xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/mem.c (revision b96fc2f3)
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 = 0;
41 module_param(use_dsgl, int, 0644);
42 MODULE_PARM_DESC(use_dsgl, "Use DSGL for PBL/FastReg (default=0)");
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 mr_exceeds_hw_limits(struct c4iw_dev *dev, u64 length)
54 {
55 	return (is_t4(dev->rdev.lldi.adapter_type) ||
56 		is_t5(dev->rdev.lldi.adapter_type)) &&
57 		length >= 8*1024*1024*1024ULL;
58 }
59 
60 static int _c4iw_write_mem_dma_aligned(struct c4iw_rdev *rdev, u32 addr,
61 				       u32 len, dma_addr_t data, int wait)
62 {
63 	struct sk_buff *skb;
64 	struct ulp_mem_io *req;
65 	struct ulptx_sgl *sgl;
66 	u8 wr_len;
67 	int ret = 0;
68 	struct c4iw_wr_wait wr_wait;
69 
70 	addr &= 0x7FFFFFF;
71 
72 	if (wait)
73 		c4iw_init_wr_wait(&wr_wait);
74 	wr_len = roundup(sizeof(*req) + sizeof(*sgl), 16);
75 
76 	skb = alloc_skb(wr_len, GFP_KERNEL);
77 	if (!skb)
78 		return -ENOMEM;
79 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
80 
81 	req = (struct ulp_mem_io *)__skb_put(skb, wr_len);
82 	memset(req, 0, wr_len);
83 	INIT_ULPTX_WR(req, wr_len, 0, 0);
84 	req->wr.wr_hi = cpu_to_be32(FW_WR_OP_V(FW_ULPTX_WR) |
85 			(wait ? FW_WR_COMPL_F : 0));
86 	req->wr.wr_lo = wait ? (__force __be64)(unsigned long) &wr_wait : 0L;
87 	req->wr.wr_mid = cpu_to_be32(FW_WR_LEN16_V(DIV_ROUND_UP(wr_len, 16)));
88 	req->cmd = cpu_to_be32(ULPTX_CMD_V(ULP_TX_MEM_WRITE));
89 	req->cmd |= cpu_to_be32(T5_ULP_MEMIO_ORDER_V(1));
90 	req->dlen = cpu_to_be32(ULP_MEMIO_DATA_LEN_V(len>>5));
91 	req->len16 = cpu_to_be32(DIV_ROUND_UP(wr_len-sizeof(req->wr), 16));
92 	req->lock_addr = cpu_to_be32(ULP_MEMIO_ADDR_V(addr));
93 
94 	sgl = (struct ulptx_sgl *)(req + 1);
95 	sgl->cmd_nsge = cpu_to_be32(ULPTX_CMD_V(ULP_TX_SC_DSGL) |
96 				    ULPTX_NSGE_V(1));
97 	sgl->len0 = cpu_to_be32(len);
98 	sgl->addr0 = cpu_to_be64(data);
99 
100 	ret = c4iw_ofld_send(rdev, skb);
101 	if (ret)
102 		return ret;
103 	if (wait)
104 		ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
105 	return ret;
106 }
107 
108 static int _c4iw_write_mem_inline(struct c4iw_rdev *rdev, u32 addr, u32 len,
109 				  void *data)
110 {
111 	struct sk_buff *skb;
112 	struct ulp_mem_io *req;
113 	struct ulptx_idata *sc;
114 	u8 wr_len, *to_dp, *from_dp;
115 	int copy_len, num_wqe, i, ret = 0;
116 	struct c4iw_wr_wait wr_wait;
117 	__be32 cmd = cpu_to_be32(ULPTX_CMD_V(ULP_TX_MEM_WRITE));
118 
119 	if (is_t4(rdev->lldi.adapter_type))
120 		cmd |= cpu_to_be32(ULP_MEMIO_ORDER_F);
121 	else
122 		cmd |= cpu_to_be32(T5_ULP_MEMIO_IMM_F);
123 
124 	addr &= 0x7FFFFFF;
125 	PDBG("%s addr 0x%x len %u\n", __func__, addr, len);
126 	num_wqe = DIV_ROUND_UP(len, C4IW_MAX_INLINE_SIZE);
127 	c4iw_init_wr_wait(&wr_wait);
128 	for (i = 0; i < num_wqe; i++) {
129 
130 		copy_len = len > C4IW_MAX_INLINE_SIZE ? C4IW_MAX_INLINE_SIZE :
131 			   len;
132 		wr_len = roundup(sizeof *req + sizeof *sc +
133 				 roundup(copy_len, T4_ULPTX_MIN_IO), 16);
134 
135 		skb = alloc_skb(wr_len, GFP_KERNEL);
136 		if (!skb)
137 			return -ENOMEM;
138 		set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
139 
140 		req = (struct ulp_mem_io *)__skb_put(skb, wr_len);
141 		memset(req, 0, wr_len);
