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