2 * Copyright(c) 2016 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 #include <linux/slab.h>
49 #include <linux/vmalloc.h>
50 #include <rdma/ib_umem.h>
51 #include <rdma/rdma_vt.h>
57 * rvt_driver_mr_init - Init MR resources per driver
58 * @rdi: rvt dev struct
60 * Do any intilization needed when a driver registers with rdmavt.
62 * Return: 0 on success or errno on failure
64 int rvt_driver_mr_init(struct rvt_dev_info *rdi)
66 unsigned int lkey_table_size = rdi->dparms.lkey_table_size;
71 * The top hfi1_lkey_table_size bits are used to index the
72 * table. The lower 8 bits can be owned by the user (copied from
73 * the LKEY). The remaining bits act as a generation number or tag.
78 spin_lock_init(&rdi->lkey_table.lock);
80 /* ensure generation is at least 4 bits */
81 if (lkey_table_size > RVT_MAX_LKEY_TABLE_BITS) {
82 rvt_pr_warn(rdi, "lkey bits %u too large, reduced to %u\n",
83 lkey_table_size, RVT_MAX_LKEY_TABLE_BITS);
84 rdi->dparms.lkey_table_size = RVT_MAX_LKEY_TABLE_BITS;
85 lkey_table_size = rdi->dparms.lkey_table_size;
87 rdi->lkey_table.max = 1 << lkey_table_size;
88 rdi->lkey_table.shift = 32 - lkey_table_size;
89 lk_tab_size = rdi->lkey_table.max * sizeof(*rdi->lkey_table.table);
90 rdi->lkey_table.table = (struct rvt_mregion __rcu **)
91 vmalloc_node(lk_tab_size, rdi->dparms.node);
92 if (!rdi->lkey_table.table)
95 RCU_INIT_POINTER(rdi->dma_mr, NULL);
96 for (i = 0; i < rdi->lkey_table.max; i++)
97 RCU_INIT_POINTER(rdi->lkey_table.table[i], NULL);
103 *rvt_mr_exit: clean up MR
104 *@rdi: rvt dev structure
106 * called when drivers have unregistered or perhaps failed to register with us
108 void rvt_mr_exit(struct rvt_dev_info *rdi)
111 rvt_pr_err(rdi, "DMA MR not null!\n");
113 vfree(rdi->lkey_table.table);
116 static void rvt_deinit_mregion(struct rvt_mregion *mr)
123 percpu_ref_exit(&mr->refcount);
126 static void __rvt_mregion_complete(struct percpu_ref *ref)
128 struct rvt_mregion *mr = container_of(ref, struct rvt_mregion,
134 static int rvt_init_mregion(struct rvt_mregion *mr, struct ib_pd *pd,
135 int count, unsigned int percpu_flags)
138 struct rvt_dev_info *dev = ib_to_rvt(pd->device);
141 m = (count + RVT_SEGSZ - 1) / RVT_SEGSZ;
143 mr->map[i] = kzalloc_node(sizeof(*mr->map[0]), GFP_KERNEL,
149 init_completion(&mr->comp);
150 /* count returning the ptr to user */
151 if (percpu_ref_init(&mr->refcount, &__rvt_mregion_complete,
152 percpu_flags, GFP_KERNEL))
155 atomic_set(&mr->lkey_invalid, 0);
157 mr->max_segs = count;
160 rvt_deinit_mregion(mr);
165 * rvt_alloc_lkey - allocate an lkey
166 * @mr: memory region that this lkey protects
167 * @dma_region: 0->normal key, 1->restricted DMA key
169 * Returns 0 if successful, otherwise returns -errno.
171 * Increments mr reference count as required.
173 * Sets the lkey field mr for non-dma regions.
