drivers/infiniband/hw/hfi1/file_ops.c

   1 /*
   2  * Copyright(c) 2015, 2016 Intel Corporation.
   3  *
   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.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   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.
  12  *
  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.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  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
  29  *    distribution.
  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.
  33  *
  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.
  45  *
  46  */
  47 #include <linux/poll.h>
  48 #include <linux/cdev.h>
  49 #include <linux/vmalloc.h>
  50 #include <linux/io.h>
  51
  52 #include <rdma/ib.h>
  53
  54 #include "hfi.h"
  55 #include "pio.h"
  56 #include "device.h"
  57 #include "common.h"
  58 #include "trace.h"
  59 #include "user_sdma.h"
  60 #include "user_exp_rcv.h"
  61 #include "eprom.h"
  62 #include "aspm.h"
  63 #include "mmu_rb.h"
  64
  65 #undef pr_fmt
  66 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
  67
  68 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
  69
  70 /*
  71  * File operation functions
  72  */
  73 static int hfi1_file_open(struct inode *, struct file *);
  74 static int hfi1_file_close(struct inode *, struct file *);
  75 static ssize_t hfi1_write_iter(struct kiocb *, struct iov_iter *);
  76 static unsigned int hfi1_poll(struct file *, struct poll_table_struct *);
  77 static int hfi1_file_mmap(struct file *, struct vm_area_struct *);
  78
  79 static u64 kvirt_to_phys(void *);
  80 static int assign_ctxt(struct file *, struct hfi1_user_info *);
  81 static int init_subctxts(struct hfi1_ctxtdata *, const struct hfi1_user_info *);
  82 static int user_init(struct file *);
  83 static int get_ctxt_info(struct file *, void __user *, __u32);
  84 static int get_base_info(struct file *, void __user *, __u32);
  85 static int setup_ctxt(struct file *);
  86 static int setup_subctxt(struct hfi1_ctxtdata *);
  87 static int get_user_context(struct file *, struct hfi1_user_info *, int);
  88 static int find_shared_ctxt(struct file *, const struct hfi1_user_info *);
  89 static int allocate_ctxt(struct file *, struct hfi1_devdata *,
  90                          struct hfi1_user_info *);
  91 static unsigned int poll_urgent(struct file *, struct poll_table_struct *);
  92 static unsigned int poll_next(struct file *, struct poll_table_struct *);
  93 static int user_event_ack(struct hfi1_ctxtdata *, int, unsigned long);
  94 static int set_ctxt_pkey(struct hfi1_ctxtdata *, unsigned, u16);
  95 static int manage_rcvq(struct hfi1_ctxtdata *, unsigned, int);
  96 static int vma_fault(struct vm_area_struct *, struct vm_fault *);
  97 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
  98                             unsigned long arg);
  99
 100 static const struct file_operations hfi1_file_ops = {
 101         .owner = THIS_MODULE,
 102         .write_iter = hfi1_write_iter,
 103         .open = hfi1_file_open,
 104         .release = hfi1_file_close,
 105         .unlocked_ioctl = hfi1_file_ioctl,
 106         .poll = hfi1_poll,
 107         .mmap = hfi1_file_mmap,
 108         .llseek = noop_llseek,
 109 };
 110
 111 static struct vm_operations_struct vm_ops = {
 112         .fault = vma_fault,
 113 };
 114
 115 /*
 116  * Types of memories mapped into user processes' space
 117  */
 118 enum mmap_types {
 119         PIO_BUFS = 1,
 120         PIO_BUFS_SOP,
 121         PIO_CRED,
 122         RCV_HDRQ,
 123         RCV_EGRBUF,
 124         UREGS,
 125         EVENTS,
 126         STATUS,
 127         RTAIL,
 128         SUBCTXT_UREGS,
 129         SUBCTXT_RCV_HDRQ,
 130         SUBCTXT_EGRBUF,
 131         SDMA_COMP
 132 };
 133
 134 /*
 135  * Masks and offsets defining the mmap tokens
 136  */
 137 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
 138 #define HFI1_MMAP_OFFSET_SHIFT  0
 139 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
 140 #define HFI1_MMAP_SUBCTXT_SHIFT 12
 141 #define HFI1_MMAP_CTXT_MASK     0xffULL
 142 #define HFI1_MMAP_CTXT_SHIFT    16
 143 #define HFI1_MMAP_TYPE_MASK     0xfULL
 144 #define HFI1_MMAP_TYPE_SHIFT    24
 145 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
 146 #define HFI1_MMAP_MAGIC_SHIFT   32
 147
 148 #define HFI1_MMAP_MAGIC         0xdabbad00
 149
 150 #define HFI1_MMAP_TOKEN_SET(field, val) \
 151         (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
 152 #define HFI1_MMAP_TOKEN_GET(field, token) \
 153         (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
 154 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
 155         (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
 156         HFI1_MMAP_TOKEN_SET(TYPE, type) | \
 157         HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
 158         HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
 159         HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
 160
 161 #define dbg(fmt, ...)                           \
 162         pr_info(fmt, ##__VA_ARGS__)
 163
 164 static inline int is_valid_mmap(u64 token)
 165 {
 166         return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
 167 }
 168
 169 static int hfi1_file_open(struct inode *inode, struct file *fp)
 170 {
 171         struct hfi1_devdata *dd = container_of(inode->i_cdev,
 172                                                struct hfi1_devdata,
 173                                                user_cdev);
 174
 175         /* Just take a ref now. Not all opens result in a context assign */
 176         kobject_get(&dd->kobj);
 177
 178         /* The real work is performed later in assign_ctxt() */
 179         fp->private_data = kzalloc(sizeof(struct hfi1_filedata), GFP_KERNEL);
 180         if (fp->private_data) /* no cpu affinity by default */
 181                 ((struct hfi1_filedata *)fp->private_data)->rec_cpu_num = -1;
 182         return fp->private_data ? 0 : -ENOMEM;
 183 }
 184
 185 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
 186                             unsigned long arg)
 187 {
 188         struct hfi1_filedata *fd = fp->private_data;
 189         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 190         struct hfi1_user_info uinfo;
 191         struct hfi1_tid_info tinfo;
 192         int ret = 0;
 193         unsigned long addr;
 194         int uval = 0;
 195         unsigned long ul_uval = 0;
 196         u16 uval16 = 0;
 197
 198         hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
 199         if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
 200             cmd != HFI1_IOCTL_GET_VERS &&
