1 /* 2 * Copyright (c) 2015 HGST, a Western Digital Company. 3 * 4 * This program is free software; you can redistribute it and/or modify it 5 * under the terms and conditions of the GNU General Public License, 6 * version 2, as published by the Free Software Foundation. 7 * 8 * This program is distributed in the hope it will be useful, but WITHOUT 9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 11 * more details. 12 */ 13 #include <linux/module.h> 14 #include <linux/err.h> 15 #include <linux/slab.h> 16 #include <rdma/ib_verbs.h> 17 18 /* # of WCs to poll for with a single call to ib_poll_cq */ 19 #define IB_POLL_BATCH 16 20 21 /* # of WCs to iterate over before yielding */ 22 #define IB_POLL_BUDGET_IRQ 256 23 #define IB_POLL_BUDGET_WORKQUEUE 65536 24 25 #define IB_POLL_FLAGS \ 26 (IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS) 27 28 static int __ib_process_cq(struct ib_cq *cq, int budget, struct ib_wc *poll_wc) 29 { 30 int i, n, completed = 0; 31 struct ib_wc *wcs = poll_wc ? : cq->wc; 32 33 /* 34 * budget might be (-1) if the caller does not 35 * want to bound this call, thus we need unsigned 36 * minimum here. 37 */ 38 while ((n = ib_poll_cq(cq, min_t(u32, IB_POLL_BATCH, 39 budget - completed), wcs)) > 0) { 40 for (i = 0; i < n; i++) { 41 struct ib_wc *wc = &wcs[i]; 42 43 if (wc->wr_cqe) 44 wc->wr_cqe->done(cq, wc); 45 else 46 WARN_ON_ONCE(wc->status == IB_WC_SUCCESS); 47 } 48 49 completed += n; 50 51 if (n != IB_POLL_BATCH || 52 (budget != -1 && completed >= budget)) 53 break; 54 } 55 56 return completed; 57 } 58 59 /** 60 * ib_process_direct_cq - process a CQ in caller context 61 * @cq: CQ to process 62 * @budget: number of CQEs to poll for 63 * 64 * This function is used to process all outstanding CQ entries. 65 * It does not offload CQ processing to a different context and does 66 * not ask for completion interrupts from the HCA. 67 * Using direct processing on CQ with non IB_POLL_DIRECT type may trigger 68 * concurrent processing. 69 * 70 * Note: do not pass -1 as %budget unless it is guaranteed that the number 71 * of completions that will be processed is small. 72 */ 73 int ib_process_cq_direct(struct ib_cq *cq, int budget) 74 { 75 struct ib_wc wcs[IB_POLL_BATCH]; 76 77 return __ib_process_cq(cq, budget, wcs); 78 } 79 EXPORT_SYMBOL(ib_process_cq_direct); 80 81 static void ib_cq_completion_direct(struct ib_cq *cq, void *private) 82 { 83 WARN_ONCE(1, "got unsolicited completion for CQ 0x%p\n", cq); 84 } 85 86 static int ib_poll_handler(struct irq_poll *iop, int budget) 87 { 88 struct ib_cq *cq = container_of(iop, struct ib_cq, iop); 89 int completed; 90 91 completed = __ib_process_cq(cq, budget, NULL); 92 if (completed < budget) { 93 irq_poll_complete(&cq->iop); 94 if (ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0) 95 irq_poll_sched(&cq->iop); 96 } 97 98 return completed; 99 } 100 101 static void ib_cq_completion_softirq(struct ib_cq *cq, void *private) 102 { 103 irq_poll_sched(&cq->iop); 104 } 105 106 static void ib_cq_poll_work(struct work_struct *work) 107 { 108 struct ib_cq *cq = container_of(work, struct ib_cq, work); 109 int completed; 110 111 completed = __ib_process_cq(cq, IB_POLL_BUDGET_WORKQUEUE, NULL); 112 if (completed >= IB_POLL_BUDGET_WORKQUEUE || 113 ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0) 114 queue_work(ib_comp_wq, &cq->work); 115 } 116 117 static void ib_cq_completion_workqueue(struct ib_cq *cq, void *private) 118 { 119 queue_work(ib_comp_wq, &cq->work); 120 } 121 122 /** 123 * __ib_alloc_cq - allocate a completion queue 124 * @dev: device to allocate the CQ for 125 * @private: driver private data, accessible from cq->cq_context 126 * @nr_cqe: number of CQEs to allocate 127 * @comp_vector: HCA completion vectors for this CQ 128 * @poll_ctx: context to poll the CQ from. 129 * @caller: module owner name. 130 * 131 * This is the proper interface to allocate a CQ for in-kernel users. A 132 * CQ allocated with this interface will automatically be polled from the 133 * specified context. The ULP must use wr->wr_cqe instead of wr->wr_id 134 * to use this CQ abstraction. 135 */ 136 struct ib_cq *__ib_alloc_cq(struct ib_device *dev, void *private, 137 int nr_cqe, int comp_vector, 138 enum ib_poll_context poll_ctx, const char *caller) 139 { 140 struct ib_cq_init_attr cq_attr = { 141 .cqe = nr_cqe, 142 .comp_vector = comp_vector, 143 }; 144 struct ib_cq *cq; 145 int ret = -ENOMEM; 146 147 cq = dev->create_cq(dev, &cq_attr, NULL, NULL); 148 if (IS_ERR(cq)) 149 return cq; 150 151 cq->device = dev; 152 cq->uobject = NULL; 153 cq->event_handler = NULL; 154 cq->cq_context = private; 155 cq->poll_ctx = poll_ctx; 156 atomic_set(&cq->usecnt, 0); 157 158 cq->wc = kmalloc_array(IB_POLL_BATCH, sizeof(*cq->wc), GFP_KERNEL); 159 if (!cq->wc) 160 goto out_destroy_cq; 161 162 cq->res.type = RDMA_RESTRACK_CQ; 163 cq->res.kern_name = caller; 164 rdma_restrack_add(&cq->res); 165 166 switch (cq->poll_ctx) { 167 case IB_POLL_DIRECT: 168 cq->comp_handler = ib_cq_completion_direct; 169 break; 170 case IB_POLL_SOFTIRQ: 171 cq->comp_handler = ib_cq_completion_softirq; 172 173 irq_poll_init(&cq->iop, IB_POLL_BUDGET_IRQ, ib_poll_handler); 174 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP); 175 break; 176 case IB_POLL_WORKQUEUE: 177 cq->comp_handler = ib_cq_completion_workqueue; 178 INIT_WORK(&cq->work, ib_cq_poll_work); 179 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP); 180 break; 181 default: 182 ret = -EINVAL; 183 goto out_free_wc; 184 } 185 186 return cq; 187 188 out_free_wc: 189 kfree(cq->wc); 190 rdma_restrack_del(&cq->res); 191 out_destroy_cq: 192 cq->device->destroy_cq(cq); 193 return ERR_PTR(ret); 194 } 195 EXPORT_SYMBOL(__ib_alloc_cq); 196 197 /** 198 * ib_free_cq - free a completion queue 199 * @cq: completion queue to free. 200 */ 201 void ib_free_cq(struct ib_cq *cq) 202 { 203 int ret; 204 205 if (WARN_ON_ONCE(atomic_read(&cq->usecnt))) 206 return; 207 208 switch (cq->poll_ctx) { 209 case IB_POLL_DIRECT: 210 break; 211 case IB_POLL_SOFTIRQ: 212 irq_poll_disable(&cq->iop); 213 break; 214 case IB_POLL_WORKQUEUE: 215 cancel_work_sync(&cq->work); 216 break; 217 default: 218 WARN_ON_ONCE(1); 219 } 220 221 kfree(cq->wc); 222 rdma_restrack_del(&cq->res); 223 ret = cq->device->destroy_cq(cq); 224 WARN_ON_ONCE(ret); 225 } 226 EXPORT_SYMBOL(ib_free_cq); 227