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 #include <linux/bitmap.h>
  53
  54 #include <rdma/ib.h>
  55
  56 #include "hfi.h"
  57 #include "pio.h"
  58 #include "device.h"
  59 #include "common.h"
  60 #include "trace.h"
  61 #include "mmu_rb.h"
  62 #include "user_sdma.h"
  63 #include "user_exp_rcv.h"
  64 #include "aspm.h"
  65
  66 #undef pr_fmt
  67 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
  68
  69 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
  70
  71 /*
  72  * File operation functions
  73  */
  74 static int hfi1_file_open(struct inode *inode, struct file *fp);
  75 static int hfi1_file_close(struct inode *inode, struct file *fp);
  76 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from);
  77 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt);
  78 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma);
  79
  80 static u64 kvirt_to_phys(void *addr);
  81 static int assign_ctxt(struct hfi1_filedata *fd, struct hfi1_user_info *uinfo);
  82 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
  83                           const struct hfi1_user_info *uinfo);
  84 static int init_user_ctxt(struct hfi1_filedata *fd,
  85                           struct hfi1_ctxtdata *uctxt);
  86 static void user_init(struct hfi1_ctxtdata *uctxt);
  87 static int get_ctxt_info(struct hfi1_filedata *fd, void __user *ubase,
  88                          __u32 len);
  89 static int get_base_info(struct hfi1_filedata *fd, void __user *ubase,
  90                          __u32 len);
  91 static int setup_base_ctxt(struct hfi1_filedata *fd,
  92                            struct hfi1_ctxtdata *uctxt);
  93 static int setup_subctxt(struct hfi1_ctxtdata *uctxt);
  94
  95 static int find_sub_ctxt(struct hfi1_filedata *fd,
  96                          const struct hfi1_user_info *uinfo);
  97 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
  98                          struct hfi1_user_info *uinfo,
  99                          struct hfi1_ctxtdata **cd);
 100 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt);
 101 static unsigned int poll_urgent(struct file *fp, struct poll_table_struct *pt);
 102 static unsigned int poll_next(struct file *fp, struct poll_table_struct *pt);
 103 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
 104                           unsigned long events);
 105 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, u16 subctxt, u16 pkey);
 106 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
 107                        int start_stop);
 108 static int vma_fault(struct vm_fault *vmf);
 109 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
 110                             unsigned long arg);
 111
 112 static const struct file_operations hfi1_file_ops = {
 113         .owner = THIS_MODULE,
 114         .write_iter = hfi1_write_iter,
 115         .open = hfi1_file_open,
 116         .release = hfi1_file_close,
 117         .unlocked_ioctl = hfi1_file_ioctl,
 118         .poll = hfi1_poll,
 119         .mmap = hfi1_file_mmap,
 120         .llseek = noop_llseek,
 121 };
 122
 123 static const struct vm_operations_struct vm_ops = {
 124         .fault = vma_fault,
 125 };
 126
 127 /*
 128  * Types of memories mapped into user processes' space
 129  */
 130 enum mmap_types {
 131         PIO_BUFS = 1,
 132         PIO_BUFS_SOP,
 133         PIO_CRED,
 134         RCV_HDRQ,
 135         RCV_EGRBUF,
 136         UREGS,
 137         EVENTS,
 138         STATUS,
 139         RTAIL,
 140         SUBCTXT_UREGS,
 141         SUBCTXT_RCV_HDRQ,
 142         SUBCTXT_EGRBUF,
 143         SDMA_COMP
 144 };
 145
 146 /*
 147  * Masks and offsets defining the mmap tokens
 148  */
 149 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
 150 #define HFI1_MMAP_OFFSET_SHIFT  0
 151 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
 152 #define HFI1_MMAP_SUBCTXT_SHIFT 12
 153 #define HFI1_MMAP_CTXT_MASK     0xffULL
 154 #define HFI1_MMAP_CTXT_SHIFT    16
 155 #define HFI1_MMAP_TYPE_MASK     0xfULL
 156 #define HFI1_MMAP_TYPE_SHIFT    24
 157 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
 158 #define HFI1_MMAP_MAGIC_SHIFT   32
 159
 160 #define HFI1_MMAP_MAGIC         0xdabbad00
 161
 162 #define HFI1_MMAP_TOKEN_SET(field, val) \
 163         (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
 164 #define HFI1_MMAP_TOKEN_GET(field, token) \
 165         (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
 166 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
 167         (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
 168         HFI1_MMAP_TOKEN_SET(TYPE, type) | \
 169         HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
 170         HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
 171         HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
 172
 173 #define dbg(fmt, ...)                           \
 174         pr_info(fmt, ##__VA_ARGS__)
 175
 176 static inline int is_valid_mmap(u64 token)
 177 {
 178         return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
 179 }
 180
 181 static int hfi1_file_open(struct inode *inode, struct file *fp)
 182 {
 183         struct hfi1_filedata *fd;
 184         struct hfi1_devdata *dd = container_of(inode->i_cdev,
 185                                                struct hfi1_devdata,
 186                                                user_cdev);
 187
 188         if (!((dd->flags & HFI1_PRESENT) && dd->kregbase1))
 189                 return -EINVAL;
 190
 191         if (!atomic_inc_not_zero(&dd->user_refcount))
 192                 return -ENXIO;
 193
 194         /* The real work is performed later in assign_ctxt() */
 195
 196         fd = kzalloc(sizeof(*fd), GFP_KERNEL);
 197
 198         if (!fd || init_srcu_struct(&fd->pq_srcu))
 199                 goto nomem;
 200         spin_lock_init(&fd->pq_rcu_lock);
 201         spin_lock_init(&fd->tid_lock);
 202         spin_lock_init(&fd->invalid_lock);
 203         fd->rec_cpu_num = -1; /* no cpu affinity by default */
 204         fd->mm = current->mm;
 205         mmgrab(fd->mm);
 206         fd->dd = dd;
 207         kobject_get(&fd->dd->kobj);
 208         fp->private_data = fd;
 209         return 0;
 210 nomem:
 211         kfree(fd);
 212         fp->private_data = NULL;
 213         if (atomic_dec_and_test(&dd->user_refcount))
 214                 complete(&dd->user_comp);
 215         return -ENOMEM;
 216 }
 217
 218 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
 219                             unsigned long arg)
 220 {
 221         struct hfi1_filedata *fd = fp->private_data;
 222         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 223         struct hfi1_user_info uinfo;
 224         struct hfi1_tid_info tinfo;
 225         int ret = 0;
 226         unsigned long addr;
 227         int uval = 0;
 228         unsigned long ul_uval = 0;
 229         u16 uval16 = 0;
 230
 231         hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
 232         if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
 233             cmd != HFI1_IOCTL_GET_VERS &&
 234             !uctxt)
 235                 return -EINVAL;
 236
 237         switch (cmd) {
 238         case HFI1_IOCTL_ASSIGN_CTXT:
 239                 if (uctxt)
 240                         return -EINVAL;
 241
 242                 if (copy_from_user(&uinfo,
 243                                    (struct hfi1_user_info __user *)arg,
 244                                    sizeof(uinfo)))
 245                         return -EFAULT;
 246
 247                 ret = assign_ctxt(fd, &uinfo);
 248                 break;
 249         case HFI1_IOCTL_CTXT_INFO:
 250                 ret = get_ctxt_info(fd, (void __user *)(unsigned long)arg,
 251                                     sizeof(struct hfi1_ctxt_info));
 252                 break;
 253         case HFI1_IOCTL_USER_INFO:
 254                 ret = get_base_info(fd, (void __user *)(unsigned long)arg,
 255                                     sizeof(struct hfi1_base_info));
 256                 break;
 257         case HFI1_IOCTL_CREDIT_UPD:
 258                 if (uctxt)
 259                         sc_return_credits(uctxt->sc);
 260                 break;
 261
 262         case HFI1_IOCTL_TID_UPDATE:
 263                 if (copy_from_user(&tinfo,
 264                                    (struct hfi11_tid_info __user *)arg,
 265                                    sizeof(tinfo)))
 266                         return -EFAULT;
 267
 268                 ret = hfi1_user_exp_rcv_setup(fd, &tinfo);
 269                 if (!ret) {
 270                         /*
 271                          * Copy the number of tidlist entries we used
 272                          * and the length of the buffer we registered.
