hns_roce_hem.c 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404
  1. /*
  2. * Copyright (c) 2016 Hisilicon Limited.
  3. * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
  4. *
  5. * This software is available to you under a choice of one of two
  6. * licenses. You may choose to be licensed under the terms of the GNU
  7. * General Public License (GPL) Version 2, available from the file
  8. * COPYING in the main directory of this source tree, or the
  9. * OpenIB.org BSD license below:
  10. *
  11. * Redistribution and use in source and binary forms, with or
  12. * without modification, are permitted provided that the following
  13. * conditions are met:
  14. *
  15. * - Redistributions of source code must retain the above
  16. * copyright notice, this list of conditions and the following
  17. * disclaimer.
  18. *
  19. * - Redistributions in binary form must reproduce the above
  20. * copyright notice, this list of conditions and the following
  21. * disclaimer in the documentation and/or other materials
  22. * provided with the distribution.
  23. *
  24. * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25. * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26. * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27. * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28. * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29. * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30. * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31. * SOFTWARE.
  32. */
  33. #include <linux/platform_device.h>
  34. #include "hns_roce_device.h"
  35. #include "hns_roce_hem.h"
  36. #include "hns_roce_common.h"
  37. #define HNS_ROCE_HEM_ALLOC_SIZE (1 << 17)
  38. #define HNS_ROCE_TABLE_CHUNK_SIZE (1 << 17)
  39. #define DMA_ADDR_T_SHIFT 12
  40. #define BT_BA_SHIFT 32
  41. struct hns_roce_hem *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev, int npages,
  42. gfp_t gfp_mask)
  43. {
  44. struct hns_roce_hem_chunk *chunk = NULL;
  45. struct hns_roce_hem *hem;
  46. struct scatterlist *mem;
  47. int order;
  48. void *buf;
  49. WARN_ON(gfp_mask & __GFP_HIGHMEM);
  50. hem = kmalloc(sizeof(*hem),
  51. gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
  52. if (!hem)
  53. return NULL;
  54. hem->refcount = 0;
  55. INIT_LIST_HEAD(&hem->chunk_list);
  56. order = get_order(HNS_ROCE_HEM_ALLOC_SIZE);
  57. while (npages > 0) {
  58. if (!chunk) {
  59. chunk = kmalloc(sizeof(*chunk),
  60. gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
  61. if (!chunk)
  62. goto fail;
  63. sg_init_table(chunk->mem, HNS_ROCE_HEM_CHUNK_LEN);
  64. chunk->npages = 0;
  65. chunk->nsg = 0;
  66. list_add_tail(&chunk->list, &hem->chunk_list);
  67. }
  68. while (1 << order > npages)
  69. --order;
  70. /*
  71. * Alloc memory one time. If failed, don't alloc small block
  72. * memory, directly return fail.
  73. */
  74. mem = &chunk->mem[chunk->npages];
  75. buf = dma_alloc_coherent(&hr_dev->pdev->dev, PAGE_SIZE << order,
  76. &sg_dma_address(mem), gfp_mask);
  77. if (!buf)
  78. goto fail;
  79. sg_set_buf(mem, buf, PAGE_SIZE << order);
  80. WARN_ON(mem->offset);
  81. sg_dma_len(mem) = PAGE_SIZE << order;
  82. ++chunk->npages;
  83. ++chunk->nsg;
  84. npages -= 1 << order;
  85. }
  86. return hem;
  87. fail:
  88. hns_roce_free_hem(hr_dev, hem);
  89. return NULL;
  90. }
  91. void hns_roce_free_hem(struct hns_roce_dev *hr_dev, struct hns_roce_hem *hem)
  92. {
  93. struct hns_roce_hem_chunk *chunk, *tmp;
  94. int i;
  95. if (!hem)
  96. return;
  97. list_for_each_entry_safe(chunk, tmp, &hem->chunk_list, list) {
  98. for (i = 0; i < chunk->npages; ++i)
  99. dma_free_coherent(&hr_dev->pdev->dev,
  100. chunk->mem[i].length,
  101. lowmem_page_address(sg_page(&chunk->mem[i])),
  102. sg_dma_address(&chunk->mem[i]));
  103. kfree(chunk);
  104. }
  105. kfree(hem);
  106. }
  107. static int hns_roce_set_hem(struct hns_roce_dev *hr_dev,
  108. struct hns_roce_hem_table *table, unsigned long obj)
  109. {
  110. struct device *dev = &hr_dev->pdev->dev;
  111. spinlock_t *lock = &hr_dev->bt_cmd_lock;
