of the FW. This area
609 	 *  needs to be split into subareas. One set of subareas is for RSS QPs
610 	 *  (in which qp number bits 6 and/or 7 are set); the other set of subareas
611 	 *  is for RAW_ETH QPs, which require that both bits 6 and 7 are zero.
612 	 *  Currently, the values returned by the FW (A0 steering area starting qp number
613 	 *  and A0 steering area size) are such that there are only two subareas -- one
614 	 *  for RSS and one for RAW_ETH.
615 	 */
616 	for (k = MLX4_QP_TABLE_ZONE_RSS + 1; k < sizeof(*bitmap)/sizeof((*bitmap)[0]);
617 	     k++) {
618 		int size;
619 		u32 offset = start_offset_rss;
620 		u32 bf_mask;
621 		u32 requested_size;
622 
623 		/* Assuming MLX4_BF_QP_SKIP_MASK is consecutive ones, this calculates
624 		 * a mask of all LSB bits set until (and not including) the first
625 		 * set bit of  MLX4_BF_QP_SKIP_MASK. For example, if MLX4_BF_QP_SKIP_MASK
626 		 * is 0xc0, bf_mask will be 0x3f.
627 		 */
628 		bf_mask = (MLX4_BF_QP_SKIP_MASK & ~(MLX4_BF_QP_SKIP_MASK - 1)) - 1;
629 		requested_size = min((u32)MLX4_QP_TABLE_RAW_ETH_SIZE, bf_mask + 1);
630 
631 		if (((last_offset & MLX4_BF_QP_SKIP_MASK) &&
632 		     ((int)(max_table_offset - last_offset)) >=
633 		     roundup_pow_of_two(MLX4_BF_QP_SKIP_MASK)) ||
634 		    (!(last_offset & MLX4_BF_QP_SKIP_MASK) &&
635 		     !((last_offset + requested_size - 1) &
636 		       MLX4_BF_QP_SKIP_MASK)))
637 			size = requested_size;
638 		else {
639 			u32 candidate_offset =
640 				(last_offset | MLX4_BF_QP_SKIP_MASK | bf_mask) + 1;
641 
642 			if (last_offset & MLX4_BF_QP_SKIP_MASK)
643 				last_offset = candidate_offset;
644 
645 			/* From this point, the BF bits are 0 */
646 
647 			if (last_offset > max_table_offset) {
648 				/* need to skip */
649 				size = -1;
650 			} else {
651 				size = min3(max_table_offset - last_offset,
652 					    bf_mask - (last_offset & bf_mask),
653 					    requested_size);
654 				if (size < requested_size) {
655 					int candidate_size;
656 
657 					candidate_size = min3(
658 						max_table_offset - candidate_offset,
659 						bf_mask - (last_offset & bf_mask),
660 						requested_size);
661 
662 					/*  We will not take this path if last_offset was
663 					 *  already set above to candidate_offset
664 					 */
665 					if (candidate_size > size) {
666 						last_offset = candidate_offset;
667 						size = candidate_size;
668 					}
669 				}
670 			}
671 		}
672 
673 		if (size > 0) {
674 			/* mlx4_bitmap_alloc_range will find a contiguous range of "size"
675 			 * QPs in which both bits 6 and 7 are zero, because we pass it the
676 			 * MLX4_BF_SKIP_MASK).