 201             !uctxt)
 202                 return -EINVAL;
 203
 204         switch (cmd) {
 205         case HFI1_IOCTL_ASSIGN_CTXT:
 206                 if (uctxt)
 207                         return -EINVAL;
 208
 209                 if (copy_from_user(&uinfo,
 210                                    (struct hfi1_user_info __user *)arg,
 211                                    sizeof(uinfo)))
 212                         return -EFAULT;
 213
 214                 ret = assign_ctxt(fp, &uinfo);
 215                 if (ret < 0)
 216                         return ret;
 217                 setup_ctxt(fp);
 218                 if (ret)
 219                         return ret;
 220                 ret = user_init(fp);
 221                 break;
 222         case HFI1_IOCTL_CTXT_INFO:
 223                 ret = get_ctxt_info(fp, (void __user *)(unsigned long)arg,
 224                                     sizeof(struct hfi1_ctxt_info));
 225                 break;
 226         case HFI1_IOCTL_USER_INFO:
 227                 ret = get_base_info(fp, (void __user *)(unsigned long)arg,
 228                                     sizeof(struct hfi1_base_info));
 229                 break;
 230         case HFI1_IOCTL_CREDIT_UPD:
 231                 if (uctxt && uctxt->sc)
 232                         sc_return_credits(uctxt->sc);
 233                 break;
 234
 235         case HFI1_IOCTL_TID_UPDATE:
 236                 if (copy_from_user(&tinfo,
 237                                    (struct hfi11_tid_info __user *)arg,
 238                                    sizeof(tinfo)))
 239                         return -EFAULT;
 240
 241                 ret = hfi1_user_exp_rcv_setup(fp, &tinfo);
 242                 if (!ret) {
 243                         /*
 244                          * Copy the number of tidlist entries we used
 245                          * and the length of the buffer we registered.
 246                          * These fields are adjacent in the structure so
 247                          * we can copy them at the same time.
 248                          */
 249                         addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
 250                         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 251                                          sizeof(tinfo.tidcnt) +
 252                                          sizeof(tinfo.length)))
 253                                 ret = -EFAULT;
 254                 }
 255                 break;
 256
 257         case HFI1_IOCTL_TID_FREE:
 258                 if (copy_from_user(&tinfo,
 259                                    (struct hfi11_tid_info __user *)arg,
 260                                    sizeof(tinfo)))
 261                         return -EFAULT;
 262
 263                 ret = hfi1_user_exp_rcv_clear(fp, &tinfo);
 264                 if (ret)
 265                         break;
 266                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
 267                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 268                                  sizeof(tinfo.tidcnt)))
 269                         ret = -EFAULT;
 270                 break;
 271
 272         case HFI1_IOCTL_TID_INVAL_READ:
 273                 if (copy_from_user(&tinfo,
 274                                    (struct hfi11_tid_info __user *)arg,
 275                                    sizeof(tinfo)))
 276                         return -EFAULT;
 277
 278                 ret = hfi1_user_exp_rcv_invalid(fp, &tinfo);
 279                 if (ret)
 280                         break;
 281                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
 282                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 283                                  sizeof(tinfo.tidcnt)))
 284                         ret = -EFAULT;
 285                 break;
 286
 287         case HFI1_IOCTL_RECV_CTRL:
 288                 ret = get_user(uval, (int __user *)arg);
 289                 if (ret != 0)
 290                         return -EFAULT;
 291                 ret = manage_rcvq(uctxt, fd->subctxt, uval);
 292                 break;
 293
 294         case HFI1_IOCTL_POLL_TYPE:
 295                 ret = get_user(uval, (int __user *)arg);
 296                 if (ret != 0)
 297                         return -EFAULT;
 298                 uctxt->poll_type = (typeof(uctxt->poll_type))uval;
 299                 break;
 300
 301         case HFI1_IOCTL_ACK_EVENT:
 302                 ret = get_user(ul_uval, (unsigned long __user *)arg);
 303                 if (ret != 0)
 304                         return -EFAULT;
 305                 ret = user_event_ack(uctxt, fd->subctxt, ul_uval);
 306                 break;
 307
 308         case HFI1_IOCTL_SET_PKEY:
 309                 ret = get_user(uval16, (u16 __user *)arg);
 310                 if (ret != 0)
 311                         return -EFAULT;
 312                 if (HFI1_CAP_IS_USET(PKEY_CHECK))
 313                         ret = set_ctxt_pkey(uctxt, fd->subctxt, uval16);
 314                 else
 315                         return -EPERM;
 316                 break;
 317
 318         case HFI1_IOCTL_CTXT_RESET: {
 319                 struct send_context *sc;
 320                 struct hfi1_devdata *dd;
 321
 322                 if (!uctxt || !uctxt->dd || !uctxt->sc)
 323                         return -EINVAL;
 324
 325                 /*
 326                  * There is no protection here. User level has to
 327                  * guarantee that no one will be writing to the send
 328                  * context while it is being re-initialized.
 329                  * If user level breaks that guarantee, it will break
 330                  * it's own context and no one else's.
 331                  */
 332                 dd = uctxt->dd;
 333                 sc = uctxt->sc;
 334                 /*
 335                  * Wait until the interrupt handler has marked the
 336                  * context as halted or frozen. Report error if we time
 337                  * out.
 338                  */
 339                 wait_event_interruptible_timeout(
 340                         sc->halt_wait, (sc->flags & SCF_HALTED),
 341                         msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 342                 if (!(sc->flags & SCF_HALTED))
 343                         return -ENOLCK;
 344
 345                 /*
 346                  * If the send context was halted due to a Freeze,
 347                  * wait until the device has been "unfrozen" before
 348                  * resetting the context.
