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 	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)
249 {
250 	u64 in_param = 0;
251 	u64 out_param;
252 	int err;
253 
254 	/* Turn off all unsupported QP allocation flags */
255 	flags &= dev->caps.alloc_res_qp_mask;
256 
257 	if (mlx4_is_mfunc(dev)) {
258 		set_param_l(&in_param, (((u32)flags) << 24) | (u32)cnt);
259 		set_param_h(&in_param, align);
260 		err = mlx4_cmd_imm(dev, in_param, &out_param,
261 				   RES_QP, RES_OP_RESERVE,
262 				   MLX4_CMD_ALLOC_RES,
263 				   MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
264 		if (err)
265 			return err;
266 
267 		*base = get_param_l(&out_param);
268 		return 0;
269 	}
270 	return __mlx4_qp_reserve_range(dev, cnt, align, base, flags);
271 }
272 EXPORT_SYMBOL_GPL(mlx4_qp_reserve_range);
273 
274 void __mlx4_qp_release_range(struct mlx4_dev *dev, int base_qpn, int cnt)
275 {
276 	struct mlx4_priv *priv = mlx4_priv(dev);
277 	struct mlx4_qp_table *qp_table = &priv->qp_table;
278 
279 	if (mlx4_is_qp_reserved(dev, (u32) base_qpn))
280 		return;
281 	mlx4_zone_free_entries_unique(qp_table->zones, base_qpn, cnt);
282 }
283 
284 void mlx4_qp_release_range(struct mlx4_dev *dev, int base_qpn, int cnt)
285 {
286 	u64 in_param = 0;
287 	int err;
288 
289 	if (mlx4_is_mfunc(dev)) {
290 		set_param_l(&in_param, base_qpn);
291 		set_param_h(&in_param, cnt);
292 		err = mlx4_cmd(dev, in_param, RES_QP, RES_OP_RESERVE,
293 			       MLX4_CMD_FREE_RES,
294 			       MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
295 		if (err) {
296 			mlx4_warn(dev, "Failed to release qp range base:%d cnt:%d\n",
297 				  base_qpn, cnt);
298 		}
299 	} else
300 		 __mlx4_qp_release_range(dev, base_qpn, cnt);
301 }
302 EXPORT_SYMBOL_GPL(mlx4_qp_release_range);
303 
304 int __mlx4_qp_alloc_icm(struct mlx4_dev *dev, int qpn)
305 {
306 	struct mlx4_priv *priv = mlx4_priv(dev);
307 	struct mlx4_qp_table *qp_table = &priv->qp_table;
308 	int err;
309 
310 	err = mlx4_table_get(dev, &qp_table->qp_table, qpn);
311 	if (err)
312 		goto err_out;
313 
314 	err = mlx4_table_get(dev, &qp_table->auxc_table, qpn);
315 	if (err)
316 		goto err_put_qp;
317 
318 	err = mlx4_table_get(dev, &qp_table->altc_table, qpn);
319 	if (err)
320 		goto err_put_auxc;
321 
322 	err = mlx4_table_get(dev, &qp_table->rdmarc_table, qpn);
323 	if (err)
324 		goto err_put_altc;
325 
326 	err = mlx4_table_get(dev, &qp_table->cmpt_table, qpn);
327 	if (err)
328 		goto err_put_rdmarc;
329 
330 	return 0;
331 
332 err_put_rdmarc:
333 	mlx4_table_put(dev, &qp_table->rdmarc_table, qpn);
334 
335 err_put_altc:
336 	mlx4_table_put(dev, &qp_table->altc_table, qpn);
337 
338 err_put_auxc:
339 	mlx4_table_put(dev, &qp_table->auxc_table, qpn);
340 
341 err_put_qp:
