smc_ib.c 17 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643
  1. // SPDX-License-Identifier: GPL-2.0
  2. /*
  3. * Shared Memory Communications over RDMA (SMC-R) and RoCE
  4. *
  5. * IB infrastructure:
  6. * Establish SMC-R as an Infiniband Client to be notified about added and
  7. * removed IB devices of type RDMA.
  8. * Determine device and port characteristics for these IB devices.
  9. *
  10. * Copyright IBM Corp. 2016
  11. *
  12. * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
  13. */
  14. #include <linux/random.h>
  15. #include <linux/workqueue.h>
  16. #include <linux/scatterlist.h>
  17. #include <linux/wait.h>
  18. #include <linux/mutex.h>
  19. #include <rdma/ib_verbs.h>
  20. #include <rdma/ib_cache.h>
  21. #include "smc_pnet.h"
  22. #include "smc_ib.h"
  23. #include "smc_core.h"
  24. #include "smc_wr.h"
  25. #include "smc.h"
  26. #define SMC_MAX_CQE 32766 /* max. # of completion queue elements */
  27. #define SMC_QP_MIN_RNR_TIMER 5
  28. #define SMC_QP_TIMEOUT 15 /* 4096 * 2 ** timeout usec */
  29. #define SMC_QP_RETRY_CNT 7 /* 7: infinite */
  30. #define SMC_QP_RNR_RETRY 7 /* 7: infinite */
  31. struct smc_ib_devices smc_ib_devices = { /* smc-registered ib devices */
  32. .mutex = __MUTEX_INITIALIZER(smc_ib_devices.mutex),
  33. .list = LIST_HEAD_INIT(smc_ib_devices.list),
  34. };
  35. u8 local_systemid[SMC_SYSTEMID_LEN]; /* unique system identifier */
  36. static int smc_ib_modify_qp_init(struct smc_link *lnk)
  37. {
  38. struct ib_qp_attr qp_attr;
  39. memset(&qp_attr, 0, sizeof(qp_attr));
  40. qp_attr.qp_state = IB_QPS_INIT;
  41. qp_attr.pkey_index = 0;
  42. qp_attr.port_num = lnk->ibport;
  43. qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
  44. | IB_ACCESS_REMOTE_WRITE;
  45. return ib_modify_qp(lnk->roce_qp, &qp_attr,
  46. IB_QP_STATE | IB_QP_PKEY_INDEX |
  47. IB_QP_ACCESS_FLAGS | IB_QP_PORT);
  48. }
  49. static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
  50. {
  51. enum ib_qp_attr_mask qp_attr_mask =
  52. IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
  53. IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
  54. struct ib_qp_attr qp_attr;
  55. memset(&qp_attr, 0, sizeof(qp_attr));
  56. qp_attr.qp_state = IB_QPS_RTR;
  57. qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
  58. qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
  59. rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
  60. rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, 1, 0);
  61. rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
  62. memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
  63. sizeof(lnk->peer_mac));
  64. qp_attr.dest_qp_num = lnk->peer_qpn;
  65. qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
  66. qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
  67. * requests
  68. */
  69. qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
  70. return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
  71. }
  72. int smc_ib_modify_qp_rts(struct smc_link *lnk)
  73. {
  74. struct ib_qp_attr qp_attr;
  75. memset(&qp_attr, 0, sizeof(qp_attr));
  76. qp_attr.qp_state = IB_QPS_RTS;
  77. qp_attr.timeout = SMC_QP_TIMEOUT; /* local ack timeout */
  78. qp_attr.retry_cnt = SMC_QP_RETRY_CNT; /* retry count */
  79. qp_attr.rnr_retry = SMC_QP_RNR_RETRY; /* RNR retries, 7=infinite */
  80. qp_attr.sq_psn = lnk->psn_initial; /* starting send packet seq # */