 349                  */
 350                 if (sc->flags & SCF_FROZEN) {
 351                         wait_event_interruptible_timeout(
 352                                 dd->event_queue,
 353                                 !(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
 354                                 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 355                         if (dd->flags & HFI1_FROZEN)
 356                                 return -ENOLCK;
 357
 358                         if (dd->flags & HFI1_FORCED_FREEZE)
 359                                 /*
 360                                  * Don't allow context reset if we are into
 361                                  * forced freeze
 362                                  */
 363                                 return -ENODEV;
 364
 365                         sc_disable(sc);
 366                         ret = sc_enable(sc);
 367                         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
 368                                      uctxt->ctxt);
 369                 } else {
 370                         ret = sc_restart(sc);
 371                 }
 372                 if (!ret)
 373                         sc_return_credits(sc);
 374                 break;
 375         }
 376
 377         case HFI1_IOCTL_GET_VERS:
 378                 uval = HFI1_USER_SWVERSION;
 379                 if (put_user(uval, (int __user *)arg))
 380                         return -EFAULT;
 381                 break;
 382
 383         default:
 384                 return -EINVAL;
 385         }
 386
 387         return ret;
 388 }
 389
 390 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
 391 {
 392         struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
 393         struct hfi1_user_sdma_pkt_q *pq = fd->pq;
 394         struct hfi1_user_sdma_comp_q *cq = fd->cq;
 395         int ret = 0, done = 0, reqs = 0;
 396         unsigned long dim = from->nr_segs;
 397
 398         if (!cq || !pq) {
 399                 ret = -EIO;
 400                 goto done;
 401         }
 402
 403         if (!iter_is_iovec(from) || !dim) {
 404                 ret = -EINVAL;
 405                 goto done;
 406         }
 407
 408         hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
 409                   fd->uctxt->ctxt, fd->subctxt, dim);
 410
 411         if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
 412                 ret = -ENOSPC;
 413                 goto done;
 414         }
 415
 416         while (dim) {
 417                 unsigned long count = 0;
 418
 419                 ret = hfi1_user_sdma_process_request(
 420                         kiocb->ki_filp, (struct iovec *)(from->iov + done),
 421                         dim, &count);
 422                 if (ret)
 423                         goto done;
 424                 dim -= count;
 425                 done += count;
 426                 reqs++;
 427         }
 428 done:
 429         return ret ? ret : reqs;
 430 }
 431
 432 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
 433 {
 434         struct hfi1_filedata *fd = fp->private_data;
 435         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 436         struct hfi1_devdata *dd;
 437         unsigned long flags, pfn;
 438         u64 token = vma->vm_pgoff << PAGE_SHIFT,
 439                 memaddr = 0;
 440         u8 subctxt, mapio = 0, vmf = 0, type;
 441         ssize_t memlen = 0;
 442         int ret = 0;
 443         u16 ctxt;
 444
 445         if (!is_valid_mmap(token) || !uctxt ||
 446             !(vma->vm_flags & VM_SHARED)) {
 447                 ret = -EINVAL;
 448                 goto done;
 449         }
 450         dd = uctxt->dd;
 451         ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
 452         subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
 453         type = HFI1_MMAP_TOKEN_GET(TYPE, token);
 454         if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
 455                 ret = -EINVAL;
 456                 goto done;
 457         }
 458
 459         flags = vma->vm_flags;
 460
 461         switch (type) {
 462         case PIO_BUFS:
 463         case PIO_BUFS_SOP:
 464                 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
 465                                 /* chip pio base */
 466                            (uctxt->sc->hw_context * BIT(16))) +
 467                                 /* 64K PIO space / ctxt */
 468                         (type == PIO_BUFS_SOP ?
 469                                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
 470                 /*
 471                  * Map only the amount allocated to the context, not the
 472                  * entire available context's PIO space.
 473                  */
 474                 memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
 475                 flags &= ~VM_MAYREAD;
 476                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 477                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 478                 mapio = 1;
 479                 break;
 480         case PIO_CRED:
 481                 if (flags & VM_WRITE) {
 482                         ret = -EPERM;
 483                         goto done;
 484                 }
 485                 /*
 486                  * The credit return location for this context could be on the
 487                  * second or third page allocated for credit returns (if number
 488                  * of enabled contexts > 64 and 128 respectively).
 489                  */
 490                 memaddr = dd->cr_base[uctxt->numa_id].pa +
 491                         (((u64)uctxt->sc->hw_free -
 492                           (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
 493                 memlen = PAGE_SIZE;
 494                 flags &= ~VM_MAYWRITE;
 495                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 496                 /*
 497                  * The driver has already allocated memory for credit
 498                  * returns and programmed it into the chip. Has that
 499                  * memory been flagged as non-cached?
 500                  */
 501                 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
 502                 mapio = 1;
 503                 break;
 504         case RCV_HDRQ:
 505                 memaddr = uctxt->rcvhdrq_phys;
 506                 memlen = uctxt->rcvhdrq_size;
 507                 break;
 508         case RCV_EGRBUF: {
 509                 unsigned long addr;
 510                 int i;
 511                 /*
 512                  * The RcvEgr buffer need to be handled differently
 513                  * as multiple non-contiguous pages need to be mapped
 514                  * into the user process.
 515                  */
 516                 memlen = uctxt->egrbufs.size;
 517                 if ((vma->vm_end - vma->vm_start) != memlen) {
 518                         dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
 519                                    (vma->vm_end - vma->vm_start), memlen);
 520                         ret = -EINVAL;
 521                         goto done;
 522                 }
 523                 if (vma->vm_flags & VM_WRITE) {
 524                         ret = -EPERM;
 525                         goto done;
 526                 }
 527                 vma->vm_flags &= ~VM_MAYWRITE;
 528                 addr = vma->vm_start;
 529                 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
 530                         ret = remap_pfn_range(
 531                                 vma, addr,
 532                                 uctxt->egrbufs.buffers[i].phys >> PAGE_SHIFT,
 533                                 uctxt->egrbufs.buffers[i].len,
 534                                 vma->vm_page_prot);
 535                         if (ret < 0)
 536                                 goto done;
 537                         addr += uctxt->egrbufs.buffers[i].len;
 538                 }
 539                 ret = 0;
 540                 goto done;
 541         }
 542         case UREGS:
 543                 /*
 544                  * Map only the page that contains this context's user
 545                  * registers.
 546                  */
 547                 memaddr = (unsigned long)
 548                         (dd->physaddr + RXE_PER_CONTEXT_USER)
 549                         + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
 550                 /*
 551                  * TidFlow table is on the same page as the rest of the
 552                  * user registers.
