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         if (!HFI1_CAP_IS_KSET(SDMA))
 428                 return -EINVAL;
 429         idx = srcu_read_lock(&fd->pq_srcu);
 430         pq = srcu_dereference(fd->pq, &fd->pq_srcu);
 431         if (!cq || !pq) {
 432                 srcu_read_unlock(&fd->pq_srcu, idx);
 433                 return -EIO;
 434         }
 435
 436         if (!iter_is_iovec(from) || !dim) {
 437                 srcu_read_unlock(&fd->pq_srcu, idx);
 438                 return -EINVAL;
 439         }
 440
 441         trace_hfi1_sdma_request(fd->dd, fd->uctxt->ctxt, fd->subctxt, dim);
 442
 443         if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
 444                 srcu_read_unlock(&fd->pq_srcu, idx);
 445                 return -ENOSPC;
 446         }
 447
 448         while (dim) {
 449                 int ret;
 450                 unsigned long count = 0;
 451
 452                 ret = hfi1_user_sdma_process_request(
 453                         fd, (struct iovec *)(from->iov + done),
 454                         dim, &count);
 455                 if (ret) {
 456                         reqs = ret;
 457                         break;
 458                 }
 459                 dim -= count;
 460                 done += count;
 461                 reqs++;
 462         }
 463
 464         srcu_read_unlock(&fd->pq_srcu, idx);
 465         return reqs;
 466 }
 467
 468 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
 469 {
 470         struct hfi1_filedata *fd = fp->private_data;
 471         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 472         struct hfi1_devdata *dd;
 473         unsigned long flags;
 474         u64 token = vma->vm_pgoff << PAGE_SHIFT,
 475                 memaddr = 0;
 476         void *memvirt = NULL;
 477         u8 subctxt, mapio = 0, vmf = 0, type;
 478         ssize_t memlen = 0;
 479         int ret = 0;
 480         u16 ctxt;
 481
 482         if (!is_valid_mmap(token) || !uctxt ||
 483             !(vma->vm_flags & VM_SHARED)) {
 484                 ret = -EINVAL;
 485                 goto done;
 486         }
 487         dd = uctxt->dd;
 488         ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
 489         subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
 490         type = HFI1_MMAP_TOKEN_GET(TYPE, token);
 491         if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
 492                 ret = -EINVAL;
 493                 goto done;
 494         }
 495
 496         flags = vma->vm_flags;
 497
 498         switch (type) {
 499         case PIO_BUFS:
 500         case PIO_BUFS_SOP:
 501                 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
 502                                 /* chip pio base */
 503                            (uctxt->sc->hw_context * BIT(16))) +
 504                                 /* 64K PIO space / ctxt */
 505                         (type == PIO_BUFS_SOP ?
 506                                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
 507                 /*
 508                  * Map only the amount allocated to the context, not the
 509                  * entire available context's PIO space.
 510                  */
 511                 memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
 512                 flags &= ~VM_MAYREAD;
 513                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 514                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 515                 mapio = 1;
 516                 break;
 517         case PIO_CRED:
 518                 if (flags & VM_WRITE) {
 519                         ret = -EPERM;
 520                         goto done;
 521                 }
 522                 /*
 523                  * The credit return location for this context could be on the
 524                  * second or third page allocated for credit returns (if number
 525                  * of enabled contexts > 64 and 128 respectively).
 526                  */
 527                 memvirt = dd->cr_base[uctxt->numa_id].va;
 528                 memaddr = virt_to_phys(memvirt) +
 529                         (((u64)uctxt->sc->hw_free -
 530                           (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
 531                 memlen = PAGE_SIZE;
 532                 flags &= ~VM_MAYWRITE;
 533                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 534                 /*
 535                  * The driver has already allocated memory for credit
 536                  * returns and programmed it into the chip. Has that
 537                  * memory been flagged as non-cached?
 538                  */
 539                 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
 540                 mapio = 1;
 541                 break;
 542         case RCV_HDRQ:
 543                 memlen = uctxt->rcvhdrq_size;
 544                 memvirt = uctxt->rcvhdrq;
 545                 break;
 546         case RCV_EGRBUF: {
 547                 unsigned long addr;
 548                 int i;
 549                 /*
 550                  * The RcvEgr buffer need to be handled differently
 551                  * as multiple non-contiguous pages need to be mapped
 552                  * into the user process.
 553                  */
 554                 memlen = uctxt->egrbufs.size;
 555                 if ((vma->vm_end - vma->vm_start) != memlen) {
 556                         dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
 557                                    (vma->vm_end - vma->vm_start), memlen);
 558                         ret = -EINVAL;
 559                         goto done;
 560                 }
 561                 if (vma->vm_flags & VM_WRITE) {
 562                         ret = -EPERM;
 563                         goto done;
 564                 }
 565                 vma->vm_flags &= ~VM_MAYWRITE;
 566                 addr = vma->vm_start;
 567                 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
 568                         memlen = uctxt->egrbufs.buffers[i].len;
 569                         memvirt = uctxt->egrbufs.buffers[i].addr;
 570                         ret = remap_pfn_range(
 571                                 vma, addr,
 572                                 /*
 573                                  * virt_to_pfn() does the same, but
 574                                  * it's not available on x86_64
 575                                  * when CONFIG_MMU is enabled.
