sys/dev/nvme/nvme.c

   1 /*-
   2  * Copyright (C) 2012 Intel Corporation
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include <sys/param.h>
  31 #include <sys/bus.h>
  32 #include <sys/conf.h>
  33 #include <sys/module.h>
  34
  35 #include <vm/uma.h>
  36
  37 #include <dev/pci/pcireg.h>
  38 #include <dev/pci/pcivar.h>
  39
  40 #include "nvme_private.h"
  41
  42 struct nvme_consumer {
  43         nvme_consumer_cb_fn_t           cb_fn;
  44         void                            *cb_arg;
  45 };
  46
  47 struct nvme_consumer nvme_consumer[NVME_MAX_CONSUMERS];
  48
  49 uma_zone_t nvme_request_zone;
  50
  51 MALLOC_DEFINE(M_NVME, "nvme", "nvme(4) memory allocations");
  52
  53 static int    nvme_probe(device_t);
  54 static int    nvme_attach(device_t);
  55 static int    nvme_detach(device_t);
  56
  57 static devclass_t nvme_devclass;
  58
  59 static device_method_t nvme_pci_methods[] = {
  60         /* Device interface */
  61         DEVMETHOD(device_probe,     nvme_probe),
  62         DEVMETHOD(device_attach,    nvme_attach),
  63         DEVMETHOD(device_detach,    nvme_detach),
  64         { 0, 0 }
  65 };
  66
  67 static driver_t nvme_pci_driver = {
  68         "nvme",
  69         nvme_pci_methods,
  70         sizeof(struct nvme_controller),
  71 };
  72
  73 DRIVER_MODULE(nvme, pci, nvme_pci_driver, nvme_devclass, 0, 0);
  74 MODULE_VERSION(nvme, 1);
  75
  76 static struct _pcsid
  77 {
  78         u_int32_t   type;
  79         const char  *desc;
  80 } pci_ids[] = {
  81         { 0x01118086,           "NVMe Controller"  },
  82         { CHATHAM_PCI_ID,       "Chatham Prototype NVMe Controller"  },
  83         { IDT_PCI_ID,           "IDT NVMe Controller"  },
  84         { 0x00000000,           NULL  }
  85 };
  86
  87 static int
  88 nvme_probe (device_t device)
  89 {
  90         struct _pcsid   *ep;
  91         u_int32_t       type;
  92
  93         type = pci_get_devid(device);
  94         ep = pci_ids;
  95
  96         while (ep->type && ep->type != type)
  97                 ++ep;
  98
  99         if (ep->desc) {
 100                 device_set_desc(device, ep->desc);
 101                 return (BUS_PROBE_DEFAULT);
 102         }
 103
 104 #if defined(PCIS_STORAGE_NVM)
 105         if (pci_get_class(device)    == PCIC_STORAGE &&
 106             pci_get_subclass(device) == PCIS_STORAGE_NVM &&
 107             pci_get_progif(device)   == PCIP_STORAGE_NVM_ENTERPRISE_NVMHCI_1_0) {
 108                 device_set_desc(device, "Generic NVMe Device");
 109                 return (BUS_PROBE_GENERIC);
 110         }
 111 #endif
 112
 113         return (ENXIO);
 114 }
 115
 116 static void
 117 nvme_init(void)
 118 {
 119         nvme_request_zone = uma_zcreate("nvme_request",
 120             sizeof(struct nvme_request), NULL, NULL, NULL, NULL, 0, 0);
 121 }
 122
 123 SYSINIT(nvme_register, SI_SUB_DRIVERS, SI_ORDER_SECOND, nvme_init, NULL);
 124
 125 static void
 126 nvme_uninit(void)
 127 {
 128         uma_zdestroy(nvme_request_zone);
 129 }
 130
 131 SYSUNINIT(nvme_unregister, SI_SUB_DRIVERS, SI_ORDER_SECOND, nvme_uninit, NULL);
 132
 133 static void
 134 nvme_load(void)
 135 {
 136 }
 137
 138 static void
 139 nvme_unload(void)
 140 {
 141 }
 142
 143 static void
 144 nvme_shutdown(void)
 145 {
 146         device_t                *devlist;
 147         struct nvme_controller  *ctrlr;
 148         union cc_register       cc;
 149         union csts_register     csts;
 150         int                     dev, devcount;
 151
 152         if (devclass_get_devices(nvme_devclass, &devlist, &devcount))
 153                 return;
 154
 155         for (dev = 0; dev < devcount; dev++) {
 156                 /*
 157                  * Only notify controller of shutdown when a real shutdown is
 158                  *  in process, not when a module unload occurs.  It seems at
 159                  *  least some controllers (Chatham at least) don't let you
 160                  *  re-enable the controller after shutdown notification has
 161                  *  been received.