  112. unsigned long end = 0;
  113. unsigned long flags;
  114. struct hns_roce_hem_iter iter;
  115. void __iomem *bt_cmd;
  116. u32 bt_cmd_h_val = 0;
  117. u32 bt_cmd_val[2];
  118. u32 bt_cmd_l = 0;
  119. u64 bt_ba = 0;
  120. int ret = 0;
  121. /* Find the HEM(Hardware Entry Memory) entry */
  122. unsigned long i = (obj & (table->num_obj - 1)) /
  123. (HNS_ROCE_TABLE_CHUNK_SIZE / table->obj_size);
  124. switch (table->type) {
  125. case HEM_TYPE_QPC:
  126. roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
  127. ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, HEM_TYPE_QPC);
  128. break;
  129. case HEM_TYPE_MTPT:
  130. roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
  131. ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S,
  132. HEM_TYPE_MTPT);
  133. break;
  134. case HEM_TYPE_CQC:
  135. roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
  136. ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, HEM_TYPE_CQC);
  137. break;
  138. case HEM_TYPE_SRQC:
  139. roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
  140. ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S,
  141. HEM_TYPE_SRQC);
  142. break;
  143. default:
  144. return ret;
  145. }
  146. roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_M,
  147. ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_S, obj);
  148. roce_set_bit(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_S, 0);
  149. roce_set_bit(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_HW_SYNS_S, 1);
  150. /* Currently iter only a chunk */
  151. for (hns_roce_hem_first(table->hem[i], &iter);
  152. !hns_roce_hem_last(&iter); hns_roce_hem_next(&iter)) {
  153. bt_ba = hns_roce_hem_addr(&iter) >> DMA_ADDR_T_SHIFT;
  154. spin_lock_irqsave(lock, flags);
  155. bt_cmd = hr_dev->reg_base + ROCEE_BT_CMD_H_REG;
  156. end = msecs_to_jiffies(HW_SYNC_TIMEOUT_MSECS) + jiffies;
  157. while (1) {
  158. if (readl(bt_cmd) >> BT_CMD_SYNC_SHIFT) {
  159. if (!(time_before(jiffies, end))) {
  160. dev_err(dev, "Write bt_cmd err,hw_sync is not zero.\n");
  161. spin_unlock_irqrestore(lock, flags);
  162. return -EBUSY;
  163. }
  164. } else {
  165. break;
  166. }
  167. msleep(HW_SYNC_SLEEP_TIME_INTERVAL);
  168. }
  169. bt_cmd_l = (u32)bt_ba;
  170. roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_M,
  171. ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_S,
  172. bt_ba >> BT_BA_SHIFT);
  173. bt_cmd_val[0] = bt_cmd_l;
  174. bt_cmd_val[1] = bt_cmd_h_val;
  175. hns_roce_write64_k(bt_cmd_val,
  176. hr_dev->reg_base + ROCEE_BT_CMD_L_REG);
  177. spin_unlock_irqrestore(lock, flags);
  178. }
  179. return ret;
  180. }
  181. int hns_roce_table_get(struct hns_roce_dev *hr_dev,
  182. struct hns_roce_hem_table *table, unsigned long obj)
  183. {
  184. struct device *dev = &hr_dev->pdev->dev;
  185. int ret = 0;
  186. unsigned long i;
  187. i = (obj & (table->num_obj - 1)) / (HNS_ROCE_TABLE_CHUNK_SIZE /
  188. table->obj_size);
  189. mutex_lock(&table->mutex);
  190. if (table->hem[i]) {
  191. ++table->hem[i]->refcount;
  192. goto out;
  193. }
  194. table->hem[i] = hns_roce_alloc_hem(hr_dev,
  195. HNS_ROCE_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
  196. (table->lowmem ? GFP_KERNEL :
  197. GFP_HIGHUSER) | __GFP_NOWARN);
  198. if (!table->hem[i]) {
  199. ret = -ENOMEM;
  200. goto out;
  201. }
  202. /* Set HEM base address(128K/page, pa) to Hardware */
  203. if (hns_roce_set_hem(hr_dev, table, obj)) {
  204. ret = -ENODEV;
  205. dev_err(dev, "set HEM base address to HW failed.\n");
  206. goto out;
  207. }
  208. ++table->hem[i]->refcount;
  209. out:
  210. mutex_unlock(&table->mutex);
  211. return ret;
  212. }
  213. void hns_roce_table_put(struct hns_roce_dev *hr_dev,
  214. struct hns_roce_hem_table *table, unsigned long obj)
  215. {
  216. struct device *dev = &hr_dev->pdev->dev;
  217. unsigned long i;
  218. i = (obj & (table->num_obj - 1)) /
  219. (HNS_ROCE_TABLE_CHUNK_SIZE / table->obj_size);
  220. mutex_lock(&table->mutex);
  221. if (--table->hem[i]->refcount == 0) {
  222. /* Clear HEM base address */
