diff options
| -rw-r--r-- | drivers/infiniband/hw/hfi1/hfi.h | 2 | ||||
| -rw-r--r-- | drivers/infiniband/hw/hfi1/init.c | 3 | ||||
| -rw-r--r-- | drivers/infiniband/hw/hfi1/tid_rdma.c | 877 | ||||
| -rw-r--r-- | drivers/infiniband/hw/hfi1/tid_rdma.h | 101 | ||||
| -rw-r--r-- | drivers/infiniband/hw/hfi1/user_exp_rcv.h | 1 | ||||
| -rw-r--r-- | drivers/infiniband/hw/hfi1/verbs.c | 29 | ||||
| -rw-r--r-- | drivers/infiniband/hw/hfi1/verbs.h | 34 | ||||
| -rw-r--r-- | drivers/infiniband/sw/rdmavt/qp.c | 2 | ||||
| -rw-r--r-- | include/rdma/rdmavt_qp.h | 2 |
9 files changed, 1033 insertions, 18 deletions
diff --git a/drivers/infiniband/hw/hfi1/hfi.h b/drivers/infiniband/hw/hfi1/hfi.h index 78aa344c7403..1412ed157c98 100644 --- a/drivers/infiniband/hw/hfi1/hfi.h +++ b/drivers/infiniband/hw/hfi1/hfi.h @@ -303,6 +303,8 @@ struct hfi1_ctxtdata { spinlock_t exp_lock; /* Queue for QP's waiting for HW TID flows */ struct tid_queue flow_queue; + /* Queue for QP's waiting for HW receive array entries */ + struct tid_queue rarr_queue; /* when waiting for rcv or pioavail */ wait_queue_head_t wait; /* uuid from PSM */ diff --git a/drivers/infiniband/hw/hfi1/init.c b/drivers/infiniband/hw/hfi1/init.c index 56830a514b92..d13304f7340d 100644 --- a/drivers/infiniband/hw/hfi1/init.c +++ b/drivers/infiniband/hw/hfi1/init.c @@ -372,6 +372,7 @@ int hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, int numa, mutex_init(&rcd->exp_mutex); spin_lock_init(&rcd->exp_lock); INIT_LIST_HEAD(&rcd->flow_queue.queue_head); + INIT_LIST_HEAD(&rcd->rarr_queue.queue_head); hfi1_cdbg(PROC, "setting up context %u\n", rcd->ctxt); @@ -1596,7 +1597,7 @@ static void cleanup_device_data(struct hfi1_devdata *dd) struct hfi1_ctxtdata *rcd = dd->rcd[ctxt]; if (rcd) { - hfi1_clear_tids(rcd); + hfi1_free_ctxt_rcv_groups(rcd); hfi1_free_ctxt(rcd); } } diff --git a/drivers/infiniband/hw/hfi1/tid_rdma.c b/drivers/infiniband/hw/hfi1/tid_rdma.c index 70671212808f..1d02b12590f6 100644 --- a/drivers/infiniband/hw/hfi1/tid_rdma.c +++ b/drivers/infiniband/hw/hfi1/tid_rdma.c @@ -8,6 +8,7 @@ #include "qp.h" #include "verbs.h" #include "tid_rdma.h" +#include "exp_rcv.h" #include "trace.h" #define RCV_TID_FLOW_TABLE_CTRL_FLOW_VALID_SMASK BIT_ULL(32) @@ -35,8 +36,14 @@ static u32 mask_generation(u32 a) #define HFI1_KERNEL_MIN_JKEY HFI1_ADMIN_JKEY_RANGE #define HFI1_KERNEL_MAX_JKEY (2 * HFI1_ADMIN_JKEY_RANGE - 1) +/* Maximum number of segments in flight per QP request. */ #define TID_RDMA_MAX_READ_SEGS_PER_REQ 6 #define TID_RDMA_MAX_WRITE_SEGS_PER_REQ 4 +#define MAX_REQ max_t(u16, TID_RDMA_MAX_READ_SEGS_PER_REQ, \ + TID_RDMA_MAX_WRITE_SEGS_PER_REQ) +#define MAX_FLOWS roundup_pow_of_two(MAX_REQ + 1) + +#define MAX_EXPECTED_PAGES (MAX_EXPECTED_BUFFER / PAGE_SIZE) #define TID_OPFN_QP_CTXT_MASK 0xff #define TID_OPFN_QP_CTXT_SHIFT 56 @@ -79,6 +86,11 @@ static u32 mask_generation(u32 a) */ static void tid_rdma_trigger_resume(struct work_struct *work); +static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req); +static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req, + gfp_t gfp); +static void hfi1_init_trdma_req(struct rvt_qp *qp, + struct tid_rdma_request *req); static u64 tid_rdma_opfn_encode(struct tid_rdma_params *p) { @@ -230,7 +242,7 @@ int hfi1_kern_exp_rcv_init(struct hfi1_ctxtdata *rcd, int reinit) BUILD_BUG_ON(TID_RDMA_JKEY > HFI1_KERNEL_MAX_JKEY); rcd->jkey = TID_RDMA_JKEY; hfi1_set_ctxt_jkey(rcd->dd, rcd, rcd->jkey); - return 0; + return hfi1_alloc_ctxt_rcv_groups(rcd); } /** @@ -266,6 +278,7 @@ int hfi1_qp_priv_init(struct rvt_dev_info *rdi, struct rvt_qp *qp, struct ib_qp_init_attr *init_attr) { struct hfi1_qp_priv *qpriv = qp->priv; + int i, ret; qpriv->rcd = qp_to_rcd(rdi, qp); @@ -278,15 +291,75 @@ int hfi1_qp_priv_init(struct rvt_dev_info *rdi, struct rvt_qp *qp, qpriv->flow_state.generation = KERN_GENERATION_RESERVED; INIT_LIST_HEAD(&qpriv->tid_wait); + if (init_attr->qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) { + struct hfi1_devdata *dd = qpriv->rcd->dd; + + qpriv->pages = kzalloc_node(TID_RDMA_MAX_PAGES * + sizeof(*qpriv->pages), + GFP_KERNEL, dd->node); + if (!qpriv->pages) + return -ENOMEM; + for (i = 0; i < qp->s_size; i++) { + struct hfi1_swqe_priv *priv; + struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, i); + + priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, + dd->node); + if (!priv) + return -ENOMEM; + + hfi1_init_trdma_req(qp, &priv->tid_req); + priv->tid_req.e.swqe = wqe; + wqe->priv = priv; + } + for (i = 0; i < rvt_max_atomic(rdi); i++) { + struct hfi1_ack_priv *priv; + + priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, + dd->node); + if (!priv) + return -ENOMEM; + + hfi1_init_trdma_req(qp, &priv->tid_req); + priv->tid_req.e.ack = &qp->s_ack_queue[i]; + + ret = hfi1_kern_exp_rcv_alloc_flows(&priv->tid_req, + GFP_KERNEL); + if (ret) { + kfree(priv); + return ret; + } + qp->s_ack_queue[i].priv = priv; + } + } + return 0; } void hfi1_qp_priv_tid_free(struct rvt_dev_info *rdi, struct rvt_qp *qp) { - struct hfi1_qp_priv *priv = qp->priv; - - if (qp->ibqp.qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) - cancel_work_sync(&priv->opfn.opfn_work); + struct hfi1_qp_priv *qpriv = qp->priv; + struct rvt_swqe *wqe; + u32 i; + + if (qp->ibqp.qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) { + for (i = 0; i < qp->s_size; i++) { + wqe = rvt_get_swqe_ptr(qp, i); + kfree(wqe->priv); + wqe->priv = NULL; + } + for (i = 0; i < rvt_max_atomic(rdi); i++) { + struct hfi1_ack_priv *priv = qp->s_ack_queue[i].priv; + + if (priv) + hfi1_kern_exp_rcv_free_flows(&priv->tid_req); + kfree(priv); + qp->s_ack_queue[i].priv = NULL; + } + cancel_work_sync(&qpriv->opfn.opfn_work); + kfree(qpriv->pages); + qpriv->pages = NULL; + } } /* Flow and tid waiter functions */ @@ -540,6 +613,7 @@ void hfi1_tid_rdma_flush_wait(struct rvt_qp *qp) struct hfi1_qp_priv *priv = qp->priv; _tid_rdma_flush_wait(qp, &priv->rcd->flow_queue); + _tid_rdma_flush_wait(qp, &priv->rcd->rarr_queue); } /* Flow functions */ @@ -702,3 +776,796 @@ void hfi1_kern_init_ctxt_generations(struct hfi1_ctxtdata *rcd) kern_set_hw_flow(rcd, KERN_GENERATION_RESERVED, i); } } + +/* TID allocation functions */ +static u8 trdma_pset_order(struct tid_rdma_pageset *s) +{ + u8 count = s->count; + + return ilog2(count) + 1; +} + +/** + * tid_rdma_find_phys_blocks_4k - get groups base on mr info + * @npages - number of pages + * @pages - pointer to an array of page structs + * @list - page set array to return + * + * This routine returns the number of groups associated with + * the current sge information. This implementation is based + * on the expected receive find_phys_blocks() adjusted to + * use the MR information vs. the pfn. + * + * Return: + * the number of RcvArray entries + */ +static u32 tid_rdma_find_phys_blocks_4k(struct tid_rdma_flow *flow, + struct page **pages, + u32 npages, + struct tid_rdma_pageset *list) +{ + u32 pagecount, pageidx, setcount = 0, i; + void *vaddr, *this_vaddr; + + if (!npages) + return 0; + + /* + * Look for sets of physically contiguous pages in the user buffer. + * This will allow us to optimize Expected RcvArray entry usage by + * using the bigger supported sizes. + */ + vaddr = page_address(pages[0]); + for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) { + this_vaddr = i < npages ? page_address(pages[i]) : NULL; + /* + * If the vaddr's are not sequential, pages are not physically + * contiguous. + */ + if (this_vaddr != (vaddr + PAGE_SIZE)) { + /* + * At this point we have to loop over the set of + * physically contiguous pages and break them down it + * sizes supported by the HW. + * There are two main constraints: + * 1. The max buffer size is MAX_EXPECTED_BUFFER. + * If the total set size is bigger than that + * program only a MAX_EXPECTED_BUFFER chunk. + * 2. The buffer size has to be a power of two. If + * it is not, round down to the closes power of + * 2 and program that size. + */ + while (pagecount) { + int maxpages = pagecount; + u32 bufsize = pagecount * PAGE_SIZE; + + if (bufsize > MAX_EXPECTED_BUFFER) + maxpages = + MAX_EXPECTED_BUFFER >> + PAGE_SHIFT; + else if (!is_power_of_2(bufsize)) + maxpages = + rounddown_pow_of_two(bufsize) >> + PAGE_SHIFT; + + list[setcount].idx = pageidx; + list[setcount].count = maxpages; + pagecount -= maxpages; + pageidx += maxpages; + setcount++; + } + pageidx = i; + pagecount = 1; + vaddr = this_vaddr; + } else { + vaddr += PAGE_SIZE; + pagecount++; + } + } + /* insure we always return an even number of sets */ + if (setcount & 1) + list[setcount++].count = 0; + return setcount; +} + +/** + * tid_flush_pages - dump out pages into pagesets + * @list - list of pagesets + * @idx - pointer to current page index + * @pages - number of pages to dump + * @sets - current number of pagesset + * + * This routine flushes out accumuated pages. + * + * To insure an even number of sets the + * code may add a filler. + * + * This can happen with when pages is not + * a power of 2 or pages is a power of 2 + * less than the maximum pages. + * + * Return: + * The new number of sets + */ + +static u32 tid_flush_pages(struct tid_rdma_pageset *list, + u32 *idx, u32 pages, u32 sets) +{ + while (pages) { + u32 maxpages = pages; + + if (maxpages > MAX_EXPECTED_PAGES) + maxpages = MAX_EXPECTED_PAGES; + else if (!is_power_of_2(maxpages)) + maxpages = rounddown_pow_of_two(maxpages); + list[sets].idx = *idx; + list[sets++].count = maxpages; + *idx += maxpages; + pages -= maxpages; + } + /* might need a filler */ + if (sets & 1) + list[sets++].count = 0; + return sets; +} + +/** + * tid_rdma_find_phys_blocks_8k - get groups base on mr info + * @pages - pointer to an array of page structs + * @npages - number of pages + * @list - page set array to return + * + * This routine parses an array of pages to compute pagesets + * in an 8k compatible way. + * + * pages are tested two at a time, i, i + 1 for contiguous + * pages and i - 1 and i contiguous pages. + * + * If any condition is false, any accumlated pages are flushed and + * v0,v1 are emitted as separate PAGE_SIZE pagesets + * + * Otherwise, the current 8k is totaled for a future flush. + * + * Return: + * The number of pagesets + * list set with the returned number of pagesets + * + */ +static u32 tid_rdma_find_phys_blocks_8k(struct tid_rdma_flow *flow, + struct page **pages, + u32 npages, + struct tid_rdma_pageset *list) +{ + u32 idx, sets = 0, i; + u32 pagecnt = 0; + void *v0, *v1, *vm1; + + if (!npages) + return 0; + for (idx = 0, i = 0, vm1 = NULL; i < npages; i += 2) { + /* get a new v0 */ + v0 = page_address(pages[i]); + v1 = i + 1 < npages ? + page_address(pages[i + 1]) : NULL; + /* compare i, i + 1 vaddr */ + if (v1 != (v0 + PAGE_SIZE)) { + /* flush out pages */ + sets = tid_flush_pages(list, &idx, pagecnt, sets); + /* output v0,v1 as two pagesets */ + list[sets].idx = idx++; + list[sets++].count = 1; + if (v1) { + list[sets].count = 1; + list[sets++].idx = idx++; + } else { + list[sets++].count = 0; + } + vm1 = NULL; + pagecnt = 0; + continue; + } + /* i,i+1 consecutive, look at i-1,i */ + if (vm1 && v0 != (vm1 + PAGE_SIZE)) { + /* flush out pages */ + sets = tid_flush_pages(list, &idx, pagecnt, sets); + pagecnt = 0; + } + /* pages will always be a multiple of 8k */ + pagecnt += 2; + /* save i-1 */ + vm1 = v1; + /* move to next pair */ + } + /* dump residual pages at end */ + sets = tid_flush_pages(list, &idx, npages - idx, sets); + /* by design cannot be odd sets */ + WARN_ON(sets & 1); + return sets; +} + +/** + * Find pages for one segment of a sge array represented by @ss. The function + * does not check the sge, the sge must have been checked for alignment with a + * prior call to hfi1_kern_trdma_ok. Other sge checking is done as part of + * rvt_lkey_ok and rvt_rkey_ok. Also, the function only modifies the local sge + * copy maintained in @ss->sge, the original sge is not modified. + * + * Unlike IB RDMA WRITE, we can't decrement ss->num_sge here because we are not + * releasing the MR reference count at the same time. Otherwise, we'll "leak" + * references to the MR. This difference requires that we keep track of progress + * into the sg_list. This is done by the cur_seg cursor in the tid_rdma_request + * structure. + */ +static u32 kern_find_pages(struct tid_rdma_flow *flow, + struct page **pages, + struct rvt_sge_state *ss, bool *last) +{ + struct tid_rdma_request *req = flow->req; + struct rvt_sge *sge = &ss->sge; + u32 length = flow->req->seg_len; + u32 len = PAGE_SIZE; + u32 i = 0; + + while (length && req->isge < ss->num_sge) { + pages[i++] = virt_to_page(sge->vaddr); + + sge->vaddr += len; + sge->length -= len; + sge->sge_length -= len; + if (!sge->sge_length) { + if (++req->isge < ss->num_sge) + *sge = ss->sg_list[req->isge - 1]; + } else if (sge->length == 0 && sge->mr->lkey) { + if (++sge->n >= RVT_SEGSZ) { + ++sge->m; + sge->n = 0; + } + sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr; + sge->length = sge->mr->map[sge->m]->segs[sge->n].length; + } + length -= len; + } + + flow->length = flow->req->seg_len - length; + *last = req->isge == ss->num_sge ? false : true; + return i; +} + +static void dma_unmap_flow(struct tid_rdma_flow *flow) +{ + struct hfi1_devdata *dd; + int i; + struct tid_rdma_pageset *pset; + + dd = flow->req->rcd->dd; + for (i = 0, pset = &flow->pagesets[0]; i < flow->npagesets; + i++, pset++) { + if (pset->count && pset->addr) { + dma_unmap_page(&dd->pcidev->dev, + pset->addr, + PAGE_SIZE * pset->count, + DMA_FROM_DEVICE); + pset->mapped = 0; + } + } +} + +static int dma_map_flow(struct tid_rdma_flow *flow, struct page **pages) +{ + int i; + struct hfi1_devdata *dd = flow->req->rcd->dd; + struct tid_rdma_pageset *pset; + + for (i = 0, pset = &flow->pagesets[0]; i < flow->npagesets; + i++, pset++) { + if (pset->count) { + pset->addr = dma_map_page(&dd->pcidev->dev, + pages[pset->idx], + 0, + PAGE_SIZE * pset->count, + DMA_FROM_DEVICE); + + if (dma_mapping_error(&dd->pcidev->dev, pset->addr)) { + dma_unmap_flow(flow); + return -ENOMEM; + } + pset->mapped = 1; + } + } + return 0; +} + +static inline bool dma_mapped(struct tid_rdma_flow *flow) +{ + return !!flow->pagesets[0].mapped; +} + +/* + * Get pages pointers and identify contiguous physical memory chunks for a + * segment. All segments are of length flow->req->seg_len. + */ +static int kern_get_phys_blocks(struct tid_rdma_flow *flow, + struct page **pages, + struct rvt_sge_state *ss, bool *last) +{ + u8 npages; + + /* Reuse previously computed pagesets, if any */ + if (flow->npagesets) { + if (!dma_mapped(flow)) + return dma_map_flow(flow, pages); + return 0; + } + + npages = kern_find_pages(flow, pages, ss, last); + + if (flow->req->qp->pmtu == enum_to_mtu(OPA_MTU_4096)) + flow->npagesets = + tid_rdma_find_phys_blocks_4k(flow, pages, npages, + flow->pagesets); + else + flow->npagesets = + tid_rdma_find_phys_blocks_8k(flow, pages, npages, + flow->pagesets); + + return dma_map_flow(flow, pages); +} + +static inline void kern_add_tid_node(struct tid_rdma_flow *flow, + struct hfi1_ctxtdata *rcd, char *s, + struct tid_group *grp, u8 cnt) +{ + struct kern_tid_node *node = &flow->tnode[flow->tnode_cnt++]; + + WARN_ON_ONCE(flow->tnode_cnt >= + (TID_RDMA_MAX_SEGMENT_SIZE >> PAGE_SHIFT)); + if (WARN_ON_ONCE(cnt & 1)) + dd_dev_err(rcd->dd, + "unexpected odd allocation cnt %u map 0x%x used %u", + cnt, grp->map, grp->used); + + node->grp = grp; + node->map = grp->map; + node->cnt = cnt; +} + +/* + * Try to allocate pageset_count TID's from TID groups for a context + * + * This function allocates TID's without moving groups between lists or + * modifying grp->map. This is done as follows, being cogizant of the lists + * between which the TID groups will move: + * 1. First allocate complete groups of 8 TID's since this is more efficient, + * these groups will move from group->full without affecting used + * 2. If more TID's are needed allocate from used (will move from used->full or + * stay in used) + * 3. If we still don't have the required number of TID's go back and look again + * at a complete group (will move from group->used) + */ +static int kern_alloc_tids(struct tid_rdma_flow *flow) +{ + struct hfi1_ctxtdata *rcd = flow->req->rcd; + struct hfi1_devdata *dd = rcd->dd; + u32 ngroups, pageidx = 0; + struct tid_group *group = NULL, *used; + u8 use; + + flow->tnode_cnt = 0; + ngroups = flow->npagesets / dd->rcv_entries.group_size; + if (!ngroups) + goto used_list; + + /* First look at complete groups */ + list_for_each_entry(group, &rcd->tid_group_list.list, list) { + kern_add_tid_node(flow, rcd, "complete groups", group, + group->size); + + pageidx += group->size; + if (!--ngroups) + break; + } + + if (pageidx >= flow->npagesets) + goto ok; + +used_list: + /* Now look at partially used groups */ + list_for_each_entry(used, &rcd->tid_used_list.list, list) { + use = min_t(u32, flow->npagesets - pageidx, + used->size - used->used); + kern_add_tid_node(flow, rcd, "used groups", used, use); + + pageidx += use; + if (pageidx >= flow->npagesets) + goto ok; + } + + /* + * Look again at a complete group, continuing from where we left. + * However, if we are at the head, we have reached the end of the + * complete groups list from the first loop above + */ + if (group && &group->list == &rcd->tid_group_list.list) + goto bail_eagain; + group = list_prepare_entry(group, &rcd->tid_group_list.list, + list); + if (list_is_last(&group->list, &rcd->tid_group_list.list)) + goto bail_eagain; + group = list_next_entry(group, list); + use = min_t(u32, flow->npagesets - pageidx, group->size); + kern_add_tid_node(flow, rcd, "complete continue", group, use); + pageidx += use; + if (pageidx >= flow->npagesets) + goto ok; +bail_eagain: + return -EAGAIN; +ok: + return 0; +} + +static void kern_program_rcv_group(struct tid_rdma_flow *flow, int grp_num, + u32 *pset_idx) +{ + struct hfi1_ctxtdata *rcd = flow->req->rcd; + struct hfi1_devdata *dd = rcd->dd; + struct kern_tid_node *node = &flow->tnode[grp_num]; + struct tid_group *grp = node->grp; + struct tid_rdma_pageset *pset; + u32 pmtu_pg = flow->req->qp->pmtu >> PAGE_SHIFT; + u32 rcventry, npages = 0, pair = 0, tidctrl; + u8 i, cnt = 0; + + for (i = 0; i < grp->size; i++) { + rcventry = grp->base + i; + + if (node->map & BIT(i) || cnt >= node->cnt) { + rcv_array_wc_fill(dd, rcventry); + continue; + } + pset = &flow->pagesets[(*pset_idx)++]; + if (pset->count) { + hfi1_put_tid(dd, rcventry, PT_EXPECTED, + pset->addr, trdma_pset_order(pset)); + } else { + hfi1_put_tid(dd, rcventry, PT_INVALID, 0, 0); + } + npages += pset->count; + + rcventry -= rcd->expected_base; + tidctrl = pair ? 0x3 : rcventry & 0x1 ? 0x2 : 0x1; + /* + * A single TID entry will be used to use a rcvarr pair (with + * tidctrl 0x3), if ALL these are true (a) the bit pos is even + * (b) the group map shows current and the next bits as free + * indicating two consecutive rcvarry entries are available (c) + * we actually need 2 more entries + */ + pair = !