142 		INIT_ULPTX_WR(req, wr_len, 0, 0);
143 
144 		if (i == (num_wqe-1)) {
145 			req->wr.wr_hi = cpu_to_be32(FW_WR_OP_V(FW_ULPTX_WR) |
146 						    FW_WR_COMPL_F);
147 			req->wr.wr_lo = (__force __be64)&wr_wait;
148 		} else
149 			req->wr.wr_hi = cpu_to_be32(FW_WR_OP_V(FW_ULPTX_WR));
150 		req->wr.wr_mid = cpu_to_be32(
151 				       FW_WR_LEN16_V(DIV_ROUND_UP(wr_len, 16)));
152 
153 		req->cmd = cmd;
154 		req->dlen = cpu_to_be32(ULP_MEMIO_DATA_LEN_V(
155 				DIV_ROUND_UP(copy_len, T4_ULPTX_MIN_IO)));
156 		req->len16 = cpu_to_be32(DIV_ROUND_UP(wr_len-sizeof(req->wr),
157 						      16));
158 		req->lock_addr = cpu_to_be32(ULP_MEMIO_ADDR_V(addr + i * 3));
159 
160 		sc = (struct ulptx_idata *)(req + 1);
161 		sc->cmd_more = cpu_to_be32(ULPTX_CMD_V(ULP_TX_SC_IMM));
162 		sc->len = cpu_to_be32(roundup(copy_len, T4_ULPTX_MIN_IO));
163 
164 		to_dp = (u8 *)(sc + 1);
165 		from_dp = (u8 *)data + i * C4IW_MAX_INLINE_SIZE;
166 		if (data)
167 			memcpy(to_dp, from_dp, copy_len);
168 		else
169 			memset(to_dp, 0, copy_len);
170 		if (copy_len % T4_ULPTX_MIN_IO)
171 			memset(to_dp + copy_len, 0, T4_ULPTX_MIN_IO -
172 			       (copy_len % T4_ULPTX_MIN_IO));
173 		ret = c4iw_ofld_send(rdev, skb);
174 		if (ret)
175 			return ret;
176 		len -= C4IW_MAX_INLINE_SIZE;
177 	}
178 
179 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
180 	return ret;
181 }
182 
183 static int _c4iw_write_mem_dma(struct c4iw_rdev *rdev, u32 addr, u32 len, void *data)
184 {
185 	u32 remain = len;
186 	u32 dmalen;
187 	int ret = 0;
188 	dma_addr_t daddr;
189 	dma_addr_t save;
190 
191 	daddr = dma_map_single(&rdev->lldi.pdev->dev, data, len, DMA_TO_DEVICE);
192 	if (dma_mapping_error(&rdev->lldi.pdev->dev, daddr))
193 		return -1;
194 	save = daddr;
195 
196 	while (remain > inline_threshold) {
197 		if (remain < T4_ULPTX_MAX_DMA) {
198 			if (remain & ~T4_ULPTX_MIN_IO)
199 				dmalen = remain & ~(T4_ULPTX_MIN_IO-1);
200 			else
201 				dmalen = remain;
202 		} else
203 			dmalen = T4_ULPTX_MAX_DMA;
204 		remain -= dmalen;
205 		ret = _c4iw_write_mem_dma_aligned(rdev, addr, dmalen, daddr,
206 						 !remain);
207 		if (ret)
208 			goto out;
209 		addr += dmalen >> 5;
210 		data += dmalen;
211 		daddr += dmalen;
212 	}
213 	if (remain)
214 		ret = _c4iw_write_mem_inline(rdev, addr, remain, data);
215 out:
216 	dma_unmap_single(&rdev->lldi.pdev->dev, save, len, DMA_TO_DEVICE);
217 	return ret;
218 }
219 
220 /*
221  * write len bytes of data into addr (32B aligned address)
222  * If data is NULL, clear len byte of memory to zero.
223  */
224 static int write_adapter_mem(struct c4iw_rdev *rdev, u32 addr, u32 len,
225 			     void *data)
226 {
227 	if (is_t5(rdev->lldi.adapter_type) && use_dsgl) {
228 		if (len > inline_threshold) {
229 			if (_c4iw_write_mem_dma(rdev, addr, len, data)) {
230 				printk_ratelimited(KERN_WARNING
231 						   "%s: dma map"
232 						   " failure (non fatal)\n",
233 						   pci_name(rdev->lldi.pdev));
234 				return _c4iw_write_mem_inline(rdev, addr, len,
235 							      data);
236 			} else
237 				return 0;
238 		} else
239 			return _c4iw_write_mem_inline(rdev, addr, len, data);
240 	} else
241 		return _c4iw_write_mem_inline(rdev, addr, len, data);
242 }
243 
244 /*
245  * Build and write a TPT entry.
