drivers/infiniband/hw/hfi1/file_ops.c

   1 /*
   2  * Copyright(c) 2015-2017 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 #include <linux/sched/mm.h>
  52
  53 #include <rdma/ib.h>
  54
  55 #include "hfi.h"
  56 #include "pio.h"
  57 #include "device.h"
  58 #include "common.h"
  59 #include "trace.h"
  60 #include "user_sdma.h"
  61 #include "user_exp_rcv.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 *inode, struct file *fp);
  74 static int hfi1_file_close(struct inode *inode, struct file *fp);
  75 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from);
  76 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt);
  77 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma);
  78
  79 static u64 kvirt_to_phys(void *addr);
  80 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo);
  81 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
  82                          const struct hfi1_user_info *uinfo);
  83 static int user_init(struct hfi1_filedata *fd);
  84 static int get_ctxt_info(struct hfi1_filedata *fd, void __user *ubase,
  85                          __u32 len);
  86 static int get_base_info(struct hfi1_filedata *fd, void __user *ubase,
  87                          __u32 len);
  88 static int setup_ctxt(struct hfi1_filedata *fd);
  89 static int setup_subctxt(struct hfi1_ctxtdata *uctxt);
  90 static int get_user_context(struct hfi1_filedata *fd,
  91                             struct hfi1_user_info *uinfo, int devno);
  92 static int find_shared_ctxt(struct hfi1_filedata *fd,
  93                             const struct hfi1_user_info *uinfo);
  94 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
  95                          struct hfi1_user_info *uinfo);
  96 static unsigned int poll_urgent(struct file *fp, struct poll_table_struct *pt);
  97 static unsigned int poll_next(struct file *fp, struct poll_table_struct *pt);
  98 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
  99                           unsigned long events);
 100 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
 101                          u16 pkey);
 102 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
 103                        int start_stop);
 104 static int vma_fault(struct vm_fault *vmf);
 105 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
 106                             unsigned long arg);
 107
 108 static const struct file_operations hfi1_file_ops = {
 109         .owner = THIS_MODULE,
 110         .write_iter = hfi1_write_iter,
 111         .open = hfi1_file_open,
 112         .release = hfi1_file_close,
 113         .unlocked_ioctl = hfi1_file_ioctl,
 114         .poll = hfi1_poll,
 115         .mmap = hfi1_file_mmap,
 116         .llseek = noop_llseek,
 117 };
 118
 119 static struct vm_operations_struct vm_ops = {
 120         .fault = vma_fault,
 121 };
 122
 123 /*
 124  * Types of memories mapped into user processes' space
 125  */
 126 enum mmap_types {
 127         PIO_BUFS = 1,
 128         PIO_BUFS_SOP,
 129         PIO_CRED,
 130         RCV_HDRQ,
 131         RCV_EGRBUF,
 132         UREGS,
 133         EVENTS,
 134         STATUS,
 135         RTAIL,
 136         SUBCTXT_UREGS,
 137         SUBCTXT_RCV_HDRQ,
 138         SUBCTXT_EGRBUF,
 139         SDMA_COMP
 140 };
 141
 142 /*
 143  * Masks and offsets defining the mmap tokens
 144  */
 145 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
 146 #define HFI1_MMAP_OFFSET_SHIFT  0
 147 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
 148 #define HFI1_MMAP_SUBCTXT_SHIFT 12
 149 #define HFI1_MMAP_CTXT_MASK     0xffULL
 150 #define HFI1_MMAP_CTXT_SHIFT    16
 151 #define HFI1_MMAP_TYPE_MASK     0xfULL
 152 #define HFI1_MMAP_TYPE_SHIFT    24
 153 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
 154 #define HFI1_MMAP_MAGIC_SHIFT   32
 155
 156 #define HFI1_MMAP_MAGIC         0xdabbad00
 157
 158 #define HFI1_MMAP_TOKEN_SET(field, val) \
 159         (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
 160 #define HFI1_MMAP_TOKEN_GET(field, token) \
 161         (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
 162 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
 163         (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
 164         HFI1_MMAP_TOKEN_SET(TYPE, type) | \
 165         HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
 166         HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
 167         HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
 168
 169 #define dbg(fmt, ...)                           \
 170         pr_info(fmt, ##__VA_ARGS__)
 171
 172 static inline int is_valid_mmap(u64 token)
 173 {
 174         return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
 175 }
 176
 177 static int hfi1_file_open(struct inode *inode, struct file *fp)
 178 {
 179         struct hfi1_filedata *fd;
 180         struct hfi1_devdata *dd = container_of(inode->i_cdev,
 181                                                struct hfi1_devdata,
 182                                                user_cdev);
 183
 184         if (!atomic_inc_not_zero(&dd->user_refcount))
 185                 return -ENXIO;
 186
 187         /* Just take a ref now. Not all opens result in a context assign */
 188         kobject_get(&dd->kobj);
 189
 190         /* The real work is performed later in assign_ctxt() */
 191
 192         fd = kzalloc(sizeof(*fd), GFP_KERNEL);
 193
 194         if (fd) {
 195                 fd->rec_cpu_num = -1; /* no cpu affinity by default */
 196                 fd->mm = current->mm;
 197                 mmgrab(fd->mm);
 198                 fp->private_data = fd;
 199         } else {
 200                 fp->private_data = NULL;
 201
 202                 if (atomic_dec_and_test(&dd->user_refcount))
 203                         complete(&dd->user_comp);
 204
 205                 return -ENOMEM;
 206         }
 207
 208         return 0;
 209 }
 210
 211 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
 212                             unsigned long arg)
 213 {
 214         struct hfi1_filedata *fd = fp->private_data;
 215         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 216         struct hfi1_user_info uinfo;
 217         struct hfi1_tid_info tinfo;
 218         int ret = 0;
 219         unsigned long addr;
 220         int uval = 0;
 221         unsigned long ul_uval = 0;
 222         u16 uval16 = 0;
 223
 224         hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
 225         if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
 226             cmd != HFI1_IOCTL_GET_VERS &&
 227             !uctxt)
 228                 return -EINVAL;
 229
 230         switch (cmd) {
 231         case HFI1_IOCTL_ASSIGN_CTXT:
 232                 if (uctxt)
 233                         return -EINVAL;
 234
 235                 if (copy_from_user(&uinfo,
 236                                    (struct hfi1_user_info __user *)arg,
 237                                    sizeof(uinfo)))
 238                         return -EFAULT;
 239
 240                 ret = assign_ctxt(fp, &uinfo);
 241                 if (ret < 0)
 242                         return ret;
 243                 ret = setup_ctxt(fd);
 244                 if (ret)
 245                         return ret;
 246                 ret = user_init(fd);
 247                 break;
 248         case HFI1_IOCTL_CTXT_INFO:
 249                 ret = get_ctxt_info(fd, (void __user *)(unsigned long)arg,
 250                                     sizeof(struct hfi1_ctxt_info));
 251                 break;
 252         case HFI1_IOCTL_USER_INFO:
 253                 ret = get_base_info(fd, (void __user *)(unsigned long)arg,
 254                                     sizeof(struct hfi1_base_info));
 255                 break;
 256         case HFI1_IOCTL_CREDIT_UPD:
 257                 if (uctxt)
 258                         sc_return_credits(uctxt->sc);
 259                 break;
 260
 261         case HFI1_IOCTL_TID_UPDATE:
 262                 if (copy_from_user(&tinfo,
 263                                    (struct hfi11_tid_info __user *)arg,
 264                                    sizeof(tinfo)))
 265                         return -EFAULT;
 266
 267                 ret = hfi1_user_exp_rcv_setup(fd, &tinfo);
 268                 if (!ret) {
 269                         /*
 270                          * Copy the number of tidlist entries we used
 271                          * and the length of the buffer we registered.
