xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/mem.c (revision 930beb5a)
1 /*
2  * Copyright (c) 2009-2010 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 
33 #include <linux/module.h>
34 #include <linux/moduleparam.h>
35 #include <rdma/ib_umem.h>
36 #include <linux/atomic.h>
37 
38 #include "iw_cxgb4.h"
39 
40 int use_dsgl = 1;
41 module_param(use_dsgl, int, 0644);
42 MODULE_PARM_DESC(use_dsgl, "Use DSGL for PBL/FastReg (default=1)");
43 
44 #define T4_ULPTX_MIN_IO 32
45 #define C4IW_MAX_INLINE_SIZE 96
46 #define T4_ULPTX_MAX_DMA 1024
47 #define C4IW_INLINE_THRESHOLD 128
48 
49 static int inline_threshold = C4IW_INLINE_THRESHOLD;
50 module_param(inline_threshold, int, 0644);
51 MODULE_PARM_DESC(inline_threshold, "inline vs dsgl threshold (default=128)");
52 
53 static int _c4iw_write_mem_dma_aligned(struct c4iw_rdev *rdev, u32 addr,
54 				       u32 len, dma_addr_t data, int wait)
55 {
56 	struct sk_buff *skb;
57 	struct ulp_mem_io *req;
58 	struct ulptx_sgl *sgl;
59 	u8 wr_len;
60 	int ret = 0;
61 	struct c4iw_wr_wait wr_wait;
62 
63 	addr &= 0x7FFFFFF;
64 
65 	if (wait)
66 		c4iw_init_wr_wait(&wr_wait);
67 	wr_len = roundup(sizeof(*req) + sizeof(*sgl), 16);
68 
69 	skb = alloc_skb(wr_len, GFP_KERNEL | __GFP_NOFAIL);
70 	if (!skb)
71 		return -ENOMEM;
72 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
73 
74 	req = (struct ulp_mem_io *)__skb_put(skb, wr_len);
75 	memset(req, 0, wr_len);
76 	INIT_ULPTX_WR(req, wr_len, 0, 0);
77 	req->wr.wr_hi = cpu_to_be32(FW_WR_OP(FW_ULPTX_WR) |
78 			(wait ? FW_WR_COMPL(1) : 0));
79 	req->wr.wr_lo = wait ? (__force __be64)&wr_wait : 0;
80 	req->wr.wr_mid = cpu_to_be32(FW_WR_LEN16(DIV_ROUND_UP(wr_len, 16)));
81 	req->cmd = cpu_to_be32(ULPTX_CMD(ULP_TX_MEM_WRITE));
82 	req->cmd |= cpu_to_be32(V_T5_ULP_MEMIO_ORDER(1));
83 	req->dlen = cpu_to_be32(ULP_MEMIO_DATA_LEN(len>>5));
84 	req->len16 = cpu_to_be32(DIV_ROUND_UP(wr_len-sizeof(req->wr), 16));
85 	req->lock_addr = cpu_to_be32(ULP_MEMIO_ADDR(addr));
86 
87 	sgl = (struct ulptx_sgl *)(req + 1);
88 	sgl->cmd_nsge = cpu_to_be32(ULPTX_CMD(ULP_TX_SC_DSGL) |
89 				    ULPTX_NSGE(1));
90 	sgl->len0 = cpu_to_be32(len);
91 	sgl->addr0 = cpu_to_be64(data);
92 
93 	ret = c4iw_ofld_send(rdev, skb);
94 	if (ret)
95 		return ret;
96 	if (wait)
97 		ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
98 	return ret;
99 }
100 
101 static int _c4iw_write_mem_inline(struct c4iw_rdev *rdev, u32 addr, u32 len,
102 				  void *data)
103 {
104 	struct sk_buff *skb;
105 	struct ulp_mem_io *req;
106 	struct ulptx_idata *sc;
107 	u8 wr_len, *to_dp, *from_dp;
108 	int copy_len, num_wqe, i, ret = 0;
109 	struct c4iw_wr_wait wr_wait;
110 	__be32 cmd = cpu_to_be32(ULPTX_CMD(ULP_TX_MEM_WRITE));
111 
112 	if (is_t4(rdev->lldi.adapter_type))
113 		cmd |= cpu_to_be32(ULP_MEMIO_ORDER(1));
114 	else
115 		cmd |= cpu_to_be32(V_T5_ULP_MEMIO_IMM(1));
116 
117 	addr &= 0x7FFFFFF;
118 	PDBG("%s addr 0x%x len %u\n", __func__, addr, len);
119 	num_wqe = DIV_ROUND_UP(len, C4IW_MAX_INLINE_SIZE);
120 	c4iw_init_wr_wait(&wr_wait);
