xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/mem.c (revision 0883c2c0)
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 #include <rdma/ib_user_verbs.h>
38 
39 #include "iw_cxgb4.h"
40 
41 int use_dsgl = 0;
42 module_param(use_dsgl, int, 0644);
43 MODULE_PARM_DESC(use_dsgl, "Use DSGL for PBL/FastReg (default=0)");
44 
45 #define T4_ULPTX_MIN_IO 32
46 #define C4IW_MAX_INLINE_SIZE 96
47 #define T4_ULPTX_MAX_DMA 1024
48 #define C4IW_INLINE_THRESHOLD 128
49 
50 static int inline_threshold = C4IW_INLINE_THRESHOLD;
51 module_param(inline_threshold, int, 0644);
52 MODULE_PARM_DESC(inline_threshold, "inline vs dsgl threshold (default=128)");
53 
54 static int mr_exceeds_hw_limits(struct c4iw_dev *dev, u64 length)
55 {
56 	return (is_t4(dev->rdev.lldi.adapter_type) ||
57 		is_t5(dev->rdev.lldi.adapter_type)) &&
58 		length >= 8*1024*1024*1024ULL;
59 }
60 
61 static int _c4iw_write_mem_dma_aligned(struct c4iw_rdev *rdev, u32 addr,
62 				       u32 len, dma_addr_t data, int wait)
63 {
64 	struct sk_buff *skb;
65 	struct ulp_mem_io *req;
66 	struct ulptx_sgl *sgl;
67 	u8 wr_len;
68 	int ret = 0;
69 	struct c4iw_wr_wait wr_wait;
70 
71 	addr &= 0x7FFFFFF;
72 
73 	if (wait)
74 		c4iw_init_wr_wait(&wr_wait);
75 	wr_len = roundup(sizeof(*req) + sizeof(*sgl), 16);
76 
77 	skb = alloc_skb(wr_len, GFP_KERNEL);
78 	if (!skb)
79 		return -ENOMEM;
80 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
81 
82 	req = (struct ulp_mem_io *)__skb_put(skb, wr_len);
83 	memset(req, 0, wr_len);
84 	INIT_ULPTX_WR(req, wr_len, 0, 0);
85 	req->wr.wr_hi = cpu_to_be32(FW_WR_OP_V(FW_ULPTX_WR) |
86 			(wait ? FW_WR_COMPL_F : 0));
87 	req->wr.wr_lo = wait ? (__force __be64)(unsigned long) &wr_wait : 0L;
88 	req->wr.wr_mid = cpu_to_be32(FW_WR_LEN16_V(DIV_ROUND_UP(wr_len, 16)));
89 	req->cmd = cpu_to_be32(ULPTX_CMD_V(ULP_TX_MEM_WRITE) |
90 			       T5_ULP_MEMIO_ORDER_V(1) |
91 			       T5_ULP_MEMIO_FID_V(rdev->lldi.rxq_ids[0]));
92 	req->dlen = cpu_to_be32(ULP_MEMIO_DATA_LEN_V(len>>5));
93 	req->len16 = cpu_to_be32(DIV_ROUND_UP(wr_len-sizeof(req->wr), 16));
94 	req->lock_addr = cpu_to_be32(ULP_MEMIO_ADDR_V(addr));
95 
96 	sgl = (struct ulptx_sgl *)(req + 1);
97 	sgl->cmd_nsge = cpu_to_be32(ULPTX_CMD_V(ULP_TX_SC_DSGL) |
98 				    ULPTX_NSGE_V(1));
99 	sgl->len0 = cpu_to_be32(len);
100 	sgl->addr0 = cpu_to_be64(data);
101 
102 	ret = c4iw_ofld_send(rdev, skb);
103 	if (ret)
104 		return ret;
105 	if (wait)
106 		ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
107 	return ret;
108 }
109 
110 static int _c4iw_write_mem_inline(struct c4iw_rdev *rdev, u32 addr, u32 len,
111 				  void *data)
112 {
113 	struct sk_buff *skb;
114 	struct ulp_mem_io *req;
115 	struct ulptx_idata *sc;
116 	u8 wr_len, *to_dp, *from_dp;
117 	int copy_len, num_wqe, i, ret = 0;
118 	struct c4iw_wr_wait wr_wait;
119 	__be32 cmd = cpu_to_be32(ULPTX_CMD_V(ULP_TX_MEM_WRITE));
120 
121 	if (is_t4(rdev->lldi.adapter_type))
122 		cmd |= cpu_to_be32(ULP_MEMIO_ORDER_F);
123 	else
124 		cmd |= cpu_to_be32(T5_ULP_MEMIO_IMM_F);
125 
126 	addr &= 0x7FFFFFF;
127 	PDBG("%s addr 0x%x len %u\n", __func__, addr, len);