 272                          * These fields are adjacent in the structure so
 273                          * we can copy them at the same time.
 274                          */
 275                         addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
 276                         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 277                                          sizeof(tinfo.tidcnt) +
 278                                          sizeof(tinfo.length)))
 279                                 ret = -EFAULT;
 280                 }
 281                 break;
 282
 283         case HFI1_IOCTL_TID_FREE:
 284                 if (copy_from_user(&tinfo,
 285                                    (struct hfi11_tid_info __user *)arg,
 286                                    sizeof(tinfo)))
 287                         return -EFAULT;
 288
 289                 ret = hfi1_user_exp_rcv_clear(fd, &tinfo);
 290                 if (ret)
 291                         break;
 292                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
 293                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 294                                  sizeof(tinfo.tidcnt)))
 295                         ret = -EFAULT;
 296                 break;
 297
 298         case HFI1_IOCTL_TID_INVAL_READ:
 299                 if (copy_from_user(&tinfo,
 300                                    (struct hfi11_tid_info __user *)arg,
 301                                    sizeof(tinfo)))
 302                         return -EFAULT;
 303
 304                 ret = hfi1_user_exp_rcv_invalid(fd, &tinfo);
 305                 if (ret)
 306                         break;
 307                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
 308                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 309                                  sizeof(tinfo.tidcnt)))
 310                         ret = -EFAULT;
 311                 break;
 312
 313         case HFI1_IOCTL_RECV_CTRL:
 314                 ret = get_user(uval, (int __user *)arg);
 315                 if (ret != 0)
 316                         return -EFAULT;
 317                 ret = manage_rcvq(uctxt, fd->subctxt, uval);
 318                 break;
 319
 320         case HFI1_IOCTL_POLL_TYPE:
 321                 ret = get_user(uval, (int __user *)arg);
 322                 if (ret != 0)
 323                         return -EFAULT;
 324                 uctxt->poll_type = (typeof(uctxt->poll_type))uval;
 325                 break;
 326
 327         case HFI1_IOCTL_ACK_EVENT:
 328                 ret = get_user(ul_uval, (unsigned long __user *)arg);
 329                 if (ret != 0)
 330                         return -EFAULT;
 331                 ret = user_event_ack(uctxt, fd->subctxt, ul_uval);
 332                 break;
 333
 334         case HFI1_IOCTL_SET_PKEY:
 335                 ret = get_user(uval16, (u16 __user *)arg);
 336                 if (ret != 0)
 337                         return -EFAULT;
 338                 if (HFI1_CAP_IS_USET(PKEY_CHECK))
 339                         ret = set_ctxt_pkey(uctxt, fd->subctxt, uval16);
 340                 else
 341                         return -EPERM;
 342                 break;
 343
 344         case HFI1_IOCTL_CTXT_RESET: {
 345                 struct send_context *sc;
 346                 struct hfi1_devdata *dd;
 347
 348                 if (!uctxt || !uctxt->dd || !uctxt->sc)
 349                         return -EINVAL;
 350
 351                 /*
 352                  * There is no protection here. User level has to
 353                  * guarantee that no one will be writing to the send
 354                  * context while it is being re-initialized.
 355                  * If user level breaks that guarantee, it will break
 356                  * it's own context and no one else's.
 357                  */
 358                 dd = uctxt->dd;
 359                 sc = uctxt->sc;
 360                 /*
 361                  * Wait until the interrupt handler has marked the
 362                  * context as halted or frozen. Report error if we time
 363                  * out.
 364                  */
 365                 wait_event_interruptible_timeout(
 366                         sc->halt_wait, (sc->flags & SCF_HALTED),
 367                         msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 368                 if (!(sc->flags & SCF_HALTED))
 369                         return -ENOLCK;
 370
 371                 /*
 372                  * If the send context was halted due to a Freeze,
 373                  * wait until the device has been "unfrozen" before
 374                  * resetting the context.