 273                          */
 274                         addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
 275                         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 276                                          sizeof(tinfo.tidcnt)))
 277                                 return -EFAULT;
 278
 279                         addr = arg + offsetof(struct hfi1_tid_info, length);
 280                         if (copy_to_user((void __user *)addr, &tinfo.length,
 281                                          sizeof(tinfo.length)))
 282                                 ret = -EFAULT;
 283                 }
 284                 break;
 285
 286         case HFI1_IOCTL_TID_FREE:
 287                 if (copy_from_user(&tinfo,
 288                                    (struct hfi11_tid_info __user *)arg,
 289                                    sizeof(tinfo)))
 290                         return -EFAULT;
 291
 292                 ret = hfi1_user_exp_rcv_clear(fd, &tinfo);
 293                 if (ret)
 294                         break;
 295                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
 296                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 297                                  sizeof(tinfo.tidcnt)))
 298                         ret = -EFAULT;
 299                 break;
 300
 301         case HFI1_IOCTL_TID_INVAL_READ:
 302                 if (copy_from_user(&tinfo,
 303                                    (struct hfi11_tid_info __user *)arg,
 304                                    sizeof(tinfo)))
 305                         return -EFAULT;
 306
 307                 ret = hfi1_user_exp_rcv_invalid(fd, &tinfo);
 308                 if (ret)
 309                         break;
 310                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
 311                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 312                                  sizeof(tinfo.tidcnt)))
 313                         ret = -EFAULT;
 314                 break;
 315
 316         case HFI1_IOCTL_RECV_CTRL:
 317                 ret = get_user(uval, (int __user *)arg);
 318                 if (ret != 0)
 319                         return -EFAULT;
 320                 ret = manage_rcvq(uctxt, fd->subctxt, uval);
 321                 break;
 322
 323         case HFI1_IOCTL_POLL_TYPE:
 324                 ret = get_user(uval, (int __user *)arg);
 325                 if (ret != 0)
 326                         return -EFAULT;
 327                 uctxt->poll_type = (typeof(uctxt->poll_type))uval;
 328                 break;
 329
 330         case HFI1_IOCTL_ACK_EVENT:
 331                 ret = get_user(ul_uval, (unsigned long __user *)arg);
 332                 if (ret != 0)
 333                         return -EFAULT;
 334                 ret = user_event_ack(uctxt, fd->subctxt, ul_uval);
 335                 break;
 336
 337         case HFI1_IOCTL_SET_PKEY:
 338                 ret = get_user(uval16, (u16 __user *)arg);
 339                 if (ret != 0)
 340                         return -EFAULT;
 341                 if (HFI1_CAP_IS_USET(PKEY_CHECK))
 342                         ret = set_ctxt_pkey(uctxt, fd->subctxt, uval16);
 343                 else
 344                         return -EPERM;
 345                 break;
 346
 347         case HFI1_IOCTL_CTXT_RESET: {
 348                 struct send_context *sc;
 349                 struct hfi1_devdata *dd;
 350
 351                 if (!uctxt || !uctxt->dd || !uctxt->sc)
 352                         return -EINVAL;
 353
 354                 /*
 355                  * There is no protection here. User level has to
 356                  * guarantee that no one will be writing to the send
 357                  * context while it is being re-initialized.
 358                  * If user level breaks that guarantee, it will break
 359                  * it's own context and no one else's.
 360                  */
 361                 dd = uctxt->dd;
 362                 sc = uctxt->sc;
 363                 /*
 364                  * Wait until the interrupt handler has marked the
 365                  * context as halted or frozen. Report error if we time
 366                  * out.
 367                  */
 368                 wait_event_interruptible_timeout(
 369                         sc->halt_wait, (sc->flags & SCF_HALTED),
 370                         msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 371                 if (!(sc->flags & SCF_HALTED))
 372                         return -ENOLCK;
 373
 374                 /*
 375                  * If the send context was halted due to a Freeze,
 376                  * wait until the device has been "unfrozen" before
 377                  * resetting the context.
