update kernel to 5.10.198

hc

2024-01-03 2f7c68cb55ecb7331f2381deb497c27155f32faf

 kernel/drivers/infiniband/hw/hns/hns_roce_alloc.c
..
..
@@ -34,6 +34,7 @@
34
34
 #include <linux/platform_device.h>
35
35
 #include <linux/vmalloc.h>
36
36
 #include "hns_roce_device.h"
37
+#include <rdma/ib_umem.h>
37
38
 
38
39
 int hns_roce_bitmap_alloc(struct hns_roce_bitmap *bitmap, unsigned long *obj)
39
40
 {
..
..
@@ -54,7 +55,7 @@
54
55
 			bitmap->last = 0;
55
56
 		*obj |= bitmap->top;
56
57
 	} else {
57
-		ret = -1;
58
+		ret = -EINVAL;
58
59
 	}
59
60
 
60
61
 	spin_unlock(&bitmap->lock);
..
..
@@ -67,7 +68,6 @@
67
68
 {
68
69
 	hns_roce_bitmap_free_range(bitmap, obj, 1, rr);
69
70
 }
70
-EXPORT_SYMBOL_GPL(hns_roce_bitmap_free);
71
71
 
72
72
 int hns_roce_bitmap_alloc_range(struct hns_roce_bitmap *bitmap, int cnt,
73
73
 				int align, unsigned long *obj)
..
..
@@ -100,7 +100,7 @@
100
100
 		}
101
101
 		*obj |= bitmap->top;
102
102
 	} else {
103
-		ret = -1;
103
+		ret = -EINVAL;
104
104
 	}
105
105
 
106
106
 	spin_unlock(&bitmap->lock);
..
..
@@ -157,88 +157,135 @@
157
157
 	kfree(bitmap->table);
158
158
 }
159
159
 
160
-void hns_roce_buf_free(struct hns_roce_dev *hr_dev, u32 size,
161
-		       struct hns_roce_buf *buf)
160
+void hns_roce_buf_free(struct hns_roce_dev *hr_dev, struct hns_roce_buf *buf)
162
161
 {
163
-	int i;
164
162
 	struct device *dev = hr_dev->dev;
163
+	u32 size = buf->size;
164
+	int i;
165
165
 
166
-	if (buf->nbufs == 1) {
166
+	if (size == 0)
167
+		return;
168
+
169
+	buf->size = 0;
170
+
171
+	if (hns_roce_buf_is_direct(buf)) {
167
172
 		dma_free_coherent(dev, size, buf->direct.buf, buf->direct.map);
168
173
 	} else {
169
-		for (i = 0; i < buf->nbufs; ++i)
174
+		for (i = 0; i < buf->npages; ++i)
170
175
 			if (buf->page_list[i].buf)
171
176
 				dma_free_coherent(dev, 1 << buf->page_shift,
172
177
 						  buf->page_list[i].buf,
173
178
 						  buf->page_list[i].map);
174
179
 		kfree(buf->page_list);
180
+		buf->page_list = NULL;
175
181
 	}
176
182
 }
177
-EXPORT_SYMBOL_GPL(hns_roce_buf_free);
178
183
 
179
184
 int hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size, u32 max_direct,
180
185
 		       struct hns_roce_buf *buf, u32 page_shift)
181
186
 {
182
-	int i = 0;
183
-	dma_addr_t t;
187
+	struct hns_roce_buf_list *buf_list;
184
188
 	struct device *dev = hr_dev->dev;
185
-	u32 page_size = 1 << page_shift;
186
-	u32 order;
189
+	u32 page_size;
190
+	int i;
187
191
 