 375                  */
 376                 if (sc->flags & SCF_FROZEN) {
 377                         wait_event_interruptible_timeout(
 378                                 dd->event_queue,
 379                                 !(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
 380                                 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 381                         if (dd->flags & HFI1_FROZEN)
 382                                 return -ENOLCK;
 383
 384                         if (dd->flags & HFI1_FORCED_FREEZE)
 385                                 /*
 386                                  * Don't allow context reset if we are into
 387                                  * forced freeze
 388                                  */
 389                                 return -ENODEV;
 390
 391                         sc_disable(sc);
 392                         ret = sc_enable(sc);
 393                         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
 394                                      uctxt->ctxt);
 395                 } else {
 396                         ret = sc_restart(sc);
 397                 }
 398                 if (!ret)
 399                         sc_return_credits(sc);
 400                 break;
 401         }
 402
 403         case HFI1_IOCTL_GET_VERS:
 404                 uval = HFI1_USER_SWVERSION;
 405                 if (put_user(uval, (int __user *)arg))
 406                         return -EFAULT;
 407                 break;
 408
 409         default:
 410                 return -EINVAL;
 411         }
 412
 413         return ret;
 414 }
 415
 416 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
 417 {
 418         struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
 419         struct hfi1_user_sdma_pkt_q *pq = fd->pq;
 420         struct hfi1_user_sdma_comp_q *cq = fd->cq;
 421         int done = 0, reqs = 0;
 422         unsigned long dim = from->nr_segs;
 423
 424         if (!cq || !pq)
 425                 return -EIO;
 426
 427         if (!iter_is_iovec(from) || !dim)
 428                 return -EINVAL;
 429
 430         hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
 431                   fd->uctxt->ctxt, fd->subctxt, dim);
 432
 433         if (atomic_read(&pq->n_reqs) == pq->n_max_reqs)
 434                 return -ENOSPC;
 435
 436         while (dim) {
 437                 int ret;
 438                 unsigned long count = 0;
 439
 440                 ret = hfi1_user_sdma_process_request(
 441                         fd, (struct iovec *)(from->iov + done),
 442                         dim, &count);
 443                 if (ret) {
 444                         reqs = ret;
 445                         break;
 446                 }
 447                 dim -= count;
 448                 done += count;
 449                 reqs++;
 450         }
 451
 452         return reqs;
 453 }
 454
 455 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
 456 {
 457         struct hfi1_filedata *fd = fp->private_data;
 458         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 459         struct hfi1_devdata *dd;
 460         unsigned long flags;
 461         u64 token = vma->vm_pgoff << PAGE_SHIFT,
 462                 memaddr = 0;
 463         void *memvirt = NULL;
 464         u8 subctxt, mapio = 0, vmf = 0, type;
 465         ssize_t memlen = 0;
 466         int ret = 0;
 467         u16 ctxt;
 468
 469         if (!is_valid_mmap(token) || !uctxt ||
 470             !(vma->vm_flags & VM_SHARED)) {
 471                 ret = -EINVAL;
 472                 goto done;
 473         }
 474         dd = uctxt->dd;
 475         ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
 476         subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
 477         type = HFI1_MMAP_TOKEN_GET(TYPE, token);
 478         if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
 479                 ret = -EINVAL;
 480                 goto done;
 481         }
 482
 483         flags = vma->vm_flags;
 484
 485         switch (type) {
 486         case PIO_BUFS:
 487         case PIO_BUFS_SOP:
 488                 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
 489                                 /* chip pio base */
 490                            (uctxt->sc->hw_context * BIT(16))) +
 491                                 /* 64K PIO space / ctxt */
 492                         (type == PIO_BUFS_SOP ?
 493                                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
 494                 /*
 495                  * Map only the amount allocated to the context, not the
 496                  * entire available context's PIO space.
 497                  */
 498                 memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
 499                 flags &= ~VM_MAYREAD;
 500                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 501                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 502                 mapio = 1;
 503                 break;
 504         case PIO_CRED:
 505                 if (flags & VM_WRITE) {
 506                         ret = -EPERM;
 507                         goto done;
 508                 }
 509                 /*
 510                  * The credit return location for this context could be on the
 511                  * second or third page allocated for credit returns (if number
 512                  * of enabled contexts > 64 and 128 respectively).
 513                  */
 514                 memvirt = dd->cr_base[uctxt->numa_id].va;
 515                 memaddr = virt_to_phys(memvirt) +
 516                         (((u64)uctxt->sc->hw_free -
 517                           (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
 518                 memlen = PAGE_SIZE;
 519                 flags &= ~VM_MAYWRITE;
 520                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 521                 /*
 522                  * The driver has already allocated memory for credit
 523                  * returns and programmed it into the chip. Has that
 524                  * memory been flagged as non-cached?
 525                  */
 526                 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
 527                 mapio = 1;
 528                 break;
 529         case RCV_HDRQ:
 530                 memlen = uctxt->rcvhdrq_size;
 531                 memvirt = uctxt->rcvhdrq;
 532                 break;
 533         case RCV_EGRBUF: {
 534                 unsigned long addr;
 535                 int i;
 536                 /*
 537                  * The RcvEgr buffer need to be handled differently
 538                  * as multiple non-contiguous pages need to be mapped
 539                  * into the user process.
 540                  */
 541                 memlen = uctxt->egrbufs.size;
 542                 if ((vma->vm_end - vma->vm_start) != memlen) {
 543                         dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
 544                                    (vma->vm_end - vma->vm_start), memlen);
 545                         ret = -EINVAL;
 546                         goto done;
 547                 }
 548                 if (vma->vm_flags & VM_WRITE) {
 549                         ret = -EPERM;
 550                         goto done;
 551                 }
 552                 vma->vm_flags &= ~VM_MAYWRITE;
 553                 addr = vma->vm_start;
 554                 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
 555                         memlen = uctxt->egrbufs.buffers[i].len;
 556                         memvirt = uctxt->egrbufs.buffers[i].addr;
 557                         ret = remap_pfn_range(
 558                                 vma, addr,
 559                                 /*
 560                                  * virt_to_pfn() does the same, but
 561                                  * it's not available on x86_64
 562                                  * when CONFIG_MMU is enabled.
 563                                  */
 564                                 PFN_DOWN(__pa(memvirt)),
 565                                 memlen,
 566                                 vma->vm_page_prot);
 567                         if (ret < 0)
 568                                 goto done;
 569                         addr += memlen;
 570                 }
 571                 ret = 0;
 572                 goto done;
 573         }
 574         case UREGS:
 575                 /*
 576                  * Map only the page that contains this context's user
 577                  * registers.
 578                  */
 579                 memaddr = (unsigned long)
 580                         (dd->physaddr + RXE_PER_CONTEXT_USER)
 581                         + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
 582                 /*
 583                  * TidFlow table is on the same page as the rest of the
 584                  * user registers.
 585                  */
 586                 memlen = PAGE_SIZE;
 587                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 588                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 589                 mapio = 1;
 590                 break;
 591         case EVENTS:
 592                 /*
 593                  * Use the page where this context's flags are. User level
 594                  * knows where it's own bitmap is within the page.
