xref: /openbmc/linux/drivers/infiniband/hw/mlx5/mem.c (revision e7bae9bb)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. 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 <rdma/ib_umem.h>
35 #include <rdma/ib_umem_odp.h>
36 #include "mlx5_ib.h"
37 #include <linux/jiffies.h>
38 
39 /* @umem: umem object to scan
40  * @addr: ib virtual address requested by the user
41  * @max_page_shift: high limit for page_shift - 0 means no limit
42  * @count: number of PAGE_SIZE pages covered by umem
43  * @shift: page shift for the compound pages found in the region
44  * @ncont: number of compund pages
45  * @order: log2 of the number of compound pages
46  */
47 void mlx5_ib_cont_pages(struct ib_umem *umem, u64 addr,
48 			unsigned long max_page_shift,
49 			int *count, int *shift,
50 			int *ncont, int *order)
51 {
52 	unsigned long tmp;
53 	unsigned long m;
54 	u64 base = ~0, p = 0;
55 	u64 len, pfn;
56 	int i = 0;
57 	struct scatterlist *sg;
58 	int entry;
59 
60 	addr = addr >> PAGE_SHIFT;
61 	tmp = (unsigned long)addr;
62 	m = find_first_bit(&tmp, BITS_PER_LONG);
63 	if (max_page_shift)
64 		m = min_t(unsigned long, max_page_shift - PAGE_SHIFT, m);
65 
66 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
67 		len = sg_dma_len(sg) >> PAGE_SHIFT;
68 		pfn = sg_dma_address(sg) >> PAGE_SHIFT;
69 		if (base + p != pfn) {
70 			/* If either the offset or the new
71 			 * base are unaligned update m
72 			 */
73 			tmp = (unsigned long)(pfn | p);
74 			if (!IS_ALIGNED(tmp, 1 << m))
75 				m = find_first_bit(&tmp, BITS_PER_LONG);
76 
77 			base = pfn;
78 			p = 0;
79 		}
80 
81 		p += len;
82 		i += len;
83 	}
84 
85 	if (i) {
86 		m = min_t(unsigned long, ilog2(roundup_pow_of_two(i)), m);
87 
88 		if (order)
89 			*order = ilog2(roundup_pow_of_two(i) >> m);
90 
91 		*ncont = DIV_ROUND_UP(i, (1 << m));
92 	} else {
93 		m  = 0;
94 
95 		if (order)
96 			*order = 0;
97 
98 		*ncont = 0;
99 	}
100 	*shift = PAGE_SHIFT + m;
101 	*count = i;
102 }
103 
104 /*
105  * Populate the given array with bus addresses from the umem.
106  *
107  * dev - mlx5_ib device
108  * umem - umem to use to fill the pages
109  * page_shift - determines the page size used in the resulting array
110  * offset - offset into the umem to start from,
111  *          only implemented for ODP umems
112  * num_pages - total number of pages to fill
113  * pas - bus addresses array to fill
114  * access_flags - access flags to set on all present pages.
115 		  use enum mlx5_ib_mtt_access_flags for this.
116  */
117 void __mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem,
118 			    int page_shift, size_t offset, size_t num_pages,
119 			    __be64 *pas, int access_flags)
120 {
121 	int shift = page_shift - PAGE_SHIFT;
122 	int mask = (1 << shift) - 1;
123 	int i, k, idx;
124 	u64 cur = 0;
125 	u64 base;
126 	int len;
127 	struct scatterlist *sg;
128 	int entry;
129 
130 	i = 0;
131 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
132 		len = sg_dma_len(sg) >> PAGE_SHIFT;
133 		base = sg_dma_address(sg);
134 
135 		/* Skip elements below offset */
136 		if (i + len < offset << shift) {
137 			i += len;
138 			continue;
139 		}
140 
141 		/* Skip pages below offset */
142 		if (i < offset << shift) {
143 			k = (offset << shift) - i;
144 			i = offset << shift;