121 	for (i = 0; i < num_wqe; i++) {
122 
123 		copy_len = len > C4IW_MAX_INLINE_SIZE ? C4IW_MAX_INLINE_SIZE :
124 			   len;
125 		wr_len = roundup(sizeof *req + sizeof *sc +
126 				 roundup(copy_len, T4_ULPTX_MIN_IO), 16);
127 
128 		skb = alloc_skb(wr_len, GFP_KERNEL);
129 		if (!skb)
130 			return -ENOMEM;
131 		set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
132 
133 		req = (struct ulp_mem_io *)__skb_put(skb, wr_len);
134 		memset(req, 0, wr_len);
135 		INIT_ULPTX_WR(req, wr_len, 0, 0);
136 
137 		if (i == (num_wqe-1)) {
138 			req->wr.wr_hi = cpu_to_be32(FW_WR_OP(FW_ULPTX_WR) |
139 						    FW_WR_COMPL(1));
140 			req->wr.wr_lo = (__force __be64)(unsigned long) &wr_wait;
141 		} else
142 			req->wr.wr_hi = cpu_to_be32(FW_WR_OP(FW_ULPTX_WR));
143 		req->wr.wr_mid = cpu_to_be32(
144 				       FW_WR_LEN16(DIV_ROUND_UP(wr_len, 16)));
145 
146 		req->cmd = cmd;
147 		req->dlen = cpu_to_be32(ULP_MEMIO_DATA_LEN(
148 				DIV_ROUND_UP(copy_len, T4_ULPTX_MIN_IO)));
149 		req->len16 = cpu_to_be32(DIV_ROUND_UP(wr_len-sizeof(req->wr),
150 						      16));
151 		req->lock_addr = cpu_to_be32(ULP_MEMIO_ADDR(addr + i * 3));
152 
153 		sc = (struct ulptx_idata *)(req + 1);
154 		sc->cmd_more = cpu_to_be32(ULPTX_CMD(ULP_TX_SC_IMM));
155 		sc->len = cpu_to_be32(roundup(copy_len, T4_ULPTX_MIN_IO));
156 
157 		to_dp = (u8 *)(sc + 1);
158 		from_dp = (u8 *)data + i * C4IW_MAX_INLINE_SIZE;
159 		if (data)
160 			memcpy(to_dp, from_dp, copy_len);
161 		else
162 			memset(to_dp, 0, copy_len);
163 		if (copy_len % T4_ULPTX_MIN_IO)
164 			memset(to_dp + copy_len, 0, T4_ULPTX_MIN_IO -
165 			       (copy_len % T4_ULPTX_MIN_IO));
166 		ret = c4iw_ofld_send(rdev, skb);
167 		if (ret)
168 			return ret;
169 		len -= C4IW_MAX_INLINE_SIZE;
170 	}
171 
172 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
173 	return ret;
174 }
175 
176 static int _c4iw_write_mem_dma(struct c4iw_rdev *rdev, u32 addr, u32 len, void *data)
177 {
178 	u32 remain = len;
179 	u32 dmalen;
180 	int ret = 0;
181 	dma_addr_t daddr;
182 	dma_addr_t save;
183 
184 	daddr = dma_map_single(&rdev->lldi.pdev->dev, data, len, DMA_TO_DEVICE);
185 	if (dma_mapping_error(&rdev->lldi.pdev->dev, daddr))
186 		return -1;
187 	save = daddr;
188 
189 	while (remain > inline_threshold) {
190 		if (remain < T4_ULPTX_MAX_DMA) {
191 			if (remain & ~T4_ULPTX_MIN_IO)
192 				dmalen = remain & ~(T4_ULPTX_MIN_IO-1);
193 			else
194 				dmalen = remain;
195 		} else
196 			dmalen = T4_ULPTX_MAX_DMA;
197 		remain -= dmalen;
198 		ret = _c4iw_write_mem_dma_aligned(rdev, addr, dmalen, daddr,
199 						 !remain);
200 		if (ret)
201 			goto out;
202 		addr += dmalen >> 5;
203 		data += dmalen;
204 		daddr += dmalen;
205 	}
206 	if (remain)
207 		ret = _c4iw_write_mem_inline(rdev, addr, remain, data);
208 out:
209 	dma_unmap_single(&rdev->lldi.pdev->dev, save, len, DMA_TO_DEVICE);
210 	return ret;
211 }
212 
213 /*
214  * write len bytes of data into addr (32B aligned address)
215  * If data is NULL, clear len byte of memory to zero.