 595                  */
 596                 memaddr = (unsigned long)(dd->events +
 597                                   ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
 598                                    HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
 599                 memlen = PAGE_SIZE;
 600                 /*
 601                  * v3.7 removes VM_RESERVED but the effect is kept by
 602                  * using VM_IO.
 603                  */
 604                 flags |= VM_IO | VM_DONTEXPAND;
 605                 vmf = 1;
 606                 break;
 607         case STATUS:
 608                 if (flags & (unsigned long)(VM_WRITE | VM_EXEC)) {
 609                         ret = -EPERM;
 610                         goto done;
 611                 }
 612                 memaddr = kvirt_to_phys((void *)dd->status);
 613                 memlen = PAGE_SIZE;
 614                 flags |= VM_IO | VM_DONTEXPAND;
 615                 break;
 616         case RTAIL:
 617                 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
 618                         /*
 619                          * If the memory allocation failed, the context alloc
 620                          * also would have failed, so we would never get here
 621                          */
 622                         ret = -EINVAL;
 623                         goto done;
 624                 }
 625                 if (flags & VM_WRITE) {
 626                         ret = -EPERM;
 627                         goto done;
 628                 }
 629                 memlen = PAGE_SIZE;
 630                 memvirt = (void *)uctxt->rcvhdrtail_kvaddr;
 631                 flags &= ~VM_MAYWRITE;
 632                 break;
 633         case SUBCTXT_UREGS:
 634                 memaddr = (u64)uctxt->subctxt_uregbase;
 635                 memlen = PAGE_SIZE;
 636                 flags |= VM_IO | VM_DONTEXPAND;
 637                 vmf = 1;
 638                 break;
 639         case SUBCTXT_RCV_HDRQ:
 640                 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
 641                 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
 642                 flags |= VM_IO | VM_DONTEXPAND;
 643                 vmf = 1;
 644                 break;
 645         case SUBCTXT_EGRBUF:
 646                 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
 647                 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
 648                 flags |= VM_IO | VM_DONTEXPAND;
 649                 flags &= ~VM_MAYWRITE;
 650                 vmf = 1;
 651                 break;
 652         case SDMA_COMP: {
 653                 struct hfi1_user_sdma_comp_q *cq = fd->cq;
 654
 655                 if (!cq) {
 656                         ret = -EFAULT;
 657                         goto done;
 658                 }
 659                 memaddr = (u64)cq->comps;
 660                 memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
 661                 flags |= VM_IO | VM_DONTEXPAND;
 662                 vmf = 1;
 663                 break;
 664         }
 665         default:
 666                 ret = -EINVAL;
 667                 break;
 668         }
 669
 670         if ((vma->vm_end - vma->vm_start) != memlen) {
 671                 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
 672                           uctxt->ctxt, fd->subctxt,
 673                           (vma->vm_end - vma->vm_start), memlen);
 674                 ret = -EINVAL;
 675                 goto done;
 676         }
 677
 678         vma->vm_flags = flags;
 679         hfi1_cdbg(PROC,
 680                   "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
 681                     ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
 682                     vma->vm_end - vma->vm_start, vma->vm_flags);
 683         if (vmf) {
 684                 vma->vm_pgoff = PFN_DOWN(memaddr);
 685                 vma->vm_ops = &vm_ops;
 686                 ret = 0;
 687         } else if (mapio) {
 688                 ret = io_remap_pfn_range(vma, vma->vm_start,
 689                                          PFN_DOWN(memaddr),
 690                                          memlen,
 691                                          vma->vm_page_prot);
 692         } else if (memvirt) {
 693                 ret = remap_pfn_range(vma, vma->vm_start,
 694                                       PFN_DOWN(__pa(memvirt)),
 695                                       memlen,
 696                                       vma->vm_page_prot);
 697         } else {
 698                 ret = remap_pfn_range(vma, vma->vm_start,
 699                                       PFN_DOWN(memaddr),
 700                                       memlen,
 701                                       vma->vm_page_prot);
 702         }
 703 done:
 704         return ret;
 705 }
 706
 707 /*
 708  * Local (non-chip) user memory is not mapped right away but as it is
 709  * accessed by the user-level code.
 710  */
 711 static int vma_fault(struct vm_fault *vmf)
 712 {
 713         struct page *page;
 714
 715         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 716         if (!page)
 717                 return VM_FAULT_SIGBUS;
 718
 719         get_page(page);
 720         vmf->page = page;
 721
 722         return 0;
 723 }
 724
 725 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
 726 {
 727         struct hfi1_ctxtdata *uctxt;
 728         unsigned pollflag;
 729
 730         uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
 731         if (!uctxt)
 732                 pollflag = POLLERR;
 733         else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
 734                 pollflag = poll_urgent(fp, pt);
 735         else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
 736                 pollflag = poll_next(fp, pt);
 737         else /* invalid */
 738                 pollflag = POLLERR;
 739
 740         return pollflag;
 741 }
 742
 743 static int hfi1_file_close(struct inode *inode, struct file *fp)
 744 {
 745         struct hfi1_filedata *fdata = fp->private_data;
 746         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 747         struct hfi1_devdata *dd = container_of(inode->i_cdev,
 748                                                struct hfi1_devdata,
 749                                                user_cdev);
 750         unsigned long flags, *ev;
 751
 752         fp->private_data = NULL;
 753
 754         if (!uctxt)
 755                 goto done;
 756
 757         hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
 758         mutex_lock(&hfi1_mutex);
 759
 760         flush_wc();
 761         /* drain user sdma queue */
 762         hfi1_user_sdma_free_queues(fdata);
 763
 764         /* release the cpu */
 765         hfi1_put_proc_affinity(fdata->rec_cpu_num);
 766
 767         /* clean up rcv side */
 768         hfi1_user_exp_rcv_free(fdata);
 769
 770         /*
 771          * Clear any left over, unhandled events so the next process that
 772          * gets this context doesn't get confused.
 773          */
 774         ev = dd->events + ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
 775                            HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
 776         *ev = 0;
 777
 778         if (--uctxt->cnt) {
 779                 uctxt->active_slaves &= ~(1 << fdata->subctxt);
 780                 mutex_unlock(&hfi1_mutex);
 781                 goto done;
 782         }
 783
 784         spin_lock_irqsave(&dd->uctxt_lock, flags);
 785         /*
 786          * Disable receive context and interrupt available, reset all
 787          * RcvCtxtCtrl bits to default values.