176 static int rvt_alloc_lkey(struct rvt_mregion *mr, int dma_region)
182 struct rvt_dev_info *dev = ib_to_rvt(mr->pd->device);
183 struct rvt_lkey_table *rkt = &dev->lkey_table;
186 spin_lock_irqsave(&rkt->lock, flags);
188 /* special case for dma_mr lkey == 0 */
190 struct rvt_mregion *tmr;
192 tmr = rcu_access_pointer(dev->dma_mr);
194 mr->lkey_published = 1;
195 /* Insure published written first */
196 rcu_assign_pointer(dev->dma_mr, mr);
202 /* Find the next available LKEY */
206 if (!rcu_access_pointer(rkt->table[r]))
208 r = (r + 1) & (rkt->max - 1);
212 rkt->next = (r + 1) & (rkt->max - 1);
214 * Make sure lkey is never zero which is reserved to indicate an
219 * bits are capped to ensure enough bits for generation number
221 mr->lkey = (r << (32 - dev->dparms.lkey_table_size)) |
222 ((((1 << (24 - dev->dparms.lkey_table_size)) - 1) & rkt->gen)
228 mr->lkey_published = 1;
229 /* Insure published written first */
230 rcu_assign_pointer(rkt->table[r], mr);
232 spin_unlock_irqrestore(&rkt->lock, flags);
237 spin_unlock_irqrestore(&rkt->lock, flags);
243 * rvt_free_lkey - free an lkey
244 * @mr: mr to free from tables
246 static void rvt_free_lkey(struct rvt_mregion *mr)
251 struct rvt_dev_info *dev = ib_to_rvt(mr->pd->device);
252 struct rvt_lkey_table *rkt = &dev->lkey_table;
255 spin_lock_irqsave(&rkt->lock, flags);
257 if (mr->lkey_published) {
258 mr->lkey_published = 0;
259 /* insure published is written before pointer */
260 rcu_assign_pointer(dev->dma_mr, NULL);
264 if (!mr->lkey_published)
266 r = lkey >> (32 - dev->dparms.lkey_table_size);
267 mr->lkey_published = 0;
268 /* insure published is written before pointer */
269 rcu_assign_pointer(rkt->table[r], NULL);
273 spin_unlock_irqrestore(&rkt->lock, flags);
275 percpu_ref_kill(&mr->refcount);
278 static struct rvt_mr *__rvt_alloc_mr(int count, struct ib_pd *pd)
284 /* Allocate struct plus pointers to first level page tables. */
285 m = (count + RVT_SEGSZ - 1) / RVT_SEGSZ;
286 mr = kzalloc(struct_size(mr, mr.map, m), GFP_KERNEL);
290 rval = rvt_init_mregion(&mr->mr, pd, count, 0);
294 * ib_reg_phys_mr() will initialize mr->ibmr except for
297 rval = rvt_alloc_lkey(&mr->mr, 0);
300 mr->ibmr.lkey = mr->mr.lkey;
301 mr->ibmr.rkey = mr->mr.lkey;
306 rvt_deinit_mregion(&mr->mr);
313 static void __rvt_free_mr(struct rvt_mr *mr)
315 rvt_free_lkey(&mr->mr);
316 rvt_deinit_mregion(&mr->mr);
321 * rvt_get_dma_mr - get a DMA memory region
322 * @pd: protection domain for this memory region
325 * Return: the memory region on success, otherwise returns an errno.
326 * Note that all DMA addresses should be created via the functions in
327 * struct dma_virt_ops.
329 struct ib_mr *rvt_get_dma_mr(struct ib_pd *pd, int acc)
335 if (ibpd_to_rvtpd(pd)->user)
336 return ERR_PTR(-EPERM);
338 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
340 ret = ERR_PTR(-ENOMEM);
344 rval = rvt_init_mregion(&mr->mr, pd, 0, 0);
350 rval = rvt_alloc_lkey(&mr->mr, 1);
356 mr->mr.access_flags = acc;
362 rvt_deinit_mregion(&mr->mr);
369 * rvt_reg_user_mr - register a userspace memory region
370 * @pd: protection domain for this memory region
371 * @start: starting userspace address
372 * @length: length of region to register
373 * @mr_access_flags: access flags for this memory region
374 * @udata: unused by the driver
376 * Return: the memory region on success, otherwise returns an errno.
378 struct ib_mr *rvt_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
379 u64 virt_addr, int mr_access_flags,
380 struct ib_udata *udata)
383 struct ib_umem *umem;
384 struct sg_page_iter sg_iter;
389 return ERR_PTR(-EINVAL);
391 umem = ib_umem_get(udata, start, length, mr_access_flags, 0);
397 mr = __rvt_alloc_mr(n, pd);
399 ret = (struct ib_mr *)mr;
403 mr->mr.user_base = start;
404 mr->mr.iova = virt_addr;
405 mr->mr.length = length;
406 mr->mr.offset = ib_umem_offset(umem);
407 mr->mr.access_flags = mr_access_flags;
410 mr->mr.page_shift = PAGE_SHIFT;
413 for_each_sg_page (umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
416 vaddr = page_address(sg_page_iter_page(&sg_iter));
418 ret = ERR_PTR(-EINVAL);
421 mr->mr.map[m]->segs[n].vaddr = vaddr;
422 mr->mr.map[m]->segs[n].length = PAGE_SIZE;
423 trace_rvt_mr_user_seg(&mr->mr, m, n, vaddr, PAGE_SIZE);
424 if (++n == RVT_SEGSZ) {
435 ib_umem_release(umem);
441 * rvt_dereg_clean_qp_cb - callback from iterator
443 * @v - the mregion (as u64)
445 * This routine fields the callback for all QPs and
446 * for QPs in the same PD as the MR will call the
447 * rvt_qp_mr_clean() to potentially cleanup references.