216  */
217 static int write_adapter_mem(struct c4iw_rdev *rdev, u32 addr, u32 len,
218 			     void *data)
219 {
220 	if (is_t5(rdev->lldi.adapter_type) && use_dsgl) {
221 		if (len > inline_threshold) {
222 			if (_c4iw_write_mem_dma(rdev, addr, len, data)) {
223 				printk_ratelimited(KERN_WARNING
224 						   "%s: dma map"
225 						   " failure (non fatal)\n",
226 						   pci_name(rdev->lldi.pdev));
227 				return _c4iw_write_mem_inline(rdev, addr, len,
228 							      data);
229 			} else
230 				return 0;
231 		} else
232 			return _c4iw_write_mem_inline(rdev, addr, len, data);
233 	} else
234 		return _c4iw_write_mem_inline(rdev, addr, len, data);
235 }
236 
237 /*
238  * Build and write a TPT entry.
239  * IN: stag key, pdid, perm, bind_enabled, zbva, to, len, page_size,
240  *     pbl_size and pbl_addr
241  * OUT: stag index
242  */
243 static int write_tpt_entry(struct c4iw_rdev *rdev, u32 reset_tpt_entry,
244 			   u32 *stag, u8 stag_state, u32 pdid,
245 			   enum fw_ri_stag_type type, enum fw_ri_mem_perms perm,
246 			   int bind_enabled, u32 zbva, u64 to,
247 			   u64 len, u8 page_size, u32 pbl_size, u32 pbl_addr)
248 {
249 	int err;
250 	struct fw_ri_tpte tpt;
251 	u32 stag_idx;
252 	static atomic_t key;
253 
254 	if (c4iw_fatal_error(rdev))
255 		return -EIO;
256 
257 	stag_state = stag_state > 0;
258 	stag_idx = (*stag) >> 8;
259 
260 	if ((!reset_tpt_entry) && (*stag == T4_STAG_UNSET)) {
261 		stag_idx = c4iw_get_resource(&rdev->resource.tpt_table);
262 		if (!stag_idx)
263 			return -ENOMEM;
264 		mutex_lock(&rdev->stats.lock);
265 		rdev->stats.stag.cur += 32;
266 		if (rdev->stats.stag.cur > rdev->stats.stag.max)
267 			rdev->stats.stag.max = rdev->stats.stag.cur;
268 		mutex_unlock(&rdev->stats.lock);
269 		*stag = (stag_idx << 8) | (atomic_inc_return(&key) & 0xff);
270 	}
271 	PDBG("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
272 	     __func__, stag_state, type, pdid, stag_idx);
273 
274 	/* write TPT entry */
275 	if (reset_tpt_entry)
276 		memset(&tpt, 0, sizeof(tpt));
277 	else {
278 		tpt.valid_to_pdid = cpu_to_be32(F_FW_RI_TPTE_VALID |
279 			V_FW_RI_TPTE_STAGKEY((*stag & M_FW_RI_TPTE_STAGKEY)) |
280 			V_FW_RI_TPTE_STAGSTATE(stag_state) |
281 			V_FW_RI_TPTE_STAGTYPE(type) | V_FW_RI_TPTE_PDID(pdid));
282 		tpt.locread_to_qpid = cpu_to_be32(V_FW_RI_TPTE_PERM(perm) |
283 			(bind_enabled ? F_FW_RI_TPTE_MWBINDEN : 0) |
284 			V_FW_RI_TPTE_ADDRTYPE((zbva ? FW_RI_ZERO_BASED_TO :
285 						      FW_RI_VA_BASED_TO))|
286 			V_FW_RI_TPTE_PS(page_size));
287 		tpt.nosnoop_pbladdr = !pbl_size ? 0 : cpu_to_be32(
288 			V_FW_RI_TPTE_PBLADDR(PBL_OFF(rdev, pbl_addr)>>3));
289 		tpt.len_lo = cpu_to_be32((u32)(len & 0xffffffffUL));
290 		tpt.va_hi = cpu_to_be32((u32)(to >> 32));
291 		tpt.va_lo_fbo = cpu_to_be32((u32)(to & 0xffffffffUL));
292 		tpt.dca_mwbcnt_pstag = cpu_to_be32(0);
293 		tpt.len_hi = cpu_to_be32((u32)(len >> 32));
294 	}
295 	err = write_adapter_mem(rdev, stag_idx +
296 				(rdev->lldi.vr->stag.start >> 5),
297 				sizeof(tpt), &tpt);
298 
299 	if (reset_tpt_entry) {
300 		c4iw_put_resource(&rdev->resource.tpt_table, stag_idx);
301 		mutex_lock(&rdev->stats.lock);
302 		rdev->stats.stag.cur -= 32;
303 		mutex_unlock(&rdev->stats.lock);
304 	}
305 	return err;
306 }
307 
308 static int write_pbl(struct c4iw_rdev *rdev, __be64 *pbl,
309 		     u32 pbl_addr, u32 pbl_size)
310 {
311 	int err;
312 
313 	PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