 788          */
 789         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
 790                      HFI1_RCVCTRL_TIDFLOW_DIS |
 791                      HFI1_RCVCTRL_INTRAVAIL_DIS |
 792                      HFI1_RCVCTRL_TAILUPD_DIS |
 793                      HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
 794                      HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
 795                      HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
 796         /* Clear the context's J_KEY */
 797         hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
 798         /*
 799          * Reset context integrity checks to default.
 800          * (writes to CSRs probably belong in chip.c)
 801          */
 802         write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
 803                         hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
 804         sc_disable(uctxt->sc);
 805         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 806
 807         dd->rcd[uctxt->ctxt] = NULL;
 808
 809         hfi1_user_exp_rcv_grp_free(uctxt);
 810         hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
 811
 812         uctxt->rcvwait_to = 0;
 813         uctxt->piowait_to = 0;
 814         uctxt->rcvnowait = 0;
 815         uctxt->pionowait = 0;
 816         uctxt->event_flags = 0;
 817
 818         hfi1_stats.sps_ctxts--;
 819         if (++dd->freectxts == dd->num_user_contexts)
 820                 aspm_enable_all(dd);
 821         mutex_unlock(&hfi1_mutex);
 822         hfi1_free_ctxtdata(dd, uctxt);
 823 done:
 824         mmdrop(fdata->mm);
 825         kobject_put(&dd->kobj);
 826
 827         if (atomic_dec_and_test(&dd->user_refcount))
 828                 complete(&dd->user_comp);
 829
 830         kfree(fdata);
 831         return 0;
 832 }
 833
 834 /*
 835  * Convert kernel *virtual* addresses to physical addresses.
 836  * This is used to vmalloc'ed addresses.
 837  */
 838 static u64 kvirt_to_phys(void *addr)
 839 {
 840         struct page *page;
 841         u64 paddr = 0;
 842
 843         page = vmalloc_to_page(addr);
 844         if (page)
 845                 paddr = page_to_pfn(page) << PAGE_SHIFT;
 846
 847         return paddr;
 848 }
 849
 850 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
 851 {
 852         int i_minor, ret = 0;
 853         unsigned int swmajor, swminor;
 854
 855         swmajor = uinfo->userversion >> 16;
 856         if (swmajor != HFI1_USER_SWMAJOR) {
 857                 ret = -ENODEV;
 858                 goto done;
 859         }
 860
 861         swminor = uinfo->userversion & 0xffff;
 862
 863         mutex_lock(&hfi1_mutex);
 864         /* First, lets check if we need to setup a shared context? */
 865         if (uinfo->subctxt_cnt) {
 866                 struct hfi1_filedata *fd = fp->private_data;
 867
 868                 ret = find_shared_ctxt(fd, uinfo);
 869                 if (ret < 0)
 870                         goto done_unlock;
 871                 if (ret) {
 872                         fd->rec_cpu_num =
 873                                 hfi1_get_proc_affinity(fd->uctxt->numa_id);
 874                 }
 875         }
 876
 877         /*
 878          * We execute the following block if we couldn't find a
 879          * shared context or if context sharing is not required.
 880          */
 881         if (!ret) {
 882                 i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
 883                 ret = get_user_context(fp->private_data, uinfo, i_minor);
 884         }
 885 done_unlock:
 886         mutex_unlock(&hfi1_mutex);
 887 done:
 888         return ret;
 889 }
 890
 891 static int get_user_context(struct hfi1_filedata *fd,
 892                             struct hfi1_user_info *uinfo, int devno)
 893 {
 894         struct hfi1_devdata *dd = NULL;
 895         int devmax, npresent, nup;
 896
 897         devmax = hfi1_count_units(&npresent, &nup);
 898         if (!npresent)
 899                 return -ENXIO;
 900
 901         if (!nup)
 902                 return -ENETDOWN;
 903
 904         dd = hfi1_lookup(devno);
 905         if (!dd)
 906                 return -ENODEV;
 907         else if (!dd->freectxts)
 908                 return -EBUSY;
 909
 910         return allocate_ctxt(fd, dd, uinfo);
 911 }
 912
 913 static int find_shared_ctxt(struct hfi1_filedata *fd,
 914                             const struct hfi1_user_info *uinfo)
 915 {
 916         int devmax, ndev, i;
 917         int ret = 0;
 918
 919         devmax = hfi1_count_units(NULL, NULL);
 920
 921         for (ndev = 0; ndev < devmax; ndev++) {
 922                 struct hfi1_devdata *dd = hfi1_lookup(ndev);
 923
 924                 if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
 925                         continue;
 926                 for (i = dd->first_dyn_alloc_ctxt;
 927                      i < dd->num_rcv_contexts; i++) {
 928                         struct hfi1_ctxtdata *uctxt = dd->rcd[i];
 929
 930                         /* Skip ctxts which are not yet open */
 931                         if (!uctxt || !uctxt->cnt)
 932                                 continue;
 933
 934                         /* Skip dynamically allocted kernel contexts */
 935                         if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
 936                                 continue;
 937
 938                         /* Skip ctxt if it doesn't match the requested one */
 939                         if (memcmp(uctxt->uuid, uinfo->uuid,
 940                                    sizeof(uctxt->uuid)) ||
 941                             uctxt->jkey != generate_jkey(current_uid()) ||
 942                             uctxt->subctxt_id != uinfo->subctxt_id ||
 943                             uctxt->subctxt_cnt != uinfo->subctxt_cnt)
 944                                 continue;
 945
 946                         /* Verify the sharing process matches the master */
 947                         if (uctxt->userversion != uinfo->userversion ||
 948                             uctxt->cnt >= uctxt->subctxt_cnt) {
 949                                 ret = -EINVAL;
 950                                 goto done;
 951                         }
 952                         fd->uctxt = uctxt;
 953                         fd->subctxt  = uctxt->cnt++;
 954                         uctxt->active_slaves |= 1 << fd->subctxt;
 955                         ret = 1;
 956                         goto done;
 957                 }
 958         }
 959
 960 done:
 961         return ret;
 962 }
 963
 964 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
 965                          struct hfi1_user_info *uinfo)
 966 {
 967         struct hfi1_ctxtdata *uctxt;
 968         unsigned ctxt;
 969         int ret, numa;
 970
 971         if (dd->flags & HFI1_FROZEN) {
 972                 /*
 973                  * Pick an error that is unique from all other errors
 974                  * that are returned so the user process knows that
 975                  * it tried to allocate while the SPC was frozen.  It
 976                  * it should be able to retry with success in a short
 977                  * while.