 553                  */
 554                 memlen = PAGE_SIZE;
 555                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 556                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 557                 mapio = 1;
 558                 break;
 559         case EVENTS:
 560                 /*
 561                  * Use the page where this context's flags are. User level
 562                  * knows where it's own bitmap is within the page.
 563                  */
 564                 memaddr = (unsigned long)(dd->events +
 565                                           ((uctxt->ctxt - dd->first_user_ctxt) *
 566                                            HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
 567                 memlen = PAGE_SIZE;
 568                 /*
 569                  * v3.7 removes VM_RESERVED but the effect is kept by
 570                  * using VM_IO.
 571                  */
 572                 flags |= VM_IO | VM_DONTEXPAND;
 573                 vmf = 1;
 574                 break;
 575         case STATUS:
 576                 memaddr = kvirt_to_phys((void *)dd->status);
 577                 memlen = PAGE_SIZE;
 578                 flags |= VM_IO | VM_DONTEXPAND;
 579                 break;
 580         case RTAIL:
 581                 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
 582                         /*
 583                          * If the memory allocation failed, the context alloc
 584                          * also would have failed, so we would never get here
 585                          */
 586                         ret = -EINVAL;
 587                         goto done;
 588                 }
 589                 if (flags & VM_WRITE) {
 590                         ret = -EPERM;
 591                         goto done;
 592                 }
 593                 memaddr = uctxt->rcvhdrqtailaddr_phys;
 594                 memlen = PAGE_SIZE;
 595                 flags &= ~VM_MAYWRITE;
 596                 break;
 597         case SUBCTXT_UREGS:
 598                 memaddr = (u64)uctxt->subctxt_uregbase;
 599                 memlen = PAGE_SIZE;
 600                 flags |= VM_IO | VM_DONTEXPAND;
 601                 vmf = 1;
 602                 break;
 603         case SUBCTXT_RCV_HDRQ:
 604                 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
 605                 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
 606                 flags |= VM_IO | VM_DONTEXPAND;
 607                 vmf = 1;
 608                 break;
 609         case SUBCTXT_EGRBUF:
 610                 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
 611                 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
 612                 flags |= VM_IO | VM_DONTEXPAND;
 613                 flags &= ~VM_MAYWRITE;
 614                 vmf = 1;
 615                 break;
 616         case SDMA_COMP: {
 617                 struct hfi1_user_sdma_comp_q *cq = fd->cq;
 618
 619                 if (!cq) {
 620                         ret = -EFAULT;
 621                         goto done;
 622                 }
 623                 memaddr = (u64)cq->comps;
 624                 memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
 625                 flags |= VM_IO | VM_DONTEXPAND;
 626                 vmf = 1;
 627                 break;
 628         }
 629         default:
 630                 ret = -EINVAL;
 631                 break;
 632         }
 633
 634         if ((vma->vm_end - vma->vm_start) != memlen) {
 635                 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
 636                           uctxt->ctxt, fd->subctxt,
 637                           (vma->vm_end - vma->vm_start), memlen);
 638                 ret = -EINVAL;
 639                 goto done;
 640         }
 641
 642         vma->vm_flags = flags;
 643         hfi1_cdbg(PROC,
 644                   "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
 645                     ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
 646                     vma->vm_end - vma->vm_start, vma->vm_flags);
 647         pfn = (unsigned long)(memaddr >> PAGE_SHIFT);
 648         if (vmf) {
 649                 vma->vm_pgoff = pfn;
 650                 vma->vm_ops = &vm_ops;
 651                 ret = 0;
 652         } else if (mapio) {
 653                 ret = io_remap_pfn_range(vma, vma->vm_start, pfn, memlen,
 654                                          vma->vm_page_prot);
 655         } else {
 656                 ret = remap_pfn_range(vma, vma->vm_start, pfn, memlen,
 657                                       vma->vm_page_prot);
 658         }
 659 done:
 660         return ret;
 661 }
 662
 663 /*
 664  * Local (non-chip) user memory is not mapped right away but as it is
 665  * accessed by the user-level code.
 666  */
 667 static int vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 668 {
 669         struct page *page;
 670
 671         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 672         if (!page)
 673                 return VM_FAULT_SIGBUS;
 674
 675         get_page(page);
 676         vmf->page = page;
 677
 678         return 0;
 679 }
 680
 681 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
 682 {
 683         struct hfi1_ctxtdata *uctxt;
 684         unsigned pollflag;
 685
 686         uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
 687         if (!uctxt)
 688                 pollflag = POLLERR;
 689         else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
 690                 pollflag = poll_urgent(fp, pt);
 691         else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
 692                 pollflag = poll_next(fp, pt);
 693         else /* invalid */
 694                 pollflag = POLLERR;
 695
 696         return pollflag;
 697 }
 698
 699 static int hfi1_file_close(struct inode *inode, struct file *fp)
 700 {
 701         struct hfi1_filedata *fdata = fp->private_data;
 702         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 703         struct hfi1_devdata *dd = container_of(inode->i_cdev,
 704                                                struct hfi1_devdata,
 705                                                user_cdev);
 706         unsigned long flags, *ev;
 707
 708         fp->private_data = NULL;
 709
 710         if (!uctxt)
 711                 goto done;
 712
 713         hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
 714         mutex_lock(&hfi1_mutex);
 715
 716         flush_wc();
 717         /* drain user sdma queue */
 718         hfi1_user_sdma_free_queues(fdata);
 719
 720         /* release the cpu */
 721         hfi1_put_proc_affinity(dd, fdata->rec_cpu_num);
 722
 723         /*
 724          * Clear any left over, unhandled events so the next process that
 725          * gets this context doesn't get confused.
 726          */
 727         ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
 728                            HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
 729         *ev = 0;
 730
 731         if (--uctxt->cnt) {
 732                 uctxt->active_slaves &= ~(1 << fdata->subctxt);
 733                 uctxt->subpid[fdata->subctxt] = 0;
 734                 mutex_unlock(&hfi1_mutex);
 735                 goto done;
 736         }
 737
 738         spin_lock_irqsave(&dd->uctxt_lock, flags);
 739         /*
 740          * Disable receive context and interrupt available, reset all
 741          * RcvCtxtCtrl bits to default values.