128 	num_wqe = DIV_ROUND_UP(len, C4IW_MAX_INLINE_SIZE);
129 	c4iw_init_wr_wait(&wr_wait);
130 	for (i = 0; i < num_wqe; i++) {
131 
132 		copy_len = len > C4IW_MAX_INLINE_SIZE ? C4IW_MAX_INLINE_SIZE :
133 			   len;
134 		wr_len = roundup(sizeof *req + sizeof *sc +
135 				 roundup(copy_len, T4_ULPTX_MIN_IO), 16);
136 
137 		skb = alloc_skb(wr_len, GFP_KERNEL);
138 		if (!skb)
139 			return -ENOMEM;
140 		set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
141 
142 		req = (struct ulp_mem_io *)__skb_put(skb, wr_len);
143 		memset(req, 0, wr_len);
144 		INIT_ULPTX_WR(req, wr_len, 0, 0);
145 
146 		if (i == (num_wqe-1)) {
147 			req->wr.wr_hi = cpu_to_be32(FW_WR_OP_V(FW_ULPTX_WR) |
148 						    FW_WR_COMPL_F);
149 			req->wr.wr_lo = (__force __be64)(unsigned long)&wr_wait;
150 		} else
151 			req->wr.wr_hi = cpu_to_be32(FW_WR_OP_V(FW_ULPTX_WR));
152 		req->wr.wr_mid = cpu_to_be32(
153 				       FW_WR_LEN16_V(DIV_ROUND_UP(wr_len, 16)));
154 
155 		req->cmd = cmd;
156 		req->dlen = cpu_to_be32(ULP_MEMIO_DATA_LEN_V(
157 				DIV_ROUND_UP(copy_len, T4_ULPTX_MIN_IO)));
158 		req->len16 = cpu_to_be32(DIV_ROUND_UP(wr_len-sizeof(req->wr),
159 						      16));
160 		req->lock_addr = cpu_to_be32(ULP_MEMIO_ADDR_V(addr + i * 3));
161 
162 		sc = (struct ulptx_idata *)(req + 1);
163 		sc->cmd_more = cpu_to_be32(ULPTX_CMD_V(ULP_TX_SC_IMM));
164 		sc->len = cpu_to_be32(roundup(copy_len, T4_ULPTX_MIN_IO));
165 
166 		to_dp = (u8 *)(sc + 1);
167 		from_dp = (u8 *)data + i * C4IW_MAX_INLINE_SIZE;
168 		if (data)
169 			memcpy(to_dp, from_dp, copy_len);
170 		else
171 			memset(to_dp, 0, copy_len);
172 		if (copy_len % T4_ULPTX_MIN_IO)
173 			memset(to_dp + copy_len, 0, T4_ULPTX_MIN_IO -
174 			       (copy_len % T4_ULPTX_MIN_IO));
175 		ret = c4iw_ofld_send(rdev, skb);
176 		if (ret)
177 			return ret;
178 		len -= C4IW_MAX_INLINE_SIZE;
179 	}
180 
181 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
182 	return ret;
183 }
184 
185 static int _c4iw_write_mem_dma(struct c4iw_rdev *rdev, u32 addr, u32 len, void *data)
186 {
187 	u32 remain = len;
188 	u32 dmalen;
189 	int ret = 0;
190 	dma_addr_t daddr;
191 	dma_addr_t save;
192 
193 	daddr = dma_map_single(&rdev->lldi.pdev->dev, data, len, DMA_TO_DEVICE);
194 	if (dma_mapping_error(&rdev->lldi.pdev->dev, daddr))
195 		return -1;
196 	save = daddr;
197 
198 	while (remain > inline_threshold) {
199 		if (remain < T4_ULPTX_MAX_DMA) {
200 			if (remain & ~T4_ULPTX_MIN_IO)
201 				dmalen = remain & ~(T4_ULPTX_MIN_IO-1);
202 			else
203 				dmalen = remain;
204 		} else
205 			dmalen = T4_ULPTX_MAX_DMA;
206 		remain -= dmalen;
207 		ret = _c4iw_write_mem_dma_aligned(rdev, addr, dmalen, daddr,
208 						 !remain);
209 		if (ret)
210 			goto out;
211 		addr += dmalen >> 5;
212 		data += dmalen;
213 		daddr += dmalen;
214 	}
215 	if (remain)
216 		ret = _c4iw_write_mem_inline(rdev, addr, remain, data);
217 out:
218 	dma_unmap_single(&rdev->lldi.pdev->dev, save, len, DMA_TO_DEVICE);
219 	return ret;
220 }
221 
222 /*
223  * write len bytes of data into addr (32B aligned address)
224  * If data is NULL, clear len byte of memory to zero.