 378                  */
 379                 if (sc->flags & SCF_FROZEN) {
 380                         wait_event_interruptible_timeout(
 381                                 dd->event_queue,
 382                                 !(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
 383                                 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 384                         if (dd->flags & HFI1_FROZEN)
 385                                 return -ENOLCK;
 386
 387                         if (dd->flags & HFI1_FORCED_FREEZE)
 388                                 /*
 389                                  * Don't allow context reset if we are into
 390                                  * forced freeze
 391                                  */
 392                                 return -ENODEV;
 393
 394                         sc_disable(sc);
 395                         ret = sc_enable(sc);
 396                         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB, uctxt);
 397                 } else {
 398                         ret = sc_restart(sc);
 399                 }
 400                 if (!ret)
 401                         sc_return_credits(sc);
 402                 break;
 403         }
 404
 405         case HFI1_IOCTL_GET_VERS:
 406                 uval = HFI1_USER_SWVERSION;
 407                 if (put_user(uval, (int __user *)arg))
 408                         return -EFAULT;
 409                 break;
 410
 411         default:
 412                 return -EINVAL;
 413         }
 414
 415         return ret;
 416 }
 417
 418 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
 419 {
 420         struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
 421         struct hfi1_user_sdma_pkt_q *pq;
 422         struct hfi1_user_sdma_comp_q *cq = fd->cq;
 423         int done = 0, reqs = 0;
 424         unsigned long dim = from->nr_segs;
 425         int idx;
 426
 427         idx = srcu_read_lock(&fd->pq_srcu);
 428         pq = srcu_dereference(fd->pq, &fd->pq_srcu);
 429         if (!cq || !pq) {
 430                 srcu_read_unlock(&fd->pq_srcu, idx);
 431                 return -EIO;
 432         }
 433
 434         if (!iter_is_iovec(from) || !dim) {
 435                 srcu_read_unlock(&fd->pq_srcu, idx);
 436                 return -EINVAL;
 437         }
 438
 439         trace_hfi1_sdma_request(fd->dd, fd->uctxt->ctxt, fd->subctxt, dim);
 440
 441         if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
 442                 srcu_read_unlock(&fd->pq_srcu, idx);
 443                 return -ENOSPC;
 444         }
 445
 446         while (dim) {
 447                 int ret;
 448                 unsigned long count = 0;
 449
 450                 ret = hfi1_user_sdma_process_request(
 451                         fd, (struct iovec *)(from->iov + done),
 452                         dim, &count);
 453                 if (ret) {
 454                         reqs = ret;
 455                         break;
 456                 }
 457                 dim -= count;
 458                 done += count;
 459                 reqs++;
 460         }
 461
 462         srcu_read_unlock(&fd->pq_srcu, idx);
 463         return reqs;
 464 }
 465
 466 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
 467 {
 468         struct hfi1_filedata *fd = fp->private_data;
 469         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 470         struct hfi1_devdata *dd;
 471         unsigned long flags;
 472         u64 token = vma->vm_pgoff << PAGE_SHIFT,
 473                 memaddr = 0;
 474         void *memvirt = NULL;
 475         u8 subctxt, mapio = 0, vmf = 0, type;
 476         ssize_t memlen = 0;
 477         int ret = 0;
 478         u16 ctxt;
 479
 480         if (!is_valid_mmap(token) || !uctxt ||
 481             !(vma->vm_flags & VM_SHARED)) {
 482                 ret = -EINVAL;
 483                 goto done;
 484         }
 485         dd = uctxt->dd;
 486         ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
 487         subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
 488         type = HFI1_MMAP_TOKEN_GET(TYPE, token);
 489         if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
 490                 ret = -EINVAL;
 491                 goto done;
 492         }
 493
 494         flags = vma->vm_flags;
 495
 496         switch (type) {
 497         case PIO_BUFS:
 498         case PIO_BUFS_SOP:
 499                 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
 500                                 /* chip pio base */
 501                            (uctxt->sc->hw_context * BIT(16))) +
 502                                 /* 64K PIO space / ctxt */
 503                         (type == PIO_BUFS_SOP ?
 504                                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
 505                 /*
 506                  * Map only the amount allocated to the context, not the
 507                  * entire available context's PIO space.
 508                  */
 509                 memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
 510                 flags &= ~VM_MAYREAD;
 511                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 512                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 513                 mapio = 1;
 514                 break;
 515         case PIO_CRED:
 516                 if (flags & VM_WRITE) {
 517                         ret = -EPERM;
 518                         goto done;
 519                 }
 520                 /*
 521                  * The credit return location for this context could be on the
 522                  * second or third page allocated for credit returns (if number
 523                  * of enabled contexts > 64 and 128 respectively).
 524                  */
 525                 memvirt = dd->cr_base[uctxt->numa_id].va;
 526                 memaddr = virt_to_phys(memvirt) +
 527                         (((u64)uctxt->sc->hw_free -
 528                           (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
 529                 memlen = PAGE_SIZE;
 530                 flags &= ~VM_MAYWRITE;
 531                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 532                 /*
 533                  * The driver has already allocated memory for credit
 534                  * returns and programmed it into the chip. Has that
 535                  * memory been flagged as non-cached?
 536                  */
 537                 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
 538                 mapio = 1;
 539                 break;
 540         case RCV_HDRQ:
 541                 memlen = uctxt->rcvhdrq_size;
 542                 memvirt = uctxt->rcvhdrq;
 543                 break;
 544         case RCV_EGRBUF: {
 545                 unsigned long addr;
 546                 int i;
 547                 /*
 548                  * The RcvEgr buffer need to be handled differently
 549                  * as multiple non-contiguous pages need to be mapped
 550                  * into the user process.
 551                  */
 552                 memlen = uctxt->egrbufs.size;
 553                 if ((vma->vm_end - vma->vm_start) != memlen) {
 554                         dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
 555                                    (vma->vm_end - vma->vm_start), memlen);
 556                         ret = -EINVAL;
 557                         goto done;
 558                 }
 559                 if (vma->vm_flags & VM_WRITE) {
 560                         ret = -EPERM;
 561                         goto done;
 562                 }
 563                 vma->vm_flags &= ~VM_MAYWRITE;
 564                 addr = vma->vm_start;
 565                 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
 566                         memlen = uctxt->egrbufs.buffers[i].len;
 567                         memvirt = uctxt->egrbufs.buffers[i].addr;
 568                         ret = remap_pfn_range(
 569                                 vma, addr,
 570                                 /*
 571                                  * virt_to_pfn() does the same, but
 572                                  * it's not available on x86_64
 573                                  * when CONFIG_MMU is enabled.
 574                                  */
 575                                 PFN_DOWN(__pa(memvirt)),
 576                                 memlen,
 577                                 vma->vm_page_prot);
 578                         if (ret < 0)
 579                                 goto done;
 580                         addr += memlen;
 581                 }
 582                 ret = 0;
 583                 goto done;
 584         }
 585         case UREGS:
 586                 /*
 587                  * Map only the page that contains this context's user
 588                  * registers.
 589                  */
 590                 memaddr = (unsigned long)
 591                         (dd->physaddr + RXE_PER_CONTEXT_USER)
 592                         + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
 593                 /*
 594                  * TidFlow table is on the same page as the rest of the
 595                  * user registers.
 596                  */
 597                 memlen = PAGE_SIZE;
 598                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 599                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 600                 mapio = 1;
 601                 break;
 602         case EVENTS:
 603                 /*
 604                  * Use the page where this context's flags are. User level
 605                  * knows where it's own bitmap is within the page.
 606                  */
 607                 memaddr = (unsigned long)(dd->events +
 608                                   ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
 609                                    HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
 610                 memlen = PAGE_SIZE;
 611                 /*
 612                  * v3.7 removes VM_RESERVED but the effect is kept by
 613                  * using VM_IO.