(i & 0x1) && !((node->map >> i) & 0x3) && + node->cnt >= cnt + 2; + if (!pair) { + if (!pset->count) + tidctrl = 0x1; + flow->tid_entry[flow->tidcnt++] = + EXP_TID_SET(IDX, rcventry >> 1) | + EXP_TID_SET(CTRL, tidctrl) | + EXP_TID_SET(LEN, npages); + /* Efficient DIV_ROUND_UP(npages, pmtu_pg) */ + flow->npkts += (npages + pmtu_pg - 1) >> ilog2(pmtu_pg); + npages = 0; + } + + if (grp->used == grp->size - 1) + tid_group_move(grp, &rcd->tid_used_list, + &rcd->tid_full_list); + else if (!grp->used) + tid_group_move(grp, &rcd->tid_group_list, + &rcd->tid_used_list); + + grp->used++; + grp->map |= BIT(i); + cnt++; + } +} + +static void kern_unprogram_rcv_group(struct tid_rdma_flow *flow, int grp_num) +{ + struct hfi1_ctxtdata *rcd = flow->req->rcd; + struct hfi1_devdata *dd = rcd->dd; + struct kern_tid_node *node = &flow->tnode[grp_num]; + struct tid_group *grp = node->grp; + u32 rcventry; + u8 i, cnt = 0; + + for (i = 0; i < grp->size; i++) { + rcventry = grp->base + i; + + if (node->map & BIT(i) || cnt >= node->cnt) { + rcv_array_wc_fill(dd, rcventry); + continue; + } + + hfi1_put_tid(dd, rcventry, PT_INVALID, 0, 0); + + grp->used--; + grp->map &= ~BIT(i); + cnt++; + + if (grp->used == grp->size - 1) + tid_group_move(grp, &rcd->tid_full_list, + &rcd->tid_used_list); + else if (!grp->used) + tid_group_move(grp, &rcd->tid_used_list, + &rcd->tid_group_list); + } + if (WARN_ON_ONCE(cnt & 1)) { + struct hfi1_ctxtdata *rcd = flow->req->rcd; + struct hfi1_devdata *dd = rcd->dd; + + dd_dev_err(dd, "unexpected odd free cnt %u map 0x%x used %u", + cnt, grp->map, grp->used); + } +} + +static void kern_program_rcvarray(struct tid_rdma_flow *flow) +{ + u32 pset_idx = 0; + int i; + + flow->npkts = 0; + flow->tidcnt = 0; + for (i = 0; i < flow->tnode_cnt; i++) + kern_program_rcv_group(flow, i, &pset_idx); +} + +/** + * hfi1_kern_exp_rcv_setup() - setup TID's and flow for one segment of a + * TID RDMA request + * + * @req: TID RDMA request for which the segment/flow is being set up + * @ss: sge state, maintains state across successive segments of a sge + * @last: set to true after the last sge segment has been processed + * + * This function + * (1) finds a free flow entry in the flow circular buffer + * (2) finds pages and continuous physical chunks constituing one segment + * of an sge + * (3) allocates TID group entries for those chunks + * (4) programs rcvarray entries in the hardware corresponding to those + * TID's + * (5) computes a tidarray with formatted TID entries which can be sent + * to the sender + * (6) Reserves and programs HW flows. + * (7) It also manages queing the QP when TID/flow resources are not + * available. + * + * @req points to struct tid_rdma_request of which the segments are a part. The + * function uses qp, rcd and seg_len members of @req. In the absence of errors, + * req->flow_idx is the index of the flow which has been prepared in this + * invocation of function call. With flow = &req->flows[req->flow_idx], + * flow->tid_entry contains the TID array which the sender can use for TID RDMA + * sends and flow->npkts contains number of packets required to send the + * segment. + * + * hfi1_check_sge_align should be called prior to calling this function and if + * it signals error TID RDMA cannot be used for this sge and this function + * should not be called. + * + * For the queuing, caller must hold the flow->req->qp s_lock from the send + * engine and the function will procure the exp_lock. + * + * Return: + * The function returns -EAGAIN if sufficient number of TID/flow resources to + * map the segment could not be allocated. In this case the function should be + * called again with previous arguments to retry the TID allocation. There are + * no other error returns. The function returns 0 on success. + */ +int hfi1_kern_exp_rcv_setup(struct tid_rdma_request *req, + struct rvt_sge_state *ss, bool *last) + __must_hold(&req->qp->s_lock) +{ + struct tid_rdma_flow *flow = &req->flows[req->setup_head]; + struct hfi1_ctxtdata *rcd = req->rcd; + struct hfi1_qp_priv *qpriv = req->qp->priv; + unsigned long flags; + struct rvt_qp *fqp; + u16 clear_tail = req->clear_tail; + + lockdep_assert_held(&req->qp->s_lock); + /* + * We return error if either (a) we don't have space in the flow + * circular buffer, or (b) we already have max entries in the buffer. + * Max entries depend on the type of request we are processing and the + * negotiated TID RDMA parameters. + */ + if (!CIRC_SPACE(req->setup_head, clear_tail, MAX_FLOWS) || + CIRC_CNT(req->setup_head, clear_tail, MAX_FLOWS) >= + req->n_flows) + return -EINVAL; + + /* + * Get