 576                                  */
 577                                 PFN_DOWN(__pa(memvirt)),
 578                                 memlen,
 579                                 vma->vm_page_prot);
 580                         if (ret < 0)
 581                                 goto done;
 582                         addr += memlen;
 583                 }
 584                 ret = 0;
 585                 goto done;
 586         }
 587         case UREGS:
 588                 /*
 589                  * Map only the page that contains this context's user
 590                  * registers.
 591                  */
 592                 memaddr = (unsigned long)
 593                         (dd->physaddr + RXE_PER_CONTEXT_USER)
 594                         + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
 595                 /*
 596                  * TidFlow table is on the same page as the rest of the
 597                  * user registers.
 598                  */
 599                 memlen = PAGE_SIZE;
 600                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 601                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 602                 mapio = 1;
 603                 break;
 604         case EVENTS:
 605                 /*
 606                  * Use the page where this context's flags are. User level
 607                  * knows where it's own bitmap is within the page.
 608                  */
 609                 memaddr = (unsigned long)(dd->events +
 610                                   ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
 611                                    HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
 612                 memlen = PAGE_SIZE;
 613                 /*
 614                  * v3.7 removes VM_RESERVED but the effect is kept by
 615                  * using VM_IO.
 616                  */
 617                 flags |= VM_IO | VM_DONTEXPAND;
 618                 vmf = 1;
 619                 break;
 620         case STATUS:
 621                 if (flags & VM_WRITE) {
 622                         ret = -EPERM;
 623                         goto done;
 624                 }
 625                 memaddr = kvirt_to_phys((void *)dd->status);
 626                 memlen = PAGE_SIZE;
 627                 flags |= VM_IO | VM_DONTEXPAND;
 628                 break;
 629         case RTAIL:
 630                 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
 631                         /*
 632                          * If the memory allocation failed, the context alloc
 633                          * also would have failed, so we would never get here
 634                          */
 635                         ret = -EINVAL;
 636                         goto done;
 637                 }
 638                 if ((flags & VM_WRITE) || !uctxt->rcvhdrtail_kvaddr) {
 639                         ret = -EPERM;
 640                         goto done;
 641                 }
 642                 memlen = PAGE_SIZE;
 643                 memvirt = (void *)uctxt->rcvhdrtail_kvaddr;
 644                 flags &= ~VM_MAYWRITE;
 645                 break;
 646         case SUBCTXT_UREGS:
 647                 memaddr = (u64)uctxt->subctxt_uregbase;
 648                 memlen = PAGE_SIZE;
 649                 flags |= VM_IO | VM_DONTEXPAND;
 650                 vmf = 1;
 651                 break;
 652         case SUBCTXT_RCV_HDRQ:
 653                 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
 654                 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
 655                 flags |= VM_IO | VM_DONTEXPAND;
 656                 vmf = 1;
 657                 break;
 658         case SUBCTXT_EGRBUF:
 659                 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
 660                 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
 661                 flags |= VM_IO | VM_DONTEXPAND;
 662                 flags &= ~VM_MAYWRITE;
 663                 vmf = 1;
 664                 break;
 665         case SDMA_COMP: {
 666                 struct hfi1_user_sdma_comp_q *cq = fd->cq;
 667
 668                 if (!cq) {
 669                         ret = -EFAULT;
 670                         goto done;
 671                 }
 672                 memaddr = (u64)cq->comps;
 673                 memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
 674                 flags |= VM_IO | VM_DONTEXPAND;
 675                 vmf = 1;
 676                 break;
 677         }
 678         default:
 679                 ret = -EINVAL;
 680                 break;
 681         }
 682
 683         if ((vma->vm_end - vma->vm_start) != memlen) {
 684                 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
 685                           uctxt->ctxt, fd->subctxt,
 686                           (vma->vm_end - vma->vm_start), memlen);
 687                 ret = -EINVAL;
 688                 goto done;
 689         }
 690
 691         vma->vm_flags = flags;
 692         hfi1_cdbg(PROC,
 693                   "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
 694                     ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
 695                     vma->vm_end - vma->vm_start, vma->vm_flags);
 696         if (vmf) {
 697                 vma->vm_pgoff = PFN_DOWN(memaddr);
 698                 vma->vm_ops = &vm_ops;
 699                 ret = 0;
 700         } else if (mapio) {
 701                 ret = io_remap_pfn_range(vma, vma->vm_start,
 702                                          PFN_DOWN(memaddr),
 703                                          memlen,
 704                                          vma->vm_page_prot);
 705         } else if (memvirt) {
 706                 ret = remap_pfn_range(vma, vma->vm_start,
 707                                       PFN_DOWN(__pa(memvirt)),
 708                                       memlen,
 709                                       vma->vm_page_prot);
 710         } else {
 711                 ret = remap_pfn_range(vma, vma->vm_start,
 712                                       PFN_DOWN(memaddr),
 713                                       memlen,
 714                                       vma->vm_page_prot);
 715         }
 716 done:
 717         return ret;
 718 }
 719
 720 /*
 721  * Local (non-chip) user memory is not mapped right away but as it is
 722  * accessed by the user-level code.