  81. qp_attr.max_rd_atomic = 1; /* # of outstanding RDMA reads and
  82. * atomic ops allowed
  83. */
  84. return ib_modify_qp(lnk->roce_qp, &qp_attr,
  85. IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
  86. IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
  87. IB_QP_MAX_QP_RD_ATOMIC);
  88. }
  89. int smc_ib_modify_qp_error(struct smc_link *lnk)
  90. {
  91. struct ib_qp_attr qp_attr;
  92. memset(&qp_attr, 0, sizeof(qp_attr));
  93. qp_attr.qp_state = IB_QPS_ERR;
  94. return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
  95. }
  96. int smc_ib_ready_link(struct smc_link *lnk)
  97. {
  98. struct smc_link_group *lgr = smc_get_lgr(lnk);
  99. int rc = 0;
  100. rc = smc_ib_modify_qp_init(lnk);
  101. if (rc)
  102. goto out;
  103. rc = smc_ib_modify_qp_rtr(lnk);
  104. if (rc)
  105. goto out;
  106. smc_wr_remember_qp_attr(lnk);
  107. rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
  108. IB_CQ_SOLICITED_MASK);
  109. if (rc)
  110. goto out;
  111. rc = smc_wr_rx_post_init(lnk);
  112. if (rc)
  113. goto out;
  114. smc_wr_remember_qp_attr(lnk);
  115. if (lgr->role == SMC_SERV) {
  116. rc = smc_ib_modify_qp_rts(lnk);
  117. if (rc)
  118. goto out;
  119. smc_wr_remember_qp_attr(lnk);
  120. }
  121. out:
  122. return rc;
  123. }
  124. static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport)
  125. {
  126. const struct ib_gid_attr *attr;
  127. int rc;
  128. attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0);
  129. if (IS_ERR(attr))
  130. return -ENODEV;
  131. rc = rdma_read_gid_l2_fields(attr, NULL, smcibdev->mac[ibport - 1]);
  132. rdma_put_gid_attr(attr);
  133. return rc;
  134. }
  135. /* Create an identifier unique for this instance of SMC-R.
  136. * The MAC-address of the first active registered IB device
  137. * plus a random 2-byte number is used to create this identifier.
  138. * This name is delivered to the peer during connection initialization.
  139. */
  140. static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
  141. u8 ibport)
  142. {
  143. memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
  144. sizeof(smcibdev->mac[ibport - 1]));
  145. }
  146. bool smc_ib_is_valid_local_systemid(void)
  147. {
  148. return !is_zero_ether_addr(&local_systemid[2]);
  149. }
  150. static void smc_ib_init_local_systemid(void)
  151. {
  152. get_random_bytes(&local_systemid[0], 2);
  153. }
  154. bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
  155. {
  156. return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
  157. }
  158. /* determine the gid for an ib-device port and vlan id */
  159. int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
  160. unsigned short vlan_id, u8 gid[], u8 *sgid_index)
  161. {
  162. const struct ib_gid_attr *attr;
  163. const struct net_device *ndev;
  164. int i;
  165. for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
  166. attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
  167. if (IS_ERR(attr))
  168. continue;
  169. rcu_read_lock();
  170. ndev = rdma_read_gid_attr_ndev_rcu(attr);
  171. if (!IS_ERR(ndev) &&
  172. ((!vlan_id && !is_vlan_dev(ndev)) ||
  173. (vlan_id && is_vlan_dev(ndev) &&
  174. vlan_dev_vlan_id(ndev) == vlan_id)) &&
  175. attr->gid_type == IB_GID_TYPE_ROCE) {
  176. rcu_read_unlock();
  177. if (gid)
  178. memcpy(gid, &attr->gid, SMC_GID_SIZE);
  179. if (sgid_index)