449 static void rvt_dereg_clean_qp_cb(struct rvt_qp *qp, u64 v)
451 struct rvt_mregion *mr = (struct rvt_mregion *)v;
453 /* skip PDs that are not ours */
454 if (mr->pd != qp->ibqp.pd)
456 rvt_qp_mr_clean(qp, mr->lkey);
460 * rvt_dereg_clean_qps - find QPs for reference cleanup
461 * @mr - the MR that is being deregistered
463 * This routine iterates RC QPs looking for references
464 * to the lkey noted in mr.
466 static void rvt_dereg_clean_qps(struct rvt_mregion *mr)
468 struct rvt_dev_info *rdi = ib_to_rvt(mr->pd->device);
470 rvt_qp_iter(rdi, (u64)mr, rvt_dereg_clean_qp_cb);
474 * rvt_check_refs - check references
476 * @t - the caller identification
478 * This routine checks MRs holding a reference during
479 * when being de-registered.
481 * If the count is non-zero, the code calls a clean routine then
482 * waits for the timeout for the count to zero.
484 static int rvt_check_refs(struct rvt_mregion *mr, const char *t)
486 unsigned long timeout;
487 struct rvt_dev_info *rdi = ib_to_rvt(mr->pd->device);
491 rvt_dereg_clean_qps(mr);
492 /* @mr was indexed on rcu protected @lkey_table */
496 timeout = wait_for_completion_timeout(&mr->comp, 5 * HZ);
499 "%s timeout mr %p pd %p lkey %x refcount %ld\n",
500 t, mr, mr->pd, mr->lkey,
501 atomic_long_read(&mr->refcount.count));
509 * rvt_mr_has_lkey - is MR
513 bool rvt_mr_has_lkey(struct rvt_mregion *mr, u32 lkey)
515 return mr && lkey == mr->lkey;
519 * rvt_ss_has_lkey - is mr in sge tests
520 * @ss - the sge state
523 * This code tests for an MR in the indicated
526 bool rvt_ss_has_lkey(struct rvt_sge_state *ss, u32 lkey)
534 rval = rvt_mr_has_lkey(ss->sge.mr, lkey);
536 for (i = 0; !rval && i < ss->num_sge - 1; i++)
537 rval = rvt_mr_has_lkey(ss->sg_list[i].mr, lkey);
542 * rvt_dereg_mr - unregister and free a memory region
543 * @ibmr: the memory region to free
546 * Note that this is called to free MRs created by rvt_get_dma_mr()
547 * or rvt_reg_user_mr().
549 * Returns 0 on success.
551 int rvt_dereg_mr(struct ib_mr *ibmr)
553 struct rvt_mr *mr = to_imr(ibmr);
556 rvt_free_lkey(&mr->mr);
558 rvt_put_mr(&mr->mr); /* will set completion if last */
559 ret = rvt_check_refs(&mr->mr, __func__);
562 rvt_deinit_mregion(&mr->mr);
564 ib_umem_release(mr->umem);
571 * rvt_alloc_mr - Allocate a memory region usable with the
572 * @pd: protection domain for this memory region
573 * @mr_type: mem region type
574 * @max_num_sg: Max number of segments allowed
576 * Return: the memory region on success, otherwise return an errno.