 614                  */
 615                 flags |= VM_IO | VM_DONTEXPAND;
 616                 vmf = 1;
 617                 break;
 618         case STATUS:
 619                 if (flags & VM_WRITE) {
 620                         ret = -EPERM;
 621                         goto done;
 622                 }
 623                 memaddr = kvirt_to_phys((void *)dd->status);
 624                 memlen = PAGE_SIZE;
 625                 flags |= VM_IO | VM_DONTEXPAND;
 626                 break;
 627         case RTAIL:
 628                 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
 629                         /*
 630                          * If the memory allocation failed, the context alloc
 631                          * also would have failed, so we would never get here
 632                          */
 633                         ret = -EINVAL;
 634                         goto done;
 635                 }
 636                 if ((flags & VM_WRITE) || !uctxt->rcvhdrtail_kvaddr) {
 637                         ret = -EPERM;
 638                         goto done;
 639                 }
 640                 memlen = PAGE_SIZE;
 641                 memvirt = (void *)uctxt->rcvhdrtail_kvaddr;
 642                 flags &= ~VM_MAYWRITE;
 643                 break;
 644         case SUBCTXT_UREGS:
 645                 memaddr = (u64)uctxt->subctxt_uregbase;
 646                 memlen = PAGE_SIZE;
 647                 flags |= VM_IO | VM_DONTEXPAND;
 648                 vmf = 1;
 649                 break;
 650         case SUBCTXT_RCV_HDRQ:
 651                 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
 652                 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
 653                 flags |= VM_IO | VM_DONTEXPAND;
 654                 vmf = 1;
 655                 break;
 656         case SUBCTXT_EGRBUF:
 657                 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
 658                 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
 659                 flags |= VM_IO | VM_DONTEXPAND;
 660                 flags &= ~VM_MAYWRITE;
 661                 vmf = 1;
 662                 break;
 663         case SDMA_COMP: {
 664                 struct hfi1_user_sdma_comp_q *cq = fd->cq;
 665
 666                 if (!cq) {
 667                         ret = -EFAULT;
 668                         goto done;
 669                 }
 670                 memaddr = (u64)cq->comps;
 671                 memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
 672                 flags |= VM_IO | VM_DONTEXPAND;
 673                 vmf = 1;
 674                 break;
 675         }
 676         default:
 677                 ret = -EINVAL;
 678                 break;
 679         }
 680
 681         if ((vma->vm_end - vma->vm_start) != memlen) {
 682                 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
 683                           uctxt->ctxt, fd->subctxt,
 684                           (vma->vm_end - vma->vm_start), memlen);
 685                 ret = -EINVAL;
 686                 goto done;
 687         }
 688
 689         vma->vm_flags = flags;
 690         hfi1_cdbg(PROC,
 691                   "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
 692                     ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
 693                     vma->vm_end - vma->vm_start, vma->vm_flags);
 694         if (vmf) {
 695                 vma->vm_pgoff = PFN_DOWN(memaddr);
 696                 vma->vm_ops = &vm_ops;
 697                 ret = 0;
 698         } else if (mapio) {
 699                 ret = io_remap_pfn_range(vma, vma->vm_start,
 700                                          PFN_DOWN(memaddr),
 701                                          memlen,
 702                                          vma->vm_page_prot);
 703         } else if (memvirt) {
 704                 ret = remap_pfn_range(vma, vma->vm_start,
 705                                       PFN_DOWN(__pa(memvirt)),
 706                                       memlen,
 707                                       vma->vm_page_prot);
 708         } else {
 709                 ret = remap_pfn_range(vma, vma->vm_start,
 710                                       PFN_DOWN(memaddr),
 711                                       memlen,
 712                                       vma->vm_page_prot);
 713         }
 714 done:
 715         return ret;
 716 }
 717
 718 /*
 719  * Local (non-chip) user memory is not mapped right away but as it is
 720  * accessed by the user-level code.
 721  */
 722 static int vma_fault(struct vm_fault *vmf)
 723 {
 724         struct page *page;
 725
 726         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 727         if (!page)
 728                 return VM_FAULT_SIGBUS;
 729
 730         get_page(page);
 731         vmf->page = page;
 732
 733         return 0;
 734 }
 735
 736 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
 737 {
 738         struct hfi1_ctxtdata *uctxt;
 739         unsigned pollflag;
 740
 741         uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
 742         if (!uctxt)
 743                 pollflag = POLLERR;
 744         else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
 745                 pollflag = poll_urgent(fp, pt);
 746         else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
 747                 pollflag = poll_next(fp, pt);
 748         else /* invalid */
 749                 pollflag = POLLERR;
 750
 751         return pollflag;
 752 }
 753
 754 static int hfi1_file_close(struct inode *inode, struct file *fp)
 755 {
 756         struct hfi1_filedata *fdata = fp->private_data;
 757         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 758         struct hfi1_devdata *dd = container_of(inode->i_cdev,
 759                                                struct hfi1_devdata,
 760                                                user_cdev);
 761         unsigned long flags, *ev;
 762
 763         fp->private_data = NULL;
 764
 765         if (!uctxt)
 766                 goto done;
 767
 768         hfi1_cdbg(PROC, "closing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
 769
 770         flush_wc();
 771         /* drain user sdma queue */
 772         hfi1_user_sdma_free_queues(fdata, uctxt);
 773
 774         /* release the cpu */
 775         hfi1_put_proc_affinity(fdata->rec_cpu_num);
 776
 777         /* clean up rcv side */
 778         hfi1_user_exp_rcv_free(fdata);
 779
 780         /*
 781          * fdata->uctxt is used in the above cleanup.  It is not ready to be
 782          * removed until here.
 783          */
 784         fdata->uctxt = NULL;
 785         hfi1_rcd_put(uctxt);
 786
 787         /*
 788          * Clear any left over, unhandled events so the next process that
 789          * gets this context doesn't get confused.
 790          */
 791         ev = dd->events + ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
 792                            HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
 793         *ev = 0;
 794
 795         spin_lock_irqsave(&dd->uctxt_lock, flags);
 796         __clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
 797         if (!bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
 798                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 799                 goto done;
 800         }
 801         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 802
 803         /*
 804          * Disable receive context and interrupt available, reset all
 805          * RcvCtxtCtrl bits to default values.
 806          */
 807         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
 808                      HFI1_RCVCTRL_TIDFLOW_DIS |
 809                      HFI1_RCVCTRL_INTRAVAIL_DIS |
 810                      HFI1_RCVCTRL_TAILUPD_DIS |
 811                      HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
 812                      HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
 813                      HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt);
 814         /* Clear the context's J_KEY */
 815         hfi1_clear_ctxt_jkey(dd, uctxt);
 816         /*
 817          * If a send context is allocated, reset context integrity
 818          * checks to default and disable the send context.
