1 /********************************************************************** 2 * Author: Cavium, Inc. 3 * 4 * Contact: support@cavium.com 5 * Please include "LiquidIO" in the subject. 6 * 7 * Copyright (c) 2003-2016 Cavium, Inc. 8 * 9 * This file is free software; you can redistribute it and/or modify 10 * it under the terms of the GNU General Public License, Version 2, as 11 * published by the Free Software Foundation. 12 * 13 * This file is distributed in the hope that it will be useful, but 14 * AS-IS and WITHOUT ANY WARRANTY; without even the implied warranty 15 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, TITLE, or 16 * NONINFRINGEMENT. See the GNU General Public License for more 17 * details. 18 **********************************************************************/ 19 #include <linux/pci.h> 20 #include <linux/netdevice.h> 21 #include "liquidio_common.h" 22 #include "octeon_droq.h" 23 #include "octeon_iq.h" 24 #include "response_manager.h" 25 #include "octeon_device.h" 26 #include "octeon_main.h" 27 28 static void oct_poll_req_completion(struct work_struct *work); 29 30 int octeon_setup_response_list(struct octeon_device *oct) 31 { 32 int i, ret = 0; 33 struct cavium_wq *cwq; 34 35 for (i = 0; i < MAX_RESPONSE_LISTS; i++) { 36 INIT_LIST_HEAD(&oct->response_list[i].head); 37 spin_lock_init(&oct->response_list[i].lock); 38 atomic_set(&oct->response_list[i].pending_req_count, 0); 39 } 40 spin_lock_init(&oct->cmd_resp_wqlock); 41 42 oct->dma_comp_wq.wq = alloc_workqueue("dma-comp", 43 WQ_MEM_RECLAIM | WQ_PERCPU, 0); 44 if (!oct->dma_comp_wq.wq) { 45 dev_err(&oct->pci_dev->dev, "failed to create wq thread\n"); 46 return -ENOMEM; 47 } 48 49 cwq = &oct->dma_comp_wq; 50 INIT_DELAYED_WORK(&cwq->wk.work, oct_poll_req_completion); 51 cwq->wk.ctxptr = oct; 52 oct->cmd_resp_state = OCT_DRV_ONLINE; 53 54 return ret; 55 } 56 EXPORT_SYMBOL_GPL(octeon_setup_response_list); 57 58 void octeon_delete_response_list(struct octeon_device *oct) 59 { 60 cancel_delayed_work_sync(&oct->dma_comp_wq.wk.work); 61 destroy_workqueue(oct->dma_comp_wq.wq); 62 } 63 EXPORT_SYMBOL_GPL(octeon_delete_response_list); 64 65 int lio_process_ordered_list(struct octeon_device *octeon_dev, 66 u32 force_quit) 67 { 68 struct octeon_response_list *ordered_sc_list; 69 struct octeon_soft_command *sc; 70 int request_complete = 0; 71 int resp_to_process = MAX_ORD_REQS_TO_PROCESS; 72 u32 status; 73 u64 status64; 74 75 octeon_free_sc_done_list(octeon_dev); 76 77 ordered_sc_list = &octeon_dev->response_list[OCTEON_ORDERED_SC_LIST]; 78 79 do { 80 spin_lock_bh(&ordered_sc_list->lock); 81 82 if (list_empty(&ordered_sc_list->head)) { 83 spin_unlock_bh(&ordered_sc_list->lock); 84 return 1; 85 } 86 87 sc = list_first_entry(&ordered_sc_list->head, 88 struct octeon_soft_command, node); 89 90 status = OCTEON_REQUEST_PENDING; 91 92 /* check if octeon has finished DMA'ing a response 93 * to where rptr is pointing to 94 */ 95 status64 = *sc->status_word; 96 97 if (status64 != COMPLETION_WORD_INIT) { 98 /* This logic ensures that all 64b have been written. 