246  * IN: stag key, pdid, perm, bind_enabled, zbva, to, len, page_size,
247  *     pbl_size and pbl_addr
248  * OUT: stag index
249  */
250 static int write_tpt_entry(struct c4iw_rdev *rdev, u32 reset_tpt_entry,
251 			   u32 *stag, u8 stag_state, u32 pdid,
252 			   enum fw_ri_stag_type type, enum fw_ri_mem_perms perm,
253 			   int bind_enabled, u32 zbva, u64 to,
254 			   u64 len, u8 page_size, u32 pbl_size, u32 pbl_addr)
255 {
256 	int err;
257 	struct fw_ri_tpte tpt;
258 	u32 stag_idx;
259 	static atomic_t key;
260 
261 	if (c4iw_fatal_error(rdev))
262 		return -EIO;
263 
264 	stag_state = stag_state > 0;
265 	stag_idx = (*stag) >> 8;
266 
267 	if ((!reset_tpt_entry) && (*stag == T4_STAG_UNSET)) {
268 		stag_idx = c4iw_get_resource(&rdev->resource.tpt_table);
269 		if (!stag_idx) {
270 			mutex_lock(&rdev->stats.lock);
271 			rdev->stats.stag.fail++;
272 			mutex_unlock(&rdev->stats.lock);
273 			return -ENOMEM;
274 		}
275 		mutex_lock(&rdev->stats.lock);
276 		rdev->stats.stag.cur += 32;
277 		if (rdev->stats.stag.cur > rdev->stats.stag.max)
278 			rdev->stats.stag.max = rdev->stats.stag.cur;
279 		mutex_unlock(&rdev->stats.lock);
280 		*stag = (stag_idx << 8) | (atomic_inc_return(&key) & 0xff);
281 	}
282 	PDBG("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
283 	     __func__, stag_state, type, pdid, stag_idx);
284 
285 	/* write TPT entry */
286 	if (reset_tpt_entry)
287 		memset(&tpt, 0, sizeof(tpt));
288 	else {
289 		tpt.valid_to_pdid = cpu_to_be32(FW_RI_TPTE_VALID_F |
290 			FW_RI_TPTE_STAGKEY_V((*stag & FW_RI_TPTE_STAGKEY_M)) |
291 			FW_RI_TPTE_STAGSTATE_V(stag_state) |
292 			FW_RI_TPTE_STAGTYPE_V(type) | FW_RI_TPTE_PDID_V(pdid));
293 		tpt.locread_to_qpid = cpu_to_be32(FW_RI_TPTE_PERM_V(perm) |
294 			(bind_enabled ? FW_RI_TPTE_MWBINDEN_F : 0) |
295 			FW_RI_TPTE_ADDRTYPE_V((zbva ? FW_RI_ZERO_BASED_TO :
296 						      FW_RI_VA_BASED_TO))|
297 			FW_RI_TPTE_PS_V(page_size));
298 		tpt.nosnoop_pbladdr = !pbl_size ? 0 : cpu_to_be32(
299 			FW_RI_TPTE_PBLADDR_V(PBL_OFF(rdev, pbl_addr)>>3));
300 		tpt.len_lo = cpu_to_be32((u32)(len & 0xffffffffUL));
301 		tpt.va_hi = cpu_to_be32((u32)(to >> 32));
302 		tpt.va_lo_fbo = cpu_to_be32((u32)(to & 0xffffffffUL));
303 		tpt.dca_mwbcnt_pstag = cpu_to_be32(0);
304 		tpt.len_hi = cpu_to_be32((u32)(len >> 32));
305 	}
306 	err = write_adapter_mem(rdev, stag_idx +
307 				(rdev->lldi.vr->stag.start >> 5),
308 				sizeof(tpt), &tpt);
309 
310 	if (reset_tpt_entry) {
311 		c4iw_put_resource(&rdev->resource.tpt_table, stag_idx);
312 		mutex_lock(&rdev->stats.lock);
313 		rdev->stats.stag.cur -= 32;
314 		mutex_unlock(&rdev->stats.lock);
315 	}
316 	return err;
317 }
318 
319 static int write_pbl(struct c4iw_rdev *rdev, __be64 *pbl,
320 		     u32 pbl_addr, u32 pbl_size)
321 {
322 	int err;
323 
324 	PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
325 	     __func__, pbl_addr, rdev->lldi.vr->pbl.start,
326 	     pbl_size);
327 
328 	err = write_adapter_mem(rdev, pbl_addr >> 5, pbl_size << 3, pbl);
329 	return err;
330 }
331 
332 static int dereg_mem(struct c4iw_rdev *rdev, u32 stag, u32 pbl_size,
333 		     u32 pbl_addr)
334 {
335 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0,
336 			       pbl_size, pbl_addr);
337 }
338 
339 static int allocate_window(struct c4iw_rdev *rdev, u32 * stag, u32 pdid)
340 {
341 	*stag = T4_STAG_UNSET;
342 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_MW, 0, 0, 0,
343 			       0UL, 0, 0, 0, 0);
344 }
345 