225  */
226 static int write_adapter_mem(struct c4iw_rdev *rdev, u32 addr, u32 len,
227 			     void *data)
228 {
229 	if (is_t5(rdev->lldi.adapter_type) && use_dsgl) {
230 		if (len > inline_threshold) {
231 			if (_c4iw_write_mem_dma(rdev, addr, len, data)) {
232 				printk_ratelimited(KERN_WARNING
233 						   "%s: dma map"
234 						   " failure (non fatal)\n",
235 						   pci_name(rdev->lldi.pdev));
236 				return _c4iw_write_mem_inline(rdev, addr, len,
237 							      data);
238 			} else
239 				return 0;
240 		} else
241 			return _c4iw_write_mem_inline(rdev, addr, len, data);
242 	} else
243 		return _c4iw_write_mem_inline(rdev, addr, len, data);
244 }
245 
246 /*
247  * Build and write a TPT entry.
248  * IN: stag key, pdid, perm, bind_enabled, zbva, to, len, page_size,
249  *     pbl_size and pbl_addr
250  * OUT: stag index
251  */
252 static int write_tpt_entry(struct c4iw_rdev *rdev, u32 reset_tpt_entry,
253 			   u32 *stag, u8 stag_state, u32 pdid,
254 			   enum fw_ri_stag_type type, enum fw_ri_mem_perms perm,
255 			   int bind_enabled, u32 zbva, u64 to,
256 			   u64 len, u8 page_size, u32 pbl_size, u32 pbl_addr)
257 {
258 	int err;
259 	struct fw_ri_tpte tpt;
260 	u32 stag_idx;
261 	static atomic_t key;
262 
263 	if (c4iw_fatal_error(rdev))
264 		return -EIO;
265 
266 	stag_state = stag_state > 0;
267 	stag_idx = (*stag) >> 8;
268 
269 	if ((!reset_tpt_entry) && (*stag == T4_STAG_UNSET)) {
270 		stag_idx = c4iw_get_resource(&rdev->resource.tpt_table);
271 		if (!stag_idx) {
272 			mutex_lock(&rdev->stats.lock);
273 			rdev->stats.stag.fail++;
274 			mutex_unlock(&rdev->stats.lock);
275 			return -ENOMEM;
276 		}
277 		mutex_lock(&rdev->stats.lock);
278 		rdev->stats.stag.cur += 32;
279 		if (rdev->stats.stag.cur > rdev->stats.stag.max)
280 			rdev->stats.stag.max = rdev->stats.stag.cur;
281 		mutex_unlock(&rdev->stats.lock);
282 		*stag = (stag_idx << 8) | (atomic_inc_return(&key) & 0xff);
283 	}
284 	PDBG("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
285 	     __func__, stag_state, type, pdid, stag_idx);
286 
287 	/* write TPT entry */
288 	if (reset_tpt_entry)
289 		memset(&tpt, 0, sizeof(tpt));
290 	else {
291 		tpt.valid_to_pdid = cpu_to_be32(FW_RI_TPTE_VALID_F |
292 			FW_RI_TPTE_STAGKEY_V((*stag & FW_RI_TPTE_STAGKEY_M)) |
293 			FW_RI_TPTE_STAGSTATE_V(stag_state) |
294 			FW_RI_TPTE_STAGTYPE_V(type) | FW_RI_TPTE_PDID_V(pdid));
295 		tpt.locread_to_qpid = cpu_to_be32(FW_RI_TPTE_PERM_V(perm) |
296 			(bind_enabled ? FW_RI_TPTE_MWBINDEN_F : 0) |
297 			FW_RI_TPTE_ADDRTYPE_V((zbva ? FW_RI_ZERO_BASED_TO :
298 						      FW_RI_VA_BASED_TO))|
299 			FW_RI_TPTE_PS_V(page_size));
300 		tpt.nosnoop_pbladdr = !pbl_size ? 0 : cpu_to_be32(