 819          */
 820         if (uctxt->sc) {
 821                 sc_disable(uctxt->sc);
 822                 set_pio_integrity(uctxt->sc);
 823         }
 824
 825         hfi1_free_ctxt_rcv_groups(uctxt);
 826         hfi1_clear_ctxt_pkey(dd, uctxt);
 827
 828         uctxt->event_flags = 0;
 829
 830         deallocate_ctxt(uctxt);
 831 done:
 832         mmdrop(fdata->mm);
 833         kobject_put(&dd->kobj);
 834
 835         if (atomic_dec_and_test(&dd->user_refcount))
 836                 complete(&dd->user_comp);
 837
 838         cleanup_srcu_struct(&fdata->pq_srcu);
 839         kfree(fdata);
 840         return 0;
 841 }
 842
 843 /*
 844  * Convert kernel *virtual* addresses to physical addresses.
 845  * This is used to vmalloc'ed addresses.
 846  */
 847 static u64 kvirt_to_phys(void *addr)
 848 {
 849         struct page *page;
 850         u64 paddr = 0;
 851
 852         page = vmalloc_to_page(addr);
 853         if (page)
 854                 paddr = page_to_pfn(page) << PAGE_SHIFT;
 855
 856         return paddr;
 857 }
 858
 859 /**
 860  * complete_subctxt
 861  * @fd: valid filedata pointer
 862  *
 863  * Sub-context info can only be set up after the base context
 864  * has been completed.  This is indicated by the clearing of the
 865  * HFI1_CTXT_BASE_UINIT bit.
 866  *
 867  * Wait for the bit to be cleared, and then complete the subcontext
 868  * initialization.
 869  *
 870  */
 871 static int complete_subctxt(struct hfi1_filedata *fd)
 872 {
 873         int ret;
 874         unsigned long flags;
 875
 876         /*
 877          * sub-context info can only be set up after the base context
 878          * has been completed.
 879          */
 880         ret = wait_event_interruptible(
 881                 fd->uctxt->wait,
 882                 !test_bit(HFI1_CTXT_BASE_UNINIT, &fd->uctxt->event_flags));
 883
 884         if (test_bit(HFI1_CTXT_BASE_FAILED, &fd->uctxt->event_flags))
 885                 ret = -ENOMEM;
 886
 887         /* Finish the sub-context init */
 888         if (!ret) {
 889                 fd->rec_cpu_num = hfi1_get_proc_affinity(fd->uctxt->numa_id);
 890                 ret = init_user_ctxt(fd, fd->uctxt);
 891         }
 892
 893         if (ret) {
 894                 spin_lock_irqsave(&fd->dd->uctxt_lock, flags);
 895                 __clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
 896                 spin_unlock_irqrestore(&fd->dd->uctxt_lock, flags);
 897                 hfi1_rcd_put(fd->uctxt);
 898                 fd->uctxt = NULL;
 899         }
 900
 901         return ret;
 902 }
 903
 904 static int assign_ctxt(struct hfi1_filedata *fd, struct hfi1_user_info *uinfo)
 905 {
 906         int ret;
 907         unsigned int swmajor, swminor;
 908         struct hfi1_ctxtdata *uctxt = NULL;
 909
 910         swmajor = uinfo->userversion >> 16;
 911         if (swmajor != HFI1_USER_SWMAJOR)
 912                 return -ENODEV;
 913
 914         if (uinfo->subctxt_cnt > HFI1_MAX_SHARED_CTXTS)
 915                 return -EINVAL;
 916
 917         swminor = uinfo->userversion & 0xffff;
 918
 919         /*
 920          * Acquire the mutex to protect against multiple creations of what
 921          * could be a shared base context.
 922          */
 923         mutex_lock(&hfi1_mutex);
 924         /*
 925          * Get a sub context if available  (fd->uctxt will be set).
 926          * ret < 0 error, 0 no context, 1 sub-context found
 927          */
 928         ret = find_sub_ctxt(fd, uinfo);
 929
 930         /*
 931          * Allocate a base context if context sharing is not required or a
 932          * sub context wasn't found.
 933          */
 934         if (!ret)
 935                 ret = allocate_ctxt(fd, fd->dd, uinfo, &uctxt);
 936
 937         mutex_unlock(&hfi1_mutex);
 938
 939         /* Depending on the context type, finish the appropriate init */
 940         switch (ret) {
 941         case 0:
 942                 ret = setup_base_ctxt(fd, uctxt);
 943                 if (ret)
 944                         deallocate_ctxt(uctxt);
 945                 break;
 946         case 1:
 947                 ret = complete_subctxt(fd);
 948                 break;
 949         default:
 950                 break;
 951         }
 952
 953         return ret;
 954 }
 955
 956 /**
 957  * match_ctxt
 958  * @fd: valid filedata pointer
 959  * @uinfo: user info to compare base context with
 960  * @uctxt: context to compare uinfo to.
 961  *
 962  * Compare the given context with the given information to see if it
 963  * can be used for a sub context.
 964  */
 965 static int match_ctxt(struct hfi1_filedata *fd,
 966                       const struct hfi1_user_info *uinfo,
 967                       struct hfi1_ctxtdata *uctxt)
 968 {
 969         struct hfi1_devdata *dd = fd->dd;
 970         unsigned long flags;
 971         u16 subctxt;
 972
 973         /* Skip dynamically allocated kernel contexts */
 974         if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
 975                 return 0;
 976
 977         /* Skip ctxt if it doesn't match the requested one */
 978         if (memcmp(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid)) ||
 979             uctxt->jkey != generate_jkey(current_uid()) ||
 980             uctxt->subctxt_id != uinfo->subctxt_id ||
 981             uctxt->subctxt_cnt != uinfo->subctxt_cnt)
 982                 return 0;
 983
 984         /* Verify the sharing process matches the base */
 985         if (uctxt->userversion != uinfo->userversion)
 986                 return -EINVAL;
 987
 988         /* Find an unused sub context */
 989         spin_lock_irqsave(&dd->uctxt_lock, flags);
 990         if (bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
 991                 /* context is being closed, do not use */
 992                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 993                 return 0;
 994         }
 995
 996         subctxt = find_first_zero_bit(uctxt->in_use_ctxts,
 997                                       HFI1_MAX_SHARED_CTXTS);
 998         if (subctxt >= uctxt->subctxt_cnt) {
 999                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1000                 return -EBUSY;
1001         }
1002
1003         fd->subctxt = subctxt;
1004         __set_bit(fd->subctxt, uctxt->in_use_ctxts);
1005         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1006
1007         fd->uctxt = uctxt;
1008         hfi1_rcd_get(uctxt);
1009
1010         return 1;
1011 }
1012
1013 /**
1014  * find_sub_ctxt
1015  * @fd: valid filedata pointer
1016  * @uinfo: matching info to use to find a possible context to share.