145 		} else {
146 			k = 0;
147 		}
148 
149 		for (; k < len; k++) {
150 			if (!(i & mask)) {
151 				cur = base + (k << PAGE_SHIFT);
152 				cur |= access_flags;
153 				idx = (i >> shift) - offset;
154 
155 				pas[idx] = cpu_to_be64(cur);
156 				mlx5_ib_dbg(dev, "pas[%d] 0x%llx\n",
157 					    i >> shift, be64_to_cpu(pas[idx]));
158 			}
159 			i++;
160 
161 			/* Stop after num_pages reached */
162 			if (i >> shift >= offset + num_pages)
163 				return;
164 		}
165 	}
166 }
167 
168 void mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem,
169 			  int page_shift, __be64 *pas, int access_flags)
170 {
171 	return __mlx5_ib_populate_pas(dev, umem, page_shift, 0,
172 				      ib_umem_num_pages(umem), pas,
173 				      access_flags);
174 }
175 int mlx5_ib_get_buf_offset(u64 addr, int page_shift, u32 *offset)
176 {
177 	u64 page_size;
178 	u64 page_mask;
179 	u64 off_size;
180 	u64 off_mask;
181 	u64 buf_off;
182 
183 	page_size = (u64)1 << page_shift;
184 	page_mask = page_size - 1;
185 	buf_off = addr & page_mask;
186 	off_size = page_size >> 6;
187 	off_mask = off_size - 1;
188 
189 	if (buf_off & off_mask)
190 		return -EINVAL;
191 
192 	*offset = buf_off >> ilog2(off_size);
193 	return 0;
194 }
195 
196 #define WR_ID_BF 0xBF
197 #define WR_ID_END 0xBAD
198 #define TEST_WC_NUM_WQES 255
199 #define TEST_WC_POLLING_MAX_TIME_JIFFIES msecs_to_jiffies(100)
200 static int post_send_nop(struct mlx5_ib_dev *dev, struct ib_qp *ibqp, u64 wr_id,
201 			 bool signaled)
202 {
203 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
204 	struct mlx5_wqe_ctrl_seg *ctrl;
205 	struct mlx5_bf *bf = &qp->bf;
206 	__be32 mmio_wqe[16] = {};
207 	unsigned long flags;
208 	unsigned int idx;
209 	int i;
210 
211 	if (unlikely(dev->mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR))
212 		return -EIO;
213 
214 	spin_lock_irqsave(&qp->sq.lock, flags);
215 
216 	idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
217 	ctrl = mlx5_frag_buf_get_wqe(&qp->sq.fbc, idx);
218 
219 	memset(ctrl, 0, sizeof(struct mlx5_wqe_ctrl_seg));
220 	ctrl->fm_ce_se = signaled ? MLX5_WQE_CTRL_CQ_UPDATE : 0;
221 	ctrl->opmod_idx_opcode =
222 		cpu_to_be32(((u32)(qp->sq.cur_post) << 8) | MLX5_OPCODE_NOP);
223 	ctrl->qpn_ds = cpu_to_be32((sizeof(struct mlx5_wqe_ctrl_seg) / 16) |
224 				   (qp->trans_qp.base.mqp.qpn << 8));
225 
226 	qp->sq.wrid[idx] = wr_id;
227 	qp->sq.w_list[idx].opcode = MLX5_OPCODE_NOP;
228 	qp->sq.wqe_head[idx] = qp->sq.head + 1;
229 	qp->sq.cur_post += DIV_ROUND_UP(sizeof(struct mlx5_wqe_ctrl_seg),
230 					MLX5_SEND_WQE_BB);
231 	qp->sq.w_list[idx].next = qp->sq.cur_post;
232 	qp->sq.head++;
233 
234 	memcpy(mmio_wqe, ctrl, sizeof(*ctrl));
235 	((struct mlx5_wqe_ctrl_seg *)&mmio_wqe)->fm_ce_se |=
236 		MLX5_WQE_CTRL_CQ_UPDATE;
237 
238 	/* Make sure that descriptors are written before
239 	 * updating doorbell record and ringing the doorbell
240 	 */
241 	wmb();
242 
243 	qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
244 
245 	/* Make sure doorbell record is visible to the HCA before
246 	 * we hit doorbell
247 	 */
248 	wmb();
249 	for (i = 0; i < 8; i++)
250 		mlx5_write64(&mmio_wqe[i * 2],
251 			     bf->bfreg->map + bf->offset + i * 8);
252 
253 	bf->offset ^= bf->buf_size;
254 
255 	spin_unlock_irqrestore(&qp->sq.lock, flags);