301 			FW_RI_TPTE_PBLADDR_V(PBL_OFF(rdev, pbl_addr)>>3));
302 		tpt.len_lo = cpu_to_be32((u32)(len & 0xffffffffUL));
303 		tpt.va_hi = cpu_to_be32((u32)(to >> 32));
304 		tpt.va_lo_fbo = cpu_to_be32((u32)(to & 0xffffffffUL));
305 		tpt.dca_mwbcnt_pstag = cpu_to_be32(0);
306 		tpt.len_hi = cpu_to_be32((u32)(len >> 32));
307 	}
308 	err = write_adapter_mem(rdev, stag_idx +
309 				(rdev->lldi.vr->stag.start >> 5),
310 				sizeof(tpt), &tpt);
311 
312 	if (reset_tpt_entry) {
313 		c4iw_put_resource(&rdev->resource.tpt_table, stag_idx);
314 		mutex_lock(&rdev->stats.lock);
315 		rdev->stats.stag.cur -= 32;
316 		mutex_unlock(&rdev->stats.lock);
317 	}
318 	return err;
319 }
320 
321 static int write_pbl(struct c4iw_rdev *rdev, __be64 *pbl,
322 		     u32 pbl_addr, u32 pbl_size)
323 {
324 	int err;
325 
326 	PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
327 	     __func__, pbl_addr, rdev->lldi.vr->pbl.start,
328 	     pbl_size);
329 
330 	err = write_adapter_mem(rdev, pbl_addr >> 5, pbl_size << 3, pbl);
331 	return err;
332 }
333 
334 static int dereg_mem(struct c4iw_rdev *rdev, u32 stag, u32 pbl_size,
335 		     u32 pbl_addr)
336 {
337 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0,
338 			       pbl_size, pbl_addr);
339 }
340 
341 static int allocate_window(struct c4iw_rdev *rdev, u32 * stag, u32 pdid)
342 {
343 	*stag = T4_STAG_UNSET;
344 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_MW, 0, 0, 0,
345 			       0UL, 0, 0, 0, 0);
346 }
347 
348 static int deallocate_window(struct c4iw_rdev *rdev, u32 stag)
349 {
350 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0, 0,
351 			       0);
352 }
353 
354 static int allocate_stag(struct c4iw_rdev *rdev, u32 *stag, u32 pdid,
355 			 u32 pbl_size, u32 pbl_addr)
356 {
357 	*stag = T4_STAG_UNSET;
358 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_NSMR, 0, 0, 0,
359 			       0UL, 0, 0, pbl_size, pbl_addr);
360 }
361 
362 static int finish_mem_reg(struct c4iw_mr *mhp, u32 stag)
363 {
364 	u32 mmid;
365 
366 	mhp->attr.state = 1;
367 	mhp->attr.stag = stag;
368 	mmid = stag >> 8;
369 	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
370 	PDBG("%s mmid 0x%x mhp %p\n", __func__, mmid, mhp);
371 	return insert_handle(mhp->rhp, &mhp->rhp->mmidr, mhp, mmid);
372 }
373 
374 static int register_mem(struct c4iw_dev *rhp, struct c4iw_pd *php,
375 		      struct c4iw_mr *mhp, int shift)
376 {
377 	u32 stag = T4_STAG_UNSET;
378 	int ret;
379 
380 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, mhp->attr.pdid,
381 			      FW_RI_STAG_NSMR, mhp->attr.len ?
382 			      mhp->attr.perms : 0,
383 			      mhp->attr.mw_bind_enable, mhp->attr.zbva,
384 			      mhp->attr.va_fbo, mhp->attr.len ?