346 static int deallocate_window(struct c4iw_rdev *rdev, u32 stag)
347 {
348 	return write_tpt_entry(rdev, 1, &stag, 0, 0, 0, 0, 0, 0, 0UL, 0, 0, 0,
349 			       0);
350 }
351 
352 static int allocate_stag(struct c4iw_rdev *rdev, u32 *stag, u32 pdid,
353 			 u32 pbl_size, u32 pbl_addr)
354 {
355 	*stag = T4_STAG_UNSET;
356 	return write_tpt_entry(rdev, 0, stag, 0, pdid, FW_RI_STAG_NSMR, 0, 0, 0,
357 			       0UL, 0, 0, pbl_size, pbl_addr);
358 }
359 
360 static int finish_mem_reg(struct c4iw_mr *mhp, u32 stag)
361 {
362 	u32 mmid;
363 
364 	mhp->attr.state = 1;
365 	mhp->attr.stag = stag;
366 	mmid = stag >> 8;
367 	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
368 	PDBG("%s mmid 0x%x mhp %p\n", __func__, mmid, mhp);
369 	return insert_handle(mhp->rhp, &mhp->rhp->mmidr, mhp, mmid);
370 }
371 
372 static int register_mem(struct c4iw_dev *rhp, struct c4iw_pd *php,
373 		      struct c4iw_mr *mhp, int shift)
374 {
375 	u32 stag = T4_STAG_UNSET;
376 	int ret;
377 
378 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, mhp->attr.pdid,
379 			      FW_RI_STAG_NSMR, mhp->attr.len ?
380 			      mhp->attr.perms : 0,
381 			      mhp->attr.mw_bind_enable, mhp->attr.zbva,
382 			      mhp->attr.va_fbo, mhp->attr.len ?
383 			      mhp->attr.len : -1, shift - 12,
384 			      mhp->attr.pbl_size, mhp->attr.pbl_addr);
385 	if (ret)
386 		return ret;
387 
388 	ret = finish_mem_reg(mhp, stag);
389 	if (ret)
390 		dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
391 		       mhp->attr.pbl_addr);
392 	return ret;
393 }
394 
395 static int reregister_mem(struct c4iw_dev *rhp, struct c4iw_pd *php,
396 			  struct c4iw_mr *mhp, int shift, int npages)
397 {
398 	u32 stag;
399 	int ret;
400 
401 	if (npages > mhp->attr.pbl_size)
402 		return -ENOMEM;
403 
404 	stag = mhp->attr.stag;
405 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, mhp->attr.pdid,
406 			      FW_RI_STAG_NSMR, mhp->attr.perms,
407 			      mhp->attr.mw_bind_enable, mhp->attr.zbva,
408 			      mhp->attr.va_fbo, mhp->attr.len, shift - 12,
409 			      mhp->attr.pbl_size, mhp->attr.pbl_addr);
410 	if (ret)
411 		return ret;
412 
413 	ret = finish_mem_reg(mhp, stag);
414 	if (ret)
415 		dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
416 		       mhp->attr.pbl_addr);
417 
418 	return ret;
419 }
420 
421 static int alloc_pbl(struct c4iw_mr *mhp, int npages)
422 {
423 	mhp->attr.pbl_addr = c4iw_pblpool_alloc(&mhp->rhp->rdev,
424 						    npages << 3);
425 
426 	if (!mhp->attr.pbl_addr)
427 		return -ENOMEM;
428 
429 	mhp->attr.pbl_size = npages;
430 
431 	return 0;
432 }
433 
434 static int build_phys_page_list(struct ib_phys_buf *buffer_list,
435 				int num_phys_buf, u64 *iova_start,
436 				u64 *total_size, int *npages,
437 				int *shift, __be64 **page_list)
438 {
439 	u64 mask;
440 	int i, j, n;
441 
442 	mask = 0;
443 	*total_size = 0;
444 	for (i = 0; i < num_phys_buf; ++i) {
445 		if (i != 0 && buffer_list[i].addr & ~PAGE_MASK)
446 			return -EINVAL;
447 		if (i != 0 && i != num_phys_buf - 1 &&
448 		    (buffer_list[i].size & ~PAGE_MASK))
449 			return -EINVAL;
450 		*total_size += buffer_list[i].size;
451 		if (i > 0)
452 			mask |= buffer_list[i].addr;
453 		else
454 			mask |= buffer_list[i].addr & PAGE_MASK;
455 		if (i != num_phys_buf - 1)
456 			mask |= buffer_list[i].addr + buffer_list[i].size;
457 		else
458 			mask |= (buffer_list[i].addr + buffer_list[i].size +
459 				PAGE_SIZE - 1) & PAGE_MASK;
460 	}
461 
462 	if (*total_size > 0xFFFFFFFFULL)
