5 * Copyright(c) 2016 - 2018 Intel Corporation.
7 * This file is provided under a dual BSD/GPLv2 license. When using or
8 * redistributing this file, you may do so under either license.
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of version 2 of the GNU General Public License as
14 * published by the Free Software Foundation.
16 * This program is distributed in the hope that it will be useful, but
17 * WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * General Public License for more details.
23 * Redistribution and use in source and binary forms, with or without
24 * modification, are permitted provided that the following conditions
27 * - Redistributions of source code must retain the above copyright
28 * notice, this list of conditions and the following disclaimer.
29 * - Redistributions in binary form must reproduce the above copyright
30 * notice, this list of conditions and the following disclaimer in
31 * the documentation and/or other materials provided with the
33 * - Neither the name of Intel Corporation nor the names of its
34 * contributors may be used to endorse or promote products derived
35 * from this software without specific prior written permission.
37 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
38 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
39 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
40 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
41 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
42 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
43 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
44 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
45 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
46 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
47 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
52 * Structure that low level drivers will populate in order to register with the
56 #include <linux/spinlock.h>
57 #include <linux/list.h>
58 #include <linux/hash.h>
59 #include <rdma/ib_verbs.h>
60 #include <rdma/ib_mad.h>
61 #include <rdma/rdmavt_mr.h>
62 #include <rdma/rdmavt_qp.h>
64 #define RVT_MAX_PKEY_VALUES 16
66 #define RVT_MAX_TRAP_LEN 100 /* Limit pending trap list */
67 #define RVT_MAX_TRAP_LISTS 5 /*((IB_NOTICE_TYPE_INFO & 0x0F) + 1)*/
68 #define RVT_TRAP_TIMEOUT 4096 /* 4.096 usec */
72 struct list_head list;
76 struct rvt_qp __rcu *qp[2];
77 struct ib_mad_agent *send_agent; /* agent for SMI (traps) */
78 struct rb_root mcast_tree;
79 spinlock_t lock; /* protect changes in this struct */
81 /* non-zero when timer is set */
82 unsigned long mkey_lease_timeout;
83 unsigned long trap_timeout;
84 __be64 gid_prefix; /* in network order */
90 u32 pma_sample_interval;
91 __be16 pma_counter_select[5];
93 u16 mkey_lease_period;
101 * Driver is expected to keep these up to date. These
102 * counters are informational only and not required to be
103 * completely accurate.
122 /* Hot-path per CPU counters to avoid cacheline trading to update */
125 u64 z_rc_delayed_comp;
126 u64 __percpu *rc_acks;
127 u64 __percpu *rc_qacks;
128 u64 __percpu *rc_delayed_comp;
130 void *priv; /* driver private data */
133 * The pkey table is allocated and maintained by the driver. Drivers
134 * need to have access to this before registering with rdmav. However
135 * rdmavt will need access to it so drivers need to proviee this during
136 * the attach port API call.
140 struct rvt_ah *sm_ah;
143 * Keep a list of traps that have not been repressed. They will be
144 * resent based on trap_timer.
146 struct trap_list trap_lists[RVT_MAX_TRAP_LISTS];
147 struct timer_list trap_timer;
150 #define RVT_CQN_MAX 16 /* maximum length of cq name */
152 #define RVT_SGE_COPY_MEMCPY 0
153 #define RVT_SGE_COPY_CACHELESS 1
154 #define RVT_SGE_COPY_ADAPTIVE 2
157 * Things that are driver specific, module parameters in hfi1 and qib
159 struct rvt_driver_params {
160 struct ib_device_attr props;
163 * Anything driver specific that is not covered by props
164 * For instance special module parameters. Goes here.
166 unsigned int lkey_table_size;
167 unsigned int qp_table_size;
168 unsigned int sge_copy_mode;
169 unsigned int wss_threshold;
170 unsigned int wss_clean_period;
185 u8 reserved_operations;
188 /* Protection domain */
197 struct rdma_ah_attr attr;
203 /* memory working set size */
205 unsigned long *entries;
206 atomic_t total_count;
207 atomic_t clean_counter;
208 atomic_t clean_entry;
213 unsigned int clean_period;
218 struct rvt_driver_provided {
220 * Which functions are required depends on which verbs rdmavt is
221 * providing and which verbs the driver is overriding. See
222 * check_support() for details.