 978                  */
 979                 return -EIO;
 980         }
 981
 982         for (ctxt = dd->first_dyn_alloc_ctxt;
 983              ctxt < dd->num_rcv_contexts; ctxt++)
 984                 if (!dd->rcd[ctxt])
 985                         break;
 986
 987         if (ctxt == dd->num_rcv_contexts)
 988                 return -EBUSY;
 989
 990         /*
 991          * If we don't have a NUMA node requested, preference is towards
 992          * device NUMA node.
 993          */
 994         fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
 995         if (fd->rec_cpu_num != -1)
 996                 numa = cpu_to_node(fd->rec_cpu_num);
 997         else
 998                 numa = numa_node_id();
 999         uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
1000         if (!uctxt) {
1001                 dd_dev_err(dd,
1002                            "Unable to allocate ctxtdata memory, failing open\n");
1003                 return -ENOMEM;
1004         }
1005         hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
1006                   uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
1007                   uctxt->numa_id);
1008
1009         /*
1010          * Allocate and enable a PIO send context.
1011          */
1012         uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
1013                              uctxt->dd->node);
1014         if (!uctxt->sc) {
1015                 ret = -ENOMEM;
1016                 goto ctxdata_free;
1017         }
1018         hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
1019                   uctxt->sc->hw_context);
1020         ret = sc_enable(uctxt->sc);
1021         if (ret)
1022                 goto ctxdata_free;
1023
1024         /*
1025          * Setup shared context resources if the user-level has requested
1026          * shared contexts and this is the 'master' process.
1027          * This has to be done here so the rest of the sub-contexts find the
1028          * proper master.
1029          */
1030         if (uinfo->subctxt_cnt && !fd->subctxt) {
1031                 ret = init_subctxts(uctxt, uinfo);
1032                 /*
1033                  * On error, we don't need to disable and de-allocate the
1034                  * send context because it will be done during file close
1035                  */
1036                 if (ret)
1037                         goto ctxdata_free;
1038         }
1039         uctxt->userversion = uinfo->userversion;
1040         uctxt->flags = hfi1_cap_mask; /* save current flag state */
1041         init_waitqueue_head(&uctxt->wait);
1042         strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1043         memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1044         uctxt->jkey = generate_jkey(current_uid());
1045         INIT_LIST_HEAD(&uctxt->sdma_queues);
1046         spin_lock_init(&uctxt->sdma_qlock);
1047         hfi1_stats.sps_ctxts++;
1048         /*
1049          * Disable ASPM when there are open user/PSM contexts to avoid
1050          * issues with ASPM L1 exit latency
1051          */
1052         if (dd->freectxts-- == dd->num_user_contexts)
1053                 aspm_disable_all(dd);
1054         fd->uctxt = uctxt;
1055
1056         return 0;
1057
1058 ctxdata_free:
1059         dd->rcd[ctxt] = NULL;
1060         hfi1_free_ctxtdata(dd, uctxt);
1061         return ret;
1062 }
1063
1064 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1065                          const struct hfi1_user_info *uinfo)
1066 {
1067         unsigned num_subctxts;
1068
1069         num_subctxts = uinfo->subctxt_cnt;
1070         if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1071                 return -EINVAL;
1072
1073         uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1074         uctxt->subctxt_id = uinfo->subctxt_id;
1075         uctxt->active_slaves = 1;
1076         uctxt->redirect_seq_cnt = 1;
1077         set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1078
1079         return 0;
1080 }
1081
1082 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1083 {
1084         int ret = 0;
1085         unsigned num_subctxts = uctxt->subctxt_cnt;
1086
1087         uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1088         if (!uctxt->subctxt_uregbase) {
1089                 ret = -ENOMEM;
1090                 goto bail;
1091         }
1092         /* We can take the size of the RcvHdr Queue from the master */
1093         uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1094                                                   num_subctxts);
1095         if (!uctxt->subctxt_rcvhdr_base) {
1096                 ret = -ENOMEM;
1097                 goto bail_ureg;
1098         }
1099
1100         uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1101                                                 num_subctxts);
1102         if (!uctxt->subctxt_rcvegrbuf) {
1103                 ret = -ENOMEM;
1104                 goto bail_rhdr;
1105         }
1106         goto bail;
1107 bail_rhdr:
1108         vfree(uctxt->subctxt_rcvhdr_base);
1109 bail_ureg:
1110         vfree(uctxt->subctxt_uregbase);
1111         uctxt->subctxt_uregbase = NULL;
1112 bail:
1113         return ret;
1114 }
1115
1116 static int user_init(struct hfi1_filedata *fd)
1117 {
1118         unsigned int rcvctrl_ops = 0;
1119         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1120
1121         /* make sure that the context has already been setup */
1122         if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags))
1123                 return -EFAULT;
1124
1125         /* initialize poll variables... */
1126         uctxt->urgent = 0;
1127         uctxt->urgent_poll = 0;
1128
1129         /*
1130          * Now enable the ctxt for receive.
1131          * For chips that are set to DMA the tail register to memory
1132          * when they change (and when the update bit transitions from
1133          * 0 to 1.  So for those chips, we turn it off and then back on.
1134          * This will (very briefly) affect any other open ctxts, but the
1135          * duration is very short, and therefore isn't an issue.  We
1136          * explicitly set the in-memory tail copy to 0 beforehand, so we
1137          * don't have to wait to be sure the DMA update has happened
1138          * (chip resets head/tail to 0 on transition to enable).
1139          */
1140         if (uctxt->rcvhdrtail_kvaddr)
1141                 clear_rcvhdrtail(uctxt);
1142
1143         /* Setup J_KEY before enabling the context */
1144         hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1145
1146         rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1147         if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1148                 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1149         /*
1150          * Ignore the bit in the flags for now until proper
1151          * support for multiple packet per rcv array entry is
1152          * added.
1153          */
1154         if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1155                 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1156         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1157                 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1158         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1159                 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1160         /*
1161          * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1162          * We can't rely on the correct value to be set from prior
1163          * uses of the chip or ctxt. Therefore, add the rcvctrl op
1164          * for both cases.