677 			 */
678 			offset = mlx4_bitmap_alloc_range(
679 					*bitmap + MLX4_QP_TABLE_ZONE_RSS,
680 					size, 1,
681 					MLX4_BF_QP_SKIP_MASK);
682 
683 			if (offset == (u32)-1) {
684 				err = -ENOMEM;
685 				break;
686 			}
687 
688 			last_offset = offset + size;
689 
690 			err = mlx4_bitmap_init(*bitmap + k, roundup_pow_of_two(size),
691 					       roundup_pow_of_two(size) - 1, 0,
692 					       roundup_pow_of_two(size) - size);
693 		} else {
694 			/* Add an empty bitmap, we'll allocate from different zones (since
695 			 * at least one is reserved)
696 			 */
697 			err = mlx4_bitmap_init(*bitmap + k, 1,
698 					       MLX4_QP_TABLE_RAW_ETH_SIZE - 1, 0,
699 					       0);
700 			mlx4_bitmap_alloc_range(*bitmap + k, 1, 1, 0);
701 		}
702 
703 		if (err)
704 			break;
705 
706 		++bitmap_initialized;
707 
708 		err = mlx4_zone_add_one(qp_table->zones, *bitmap + k,
709 					MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO |
710 					MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO |
711 					MLX4_ZONE_USE_RR, MLX4_QP_TABLE_RAW_ETH_PRIORITY,
712 					offset, qp_table->zones_uids + k);
713 
714 		if (err)
715 			break;
716 	}
717 
718 	if (err)
719 		goto free_bitmap;
720 
721 	qp_table->bitmap_gen = *bitmap;
722 
723 	return err;
724 
725 free_bitmap:
726 	for (k = 0; k < bitmap_initialized; k++)
727 		mlx4_bitmap_cleanup(*bitmap + k);
728 	kfree(bitmap);
729 free_zone:
730 	mlx4_zone_allocator_destroy(qp_table->zones);
731 	return err;
732 }
733 
734 static void mlx4_cleanup_qp_zones(struct mlx4_dev *dev)
735 {
736 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
737 
738 	if (qp_table->zones) {
739 		int i;
740 
741 		for (i = 0;
742 		     i < ARRAY_SIZE(qp_table->zones_uids);
743 		     i++) {
744 			struct mlx4_bitmap *bitmap =
745 				mlx4_zone_get_bitmap(qp_table->zones,
746 						     qp_table->zones_uids[i]);
747 
748 			mlx4_zone_remove_one(qp_table->zones, qp_table->zones_uids[i]);
749 			if (NULL == bitmap)
750 				continue;
751 
752 			mlx4_bitmap_cleanup(bitmap);
753 		}
754 		mlx4_zone_allocator_destroy(qp_table->zones);
755 		kfree(qp_table->bitmap_gen);
756 		qp_table->bitmap_gen = NULL;
757 		qp_table->zones = NULL;
758 	}
759 }
760 
761 int mlx4_init_qp_table(struct mlx4_dev *dev)
762 {
763 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
764 	int err;
765 	int reserved_from_top = 0;
766 	int reserved_from_bot;
767 	int k;
768 	int fixed_reserved_from_bot_rv = 0;
769 	int bottom_reserved_for_rss_bitmap;
770 	u32 max_table_offset = dev->caps.dmfs_high_rate_qpn_base +
771 			dev->caps.dmfs_high_rate_qpn_range;
772 
773 	spin_lock_init(&qp_table->lock);
774 	INIT_RADIX_TREE(&dev->qp_table_tree, GFP_ATOMIC);
775 	if (mlx4_is_slave(dev))
776 		return 0;
777 
778 	/* We reserve 2 extra QPs per port for the special QPs.  The
779 	 * block of special QPs must be aligned to a multiple of 8, so
780 	 * round up.
781 	 *
782 	 * We also reserve the MSB of the 24-bit QP number to indicate
783 	 * that a QP is an XRC QP.