463 		return -ENOMEM;
464 
465 	/* Find largest page shift we can use to cover buffers */
466 	for (*shift = PAGE_SHIFT; *shift < 27; ++(*shift))
467 		if ((1ULL << *shift) & mask)
468 			break;
469 
470 	buffer_list[0].size += buffer_list[0].addr & ((1ULL << *shift) - 1);
471 	buffer_list[0].addr &= ~0ull << *shift;
472 
473 	*npages = 0;
474 	for (i = 0; i < num_phys_buf; ++i)
475 		*npages += (buffer_list[i].size +
476 			(1ULL << *shift) - 1) >> *shift;
477 
478 	if (!*npages)
479 		return -EINVAL;
480 
481 	*page_list = kmalloc(sizeof(u64) * *npages, GFP_KERNEL);
482 	if (!*page_list)
483 		return -ENOMEM;
484 
485 	n = 0;
486 	for (i = 0; i < num_phys_buf; ++i)
487 		for (j = 0;
488 		     j < (buffer_list[i].size + (1ULL << *shift) - 1) >> *shift;
489 		     ++j)
490 			(*page_list)[n++] = cpu_to_be64(buffer_list[i].addr +
491 			    ((u64) j << *shift));
492 
493 	PDBG("%s va 0x%llx mask 0x%llx shift %d len %lld pbl_size %d\n",
494 	     __func__, (unsigned long long)*iova_start,
495 	     (unsigned long long)mask, *shift, (unsigned long long)*total_size,
496 	     *npages);
497 
498 	return 0;
499 
500 }
501 
502 int c4iw_reregister_phys_mem(struct ib_mr *mr, int mr_rereg_mask,
503 			     struct ib_pd *pd, struct ib_phys_buf *buffer_list,
504 			     int num_phys_buf, int acc, u64 *iova_start)
505 {
506 
507 	struct c4iw_mr mh, *mhp;
508 	struct c4iw_pd *php;
509 	struct c4iw_dev *rhp;
510 	__be64 *page_list = NULL;
511 	int shift = 0;
512 	u64 total_size;
513 	int npages;
514 	int ret;
515 
516 	PDBG("%s ib_mr %p ib_pd %p\n", __func__, mr, pd);
517 
518 	/* There can be no memory windows */
519 	if (atomic_read(&mr->usecnt))
520 		return -EINVAL;
521 
522 	mhp = to_c4iw_mr(mr);
523 	rhp = mhp->rhp;
524 	php = to_c4iw_pd(mr->pd);
525 
526 	/* make sure we are on the same adapter */
527 	if (rhp != php->rhp)
528 		return -EINVAL;
529 
530 	memcpy(&mh, mhp, sizeof *mhp);
531 
532 	if (mr_rereg_mask & IB_MR_REREG_PD)
533 		php = to_c4iw_pd(pd);
534 	if (mr_rereg_mask & IB_MR_REREG_ACCESS) {
535 		mh.attr.perms = c4iw_ib_to_tpt_access(acc);
536 		mh.attr.mw_bind_enable = (acc & IB_ACCESS_MW_BIND) ==
537 					 IB_ACCESS_MW_BIND;
538 	}
539 	if (mr_rereg_mask & IB_MR_REREG_TRANS) {
540 		ret = build_phys_page_list(buffer_list, num_phys_buf,
541 						iova_start,
542 						&total_size, &npages,
543 						&shift, &page_list);
544 		if (ret)
545 			return ret;
546 	}
547 
548 	if (mr_exceeds_hw_limits(rhp, total_size)) {
549 		kfree(page_list);
550 		return -EINVAL;
551 	}
552 
553 	ret = reregister_mem(rhp, php, &mh, shift, npages);
554 	kfree(page_list);
555 	if (ret)
556 		return ret;
557 	if (mr_rereg_mask & IB_MR_REREG_PD)
558 		mhp->attr.pdid = php->pdid;
559 	if (mr_rereg_mask & IB_MR_REREG_ACCESS)
560 		mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
561 	if (mr_rereg_mask & IB_MR_REREG_TRANS) {
562 		mhp->attr.zbva = 0;
563 		mhp->attr.va_fbo = *iova_start;
564 		mhp->attr.page_size = shift - 12;
565 		mhp->attr.len = (u32) total_size;
566 		mhp->attr.pbl_size = npages;
567 	}
568 
569 	return 0;
570 }
571 
572 struct ib_mr *c4iw_register_phys_mem(struct ib_pd *pd,
573 				     struct ib_phys_buf *buffer_list,
574 				     int num_phys_buf, int acc, u64 *iova_start)
575 {
576 	__be64 *page_list;
577 	int shift;
578 	u64 total_size;
579 	int npages;
580 	struct c4iw_dev *rhp;
581 	struct c4iw_pd *php;
582 	struct c4iw_mr *mhp;
583 	int ret;
584 
585 	PDBG("%s ib_pd %p\n", __func__, pd);
586 	php = to_c4iw_pd(pd);
587 	rhp = php->rhp;