 742          */
 743         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
 744                      HFI1_RCVCTRL_TIDFLOW_DIS |
 745                      HFI1_RCVCTRL_INTRAVAIL_DIS |
 746                      HFI1_RCVCTRL_TAILUPD_DIS |
 747                      HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
 748                      HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
 749                      HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
 750         /* Clear the context's J_KEY */
 751         hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
 752         /*
 753          * Reset context integrity checks to default.
 754          * (writes to CSRs probably belong in chip.c)
 755          */
 756         write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
 757                         hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
 758         sc_disable(uctxt->sc);
 759         uctxt->pid = 0;
 760         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 761
 762         dd->rcd[uctxt->ctxt] = NULL;
 763
 764         hfi1_user_exp_rcv_free(fdata);
 765         hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
 766
 767         uctxt->rcvwait_to = 0;
 768         uctxt->piowait_to = 0;
 769         uctxt->rcvnowait = 0;
 770         uctxt->pionowait = 0;
 771         uctxt->event_flags = 0;
 772
 773         hfi1_stats.sps_ctxts--;
 774         if (++dd->freectxts == dd->num_user_contexts)
 775                 aspm_enable_all(dd);
 776         mutex_unlock(&hfi1_mutex);
 777         hfi1_free_ctxtdata(dd, uctxt);
 778 done:
 779         kobject_put(&dd->kobj);
 780         kfree(fdata);
 781         return 0;
 782 }
 783
 784 /*
 785  * Convert kernel *virtual* addresses to physical addresses.
 786  * This is used to vmalloc'ed addresses.
 787  */
 788 static u64 kvirt_to_phys(void *addr)
 789 {
 790         struct page *page;
 791         u64 paddr = 0;
 792
 793         page = vmalloc_to_page(addr);
 794         if (page)
 795                 paddr = page_to_pfn(page) << PAGE_SHIFT;
 796
 797         return paddr;
 798 }
 799
 800 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
 801 {
 802         int i_minor, ret = 0;
 803         unsigned int swmajor, swminor;
 804
 805         swmajor = uinfo->userversion >> 16;
 806         if (swmajor != HFI1_USER_SWMAJOR) {
 807                 ret = -ENODEV;
 808                 goto done;
 809         }
 810
 811         swminor = uinfo->userversion & 0xffff;
 812
 813         mutex_lock(&hfi1_mutex);
 814         /* First, lets check if we need to setup a shared context? */
 815         if (uinfo->subctxt_cnt) {
 816                 struct hfi1_filedata *fd = fp->private_data;
 817
 818                 ret = find_shared_ctxt(fp, uinfo);
 819                 if (ret < 0)
 820                         goto done_unlock;
 821                 if (ret)
 822                         fd->rec_cpu_num = hfi1_get_proc_affinity(
 823                                 fd->uctxt->dd, fd->uctxt->numa_id);
 824         }
 825
 826         /*
 827          * We execute the following block if we couldn't find a
 828          * shared context or if context sharing is not required.
 829          */
 830         if (!ret) {
 831                 i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
 832                 ret = get_user_context(fp, uinfo, i_minor);
 833         }
 834 done_unlock:
 835         mutex_unlock(&hfi1_mutex);
 836 done:
 837         return ret;
 838 }
 839
 840 static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
 841                             int devno)
 842 {
 843         struct hfi1_devdata *dd = NULL;
 844         int devmax, npresent, nup;
 845
 846         devmax = hfi1_count_units(&npresent, &nup);
 847         if (!npresent)
 848                 return -ENXIO;
 849
 850         if (!nup)
 851                 return -ENETDOWN;
 852
 853         dd = hfi1_lookup(devno);
 854         if (!dd)
 855                 return -ENODEV;
 856         else if (!dd->freectxts)
 857                 return -EBUSY;
 858
 859         return allocate_ctxt(fp, dd, uinfo);
 860 }
 861
 862 static int find_shared_ctxt(struct file *fp,
 863                             const struct hfi1_user_info *uinfo)
 864 {
 865         int devmax, ndev, i;
 866         int ret = 0;
 867         struct hfi1_filedata *fd = fp->private_data;
 868
 869         devmax = hfi1_count_units(NULL, NULL);
 870
 871         for (ndev = 0; ndev < devmax; ndev++) {
 872                 struct hfi1_devdata *dd = hfi1_lookup(ndev);
 873
 874                 if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
 875                         continue;
 876                 for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
 877                         struct hfi1_ctxtdata *uctxt = dd->rcd[i];
 878
 879                         /* Skip ctxts which are not yet open */
 880                         if (!uctxt || !uctxt->cnt)
 881                                 continue;
 882                         /* Skip ctxt if it doesn't match the requested one */
 883                         if (memcmp(uctxt->uuid, uinfo->uuid,
 884                                    sizeof(uctxt->uuid)) ||
 885                             uctxt->jkey != generate_jkey(current_uid()) ||
 886                             uctxt->subctxt_id != uinfo->subctxt_id ||
 887                             uctxt->subctxt_cnt != uinfo->subctxt_cnt)
 888                                 continue;
 889
 890                         /* Verify the sharing process matches the master */
 891                         if (uctxt->userversion != uinfo->userversion ||
 892                             uctxt->cnt >= uctxt->subctxt_cnt) {
 893                                 ret = -EINVAL;
 894                                 goto done;
 895                         }
 896                         fd->uctxt = uctxt;
 897                         fd->subctxt  = uctxt->cnt++;
 898                         uctxt->subpid[fd->subctxt] = current->pid;
 899                         uctxt->active_slaves |= 1 << fd->subctxt;
 900                         ret = 1;
 901                         goto done;
 902                 }
 903         }
 904
 905 done:
 906         return ret;
 907 }
 908
 909 static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
 910                          struct hfi1_user_info *uinfo)
 911 {
 912         struct hfi1_filedata *fd = fp->private_data;
 913         struct hfi1_ctxtdata *uctxt;
 914         unsigned ctxt;
 915         int ret, numa;
 916
 917         if (dd->flags & HFI1_FROZEN) {
 918                 /*
 919                  * Pick an error that is unique from all other errors
 920                  * that are returned so the user process knows that
 921                  * it tried to allocate while the SPC was frozen.  It
 922                  * it should be able to retry with success in a short
 923                  * while.