342 	mlx4_table_put(dev, &qp_table->qp_table, qpn);
343 
344 err_out:
345 	return err;
346 }
347 
348 static int mlx4_qp_alloc_icm(struct mlx4_dev *dev, int qpn)
349 {
350 	u64 param = 0;
351 
352 	if (mlx4_is_mfunc(dev)) {
353 		set_param_l(&param, qpn);
354 		return mlx4_cmd_imm(dev, param, &param, RES_QP, RES_OP_MAP_ICM,
355 				    MLX4_CMD_ALLOC_RES, MLX4_CMD_TIME_CLASS_A,
356 				    MLX4_CMD_WRAPPED);
357 	}
358 	return __mlx4_qp_alloc_icm(dev, qpn);
359 }
360 
361 void __mlx4_qp_free_icm(struct mlx4_dev *dev, int qpn)
362 {
363 	struct mlx4_priv *priv = mlx4_priv(dev);
364 	struct mlx4_qp_table *qp_table = &priv->qp_table;
365 
366 	mlx4_table_put(dev, &qp_table->cmpt_table, qpn);
367 	mlx4_table_put(dev, &qp_table->rdmarc_table, qpn);
368 	mlx4_table_put(dev, &qp_table->altc_table, qpn);
369 	mlx4_table_put(dev, &qp_table->auxc_table, qpn);
370 	mlx4_table_put(dev, &qp_table->qp_table, qpn);
371 }
372 
373 static void mlx4_qp_free_icm(struct mlx4_dev *dev, int qpn)
374 {
375 	u64 in_param = 0;
376 
377 	if (mlx4_is_mfunc(dev)) {
378 		set_param_l(&in_param, qpn);
379 		if (mlx4_cmd(dev, in_param, RES_QP, RES_OP_MAP_ICM,
380 			     MLX4_CMD_FREE_RES, MLX4_CMD_TIME_CLASS_A,
381 			     MLX4_CMD_WRAPPED))
382 			mlx4_warn(dev, "Failed to free icm of qp:%d\n", qpn);
383 	} else
384 		__mlx4_qp_free_icm(dev, qpn);
385 }
386 
387 struct mlx4_qp *mlx4_qp_lookup(struct mlx4_dev *dev, u32 qpn)
388 {
389 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
390 	struct mlx4_qp *qp;
391 
392 	spin_lock(&qp_table->lock);
393 
394 	qp = __mlx4_qp_lookup(dev, qpn);
395 
396 	spin_unlock(&qp_table->lock);
397 	return qp;
398 }
399 
400 int mlx4_qp_alloc(struct mlx4_dev *dev, int qpn, struct mlx4_qp *qp)
401 {
402 	struct mlx4_priv *priv = mlx4_priv(dev);
403 	struct mlx4_qp_table *qp_table = &priv->qp_table;
404 	int err;
405 
406 	if (!qpn)
407 		return -EINVAL;
408 
409 	qp->qpn = qpn;
410 
411 	err = mlx4_qp_alloc_icm(dev, qpn);
412 	if (err)
413 		return err;
414 
415 	spin_lock_irq(&qp_table->lock);
416 	err = radix_tree_insert(&dev->qp_table_tree, qp->qpn &
417 				(dev->caps.num_qps - 1), qp);
418 	spin_unlock_irq(&qp_table->lock);
419 	if (err)
420 		goto err_icm;
421 
422 	atomic_set(&qp->refcount, 1);
423 	init_completion(&qp->free);
424 
425 	return 0;
426 
427 err_icm:
428 	mlx4_qp_free_icm(dev, qpn);
429 	return err;
430 }
431 
432 EXPORT_SYMBOL_GPL(mlx4_qp_alloc);
433 
434 int mlx4_update_qp(struct mlx4_dev *dev, u32 qpn,
435 		   enum mlx4_update_qp_attr attr,
436 		   struct mlx4_update_qp_params *params)
437 {
438 	struct mlx4_cmd_mailbox *mailbox;
439 	struct mlx4_update_qp_context *cmd;
440 	u64 pri_addr_path_mask = 0;
441 	u64 qp_mask = 0;
442 	int err = 0;
443 
444 	if (!attr || (attr & ~MLX4_UPDATE_QP_SUPPORTED_ATTRS))