188
-	/* SQ/RQ buf lease than one page, SQ + RQ = 8K */
192
+	/* The minimum shift of the page accessed by hw is HNS_HW_PAGE_SHIFT */
193
+	buf->page_shift = max_t(int, HNS_HW_PAGE_SHIFT, page_shift);
194
+
195
+	page_size = 1 << buf->page_shift;
196
+	buf->npages = DIV_ROUND_UP(size, page_size);
197
+
198
+	/* required size is not bigger than one trunk size */
189
199
 	if (size <= max_direct) {
190
-		buf->nbufs = 1;
191
-		/* Npages calculated by page_size */
192
-		order = get_order(size);
193
-		if (order <= page_shift - PAGE_SHIFT)
194
-			order = 0;
195
-		else
196
-			order -= page_shift - PAGE_SHIFT;
197
-		buf->npages = 1 << order;
198
-		buf->page_shift = page_shift;
199
-		/* MTT PA must be recorded in 4k alignment, t is 4k aligned */
200
-		buf->direct.buf = dma_zalloc_coherent(dev,
201
-						      size, &t, GFP_KERNEL);
200
+		buf->page_list = NULL;
201
+		buf->direct.buf = dma_alloc_coherent(dev, size,
202
+						     &buf->direct.map,
203
+						     GFP_KERNEL);
202
204
 		if (!buf->direct.buf)
203
205
 			return -ENOMEM;
204
-
205
-		buf->direct.map = t;
206
-
207
-		while (t & ((1 << buf->page_shift) - 1)) {
208
-			--buf->page_shift;
209
-			buf->npages *= 2;
210
-		}
211
206
 	} else {
212
-		buf->nbufs = (size + page_size - 1) / page_size;
213
-		buf->npages = buf->nbufs;
214
-		buf->page_shift = page_shift;
215
-		buf->page_list = kcalloc(buf->nbufs, sizeof(*buf->page_list),
216
-					 GFP_KERNEL);
217
-
218
-		if (!buf->page_list)
207
+		buf_list = kcalloc(buf->npages, sizeof(*buf_list), GFP_KERNEL);
208
+		if (!buf_list)
219
209
 			return -ENOMEM;
220
210
 
221
-		for (i = 0; i < buf->nbufs; ++i) {
222
-			buf->page_list[i].buf = dma_zalloc_coherent(dev,
223
-								  page_size, &t,
224
-								  GFP_KERNEL);
225
-
226
-			if (!buf->page_list[i].buf)
227
-				goto err_free;
228
-
229
-			buf->page_list[i].map = t;
211
+		for (i = 0; i < buf->npages; i++) {
212
+			buf_list[i].buf = dma_alloc_coherent(dev, page_size,
213
+							     &buf_list[i].map,
214
+							     GFP_KERNEL);
215
+			if (!buf_list[i].buf)
216
+				break;
230
217
 		}
218
+
219
+		if (i != buf->npages && i > 0) {
220
+			while (i-- > 0)
221
+				dma_free_coherent(dev, page_size,
222
+						  buf_list[i].buf,
223
+						  buf_list[i].map);
224
+			kfree(buf_list);
225
+			return -ENOMEM;
226
+		}
227
+		buf->page_list = buf_list;
231
228
 	}
229
+	buf->size = size;
232
230
 
233
231
 	return 0;
232
+}
234
233
 
235
-err_free:
236
-	hns_roce_buf_free(hr_dev, size, buf);
237
-	return -ENOMEM;
234
+int hns_roce_get_kmem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
235
+			   int buf_cnt, int start, struct hns_roce_buf *buf)
236
+{
237
+	int i, end;
238
+	int total;
239
+
240
+	end = start + buf_cnt;
241
+	if (end > buf->npages) {
242
+		dev_err(hr_dev->dev,
243
+			"failed to check kmem bufs, end %d + %d total %u!\n",
244
+			start, buf_cnt, buf->npages);
245
+		return -EINVAL;
246
+	}
247
+
248
+	total = 0;
249
+	for (i = start; i < end; i++)
250
+		bufs[total++] = hns_roce_buf_page(buf, i);
251
+
252
+	return total;
253
+}
254
+
255
+int hns_roce_get_umem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
256
+			   int buf_cnt, int start, struct ib_umem *umem,
257
+			   unsigned int page_shift)
258
+{
259
+	struct ib_block_iter biter;
260
+	int total = 0;
261
+	int idx = 0;
262
+	u64 addr;
263
+
264
+	if (page_shift < HNS_HW_PAGE_SHIFT) {
265
+		dev_err(hr_dev->dev, "failed to check umem page shift %u!\n",
266
+			page_shift);
267
+		return -EINVAL;
268
+	}
269
+
270
+	/* convert system page cnt to hw page cnt */
271
+	rdma_umem_for_each_dma_block(umem, &biter, 1 << page_shift) {
272
+		addr = rdma_block_iter_dma_address(&biter);
273
+		if (idx >= start) {
274
+			bufs[total++] = addr;
275
+			if (total >= buf_cnt)
276
+				goto done;
277
+		}
278
+		idx++;
279
+	}
280
+
281
+done:
282
+	return total;
238
283
 }
239
284
 
240
285
 void hns_roce_cleanup_bitmap(struct hns_roce_dev *hr_dev)
241
286
 {
287
+	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
288
+		hns_roce_cleanup_srq_table(hr_dev);
242
289
 	hns_roce_cleanup_qp_table(hr_dev);
243
290
 	hns_roce_cleanup_cq_table(hr_dev);
244
291
 	hns_roce_cleanup_mr_table(hr_dev);