 162                  */
 163                 ctrlr = DEVICE2SOFTC(devlist[dev]);
 164                 cc.raw = nvme_mmio_read_4(ctrlr, cc);
 165                 cc.bits.shn = NVME_SHN_NORMAL;
 166                 nvme_mmio_write_4(ctrlr, cc, cc.raw);
 167                 csts.raw = nvme_mmio_read_4(ctrlr, csts);
 168                 while (csts.bits.shst != NVME_SHST_COMPLETE) {
 169                         DELAY(5);
 170                         csts.raw = nvme_mmio_read_4(ctrlr, csts);
 171                 }
 172         }
 173
 174         free(devlist, M_TEMP);
 175 }
 176
 177 static int
 178 nvme_modevent(module_t mod, int type, void *arg)
 179 {
 180
 181         switch (type) {
 182         case MOD_LOAD:
 183                 nvme_load();
 184                 break;
 185         case MOD_UNLOAD:
 186                 nvme_unload();
 187                 break;
 188         case MOD_SHUTDOWN:
 189                 nvme_shutdown();
 190                 break;
 191         default:
 192                 break;
 193         }
 194
 195         return (0);
 196 }
 197
 198 moduledata_t nvme_mod = {
 199         "nvme",
 200         (modeventhand_t)nvme_modevent,
 201         0
 202 };
 203
 204 DECLARE_MODULE(nvme, nvme_mod, SI_SUB_DRIVERS, SI_ORDER_FIRST);
 205
 206 void
 207 nvme_dump_command(struct nvme_command *cmd)
 208 {
 209         printf("opc:%x f:%x r1:%x cid:%x nsid:%x r2:%x r3:%x "
 210             "mptr:%qx prp1:%qx prp2:%qx cdw:%x %x %x %x %x %x\n",
 211             cmd->opc, cmd->fuse, cmd->rsvd1, cmd->cid, cmd->nsid,
 212             cmd->rsvd2, cmd->rsvd3,
 213             (long long unsigned int)cmd->mptr,
 214             (long long unsigned int)cmd->prp1,
 215             (long long unsigned int)cmd->prp2,
 216             cmd->cdw10, cmd->cdw11, cmd->cdw12, cmd->cdw13, cmd->cdw14,
 217             cmd->cdw15);
 218 }
 219
 220 void
 221 nvme_dump_completion(struct nvme_completion *cpl)
 222 {
 223         printf("cdw0:%08x sqhd:%04x sqid:%04x "
 224             "cid:%04x p:%x sc:%02x sct:%x m:%x dnr:%x\n",
 225             cpl->cdw0, cpl->sqhd, cpl->sqid,
 226             cpl->cid, cpl->p, cpl->sf_sc, cpl->sf_sct, cpl->sf_m,
 227             cpl->sf_dnr);
 228 }
 229
 230 void
 231 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
 232 {
 233         struct nvme_tracker     *tr = arg;
 234         uint32_t                cur_nseg;
 235
 236         KASSERT(error == 0, ("nvme_payload_map error != 0\n"));
 237
 238         /*
 239          * Note that we specified PAGE_SIZE for alignment and max
 240          *  segment size when creating the bus dma tags.  So here
 241          *  we can safely just transfer each segment to its
 242          *  associated PRP entry.
 243          */
 244         tr->req->cmd.prp1 = seg[0].ds_addr;
 245
 246         if (nseg == 2) {
 247                 tr->req->cmd.prp2 = seg[1].ds_addr;
 248         } else if (nseg > 2) {
 249                 cur_nseg = 1;
 250                 tr->req->cmd.prp2 = (uint64_t)tr->prp_bus_addr;
 251                 while (cur_nseg < nseg) {
 252                         tr->prp[cur_nseg-1] =
 253                             (uint64_t)seg[cur_nseg].ds_addr;
 254                         cur_nseg++;
 255                 }
 256         }
 257
 258         nvme_qpair_submit_cmd(tr->qpair, tr);
 259 }
 260
 261 static int
 262 nvme_attach(device_t dev)
 263 {
 264         struct nvme_controller  *ctrlr = DEVICE2SOFTC(dev);
 265         int                     status;
 266
 267         status = nvme_ctrlr_construct(ctrlr, dev);
 268
 269         if (status != 0)
 270                 return (status);
 271
 272         /*
 273          * Reset controller twice to ensure we do a transition from cc.en==1
 274          *  to cc.en==0.  This is because we don't really know what status
 275          *  the controller was left in when boot handed off to OS.