 924                  */
 925                 return -EIO;
 926         }
 927
 928         for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
 929                 if (!dd->rcd[ctxt])
 930                         break;
 931
 932         if (ctxt == dd->num_rcv_contexts)
 933                 return -EBUSY;
 934
 935         fd->rec_cpu_num = hfi1_get_proc_affinity(dd, -1);
 936         if (fd->rec_cpu_num != -1)
 937                 numa = cpu_to_node(fd->rec_cpu_num);
 938         else
 939                 numa = numa_node_id();
 940         uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
 941         if (!uctxt) {
 942                 dd_dev_err(dd,
 943                            "Unable to allocate ctxtdata memory, failing open\n");
 944                 return -ENOMEM;
 945         }
 946         hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
 947                   uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
 948                   uctxt->numa_id);
 949
 950         /*
 951          * Allocate and enable a PIO send context.
 952          */
 953         uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
 954                              uctxt->dd->node);
 955         if (!uctxt->sc)
 956                 return -ENOMEM;
 957
 958         hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
 959                   uctxt->sc->hw_context);
 960         ret = sc_enable(uctxt->sc);
 961         if (ret)
 962                 return ret;
 963         /*
 964          * Setup shared context resources if the user-level has requested
 965          * shared contexts and this is the 'master' process.
 966          * This has to be done here so the rest of the sub-contexts find the
 967          * proper master.
 968          */
 969         if (uinfo->subctxt_cnt && !fd->subctxt) {
 970                 ret = init_subctxts(uctxt, uinfo);
 971                 /*
 972                  * On error, we don't need to disable and de-allocate the
 973                  * send context because it will be done during file close
 974                  */
 975                 if (ret)
 976                         return ret;
 977         }
 978         uctxt->userversion = uinfo->userversion;
 979         uctxt->pid = current->pid;
 980         uctxt->flags = HFI1_CAP_UGET(MASK);
 981         init_waitqueue_head(&uctxt->wait);
 982         strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
 983         memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
 984         uctxt->jkey = generate_jkey(current_uid());
 985         INIT_LIST_HEAD(&uctxt->sdma_queues);
 986         spin_lock_init(&uctxt->sdma_qlock);
 987         hfi1_stats.sps_ctxts++;
 988         /*
 989          * Disable ASPM when there are open user/PSM contexts to avoid
 990          * issues with ASPM L1 exit latency
 991          */
 992         if (dd->freectxts-- == dd->num_user_contexts)
 993                 aspm_disable_all(dd);
 994         fd->uctxt = uctxt;
 995
 996         return 0;
 997 }
 998
 999 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1000                          const struct hfi1_user_info *uinfo)
1001 {
1002         unsigned num_subctxts;
1003
1004         num_subctxts = uinfo->subctxt_cnt;
1005         if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1006                 return -EINVAL;
1007
1008         uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1009         uctxt->subctxt_id = uinfo->subctxt_id;
1010         uctxt->active_slaves = 1;
1011         uctxt->redirect_seq_cnt = 1;
1012         set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1013
1014         return 0;
1015 }
1016
1017 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1018 {
1019         int ret = 0;
1020         unsigned num_subctxts = uctxt->subctxt_cnt;
1021
1022         uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1023         if (!uctxt->subctxt_uregbase) {
1024                 ret = -ENOMEM;
1025                 goto bail;
1026         }
1027         /* We can take the size of the RcvHdr Queue from the master */
1028         uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1029                                                   num_subctxts);
1030         if (!uctxt->subctxt_rcvhdr_base) {
1031                 ret = -ENOMEM;
1032                 goto bail_ureg;
1033         }
1034
1035         uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1036                                                 num_subctxts);
1037         if (!uctxt->subctxt_rcvegrbuf) {
1038                 ret = -ENOMEM;
1039                 goto bail_rhdr;
1040         }
1041         goto bail;
1042 bail_rhdr:
1043         vfree(uctxt->subctxt_rcvhdr_base);
1044 bail_ureg:
1045         vfree(uctxt->subctxt_uregbase);
1046         uctxt->subctxt_uregbase = NULL;
1047 bail:
1048         return ret;
1049 }
1050
1051 static int user_init(struct file *fp)
1052 {
1053         unsigned int rcvctrl_ops = 0;
1054         struct hfi1_filedata *fd = fp->private_data;
1055         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1056
1057         /* make sure that the context has already been setup */
1058         if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags))
1059                 return -EFAULT;
1060
1061         /* initialize poll variables... */
1062         uctxt->urgent = 0;
1063         uctxt->urgent_poll = 0;
1064
1065         /*
1066          * Now enable the ctxt for receive.
1067          * For chips that are set to DMA the tail register to memory
1068          * when they change (and when the update bit transitions from
1069          * 0 to 1.  So for those chips, we turn it off and then back on.
1070          * This will (very briefly) affect any other open ctxts, but the
1071          * duration is very short, and therefore isn't an issue.  We
1072          * explicitly set the in-memory tail copy to 0 beforehand, so we
1073          * don't have to wait to be sure the DMA update has happened
1074          * (chip resets head/tail to 0 on transition to enable).
1075          */
1076         if (uctxt->rcvhdrtail_kvaddr)
1077                 clear_rcvhdrtail(uctxt);
1078
1079         /* Setup J_KEY before enabling the context */
1080         hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1081
1082         rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1083         if (HFI1_CAP_KGET_MASK(uctxt->flags, HDRSUPP))
1084                 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1085         /*
1086          * Ignore the bit in the flags for now until proper
1087          * support for multiple packet per rcv array entry is
1088          * added.
1089          */
1090         if (!HFI1_CAP_KGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1091                 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1092         if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1093                 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1094         if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1095                 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1096         /*
1097          * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1098          * We can't rely on the correct value to be set from prior
1099          * uses of the chip or ctxt. Therefore, add the rcvctrl op
1100          * for both cases.