314 	     __func__, pbl_addr, rdev->lldi.vr->pbl.start,
315 	     pbl_size);
316 
317 	err = write_adapter_mem(rdev, pbl_addr >> 5, pbl_size << 3, pbl);
318 	return err;
319 }
320 
321 static int dereg_mem(struct c4iw_rdev *rdev, u32 stag, u32 pbl_size,
322 		     u32 pbl_addr)
323 {
324 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0,
325 			       pbl_size, pbl_addr);
326 }
327 
328 static int allocate_window(struct c4iw_rdev *rdev, u32 * stag, u32 pdid)
329 {
330 	*stag = T4_STAG_UNSET;
331 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_MW, 0, 0, 0,
332 			       0UL, 0, 0, 0, 0);
333 }
334 
335 static int deallocate_window(struct c4iw_rdev *rdev, u32 stag)
336 {
337 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0, 0,
338 			       0);
339 }
340 
341 static int allocate_stag(struct c4iw_rdev *rdev, u32 *stag, u32 pdid,
342 			 u32 pbl_size, u32 pbl_addr)
343 {
344 	*stag = T4_STAG_UNSET;
345 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_NSMR, 0, 0, 0,
346 			       0UL, 0, 0, pbl_size, pbl_addr);
347 }
348 
349 static int finish_mem_reg(struct c4iw_mr *mhp, u32 stag)
350 {
351 	u32 mmid;
352 
353 	mhp->attr.state = 1;
354 	mhp->attr.stag = stag;
355 	mmid = stag >> 8;
356 	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
357 	PDBG("%s mmid 0x%x mhp %p\n", __func__, mmid, mhp);
358 	return insert_handle(mhp->rhp, &mhp->rhp->mmidr, mhp, mmid);
359 }
360 
361 static int register_mem(struct c4iw_dev *rhp, struct c4iw_pd *php,
362 		      struct c4iw_mr *mhp, int shift)
363 {
364 	u32 stag = T4_STAG_UNSET;
365 	int ret;
366 
367 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, mhp->attr.pdid,
368 			      FW_RI_STAG_NSMR, mhp->attr.perms,
369 			      mhp->attr.mw_bind_enable, mhp->attr.zbva,
370 			      mhp->attr.va_fbo, mhp->attr.len, shift - 12,
371 			      mhp->attr.pbl_size, mhp->attr.pbl_addr);
372 	if (ret)
373 		return ret;
374 
375 	ret = finish_mem_reg(mhp, stag);
376 	if (ret)
377 		dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
378 		       mhp->attr.pbl_addr);
379 	return ret;
380 }
381 
382 static int reregister_mem(struct c4iw_dev *rhp, struct c4iw_pd *php,
383 			  struct c4iw_mr *mhp, int shift, int npages)
384 {
385 	u32 stag;
386 	int ret;
387 
388 	if (npages > mhp->attr.pbl_size)
389 		return -ENOMEM;
390 
391 	stag = mhp->attr.stag;
392 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, mhp->attr.pdid,
393 			      FW_RI_STAG_NSMR, mhp->attr.perms,
394 			      mhp->attr.mw_bind_enable, mhp->attr.zbva,
395 			      mhp->attr.va_fbo, mhp->attr.len, shift - 12,
396 			      mhp->attr.pbl_size, mhp->attr.pbl_addr);
397 	if (ret)
398 		return ret;
399 
400 	ret = finish_mem_reg(mhp, stag);
401 	if (ret)
402 		dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
403 		       mhp->attr.pbl_addr);
404 
405 	return ret;
406 }
407 
408 static int alloc_pbl(struct c4iw_mr *mhp, int npages)
409 {
410 	mhp->attr.pbl_addr = c4iw_pblpool_alloc(&mhp->rhp->rdev,
411 						    npages << 3);
412 
413 	if (!mhp->attr.pbl_addr)
414 		return -ENOMEM;
415 
416 	mhp->attr.pbl_size = npages;
417 
418 	return 0;
419 }
420 
421 static int build_phys_page_list(struct ib_phys_buf *buffer_list,
422 				int num_phys_buf, u64 *iova_start,
423 				u64 *total_size, int *npages,
424 				int *shift, __be64 **page_list)
425 {
426 	u64 mask;
427 	int i, j, n;
428 
429 	mask = 0;
430 	*total_size = 0;
431 	for (i = 0; i < num_phys_buf; ++i) {
432 		if (i != 0 && buffer_list[i].addr & ~PAGE_MASK)