225 /* hot path calldowns in a single cacheline */
228 * Give the driver a notice that there is send work to do. It is up to
229 * the driver to generally push the packets out, this just queues the
230 * work with the driver. There are two variants here. The no_lock
231 * version requires the s_lock not to be held. The other assumes the
234 bool (*schedule_send)(struct rvt_qp *qp);
235 bool (*schedule_send_no_lock)(struct rvt_qp *qp);
238 * Driver specific work request setup and checking.
239 * This function is allowed to perform any setup, checks, or
240 * adjustments required to the SWQE in order to be usable by
241 * underlying protocols. This includes private data structure
244 int (*setup_wqe)(struct rvt_qp *qp, struct rvt_swqe *wqe,
248 * Sometimes rdmavt needs to kick the driver's send progress. That is
249 * done by this call back.
251 void (*do_send)(struct rvt_qp *qp);
253 /* Passed to ib core registration. Callback to create syfs files */
254 int (*port_callback)(struct ib_device *, u8, struct kobject *);
257 * Returns a pointer to the undelying hardware's PCI device. This is
258 * used to display information as to what hardware is being referenced
259 * in an output message
261 struct pci_dev * (*get_pci_dev)(struct rvt_dev_info *rdi);
264 * Allocate a private queue pair data structure for driver specific
265 * information which is opaque to rdmavt. Errors are returned via
266 * ERR_PTR(err). The driver is free to return NULL or a valid
269 void * (*qp_priv_alloc)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
272 * Free the driver's private qp structure.
274 void (*qp_priv_free)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
277 * Inform the driver the particular qp in quesiton has been reset so
278 * that it can clean up anything it needs to.
280 void (*notify_qp_reset)(struct rvt_qp *qp);
283 * Get a path mtu from the driver based on qp attributes.
285 int (*get_pmtu_from_attr)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
286 struct ib_qp_attr *attr);
289 * Notify driver that it needs to flush any outstanding IO requests that
290 * are waiting on a qp.
292 void (*flush_qp_waiters)(struct rvt_qp *qp);
295 * Notify driver to stop its queue of sending packets. Nothing else
296 * should be posted to the queue pair after this has been called.
298 void (*stop_send_queue)(struct rvt_qp *qp);
301 * Have the drivr drain any in progress operations
303 void (*quiesce_qp)(struct rvt_qp *qp);
306 * Inform the driver a qp has went to error state.
308 void (*notify_error_qp)(struct rvt_qp *qp);
311 * Get an MTU for a qp.
313 u32 (*mtu_from_qp)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
316 * Convert an mtu to a path mtu
318 int (*mtu_to_path_mtu)(u32 mtu);
321 * Get the guid of a port in big endian byte order
323 int (*get_guid_be)(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
324 int guid_index, __be64 *guid);
327 * Query driver for the state of the port.
329 int (*query_port_state)(struct rvt_dev_info *rdi, u8 port_num,
330 struct ib_port_attr *props);
333 * Tell driver to shutdown a port
335 int (*shut_down_port)(struct rvt_dev_info *rdi, u8 port_num);
337 /* Tell driver to send a trap for changed port capabilities */
338 void (*cap_mask_chg)(struct rvt_dev_info *rdi, u8 port_num);
341 * The following functions can be safely ignored completely. Any use of
342 * these is checked for NULL before blindly calling. Rdmavt should also
343 * be functional if drivers omit these.
