xref: /freebsd/sys/dev/cxgbe/iw_cxgbe/mem.c (revision 3fc36ee018bb836bd1796067cf4ef8683f166ebc)
1 /*
2  * Copyright (c) 2009-2013 Chelsio, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_inet.h"
36 
37 #ifdef TCP_OFFLOAD
38 #include <linux/types.h>
39 #include <linux/kref.h>
40 #include <rdma/ib_umem.h>
41 #include <asm/atomic.h>
42 
43 #include <common/t4_msg.h>
44 #include "iw_cxgbe.h"
45 
46 #define T4_ULPTX_MIN_IO 32
47 #define C4IW_MAX_INLINE_SIZE 96
48 
49 static int mr_exceeds_hw_limits(struct c4iw_dev *dev, u64 length)
50 {
51 	return (is_t4(dev->rdev.adap) ||
52 		is_t5(dev->rdev.adap)) &&
53 		length >= 8*1024*1024*1024ULL;
54 }
55 static int
56 write_adapter_mem(struct c4iw_rdev *rdev, u32 addr, u32 len, void *data)
57 {
58 	struct adapter *sc = rdev->adap;
59 	struct ulp_mem_io *ulpmc;
60 	struct ulptx_idata *ulpsc;
61 	u8 wr_len, *to_dp, *from_dp;
62 	int copy_len, num_wqe, i, ret = 0;
63 	struct c4iw_wr_wait wr_wait;
64 	struct wrqe *wr;
65 	u32 cmd;
66 
67 	cmd = cpu_to_be32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
68 	if (is_t4(sc))
69 		cmd |= cpu_to_be32(F_ULP_MEMIO_ORDER);
70 	else
71 		cmd |= cpu_to_be32(F_T5_ULP_MEMIO_IMM);
72 
73 	addr &= 0x7FFFFFF;
74 	CTR3(KTR_IW_CXGBE, "%s addr 0x%x len %u", __func__, addr, len);
75 	num_wqe = DIV_ROUND_UP(len, C4IW_MAX_INLINE_SIZE);
76 	c4iw_init_wr_wait(&wr_wait);
77 	for (i = 0; i < num_wqe; i++) {
78 
79 		copy_len = min(len, C4IW_MAX_INLINE_SIZE);
80 		wr_len = roundup(sizeof *ulpmc + sizeof *ulpsc +
81 				 roundup(copy_len, T4_ULPTX_MIN_IO), 16);
82 
83 		wr = alloc_wrqe(wr_len, &sc->sge.mgmtq);
84 		if (wr == NULL)
85 			return (0);
86 		ulpmc = wrtod(wr);
87 
88 		memset(ulpmc, 0, wr_len);
89 		INIT_ULPTX_WR(ulpmc, wr_len, 0, 0);
90 
91 		if (i == (num_wqe-1)) {
92 			ulpmc->wr.wr_hi = cpu_to_be32(V_FW_WR_OP(FW_ULPTX_WR) |
93 						    F_FW_WR_COMPL);
94 			ulpmc->wr.wr_lo = (__force __be64)(unsigned long) &wr_wait;
95 		} else
96 			ulpmc->wr.wr_hi = cpu_to_be32(V_FW_WR_OP(FW_ULPTX_WR));
97 		ulpmc->wr.wr_mid = cpu_to_be32(
98 				       V_FW_WR_LEN16(DIV_ROUND_UP(wr_len, 16)));
99 
100 		ulpmc->cmd = cmd;
101 		ulpmc->dlen = cpu_to_be32(V_ULP_MEMIO_DATA_LEN(
102 		    DIV_ROUND_UP(copy_len, T4_ULPTX_MIN_IO)));
103 		ulpmc->len16 = cpu_to_be32(DIV_ROUND_UP(wr_len-sizeof(ulpmc->wr),
104 						      16));
105 		ulpmc->lock_addr = cpu_to_be32(V_ULP_MEMIO_ADDR(addr + i * 3));
106 
107 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
108 		ulpsc->cmd_more = cpu_to_be32(V_ULPTX_CMD(ULP_TX_SC_IMM));
109 		ulpsc->len = cpu_to_be32(roundup(copy_len, T4_ULPTX_MIN_IO));
110 
111 		to_dp = (u8 *)(ulpsc + 1);
112 		from_dp = (u8 *)data + i * C4IW_MAX_INLINE_SIZE;
113 		if (data)
114 			memcpy(to_dp, from_dp, copy_len);
115 		else
116 			memset(to_dp, 0, copy_len);
117 		if (copy_len % T4_ULPTX_MIN_IO)
118 			memset(to_dp + copy_len, 0, T4_ULPTX_MIN_IO -
119 			       (copy_len % T4_ULPTX_MIN_IO));
120 		t4_wrq_tx(sc, wr);
121 		len -= C4IW_MAX_INLINE_SIZE;
122 	}
123 
124 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
125 	return ret;
126 }
127 
128 /*
129  * Build and write a TPT entry.