433 			return -EINVAL;
434 		if (i != 0 && i != num_phys_buf - 1 &&
435 		    (buffer_list[i].size & ~PAGE_MASK))
436 			return -EINVAL;
437 		*total_size += buffer_list[i].size;
438 		if (i > 0)
439 			mask |= buffer_list[i].addr;
440 		else
441 			mask |= buffer_list[i].addr & PAGE_MASK;
442 		if (i != num_phys_buf - 1)
443 			mask |= buffer_list[i].addr + buffer_list[i].size;
444 		else
445 			mask |= (buffer_list[i].addr + buffer_list[i].size +
446 				PAGE_SIZE - 1) & PAGE_MASK;
447 	}
448 
449 	if (*total_size > 0xFFFFFFFFULL)
450 		return -ENOMEM;
451 
452 	/* Find largest page shift we can use to cover buffers */
453 	for (*shift = PAGE_SHIFT; *shift < 27; ++(*shift))
454 		if ((1ULL << *shift) & mask)
455 			break;
456 
457 	buffer_list[0].size += buffer_list[0].addr & ((1ULL << *shift) - 1);
458 	buffer_list[0].addr &= ~0ull << *shift;
459 
460 	*npages = 0;
461 	for (i = 0; i < num_phys_buf; ++i)
462 		*npages += (buffer_list[i].size +
463 			(1ULL << *shift) - 1) >> *shift;
464 
465 	if (!*npages)
466 		return -EINVAL;
467 
468 	*page_list = kmalloc(sizeof(u64) * *npages, GFP_KERNEL);
469 	if (!*page_list)
470 		return -ENOMEM;
471 
472 	n = 0;
473 	for (i = 0; i < num_phys_buf; ++i)
474 		for (j = 0;
475 		     j < (buffer_list[i].size + (1ULL << *shift) - 1) >> *shift;
476 		     ++j)
477 			(*page_list)[n++] = cpu_to_be64(buffer_list[i].addr +
478 			    ((u64) j << *shift));
479 
480 	PDBG("%s va 0x%llx mask 0x%llx shift %d len %lld pbl_size %d\n",
481 	     __func__, (unsigned long long)*iova_start,
482 	     (unsigned long long)mask, *shift, (unsigned long long)*total_size,
483 	     *npages);
484 
485 	return 0;
486 
487 }
488 
489 int c4iw_reregister_phys_mem(struct ib_mr *mr, int mr_rereg_mask,
490 			     struct ib_pd *pd, struct ib_phys_buf *buffer_list,
491 			     int num_phys_buf, int acc, u64 *iova_start)
492 {
493 
494 	struct c4iw_mr mh, *mhp;
495 	struct c4iw_pd *php;
496 	struct c4iw_dev *rhp;
497 	__be64 *page_list = NULL;
498 	int shift = 0;
499 	u64 total_size;
500 	int npages;
501 	int ret;
502 
503 	PDBG("%s ib_mr %p ib_pd %p\n", __func__, mr, pd);
504 
505 	/* There can be no memory windows */
506 	if (atomic_read(&mr->usecnt))
507 		return -EINVAL;
508 
509 	mhp = to_c4iw_mr(mr);
510 	rhp = mhp->rhp;
511 	php = to_c4iw_pd(mr->pd);
512 
513 	/* make sure we are on the same adapter */
514 	if (rhp != php->rhp)
515 		return -EINVAL;
516 
517 	memcpy(&mh, mhp, sizeof *mhp);
518 
519 	if (mr_rereg_mask & IB_MR_REREG_PD)
520 		php = to_c4iw_pd(pd);
521 	if (mr_rereg_mask & IB_MR_REREG_ACCESS) {
522 		mh.attr.perms = c4iw_ib_to_tpt_access(acc);
523 		mh.attr.mw_bind_enable = (acc & IB_ACCESS_MW_BIND) ==
524 					 IB_ACCESS_MW_BIND;
525 	}
526 	if (mr_rereg_mask & IB_MR_REREG_TRANS) {
527 		ret = build_phys_page_list(buffer_list, num_phys_buf,
528 						iova_start,
529 						&total_size, &npages,
530 						&shift, &page_list);
531 		if (ret)
532 			return ret;
533 	}
534 
535 	ret = reregister_mem(rhp, php, &mh, shift, npages);
536 	kfree(page_list);
537 	if (ret)
538 		return ret;
539 	if (mr_rereg_mask & IB_MR_REREG_PD)
540 		mhp->attr.pdid = php->pdid;
541 	if (mr_rereg_mask & IB_MR_REREG_ACCESS)
542 		mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
543 	if (mr_rereg_mask & IB_MR_REREG_TRANS) {
544 		mhp->attr.zbva = 0;
545 		mhp->attr.va_fbo = *iova_start;
546 		mhp->attr.page_size = shift - 12;
547 		mhp->attr.len = (u32) total_size;
548 		mhp->attr.pbl_size = npages;
549 	}
550 
