Loading...
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 <rdma/ib_umem_odp.h>
34#include "mlx5_ib.h"
35#include <linux/jiffies.h>
36
37/*
38 * Fill in a physical address list. ib_umem_num_dma_blocks() entries will be
39 * filled in the pas array.
40 */
41void mlx5_ib_populate_pas(struct ib_umem *umem, size_t page_size, __be64 *pas,
42 u64 access_flags)
43{
44 struct ib_block_iter biter;
45
46 rdma_umem_for_each_dma_block (umem, &biter, page_size) {
47 *pas = cpu_to_be64(rdma_block_iter_dma_address(&biter) |
48 access_flags);
49 pas++;
50 }
51}
52
53/*
54 * Compute the page shift and page_offset for mailboxes that use a quantized
55 * page_offset. The granulatity of the page offset scales according to page
56 * size.
57 */
58unsigned long __mlx5_umem_find_best_quantized_pgoff(
59 struct ib_umem *umem, unsigned long pgsz_bitmap,
60 unsigned int page_offset_bits, u64 pgoff_bitmask, unsigned int scale,
61 unsigned int *page_offset_quantized)
62{
63 const u64 page_offset_mask = (1UL << page_offset_bits) - 1;
64 unsigned long page_size;
65 u64 page_offset;
66
67 page_size = ib_umem_find_best_pgoff(umem, pgsz_bitmap, pgoff_bitmask);
68 if (!page_size)
69 return 0;
70
71 /*
72 * page size is the largest possible page size.
73 *
74 * Reduce the page_size, and thus the page_offset and quanta, until the
75 * page_offset fits into the mailbox field. Once page_size < scale this
76 * loop is guaranteed to terminate.
77 */
78 page_offset = ib_umem_dma_offset(umem, page_size);
79 while (page_offset & ~(u64)(page_offset_mask * (page_size / scale))) {
80 page_size /= 2;
81 page_offset = ib_umem_dma_offset(umem, page_size);
82 }
83
84 /*
85 * The address is not aligned, or otherwise cannot be represented by the
86 * page_offset.
87 */
88 if (!(pgsz_bitmap & page_size))
89 return 0;
90
91 *page_offset_quantized =
92 (unsigned long)page_offset / (page_size / scale);
93 if (WARN_ON(*page_offset_quantized > page_offset_mask))
94 return 0;
95 return page_size;
96}
97
98#define WR_ID_BF 0xBF
99#define WR_ID_END 0xBAD
100#define TEST_WC_NUM_WQES 255
101#define TEST_WC_POLLING_MAX_TIME_JIFFIES msecs_to_jiffies(100)
102static int post_send_nop(struct mlx5_ib_dev *dev, struct ib_qp *ibqp, u64 wr_id,
103 bool signaled)
104{
105 struct mlx5_ib_qp *qp = to_mqp(ibqp);
106 struct mlx5_wqe_ctrl_seg *ctrl;
107 struct mlx5_bf *bf = &qp->bf;
108 __be32 mmio_wqe[16] = {};
109 unsigned long flags;
110 unsigned int idx;
111 int i;
112
113 if (unlikely(dev->mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR))
114 return -EIO;
115
116 spin_lock_irqsave(&qp->sq.lock, flags);
117
118 idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
119 ctrl = mlx5_frag_buf_get_wqe(&qp->sq.fbc, idx);
120
121 memset(ctrl, 0, sizeof(struct mlx5_wqe_ctrl_seg));
122 ctrl->fm_ce_se = signaled ? MLX5_WQE_CTRL_CQ_UPDATE : 0;
123 ctrl->opmod_idx_opcode =
124 cpu_to_be32(((u32)(qp->sq.cur_post) << 8) | MLX5_OPCODE_NOP);
125 ctrl->qpn_ds = cpu_to_be32((sizeof(struct mlx5_wqe_ctrl_seg) / 16) |
126 (qp->trans_qp.base.mqp.qpn << 8));
127
128 qp->sq.wrid[idx] = wr_id;
129 qp->sq.w_list[idx].opcode = MLX5_OPCODE_NOP;
130 qp->sq.wqe_head[idx] = qp->sq.head + 1;