445 		return -EINVAL;
446 
447 	mailbox = mlx4_alloc_cmd_mailbox(dev);
448 	if (IS_ERR(mailbox))
449 		return PTR_ERR(mailbox);
450 
451 	cmd = (struct mlx4_update_qp_context *)mailbox->buf;
452 
453 	if (attr & MLX4_UPDATE_QP_SMAC) {
454 		pri_addr_path_mask |= 1ULL << MLX4_UPD_QP_PATH_MASK_MAC_INDEX;
455 		cmd->qp_context.pri_path.grh_mylmc = params->smac_index;
456 	}
457 
458 	if (attr & MLX4_UPDATE_QP_ETH_SRC_CHECK_MC_LB) {
459 		if (!(dev->caps.flags2
460 		      & MLX4_DEV_CAP_FLAG2_UPDATE_QP_SRC_CHECK_LB)) {
461 			mlx4_warn(dev,
462 				  "Trying to set src check LB, but it isn't supported\n");
463 			err = -EOPNOTSUPP;
464 			goto out;
465 		}
466 		pri_addr_path_mask |=
467 			1ULL << MLX4_UPD_QP_PATH_MASK_ETH_SRC_CHECK_MC_LB;
468 		if (params->flags &
469 		    MLX4_UPDATE_QP_PARAMS_FLAGS_ETH_CHECK_MC_LB) {
470 			cmd->qp_context.pri_path.fl |=
471 				MLX4_FL_ETH_SRC_CHECK_MC_LB;
472 		}
473 	}
474 
475 	if (attr & MLX4_UPDATE_QP_VSD) {
476 		qp_mask |= 1ULL << MLX4_UPD_QP_MASK_VSD;
477 		if (params->flags & MLX4_UPDATE_QP_PARAMS_FLAGS_VSD_ENABLE)
478 			cmd->qp_context.param3 |= cpu_to_be32(MLX4_STRIP_VLAN);
479 	}
480 
481 	if (attr & MLX4_UPDATE_QP_RATE_LIMIT) {
482 		qp_mask |= 1ULL << MLX4_UPD_QP_MASK_RATE_LIMIT;
483 		cmd->qp_context.rate_limit_params = cpu_to_be16((params->rate_unit << 14) | params->rate_val);
484 	}
485 
486 	if (attr & MLX4_UPDATE_QP_QOS_VPORT) {
487 		if (!(dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_QOS_VPP)) {
488 			mlx4_warn(dev, "Granular QoS per VF is not enabled\n");
489 			err = -EOPNOTSUPP;
490 			goto out;
491 		}
492 
493 		qp_mask |= 1ULL << MLX4_UPD_QP_MASK_QOS_VPP;
494 		cmd->qp_context.qos_vport = params->qos_vport;
495 	}
496 
497 	cmd->primary_addr_path_mask = cpu_to_be64(pri_addr_path_mask);
498 	cmd->qp_mask = cpu_to_be64(qp_mask);
499 
500 	err = mlx4_cmd(dev, mailbox->dma, qpn & 0xffffff, 0,
501 		       MLX4_CMD_UPDATE_QP, MLX4_CMD_TIME_CLASS_A,
502 		       MLX4_CMD_NATIVE);
503 out:
504 	mlx4_free_cmd_mailbox(dev, mailbox);
505 	return err;
506 }
507 EXPORT_SYMBOL_GPL(mlx4_update_qp);
508 
509 void mlx4_qp_remove(struct mlx4_dev *dev, struct mlx4_qp *qp)
510 {
511 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
512 	unsigned long flags;
513 
514 	spin_lock_irqsave(&qp_table->lock, flags);
515 	radix_tree_delete(&dev->qp_table_tree, qp->qpn & (dev->caps.num_qps - 1));
516 	spin_unlock_irqrestore(&qp_table->lock, flags);
517 }
518 EXPORT_SYMBOL_GPL(mlx4_qp_remove);
519 
520 void mlx4_qp_free(struct mlx4_dev *dev, struct mlx4_qp *qp)
521 {
522 	if (atomic_dec_and_test(&qp->refcount))
523 		complete(&qp->free);
524 	wait_for_completion(&qp->free);
525 
526 	mlx4_qp_free_icm(dev, qp->qpn);
527 }
528 EXPORT_SYMBOL_GPL(mlx4_qp_free);
529 