578 struct ib_mr *rvt_alloc_mr(struct ib_pd *pd,
579 enum ib_mr_type mr_type,
584 if (mr_type != IB_MR_TYPE_MEM_REG)
585 return ERR_PTR(-EINVAL);
587 mr = __rvt_alloc_mr(max_num_sg, pd);
589 return (struct ib_mr *)mr;
595 * rvt_set_page - page assignment function called by ib_sg_to_pages
596 * @ibmr: memory region
597 * @addr: dma address of mapped page
599 * Return: 0 on success
601 static int rvt_set_page(struct ib_mr *ibmr, u64 addr)
603 struct rvt_mr *mr = to_imr(ibmr);
604 u32 ps = 1 << mr->mr.page_shift;
605 u32 mapped_segs = mr->mr.length >> mr->mr.page_shift;
608 if (unlikely(mapped_segs == mr->mr.max_segs))
611 m = mapped_segs / RVT_SEGSZ;
612 n = mapped_segs % RVT_SEGSZ;
613 mr->mr.map[m]->segs[n].vaddr = (void *)addr;
614 mr->mr.map[m]->segs[n].length = ps;
615 trace_rvt_mr_page_seg(&mr->mr, m, n, (void *)addr, ps);
622 * rvt_map_mr_sg - map sg list and set it the memory region
623 * @ibmr: memory region
624 * @sg: dma mapped scatterlist
625 * @sg_nents: number of entries in sg
626 * @sg_offset: offset in bytes into sg
628 * Overwrite rvt_mr length with mr length calculated by ib_sg_to_pages.
630 * Return: number of sg elements mapped to the memory region
632 int rvt_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
633 int sg_nents, unsigned int *sg_offset)
635 struct rvt_mr *mr = to_imr(ibmr);
639 mr->mr.page_shift = PAGE_SHIFT;
640 ret = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rvt_set_page);
641 mr->mr.user_base = ibmr->iova;
642 mr->mr.iova = ibmr->iova;
643 mr->mr.offset = ibmr->iova - (u64)mr->mr.map[0]->segs[0].vaddr;
644 mr->mr.length = (size_t)ibmr->length;
649 * rvt_fast_reg_mr - fast register physical MR
650 * @qp: the queue pair where the work request comes from
651 * @ibmr: the memory region to be registered
652 * @key: updated key for this memory region
653 * @access: access flags for this memory region
655 * Returns 0 on success.
657 int rvt_fast_reg_mr(struct rvt_qp *qp, struct ib_mr *ibmr, u32 key,
660 struct rvt_mr *mr = to_imr(ibmr);
662 if (qp->ibqp.pd != mr->mr.pd)
665 /* not applicable to dma MR or user MR */
666 if (!mr->mr.lkey || mr->umem)
669 if ((key & 0xFFFFFF00) != (mr->mr.lkey & 0xFFFFFF00))
675 mr->mr.access_flags = access;
676 mr->mr.iova = ibmr->iova;
677 atomic_set(&mr->mr.lkey_invalid, 0);
681 EXPORT_SYMBOL(rvt_fast_reg_mr);
684 * rvt_invalidate_rkey - invalidate an MR rkey
685 * @qp: queue pair associated with the invalidate op
686 * @rkey: rkey to invalidate
688 * Returns 0 on success.
690 int rvt_invalidate_rkey(struct rvt_qp *qp, u32 rkey)
692 struct rvt_dev_info *dev = ib_to_rvt(qp->ibqp.device);
693 struct rvt_lkey_table *rkt = &dev->lkey_table;
694 struct rvt_mregion *mr;
700 mr = rcu_dereference(
701 rkt->table[(rkey >> (32 - dev->dparms.lkey_table_size))]);
702 if (unlikely(!mr || mr->lkey != rkey || qp->ibqp.pd != mr->pd))
705 atomic_set(&mr->lkey_invalid, 1);
713 EXPORT_SYMBOL(rvt_invalidate_rkey);
716 * rvt_alloc_fmr - allocate a fast memory region
717 * @pd: the protection domain for this memory region
718 * @mr_access_flags: access flags for this memory region
719 * @fmr_attr: fast memory region attributes
721 * Return: the memory region on success, otherwise returns an errno.
723 struct ib_fmr *rvt_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
724 struct ib_fmr_attr *fmr_attr)
731 /* Allocate struct plus pointers to first level page tables. */
732 m = (fmr_attr->max_pages + RVT_SEGSZ - 1) / RVT_SEGSZ;
733 fmr = kzalloc(struct_size(fmr, mr.map, m), GFP_KERNEL);
737 rval = rvt_init_mregion(&fmr->mr, pd, fmr_attr->max_pages,
738 PERCPU_REF_INIT_ATOMIC);
743 * ib_alloc_fmr() will initialize fmr->ibfmr except for lkey &
746 rval = rvt_alloc_lkey(&fmr->mr, 0);
749 fmr->ibfmr.rkey = fmr->mr.lkey;
750 fmr->ibfmr.lkey = fmr->mr.lkey;
752 * Resources are allocated but no valid mapping (RKEY can't be
755 fmr->mr.access_flags = mr_access_flags;
756 fmr->mr.max_segs = fmr_attr->max_pages;
757 fmr->mr.page_shift = fmr_attr->page_shift;
764 rvt_deinit_mregion(&fmr->mr);
772 * rvt_map_phys_fmr - set up a fast memory region
773 * @ibfmr: the fast memory region to set up
774 * @page_list: the list of pages to associate with the fast memory region
775 * @list_len: the number of pages to associate with the fast memory region
776 * @iova: the virtual address of the start of the fast memory region
778 * This may be called from interrupt context.