551 	return 0;
552 }
553 
554 struct ib_mr *c4iw_register_phys_mem(struct ib_pd *pd,
555 				     struct ib_phys_buf *buffer_list,
556 				     int num_phys_buf, int acc, u64 *iova_start)
557 {
558 	__be64 *page_list;
559 	int shift;
560 	u64 total_size;
561 	int npages;
562 	struct c4iw_dev *rhp;
563 	struct c4iw_pd *php;
564 	struct c4iw_mr *mhp;
565 	int ret;
566 
567 	PDBG("%s ib_pd %p\n", __func__, pd);
568 	php = to_c4iw_pd(pd);
569 	rhp = php->rhp;
570 
571 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
572 	if (!mhp)
573 		return ERR_PTR(-ENOMEM);
574 
575 	mhp->rhp = rhp;
576 
577 	/* First check that we have enough alignment */
578 	if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK)) {
579 		ret = -EINVAL;
580 		goto err;
581 	}
582 
583 	if (num_phys_buf > 1 &&
584 	    ((buffer_list[0].addr + buffer_list[0].size) & ~PAGE_MASK)) {
585 		ret = -EINVAL;
586 		goto err;
587 	}
588 
589 	ret = build_phys_page_list(buffer_list, num_phys_buf, iova_start,
590 					&total_size, &npages, &shift,
591 					&page_list);
592 	if (ret)
593 		goto err;
594 
595 	ret = alloc_pbl(mhp, npages);
596 	if (ret) {
597 		kfree(page_list);
598 		goto err;
599 	}
600 
601 	ret = write_pbl(&mhp->rhp->rdev, page_list, mhp->attr.pbl_addr,
602 			     npages);
603 	kfree(page_list);
604 	if (ret)
605 		goto err_pbl;
606 
607 	mhp->attr.pdid = php->pdid;
608 	mhp->attr.zbva = 0;
609 
610 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
611 	mhp->attr.va_fbo = *iova_start;
612 	mhp->attr.page_size = shift - 12;
613 
614 	mhp->attr.len = (u32) total_size;
615 	mhp->attr.pbl_size = npages;
616 	ret = register_mem(rhp, php, mhp, shift);
617 	if (ret)
618 		goto err_pbl;
619 
620 	return &mhp->ibmr;
621 
622 err_pbl:
623 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
624 			      mhp->attr.pbl_size << 3);
625 
626 err:
627 	kfree(mhp);
628 	return ERR_PTR(ret);
629 
630 }
631 
632 struct ib_mr *c4iw_get_dma_mr(struct ib_pd *pd, int acc)
633 {
634 	struct c4iw_dev *rhp;
635 	struct c4iw_pd *php;
636 	struct c4iw_mr *mhp;
637 	int ret;
638 	u32 stag = T4_STAG_UNSET;
639 
640 	PDBG("%s ib_pd %p\n", __func__, pd);
641 	php = to_c4iw_pd(pd);
642 	rhp = php->rhp;
643 
644 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
645 	if (!mhp)
646 		return ERR_PTR(-ENOMEM);
647 
648 	mhp->rhp = rhp;
649 	mhp->attr.pdid = php->pdid;
650 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
651 	mhp->attr.mw_bind_enable = (acc&IB_ACCESS_MW_BIND) == IB_ACCESS_MW_BIND;
652 	mhp->attr.zbva = 0;
653 	mhp->attr.va_fbo = 0;
654 	mhp->attr.page_size = 0;
655 	mhp->attr.len = ~0UL;
656 	mhp->attr.pbl_size = 0;
657 
658 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, php->pdid,
659 			      FW_RI_STAG_NSMR, mhp->attr.perms,
660 			      mhp->attr.mw_bind_enable, 0, 0, ~0UL, 0, 0, 0);
661 	if (ret)
662 		goto err1;
663 
664 	ret = finish_mem_reg(mhp, stag);
665 	if (ret)
666 		goto err2;
667 	return &mhp->ibmr;
668 err2:
669 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
670 		  mhp->attr.pbl_addr);
671 err1:
672 	kfree(mhp);
673 	return ERR_PTR(ret);
674 }
675 
676 struct ib_mr *c4iw_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
677 			       u64 virt, int acc, struct ib_udata *udata)
678 {
679 	__be64 *pages;
680 	int shift, n, len;
681 	int i, j, k;
682 	int err = 0;
683 	struct ib_umem_chunk *chunk;
684 	struct c4iw_dev *rhp;
685 	struct c4iw_pd *php;
686 	struct c4iw_mr *mhp;
687 
688 	PDBG("%s ib_pd %p\n", __func__, pd);