130  * IN: stag key, pdid, perm, bind_enabled, zbva, to, len, page_size,
131  *     pbl_size and pbl_addr
132  * OUT: stag index
133  */
134 static int write_tpt_entry(struct c4iw_rdev *rdev, u32 reset_tpt_entry,
135 			   u32 *stag, u8 stag_state, u32 pdid,
136 			   enum fw_ri_stag_type type, enum fw_ri_mem_perms perm,
137 			   int bind_enabled, u32 zbva, u64 to,
138 			   u64 len, u8 page_size, u32 pbl_size, u32 pbl_addr)
139 {
140 	int err;
141 	struct fw_ri_tpte tpt;
142 	u32 stag_idx;
143 	static atomic_t key;
144 
145 	if (c4iw_fatal_error(rdev))
146 		return -EIO;
147 
148 	stag_state = stag_state > 0;
149 	stag_idx = (*stag) >> 8;
150 
151 	if ((!reset_tpt_entry) && (*stag == T4_STAG_UNSET)) {
152 		stag_idx = c4iw_get_resource(&rdev->resource.tpt_table);
153 		if (!stag_idx) {
154 			mutex_lock(&rdev->stats.lock);
155 			rdev->stats.stag.fail++;
156 			mutex_unlock(&rdev->stats.lock);
157 			return -ENOMEM;
158 		}
159 		mutex_lock(&rdev->stats.lock);
160 		rdev->stats.stag.cur += 32;
161 		if (rdev->stats.stag.cur > rdev->stats.stag.max)
162 			rdev->stats.stag.max = rdev->stats.stag.cur;
163 		mutex_unlock(&rdev->stats.lock);
164 		*stag = (stag_idx << 8) | (atomic_inc_return(&key) & 0xff);
165 	}
166 	CTR5(KTR_IW_CXGBE,
167 	    "%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x",
168 	    __func__, stag_state, type, pdid, stag_idx);
169 
170 	/* write TPT entry */
171 	if (reset_tpt_entry)
172 		memset(&tpt, 0, sizeof(tpt));
173 	else {
174 		tpt.valid_to_pdid = cpu_to_be32(F_FW_RI_TPTE_VALID |
175 			V_FW_RI_TPTE_STAGKEY((*stag & M_FW_RI_TPTE_STAGKEY)) |
176 			V_FW_RI_TPTE_STAGSTATE(stag_state) |
177 			V_FW_RI_TPTE_STAGTYPE(type) | V_FW_RI_TPTE_PDID(pdid));
178 		tpt.locread_to_qpid = cpu_to_be32(V_FW_RI_TPTE_PERM(perm) |
179 			(bind_enabled ? F_FW_RI_TPTE_MWBINDEN : 0) |
180 			V_FW_RI_TPTE_ADDRTYPE((zbva ? FW_RI_ZERO_BASED_TO :
181 						      FW_RI_VA_BASED_TO))|
182 			V_FW_RI_TPTE_PS(page_size));
183 		tpt.nosnoop_pbladdr = !pbl_size ? 0 : cpu_to_be32(
184 			V_FW_RI_TPTE_PBLADDR(PBL_OFF(rdev, pbl_addr)>>3));
185 		tpt.len_lo = cpu_to_be32((u32)(len & 0xffffffffUL));
186 		tpt.va_hi = cpu_to_be32((u32)(to >> 32));
187 		tpt.va_lo_fbo = cpu_to_be32((u32)(to & 0xffffffffUL));
188 		tpt.dca_mwbcnt_pstag = cpu_to_be32(0);
189 		tpt.len_hi = cpu_to_be32((u32)(len >> 32));
190 	}
191 	err = write_adapter_mem(rdev, stag_idx +
192 				(rdev->adap->vres.stag.start >> 5),
193 				sizeof(tpt), &tpt);
194 
195 	if (reset_tpt_entry) {
196 		c4iw_put_resource(&rdev->resource.tpt_table, stag_idx);
197 		mutex_lock(&rdev->stats.lock);
198 		rdev->stats.stag.cur -= 32;
199 		mutex_unlock(&rdev->stats.lock);
200 	}
201 	return err;
202 }
203 
204 static int write_pbl(struct c4iw_rdev *rdev, __be64 *pbl,
205 		     u32 pbl_addr, u32 pbl_size)
206 {
207 	int err;
208 
209 	CTR4(KTR_IW_CXGBE, "%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d",