530 static int mlx4_CONF_SPECIAL_QP(struct mlx4_dev *dev, u32 base_qpn)
531 {
532 	return mlx4_cmd(dev, 0, base_qpn, 0, MLX4_CMD_CONF_SPECIAL_QP,
533 			MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
534 }
535 
536 #define MLX4_QP_TABLE_RSS_ETH_PRIORITY 2
537 #define MLX4_QP_TABLE_RAW_ETH_PRIORITY 1
538 #define MLX4_QP_TABLE_RAW_ETH_SIZE     256
539 
540 static int mlx4_create_zones(struct mlx4_dev *dev,
541 			     u32 reserved_bottom_general,
542 			     u32 reserved_top_general,
543 			     u32 reserved_bottom_rss,
544 			     u32 start_offset_rss,
545 			     u32 max_table_offset)
546 {
547 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
548 	struct mlx4_bitmap (*bitmap)[MLX4_QP_TABLE_ZONE_NUM] = NULL;
549 	int bitmap_initialized = 0;
550 	u32 last_offset;
551 	int k;
552 	int err;
553 
554 	qp_table->zones = mlx4_zone_allocator_create(MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP);
555 
556 	if (NULL == qp_table->zones)
557 		return -ENOMEM;
558 
559 	bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
560 
561 	if (NULL == bitmap) {
562 		err = -ENOMEM;
563 		goto free_zone;
564 	}
565 
566 	err = mlx4_bitmap_init(*bitmap + MLX4_QP_TABLE_ZONE_GENERAL, dev->caps.num_qps,
567 			       (1 << 23) - 1, reserved_bottom_general,
568 			       reserved_top_general);
569 
570 	if (err)
571 		goto free_bitmap;
572 
573 	++bitmap_initialized;
574 
575 	err = mlx4_zone_add_one(qp_table->zones, *bitmap + MLX4_QP_TABLE_ZONE_GENERAL,
576 				MLX4_ZONE_FALLBACK_TO_HIGHER_PRIO |
577 				MLX4_ZONE_USE_RR, 0,
578 				0, qp_table->zones_uids + MLX4_QP_TABLE_ZONE_GENERAL);
579 
580 	if (err)
581 		goto free_bitmap;
582 
583 	err = mlx4_bitmap_init(*bitmap + MLX4_QP_TABLE_ZONE_RSS,
584 			       reserved_bottom_rss,
585 			       reserved_bottom_rss - 1,
586 			       dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW],
587 			       reserved_bottom_rss - start_offset_rss);
588 
589 	if (err)
590 		goto free_bitmap;
591 
592 	++bitmap_initialized;
593 
594 	err = mlx4_zone_add_one(qp_table->zones, *bitmap + MLX4_QP_TABLE_ZONE_RSS,
595 				MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO |
596 				MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO |
597 				MLX4_ZONE_USE_RR, MLX4_QP_TABLE_RSS_ETH_PRIORITY,
598 				0, qp_table->zones_uids + MLX4_QP_TABLE_ZONE_RSS);
599 
600 	if (err)
601 		goto free_bitmap;
602 
603 	last_offset = dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW];
604 	/*  We have a single zone for the A0 steering QPs area of the FW. This area
605 	 *  needs to be split into subareas. One set of subareas is for RSS QPs
606 	 *  (in which qp number bits 6 and/or 7 are set); the other set of subareas
607 	 *  is for RAW_ETH QPs, which require that both bits 6 and 7 are zero.