689 
690 	if (length == ~0ULL)
691 		return ERR_PTR(-EINVAL);
692 
693 	if ((length + start) < start)
694 		return ERR_PTR(-EINVAL);
695 
696 	php = to_c4iw_pd(pd);
697 	rhp = php->rhp;
698 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
699 	if (!mhp)
700 		return ERR_PTR(-ENOMEM);
701 
702 	mhp->rhp = rhp;
703 
704 	mhp->umem = ib_umem_get(pd->uobject->context, start, length, acc, 0);
705 	if (IS_ERR(mhp->umem)) {
706 		err = PTR_ERR(mhp->umem);
707 		kfree(mhp);
708 		return ERR_PTR(err);
709 	}
710 
711 	shift = ffs(mhp->umem->page_size) - 1;
712 
713 	n = 0;
714 	list_for_each_entry(chunk, &mhp->umem->chunk_list, list)
715 		n += chunk->nents;
716 
717 	err = alloc_pbl(mhp, n);
718 	if (err)
719 		goto err;
720 
721 	pages = (__be64 *) __get_free_page(GFP_KERNEL);
722 	if (!pages) {
723 		err = -ENOMEM;
724 		goto err_pbl;
725 	}
726 
727 	i = n = 0;
728 
729 	list_for_each_entry(chunk, &mhp->umem->chunk_list, list)
730 		for (j = 0; j < chunk->nmap; ++j) {
731 			len = sg_dma_len(&chunk->page_list[j]) >> shift;
732 			for (k = 0; k < len; ++k) {
733 				pages[i++] = cpu_to_be64(sg_dma_address(
734 					&chunk->page_list[j]) +
735 					mhp->umem->page_size * k);
736 				if (i == PAGE_SIZE / sizeof *pages) {
737 					err = write_pbl(&mhp->rhp->rdev,
738 					      pages,
739 					      mhp->attr.pbl_addr + (n << 3), i);
740 					if (err)
741 						goto pbl_done;
742 					n += i;
743 					i = 0;
744 				}
745 			}
746 		}
747 
748 	if (i)
749 		err = write_pbl(&mhp->rhp->rdev, pages,
750 				     mhp->attr.pbl_addr + (n << 3), i);
751 
752 pbl_done:
753 	free_page((unsigned long) pages);
754 	if (err)
755 		goto err_pbl;
756 
757 	mhp->attr.pdid = php->pdid;
758 	mhp->attr.zbva = 0;
759 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
760 	mhp->attr.va_fbo = virt;
761 	mhp->attr.page_size = shift - 12;
762 	mhp->attr.len = length;
763 
764 	err = register_mem(rhp, php, mhp, shift);
765 	if (err)
766 		goto err_pbl;
767 
768 	return &mhp->ibmr;
769 
770 err_pbl:
771 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
772 			      mhp->attr.pbl_size << 3);
773 
774 err:
775 	ib_umem_release(mhp->umem);
776 	kfree(mhp);
777 	return ERR_PTR(err);
778 }
779 
780 struct ib_mw *c4iw_alloc_mw(struct ib_pd *pd, enum ib_mw_type type)
781 {
782 	struct c4iw_dev *rhp;
783 	struct c4iw_pd *php;
784 	struct c4iw_mw *mhp;
785 	u32 mmid;
786 	u32 stag = 0;
787 	int ret;
788 
789 	if (type != IB_MW_TYPE_1)
790 		return ERR_PTR(-EINVAL);
791 
792 	php = to_c4iw_pd(pd);
793 	rhp = php->rhp;
794 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
795 	if (!mhp)
796 		return ERR_PTR(-ENOMEM);
797 	ret = allocate_window(&rhp->rdev, &stag, php->pdid);
798 	if (ret) {
799 		kfree(mhp);
800 		return ERR_PTR(ret);
801 	}
802 	mhp->rhp = rhp;
803 	mhp->attr.pdid = php->pdid;
804 	mhp->attr.type = FW_RI_STAG_MW;
805 	mhp->attr.stag = stag;
806 	mmid = (stag) >> 8;
807 	mhp->ibmw.rkey = stag;
808 	if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
809 		deallocate_window(&rhp->rdev, mhp->attr.stag);
810 		kfree(mhp);
811 		return ERR_PTR(-ENOMEM);
812 	}
813 	PDBG("%s mmid 0x%x mhp %p stag 0x%x\n", __func__, mmid, mhp, stag);
814 	return &(mhp->ibmw);
815 }
816 
817 int c4iw_dealloc_mw(struct ib_mw *mw)
818 {
819 	struct c4iw_dev *rhp;
820 	struct c4iw_mw *mhp;
821 	u32 mmid;
822 
823 	mhp = to_c4iw_mw(mw);
824 	rhp = mhp->rhp;
825 	mmid = (mw->rkey) >> 8;
826 	remove_handle(rhp, &rhp->mmidr, mmid);