  180. *sgid_index = attr->index;
  181. rdma_put_gid_attr(attr);
  182. return 0;
  183. }
  184. rcu_read_unlock();
  185. rdma_put_gid_attr(attr);
  186. }
  187. return -ENODEV;
  188. }
  189. static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
  190. {
  191. int rc;
  192. memset(&smcibdev->pattr[ibport - 1], 0,
  193. sizeof(smcibdev->pattr[ibport - 1]));
  194. rc = ib_query_port(smcibdev->ibdev, ibport,
  195. &smcibdev->pattr[ibport - 1]);
  196. if (rc)
  197. goto out;
  198. /* the SMC protocol requires specification of the RoCE MAC address */
  199. rc = smc_ib_fill_mac(smcibdev, ibport);
  200. if (rc)
  201. goto out;
  202. if (!smc_ib_is_valid_local_systemid() &&
  203. smc_ib_port_active(smcibdev, ibport))
  204. /* create unique system identifier */
  205. smc_ib_define_local_systemid(smcibdev, ibport);
  206. out:
  207. return rc;
  208. }
  209. /* process context wrapper for might_sleep smc_ib_remember_port_attr */
  210. static void smc_ib_port_event_work(struct work_struct *work)
  211. {
  212. struct smc_ib_device *smcibdev = container_of(
  213. work, struct smc_ib_device, port_event_work);
  214. u8 port_idx;
  215. for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
  216. smc_ib_remember_port_attr(smcibdev, port_idx + 1);
  217. clear_bit(port_idx, &smcibdev->port_event_mask);
  218. if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
  219. set_bit(port_idx, smcibdev->ports_going_away);
  220. smcr_port_err(smcibdev, port_idx + 1);
  221. } else {
  222. clear_bit(port_idx, smcibdev->ports_going_away);
  223. smcr_port_add(smcibdev, port_idx + 1);
  224. }
  225. }
  226. }
  227. /* can be called in IRQ context */
  228. static void smc_ib_global_event_handler(struct ib_event_handler *handler,
  229. struct ib_event *ibevent)
  230. {
  231. struct smc_ib_device *smcibdev;
  232. bool schedule = false;
  233. u8 port_idx;
  234. smcibdev = container_of(handler, struct smc_ib_device, event_handler);
  235. switch (ibevent->event) {
  236. case IB_EVENT_DEVICE_FATAL:
  237. /* terminate all ports on device */
  238. for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
  239. set_bit(port_idx, &smcibdev->port_event_mask);
  240. if (!test_and_set_bit(port_idx,
  241. smcibdev->ports_going_away))
  242. schedule = true;
  243. }
  244. if (schedule)
  245. schedule_work(&smcibdev->port_event_work);
  246. break;
  247. case IB_EVENT_PORT_ACTIVE:
  248. port_idx = ibevent->element.port_num - 1;
  249. if (port_idx >= SMC_MAX_PORTS)
  250. break;
  251. set_bit(port_idx, &smcibdev->port_event_mask);
  252. if (test_and_clear_bit(port_idx, smcibdev->ports_going_away))
  253. schedule_work(&smcibdev->port_event_work);
  254. break;
  255. case IB_EVENT_PORT_ERR:
  256. port_idx = ibevent->element.port_num - 1;
  257. if (port_idx >= SMC_MAX_PORTS)
  258. break;
  259. set_bit(port_idx, &smcibdev->port_event_mask);
  260. if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
  261. schedule_work(&smcibdev->port_event_work);
  262. break;
  263. case IB_EVENT_GID_CHANGE:
  264. port_idx = ibevent->element.port_num - 1;
  265. if (port_idx >= SMC_MAX_PORTS)
  266. break;
  267. set_bit(port_idx, &smcibdev->port_event_mask);
  268. schedule_work(&smcibdev->port_event_work);
  269. break;
  270. default:
  271. break;
  272. }
  273. }
  274. void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
  275. {
  276. if (lnk->roce_pd)