588 
589 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
590 	if (!mhp)
591 		return ERR_PTR(-ENOMEM);
592 
593 	mhp->rhp = rhp;
594 
595 	/* First check that we have enough alignment */
596 	if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK)) {
597 		ret = -EINVAL;
598 		goto err;
599 	}
600 
601 	if (num_phys_buf > 1 &&
602 	    ((buffer_list[0].addr + buffer_list[0].size) & ~PAGE_MASK)) {
603 		ret = -EINVAL;
604 		goto err;
605 	}
606 
607 	ret = build_phys_page_list(buffer_list, num_phys_buf, iova_start,
608 					&total_size, &npages, &shift,
609 					&page_list);
610 	if (ret)
611 		goto err;
612 
613 	if (mr_exceeds_hw_limits(rhp, total_size)) {
614 		kfree(page_list);
615 		ret = -EINVAL;
616 		goto err;
617 	}
618 
619 	ret = alloc_pbl(mhp, npages);
620 	if (ret) {
621 		kfree(page_list);
622 		goto err;
623 	}
624 
625 	ret = write_pbl(&mhp->rhp->rdev, page_list, mhp->attr.pbl_addr,
626 			     npages);
627 	kfree(page_list);
628 	if (ret)
629 		goto err_pbl;
630 
631 	mhp->attr.pdid = php->pdid;
632 	mhp->attr.zbva = 0;
633 
634 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
635 	mhp->attr.va_fbo = *iova_start;
636 	mhp->attr.page_size = shift - 12;
637 
638 	mhp->attr.len = (u32) total_size;
639 	mhp->attr.pbl_size = npages;
640 	ret = register_mem(rhp, php, mhp, shift);
641 	if (ret)
642 		goto err_pbl;
643 
644 	return &mhp->ibmr;
645 
646 err_pbl:
647 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
648 			      mhp->attr.pbl_size << 3);
649 
650 err:
651 	kfree(mhp);
652 	return ERR_PTR(ret);
653 
654 }
655 
656 struct ib_mr *c4iw_get_dma_mr(struct ib_pd *pd, int acc)
657 {
658 	struct c4iw_dev *rhp;
659 	struct c4iw_pd *php;
660 	struct c4iw_mr *mhp;
661 	int ret;
662 	u32 stag = T4_STAG_UNSET;
663 
664 	PDBG("%s ib_pd %p\n", __func__, pd);
665 	php = to_c4iw_pd(pd);
666 	rhp = php->rhp;
667 
668 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
669 	if (!mhp)
670 		return ERR_PTR(-ENOMEM);
671 
672 	mhp->rhp = rhp;
673 	mhp->attr.pdid = php->pdid;
674 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
675 	mhp->attr.mw_bind_enable = (acc&IB_ACCESS_MW_BIND) == IB_ACCESS_MW_BIND;
676 	mhp->attr.zbva = 0;
677 	mhp->attr.va_fbo = 0;
678 	mhp->attr.page_size = 0;
679 	mhp->attr.len = ~0ULL;
680 	mhp->attr.pbl_size = 0;
681 
682 	ret = write_tpt_entry(&rhp->rdev, 0, &stag, 1, php->pdid,
683 			      FW_RI_STAG_NSMR, mhp->attr.perms,
684 			      mhp->attr.mw_bind_enable, 0, 0, ~0ULL, 0, 0, 0);
685 	if (ret)
686 		goto err1;
687 
688 	ret = finish_mem_reg(mhp, stag);
689 	if (ret)
690 		goto err2;
691 	return &mhp->ibmr;
692 err2:
693 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
694 		  mhp->attr.pbl_addr);
695 err1:
696 	kfree(mhp);
697 	return ERR_PTR(ret);
698 }
699 
700 struct ib_mr *c4iw_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
701 			       u64 virt, int acc, struct ib_udata *udata)
702 {
703 	__be64 *pages;
704 	int shift, n, len;
705 	int i, k, entry;
706 	int err = 0;
707 	struct scatterlist *sg;
708 	struct c4iw_dev *rhp;
709 	struct c4iw_pd *php;
710 	struct c4iw_mr *mhp;
711 
712 	PDBG("%s ib_pd %p\n", __func__, pd);
713 
714 	if (length == ~0ULL)
715 		return ERR_PTR(-EINVAL);
716 
717 	if ((length + start) < start)
718 		return ERR_PTR(-EINVAL);
719 
720 	php = to_c4iw_pd(pd);
721 	rhp = php->rhp;
722 
723 	if (mr_exceeds_hw_limits(rhp, length))
724 		return ERR_PTR(-EINVAL);
725 
726 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