780 * Return: 0 on success
783 int rvt_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
784 int list_len, u64 iova)
786 struct rvt_fmr *fmr = to_ifmr(ibfmr);
787 struct rvt_lkey_table *rkt;
792 struct rvt_dev_info *rdi = ib_to_rvt(ibfmr->device);
794 i = atomic_long_read(&fmr->mr.refcount.count);
798 if (list_len > fmr->mr.max_segs)
801 rkt = &rdi->lkey_table;
802 spin_lock_irqsave(&rkt->lock, flags);
803 fmr->mr.user_base = iova;
805 ps = 1 << fmr->mr.page_shift;
806 fmr->mr.length = list_len * ps;
809 for (i = 0; i < list_len; i++) {
810 fmr->mr.map[m]->segs[n].vaddr = (void *)page_list[i];
811 fmr->mr.map[m]->segs[n].length = ps;
812 trace_rvt_mr_fmr_seg(&fmr->mr, m, n, (void *)page_list[i], ps);
813 if (++n == RVT_SEGSZ) {
818 spin_unlock_irqrestore(&rkt->lock, flags);
823 * rvt_unmap_fmr - unmap fast memory regions
824 * @fmr_list: the list of fast memory regions to unmap
826 * Return: 0 on success.
828 int rvt_unmap_fmr(struct list_head *fmr_list)
831 struct rvt_lkey_table *rkt;
833 struct rvt_dev_info *rdi;
835 list_for_each_entry(fmr, fmr_list, ibfmr.list) {
836 rdi = ib_to_rvt(fmr->ibfmr.device);
837 rkt = &rdi->lkey_table;
838 spin_lock_irqsave(&rkt->lock, flags);
839 fmr->mr.user_base = 0;
842 spin_unlock_irqrestore(&rkt->lock, flags);
848 * rvt_dealloc_fmr - deallocate a fast memory region
849 * @ibfmr: the fast memory region to deallocate
851 * Return: 0 on success.
853 int rvt_dealloc_fmr(struct ib_fmr *ibfmr)
855 struct rvt_fmr *fmr = to_ifmr(ibfmr);
858 rvt_free_lkey(&fmr->mr);
859 rvt_put_mr(&fmr->mr); /* will set completion if last */
860 ret = rvt_check_refs(&fmr->mr, __func__);
863 rvt_deinit_mregion(&fmr->mr);
870 * rvt_sge_adjacent - is isge compressible
871 * @last_sge: last outgoing SGE written
874 * If adjacent will update last_sge to add length.
876 * Return: true if isge is adjacent to last sge
878 static inline bool rvt_sge_adjacent(struct rvt_sge *last_sge,
881 if (last_sge && sge->lkey == last_sge->mr->lkey &&
882 ((uint64_t)(last_sge->vaddr + last_sge->length) == sge->addr)) {
884 if (unlikely((sge->addr - last_sge->mr->user_base +
885 sge->length > last_sge->mr->length)))
886 return false; /* overrun, caller will catch */
888 last_sge->length += sge->length;
890 last_sge->sge_length += sge->length;
891 trace_rvt_sge_adjacent(last_sge, sge);
898 * rvt_lkey_ok - check IB SGE for validity and initialize
899 * @rkt: table containing lkey to check SGE against
900 * @pd: protection domain
901 * @isge: outgoing internal SGE
902 * @last_sge: last outgoing SGE written
906 * Check the IB SGE for validity and initialize our internal version
909 * Increments the reference count when a new sge is stored.
911 * Return: 0 if compressed, 1 if added , otherwise returns -errno.
913 int rvt_lkey_ok(struct rvt_lkey_table *rkt, struct rvt_pd *pd,
914 struct rvt_sge *isge, struct rvt_sge *last_sge,
915 struct ib_sge *sge, int acc)
917 struct rvt_mregion *mr;
922 * We use LKEY == zero for kernel virtual addresses
923 * (see rvt_get_dma_mr() and dma_virt_ops).