1101          */
1102         if (HFI1_CAP_KGET_MASK(uctxt->flags, DMA_RTAIL))
1103                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1104         else
1105                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1106         hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1107
1108         /* Notify any waiting slaves */
1109         if (uctxt->subctxt_cnt) {
1110                 clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1111                 wake_up(&uctxt->wait);
1112         }
1113
1114         return 0;
1115 }
1116
1117 static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1118 {
1119         struct hfi1_ctxt_info cinfo;
1120         struct hfi1_filedata *fd = fp->private_data;
1121         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1122         int ret = 0;
1123
1124         memset(&cinfo, 0, sizeof(cinfo));
1125         ret = hfi1_get_base_kinfo(uctxt, &cinfo);
1126         if (ret < 0)
1127                 goto done;
1128         cinfo.num_active = hfi1_count_active_units();
1129         cinfo.unit = uctxt->dd->unit;
1130         cinfo.ctxt = uctxt->ctxt;
1131         cinfo.subctxt = fd->subctxt;
1132         cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1133                                 uctxt->dd->rcv_entries.group_size) +
1134                 uctxt->expected_count;
1135         cinfo.credits = uctxt->sc->credits;
1136         cinfo.numa_node = uctxt->numa_id;
1137         cinfo.rec_cpu = fd->rec_cpu_num;
1138         cinfo.send_ctxt = uctxt->sc->hw_context;
1139
1140         cinfo.egrtids = uctxt->egrbufs.alloced;
1141         cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1142         cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1143         cinfo.sdma_ring_size = fd->cq->nentries;
1144         cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1145
1146         trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1147         if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1148                 ret = -EFAULT;
1149 done:
1150         return ret;
1151 }
1152
1153 static int setup_ctxt(struct file *fp)
1154 {
1155         struct hfi1_filedata *fd = fp->private_data;
1156         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1157         struct hfi1_devdata *dd = uctxt->dd;
1158         int ret = 0;
1159
1160         /*
1161          * Context should be set up only once, including allocation and
1162          * programming of eager buffers. This is done if context sharing
1163          * is not requested or by the master process.
1164          */
1165         if (!uctxt->subctxt_cnt || !fd->subctxt) {
1166                 ret = hfi1_init_ctxt(uctxt->sc);
1167                 if (ret)
1168                         goto done;
1169
1170                 /* Now allocate the RcvHdr queue and eager buffers. */
1171                 ret = hfi1_create_rcvhdrq(dd, uctxt);
1172                 if (ret)
1173                         goto done;
1174                 ret = hfi1_setup_eagerbufs(uctxt);
1175                 if (ret)
1176                         goto done;
1177                 if (uctxt->subctxt_cnt && !fd->subctxt) {
1178                         ret = setup_subctxt(uctxt);
1179                         if (ret)
1180                                 goto done;
1181                 }
1182         } else {
1183                 ret = wait_event_interruptible(uctxt->wait, !test_bit(
1184                                                HFI1_CTXT_MASTER_UNINIT,
1185                                                &uctxt->event_flags));
1186                 if (ret)
1187                         goto done;
1188         }
1189
1190         ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1191         if (ret)
1192                 goto done;
1193         /*
1194          * Expected receive has to be setup for all processes (including
1195          * shared contexts). However, it has to be done after the master
1196          * context has been fully configured as it depends on the
1197          * eager/expected split of the RcvArray entries.
1198          * Setting it up here ensures that the subcontexts will be waiting
1199          * (due to the above wait_event_interruptible() until the master
1200          * is setup.
1201          */
1202         ret = hfi1_user_exp_rcv_init(fp);
1203         if (ret)
1204                 goto done;
1205
1206         set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1207 done:
1208         return ret;
1209 }
1210
1211 static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1212 {
1213         struct hfi1_base_info binfo;
1214         struct hfi1_filedata *fd = fp->private_data;
1215         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1216         struct hfi1_devdata *dd = uctxt->dd;
1217         ssize_t sz;
1218         unsigned offset;
1219         int ret = 0;
1220
1221         trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1222
1223         memset(&binfo, 0, sizeof(binfo));
1224         binfo.hw_version = dd->revision;
1225         binfo.sw_version = HFI1_KERN_SWVERSION;
1226         binfo.bthqp = kdeth_qp;
1227         binfo.jkey = uctxt->jkey;
1228         /*
1229          * If more than 64 contexts are enabled the allocated credit
1230          * return will span two or three contiguous pages. Since we only
1231          * map the page containing the context's credit return address,
1232          * we need to calculate the offset in the proper page.
1233          */
1234         offset = ((u64)uctxt->sc->hw_free -
1235                   (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1236         binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1237                                                 fd->subctxt, offset);
1238         binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1239                                             fd->subctxt,
1240                                             uctxt->sc->base_addr);
1241         binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1242                                                 uctxt->ctxt,
1243                                                 fd->subctxt,
1244                                                 uctxt->sc->base_addr);
1245         binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1246                                                fd->subctxt,
1247                                                uctxt->rcvhdrq);
1248         binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1249                                                fd->subctxt,
1250                                                uctxt->egrbufs.rcvtids[0].phys);
1251         binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1252                                                  fd->subctxt, 0);
1253         /*
1254          * user regs are at
1255          * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1256          */
1257         binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1258                                             fd->subctxt, 0);
1259         offset = offset_in_page((((uctxt->ctxt - dd->first_user_ctxt) *
1260                     HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1261                   sizeof(*dd->events));
1262         binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1263                                               fd->subctxt,
1264                                               offset);
1265         binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1266                                               fd->subctxt,
1267                                               dd->status);
1268         if (HFI1_CAP_IS_USET(DMA_RTAIL))
1269                 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1270                                                        fd->subctxt, 0);
1271         if (uctxt->subctxt_cnt) {
1272                 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1273                                                         uctxt->ctxt,
1274                                                         fd->subctxt, 0);
1275                 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1276                                                          uctxt->ctxt,
1277                                                          fd->subctxt, 0);
1278                 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1279                                                          uctxt->ctxt,
1280                                                          fd->subctxt, 0);
1281         }
1282         sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1283         if (copy_to_user(ubase, &binfo, sz))
1284                 ret = -EFAULT;
1285         return ret;
1286 }
1287
1288 static unsigned int poll_urgent(struct file *fp,
1289                                 struct poll_table_struct *pt)
1290 {
1291         struct hfi1_filedata *fd = fp->private_data;
1292         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1293         struct hfi1_devdata *dd = uctxt->dd;
1294         unsigned pollflag;
1295
1296         poll_wait(fp, &uctxt->wait, pt);
1297
1298         spin_lock_irq(&dd->uctxt_lock);
1299         if (uctxt->urgent != uctxt->urgent_poll) {
1300                 pollflag = POLLIN | POLLRDNORM;
1301                 uctxt->urgent_poll = uctxt->urgent;
1302         } else {
1303                 pollflag = 0;
1304                 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1305         }
1306         spin_unlock_irq(&dd->uctxt_lock);
1307
1308         return pollflag;
1309 }
1310
1311 static unsigned int poll_next(struct file *fp,
1312                               struct poll_table_struct *pt)
1313 {
1314         struct hfi1_filedata *fd = fp->private_data;
1315         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1316         struct hfi1_devdata *dd = uctxt->dd;
1317         unsigned pollflag;
1318
1319         poll_wait(fp, &uctxt->wait, pt);
1320
1321         spin_lock_irq(&dd->uctxt_lock);
1322         if (hdrqempty(uctxt)) {
1323                 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1324                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1325                 pollflag = 0;
1326         } else {
1327                 pollflag = POLLIN | POLLRDNORM;
1328         }
1329         spin_unlock_irq(&dd->uctxt_lock);
1330
1331         return pollflag;
1332 }
1333
1334 /*
1335  * Find all user contexts in use, and set the specified bit in their
1336  * event mask.