608 	 *  Currently, the values returned by the FW (A0 steering area starting qp number
609 	 *  and A0 steering area size) are such that there are only two subareas -- one
610 	 *  for RSS and one for RAW_ETH.
611 	 */
612 	for (k = MLX4_QP_TABLE_ZONE_RSS + 1; k < sizeof(*bitmap)/sizeof((*bitmap)[0]);
613 	     k++) {
614 		int size;
615 		u32 offset = start_offset_rss;
616 		u32 bf_mask;
617 		u32 requested_size;
618 
619 		/* Assuming MLX4_BF_QP_SKIP_MASK is consecutive ones, this calculates
620 		 * a mask of all LSB bits set until (and not including) the first
621 		 * set bit of  MLX4_BF_QP_SKIP_MASK. For example, if MLX4_BF_QP_SKIP_MASK
622 		 * is 0xc0, bf_mask will be 0x3f.
623 		 */
624 		bf_mask = (MLX4_BF_QP_SKIP_MASK & ~(MLX4_BF_QP_SKIP_MASK - 1)) - 1;
625 		requested_size = min((u32)MLX4_QP_TABLE_RAW_ETH_SIZE, bf_mask + 1);
626 
627 		if (((last_offset & MLX4_BF_QP_SKIP_MASK) &&
628 		     ((int)(max_table_offset - last_offset)) >=
629 		     roundup_pow_of_two(MLX4_BF_QP_SKIP_MASK)) ||
630 		    (!(last_offset & MLX4_BF_QP_SKIP_MASK) &&
631 		     !((last_offset + requested_size - 1) &
632 		       MLX4_BF_QP_SKIP_MASK)))
633 			size = requested_size;
634 		else {
635 			u32 candidate_offset =
636 				(last_offset | MLX4_BF_QP_SKIP_MASK | bf_mask) + 1;
637 
638 			if (last_offset & MLX4_BF_QP_SKIP_MASK)
639 				last_offset = candidate_offset;
640 
641 			/* From this point, the BF bits are 0 */
642 
643 			if (last_offset > max_table_offset) {
644 				/* need to skip */
645 				size = -1;
646 			} else {
647 				size = min3(max_table_offset - last_offset,
648 					    bf_mask - (last_offset & bf_mask),
649 					    requested_size);
650 				if (size < requested_size) {
651 					int candidate_size;
652 
653 					candidate_size = min3(
654 						max_table_offset - candidate_offset,
655 						bf_mask - (last_offset & bf_mask),
656 						requested_size);
657 
658 					/*  We will not take this path if last_offset was
659 					 *  already set above to candidate_offset
660 					 */
661 					if (candidate_size > size) {
662 						last_offset = candidate_offset;
663 						size = candidate_size;
664 					}
665 				}
666 			}
667 		}
668 
669 		if (size > 0) {
670 			/* mlx4_bitmap_alloc_range will find a contiguous range of "size"
671 			 * QPs in which both bits 6 and 7 are zero, because we pass it the
672 			 * MLX4_BF_SKIP_MASK).