1165          */
1166         if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1167                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1168         else
1169                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1170         hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1171
1172         /* Notify any waiting slaves */
1173         if (uctxt->subctxt_cnt) {
1174                 clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1175                 wake_up(&uctxt->wait);
1176         }
1177
1178         return 0;
1179 }
1180
1181 static int get_ctxt_info(struct hfi1_filedata *fd, void __user *ubase,
1182                          __u32 len)
1183 {
1184         struct hfi1_ctxt_info cinfo;
1185         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1186         int ret = 0;
1187
1188         memset(&cinfo, 0, sizeof(cinfo));
1189         cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1190                                 HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1191                         HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1192                         HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1193         /* adjust flag if this fd is not able to cache */
1194         if (!fd->handler)
1195                 cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1196
1197         cinfo.num_active = hfi1_count_active_units();
1198         cinfo.unit = uctxt->dd->unit;
1199         cinfo.ctxt = uctxt->ctxt;
1200         cinfo.subctxt = fd->subctxt;
1201         cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1202                                 uctxt->dd->rcv_entries.group_size) +
1203                 uctxt->expected_count;
1204         cinfo.credits = uctxt->sc->credits;
1205         cinfo.numa_node = uctxt->numa_id;
1206         cinfo.rec_cpu = fd->rec_cpu_num;
1207         cinfo.send_ctxt = uctxt->sc->hw_context;
1208
1209         cinfo.egrtids = uctxt->egrbufs.alloced;
1210         cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1211         cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1212         cinfo.sdma_ring_size = fd->cq->nentries;
1213         cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1214
1215         trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1216         if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1217                 ret = -EFAULT;
1218
1219         return ret;
1220 }
1221
1222 static int setup_ctxt(struct hfi1_filedata *fd)
1223 {
1224         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1225         struct hfi1_devdata *dd = uctxt->dd;
1226         int ret = 0;
1227
1228         /*
1229          * Context should be set up only once, including allocation and
1230          * programming of eager buffers. This is done if context sharing
1231          * is not requested or by the master process.
1232          */
1233         if (!uctxt->subctxt_cnt || !fd->subctxt) {
1234                 ret = hfi1_init_ctxt(uctxt->sc);
1235                 if (ret)
1236                         goto done;
1237
1238                 /* Now allocate the RcvHdr queue and eager buffers. */
1239                 ret = hfi1_create_rcvhdrq(dd, uctxt);
1240                 if (ret)
1241                         goto done;
1242                 ret = hfi1_setup_eagerbufs(uctxt);
1243                 if (ret)
1244                         goto done;
1245                 if (uctxt->subctxt_cnt && !fd->subctxt) {
1246                         ret = setup_subctxt(uctxt);
1247                         if (ret)
1248                                 goto done;
1249                 }
1250         } else {
1251                 ret = wait_event_interruptible(uctxt->wait, !test_bit(
1252                                                HFI1_CTXT_MASTER_UNINIT,
1253                                                &uctxt->event_flags));
1254                 if (ret)
1255                         goto done;
1256         }
1257
1258         ret = hfi1_user_sdma_alloc_queues(uctxt, fd);
1259         if (ret)
1260                 goto done;
1261         /*
1262          * Expected receive has to be setup for all processes (including
1263          * shared contexts). However, it has to be done after the master
1264          * context has been fully configured as it depends on the
1265          * eager/expected split of the RcvArray entries.
1266          * Setting it up here ensures that the subcontexts will be waiting
1267          * (due to the above wait_event_interruptible() until the master
1268          * is setup.
1269          */
1270         ret = hfi1_user_exp_rcv_init(fd);
1271         if (ret)
1272                 goto done;
1273
1274         set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1275 done:
1276         return ret;
1277 }
1278
1279 static int get_base_info(struct hfi1_filedata *fd, void __user *ubase,
1280                          __u32 len)
1281 {
1282         struct hfi1_base_info binfo;
1283         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1284         struct hfi1_devdata *dd = uctxt->dd;
1285         ssize_t sz;
1286         unsigned offset;
1287         int ret = 0;
1288
1289         trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1290
1291         memset(&binfo, 0, sizeof(binfo));
1292         binfo.hw_version = dd->revision;
1293         binfo.sw_version = HFI1_KERN_SWVERSION;
1294         binfo.bthqp = kdeth_qp;
1295         binfo.jkey = uctxt->jkey;
1296         /*
1297          * If more than 64 contexts are enabled the allocated credit
1298          * return will span two or three contiguous pages. Since we only
1299          * map the page containing the context's credit return address,
1300          * we need to calculate the offset in the proper page.
1301          */
1302         offset = ((u64)uctxt->sc->hw_free -
1303                   (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1304         binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1305                                                 fd->subctxt, offset);
1306         binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1307                                             fd->subctxt,
1308                                             uctxt->sc->base_addr);
1309         binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1310                                                 uctxt->ctxt,
1311                                                 fd->subctxt,
1312                                                 uctxt->sc->base_addr);
1313         binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1314                                                fd->subctxt,
1315                                                uctxt->rcvhdrq);
1316         binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1317                                                fd->subctxt,
1318                                                uctxt->egrbufs.rcvtids[0].dma);
1319         binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1320                                                  fd->subctxt, 0);
1321         /*
1322          * user regs are at
1323          * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1324          */
1325         binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1326                                             fd->subctxt, 0);
1327         offset = offset_in_page((((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1328                     HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1329                   sizeof(*dd->events));
1330         binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1331                                               fd->subctxt,
1332                                               offset);
1333         binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1334                                               fd->subctxt,
1335                                               dd->status);
1336         if (HFI1_CAP_IS_USET(DMA_RTAIL))
1337                 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1338                                                        fd->subctxt, 0);
1339         if (uctxt->subctxt_cnt) {
1340                 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1341                                                         uctxt->ctxt,
1342                                                         fd->subctxt, 0);
1343                 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1344                                                          uctxt->ctxt,
1345                                                          fd->subctxt, 0);
1346                 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1347                                                          uctxt->ctxt,
1348                                                          fd->subctxt, 0);
1349         }
1350         sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1351         if (copy_to_user(ubase, &binfo, sz))
1352                 ret = -EFAULT;
1353         return ret;
1354 }
1355
1356 static unsigned int poll_urgent(struct file *fp,
1357                                 struct poll_table_struct *pt)
1358 {
1359         struct hfi1_filedata *fd = fp->private_data;
1360         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1361         struct hfi1_devdata *dd = uctxt->dd;
1362         unsigned pollflag;
1363
1364         poll_wait(fp, &uctxt->wait, pt);
1365
1366         spin_lock_irq(&dd->uctxt_lock);
1367         if (uctxt->urgent != uctxt->urgent_poll) {
1368                 pollflag = POLLIN | POLLRDNORM;
1369                 uctxt->urgent_poll = uctxt->urgent;
1370         } else {
1371                 pollflag = 0;
1372                 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1373         }
1374         spin_unlock_irq(&dd->uctxt_lock);
1375
1376         return pollflag;
1377 }
1378
1379 static unsigned int poll_next(struct file *fp,
1380                               struct poll_table_struct *pt)
1381 {
1382         struct hfi1_filedata *fd = fp->private_data;
1383         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1384         struct hfi1_devdata *dd = uctxt->dd;
1385         unsigned pollflag;
1386
1387         poll_wait(fp, &uctxt->wait, pt);
1388
1389         spin_lock_irq(&dd->uctxt_lock);
1390         if (hdrqempty(uctxt)) {
1391                 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1392                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1393                 pollflag = 0;
1394         } else {
1395                 pollflag = POLLIN | POLLRDNORM;
1396         }
1397         spin_unlock_irq(&dd->uctxt_lock);
1398
1399         return pollflag;
1400 }
1401
1402 /*
1403  * Find all user contexts in use, and set the specified bit in their
1404  * event mask.