1017  *
1018  * The hfi1_mutex must be held when this function is called.  It is
1019  * necessary to ensure serialized creation of shared contexts.
1020  *
1021  * Return:
1022  *    0      No sub-context found
1023  *    1      Subcontext found and allocated
1024  *    errno  EINVAL (incorrect parameters)
1025  *           EBUSY (all sub contexts in use)
1026  */
1027 static int find_sub_ctxt(struct hfi1_filedata *fd,
1028                          const struct hfi1_user_info *uinfo)
1029 {
1030         struct hfi1_ctxtdata *uctxt;
1031         struct hfi1_devdata *dd = fd->dd;
1032         u16 i;
1033         int ret;
1034
1035         if (!uinfo->subctxt_cnt)
1036                 return 0;
1037
1038         for (i = dd->first_dyn_alloc_ctxt; i < dd->num_rcv_contexts; i++) {
1039                 uctxt = hfi1_rcd_get_by_index(dd, i);
1040                 if (uctxt) {
1041                         ret = match_ctxt(fd, uinfo, uctxt);
1042                         hfi1_rcd_put(uctxt);
1043                         /* value of != 0 will return */
1044                         if (ret)
1045                                 return ret;
1046                 }
1047         }
1048
1049         return 0;
1050 }
1051
1052 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
1053                          struct hfi1_user_info *uinfo,
1054                          struct hfi1_ctxtdata **rcd)
1055 {
1056         struct hfi1_ctxtdata *uctxt;
1057         int ret, numa;
1058
1059         if (dd->flags & HFI1_FROZEN) {
1060                 /*
1061                  * Pick an error that is unique from all other errors
1062                  * that are returned so the user process knows that
1063                  * it tried to allocate while the SPC was frozen.  It
1064                  * it should be able to retry with success in a short
1065                  * while.
1066                  */
1067                 return -EIO;
1068         }
1069
1070         if (!dd->freectxts)
1071                 return -EBUSY;
1072
1073         /*
1074          * If we don't have a NUMA node requested, preference is towards
1075          * device NUMA node.
1076          */
1077         fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
1078         if (fd->rec_cpu_num != -1)
1079                 numa = cpu_to_node(fd->rec_cpu_num);
1080         else
1081                 numa = numa_node_id();
1082         ret = hfi1_create_ctxtdata(dd->pport, numa, &uctxt);
1083         if (ret < 0) {
1084                 dd_dev_err(dd, "user ctxtdata allocation failed\n");
1085                 return ret;
1086         }
1087         hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
1088                   uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
1089                   uctxt->numa_id);
1090
1091         /*
1092          * Allocate and enable a PIO send context.
1093          */
1094         uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize, dd->node);
1095         if (!uctxt->sc) {
1096                 ret = -ENOMEM;
1097                 goto ctxdata_free;
1098         }
1099         hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
1100                   uctxt->sc->hw_context);
1101         ret = sc_enable(uctxt->sc);
1102         if (ret)
1103                 goto ctxdata_free;
1104
1105         /*
1106          * Setup sub context information if the user-level has requested
1107          * sub contexts.
1108          * This has to be done here so the rest of the sub-contexts find the
1109          * proper base context.
1110          */
1111         if (uinfo->subctxt_cnt)
1112                 init_subctxts(uctxt, uinfo);
1113         uctxt->userversion = uinfo->userversion;
1114         uctxt->flags = hfi1_cap_mask; /* save current flag state */
1115         init_waitqueue_head(&uctxt->wait);
1116         strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1117         memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1118         uctxt->jkey = generate_jkey(current_uid());
1119         hfi1_stats.sps_ctxts++;
1120         /*
1121          * Disable ASPM when there are open user/PSM contexts to avoid
1122          * issues with ASPM L1 exit latency
1123          */
1124         if (dd->freectxts-- == dd->num_user_contexts)
1125                 aspm_disable_all(dd);
1126
1127         *rcd = uctxt;
1128
1129         return 0;
1130
1131 ctxdata_free:
1132         hfi1_free_ctxt(uctxt);
1133         return ret;
1134 }
1135
1136 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt)
1137 {
1138         mutex_lock(&hfi1_mutex);
1139         hfi1_stats.sps_ctxts--;
1140         if (++uctxt->dd->freectxts == uctxt->dd->num_user_contexts)
1141                 aspm_enable_all(uctxt->dd);
1142         mutex_unlock(&hfi1_mutex);
1143
1144         hfi1_free_ctxt(uctxt);
1145 }
1146
1147 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
1148                           const struct hfi1_user_info *uinfo)
1149 {
1150         uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1151         uctxt->subctxt_id = uinfo->subctxt_id;
1152         set_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1153 }
1154
1155 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1156 {
1157         int ret = 0;
1158         u16 num_subctxts = uctxt->subctxt_cnt;
1159
1160         uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1161         if (!uctxt->subctxt_uregbase)
1162                 return -ENOMEM;
1163
1164         /* We can take the size of the RcvHdr Queue from the master */
1165         uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1166                                                   num_subctxts);
1167         if (!uctxt->subctxt_rcvhdr_base) {
1168                 ret = -ENOMEM;
1169                 goto bail_ureg;
1170         }
1171
1172         uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1173                                                 num_subctxts);
1174         if (!uctxt->subctxt_rcvegrbuf) {
1175                 ret = -ENOMEM;
1176                 goto bail_rhdr;
1177         }
1178
1179         return 0;
1180
1181 bail_rhdr:
1182         vfree(uctxt->subctxt_rcvhdr_base);
1183         uctxt->subctxt_rcvhdr_base = NULL;
1184 bail_ureg:
1185         vfree(uctxt->subctxt_uregbase);
1186         uctxt->subctxt_uregbase = NULL;
1187
1188         return ret;
1189 }
1190
1191 static void user_init(struct hfi1_ctxtdata *uctxt)
1192 {
1193         unsigned int rcvctrl_ops = 0;
1194
1195         /* initialize poll variables... */
1196         uctxt->urgent = 0;
1197         uctxt->urgent_poll = 0;
1198
1199         /*
1200          * Now enable the ctxt for receive.
1201          * For chips that are set to DMA the tail register to memory
1202          * when they change (and when the update bit transitions from
1203          * 0 to 1.  So for those chips, we turn it off and then back on.
1204          * This will (very briefly) affect any other open ctxts, but the
1205          * duration is very short, and therefore isn't an issue.  We
1206          * explicitly set the in-memory tail copy to 0 beforehand, so we
1207          * don't have to wait to be sure the DMA update has happened
1208          * (chip resets head/tail to 0 on transition to enable).