784 	 */
785 	for (k = 0; k <= MLX4_QP_REGION_BOTTOM; k++)
786 		fixed_reserved_from_bot_rv += dev->caps.reserved_qps_cnt[k];
787 
788 	if (fixed_reserved_from_bot_rv < max_table_offset)
789 		fixed_reserved_from_bot_rv = max_table_offset;
790 
791 	/* We reserve at least 1 extra for bitmaps that we don't have enough space for*/
792 	bottom_reserved_for_rss_bitmap =
793 		roundup_pow_of_two(fixed_reserved_from_bot_rv + 1);
794 	dev->phys_caps.base_sqpn = ALIGN(bottom_reserved_for_rss_bitmap, 8);
795 
796 	{
797 		int sort[MLX4_NUM_QP_REGION];
798 		int i, j;
799 		int last_base = dev->caps.num_qps;
800 
801 		for (i = 1; i < MLX4_NUM_QP_REGION; ++i)
802 			sort[i] = i;
803 
804 		for (i = MLX4_NUM_QP_REGION; i > MLX4_QP_REGION_BOTTOM; --i) {
805 			for (j = MLX4_QP_REGION_BOTTOM + 2; j < i; ++j) {
806 				if (dev->caps.reserved_qps_cnt[sort[j]] >
807 				    dev->caps.reserved_qps_cnt[sort[j - 1]])
808 					swap(sort[j], sort[j - 1]);
809 			}
810 		}
811 
812 		for (i = MLX4_QP_REGION_BOTTOM + 1; i < MLX4_NUM_QP_REGION; ++i) {
813 			last_base -= dev->caps.reserved_qps_cnt[sort[i]];
814 			dev->caps.reserved_qps_base[sort[i]] = last_base;
815 			reserved_from_top +=
816 				dev->caps.reserved_qps_cnt[sort[i]];
817 		}
818 	}
819 
820        /* Reserve 8 real SQPs in both native and SRIOV modes.
821 	* In addition, in SRIOV mode, reserve 8 proxy SQPs per function
822 	* (for all PFs and VFs), and 8 corresponding tunnel QPs.
823 	* Each proxy SQP works opposite its own tunnel QP.
824 	*
825 	* The QPs are arranged as follows:
826 	* a. 8 real SQPs
827 	* b. All the proxy SQPs (8 per function)
828 	* c. All the tunnel QPs (8 per function)
829 	*/
830 	reserved_from_bot = mlx4_num_reserved_sqps(dev);
831 	if (reserved_from_bot + reserved_from_top > dev->caps.num_qps) {
832 		mlx4_err(dev, "Number of reserved QPs is higher than number of QPs\n");
833 		return -EINVAL;
834 	}
835 
836 	err = mlx4_create_zones(dev, reserved_from_bot, reserved_from_bot,
837 				bottom_reserved_for_rss_bitmap,
838 				fixed_reserved_from_bot_rv,
839 				max_table_offset);
840 
841 	if (err)
842 		return err;
843 
844 	if (mlx4_is_mfunc(dev)) {
845 		/* for PPF use */
846 		dev->phys_caps.base_proxy_sqpn = dev->phys_caps.base_sqpn + 8;
847 		dev->phys_caps.base_tunnel_sqpn = dev->phys_caps.base_sqpn + 8 + 8 * MLX4_MFUNC_MAX;
848 
849 		/* In mfunc, calculate proxy and tunnel qp offsets for the PF here,
850 		 * since the PF does not call mlx4_slave_caps */
851 		dev->caps.spec_qps = kcalloc(dev->caps.num_ports,
852 					     sizeof(*dev->caps.spec_qps),
853 					     GFP_KERNEL);
854 		if (!dev->caps.spec_qps) {
855 			err = -ENOMEM;
856 			goto err_mem;
857 		}
858 
859 		for (k = 0; k < dev->caps.num_ports; k++) {
860 			dev->caps.spec_qps[k].qp0_proxy = dev->phys_caps.base_proxy_sqpn +
861 				8 * mlx4_master_func_num(dev) + k;