 276          */
 277         status = nvme_ctrlr_reset(ctrlr);
 278         if (status != 0)
 279                 return (status);
 280
 281         status = nvme_ctrlr_reset(ctrlr);
 282         if (status != 0)
 283                 return (status);
 284
 285         ctrlr->config_hook.ich_func = nvme_ctrlr_start;
 286         ctrlr->config_hook.ich_arg = ctrlr;
 287
 288         config_intrhook_establish(&ctrlr->config_hook);
 289
 290         return (0);
 291 }
 292
 293 static int
 294 nvme_detach (device_t dev)
 295 {
 296         struct nvme_controller  *ctrlr = DEVICE2SOFTC(dev);
 297         struct nvme_namespace   *ns;
 298         int                     i;
 299
 300         if (ctrlr->taskqueue) {
 301                 taskqueue_drain(ctrlr->taskqueue, &ctrlr->task);
 302                 taskqueue_free(ctrlr->taskqueue);
 303         }
 304
 305         for (i = 0; i < NVME_MAX_NAMESPACES; i++) {
 306                 ns = &ctrlr->ns[i];
 307                 if (ns->cdev)
 308                         destroy_dev(ns->cdev);
 309         }
 310
 311         if (ctrlr->cdev)
 312                 destroy_dev(ctrlr->cdev);
 313
 314         for (i = 0; i < ctrlr->num_io_queues; i++) {
 315                 nvme_io_qpair_destroy(&ctrlr->ioq[i]);
 316         }
 317
 318         free(ctrlr->ioq, M_NVME);
 319
 320         nvme_admin_qpair_destroy(&ctrlr->adminq);
 321
 322         if (ctrlr->resource != NULL) {
 323                 bus_release_resource(dev, SYS_RES_MEMORY,
 324                     ctrlr->resource_id, ctrlr->resource);
 325         }
 326
 327 #ifdef CHATHAM2
 328         if (ctrlr->chatham_resource != NULL) {
 329                 bus_release_resource(dev, SYS_RES_MEMORY,
 330                     ctrlr->chatham_resource_id, ctrlr->chatham_resource);
 331         }
 332 #endif
 333
 334         if (ctrlr->tag)
 335                 bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
 336
 337         if (ctrlr->res)
 338                 bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
 339                     rman_get_rid(ctrlr->res), ctrlr->res);
 340
 341         if (ctrlr->msix_enabled)
 342                 pci_release_msi(dev);
 343
 344         return (0);
 345 }
 346
 347 static void
 348 nvme_notify_consumer(struct nvme_consumer *consumer)
 349 {
 350         device_t                *devlist;
 351         struct nvme_controller  *ctrlr;
 352         int                     dev, ns, devcount;
 353
 354         if (devclass_get_devices(nvme_devclass, &devlist, &devcount))
 355                 return;
 356
 357         for (dev = 0; dev < devcount; dev++) {
 358                 ctrlr = DEVICE2SOFTC(devlist[dev]);
 359                 for (ns = 0; ns < ctrlr->cdata.nn; ns++)
 360                         (*consumer->cb_fn)(consumer->cb_arg, &ctrlr->ns[ns]);
 361         }
 362
 363         free(devlist, M_TEMP);
 364 }
 365
 366 struct nvme_consumer *
 367 nvme_register_consumer(nvme_consumer_cb_fn_t cb_fn, void *cb_arg)
 368 {
 369         int i;
 370
 371         /*
 372          * TODO: add locking around consumer registration.  Not an issue
 373          *  right now since we only have one nvme consumer - nvd(4).
 374          */
 375         for (i = 0; i < NVME_MAX_CONSUMERS; i++)
 376                 if (nvme_consumer[i].cb_fn == NULL) {
 377                         nvme_consumer[i].cb_fn = cb_fn;
 378                         nvme_consumer[i].cb_arg = cb_arg;
 379
 380                         nvme_notify_consumer(&nvme_consumer[i]);
 381                         return (&nvme_consumer[i]);
 382                 }
 383
 384         printf("nvme(4): consumer not registered - no slots available\n");
 385         return (NULL);
 386 }
 387
 388 void
 389 nvme_unregister_consumer(struct nvme_consumer *consumer)
 390 {
 391
 392         consumer->cb_fn = NULL;
 393         consumer->cb_arg = NULL;
 394 }
 395