1405  * See also find_ctxt() for a similar use, that is specific to send buffers.
1406  */
1407 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1408 {
1409         struct hfi1_ctxtdata *uctxt;
1410         struct hfi1_devdata *dd = ppd->dd;
1411         unsigned ctxt;
1412         int ret = 0;
1413         unsigned long flags;
1414
1415         if (!dd->events) {
1416                 ret = -EINVAL;
1417                 goto done;
1418         }
1419
1420         spin_lock_irqsave(&dd->uctxt_lock, flags);
1421         for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
1422              ctxt++) {
1423                 uctxt = dd->rcd[ctxt];
1424                 if (uctxt) {
1425                         unsigned long *evs = dd->events +
1426                                 (uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1427                                 HFI1_MAX_SHARED_CTXTS;
1428                         int i;
1429                         /*
1430                          * subctxt_cnt is 0 if not shared, so do base
1431                          * separately, first, then remaining subctxt, if any
1432                          */
1433                         set_bit(evtbit, evs);
1434                         for (i = 1; i < uctxt->subctxt_cnt; i++)
1435                                 set_bit(evtbit, evs + i);
1436                 }
1437         }
1438         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1439 done:
1440         return ret;
1441 }
1442
1443 /**
1444  * manage_rcvq - manage a context's receive queue
1445  * @uctxt: the context
1446  * @subctxt: the sub-context
1447  * @start_stop: action to carry out
1448  *
1449  * start_stop == 0 disables receive on the context, for use in queue
1450  * overflow conditions.  start_stop==1 re-enables, to be used to
1451  * re-init the software copy of the head register
1452  */
1453 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1454                        int start_stop)
1455 {
1456         struct hfi1_devdata *dd = uctxt->dd;
1457         unsigned int rcvctrl_op;
1458
1459         if (subctxt)
1460                 goto bail;
1461         /* atomically clear receive enable ctxt. */
1462         if (start_stop) {
1463                 /*
1464                  * On enable, force in-memory copy of the tail register to
1465                  * 0, so that protocol code doesn't have to worry about
1466                  * whether or not the chip has yet updated the in-memory
1467                  * copy or not on return from the system call. The chip
1468                  * always resets it's tail register back to 0 on a
1469                  * transition from disabled to enabled.
1470                  */
1471                 if (uctxt->rcvhdrtail_kvaddr)
1472                         clear_rcvhdrtail(uctxt);
1473                 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1474         } else {
1475                 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1476         }
1477         hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1478         /* always; new head should be equal to new tail; see above */
1479 bail:
1480         return 0;
1481 }
1482
1483 /*
1484  * clear the event notifier events for this context.
1485  * User process then performs actions appropriate to bit having been
1486  * set, if desired, and checks again in future.
1487  */
1488 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1489                           unsigned long events)
1490 {
1491         int i;
1492         struct hfi1_devdata *dd = uctxt->dd;
1493         unsigned long *evs;
1494
1495         if (!dd->events)
1496                 return 0;
1497
1498         evs = dd->events + ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1499                             HFI1_MAX_SHARED_CTXTS) + subctxt;
1500
1501         for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1502                 if (!test_bit(i, &events))
1503                         continue;
1504                 clear_bit(i, evs);
1505         }
1506         return 0;
1507 }
1508
1509 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1510                          u16 pkey)
1511 {
1512         int ret = -ENOENT, i, intable = 0;
1513         struct hfi1_pportdata *ppd = uctxt->ppd;
1514         struct hfi1_devdata *dd = uctxt->dd;
1515
1516         if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1517                 ret = -EINVAL;
1518                 goto done;
1519         }
1520
1521         for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1522                 if (pkey == ppd->pkeys[i]) {
1523                         intable = 1;
1524                         break;
1525                 }
1526
1527         if (intable)
1528                 ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1529 done:
1530         return ret;
1531 }
1532
1533 static void user_remove(struct hfi1_devdata *dd)
1534 {
1535
1536         hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1537 }
1538
1539 static int user_add(struct hfi1_devdata *dd)
1540 {
1541         char name[10];
1542         int ret;
1543
1544         snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1545         ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1546                              &dd->user_cdev, &dd->user_device,
1547                              true, &dd->kobj);
1548         if (ret)
1549                 user_remove(dd);
1550
1551         return ret;
1552 }
1553
1554 /*
1555  * Create per-unit files in /dev
1556  */
1557 int hfi1_device_create(struct hfi1_devdata *dd)
1558 {
1559         return user_add(dd);
1560 }
1561
1562 /*
1563  * Remove per-unit files in /dev
1564  * void, core kernel returns no errors for this stuff
1565  */
1566 void hfi1_device_remove(struct hfi1_devdata *dd)
1567 {
1568         user_remove(dd);
1569 }