210 	     __func__, pbl_addr, rdev->adap->vres.pbl.start, pbl_size);
211 
212 	err = write_adapter_mem(rdev, pbl_addr >> 5, pbl_size << 3, pbl);
213 	return err;
214 }
215 
216 static int dereg_mem(struct c4iw_rdev *rdev, u32 stag, u32 pbl_size,
217 		     u32 pbl_addr)
218 {
219 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0,
220 			       pbl_size, pbl_addr);
221 }
222 
223 static int allocate_window(struct c4iw_rdev *rdev, u32 * stag, u32 pdid)
224 {
225 	*stag = T4_STAG_UNSET;
226 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_MW, 0, 0, 0,
227 			       0UL, 0, 0, 0, 0);
228 }
229 
230 static int deallocate_window(struct c4iw_rdev *rdev, u32 stag)
231 {
232 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0, 0,
233 			       0);
234 }
235 
236 static int allocate_stag(struct c4iw_rdev *rdev, u32 *stag, u32 pdid,
237 			 u32 pbl_size, u32 pbl_addr)
238 {
239 	*stag = T4_STAG_UNSET;
240 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_NSMR, 0, 0, 0,
241 			       0UL, 0, 0, pbl_size, pbl_addr);
242 }
243 
244 static int finish_mem_reg(struct c4iw_mr *mhp, u32 stag)
245 {
246 	u32 mmid;
247 
248 	mhp->attr.state = 1;
249 	mhp->attr.stag = stag;
250 	mmid = stag >> 8;
251 	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
252 	CTR3(KTR_IW_CXGBE, "%s mmid 0x%x mhp %p", __func__, mmid, mhp);
253 	return insert_handle(mhp->rhp, &mhp->rhp->mmidr, mhp, mmid);
254 }
255 
256 static int register_mem(struct c4iw_dev *rhp, struct c4iw_pd *php,
257 		      struct c4iw_mr *mhp, int shift)
258 {
259 	u32 stag = T4_STAG_UNSET;
260 	int ret;
261 
262 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, mhp->attr.pdid,
263 			      FW_RI_STAG_NSMR, mhp->attr.len ? mhp->attr.perms : 0,
264 			      mhp->attr.mw_bind_enable, mhp->attr.zbva,
265 			      mhp->attr.va_fbo, mhp->attr.len ? mhp->attr.len : -1, shift - 12,
266 			      mhp->attr.pbl_size, mhp->attr.pbl_addr);
267 	if (ret)
268 		return ret;
269 
270 	ret = finish_mem_reg(mhp, stag);
271 	if (ret)
272 		dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
273 		       mhp->attr.pbl_addr);
274 	return ret;
275 }
276 
277 static int reregister_mem(struct c4iw_dev *rhp, struct c4iw_pd *php,
278 			  struct c4iw_mr *mhp, int shift, int npages)
279 {
280 	u32 stag;
281 	int ret;
282 
283 	if (npages > mhp->attr.pbl_size)
284 		return -ENOMEM;
285 
286 	stag = mhp->attr.stag;
287 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, mhp->attr.pdid,
288 			      FW_RI_STAG_NSMR, mhp->attr.perms,
289 			      mhp->attr.mw_bind_enable, mhp->attr.zbva,
290 			      mhp->attr.va_fbo, mhp->attr.len, shift - 12,
291 			      mhp->attr.pbl_size, mhp->attr.pbl_addr);
292 	if (ret)
293 		return ret;
294 
295 	ret = finish_mem_reg(mhp, stag);
296 	if (ret)
297 		dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
298 		       mhp->attr.pbl_addr);
299 
300 	return ret;
301 }
302 
303 static int alloc_pbl(struct c4iw_mr *mhp, int npages)