385 			      mhp->attr.len : -1, shift - 12,
386 			      mhp->attr.pbl_size, mhp->attr.pbl_addr);
387 	if (ret)
388 		return ret;
389 
390 	ret = finish_mem_reg(mhp, stag);
391 	if (ret)
392 		dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
393 		       mhp->attr.pbl_addr);
394 	return ret;
395 }
396 
397 static int alloc_pbl(struct c4iw_mr *mhp, int npages)
398 {
399 	mhp->attr.pbl_addr = c4iw_pblpool_alloc(&mhp->rhp->rdev,
400 						    npages << 3);
401 
402 	if (!mhp->attr.pbl_addr)
403 		return -ENOMEM;
404 
405 	mhp->attr.pbl_size = npages;
406 
407 	return 0;
408 }
409 
410 struct ib_mr *c4iw_get_dma_mr(struct ib_pd *pd, int acc)
411 {
412 	struct c4iw_dev *rhp;
413 	struct c4iw_pd *php;
414 	struct c4iw_mr *mhp;
415 	int ret;
416 	u32 stag = T4_STAG_UNSET;
417 
418 	PDBG("%s ib_pd %p\n", __func__, pd);
419 	php = to_c4iw_pd(pd);
420 	rhp = php->rhp;
421 
422 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
423 	if (!mhp)
424 		return ERR_PTR(-ENOMEM);
425 
426 	mhp->rhp = rhp;
427 	mhp->attr.pdid = php->pdid;
428 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
429 	mhp->attr.mw_bind_enable = (acc&IB_ACCESS_MW_BIND) == IB_ACCESS_MW_BIND;
430 	mhp->attr.zbva = 0;
431 	mhp->attr.va_fbo = 0;
432 	mhp->attr.page_size = 0;
433 	mhp->attr.len = ~0ULL;
434 	mhp->attr.pbl_size = 0;
435 
436 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, php->pdid,
437 			      FW_RI_STAG_NSMR, mhp->attr.perms,
438 			      mhp->attr.mw_bind_enable, 0, 0, ~0ULL, 0, 0, 0);
439 	if (ret)
440 		goto err1;
441 
442 	ret = finish_mem_reg(mhp, stag);
443 	if (ret)
444 		goto err2;
445 	return &mhp->ibmr;
446 err2:
447 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
448 		  mhp->attr.pbl_addr);
449 err1:
450 	kfree(mhp);
451 	return ERR_PTR(ret);
452 }
453 
454 struct ib_mr *c4iw_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
455 			       u64 virt, int acc, struct ib_udata *udata)
456 {
457 	__be64 *pages;
458 	int shift, n, len;
459 	int i, k, entry;
460 	int err = 0;
461 	struct scatterlist *sg;
462 	struct c4iw_dev *rhp;
463 	struct c4iw_pd *php;
464 	struct c4iw_mr *mhp;
465 
466 	PDBG("%s ib_pd %p\n", __func__, pd);
467 
468 	if (length == ~0ULL)
469 		return ERR_PTR(-EINVAL);
470 
471 	if ((length + start) < start)
472 		return ERR_PTR(-EINVAL);
473 
474 	php = to_c4iw_pd(pd);
475 	rhp = php->rhp;
476 
477 	if (mr_exceeds_hw_limits(rhp, length))
478 		return ERR_PTR(-EINVAL);
479 
480 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
481 	if (!mhp)
482 		return ERR_PTR(-ENOMEM);
483 
484 	mhp->rhp = rhp;
485 
486 	mhp->umem = ib_umem_get(pd->uobject->context, start, length, acc, 0);
487 	if (IS_ERR(mhp->umem)) {
488 		err = PTR_ERR(mhp->umem);
489 		kfree(mhp);
490 		return ERR_PTR(err);
491 	}
492 
493 	shift = ffs(mhp->umem->page_size) - 1;
494 
495 	n = mhp->umem->nmap;
496 	err = alloc_pbl(mhp, n);
497 	if (err)
498 		goto err;
499 
500 	pages = (__be64 *) __get_free_page(GFP_KERNEL);
501 	if (!pages) {
502 		err = -ENOMEM;
503 		goto err_pbl;
504 	}
505 