727 	if (!mhp)
728 		return ERR_PTR(-ENOMEM);
729 
730 	mhp->rhp = rhp;
731 
732 	mhp->umem = ib_umem_get(pd->uobject->context, start, length, acc, 0);
733 	if (IS_ERR(mhp->umem)) {
734 		err = PTR_ERR(mhp->umem);
735 		kfree(mhp);
736 		return ERR_PTR(err);
737 	}
738 
739 	shift = ffs(mhp->umem->page_size) - 1;
740 
741 	n = mhp->umem->nmap;
742 	err = alloc_pbl(mhp, n);
743 	if (err)
744 		goto err;
745 
746 	pages = (__be64 *) __get_free_page(GFP_KERNEL);
747 	if (!pages) {
748 		err = -ENOMEM;
749 		goto err_pbl;
750 	}
751 
752 	i = n = 0;
753 
754 	for_each_sg(mhp->umem->sg_head.sgl, sg, mhp->umem->nmap, entry) {
755 		len = sg_dma_len(sg) >> shift;
756 		for (k = 0; k < len; ++k) {
757 			pages[i++] = cpu_to_be64(sg_dma_address(sg) +
758 				mhp->umem->page_size * k);
759 			if (i == PAGE_SIZE / sizeof *pages) {
760 				err = write_pbl(&mhp->rhp->rdev,
761 				      pages,
762 				      mhp->attr.pbl_addr + (n << 3), i);
763 				if (err)
764 					goto pbl_done;
765 				n += i;
766 				i = 0;
767 			}
768 		}
769 	}
770 
771 	if (i)
772 		err = write_pbl(&mhp->rhp->rdev, pages,
773 				     mhp->attr.pbl_addr + (n << 3), i);
774 
775 pbl_done:
776 	free_page((unsigned long) pages);
777 	if (err)
778 		goto err_pbl;
779 
780 	mhp->attr.pdid = php->pdid;
781 	mhp->attr.zbva = 0;
782 	mhp->attr.perms = c4iw_ib_to_tpt_access(acc);
783 	mhp->attr.va_fbo = virt;
784 	mhp->attr.page_size = shift - 12;
785 	mhp->attr.len = length;
786 
787 	err = register_mem(rhp, php, mhp, shift);
788 	if (err)
789 		goto err_pbl;
790 
791 	return &mhp->ibmr;
792 
793 err_pbl:
794 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
795 			      mhp->attr.pbl_size << 3);
796 
797 err:
798 	ib_umem_release(mhp->umem);
799 	kfree(mhp);
800 	return ERR_PTR(err);
801 }
802 
803 struct ib_mw *c4iw_alloc_mw(struct ib_pd *pd, enum ib_mw_type type)
804 {
805 	struct c4iw_dev *rhp;
806 	struct c4iw_pd *php;
807 	struct c4iw_mw *mhp;
808 	u32 mmid;
809 	u32 stag = 0;
810 	int ret;
811 
812 	if (type != IB_MW_TYPE_1)
813 		return ERR_PTR(-EINVAL);
814 
815 	php = to_c4iw_pd(pd);
816 	rhp = php->rhp;
817 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
818 	if (!mhp)
819 		return ERR_PTR(-ENOMEM);
820 	ret = allocate_window(&rhp->rdev, &stag, php->pdid);
821 	if (ret) {
822 		kfree(mhp);
823 		return ERR_PTR(ret);
824 	}
825 	mhp->rhp = rhp;
826 	mhp->attr.pdid = php->pdid;
827 	mhp->attr.type = FW_RI_STAG_MW;
828 	mhp->attr.stag = stag;
829 	mmid = (stag) >> 8;
830 	mhp->ibmw.rkey = stag;
831 	if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
832 		deallocate_window(&rhp->rdev, mhp->attr.stag);
833 		kfree(mhp);
834 		return ERR_PTR(-ENOMEM);
835 	}
836 	PDBG("%s mmid 0x%x mhp %p stag 0x%x\n", __func__, mmid, mhp, stag);
837 	return &(mhp->ibmw);
838 }
839 
840 int c4iw_dealloc_mw(struct ib_mw *mw)
841 {
842 	struct c4iw_dev *rhp;
843 	struct c4iw_mw *mhp;
844 	u32 mmid;
845 
846 	mhp = to_c4iw_mw(mw);
847 	rhp = mhp->rhp;
848 	mmid = (mw->rkey) >> 8;
849 	remove_handle(rhp, &rhp->mmidr, mmid);
850 	deallocate_window(&rhp->rdev, mhp->attr.stag);
851 	kfree(mhp);
852 	PDBG("%s ib_mw %p mmid 0x%x ptr %p\n", __func__, mw, mmid, mhp);
853 	return 0;
854 }
855 
856 struct ib_mr *c4iw_alloc_mr(struct ib_pd *pd,
857 			    enum ib_mr_type mr_type,
858 			    u32 max_num_sg)
859 {
860 	struct c4iw_dev *rhp;
861 	struct c4iw_pd *php;
862 	struct c4iw_mr *mhp;
863 	u32 mmid;
864 	u32 stag = 0;
865 	int ret = 0;
866 
867 	if (mr_type != IB_MR_TYPE_MEM_REG ||