304 {
305 	mhp->attr.pbl_addr = c4iw_pblpool_alloc(&mhp->rhp->rdev,
306 						    npages << 3);
307 
308 	if (!mhp->attr.pbl_addr)
309 		return -ENOMEM;
310 
311 	mhp->attr.pbl_size = npages;
312 
313 	return 0;
314 }
315 
316 static int build_phys_page_list(struct ib_phys_buf *buffer_list,
317 				int num_phys_buf, u64 *iova_start,
318 				u64 *total_size, int *npages,
319 				int *shift, __be64 **page_list)
320 {
321 	u64 mask;
322 	int i, j, n;
323 
324 	mask = 0;
325 	*total_size = 0;
326 	for (i = 0; i < num_phys_buf; ++i) {
327 		if (i != 0 && buffer_list[i].addr & ~PAGE_MASK)
328 			return -EINVAL;
329 		if (i != 0 && i != num_phys_buf - 1 &&
330 		    (buffer_list[i].size & ~PAGE_MASK))
331 			return -EINVAL;
332 		*total_size += buffer_list[i].size;
333 		if (i > 0)
334 			mask |= buffer_list[i].addr;
335 		else
336 			mask |= buffer_list[i].addr & PAGE_MASK;
337 		if (i != num_phys_buf - 1)
338 			mask |= buffer_list[i].addr + buffer_list[i].size;
339 		else
340 			mask |= (buffer_list[i].addr + buffer_list[i].size +
341 				PAGE_SIZE - 1) & PAGE_MASK;
342 	}
343 
344 	if (*total_size > 0xFFFFFFFFULL)
345 		return -ENOMEM;
346 
347 	/* Find largest page shift we can use to cover buffers */
348 	for (*shift = PAGE_SHIFT; *shift < 27; ++(*shift))
349 		if ((1ULL << *shift) & mask)
350 			break;
351 
352 	buffer_list[0].size += buffer_list[0].addr & ((1ULL << *shift) - 1);
353 	buffer_list[0].addr &= ~0ull << *shift;
354 
355 	*npages = 0;
356 	for (i = 0; i < num_phys_buf; ++i)
357 		*npages += (buffer_list[i].size +
358 			(1ULL << *shift) - 1) >> *shift;
359 
360 	if (!*npages)
361 		return -EINVAL;
362 
363 	*page_list = kmalloc(sizeof(u64) * *npages, GFP_KERNEL);
364 	if (!*page_list)
365 		return -ENOMEM;
366 
367 	n = 0;
368 	for (i = 0; i < num_phys_buf; ++i)
369 		for (j = 0;
370 		     j < (buffer_list[i].size + (1ULL << *shift) - 1) >> *shift;
371 		     ++j)
372 			(*page_list)[n++] = cpu_to_be64(buffer_list[i].addr +
373 			    ((u64) j << *shift));
374 
375 	CTR6(KTR_IW_CXGBE,
376 	    "%s va 0x%llx mask 0x%llx shift %d len %lld pbl_size %d", __func__,
377 	    (unsigned long long)*iova_start, (unsigned long long)mask, *shift,
378 	    (unsigned long long)*total_size, *npages);
379 
380 	return 0;
381 
382 }
383 
384 int c4iw_reregister_phys_mem(struct ib_mr *mr, int mr_rereg_mask,
385 			     struct ib_pd *pd, struct ib_phys_buf *buffer_list,
386 			     int num_phys_buf, int acc, u64 *iova_start)
387 {
388 
389 	struct c4iw_mr mh, *mhp;
390 	struct c4iw_pd *php;
391 	struct c4iw_dev *rhp;
392 	__be64 *page_list = NULL;
393 	int shift = 0;
394 	u64 total_size = 0;
395 	int npages = 0;
396 	int ret;
397 
398 	CTR3(KTR_IW_CXGBE, "%s ib_mr %p ib_pd %p", __func__, mr, pd);
399 
400 	/* There can be no memory windows */
401 	if (atomic_read(&mr->usecnt))
402 		return -EINVAL;
403 
404 	mhp = to_c4iw_mr(mr);
405 	rhp = mhp->rhp;