346 /* Called to inform the driver that all qps should now be freed. */
347 unsigned (*free_all_qps)(struct rvt_dev_info *rdi);
349 /* Driver specific AH validation */
350 int (*check_ah)(struct ib_device *, struct rdma_ah_attr *);
352 /* Inform the driver a new AH has been created */
353 void (*notify_new_ah)(struct ib_device *, struct rdma_ah_attr *,
356 /* Let the driver pick the next queue pair number*/
357 int (*alloc_qpn)(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
358 enum ib_qp_type type, u8 port_num);
360 /* Determine if its safe or allowed to modify the qp */
361 int (*check_modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr,
362 int attr_mask, struct ib_udata *udata);
364 /* Driver specific QP modification/notification-of */
365 void (*modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr,
366 int attr_mask, struct ib_udata *udata);
368 /* Notify driver a mad agent has been created */
369 void (*notify_create_mad_agent)(struct rvt_dev_info *rdi, int port_idx);
371 /* Notify driver a mad agent has been removed */
372 void (*notify_free_mad_agent)(struct rvt_dev_info *rdi, int port_idx);
374 /* Notify driver to restart rc */
375 void (*notify_restart_rc)(struct rvt_qp *qp, u32 psn, int wait);
377 /* Get and return CPU to pin CQ processing thread */
378 int (*comp_vect_cpu_lookup)(struct rvt_dev_info *rdi, int comp_vect);
381 struct rvt_dev_info {
382 struct ib_device ibdev; /* Keep this first. Nothing above here */
385 * Prior to calling for registration the driver will be responsible for
386 * allocating space for this structure.
388 * The driver will also be responsible for filling in certain members of
389 * dparms.props. The driver needs to fill in dparms exactly as it would
390 * want values reported to a ULP. This will be returned to the caller
391 * in rdmavt's device. The driver should also therefore refrain from
392 * modifying this directly after registration with rdmavt.
395 /* Driver specific properties */
396 struct rvt_driver_params dparms;
398 /* post send table */
399 const struct rvt_operation_params *post_parms;
401 /* opcode translation table */
402 const enum ib_wc_opcode *wc_opcode;
404 /* Driver specific helper functions */
405 struct rvt_driver_provided driver_f;
407 struct rvt_mregion __rcu *dma_mr;
408 struct rvt_lkey_table lkey_table;
412 spinlock_t n_pds_lock; /* Protect pd allocated count */
415 spinlock_t n_ahs_lock; /* Protect ah allocated count */
417 u32 n_srqs_allocated;
418 spinlock_t n_srqs_lock; /* Protect srqs allocated count */
421 struct rvt_ibport **ports;
424 struct rvt_qp_ibdev *qp_dev;
425 u32 n_qps_allocated; /* number of QPs allocated for device */
426 u32 n_rc_qps; /* number of RC QPs allocated for device */
427 u32 busy_jiffies; /* timeout scaling based on RC QP count */
428 spinlock_t n_qps_lock; /* protect qps, rc qps and busy jiffy counts */
431 struct list_head pending_mmaps;
432 spinlock_t mmap_offset_lock; /* protect mmap_offset */
434 spinlock_t pending_lock; /* protect pending mmap list */
437 u32 n_cqs_allocated; /* number of CQs allocated for device */
438 spinlock_t n_cqs_lock; /* protect count of in use cqs */
441 u32 n_mcast_grps_allocated; /* number of mcast groups allocated */
442 spinlock_t n_mcast_grps_lock;
444 /* Memory Working Set Size */
449 * rvt_set_ibdev_name - Craft an IB device name from client info
450 * @rdi: pointer to the client rvt_dev_info structure
451 * @name: client specific name
452 * @unit: client specific unit number.
454 static inline void rvt_set_ibdev_name(struct rvt_dev_info *rdi,
455 const char *fmt, const char *name,
459 * FIXME: rvt and its users want to touch the ibdev before
460 * registration and have things like the name work. We don't have the
461 * infrastructure in the core to support this directly today, hack it
462 * to work by setting the name manually here.
464 dev_set_name(&rdi->ibdev.dev, fmt, name, unit);
465 strlcpy(rdi->ibdev.name, dev_name(&rdi->ibdev.dev), IB_DEVICE_NAME_MAX);
469 * rvt_get_ibdev_name - return the IB name
470 * @rdi: rdmavt device
472 * Return the registered name of the device.