925 if (sge->lkey == 0) {
926 struct rvt_dev_info *dev = ib_to_rvt(pd->ibpd.device);
930 if (rvt_sge_adjacent(last_sge, sge))
933 mr = rcu_dereference(dev->dma_mr);
940 isge->vaddr = (void *)sge->addr;
941 isge->length = sge->length;
942 isge->sge_length = sge->length;
947 if (rvt_sge_adjacent(last_sge, sge))
950 mr = rcu_dereference(rkt->table[sge->lkey >> rkt->shift]);
954 if (!READ_ONCE(mr->lkey_published))
957 if (unlikely(atomic_read(&mr->lkey_invalid) ||
958 mr->lkey != sge->lkey || mr->pd != &pd->ibpd))
961 off = sge->addr - mr->user_base;
962 if (unlikely(sge->addr < mr->user_base ||
963 off + sge->length > mr->length ||
964 (mr->access_flags & acc) != acc))
969 if (mr->page_shift) {
971 * page sizes are uniform power of 2 so no loop is necessary
972 * entries_spanned_by_off is the number of times the loop below
973 * would have executed.
975 size_t entries_spanned_by_off;
977 entries_spanned_by_off = off >> mr->page_shift;
978 off -= (entries_spanned_by_off << mr->page_shift);
979 m = entries_spanned_by_off / RVT_SEGSZ;
980 n = entries_spanned_by_off % RVT_SEGSZ;
984 while (off >= mr->map[m]->segs[n].length) {
985 off -= mr->map[m]->segs[n].length;
987 if (n >= RVT_SEGSZ) {
994 isge->vaddr = mr->map[m]->segs[n].vaddr + off;
995 isge->length = mr->map[m]->segs[n].length - off;
996 isge->sge_length = sge->length;
1000 trace_rvt_sge_new(isge, sge);
1008 EXPORT_SYMBOL(rvt_lkey_ok);
1011 * rvt_rkey_ok - check the IB virtual address, length, and RKEY
1012 * @qp: qp for validation
1014 * @len: length of data
1015 * @vaddr: virtual address to place data
1016 * @rkey: rkey to check
1017 * @acc: access flags
1019 * Return: 1 if successful, otherwise 0.
1021 * increments the reference count upon success
1023 int rvt_rkey_ok(struct rvt_qp *qp, struct rvt_sge *sge,
1024 u32 len, u64 vaddr, u32 rkey, int acc)
1026 struct rvt_dev_info *dev = ib_to_rvt(qp->ibqp.device);
1027 struct rvt_lkey_table *rkt = &dev->lkey_table;
1028 struct rvt_mregion *mr;
1033 * We use RKEY == zero for kernel virtual addresses
1034 * (see rvt_get_dma_mr() and dma_virt_ops).
1038 struct rvt_pd *pd = ibpd_to_rvtpd(qp->ibqp.pd);
1039 struct rvt_dev_info *rdi = ib_to_rvt(pd->ibpd.device);
1043 mr = rcu_dereference(rdi->dma_mr);
1050 sge->vaddr = (void *)vaddr;
1052 sge->sge_length = len;
1058 mr = rcu_dereference(rkt->table[rkey >> rkt->shift]);
1062 /* insure mr read is before test */
1063 if (!READ_ONCE(mr->lkey_published))
1065 if (unlikely(atomic_read(&mr->lkey_invalid) ||
1066 mr->lkey != rkey || qp->ibqp.pd != mr->pd))
1069 off = vaddr - mr->iova;
1070 if (unlikely(vaddr < mr->iova || off + len > mr->length ||
1071 (mr->access_flags & acc) == 0))
1076 if (mr->page_shift) {
1078 * page sizes are uniform power of 2 so no loop is necessary
1079 * entries_spanned_by_off is the number of times the loop below
1080 * would have executed.
1082 size_t entries_spanned_by_off;
1084 entries_spanned_by_off = off >> mr->page_shift;
1085 off -= (entries_spanned_by_off << mr->page_shift);
1086 m = entries_spanned_by_off / RVT_SEGSZ;
1087 n = entries_spanned_by_off % RVT_SEGSZ;
1091 while (off >= mr->map[m]->segs[n].length) {
1092 off -= mr->map[m]->segs[n].length;
1094 if (n >= RVT_SEGSZ) {
1101 sge->vaddr = mr->map[m]->segs[n].vaddr + off;
1102 sge->length = mr->map[m]->segs[n].length - off;
1103 sge->sge_length = len;
1114 EXPORT_SYMBOL(rvt_rkey_ok);