  223. if (hr_dev->hw->clear_hem(hr_dev, table, obj))
  224. dev_warn(dev, "Clear HEM base address failed.\n");
  225. hns_roce_free_hem(hr_dev, table->hem[i]);
  226. table->hem[i] = NULL;
  227. }
  228. mutex_unlock(&table->mutex);
  229. }
  230. void *hns_roce_table_find(struct hns_roce_hem_table *table, unsigned long obj,
  231. dma_addr_t *dma_handle)
  232. {
  233. struct hns_roce_hem_chunk *chunk;
  234. unsigned long idx;
  235. int i;
  236. int offset, dma_offset;
  237. struct hns_roce_hem *hem;
  238. struct page *page = NULL;
  239. if (!table->lowmem)
  240. return NULL;
  241. mutex_lock(&table->mutex);
  242. idx = (obj & (table->num_obj - 1)) * table->obj_size;
  243. hem = table->hem[idx / HNS_ROCE_TABLE_CHUNK_SIZE];
  244. dma_offset = offset = idx % HNS_ROCE_TABLE_CHUNK_SIZE;
  245. if (!hem)
  246. goto out;
  247. list_for_each_entry(chunk, &hem->chunk_list, list) {
  248. for (i = 0; i < chunk->npages; ++i) {
  249. if (dma_handle && dma_offset >= 0) {
  250. if (sg_dma_len(&chunk->mem[i]) >
  251. (u32)dma_offset)
  252. *dma_handle = sg_dma_address(
  253. &chunk->mem[i]) + dma_offset;
  254. dma_offset -= sg_dma_len(&chunk->mem[i]);
  255. }
  256. if (chunk->mem[i].length > (u32)offset) {
  257. page = sg_page(&chunk->mem[i]);
  258. goto out;
  259. }
  260. offset -= chunk->mem[i].length;
  261. }
  262. }
  263. out:
  264. mutex_unlock(&table->mutex);
  265. return page ? lowmem_page_address(page) + offset : NULL;
  266. }
  267. int hns_roce_table_get_range(struct hns_roce_dev *hr_dev,
  268. struct hns_roce_hem_table *table,
  269. unsigned long start, unsigned long end)
  270. {
  271. unsigned long inc = HNS_ROCE_TABLE_CHUNK_SIZE / table->obj_size;
  272. unsigned long i = 0;
  273. int ret = 0;
  274. /* Allocate MTT entry memory according to chunk(128K) */
  275. for (i = start; i <= end; i += inc) {
  276. ret = hns_roce_table_get(hr_dev, table, i);
  277. if (ret)
  278. goto fail;
  279. }
  280. return 0;
  281. fail:
  282. while (i > start) {
  283. i -= inc;
  284. hns_roce_table_put(hr_dev, table, i);
  285. }
  286. return ret;
  287. }
  288. void hns_roce_table_put_range(struct hns_roce_dev *hr_dev,
  289. struct hns_roce_hem_table *table,
  290. unsigned long start, unsigned long end)
  291. {
  292. unsigned long i;
  293. for (i = start; i <= end;
  294. i += HNS_ROCE_TABLE_CHUNK_SIZE / table->obj_size)
  295. hns_roce_table_put(hr_dev, table, i);
  296. }
  297. int hns_roce_init_hem_table(struct hns_roce_dev *hr_dev,
  298. struct hns_roce_hem_table *table, u32 type,
  299. unsigned long obj_size, unsigned long nobj,
  300. int use_lowmem)
  301. {
  302. unsigned long obj_per_chunk;
  303. unsigned long num_hem;
  304. obj_per_chunk = HNS_ROCE_TABLE_CHUNK_SIZE / obj_size;
  305. num_hem = (nobj + obj_per_chunk - 1) / obj_per_chunk;
  306. table->hem = kcalloc(num_hem, sizeof(*table->hem), GFP_KERNEL);
  307. if (!table->hem)
  308. return -ENOMEM;
  309. table->type = type;
  310. table->num_hem = num_hem;
  311. table->num_obj = nobj;
  312. table->obj_size = obj_size;
  313. table->lowmem = use_lowmem;
  314. mutex_init(&table->mutex);
  315. return 0;
  316. }
  317. void hns_roce_cleanup_hem_table(struct hns_roce_dev *hr_dev,
  318. struct hns_roce_hem_table *table)
  319. {
  320. struct device *dev = &hr_dev->pdev->dev;
  321. unsigned long i;
  322. for (i = 0; i < table->num_hem; ++i)
  323. if (table->hem[i]) {
  324. if (hr_dev->hw->clear_hem(hr_dev, table,
  325. i * HNS_ROCE_TABLE_CHUNK_SIZE / table->obj_size))
  326. dev_err(dev, "Clear HEM base address failed.\n");
  327. hns_roce_free_hem(hr_dev, table->hem[i]);
  328. }
  329. kfree(table->hem);
  330. }
  331. void hns_roce_cleanup_hem(struct hns_roce_dev *hr_dev)
  332. {
  333. hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cq_table.table);
  334. hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.irrl_table);
  335. hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.qp_table);
  336. hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table);
  337. hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtt_table);
  338. }