256 
257 	return 0;
258 }
259 
260 static int test_wc_poll_cq_result(struct mlx5_ib_dev *dev, struct ib_cq *cq)
261 {
262 	int ret;
263 	struct ib_wc wc = {};
264 	unsigned long end = jiffies + TEST_WC_POLLING_MAX_TIME_JIFFIES;
265 
266 	do {
267 		ret = ib_poll_cq(cq, 1, &wc);
268 		if (ret < 0 || wc.status)
269 			return ret < 0 ? ret : -EINVAL;
270 		if (ret)
271 			break;
272 	} while (!time_after(jiffies, end));
273 
274 	if (!ret)
275 		return -ETIMEDOUT;
276 
277 	if (wc.wr_id != WR_ID_BF)
278 		ret = 0;
279 
280 	return ret;
281 }
282 
283 static int test_wc_do_send(struct mlx5_ib_dev *dev, struct ib_qp *qp)
284 {
285 	int err, i;
286 
287 	for (i = 0; i < TEST_WC_NUM_WQES; i++) {
288 		err = post_send_nop(dev, qp, WR_ID_BF, false);
289 		if (err)
290 			return err;
291 	}
292 
293 	return post_send_nop(dev, qp, WR_ID_END, true);
294 }
295 
296 int mlx5_ib_test_wc(struct mlx5_ib_dev *dev)
297 {
298 	struct ib_cq_init_attr cq_attr = { .cqe = TEST_WC_NUM_WQES + 1 };
299 	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
300 	struct ib_qp_init_attr qp_init_attr = {
301 		.cap = { .max_send_wr = TEST_WC_NUM_WQES },
302 		.qp_type = IB_QPT_UD,
303 		.sq_sig_type = IB_SIGNAL_REQ_WR,
304 		.create_flags = MLX5_IB_QP_CREATE_WC_TEST,
305 	};
306 	struct ib_qp_attr qp_attr = { .port_num = 1 };
307 	struct ib_device *ibdev = &dev->ib_dev;
308 	struct ib_qp *qp;
309 	struct ib_cq *cq;
310 	struct ib_pd *pd;
311 	int ret;
312 
313 	if (!MLX5_CAP_GEN(dev->mdev, bf))
314 		return 0;
315 
316 	if (!dev->mdev->roce.roce_en &&
317 	    port_type_cap == MLX5_CAP_PORT_TYPE_ETH) {
318 		if (mlx5_core_is_pf(dev->mdev))
319 			dev->wc_support = arch_can_pci_mmap_wc();
320 		return 0;
321 	}
322 
323 	ret = mlx5_alloc_bfreg(dev->mdev, &dev->wc_bfreg, true, false);
324 	if (ret)
325 		goto print_err;
326 
327 	if (!dev->wc_bfreg.wc)
328 		goto out1;
329 
330 	pd = ib_alloc_pd(ibdev, 0);
331 	if (IS_ERR(pd)) {
332 		ret = PTR_ERR(pd);
333 		goto out1;
334 	}
335 
336 	cq = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr);
337 	if (IS_ERR(cq)) {
338 		ret = PTR_ERR(cq);
339 		goto out2;
340 	}
341 
342 	qp_init_attr.recv_cq = cq;
343 	qp_init_attr.send_cq = cq;
344 	qp = ib_create_qp(pd, &qp_init_attr);
345 	if (IS_ERR(qp)) {
346 		ret = PTR_ERR(qp);
347 		goto out3;
348 	}
349 
350 	qp_attr.qp_state = IB_QPS_INIT;
351 	ret = ib_modify_qp(qp, &qp_attr,
352 			   IB_QP_STATE | IB_QP_PORT | IB_QP_PKEY_INDEX |
353 				   IB_QP_QKEY);
354 	if (ret)
355 		goto out4;
356 
357 	qp_attr.qp_state = IB_QPS_RTR;
358 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
359 	if (ret)
360 		goto out4;
361 
362 	qp_attr.qp_state = IB_QPS_RTS;
363 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
364 	if (ret)
365 		goto out4;
366 
367 	ret = test_wc_do_send(dev, qp);
368 	if (ret < 0)
369 		goto out4;
370 
371 	ret = test_wc_poll_cq_result(dev, cq);
372 	if (ret > 0) {
373 		dev->wc_support = true;
374 		ret = 0;
375 	}
376 
377 out4:
378 	ib_destroy_qp(qp);
379 out3:
380 	ib_destroy_cq(cq);
381 out2:
382 	ib_dealloc_pd(pd);
383 out1:
384 	mlx5_free_bfreg(dev->mdev, &dev->wc_bfreg);
385 print_err:
386 	if (ret)
387 		mlx5_ib_err(
388 			dev,
389 			"Error %d while trying to test write-combining support\n",
390 			ret);
391 	return ret;
392 }
393