131 qp->sq.cur_post += DIV_ROUND_UP(sizeof(struct mlx5_wqe_ctrl_seg),
132 MLX5_SEND_WQE_BB);
133 qp->sq.w_list[idx].next = qp->sq.cur_post;
134 qp->sq.head++;
135
136 memcpy(mmio_wqe, ctrl, sizeof(*ctrl));
137 ((struct mlx5_wqe_ctrl_seg *)&mmio_wqe)->fm_ce_se |=
138 MLX5_WQE_CTRL_CQ_UPDATE;
139
140 /* Make sure that descriptors are written before
141 * updating doorbell record and ringing the doorbell
142 */
143 wmb();
144
145 qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
146
147 /* Make sure doorbell record is visible to the HCA before
148 * we hit doorbell
149 */
150 wmb();
151 for (i = 0; i < 8; i++)
152 mlx5_write64(&mmio_wqe[i * 2],
153 bf->bfreg->map + bf->offset + i * 8);
154 io_stop_wc();
155
156 bf->offset ^= bf->buf_size;
157
158 spin_unlock_irqrestore(&qp->sq.lock, flags);
159
160 return 0;
161}
162
163static int test_wc_poll_cq_result(struct mlx5_ib_dev *dev, struct ib_cq *cq)
164{
165 int ret;
166 struct ib_wc wc = {};
167 unsigned long end = jiffies + TEST_WC_POLLING_MAX_TIME_JIFFIES;
168
169 do {
170 ret = ib_poll_cq(cq, 1, &wc);
171 if (ret < 0 || wc.status)
172 return ret < 0 ? ret : -EINVAL;
173 if (ret)
174 break;
175 } while (!time_after(jiffies, end));
176
177 if (!ret)
178 return -ETIMEDOUT;
179
180 if (wc.wr_id != WR_ID_BF)
181 ret = 0;
182
183 return ret;
184}
185
186static int test_wc_do_send(struct mlx5_ib_dev *dev, struct ib_qp *qp)
187{
188 int err, i;
189
190 for (i = 0; i < TEST_WC_NUM_WQES; i++) {
191 err = post_send_nop(dev, qp, WR_ID_BF, false);
192 if (err)
193 return err;
194 }
195
196 return post_send_nop(dev, qp, WR_ID_END, true);
197}
198
199int mlx5_ib_test_wc(struct mlx5_ib_dev *dev)
200{
201 struct ib_cq_init_attr cq_attr = { .cqe = TEST_WC_NUM_WQES + 1 };
202 int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
203 struct ib_qp_init_attr qp_init_attr = {
204 .cap = { .max_send_wr = TEST_WC_NUM_WQES },
205 .qp_type = IB_QPT_UD,
206 .sq_sig_type = IB_SIGNAL_REQ_WR,
207 .create_flags = MLX5_IB_QP_CREATE_WC_TEST,
208 };
209 struct ib_qp_attr qp_attr = { .port_num = 1 };
210 struct ib_device *ibdev = &dev->ib_dev;
211 struct ib_qp *qp;
212 struct ib_cq *cq;
213 struct ib_pd *pd;
214 int ret;
215
216 if (!MLX5_CAP_GEN(dev->mdev, bf))
217 return 0;
218
219 if (!dev->mdev->roce.roce_en &&
220 port_type_cap == MLX5_CAP_PORT_TYPE_ETH) {
221 if (mlx5_core_is_pf(dev->mdev))
222 dev->wc_support = arch_can_pci_mmap_wc();
223 return 0;
224 }
225
226 ret = mlx5_alloc_bfreg(dev->mdev, &dev->wc_bfreg, true, false);
227 if (ret)
228 goto print_err;
229
230 if (!dev->wc_bfreg.wc)
231 goto out1;
232
233 pd = ib_alloc_pd(ibdev, 0);
234 if (IS_ERR(pd)) {
235 ret = PTR_ERR(pd);
236 goto out1;
237 }
238
239 cq = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr);
240 if (IS_ERR(cq)) {
241 ret = PTR_ERR(cq);
242 goto out2;
243 }
244
245 qp_init_attr.recv_cq = cq;
246 qp_init_attr.send_cq = cq;
247 qp = ib_create_qp(pd, &qp_init_attr);
248 if (IS_ERR(qp)) {
249 ret = PTR_ERR(qp);
250 goto out3;
251 }
252
253 qp_attr.qp_state = IB_QPS_INIT;
254 ret = ib_modify_qp(qp, &qp_attr,
255 IB_QP_STATE | IB_QP_PORT | IB_QP_PKEY_INDEX |
256 IB_QP_QKEY);
257 if (ret)
258 goto out4;
259