506 	i = n = 0;
507 
508 	for_each_sg(mhp->umem->sg_head.sgl, sg, mhp->umem->nmap, entry) {
509 		len = sg_dma_len(sg) >> shift;
510 		for (k = 0; k < len; ++k) {
511 			pages[i++] = cpu_to_be64(sg_dma_address(sg) +
512 				mhp->umem->page_size * k);
513 			if (i == PAGE_SIZE / sizeof *pages) {
514 				err = write_pbl(&mhp->rhp->rdev,
515 				      pages,
516 				      mhp->attr.pbl_addr + (n << 3), i);
517 				if (err)
518 					goto pbl_done;
519 				n += i;
520 				i = 0;
521 			}
522 		}
523 	}
524 
525 	if (i)
526 		err = write_pbl(&mhp->rhp->rdev, pages,
527 				     mhp->attr.pbl_addr + (n << 3), i);
528 
529 pbl_done:
530 	free_page((unsigned long) pages);
531 	if (err)
532 		goto err_pbl;
533 
534 	mhp->attr.pdid = php->pdid;
535 	mhp->attr.zbva = 0;
536 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
537 	mhp->attr.va_fbo = virt;
538 	mhp->attr.page_size = shift - 12;
539 	mhp->attr.len = length;
540 
541 	err = register_mem(rhp, php, mhp, shift);
542 	if (err)
543 		goto err_pbl;
544 
545 	return &mhp->ibmr;
546 
547 err_pbl:
548 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
549 			      mhp->attr.pbl_size << 3);
550 
551 err:
552 	ib_umem_release(mhp->umem);
553 	kfree(mhp);
554 	return ERR_PTR(err);
555 }
556 
557 struct ib_mw *c4iw_alloc_mw(struct ib_pd *pd, enum ib_mw_type type,
558 			    struct ib_udata *udata)
559 {
560 	struct c4iw_dev *rhp;
561 	struct c4iw_pd *php;
562 	struct c4iw_mw *mhp;
563 	u32 mmid;
564 	u32 stag = 0;
565 	int ret;
566 
567 	if (type != IB_MW_TYPE_1)
568 		return ERR_PTR(-EINVAL);
569 
570 	php = to_c4iw_pd(pd);
571 	rhp = php->rhp;
572 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
573 	if (!mhp)
574 		return ERR_PTR(-ENOMEM);
575 	ret = allocate_window(&rhp->rdev, &stag, php->pdid);
576 	if (ret) {
577 		kfree(mhp);
578 		return ERR_PTR(ret);
579 	}
580 	mhp->rhp = rhp;
581 	mhp->attr.pdid = php->pdid;
582 	mhp->attr.type = FW_RI_STAG_MW;
583 	mhp->attr.stag = stag;
584 	mmid = (stag) >> 8;
585 	mhp->ibmw.rkey = stag;
586 	if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
587 		deallocate_window(&rhp->rdev, mhp->attr.stag);
588 		kfree(mhp);
589 		return ERR_PTR(-ENOMEM);
590 	}
591 	PDBG("%s mmid 0x%x mhp %p stag 0x%x\n", __func__, mmid, mhp, stag);
592 	return &(mhp->ibmw);
593 }
594 
595 int c4iw_dealloc_mw(struct ib_mw *mw)
596 {
597 	struct c4iw_dev *rhp;
598 	struct c4iw_mw *mhp;
599 	u32 mmid;
600 
601 	mhp = to_c4iw_mw(mw);
602 	rhp = mhp->rhp;
603 	mmid = (mw->rkey) >> 8;
604 	remove_handle(rhp, &rhp->mmidr, mmid);
605 	deallocate_window(&rhp->rdev, mhp->attr.stag);
606 	kfree(mhp);
607 	PDBG("%s ib_mw %p mmid 0x%x ptr %p\n", __func__, mw, mmid, mhp);
608 	return 0;
609 }
610 
611 struct ib_mr *c4iw_alloc_mr(struct ib_pd *pd,
612 			    enum ib_mr_type mr_type,
613 			    u32 max_num_sg)
614 {
615 	struct c4iw_dev *rhp;
616 	struct c4iw_pd *php;
617 	struct c4iw_mr *mhp;
618 	u32 mmid;
619 	u32 stag = 0;
620 	int ret = 0;
621 	int length = roundup(max_num_sg * sizeof(u64), 32);
622 
623 	php = to_c4iw_pd(pd);