406 	php = to_c4iw_pd(mr->pd);
407 
408 	/* make sure we are on the same adapter */
409 	if (rhp != php->rhp)
410 		return -EINVAL;
411 
412 	memcpy(&mh, mhp, sizeof *mhp);
413 
414 	if (mr_rereg_mask & IB_MR_REREG_PD)
415 		php = to_c4iw_pd(pd);
416 	if (mr_rereg_mask & IB_MR_REREG_ACCESS) {
417 		mh.attr.perms = c4iw_ib_to_tpt_access(acc);
418 		mh.attr.mw_bind_enable = (acc & IB_ACCESS_MW_BIND) ==
419 					 IB_ACCESS_MW_BIND;
420 	}
421 	if (mr_rereg_mask & IB_MR_REREG_TRANS) {
422 		ret = build_phys_page_list(buffer_list, num_phys_buf,
423 						iova_start,
424 						&total_size, &npages,
425 						&shift, &page_list);
426 		if (ret)
427 			return ret;
428 	}
429 	if (mr_exceeds_hw_limits(rhp, total_size)) {
430 		kfree(page_list);
431 		return -EINVAL;
432 	}
433 	ret = reregister_mem(rhp, php, &mh, shift, npages);
434 	kfree(page_list);
435 	if (ret)
436 		return ret;
437 	if (mr_rereg_mask & IB_MR_REREG_PD)
438 		mhp->attr.pdid = php->pdid;
439 	if (mr_rereg_mask & IB_MR_REREG_ACCESS)
440 		mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
441 	if (mr_rereg_mask & IB_MR_REREG_TRANS) {
442 		mhp->attr.zbva = 0;
443 		mhp->attr.va_fbo = *iova_start;
444 		mhp->attr.page_size = shift - 12;
445 		mhp->attr.len = (u32) total_size;
446 		mhp->attr.pbl_size = npages;
447 	}
448 
449 	return 0;
450 }
451 
452 struct ib_mr *c4iw_register_phys_mem(struct ib_pd *pd,
453 				     struct ib_phys_buf *buffer_list,
454 				     int num_phys_buf, int acc, u64 *iova_start)
455 {
456 	__be64 *page_list;
457 	int shift;
458 	u64 total_size;
459 	int npages;
460 	struct c4iw_dev *rhp;
461 	struct c4iw_pd *php;
462 	struct c4iw_mr *mhp;
463 	int ret;
464 
465 	CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd);
466 	php = to_c4iw_pd(pd);
467 	rhp = php->rhp;
468 
469 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
470 	if (!mhp)
471 		return ERR_PTR(-ENOMEM);
472 
473 	mhp->rhp = rhp;
474 
475 	/* First check that we have enough alignment */
476 	if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK)) {
477 		ret = -EINVAL;
478 		goto err;
479 	}
480 
481 	if (num_phys_buf > 1 &&
482 	    ((buffer_list[0].addr + buffer_list[0].size) & ~PAGE_MASK)) {
483 		ret = -EINVAL;
484 		goto err;
485 	}
486 
487 	ret = build_phys_page_list(buffer_list, num_phys_buf, iova_start,
488 					&total_size, &npages, &shift,
489 					&page_list);
490 	if (ret)
491 		goto err;
492 
493 	if (mr_exceeds_hw_limits(rhp, total_size)) {
494 		kfree(page_list);
495 		ret = -EINVAL;
496 		goto err;
497 	}
498 	ret = alloc_pbl(mhp, npages);
499 	if (ret) {
500 		kfree(page_list);
501 		goto err;
502 	}
503 
504 	ret = write_pbl(&mhp->rhp->rdev, page_list, mhp->attr.pbl_addr,
505 			     npages);
506 	kfree(page_list);
507 	if (ret)
508 		goto err_pbl;
509 
510 	mhp->attr.pdid = php->pdid;
511 	mhp->attr.zbva = 0;
512 
513 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
514 	mhp->attr.va_fbo = *iova_start;