1209          */
1210         if (uctxt->rcvhdrtail_kvaddr)
1211                 clear_rcvhdrtail(uctxt);
1212
1213         /* Setup J_KEY before enabling the context */
1214         hfi1_set_ctxt_jkey(uctxt->dd, uctxt, uctxt->jkey);
1215
1216         rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1217         if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1218                 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1219         /*
1220          * Ignore the bit in the flags for now until proper
1221          * support for multiple packet per rcv array entry is
1222          * added.
1223          */
1224         if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1225                 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1226         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1227                 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1228         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1229                 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1230         /*
1231          * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1232          * We can't rely on the correct value to be set from prior
1233          * uses of the chip or ctxt. Therefore, add the rcvctrl op
1234          * for both cases.
1235          */
1236         if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1237                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1238         else
1239                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1240         hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt);
1241 }
1242
1243 static int get_ctxt_info(struct hfi1_filedata *fd, void __user *ubase,
1244                          __u32 len)
1245 {
1246         struct hfi1_ctxt_info cinfo;
1247         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1248         int ret = 0;
1249
1250         memset(&cinfo, 0, sizeof(cinfo));
1251         cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1252                                 HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1253                         HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1254                         HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1255         /* adjust flag if this fd is not able to cache */
1256         if (!fd->handler)
1257                 cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1258
1259         cinfo.num_active = hfi1_count_active_units();
1260         cinfo.unit = uctxt->dd->unit;
1261         cinfo.ctxt = uctxt->ctxt;
1262         cinfo.subctxt = fd->subctxt;
1263         cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1264                                 uctxt->dd->rcv_entries.group_size) +
1265                 uctxt->expected_count;
1266         cinfo.credits = uctxt->sc->credits;
1267         cinfo.numa_node = uctxt->numa_id;
1268         cinfo.rec_cpu = fd->rec_cpu_num;
1269         cinfo.send_ctxt = uctxt->sc->hw_context;
1270
1271         cinfo.egrtids = uctxt->egrbufs.alloced;
1272         cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1273         cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1274         cinfo.sdma_ring_size = fd->cq->nentries;
1275         cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1276
1277         trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1278         if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1279                 ret = -EFAULT;
1280
1281         return ret;
1282 }
1283
1284 static int init_user_ctxt(struct hfi1_filedata *fd,
1285                           struct hfi1_ctxtdata *uctxt)
1286 {
1287         int ret;
1288
1289         ret = hfi1_user_sdma_alloc_queues(uctxt, fd);
1290         if (ret)
1291                 return ret;
1292
1293         ret = hfi1_user_exp_rcv_init(fd, uctxt);
1294         if (ret)
1295                 hfi1_user_sdma_free_queues(fd, uctxt);
1296
1297         return ret;
1298 }
1299
1300 static int setup_base_ctxt(struct hfi1_filedata *fd,
1301                            struct hfi1_ctxtdata *uctxt)
1302 {
1303         struct hfi1_devdata *dd = uctxt->dd;
1304         int ret = 0;
1305
1306         hfi1_init_ctxt(uctxt->sc);
1307
1308         /* Now allocate the RcvHdr queue and eager buffers. */
1309         ret = hfi1_create_rcvhdrq(dd, uctxt);
1310         if (ret)
1311                 goto done;
1312
1313         ret = hfi1_setup_eagerbufs(uctxt);
1314         if (ret)
1315                 goto done;
1316
1317         /* If sub-contexts are enabled, do the appropriate setup */
1318         if (uctxt->subctxt_cnt)
1319                 ret = setup_subctxt(uctxt);
1320         if (ret)
1321                 goto done;
1322
1323         ret = hfi1_alloc_ctxt_rcv_groups(uctxt);
1324         if (ret)
1325                 goto done;
1326
1327         ret = init_user_ctxt(fd, uctxt);
1328         if (ret)
1329                 goto done;
1330
1331         user_init(uctxt);
1332
1333         /* Now that the context is set up, the fd can get a reference. */
1334         fd->uctxt = uctxt;
1335         hfi1_rcd_get(uctxt);
1336
1337 done:
1338         if (uctxt->subctxt_cnt) {
1339                 /*
1340                  * On error, set the failed bit so sub-contexts will clean up
1341                  * correctly.
1342                  */
1343                 if (ret)
1344                         set_bit(HFI1_CTXT_BASE_FAILED, &uctxt->event_flags);
1345
1346                 /*
1347                  * Base context is done (successfully or not), notify anybody
1348                  * using a sub-context that is waiting for this completion.
1349                  */
1350                 clear_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1351                 wake_up(&uctxt->wait);
1352         }
1353
1354         return ret;
1355 }
1356
1357 static int get_base_info(struct hfi1_filedata *fd, void __user *ubase,
1358                          __u32 len)
1359 {
1360         struct hfi1_base_info binfo;
1361         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1362         struct hfi1_devdata *dd = uctxt->dd;
1363         ssize_t sz;
1364         unsigned offset;
1365         int ret = 0;
1366
1367         trace_hfi1_uctxtdata(uctxt->dd, uctxt, fd->subctxt);
1368
1369         memset(&binfo, 0, sizeof(binfo));
1370         binfo.hw_version = dd->revision;
1371         binfo.sw_version = HFI1_KERN_SWVERSION;
1372         binfo.bthqp = kdeth_qp;
1373         binfo.jkey = uctxt->jkey;
1374         /*
1375          * If more than 64 contexts are enabled the allocated credit
1376          * return will span two or three contiguous pages. Since we only
1377          * map the page containing the context's credit return address,
1378          * we need to calculate the offset in the proper page.