827 	deallocate_window(&rhp->rdev, mhp->attr.stag);
828 	kfree(mhp);
829 	PDBG("%s ib_mw %p mmid 0x%x ptr %p\n", __func__, mw, mmid, mhp);
830 	return 0;
831 }
832 
833 struct ib_mr *c4iw_alloc_fast_reg_mr(struct ib_pd *pd, int pbl_depth)
834 {
835 	struct c4iw_dev *rhp;
836 	struct c4iw_pd *php;
837 	struct c4iw_mr *mhp;
838 	u32 mmid;
839 	u32 stag = 0;
840 	int ret = 0;
841 
842 	php = to_c4iw_pd(pd);
843 	rhp = php->rhp;
844 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
845 	if (!mhp) {
846 		ret = -ENOMEM;
847 		goto err;
848 	}
849 
850 	mhp->rhp = rhp;
851 	ret = alloc_pbl(mhp, pbl_depth);
852 	if (ret)
853 		goto err1;
854 	mhp->attr.pbl_size = pbl_depth;
855 	ret = allocate_stag(&rhp->rdev, &stag, php->pdid,
856 				 mhp->attr.pbl_size, mhp->attr.pbl_addr);
857 	if (ret)
858 		goto err2;
859 	mhp->attr.pdid = php->pdid;
860 	mhp->attr.type = FW_RI_STAG_NSMR;
861 	mhp->attr.stag = stag;
862 	mhp->attr.state = 1;
863 	mmid = (stag) >> 8;
864 	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
865 	if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
866 		ret = -ENOMEM;
867 		goto err3;
868 	}
869 
870 	PDBG("%s mmid 0x%x mhp %p stag 0x%x\n", __func__, mmid, mhp, stag);
871 	return &(mhp->ibmr);
872 err3:
873 	dereg_mem(&rhp->rdev, stag, mhp->attr.pbl_size,
874 		       mhp->attr.pbl_addr);
875 err2:
876 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
877 			      mhp->attr.pbl_size << 3);
878 err1:
879 	kfree(mhp);
880 err:
881 	return ERR_PTR(ret);
882 }
883 
884 struct ib_fast_reg_page_list *c4iw_alloc_fastreg_pbl(struct ib_device *device,
885 						     int page_list_len)
886 {
887 	struct c4iw_fr_page_list *c4pl;
888 	struct c4iw_dev *dev = to_c4iw_dev(device);
889 	dma_addr_t dma_addr;
890 	int pll_len = roundup(page_list_len * sizeof(u64), 32);
891 
892 	c4pl = kmalloc(sizeof(*c4pl), GFP_KERNEL);
893 	if (!c4pl)
894 		return ERR_PTR(-ENOMEM);
895 
896 	c4pl->ibpl.page_list = dma_alloc_coherent(&dev->rdev.lldi.pdev->dev,
897 						  pll_len, &dma_addr,
898 						  GFP_KERNEL);
899 	if (!c4pl->ibpl.page_list) {
900 		kfree(c4pl);
901 		return ERR_PTR(-ENOMEM);
902 	}
903 	dma_unmap_addr_set(c4pl, mapping, dma_addr);
904 	c4pl->dma_addr = dma_addr;
905 	c4pl->dev = dev;
906 	c4pl->ibpl.max_page_list_len = pll_len;
907 
908 	return &c4pl->ibpl;
909 }
910 
911 void c4iw_free_fastreg_pbl(struct ib_fast_reg_page_list *ibpl)
912 {
913 	struct c4iw_fr_page_list *c4pl = to_c4iw_fr_page_list(ibpl);
914 
915 	dma_free_coherent(&c4pl->dev->rdev.lldi.pdev->dev,
916 			  c4pl->ibpl.max_page_list_len,
917 			  c4pl->ibpl.page_list, dma_unmap_addr(c4pl, mapping));
918 	kfree(c4pl);
919 }
920 
921 int c4iw_dereg_mr(struct ib_mr *ib_mr)
922 {
923 	struct c4iw_dev *rhp;
924 	struct c4iw_mr *mhp;
925 	u32 mmid;
926 
927 	PDBG("%s ib_mr %p\n", __func__, ib_mr);
928 	/* There can be no memory windows */
929 	if (atomic_read(&ib_mr->usecnt))
930 		return -EINVAL;
931 
932 	mhp = to_c4iw_mr(ib_mr);
933 	rhp = mhp->rhp;
934 	mmid = mhp->attr.stag >> 8;
935 	remove_handle(rhp, &rhp->mmidr, mmid);
936 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
937 		       mhp->attr.pbl_addr);
938 	if (mhp->attr.pbl_size)
939 		c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
940 				  mhp->attr.pbl_size << 3);
941 	if (mhp->kva)
942 		kfree((void *) (unsigned long) mhp->kva);
943 	if (mhp->umem)
944 		ib_umem_release(mhp->umem);
945 	PDBG("%s mmid 0x%x ptr %p\n", __func__, mmid, mhp);
946 	kfree(mhp);
947 	return 0;
948 }
949