474 static inline const char *rvt_get_ibdev_name(const struct rvt_dev_info *rdi)
476 return dev_name(&rdi->ibdev.dev);
479 static inline struct rvt_pd *ibpd_to_rvtpd(struct ib_pd *ibpd)
481 return container_of(ibpd, struct rvt_pd, ibpd);
484 static inline struct rvt_ah *ibah_to_rvtah(struct ib_ah *ibah)
486 return container_of(ibah, struct rvt_ah, ibah);
489 static inline struct rvt_dev_info *ib_to_rvt(struct ib_device *ibdev)
491 return container_of(ibdev, struct rvt_dev_info, ibdev);
494 static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq)
496 return container_of(ibsrq, struct rvt_srq, ibsrq);
499 static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp)
501 return container_of(ibqp, struct rvt_qp, ibqp);
504 static inline unsigned rvt_get_npkeys(struct rvt_dev_info *rdi)
507 * All ports have same number of pkeys.
509 return rdi->dparms.npkeys;
513 * Return the max atomic suitable for determining
514 * the size of the ack ring buffer in a QP.
516 static inline unsigned int rvt_max_atomic(struct rvt_dev_info *rdi)
518 return rdi->dparms.max_rdma_atomic + 1;
522 * Return the indexed PKEY from the port PKEY table.
524 static inline u16 rvt_get_pkey(struct rvt_dev_info *rdi,
528 if (index >= rvt_get_npkeys(rdi))
531 return rdi->ports[port_index]->pkey_table[index];
535 * rvt_lookup_qpn - return the QP with the given QPN
537 * @qpn: the QP number to look up
539 * The caller must hold the rcu_read_lock(), and keep the lock until
540 * the returned qp is no longer in use.
542 /* TODO: Remove this and put in rdmavt/qp.h when no longer needed by drivers */
543 static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi,
544 struct rvt_ibport *rvp,
545 u32 qpn) __must_hold(RCU)
547 struct rvt_qp *qp = NULL;
549 if (unlikely(qpn <= 1)) {
550 qp = rcu_dereference(rvp->qp[qpn]);
552 u32 n = hash_32(qpn, rdi->qp_dev->qp_table_bits);
554 for (qp = rcu_dereference(rdi->qp_dev->qp_table[n]); qp;
555 qp = rcu_dereference(qp->next))
556 if (qp->ibqp.qp_num == qpn)
563 * rvt_mod_retry_timer - mod a retry timer
565 * Modify a potentially already running retry timer
567 static inline void rvt_mod_retry_timer(struct rvt_qp *qp)
569 struct ib_qp *ibqp = &qp->ibqp;
570 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
572 lockdep_assert_held(&qp->s_lock);
573 qp->s_flags |= RVT_S_TIMER;
574 /* 4.096 usec. * (1 << qp->timeout) */
575 mod_timer(&qp->s_timer, jiffies + qp->timeout_jiffies +
579 struct rvt_dev_info *rvt_alloc_device(size_t size, int nports);
580 void rvt_dealloc_device(struct rvt_dev_info *rdi);
581 int rvt_register_device(struct rvt_dev_info *rvd, u32 driver_id);
582 void rvt_unregister_device(struct rvt_dev_info *rvd);
583 int rvt_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr);
584 int rvt_init_port(struct rvt_dev_info *rdi, struct rvt_ibport *port,
585 int port_index, u16 *pkey_table);
586 int rvt_fast_reg_mr(struct rvt_qp *qp, struct ib_mr *ibmr, u32 key,
588 int rvt_invalidate_rkey(struct rvt_qp *qp, u32 rkey);
589 int rvt_rkey_ok(struct rvt_qp *qp, struct rvt_sge *sge,
590 u32 len, u64 vaddr, u32 rkey, int acc);
591 int rvt_lkey_ok(struct rvt_lkey_table *rkt, struct rvt_pd *pd,
592 struct rvt_sge *isge, struct rvt_sge *last_sge,
593 struct ib_sge *sge, int acc);
594 struct rvt_mcast *rvt_mcast_find(struct rvt_ibport *ibp, union ib_gid *mgid,
597 #endif /* DEF_RDMA_VT_H */