pages, identify contiguous physical memory chunks for the segment + * If we can not determine a DMA address mapping we will treat it just + * like if we ran out of space above. + */ + if (kern_get_phys_blocks(flow, qpriv->pages, ss, last)) { + hfi1_wait_kmem(flow->req->qp); + return -ENOMEM; + } + + spin_lock_irqsave(&rcd->exp_lock, flags); + if (kernel_tid_waiters(rcd, &rcd->rarr_queue, flow->req->qp)) + goto queue; + + /* + * At this point we know the number of pagesets and hence the number of + * TID's to map the segment. Allocate the TID's from the TID groups. If + * we cannot allocate the required number we exit and try again later + */ + if (kern_alloc_tids(flow)) + goto queue; + /* + * Finally program the TID entries with the pagesets, compute the + * tidarray and enable the HW flow + */ + kern_program_rcvarray(flow); + + /* + * Setup the flow state with relevant information. + * This information is used for tracking the sequence of data packets + * for the segment. + * The flow is setup here as this is the most accurate time and place + * to do so. Doing at a later time runs the risk of the flow data in + * qpriv getting out of sync. + */ + memset(&flow->flow_state, 0x0, sizeof(flow->flow_state)); + flow->idx = qpriv->flow_state.index; + flow->flow_state.generation = qpriv->flow_state.generation; + flow->flow_state.spsn = qpriv->flow_state.psn; + flow->flow_state.lpsn = flow->flow_state.spsn + flow->npkts - 1; + flow->flow_state.r_next_psn = + full_flow_psn(flow, flow->flow_state.spsn); + qpriv->flow_state.psn += flow->npkts; + + dequeue_tid_waiter(rcd, &rcd->rarr_queue, flow->req->qp); + /* get head before dropping lock */ + fqp = first_qp(rcd, &rcd->rarr_queue); + spin_unlock_irqrestore(&rcd->exp_lock, flags); + tid_rdma_schedule_tid_wakeup(fqp); + + req->setup_head = (req->setup_head + 1) & (MAX_FLOWS - 1); + return 0; +queue: + queue_qp_for_tid_wait(rcd, &rcd->rarr_queue, flow->req->qp); + spin_unlock_irqrestore(&rcd->exp_lock, flags); + return -EAGAIN; +} + +static void hfi1_tid_rdma_reset_flow(struct tid_rdma_flow *flow) +{ + flow->npagesets = 0; +} + +/* + * This function is called after one segment has been successfully sent to + * release the flow and TID HW/SW resources for that segment. The segments for a + * TID RDMA request are setup and cleared in FIFO order which is managed using a + * circular buffer. + */ +int hfi1_kern_exp_rcv_clear(struct tid_rdma_request *req) + __must_hold(&req->qp->s_lock) +{ + struct tid_rdma_flow *flow = &req->flows[req->clear_tail]; + struct hfi1_ctxtdata *rcd = req->rcd; + unsigned long flags; + int i; + struct rvt_qp *fqp; + + lockdep_assert_held(&req->qp->s_lock); + /* Exit if we have nothing in the flow circular buffer */ + if (!CIRC_CNT(req->setup_head, req->clear_tail, MAX_FLOWS)) + return -EINVAL; + + spin_lock_irqsave(&rcd->exp_lock, flags); + + for (i = 0; i < flow->tnode_cnt; i++) + kern_unprogram_rcv_group(flow, i); + /* To prevent double unprogramming */ + flow->tnode_cnt = 0; + /* get head before dropping lock */ + fqp = first_qp(rcd, &rcd->rarr_queue); + spin_unlock_irqrestore(&rcd->exp_lock, flags); + + dma_unmap_flow(flow); + + hfi1_tid_rdma_reset_flow(flow); + req->clear_tail = (req->clear_tail + 1) & (MAX_FLOWS - 1); + + if (fqp == req->qp) { + __trigger_tid_waiter(fqp); + rvt_put_qp(fqp); + } else { + tid_rdma_schedule_tid_wakeup(fqp); + } + + return 0; +} + +/* + * This function is called to release all the tid entries for + * a request. + */ +void hfi1_kern_exp_rcv_clear_all(struct tid_rdma_request *req) + __must_hold(&req->qp->s_lock) +{ + /* Use memory barrier for proper ordering */ + while (CIRC_CNT(req->setup_head, req->clear_tail, MAX_FLOWS)) { + if (hfi1_kern_exp_rcv_clear(req)) + break; + } +} + +/** + * hfi1_kern_exp_rcv_free_flows - free priviously allocated flow information + * @req - the tid rdma request to be cleaned + */ +static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req) +{ + kfree(req->flows); + req->flows = NULL; +} + +/** + * __trdma_clean_swqe - clean up for large sized QPs + * @qp: the queue patch + * @wqe: the send wqe + */ +void __trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe) +{ + struct hfi1_swqe_priv *p = wqe->priv; + + hfi1_kern_exp_rcv_free_flows(&p->tid_req); +} + +/* + * This can be called at QP create time or in the data path. + */ +static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req, + gfp_t gfp) +{ + struct tid_rdma_flow *flows; + int i; + + if (likely(req->flows)) + return 0; + flows = kmalloc_node(MAX_FLOWS * sizeof(*flows), gfp, + req->rcd->numa_id); + if (!flows) + return -ENOMEM; + /* mini init */ + for (i = 0; i < MAX_FLOWS; i++) { + flows[i].req = req; + flows[i].npagesets = 0; + flows[i].pagesets[0].mapped = 0; + } + req->flows = flows; + return 0; +} + +static void hfi1_init_trdma_req(struct rvt_qp *qp, + struct tid_rdma_request *req) +{ + struct hfi1_qp_priv *qpriv = qp->priv; + + /* + * Initialize various TID RDMA request variables. + * These variables are "static", which is why they + * can be pre-initialized here before the WRs has + * even been submitted. + * However, non-NULL values for these variables do not + * imply that this WQE has been enabled for TID RDMA. + * Drivers should check the WQE's opcode to determine + * if a request is a TID RDMA one or not. + */ + req->qp = qp; + req->rcd = qpriv->rcd; +} diff --git a/drivers/infiniband/hw/hfi1/tid_rdma.h b/drivers/infiniband/hw/hfi1/tid_rdma.h index 3bc0aaf9568f..524baf8c8fac 100644 --- a/drivers/infiniband/hw/hfi1/tid_rdma.h +++ b/drivers/infiniband/hw/hfi1/tid_rdma.h @@ -6,7 +6,16 @@ #ifndef HFI1_TID_RDMA_H #define HFI1_TID_RDMA_H +#include <linux/circ_buf.h> +#include "common.h" + +/* Add a convenience helper */ +#define CIRC_ADD(val, add, size) (((val) + (add)) & ((size) - 1)) +#define CIRC_NEXT(val, size) CIRC_ADD(val, 1, size) +#define CIRC_PREV(val, size) CIRC_ADD(val, -1, size) + #define TID_RDMA_MAX_SEGMENT_SIZE BIT(18) /* 256 KiB (for now) */ +#define TID_RDMA_MAX_PAGES (BIT(18) >> PAGE_SHIFT) struct tid_rdma_params { struct rcu_head rcu_head; @@ -36,6 +45,81 @@ struct tid_flow_state { u8 flags; }; +struct tid_rdma_request { + struct rvt_qp *qp; + struct hfi1_ctxtdata *rcd; + union { + struct rvt_swqe *swqe; + struct rvt_ack_entry *ack; + } e; + + struct tid_rdma_flow *flows; /* array of tid flows */ + u16 n_flows; /* size of the flow buffer window */ + u16 setup_head; /* flow index we are setting up */ + u16 clear_tail; /* flow index we are clearing */ + u16 flow_idx; /* flow index most recently set up */ + + u32 seg_len; + + u32 isge; /* index of "current" sge */ +}; + +/* + * When header suppression is used, PSNs associated with a "flow" are + * relevant (and not the PSNs maintained by verbs). Track per-flow + * PSNs here for a TID RDMA segment. + * + */ +struct flow_state { + u32 flags; + u32 resp_ib_psn; /* The IB PSN of the response for this flow */ + u32 generation; /* generation of flow */ + u32 spsn; /* starting PSN in TID space */ + u32 lpsn; /* last PSN in TID space */ + u32 r_next_psn; /* next PSN to be received (in TID space) */ +}; + +struct tid_rdma_pageset { + dma_addr_t addr : 48; /* Only needed for the first page */ + u8 idx: 8; + u8 count : 7; + u8 mapped: 1; +}; + +/** + * kern_tid_node - used for managing TID's in TID groups + * + * @grp_idx: rcd relative index to tid_group + * @map: grp->map captured prior to programming this TID group in HW + * @cnt: Only @cnt of available group entries are actually programmed + */ +struct kern_tid_node { + struct tid_group *grp; + u8 map; + u8 cnt; +}; + +/* Overall info for a TID RDMA segment */ +struct tid_rdma_flow { + /* + * While a TID RDMA segment is being transferred, it uses a QP number + * from the "KDETH section of QP numbers" (which is different from the + * QP number that originated the request). Bits 11-15 of these QP + * numbers identify the "TID flow" for the segment. + */ + struct flow_state flow_state; + struct tid_rdma_request *req; + u32 length; + u8 tnode_cnt; + u8 tidcnt; + u8 idx; + u8 npagesets; + u8 npkts; + struct kern_tid_node tnode[TID_RDMA_MAX_PAGES]; + struct tid_rdma_pageset pagesets[TID_RDMA_MAX_PAGES]; + u32 tid_entry[TID_RDMA_MAX_PAGES]; +}; + bool tid_rdma_conn_req(struct rvt_qp *qp, u64 *data); bool tid_rdma_conn_reply(struct rvt_qp *qp, u64 data); bool tid_rdma_conn_resp(struct rvt_qp *qp, u64 *data); @@ -43,6 +127,23 @@ void tid_rdma_conn_error(struct rvt_qp *qp); void tid_rdma_opfn_init(struct rvt_qp *qp, struct tid_rdma_params *p); int hfi1_kern_exp_rcv_init(struct hfi1_ctxtdata *rcd, int reinit); +int hfi1_kern_exp_rcv_setup(struct tid_rdma_request *req, + struct rvt_sge_state *ss, bool *last); +int hfi1_kern_exp_rcv_clear(struct tid_rdma_request *req); +void hfi1_kern_exp_rcv_clear_all(struct tid_rdma_request *req); +void __trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe); + +/** + * trdma_clean_swqe - clean flows for swqe if large send queue + * @qp: the qp + * @wqe: the send wqe + */ +static inline void trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe) +{ + if (!wqe->priv) + return; + __trdma_clean_swqe(qp, wqe); +} int hfi1_qp_priv_init(struct rvt_dev_info *rdi, struct rvt_qp *qp, struct ib_qp_init_attr *init_attr); diff --git a/drivers/infiniband/hw/hfi1/user_exp_rcv.h b/drivers/infiniband/hw/hfi1/user_exp_rcv.h index e383cc01a2bf..43b105de1d54 100644 --- a/drivers/infiniband/hw/hfi1/user_exp_rcv.h +++ b/drivers/infiniband/hw/hfi1/user_exp_rcv.h @@ -48,7 +48,6 @@ */ #include "hfi.h" - #include "exp_rcv.h" struct tid_pageset { diff --git a/drivers/infiniband/hw/hfi1/verbs.c b/drivers/infiniband/hw/hfi1/verbs.c index 571bfd549c2a..02c1873a976c 100644 --- a/drivers/infiniband/hw/hfi1/verbs.c +++ b/drivers/infiniband/hw/hfi1/verbs.c @@ -504,11 +504,28 @@ static void verbs_sdma_complete( hfi1_put_txreq(tx); } +void hfi1_wait_kmem(struct rvt_qp *qp) +{ + struct hfi1_qp_priv *priv = qp->priv; + struct ib_qp *ibqp = &qp->ibqp; + struct ib_device *ibdev = ibqp->device; + struct hfi1_ibdev *dev = to_idev(ibdev); + + if (list_empty(&priv->s_iowait.list)) { + if (list_empty(&dev->memwait)) + mod_timer(&dev->mem_timer, jiffies + 1); + qp->s_flags |= RVT_S_WAIT_KMEM; + list_add_tail(&priv->s_iowait.list, &dev->memwait); + priv->s_iowait.lock = &dev->iowait_lock; + trace_hfi1_qpsleep(qp, RVT_S_WAIT_KMEM); + rvt_get_qp(qp); + } +} + static int wait_kmem(struct hfi1_ibdev *dev, struct rvt_qp *qp, struct hfi1_pkt_state *ps) { - struct hfi1_qp_priv *priv = qp->priv; unsigned long flags; int ret = 0; @@ -517,15 +534,7 @@ static int wait_kmem(struct hfi1_ibdev *dev, write_seqlock(&dev->iowait_lock); list_add_tail(&ps->s_txreq->txreq.list, &ps->wait->tx_head); - if (list_empty(&priv->s_iowait.list)) { - if (list_empty(&dev->memwait)) - mod_timer(&dev->mem_timer, jiffies + 1); - qp->s_flags |= RVT_S_WAIT_KMEM; - list_add_tail(&priv->s_iowait.list, &dev->memwait); - priv->s_iowait.lock = &dev->iowait_lock; - trace_hfi1_qpsleep(qp, RVT_S_WAIT_KMEM); - rvt_get_qp(qp); - } + hfi1_wait_kmem(qp); write_sequnlock(&dev->iowait_lock); hfi1_qp_unbusy(qp, ps->wait); ret = -EBUSY; diff --git a/drivers/infiniband/hw/hfi1/verbs.h b/drivers/infiniband/hw/hfi1/verbs.h index 9065e470bebb..94f198b47239 100644 --- a/drivers/infiniband/hw/hfi1/verbs.h +++ b/drivers/infiniband/hw/hfi1/verbs.h @@ -159,6 +159,7 @@ struct hfi1_qp_priv { struct sdma_engine *s_sde; /* current sde */ struct send_context *s_sendcontext; /* current sendcontext */ struct hfi1_ctxtdata *rcd; /* QP's receive context */ + struct page **pages; /* for TID page scan */ u32 tid_enqueue; /* saved when tid waited */ u8 s_sc; /* SC[0..4] for next packet */ struct iowait s_iowait; @@ -173,6 +174,14 @@ struct hfi1_qp_priv { u8 timeout_shift; /* account for number of packets per segment */ }; +struct hfi1_swqe_priv { + struct tid_rdma_request tid_req; +}; + +struct hfi1_ack_priv { + struct tid_rdma_request tid_req; +}; + /* * This structure is used to hold commonly lookedup and computed values during * the send engine progress. @@ -321,6 +330,21 @@ static inline u32 delta_psn(u32 a, u32 b) return (((int)a - (int)b) << PSN_SHIFT) >> PSN_SHIFT; } +/* + * Look through all the active flows for a TID RDMA request and find + * the one (if it exists) that contains the specified PSN. + */ +static inline u32 __full_flow_psn(struct flow_state *state, u32 psn) +{ + return mask_psn((state->generation << HFI1_KDETH_BTH_SEQ_SHIFT) | + (psn & HFI1_KDETH_BTH_SEQ_MASK)); +} + +static inline u32 full_flow_psn(struct tid_rdma_flow *flow, u32 psn) +{ + return __full_flow_psn(&flow->flow_state, psn); +} + struct verbs_txreq; void hfi1_put_txreq(struct verbs_txreq *tx); @@ -403,6 +427,16 @@ static inline bool opa_bth_is_migration(struct ib_other_headers *ohdr) return ohdr->bth[1] & cpu_to_be32(OPA_BTH_MIG_REQ); } +void hfi1_wait_kmem(struct rvt_qp *qp); + +static inline void hfi1_trdma_send_complete(struct rvt_qp *qp, + struct rvt_swqe *wqe, + enum ib_wc_status status) +{ + trdma_clean_swqe(qp, wqe); + rvt_send_complete(qp, wqe, status); +} + extern const enum ib_wc_opcode ib_hfi1_wc_opcode[]; extern const u8 hdr_len_by_opcode[]; diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c index 16247d2a671d..c8e70cf69a8a 100644 --- a/drivers/infiniband/sw/rdmavt/qp.c +++ b/drivers/infiniband/sw/rdmavt/qp.c @@ -1642,11 +1642,11 @@ int rvt_destroy_qp(struct ib_qp *ibqp) kref_put(&qp->ip->ref, rvt_release_mmap_info); else vfree(qp->r_rq.wq); - vfree(qp->s_wq); rdi->driver_f.qp_priv_free(rdi, qp); kfree(qp->s_ack_queue); rdma_destroy_ah_attr(&qp->remote_ah_attr); rdma_destroy_ah_attr(&qp->alt_ah_attr); + vfree(qp->s_wq); kfree(qp); return 0; } diff --git a/include/rdma/rdmavt_qp.h b/include/rdma/rdmavt_qp.h index 56a9221378d9..9095a0b71250 100644 --- a/include/rdma/rdmavt_qp.h +++ b/include/rdma/rdmavt_qp.h @@ -174,6 +174,7 @@ struct rvt_swqe { u32 lpsn; /* last packet sequence number */ u32 ssn; /* send sequence number */ u32 length; /* total length of data in sg_list */ + void *priv; /* driver dependent field */ struct rvt_sge sg_list[0]; }; @@ -235,6 +236,7 @@ struct rvt_ack_entry { u32 lpsn; u8 opcode; u8 sent; + void *priv; }; #define RC_QP_SCALING_INTERVAL 5 |