99 * 1. check byte 0 for non-FF 100 * 2. if non-FF, then swap result from BE to host order 101 * 3. check byte 7 (swapped to 0) for non-FF 102 * 4. if non-FF, use the low 32-bit status code 103 * 5. if either byte 0 or byte 7 is FF, don't use status 104 */ 105 if ((status64 & 0xff) != 0xff) { 106 octeon_swap_8B_data(&status64, 1); 107 if (((status64 & 0xff) != 0xff)) { 108 /* retrieve 16-bit firmware status */ 109 status = (u32)(status64 & 0xffffULL); 110 if (status) { 111 status = 112 FIRMWARE_STATUS_CODE(status); 113 } else { 114 /* i.e. no error */ 115 status = OCTEON_REQUEST_DONE; 116 } 117 } 118 } 119 } else if (unlikely(force_quit) || (sc->expiry_time && 120 time_after(jiffies, (unsigned long)sc->expiry_time))) { 121 struct octeon_instr_irh *irh = 122 (struct octeon_instr_irh *)&sc->cmd.cmd3.irh; 123 124 dev_err(&octeon_dev->pci_dev->dev, "%s: ", __func__); 125 dev_err(&octeon_dev->pci_dev->dev, 126 "cmd %x/%x/%llx/%llx failed, ", 127 irh->opcode, irh->subcode, 128 sc->cmd.cmd3.ossp[0], sc->cmd.cmd3.ossp[1]); 129 dev_err(&octeon_dev->pci_dev->dev, 130 "timeout (%ld, %ld)\n", 131 (long)jiffies, (long)sc->expiry_time); 132 status = OCTEON_REQUEST_TIMEOUT; 133 } 134 135 if (status != OCTEON_REQUEST_PENDING) { 136 sc->sc_status = status; 137 138 /* we have received a response or we have timed out */ 139 /* remove node from linked list */ 140 list_del(&sc->node); 141 atomic_dec(&octeon_dev->response_list 142 [OCTEON_ORDERED_SC_LIST]. 143 pending_req_count); 144 145 if (!sc->callback) { 146 atomic_inc(&octeon_dev->response_list 147 [OCTEON_DONE_SC_LIST]. 148 pending_req_count); 149 list_add_tail(&sc->node, 150 &octeon_dev->response_list 151 [OCTEON_DONE_SC_LIST].head); 152 153 if (unlikely(READ_ONCE(sc->caller_is_done))) { 154 /* caller does not wait for response 155 * from firmware 156 */ 157 if (status != OCTEON_REQUEST_DONE) { 158 struct octeon_instr_irh *irh; 159 160 irh = 161 (struct octeon_instr_irh *) 162 &sc->cmd.cmd3.irh; 163 dev_dbg 164 (&octeon_dev->pci_dev->dev, 165 "%s: sc failed: opcode=%x, ", 166 __func__, irh->opcode); 167 dev_dbg 168 (&octeon_dev->pci_dev->dev, 169 "subcode=%x, ossp[0]=%llx, ", 170 irh->subcode, 171 sc->cmd.cmd3.ossp[0]); 172 dev_dbg 173 (&octeon_dev->pci_dev->dev, 174 "ossp[1]=%llx, status=%d\n", 175 sc->cmd.cmd3.ossp[1], 176 status); 177 } 178 } else { 179 complete(&sc->complete); 180 } 181 182 spin_unlock_bh(&ordered_sc_list->lock); 183 } else { 184 /* sc with callback function */ 185 if (status == OCTEON_REQUEST_TIMEOUT) { 186 atomic_inc(&octeon_dev->response_list 187 [OCTEON_ZOMBIE_SC_LIST]. 188 pending_req_count); 189 list_add_tail(&sc->node, 190 &octeon_dev->response_list 191 [OCTEON_ZOMBIE_SC_LIST]. 192 head); 193 } 194 195 spin_unlock_bh(&ordered_sc_list->lock); 196 197 sc->callback(octeon_dev, status, 198 sc->callback_arg); 199 /* sc is freed by caller */ 200 } 201 202 request_complete++; 203 204 } else { 205 /* no response yet */ 206 request_complete = 0; 207 spin_unlock_bh 208 (&ordered_sc_list->lock); 209 } 210 211 /* If we hit the Max Ordered requests to process every loop, 212 * we quit 213 * and let this function be invoked the next time the poll 214 * thread runs 215 * to process the remaining requests. This function can take up 216 * the entire CPU if there is no upper limit to the requests 217 * processed. 218 */ 219 if (request_complete >= resp_to_process) 220 break; 221 } while (request_complete); 222 223 return 0; 224 } 225 EXPORT_SYMBOL_GPL(lio_process_ordered_list); 226 227 static void oct_poll_req_completion(struct work_struct *work) 228 { 229 struct cavium_wk *wk = (struct cavium_wk *)work; 230 struct octeon_device *oct = (struct octeon_device *)wk->ctxptr; 231 struct cavium_wq *cwq = &oct->dma_comp_wq; 232 233 lio_process_ordered_list(oct, 0); 234 235 if (atomic_read(&oct->response_list 236 [OCTEON_ORDERED_SC_LIST].pending_req_count)) 237 queue_delayed_work(cwq->wq, &cwq->wk.work, msecs_to_jiffies(1)); 238 } 239