673 			 */
674 			offset = mlx4_bitmap_alloc_range(
675 					*bitmap + MLX4_QP_TABLE_ZONE_RSS,
676 					size, 1,
677 					MLX4_BF_QP_SKIP_MASK);
678 
679 			if (offset == (u32)-1) {
680 				err = -ENOMEM;
681 				break;
682 			}
683 
684 			last_offset = offset + size;
685 
686 			err = mlx4_bitmap_init(*bitmap + k, roundup_pow_of_two(size),
687 					       roundup_pow_of_two(size) - 1, 0,
688 					       roundup_pow_of_two(size) - size);
689 		} else {
690 			/* Add an empty bitmap, we'll allocate from different zones (since
691 			 * at least one is reserved)
692 			 */
693 			err = mlx4_bitmap_init(*bitmap + k, 1,
694 					       MLX4_QP_TABLE_RAW_ETH_SIZE - 1, 0,
695 					       0);
696 			mlx4_bitmap_alloc_range(*bitmap + k, 1, 1, 0);
697 		}
698 
699 		if (err)
700 			break;
701 
702 		++bitmap_initialized;
703 
704 		err = mlx4_zone_add_one(qp_table->zones, *bitmap + k,
705 					MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO |
706 					MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO |
707 					MLX4_ZONE_USE_RR, MLX4_QP_TABLE_RAW_ETH_PRIORITY,
708 					offset, qp_table->zones_uids + k);
709 
710 		if (err)
711 			break;
712 	}
713 
714 	if (err)
715 		goto free_bitmap;
716 
717 	qp_table->bitmap_gen = *bitmap;
718 
719 	return err;
720 
721 free_bitmap:
722 	for (k = 0; k < bitmap_initialized; k++)
723 		mlx4_bitmap_cleanup(*bitmap + k);
724 	kfree(bitmap);
725 free_zone:
726 	mlx4_zone_allocator_destroy(qp_table->zones);
727 	return err;
728 }
729 
730 static void mlx4_cleanup_qp_zones(struct mlx4_dev *dev)
731 {
732 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
733 
734 	if (qp_table->zones) {
735 		int i;
736 
737 		for (i = 0;
738 		     i < sizeof(qp_table->zones_uids)/sizeof(qp_table->zones_uids[0]);
739 		     i++) {
740 			struct mlx4_bitmap *bitmap =
741 				mlx4_zone_get_bitmap(qp_table->zones,
742 						     qp_table->zones_uids[i]);
743 
744 			mlx4_zone_remove_one(qp_table->zones, qp_table->zones_uids[i]);
745 			if (NULL == bitmap)
746 				continue;
747 
748 			mlx4_bitmap_cleanup(bitmap);
749 		}
750 		mlx4_zone_allocator_destroy(qp_table->zones);
751 		kfree(qp_table->bitmap_gen);
752 		qp_table->bitmap_gen = NULL;
753 		qp_table->zones = NULL;
754 	}
755 }
756 
757 int mlx4_init_qp_table(struct mlx4_dev *dev)
758 {
759 	struct mlx4_qp_table *qp_table = &mlx4_priv(dev)->qp_table;
760 	int err;
761 	int reserved_from_top = 0;
762 	int reserved_from_bot;
763 	int k;
764 	int fixed_reserved_from_bot_rv = 0;
765 	int bottom_reserved_for_rss_bitmap;
766 	u32 max_table_offset = dev->caps.dmfs_high_rate_qpn_base +
767 			dev->caps.dmfs_high_rate_qpn_range;
768 
769 	spin_lock_init(&qp_table->lock);
770 	INIT_RADIX_TREE(&dev->qp_table_tree, GFP_ATOMIC);
771 	if (mlx4_is_slave(dev))
772 		return 0;
773 
774 	/* We reserve 2 extra QPs per port for the special QPs.  The
775 	 * block of special QPs must be aligned to a multiple of 8, so
776 	 * round up.
777 	 *
778 	 * We also reserve the MSB of the 24-bit QP number to indicate
779 	 * that a QP is an XRC QP.