515 	mhp->attr.page_size = shift - 12;
516 
517 	mhp->attr.len = (u32) total_size;
518 	mhp->attr.pbl_size = npages;
519 	ret = register_mem(rhp, php, mhp, shift);
520 	if (ret)
521 		goto err_pbl;
522 
523 	return &mhp->ibmr;
524 
525 err_pbl:
526 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
527 			      mhp->attr.pbl_size << 3);
528 
529 err:
530 	kfree(mhp);
531 	return ERR_PTR(ret);
532 
533 }
534 
535 struct ib_mr *c4iw_get_dma_mr(struct ib_pd *pd, int acc)
536 {
537 	struct c4iw_dev *rhp;
538 	struct c4iw_pd *php;
539 	struct c4iw_mr *mhp;
540 	int ret;
541 	u32 stag = T4_STAG_UNSET;
542 
543 	CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd);
544 	php = to_c4iw_pd(pd);
545 	rhp = php->rhp;
546 
547 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
548 	if (!mhp)
549 		return ERR_PTR(-ENOMEM);
550 
551 	mhp->rhp = rhp;
552 	mhp->attr.pdid = php->pdid;
553 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
554 	mhp->attr.mw_bind_enable = (acc&IB_ACCESS_MW_BIND) == IB_ACCESS_MW_BIND;
555 	mhp->attr.zbva = 0;
556 	mhp->attr.va_fbo = 0;
557 	mhp->attr.page_size = 0;
558 	mhp->attr.len = ~0UL;
559 	mhp->attr.pbl_size = 0;
560 
561 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, php->pdid,
562 			      FW_RI_STAG_NSMR, mhp->attr.perms,
563 			      mhp->attr.mw_bind_enable, 0, 0, ~0UL, 0, 0, 0);
564 	if (ret)
565 		goto err1;
566 
567 	ret = finish_mem_reg(mhp, stag);
568 	if (ret)
569 		goto err2;
570 	return &mhp->ibmr;
571 err2:
572 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
573 		  mhp->attr.pbl_addr);
574 err1:
575 	kfree(mhp);
576 	return ERR_PTR(ret);
577 }
578 
579 struct ib_mr *c4iw_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
580     u64 virt, int acc, struct ib_udata *udata, int mr_id)
581 {
582 	__be64 *pages;
583 	int shift, n, len;
584 	int i, k, entry;
585 	int err = 0;
586 	struct scatterlist *sg;
587 	struct c4iw_dev *rhp;
588 	struct c4iw_pd *php;
589 	struct c4iw_mr *mhp;
590 
591 	CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd);
592 
593 	if (length == ~0ULL)
594 		return ERR_PTR(-EINVAL);
595 
596 	if ((length + start) < start)
597 		return ERR_PTR(-EINVAL);
598 
599 	php = to_c4iw_pd(pd);
600 	rhp = php->rhp;
601 
602 	if (mr_exceeds_hw_limits(rhp, length))
603 		return ERR_PTR(-EINVAL);
604 
605 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
606 	if (!mhp)
607 		return ERR_PTR(-ENOMEM);
608 
609 	mhp->rhp = rhp;
610 
611 	mhp->umem = ib_umem_get(pd->uobject->context, start, length, acc, 0);
612 	if (IS_ERR(mhp->umem)) {
613 		err = PTR_ERR(mhp->umem);
614 		kfree(mhp);
615 		return ERR_PTR(err);
616 	}
617 
618 	shift = ffs(mhp->umem->page_size) - 1;
619 
620 	n = mhp->umem->nmap;
621 	err = alloc_pbl(mhp, n);
622 	if (err)
623 		goto err;
624 
625 	pages = (__be64 *) __get_free_page(GFP_KERNEL);
626 	if (!pages) {
627 		err = -ENOMEM;
628 		goto err_pbl;
629 	}
630 
631 	i = n = 0;