1379          */
1380         offset = ((u64)uctxt->sc->hw_free -
1381                   (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1382         binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1383                                                 fd->subctxt, offset);
1384         binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1385                                             fd->subctxt,
1386                                             uctxt->sc->base_addr);
1387         binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1388                                                 uctxt->ctxt,
1389                                                 fd->subctxt,
1390                                                 uctxt->sc->base_addr);
1391         binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1392                                                fd->subctxt,
1393                                                uctxt->rcvhdrq);
1394         binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1395                                                fd->subctxt,
1396                                                uctxt->egrbufs.rcvtids[0].dma);
1397         binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1398                                                  fd->subctxt, 0);
1399         /*
1400          * user regs are at
1401          * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1402          */
1403         binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1404                                             fd->subctxt, 0);
1405         offset = offset_in_page((((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1406                     HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1407                   sizeof(*dd->events));
1408         binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1409                                               fd->subctxt,
1410                                               offset);
1411         binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1412                                               fd->subctxt,
1413                                               dd->status);
1414         if (HFI1_CAP_IS_USET(DMA_RTAIL))
1415                 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1416                                                        fd->subctxt, 0);
1417         if (uctxt->subctxt_cnt) {
1418                 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1419                                                         uctxt->ctxt,
1420                                                         fd->subctxt, 0);
1421                 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1422                                                          uctxt->ctxt,
1423                                                          fd->subctxt, 0);
1424                 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1425                                                          uctxt->ctxt,
1426                                                          fd->subctxt, 0);
1427         }
1428         sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1429         if (copy_to_user(ubase, &binfo, sz))
1430                 ret = -EFAULT;
1431         return ret;
1432 }
1433
1434 static unsigned int poll_urgent(struct file *fp,
1435                                 struct poll_table_struct *pt)
1436 {
1437         struct hfi1_filedata *fd = fp->private_data;
1438         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1439         struct hfi1_devdata *dd = uctxt->dd;
1440         unsigned pollflag;
1441
1442         poll_wait(fp, &uctxt->wait, pt);
1443
1444         spin_lock_irq(&dd->uctxt_lock);
1445         if (uctxt->urgent != uctxt->urgent_poll) {
1446                 pollflag = POLLIN | POLLRDNORM;
1447                 uctxt->urgent_poll = uctxt->urgent;
1448         } else {
1449                 pollflag = 0;
1450                 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1451         }
1452         spin_unlock_irq(&dd->uctxt_lock);
1453
1454         return pollflag;
1455 }
1456
1457 static unsigned int poll_next(struct file *fp,
1458                               struct poll_table_struct *pt)
1459 {
1460         struct hfi1_filedata *fd = fp->private_data;
1461         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1462         struct hfi1_devdata *dd = uctxt->dd;
1463         unsigned pollflag;
1464
1465         poll_wait(fp, &uctxt->wait, pt);
1466
1467         spin_lock_irq(&dd->uctxt_lock);
1468         if (hdrqempty(uctxt)) {
1469                 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1470                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt);
1471                 pollflag = 0;
1472         } else {
1473                 pollflag = POLLIN | POLLRDNORM;
1474         }
1475         spin_unlock_irq(&dd->uctxt_lock);
1476
1477         return pollflag;
1478 }
1479
1480 /*
1481  * Find all user contexts in use, and set the specified bit in their
1482  * event mask.
1483  * See also find_ctxt() for a similar use, that is specific to send buffers.
1484  */
1485 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1486 {
1487         struct hfi1_ctxtdata *uctxt;
1488         struct hfi1_devdata *dd = ppd->dd;
1489         u16 ctxt;
1490
1491         if (!dd->events)
1492                 return -EINVAL;
1493
1494         for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
1495              ctxt++) {
1496                 uctxt = hfi1_rcd_get_by_index(dd, ctxt);
1497                 if (uctxt) {
1498                         unsigned long *evs = dd->events +
1499                                 (uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1500                                 HFI1_MAX_SHARED_CTXTS;
1501                         int i;
1502                         /*
1503                          * subctxt_cnt is 0 if not shared, so do base
1504                          * separately, first, then remaining subctxt, if any
1505                          */
1506                         set_bit(evtbit, evs);
1507                         for (i = 1; i < uctxt->subctxt_cnt; i++)
1508                                 set_bit(evtbit, evs + i);
1509                         hfi1_rcd_put(uctxt);
1510                 }
1511         }
1512
1513         return 0;
1514 }
1515
1516 /**
1517  * manage_rcvq - manage a context's receive queue
1518  * @uctxt: the context
1519  * @subctxt: the sub-context
1520  * @start_stop: action to carry out
1521  *
1522  * start_stop == 0 disables receive on the context, for use in queue
1523  * overflow conditions.  start_stop==1 re-enables, to be used to
1524  * re-init the software copy of the head register
1525  */
1526 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1527                        int start_stop)
1528 {
1529         struct hfi1_devdata *dd = uctxt->dd;
1530         unsigned int rcvctrl_op;
1531
1532         if (subctxt)
1533                 goto bail;
1534         /* atomically clear receive enable ctxt. */
1535         if (start_stop) {
1536                 /*
1537                  * On enable, force in-memory copy of the tail register to
1538                  * 0, so that protocol code doesn't have to worry about
1539                  * whether or not the chip has yet updated the in-memory
1540                  * copy or not on return from the system call. The chip
1541                  * always resets it's tail register back to 0 on a
1542                  * transition from disabled to enabled.
1543                  */
1544                 if (uctxt->rcvhdrtail_kvaddr)
1545                         clear_rcvhdrtail(uctxt);
1546                 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1547         } else {
1548                 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1549         }
1550         hfi1_rcvctrl(dd, rcvctrl_op, uctxt);
1551         /* always; new head should be equal to new tail; see above */
1552 bail:
1553         return 0;
1554 }
1555
1556 /*
1557  * clear the event notifier events for this context.
1558  * User process then performs actions appropriate to bit having been
1559  * set, if desired, and checks again in future.
1560  */
1561 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1562                           unsigned long events)
1563 {
1564         int i;
1565         struct hfi1_devdata *dd = uctxt->dd;
1566         unsigned long *evs;
1567
1568         if (!dd->events)
1569                 return 0;
1570
1571         evs = dd->events + ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1572                             HFI1_MAX_SHARED_CTXTS) + subctxt;
1573
1574         for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1575                 if (!test_bit(i, &events))
1576                         continue;
1577                 clear_bit(i, evs);
1578         }
1579         return 0;
1580 }
1581
1582 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, u16 subctxt, u16 pkey)
1583 {
1584         int ret = -ENOENT, i, intable = 0;
1585         struct hfi1_pportdata *ppd = uctxt->ppd;
1586         struct hfi1_devdata *dd = uctxt->dd;
1587
1588         if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1589                 ret = -EINVAL;
1590                 goto done;
1591         }
1592
1593         for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1594                 if (pkey == ppd->pkeys[i]) {
1595                         intable = 1;
1596                         break;
1597                 }
1598
1599         if (intable)
1600                 ret = hfi1_set_ctxt_pkey(dd, uctxt, pkey);
1601 done:
1602         return ret;
1603 }
1604
1605 static void user_remove(struct hfi1_devdata *dd)
1606 {
1607
1608         hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1609 }
1610
1611 static int user_add(struct hfi1_devdata *dd)
1612 {
1613         char name[10];
1614         int ret;
1615
1616         snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1617         ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1618                              &dd->user_cdev, &dd->user_device,
1619                              true, &dd->kobj);
1620         if (ret)
1621                 user_remove(dd);
1622
1623         return ret;
1624 }
1625
1626 /*
1627  * Create per-unit files in /dev
1628  */
1629 int hfi1_device_create(struct hfi1_devdata *dd)
1630 {
1631         return user_add(dd);
1632 }
1633
1634 /*
1635  * Remove per-unit files in /dev
1636  * void, core kernel returns no errors for this stuff
1637  */
1638 void hfi1_device_remove(struct hfi1_devdata *dd)
1639 {
1640         user_remove(dd);
1641 }