780 	 */
781 	for (k = 0; k <= MLX4_QP_REGION_BOTTOM; k++)
782 		fixed_reserved_from_bot_rv += dev->caps.reserved_qps_cnt[k];
783 
784 	if (fixed_reserved_from_bot_rv < max_table_offset)
785 		fixed_reserved_from_bot_rv = max_table_offset;
786 
787 	/* We reserve at least 1 extra for bitmaps that we don't have enough space for*/
788 	bottom_reserved_for_rss_bitmap =
789 		roundup_pow_of_two(fixed_reserved_from_bot_rv + 1);
790 	dev->phys_caps.base_sqpn = ALIGN(bottom_reserved_for_rss_bitmap, 8);
791 
792 	{
793 		int sort[MLX4_NUM_QP_REGION];
794 		int i, j;
795 		int last_base = dev->caps.num_qps;
796 
797 		for (i = 1; i < MLX4_NUM_QP_REGION; ++i)
798 			sort[i] = i;
799 
800 		for (i = MLX4_NUM_QP_REGION; i > MLX4_QP_REGION_BOTTOM; --i) {
801 			for (j = MLX4_QP_REGION_BOTTOM + 2; j < i; ++j) {
802 				if (dev->caps.reserved_qps_cnt[sort[j]] >
803 				    dev->caps.reserved_qps_cnt[sort[j - 1]])
804 					swap(sort[j], sort[j - 1]);
805 			}
806 		}
807 
808 		for (i = MLX4_QP_REGION_BOTTOM + 1; i < MLX4_NUM_QP_REGION; ++i) {
809 			last_base -= dev->caps.reserved_qps_cnt[sort[i]];
810 			dev->caps.reserved_qps_base[sort[i]] = last_base;
811 			reserved_from_top +=
812 				dev->caps.reserved_qps_cnt[sort[i]];
813 		}
814 	}
815 
816        /* Reserve 8 real SQPs in both native and SRIOV modes.
817 	* In addition, in SRIOV mode, reserve 8 proxy SQPs per function
818 	* (for all PFs and VFs), and 8 corresponding tunnel QPs.
819 	* Each proxy SQP works opposite its own tunnel QP.
820 	*
821 	* The QPs are arranged as follows:
822 	* a. 8 real SQPs
823 	* b. All the proxy SQPs (8 per function)
824 	* c. All the tunnel QPs (8 per function)
825 	*/
826 	reserved_from_bot = mlx4_num_reserved_sqps(dev);
827 	if (reserved_from_bot + reserved_from_top > dev->caps.num_qps) {
828 		mlx4_err(dev, "Number of reserved QPs is higher than number of QPs\n");
829 		return -EINVAL;
830 	}
831 
832 	err = mlx4_create_zones(dev, reserved_from_bot, reserved_from_bot,
833 				bottom_reserved_for_rss_bitmap,
834 				fixed_reserved_from_bot_rv,
835 				max_table_offset);
836 
837 	if (err)
838 		return err;
839 
840 	if (mlx4_is_mfunc(dev)) {
841 		/* for PPF use */
842 		dev->phys_caps.base_proxy_sqpn = dev->phys_caps.base_sqpn + 8;
843 		dev->phys_caps.base_tunnel_sqpn = dev->phys_caps.base_sqpn + 8 + 8 * MLX4_MFUNC_MAX;
844 
845 		/* In mfunc, calculate proxy and tunnel qp offsets for the PF here,
846 		 * since the PF does not call mlx4_slave_caps */
847 		dev->caps.qp0_tunnel = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
848 		dev->caps.qp0_proxy = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
849 		dev->caps.qp1_tunnel = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
850 		dev->caps.qp1_proxy = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
851 
852 		if (!dev->caps.qp0_tunnel || !dev->caps.qp0_proxy ||
853 		    !dev->caps.qp1_tunnel || !dev->caps.qp1_proxy) {
854 			err = -ENOMEM;
855 			goto err_mem;
856 		}
857 
858 		for (k = 0; k < dev->caps.num_ports; k++) {
859 			dev->caps.qp0_proxy[k] = dev->phys_caps.base_proxy_sqpn +