  277. ib_dealloc_pd(lnk->roce_pd);
  278. lnk->roce_pd = NULL;
  279. }
  280. int smc_ib_create_protection_domain(struct smc_link *lnk)
  281. {
  282. int rc;
  283. lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
  284. rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
  285. if (IS_ERR(lnk->roce_pd))
  286. lnk->roce_pd = NULL;
  287. return rc;
  288. }
  289. static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
  290. {
  291. struct smc_link *lnk = (struct smc_link *)priv;
  292. struct smc_ib_device *smcibdev = lnk->smcibdev;
  293. u8 port_idx;
  294. switch (ibevent->event) {
  295. case IB_EVENT_QP_FATAL:
  296. case IB_EVENT_QP_ACCESS_ERR:
  297. port_idx = ibevent->element.qp->port - 1;
  298. if (port_idx >= SMC_MAX_PORTS)
  299. break;
  300. set_bit(port_idx, &smcibdev->port_event_mask);
  301. if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
  302. schedule_work(&smcibdev->port_event_work);
  303. break;
  304. default:
  305. break;
  306. }
  307. }
  308. void smc_ib_destroy_queue_pair(struct smc_link *lnk)
  309. {
  310. if (lnk->roce_qp)
  311. ib_destroy_qp(lnk->roce_qp);
  312. lnk->roce_qp = NULL;
  313. }
  314. /* create a queue pair within the protection domain for a link */
  315. int smc_ib_create_queue_pair(struct smc_link *lnk)
  316. {
  317. struct ib_qp_init_attr qp_attr = {
  318. .event_handler = smc_ib_qp_event_handler,
  319. .qp_context = lnk,
  320. .send_cq = lnk->smcibdev->roce_cq_send,
  321. .recv_cq = lnk->smcibdev->roce_cq_recv,
  322. .srq = NULL,
  323. .cap = {
  324. /* include unsolicited rdma_writes as well,
  325. * there are max. 2 RDMA_WRITE per 1 WR_SEND
  326. */
  327. .max_send_wr = SMC_WR_BUF_CNT * 3,
  328. .max_recv_wr = SMC_WR_BUF_CNT * 3,
  329. .max_send_sge = SMC_IB_MAX_SEND_SGE,
  330. .max_recv_sge = 1,
  331. },
  332. .sq_sig_type = IB_SIGNAL_REQ_WR,
  333. .qp_type = IB_QPT_RC,
  334. };
  335. int rc;
  336. lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
  337. rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
  338. if (IS_ERR(lnk->roce_qp))
  339. lnk->roce_qp = NULL;
  340. else
  341. smc_wr_remember_qp_attr(lnk);
  342. return rc;
  343. }
  344. void smc_ib_put_memory_region(struct ib_mr *mr)
  345. {
  346. ib_dereg_mr(mr);
  347. }
  348. static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot, u8 link_idx)
  349. {
  350. unsigned int offset = 0;
  351. int sg_num;
  352. /* map the largest prefix of a dma mapped SG list */
  353. sg_num = ib_map_mr_sg(buf_slot->mr_rx[link_idx],
  354. buf_slot->sgt[link_idx].sgl,
  355. buf_slot->sgt[link_idx].orig_nents,
  356. &offset, PAGE_SIZE);
  357. return sg_num;
  358. }
  359. /* Allocate a memory region and map the dma mapped SG list of buf_slot */
  360. int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
  361. struct smc_buf_desc *buf_slot, u8 link_idx)
  362. {
  363. if (buf_slot->mr_rx[link_idx])
  364. return 0; /* already done */
  365. buf_slot->mr_rx[link_idx] =
  366. ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
  367. if (IS_ERR(buf_slot->mr_rx[link_idx])) {
  368. int rc;
  369. rc = PTR_ERR(buf_slot->mr_rx[link_idx]);
  370. buf_slot->mr_rx[link_idx] = NULL;
  371. return rc;
  372. }
  373. if (smc_ib_map_mr_sg(buf_slot, link_idx) != 1)
  374. return -EINVAL;
  375. return 0;
  376. }
  377. /* synchronize buffer usage for cpu access */
  378. void smc_ib_sync_sg_for_cpu(struct smc_link *lnk,