624 	rhp = php->rhp;
625 
626 	if (mr_type != IB_MR_TYPE_MEM_REG ||
627 	    max_num_sg > t4_max_fr_depth(&rhp->rdev.lldi.ulptx_memwrite_dsgl &&
628 					 use_dsgl))
629 		return ERR_PTR(-EINVAL);
630 
631 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
632 	if (!mhp) {
633 		ret = -ENOMEM;
634 		goto err;
635 	}
636 
637 	mhp->mpl = dma_alloc_coherent(&rhp->rdev.lldi.pdev->dev,
638 				      length, &mhp->mpl_addr, GFP_KERNEL);
639 	if (!mhp->mpl) {
640 		ret = -ENOMEM;
641 		goto err_mpl;
642 	}
643 	mhp->max_mpl_len = length;
644 
645 	mhp->rhp = rhp;
646 	ret = alloc_pbl(mhp, max_num_sg);
647 	if (ret)
648 		goto err1;
649 	mhp->attr.pbl_size = max_num_sg;
650 	ret = allocate_stag(&rhp->rdev, &stag, php->pdid,
651 				 mhp->attr.pbl_size, mhp->attr.pbl_addr);
652 	if (ret)
653 		goto err2;
654 	mhp->attr.pdid = php->pdid;
655 	mhp->attr.type = FW_RI_STAG_NSMR;
656 	mhp->attr.stag = stag;
657 	mhp->attr.state = 1;
658 	mmid = (stag) >> 8;
659 	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
660 	if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
661 		ret = -ENOMEM;
662 		goto err3;
663 	}
664 
665 	PDBG("%s mmid 0x%x mhp %p stag 0x%x\n", __func__, mmid, mhp, stag);
666 	return &(mhp->ibmr);
667 err3:
668 	dereg_mem(&rhp->rdev, stag, mhp->attr.pbl_size,
669 		       mhp->attr.pbl_addr);
670 err2:
671 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
672 			      mhp->attr.pbl_size << 3);
673 err1:
674 	dma_free_coherent(&mhp->rhp->rdev.lldi.pdev->dev,
675 			  mhp->max_mpl_len, mhp->mpl, mhp->mpl_addr);
676 err_mpl:
677 	kfree(mhp);
678 err:
679 	return ERR_PTR(ret);
680 }
681 
682 static int c4iw_set_page(struct ib_mr *ibmr, u64 addr)
683 {
684 	struct c4iw_mr *mhp = to_c4iw_mr(ibmr);
685 
686 	if (unlikely(mhp->mpl_len == mhp->max_mpl_len))
687 		return -ENOMEM;
688 
689 	mhp->mpl[mhp->mpl_len++] = addr;
690 
691 	return 0;
692 }
693 
694 int c4iw_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
695 		   unsigned int *sg_offset)
696 {
697 	struct c4iw_mr *mhp = to_c4iw_mr(ibmr);
698 
699 	mhp->mpl_len = 0;
700 
701 	return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, c4iw_set_page);
702 }
703 
704 int c4iw_dereg_mr(struct ib_mr *ib_mr)
705 {
706 	struct c4iw_dev *rhp;
707 	struct c4iw_mr *mhp;
708 	u32 mmid;
709 
710 	PDBG("%s ib_mr %p\n", __func__, ib_mr);
711 
712 	mhp = to_c4iw_mr(ib_mr);
713 	rhp = mhp->rhp;
714 	mmid = mhp->attr.stag >> 8;
715 	remove_handle(rhp, &rhp->mmidr, mmid);
716 	if (mhp->mpl)
717 		dma_free_coherent(&mhp->rhp->rdev.lldi.pdev->dev,
718 				  mhp->max_mpl_len, mhp->mpl, mhp->mpl_addr);
719 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
720 		       mhp->attr.pbl_addr);
721 	if (mhp->attr.pbl_size)
722 		c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
723 				  mhp->attr.pbl_size << 3);
724 	if (mhp->kva)
725 		kfree((void *) (unsigned long) mhp->kva);
726 	if (mhp->umem)
727 		ib_umem_release(mhp->umem);
728 	PDBG("%s mmid 0x%x ptr %p\n", __func__, mmid, mhp);
729 	kfree(mhp);
730 	return 0;
731 }
732