862 			dev->caps.spec_qps[k].qp0_tunnel = dev->caps.spec_qps[k].qp0_proxy + 8 * MLX4_MFUNC_MAX;
863 			dev->caps.spec_qps[k].qp1_proxy = dev->phys_caps.base_proxy_sqpn +
864 				8 * mlx4_master_func_num(dev) + MLX4_MAX_PORTS + k;
865 			dev->caps.spec_qps[k].qp1_tunnel = dev->caps.spec_qps[k].qp1_proxy + 8 * MLX4_MFUNC_MAX;
866 		}
867 	}
868 
869 
870 	err = mlx4_CONF_SPECIAL_QP(dev, dev->phys_caps.base_sqpn);
871 	if (err)
872 		goto err_mem;
873 
874 	return err;
875 
876 err_mem:
877 	kfree(dev->caps.spec_qps);
878 	dev->caps.spec_qps = NULL;
879 	mlx4_cleanup_qp_zones(dev);
880 	return err;
881 }
882 
883 void mlx4_cleanup_qp_table(struct mlx4_dev *dev)
884 {
885 	if (mlx4_is_slave(dev))
886 		return;
887 
888 	mlx4_CONF_SPECIAL_QP(dev, 0);
889 
890 	mlx4_cleanup_qp_zones(dev);
891 }
892 
893 int mlx4_qp_query(struct mlx4_dev *dev, struct mlx4_qp *qp,
894 		  struct mlx4_qp_context *context)
895 {
896 	struct mlx4_cmd_mailbox *mailbox;
897 	int err;
898 
899 	mailbox = mlx4_alloc_cmd_mailbox(dev);
900 	if (IS_ERR(mailbox))
901 		return PTR_ERR(mailbox);
902 
903 	err = mlx4_cmd_box(dev, 0, mailbox->dma, qp->qpn, 0,
904 			   MLX4_CMD_QUERY_QP, MLX4_CMD_TIME_CLASS_A,
905 			   MLX4_CMD_WRAPPED);
906 	if (!err)
907 		memcpy(context, mailbox->buf + 8, sizeof(*context));
908 
909 	mlx4_free_cmd_mailbox(dev, mailbox);
910 	return err;
911 }
912 EXPORT_SYMBOL_GPL(mlx4_qp_query);
913 
914 int mlx4_qp_to_ready(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
915 		     struct mlx4_qp_context *context,
916 		     struct mlx4_qp *qp, enum mlx4_qp_state *qp_state)
917 {
918 	int err;
919 	int i;
920 	static const enum mlx4_qp_state states[] = {
921 		MLX4_QP_STATE_RST,
922 		MLX4_QP_STATE_INIT,
923 		MLX4_QP_STATE_RTR,
924 		MLX4_QP_STATE_RTS
925 	};
926 
927 	for (i = 0; i < ARRAY_SIZE(states) - 1; i++) {
928 		context->flags &= cpu_to_be32(~(0xf << 28));
929 		context->flags |= cpu_to_be32(states[i + 1] << 28);
930 		if (states[i + 1] != MLX4_QP_STATE_RTR)
931 			context->params2 &= ~cpu_to_be32(MLX4_QP_BIT_FPP);
932 		err = mlx4_qp_modify(dev, mtt, states[i], states[i + 1],
933 				     context, 0, 0, qp);
934 		if (err) {
935 			mlx4_err(dev, "Failed to bring QP to state: %d with error: %d\n",
936 				 states[i + 1], err);
937 			return err;
938 		}
939 
940 		*qp_state = states[i + 1];
941 	}
942 
943 	return 0;
944 }
945 EXPORT_SYMBOL_GPL(mlx4_qp_to_ready);
946 
947 u16 mlx4_qp_roce_entropy(struct mlx4_dev *dev, u32 qpn)
948 {
949 	struct mlx4_qp_context context;
950 	struct mlx4_qp qp;
951 	int err;
952 
953 	qp.qpn = qpn;
954 	err = mlx4_qp_query(dev, &qp, &context);
955 	if (!err) {
956 		u32 dest_qpn = be32_to_cpu(context.remote_qpn) & 0xffffff;
957 		u16 folded_dst = folded_qp(dest_qpn);
958 		u16 folded_src = folded_qp(qpn);
959 
960 		return (dest_qpn != qpn) ?
961 			((folded_dst ^ folded_src) | 0xC000) :
962 			folded_src | 0xC000;
963 	}
964 	return 0xdead;
965 }
966