  379. struct smc_buf_desc *buf_slot,
  380. enum dma_data_direction data_direction)
  381. {
  382. struct scatterlist *sg;
  383. unsigned int i;
  384. /* for now there is just one DMA address */
  385. for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
  386. buf_slot->sgt[lnk->link_idx].nents, i) {
  387. if (!sg_dma_len(sg))
  388. break;
  389. ib_dma_sync_single_for_cpu(lnk->smcibdev->ibdev,
  390. sg_dma_address(sg),
  391. sg_dma_len(sg),
  392. data_direction);
  393. }
  394. }
  395. /* synchronize buffer usage for device access */
  396. void smc_ib_sync_sg_for_device(struct smc_link *lnk,
  397. struct smc_buf_desc *buf_slot,
  398. enum dma_data_direction data_direction)
  399. {
  400. struct scatterlist *sg;
  401. unsigned int i;
  402. /* for now there is just one DMA address */
  403. for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
  404. buf_slot->sgt[lnk->link_idx].nents, i) {
  405. if (!sg_dma_len(sg))
  406. break;
  407. ib_dma_sync_single_for_device(lnk->smcibdev->ibdev,
  408. sg_dma_address(sg),
  409. sg_dma_len(sg),
  410. data_direction);
  411. }
  412. }
  413. /* Map a new TX or RX buffer SG-table to DMA */
  414. int smc_ib_buf_map_sg(struct smc_link *lnk,
  415. struct smc_buf_desc *buf_slot,
  416. enum dma_data_direction data_direction)
  417. {
  418. int mapped_nents;
  419. mapped_nents = ib_dma_map_sg(lnk->smcibdev->ibdev,
  420. buf_slot->sgt[lnk->link_idx].sgl,
  421. buf_slot->sgt[lnk->link_idx].orig_nents,
  422. data_direction);
  423. if (!mapped_nents)
  424. return -ENOMEM;
  425. return mapped_nents;
  426. }
  427. void smc_ib_buf_unmap_sg(struct smc_link *lnk,
  428. struct smc_buf_desc *buf_slot,
  429. enum dma_data_direction data_direction)
  430. {
  431. if (!buf_slot->sgt[lnk->link_idx].sgl->dma_address)
  432. return; /* already unmapped */
  433. ib_dma_unmap_sg(lnk->smcibdev->ibdev,
  434. buf_slot->sgt[lnk->link_idx].sgl,
  435. buf_slot->sgt[lnk->link_idx].orig_nents,
  436. data_direction);
  437. buf_slot->sgt[lnk->link_idx].sgl->dma_address = 0;
  438. }
  439. long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
  440. {
  441. struct ib_cq_init_attr cqattr = {
  442. .cqe = SMC_MAX_CQE, .comp_vector = 0 };
  443. int cqe_size_order, smc_order;
  444. long rc;
  445. mutex_lock(&smcibdev->mutex);
  446. rc = 0;
  447. if (smcibdev->initialized)
  448. goto out;
  449. /* the calculated number of cq entries fits to mlx5 cq allocation */
  450. cqe_size_order = cache_line_size() == 128 ? 7 : 6;
  451. smc_order = MAX_ORDER - cqe_size_order - 1;
  452. if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
  453. cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
  454. smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
  455. smc_wr_tx_cq_handler, NULL,
  456. smcibdev, &cqattr);
  457. rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
  458. if (IS_ERR(smcibdev->roce_cq_send)) {
  459. smcibdev->roce_cq_send = NULL;
  460. goto out;
  461. }
  462. smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
  463. smc_wr_rx_cq_handler, NULL,
  464. smcibdev, &cqattr);
  465. rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
  466. if (IS_ERR(smcibdev->roce_cq_recv)) {
  467. smcibdev->roce_cq_recv = NULL;
  468. goto err;
  469. }
  470. smc_wr_add_dev(smcibdev);
  471. smcibdev->initialized = 1;
  472. goto out;
  473. err:
  474. ib_destroy_cq(smcibdev->roce_cq_send);
  475. out:
  476. mutex_unlock(&smcibdev->mutex);
  477. return rc;
  478. }
  479. static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
  480. {
  481. mutex_lock(&smcibdev->mutex);
  482. if (!smcibdev->initialized)
  483. goto out;
  484. smcibdev->initialized = 0;
  485. ib_destroy_cq(smcibdev->roce_cq_recv);
  486. ib_destroy_cq(smcibdev->roce_cq_send);
  487. smc_wr_remove_dev(smcibdev);
  488. out:
  489. mutex_unlock(&smcibdev->mutex);
  490. }
  491. static struct ib_client smc_ib_client;
  492. /* callback function for ib_register_client() */
  493. static int smc_ib_add_dev(struct ib_device *ibdev)
  494. {
  495. struct smc_ib_device *smcibdev;
  496. u8 port_cnt;
  497. int i;
  498. if (ibdev->node_type != RDMA_NODE_IB_CA)
  499. return -EOPNOTSUPP;
  500. smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
  501. if (!smcibdev)
  502. return -ENOMEM;
  503. smcibdev->ibdev = ibdev;
  504. INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
  505. atomic_set(&smcibdev->lnk_cnt, 0);
  506. init_waitqueue_head(&smcibdev->lnks_deleted);
  507. mutex_init(&smcibdev->mutex);
  508. mutex_lock(&smc_ib_devices.mutex);
  509. list_add_tail(&smcibdev->list, &smc_ib_devices.list);
  510. mutex_unlock(&smc_ib_devices.mutex);
  511. ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
  512. INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
  513. smc_ib_global_event_handler);
  514. ib_register_event_handler(&smcibdev->event_handler);
  515. /* trigger reading of the port attributes */
  516. port_cnt = smcibdev->ibdev->phys_port_cnt;
  517. pr_warn_ratelimited("smc: adding ib device %s with port count %d\n",
  518. smcibdev->ibdev->name, port_cnt);
  519. for (i = 0;
  520. i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
  521. i++) {
  522. set_bit(i, &smcibdev->port_event_mask);
  523. /* determine pnetids of the port */
  524. if (smc_pnetid_by_dev_port(ibdev->dev.parent, i,
  525. smcibdev->pnetid[i]))
  526. smc_pnetid_by_table_ib(smcibdev, i + 1);
  527. pr_warn_ratelimited("smc: ib device %s port %d has pnetid "
  528. "%.16s%s\n",
  529. smcibdev->ibdev->name, i + 1,
  530. smcibdev->pnetid[i],
  531. smcibdev->pnetid_by_user[i] ?
  532. " (user defined)" :
  533. "");
  534. }
  535. schedule_work(&smcibdev->port_event_work);
  536. return 0;
  537. }
  538. /* callback function for ib_unregister_client() */
  539. static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
  540. {
  541. struct smc_ib_device *smcibdev = client_data;
  542. mutex_lock(&smc_ib_devices.mutex);
  543. list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
  544. mutex_unlock(&smc_ib_devices.mutex);
  545. pr_warn_ratelimited("smc: removing ib device %s\n",
  546. smcibdev->ibdev->name);
  547. smc_smcr_terminate_all(smcibdev);
  548. smc_ib_cleanup_per_ibdev(smcibdev);
  549. ib_unregister_event_handler(&smcibdev->event_handler);
  550. cancel_work_sync(&smcibdev->port_event_work);
  551. kfree(smcibdev);
  552. }
  553. static struct ib_client smc_ib_client = {
  554. .name = "smc_ib",
  555. .add = smc_ib_add_dev,
  556. .remove = smc_ib_remove_dev,
  557. };
  558. int __init smc_ib_register_client(void)
  559. {
  560. smc_ib_init_local_systemid();
  561. return ib_register_client(&smc_ib_client);
  562. }
  563. void smc_ib_unregister_client(void)
  564. {
  565. ib_unregister_client(&smc_ib_client);
  566. }