632 	for_each_sg(mhp->umem->sg_head.sgl, sg, mhp->umem->nmap, entry) {
633 		len = sg_dma_len(sg) >> shift;
634 		for (k = 0; k < len; ++k) {
635 			pages[i++] = cpu_to_be64(sg_dma_address(sg) +
636 					mhp->umem->page_size * k);
637 			if (i == PAGE_SIZE / sizeof *pages) {
638 				err = write_pbl(&mhp->rhp->rdev,
639 						pages,
640 						mhp->attr.pbl_addr + (n << 3), i);
641 				if (err)
642 					goto pbl_done;
643 				n += i;
644 				i = 0;
645 
646 			}
647 		}
648 	}
649 
650 	if (i)
651 		err = write_pbl(&mhp->rhp->rdev, pages,
652 				     mhp->attr.pbl_addr + (n << 3), i);
653 
654 pbl_done:
655 	free_page((unsigned long) pages);
656 	if (err)
657 		goto err_pbl;
658 
659 	mhp->attr.pdid = php->pdid;
660 	mhp->attr.zbva = 0;
661 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
662 	mhp->attr.va_fbo = virt;
663 	mhp->attr.page_size = shift - 12;
664 	mhp->attr.len = length;
665 
666 	err = register_mem(rhp, php, mhp, shift);
667 	if (err)
668 		goto err_pbl;
669 
670 	return &mhp->ibmr;
671 
672 err_pbl:
673 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
674 			      mhp->attr.pbl_size << 3);
675 
676 err:
677 	ib_umem_release(mhp->umem);
678 	kfree(mhp);
679 	return ERR_PTR(err);
680 }
681 
682 struct ib_mw *c4iw_alloc_mw(struct ib_pd *pd, enum ib_mw_type type)
683 {
684 	struct c4iw_dev *rhp;
685 	struct c4iw_pd *php;
686 	struct c4iw_mw *mhp;
687 	u32 mmid;
688 	u32 stag = 0;
689 	int ret;
690 
691 	php = to_c4iw_pd(pd);
692 	rhp = php->rhp;
693 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
694 	if (!mhp)
695 		return ERR_PTR(-ENOMEM);
696 	ret = allocate_window(&rhp->rdev, &stag, php->pdid);
697 	if (ret) {
698 		kfree(mhp);
699 		return ERR_PTR(ret);
700 	}
701 	mhp->rhp = rhp;
702 	mhp->attr.pdid = php->pdid;
703 	mhp->attr.type = FW_RI_STAG_MW;
704 	mhp->attr.stag = stag;
705 	mmid = (stag) >> 8;
706 	mhp->ibmw.rkey = stag;
707 	if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
708 		deallocate_window(&rhp->rdev, mhp->attr.stag);
709 		kfree(mhp);
710 		return ERR_PTR(-ENOMEM);
711 	}
712 	CTR4(KTR_IW_CXGBE, "%s mmid 0x%x mhp %p stag 0x%x", __func__, mmid, mhp,
713 	    stag);
714 	return &(mhp->ibmw);
715 }
716 
717 int c4iw_dealloc_mw(struct ib_mw *mw)
718 {
719 	struct c4iw_dev *rhp;
720 	struct c4iw_mw *mhp;
721 	u32 mmid;
722 
723 	mhp = to_c4iw_mw(mw);
724 	rhp = mhp->rhp;
725 	mmid = (mw->rkey) >> 8;
726 	remove_handle(rhp, &rhp->mmidr, mmid);
727 	deallocate_window(&rhp->rdev, mhp->attr.stag);
728 	kfree(mhp);
729 	CTR4(KTR_IW_CXGBE, "%s ib_mw %p mmid 0x%x ptr %p", __func__, mw, mmid,
730 	    mhp);
731 	return 0;
732 }
733 
734 struct ib_mr *c4iw_alloc_fast_reg_mr(struct ib_pd *pd, int pbl_depth)
735 {
736 	struct c4iw_dev *rhp;
737 	struct c4iw_pd *php;
738 	struct c4iw_mr *mhp;
739 	u32 mmid;
740 	u32 stag = 0;
741 	int ret = 0;
742 
743 	php = to_c4iw_pd(pd);
744 	rhp = php->rhp;
745 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