1337  * See also find_ctxt() for a similar use, that is specific to send buffers.
1338  */
1339 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1340 {
1341         struct hfi1_ctxtdata *uctxt;
1342         struct hfi1_devdata *dd = ppd->dd;
1343         unsigned ctxt;
1344         int ret = 0;
1345         unsigned long flags;
1346
1347         if (!dd->events) {
1348                 ret = -EINVAL;
1349                 goto done;
1350         }
1351
1352         spin_lock_irqsave(&dd->uctxt_lock, flags);
1353         for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1354              ctxt++) {
1355                 uctxt = dd->rcd[ctxt];
1356                 if (uctxt) {
1357                         unsigned long *evs = dd->events +
1358                                 (uctxt->ctxt - dd->first_user_ctxt) *
1359                                 HFI1_MAX_SHARED_CTXTS;
1360                         int i;
1361                         /*
1362                          * subctxt_cnt is 0 if not shared, so do base
1363                          * separately, first, then remaining subctxt, if any
1364                          */
1365                         set_bit(evtbit, evs);
1366                         for (i = 1; i < uctxt->subctxt_cnt; i++)
1367                                 set_bit(evtbit, evs + i);
1368                 }
1369         }
1370         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1371 done:
1372         return ret;
1373 }
1374
1375 /**
1376  * manage_rcvq - manage a context's receive queue
1377  * @uctxt: the context
1378  * @subctxt: the sub-context
1379  * @start_stop: action to carry out
1380  *
1381  * start_stop == 0 disables receive on the context, for use in queue
1382  * overflow conditions.  start_stop==1 re-enables, to be used to
1383  * re-init the software copy of the head register
1384  */
1385 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1386                        int start_stop)
1387 {
1388         struct hfi1_devdata *dd = uctxt->dd;
1389         unsigned int rcvctrl_op;
1390
1391         if (subctxt)
1392                 goto bail;
1393         /* atomically clear receive enable ctxt. */
1394         if (start_stop) {
1395                 /*
1396                  * On enable, force in-memory copy of the tail register to
1397                  * 0, so that protocol code doesn't have to worry about
1398                  * whether or not the chip has yet updated the in-memory
1399                  * copy or not on return from the system call. The chip
1400                  * always resets it's tail register back to 0 on a
1401                  * transition from disabled to enabled.
1402                  */
1403                 if (uctxt->rcvhdrtail_kvaddr)
1404                         clear_rcvhdrtail(uctxt);
1405                 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1406         } else {
1407                 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1408         }
1409         hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1410         /* always; new head should be equal to new tail; see above */
1411 bail:
1412         return 0;
1413 }
1414
1415 /*
1416  * clear the event notifier events for this context.
1417  * User process then performs actions appropriate to bit having been
1418  * set, if desired, and checks again in future.
1419  */
1420 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1421                           unsigned long events)
1422 {
1423         int i;
1424         struct hfi1_devdata *dd = uctxt->dd;
1425         unsigned long *evs;
1426
1427         if (!dd->events)
1428                 return 0;
1429
1430         evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1431                             HFI1_MAX_SHARED_CTXTS) + subctxt;
1432
1433         for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1434                 if (!test_bit(i, &events))
1435                         continue;
1436                 clear_bit(i, evs);
1437         }
1438         return 0;
1439 }
1440
1441 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1442                          u16 pkey)
1443 {
1444         int ret = -ENOENT, i, intable = 0;
1445         struct hfi1_pportdata *ppd = uctxt->ppd;
1446         struct hfi1_devdata *dd = uctxt->dd;
1447
1448         if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1449                 ret = -EINVAL;
1450                 goto done;
1451         }
1452
1453         for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1454                 if (pkey == ppd->pkeys[i]) {
1455                         intable = 1;
1456                         break;
1457                 }
1458
1459         if (intable)
1460                 ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1461 done:
1462         return ret;
1463 }
1464
1465 static void user_remove(struct hfi1_devdata *dd)
1466 {
1467
1468         hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1469 }
1470
1471 static int user_add(struct hfi1_devdata *dd)
1472 {
1473         char name[10];
1474         int ret;
1475
1476         snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1477         ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1478                              &dd->user_cdev, &dd->user_device,
1479                              true, &dd->kobj);
1480         if (ret)
1481                 user_remove(dd);
1482
1483         return ret;
1484 }
1485
1486 /*
1487  * Create per-unit files in /dev
1488  */
1489 int hfi1_device_create(struct hfi1_devdata *dd)
1490 {
1491         return user_add(dd);
1492 }
1493
1494 /*
1495  * Remove per-unit files in /dev
1496  * void, core kernel returns no errors for this stuff
1497  */
1498 void hfi1_device_remove(struct hfi1_devdata *dd)
1499 {
1500         user_remove(dd);
1501 }