260 qp_attr.qp_state = IB_QPS_RTR;
261 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
262 if (ret)
263 goto out4;
264
265 qp_attr.qp_state = IB_QPS_RTS;
266 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
267 if (ret)
268 goto out4;
269
270 ret = test_wc_do_send(dev, qp);
271 if (ret < 0)
272 goto out4;
273
274 ret = test_wc_poll_cq_result(dev, cq);
275 if (ret > 0) {
276 dev->wc_support = true;
277 ret = 0;
278 }
279
280out4:
281 ib_destroy_qp(qp);
282out3:
283 ib_destroy_cq(cq);
284out2:
285 ib_dealloc_pd(pd);
286out1:
287 mlx5_free_bfreg(dev->mdev, &dev->wc_bfreg);
288print_err:
289 if (ret)
290 mlx5_ib_err(
291 dev,
292 "Error %d while trying to test write-combining support\n",
293 ret);
294 return ret;
295}
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
38/* @umem: umem object to scan
39 * @addr: ib virtual address requested by the user
40 * @count: number of PAGE_SIZE pages covered by umem
41 * @shift: page shift for the compound pages found in the region
42 * @ncont: number of compund pages
43 * @order: log2 of the number of compound pages
44 */
45void mlx5_ib_cont_pages(struct ib_umem *umem, u64 addr, int *count, int *shift,
46 int *ncont, int *order)
47{
48 unsigned long tmp;
49 unsigned long m;
50 int i, k;
51 u64 base = 0;
52 int p = 0;
53 int skip;
54 int mask;
55 u64 len;
56 u64 pfn;
57 struct scatterlist *sg;
58 int entry;
59 unsigned long page_shift = ilog2(umem->page_size);
60
61 /* With ODP we must always match OS page size. */
62 if (umem->odp_data) {
63 *count = ib_umem_page_count(umem);
64 *shift = PAGE_SHIFT;
65 *ncont = *count;
66 if (order)
67 *order = ilog2(roundup_pow_of_two(*count));
68
69 return;
70 }
71
72 addr = addr >> page_shift;
73 tmp = (unsigned long)addr;
74 m = find_first_bit(&tmp, sizeof(tmp));
75 skip = 1 << m;
76 mask = skip - 1;
77 i = 0;
78 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
79 len = sg_dma_len(sg) >> page_shift;
80 pfn = sg_dma_address(sg) >> page_shift;
81 for (k = 0; k < len; k++) {
82 if (!(i & mask)) {
83 tmp = (unsigned long)pfn;
84 m = min_t(unsigned long, m, find_first_bit(&tmp, sizeof(tmp)));
85 skip = 1 << m;
86 mask = skip - 1;
87 base = pfn;
88 p = 0;
89 } else {
90 if (base + p != pfn) {
91 tmp = (unsigned long)p;
92 m = find_first_bit(&tmp, sizeof(tmp));
93 skip = 1 << m;
94 mask = skip - 1;
95 base = pfn;
96 p = 0;
97 }
98 }
99 p++;
100 i++;
101 }
102 }
103
104 if (i) {
105 m = min_t(unsigned long, ilog2(roundup_pow_of_two(i)), m);
106
107 if (order)
108 *order = ilog2(roundup_pow_of_two(i) >> m);
109
110 *ncont = DIV_ROUND_UP(i, (1 << m));
111 } else {
112 m = 0;
113
114 if (order)
115 *order = 0;
116
117 *ncont = 0;
118 }
119 *shift = page_shift + m;
120 *count = i;
121}
122
123#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
124static u64 umem_dma_to_mtt(dma_addr_t umem_dma)
125{
126 u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK;
127
128 if (umem_dma & ODP_READ_ALLOWED_BIT)
129 mtt_entry |= MLX5_IB_MTT_READ;
130 if (umem_dma & ODP_WRITE_ALLOWED_BIT)
131 mtt_entry |= MLX5_IB_MTT_WRITE;
132
133 return mtt_entry;
134}
135#endif
136
137/*
138 * Populate the given array with bus addresses from the umem.