746 	if (!mhp) {
747 		ret = -ENOMEM;
748 		goto err;
749 	}
750 
751 	mhp->rhp = rhp;
752 	ret = alloc_pbl(mhp, pbl_depth);
753 	if (ret)
754 		goto err1;
755 	mhp->attr.pbl_size = pbl_depth;
756 	ret = allocate_stag(&rhp->rdev, &stag, php->pdid,
757 				 mhp->attr.pbl_size, mhp->attr.pbl_addr);
758 	if (ret)
759 		goto err2;
760 	mhp->attr.pdid = php->pdid;
761 	mhp->attr.type = FW_RI_STAG_NSMR;
762 	mhp->attr.stag = stag;
763 	mhp->attr.state = 1;
764 	mmid = (stag) >> 8;
765 	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
766 	if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
767 		ret = -ENOMEM;
768 		goto err3;
769 	}
770 
771 	CTR4(KTR_IW_CXGBE, "%s mmid 0x%x mhp %p stag 0x%x", __func__, mmid, mhp,
772 	    stag);
773 	return &(mhp->ibmr);
774 err3:
775 	dereg_mem(&rhp->rdev, stag, mhp->attr.pbl_size,
776 		       mhp->attr.pbl_addr);
777 err2:
778 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
779 			      mhp->attr.pbl_size << 3);
780 err1:
781 	kfree(mhp);
782 err:
783 	return ERR_PTR(ret);
784 }
785 
786 struct ib_fast_reg_page_list *c4iw_alloc_fastreg_pbl(struct ib_device *device,
787 						     int page_list_len)
788 {
789 	struct c4iw_fr_page_list *c4pl;
790 	struct c4iw_dev *dev = to_c4iw_dev(device);
791 	bus_addr_t dma_addr;
792 	int size = sizeof *c4pl + page_list_len * sizeof(u64);
793 
794 	c4pl = contigmalloc(size,
795             M_DEVBUF, M_NOWAIT, 0ul, ~0ul, 4096, 0);
796         if (c4pl)
797                 dma_addr = vtophys(c4pl);
798         else
799                 return ERR_PTR(-ENOMEM);
800 
801 	pci_unmap_addr_set(c4pl, mapping, dma_addr);
802 	c4pl->dma_addr = dma_addr;
803 	c4pl->dev = dev;
804 	c4pl->size = size;
805 	c4pl->ibpl.page_list = (u64 *)(c4pl + 1);
806 	c4pl->ibpl.max_page_list_len = page_list_len;
807 
808 	return &c4pl->ibpl;
809 }
810 
811 void c4iw_free_fastreg_pbl(struct ib_fast_reg_page_list *ibpl)
812 {
813 	struct c4iw_fr_page_list *c4pl = to_c4iw_fr_page_list(ibpl);
814 	contigfree(c4pl, c4pl->size, M_DEVBUF);
815 }
816 
817 int c4iw_dereg_mr(struct ib_mr *ib_mr)
818 {
819 	struct c4iw_dev *rhp;
820 	struct c4iw_mr *mhp;
821 	u32 mmid;
822 
823 	CTR2(KTR_IW_CXGBE, "%s ib_mr %p", __func__, ib_mr);
824 	/* There can be no memory windows */
825 	if (atomic_read(&ib_mr->usecnt))
826 		return -EINVAL;
827 
828 	mhp = to_c4iw_mr(ib_mr);
829 	rhp = mhp->rhp;
830 	mmid = mhp->attr.stag >> 8;
831 	remove_handle(rhp, &rhp->mmidr, mmid);
832 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
833 		       mhp->attr.pbl_addr);
834 	if (mhp->attr.pbl_size)
835 		c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
836 				  mhp->attr.pbl_size << 3);
837 	if (mhp->kva)
838 		kfree((void *) (unsigned long) mhp->kva);
839 	if (mhp->umem)
840 		ib_umem_release(mhp->umem);
841 	CTR3(KTR_IW_CXGBE, "%s mmid 0x%x ptr %p", __func__, mmid, mhp);
842 	kfree(mhp);
843 	return 0;
844 }
845 #endif
846