860 				8 * mlx4_master_func_num(dev) + k;
861 			dev->caps.qp0_tunnel[k] = dev->caps.qp0_proxy[k] + 8 * MLX4_MFUNC_MAX;
862 			dev->caps.qp1_proxy[k] = dev->phys_caps.base_proxy_sqpn +
863 				8 * mlx4_master_func_num(dev) + MLX4_MAX_PORTS + k;
864 			dev->caps.qp1_tunnel[k] = dev->caps.qp1_proxy[k] + 8 * MLX4_MFUNC_MAX;
865 		}
866 	}
867 
868 
869 	err = mlx4_CONF_SPECIAL_QP(dev, dev->phys_caps.base_sqpn);
870 	if (err)
871 		goto err_mem;
872 
873 	return err;
874 
875 err_mem:
876 	kfree(dev->caps.qp0_tunnel);
877 	kfree(dev->caps.qp0_proxy);
878 	kfree(dev->caps.qp1_tunnel);
879 	kfree(dev->caps.qp1_proxy);
880 	dev->caps.qp0_tunnel = dev->caps.qp0_proxy =
881 		dev->caps.qp1_tunnel = dev->caps.qp1_proxy = NULL;
882 	mlx4_cleanup_qp_zones(dev);
883 	return err;
884 }
885 
886 void mlx4_cleanup_qp_table(struct mlx4_dev *dev)
887 {
888 	if (mlx4_is_slave(dev))
889 		return;
890 
891 	mlx4_CONF_SPECIAL_QP(dev, 0);
892 
893 	mlx4_cleanup_qp_zones(dev);
894 }
895 
896 int mlx4_qp_query(struct mlx4_dev *dev, struct mlx4_qp *qp,
897 		  struct mlx4_qp_context *context)
898 {
899 	struct mlx4_cmd_mailbox *mailbox;
900 	int err;
901 
902 	mailbox = mlx4_alloc_cmd_mailbox(dev);
903 	if (IS_ERR(mailbox))
904 		return PTR_ERR(mailbox);
905 
906 	err = mlx4_cmd_box(dev, 0, mailbox->dma, qp->qpn, 0,
907 			   MLX4_CMD_QUERY_QP, MLX4_CMD_TIME_CLASS_A,
908 			   MLX4_CMD_WRAPPED);
909 	if (!err)
910 		memcpy(context, mailbox->buf + 8, sizeof *context);
911 
912 	mlx4_free_cmd_mailbox(dev, mailbox);
913 	return err;
914 }
915 EXPORT_SYMBOL_GPL(mlx4_qp_query);
916 
917 int mlx4_qp_to_ready(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
918 		     struct mlx4_qp_context *context,
919 		     struct mlx4_qp *qp, enum mlx4_qp_state *qp_state)
920 {
921 	int err;
922 	int i;
923 	enum mlx4_qp_state states[] = {
924 		MLX4_QP_STATE_RST,
925 		MLX4_QP_STATE_INIT,
926 		MLX4_QP_STATE_RTR,
927 		MLX4_QP_STATE_RTS
928 	};
929 
930 	for (i = 0; i < ARRAY_SIZE(states) - 1; i++) {
931 		context->flags &= cpu_to_be32(~(0xf << 28));
932 		context->flags |= cpu_to_be32(states[i + 1] << 28);
933 		if (states[i + 1] != MLX4_QP_STATE_RTR)
934 			context->params2 &= ~MLX4_QP_BIT_FPP;
935 		err = mlx4_qp_modify(dev, mtt, states[i], states[i + 1],
936 				     context, 0, 0, qp);
937 		if (err) {
938 			mlx4_err(dev, "Failed to bring QP to state: %d with error: %d\n",
939 				 states[i + 1], err);
940 			return err;
941 		}
942 
943 		*qp_state = states[i + 1];
944 	}
945 
946 	return 0;
947 }
948 EXPORT_SYMBOL_GPL(mlx4_qp_to_ready);
949 
950 u16 mlx4_qp_roce_entropy(struct mlx4_dev *dev, u32 qpn)
951 {
952 	struct mlx4_qp_context context;
953 	struct mlx4_qp qp;
954 	int err;
955 
956 	qp.qpn = qpn;
957 	err = mlx4_qp_query(dev, &qp, &context);
958 	if (!err) {
959 		u32 dest_qpn = be32_to_cpu(context.remote_qpn) & 0xffffff;
960 		u16 folded_dst = folded_qp(dest_qpn);
961 		u16 folded_src = folded_qp(qpn);
962 
963 		return (dest_qpn != qpn) ?
964 			((folded_dst ^ folded_src) | 0xC000) :
965 			folded_src | 0xC000;
966 	}
967 	return 0xdead;
968 }
969