868 	    max_num_sg > t4_max_fr_depth(use_dsgl))
869 		return ERR_PTR(-EINVAL);
870 
871 	php = to_c4iw_pd(pd);
872 	rhp = php->rhp;
873 	mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
874 	if (!mhp) {
875 		ret = -ENOMEM;
876 		goto err;
877 	}
878 
879 	mhp->rhp = rhp;
880 	ret = alloc_pbl(mhp, max_num_sg);
881 	if (ret)
882 		goto err1;
883 	mhp->attr.pbl_size = max_num_sg;
884 	ret = allocate_stag(&rhp->rdev, &stag, php->pdid,
885 				 mhp->attr.pbl_size, mhp->attr.pbl_addr);
886 	if (ret)
887 		goto err2;
888 	mhp->attr.pdid = php->pdid;
889 	mhp->attr.type = FW_RI_STAG_NSMR;
890 	mhp->attr.stag = stag;
891 	mhp->attr.state = 1;
892 	mmid = (stag) >> 8;
893 	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
894 	if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
895 		ret = -ENOMEM;
896 		goto err3;
897 	}
898 
899 	PDBG("%s mmid 0x%x mhp %p stag 0x%x\n", __func__, mmid, mhp, stag);
900 	return &(mhp->ibmr);
901 err3:
902 	dereg_mem(&rhp->rdev, stag, mhp->attr.pbl_size,
903 		       mhp->attr.pbl_addr);
904 err2:
905 	c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
906 			      mhp->attr.pbl_size << 3);
907 err1:
908 	kfree(mhp);
909 err:
910 	return ERR_PTR(ret);
911 }
912 
913 struct ib_fast_reg_page_list *c4iw_alloc_fastreg_pbl(struct ib_device *device,
914 						     int page_list_len)
915 {
916 	struct c4iw_fr_page_list *c4pl;
917 	struct c4iw_dev *dev = to_c4iw_dev(device);
918 	dma_addr_t dma_addr;
919 	int pll_len = roundup(page_list_len * sizeof(u64), 32);
920 
921 	c4pl = kmalloc(sizeof(*c4pl), GFP_KERNEL);
922 	if (!c4pl)
923 		return ERR_PTR(-ENOMEM);
924 
925 	c4pl->ibpl.page_list = dma_alloc_coherent(&dev->rdev.lldi.pdev->dev,
926 						  pll_len, &dma_addr,
927 						  GFP_KERNEL);
928 	if (!c4pl->ibpl.page_list) {
929 		kfree(c4pl);
930 		return ERR_PTR(-ENOMEM);
931 	}
932 	dma_unmap_addr_set(c4pl, mapping, dma_addr);
933 	c4pl->dma_addr = dma_addr;
934 	c4pl->dev = dev;
935 	c4pl->pll_len = pll_len;
936 
937 	PDBG("%s c4pl %p pll_len %u page_list %p dma_addr %pad\n",
938 	     __func__, c4pl, c4pl->pll_len, c4pl->ibpl.page_list,
939 	     &c4pl->dma_addr);
940 
941 	return &c4pl->ibpl;
942 }
943 
944 void c4iw_free_fastreg_pbl(struct ib_fast_reg_page_list *ibpl)
945 {
946 	struct c4iw_fr_page_list *c4pl = to_c4iw_fr_page_list(ibpl);
947 
948 	PDBG("%s c4pl %p pll_len %u page_list %p dma_addr %pad\n",
949 	     __func__, c4pl, c4pl->pll_len, c4pl->ibpl.page_list,
950 	     &c4pl->dma_addr);
951 
952 	dma_free_coherent(&c4pl->dev->rdev.lldi.pdev->dev,
953 			  c4pl->pll_len,
954 			  c4pl->ibpl.page_list, dma_unmap_addr(c4pl, mapping));
955 	kfree(c4pl);
956 }
957 
958 int c4iw_dereg_mr(struct ib_mr *ib_mr)
959 {
960 	struct c4iw_dev *rhp;
961 	struct c4iw_mr *mhp;
962 	u32 mmid;
963 
964 	PDBG("%s ib_mr %p\n", __func__, ib_mr);
965 	/* There can be no memory windows */
966 	if (atomic_read(&ib_mr->usecnt))
967 		return -EINVAL;
968 
969 	mhp = to_c4iw_mr(ib_mr);
970 	rhp = mhp->rhp;
971 	mmid = mhp->attr.stag >> 8;
972 	remove_handle(rhp, &rhp->mmidr, mmid);
973 	dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
974 		       mhp->attr.pbl_addr);
975 	if (mhp->attr.pbl_size)
976 		c4iw_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
977 				  mhp->attr.pbl_size << 3);
978 	if (mhp->kva)
979 		kfree((void *) (unsigned long) mhp->kva);
980 	if (mhp->umem)
981 		ib_umem_release(mhp->umem);
982 	PDBG("%s mmid 0x%x ptr %p\n", __func__, mmid, mhp);
983 	kfree(mhp);
984 	return 0;
985 }
986