139 *
140 * dev - mlx5_ib device
141 * umem - umem to use to fill the pages
142 * page_shift - determines the page size used in the resulting array
143 * offset - offset into the umem to start from,
144 * only implemented for ODP umems
145 * num_pages - total number of pages to fill
146 * pas - bus addresses array to fill
147 * access_flags - access flags to set on all present pages.
148 use enum mlx5_ib_mtt_access_flags for this.
149 */
150void __mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem,
151 int page_shift, size_t offset, size_t num_pages,
152 __be64 *pas, int access_flags)
153{
154 unsigned long umem_page_shift = ilog2(umem->page_size);
155 int shift = page_shift - umem_page_shift;
156 int mask = (1 << shift) - 1;
157 int i, k;
158 u64 cur = 0;
159 u64 base;
160 int len;
161 struct scatterlist *sg;
162 int entry;
163#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
164 const bool odp = umem->odp_data != NULL;
165
166 if (odp) {
167 WARN_ON(shift != 0);
168 WARN_ON(access_flags != (MLX5_IB_MTT_READ | MLX5_IB_MTT_WRITE));
169
170 for (i = 0; i < num_pages; ++i) {
171 dma_addr_t pa = umem->odp_data->dma_list[offset + i];
172
173 pas[i] = cpu_to_be64(umem_dma_to_mtt(pa));
174 }
175 return;
176 }
177#endif
178
179 i = 0;
180 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
181 len = sg_dma_len(sg) >> umem_page_shift;
182 base = sg_dma_address(sg);
183 for (k = 0; k < len; k++) {
184 if (!(i & mask)) {
185 cur = base + (k << umem_page_shift);
186 cur |= access_flags;
187
188 pas[i >> shift] = cpu_to_be64(cur);
189 mlx5_ib_dbg(dev, "pas[%d] 0x%llx\n",
190 i >> shift, be64_to_cpu(pas[i >> shift]));
191 } else
192 mlx5_ib_dbg(dev, "=====> 0x%llx\n",
193 base + (k << umem_page_shift));
194 i++;
195 }
196 }
197}
198
199void mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem,
200 int page_shift, __be64 *pas, int access_flags)
201{
202 return __mlx5_ib_populate_pas(dev, umem, page_shift, 0,
203 ib_umem_num_pages(umem), pas,
204 access_flags);
205}
206int mlx5_ib_get_buf_offset(u64 addr, int page_shift, u32 *offset)
207{
208 u64 page_size;
209 u64 page_mask;
210 u64 off_size;
211 u64 off_mask;
212 u64 buf_off;
213
214 page_size = (u64)1 << page_shift;
215 page_mask = page_size - 1;
216 buf_off = addr & page_mask;
217 off_size = page_size >> 6;
218 off_mask = off_size - 1;
219
220 if (buf_off & off_mask)
221 return -EINVAL;
222
223 *offset = buf_off >> ilog2(off_size);
224 return 0;
225}