 723  */
 724 static int vma_fault(struct vm_fault *vmf)
 725 {
 726         struct page *page;
 727
 728         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 729         if (!page)
 730                 return VM_FAULT_SIGBUS;
 731
 732         get_page(page);
 733         vmf->page = page;
 734
 735         return 0;
 736 }
 737
 738 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
 739 {
 740         struct hfi1_ctxtdata *uctxt;
 741         unsigned pollflag;
 742
 743         uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
 744         if (!uctxt)
 745                 pollflag = POLLERR;
 746         else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
 747                 pollflag = poll_urgent(fp, pt);
 748         else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
 749                 pollflag = poll_next(fp, pt);
 750         else /* invalid */
 751                 pollflag = POLLERR;
 752
 753         return pollflag;
 754 }
 755
 756 static int hfi1_file_close(struct inode *inode, struct file *fp)
 757 {
 758         struct hfi1_filedata *fdata = fp->private_data;
 759         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 760         struct hfi1_devdata *dd = container_of(inode->i_cdev,
 761                                                struct hfi1_devdata,
 762                                                user_cdev);
 763         unsigned long flags, *ev;
 764
 765         fp->private_data = NULL;
 766
 767         if (!uctxt)
 768                 goto done;
 769
 770         hfi1_cdbg(PROC, "closing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
 771
 772         flush_wc();
 773         /* drain user sdma queue */
 774         hfi1_user_sdma_free_queues(fdata, uctxt);
 775
 776         /* release the cpu */
 777         hfi1_put_proc_affinity(fdata->rec_cpu_num);
 778
 779         /* clean up rcv side */
 780         hfi1_user_exp_rcv_free(fdata);
 781
 782         /*
 783          * fdata->uctxt is used in the above cleanup.  It is not ready to be
 784          * removed until here.
 785          */
 786         fdata->uctxt = NULL;
 787         hfi1_rcd_put(uctxt);
 788
 789         /*
 790          * Clear any left over, unhandled events so the next process that
 791          * gets this context doesn't get confused.
 792          */
 793         ev = dd->events + ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
 794                            HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
 795         *ev = 0;
 796
 797         spin_lock_irqsave(&dd->uctxt_lock, flags);
 798         __clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
 799         if (!bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
 800                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 801                 goto done;
 802         }
 803         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 804
 805         /*
 806          * Disable receive context and interrupt available, reset all
 807          * RcvCtxtCtrl bits to default values.
 808          */
 809         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
 810                      HFI1_RCVCTRL_TIDFLOW_DIS |
 811                      HFI1_RCVCTRL_INTRAVAIL_DIS |
 812                      HFI1_RCVCTRL_TAILUPD_DIS |
 813                      HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
 814                      HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
 815                      HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt);
 816         /* Clear the context's J_KEY */
 817         hfi1_clear_ctxt_jkey(dd, uctxt);
 818         /*
 819          * If a send context is allocated, reset context integrity
 820          * checks to default and disable the send context.
 821          */
 822         if (uctxt->sc) {
 823                 sc_disable(uctxt->sc);
 824                 set_pio_integrity(uctxt->sc);
 825         }
 826
 827         hfi1_free_ctxt_rcv_groups(uctxt);
 828         hfi1_clear_ctxt_pkey(dd, uctxt);
 829
 830         uctxt->event_flags = 0;
 831
 832         deallocate_ctxt(uctxt);
 833 done:
 834         mmdrop(fdata->mm);
 835         kobject_put(&dd->kobj);
 836
 837         if (atomic_dec_and_test(&dd->user_refcount))
 838                 complete(&dd->user_comp);
 839
 840         cleanup_srcu_struct(&fdata->pq_srcu);
 841         kfree(fdata);
 842         return 0;
 843 }
 844
 845 /*
 846  * Convert kernel *virtual* addresses to physical addresses.
 847  * This is used to vmalloc'ed addresses.
 848  */
 849 static u64 kvirt_to_phys(void *addr)
 850 {
 851         struct page *page;
 852         u64 paddr = 0;
 853
 854         page = vmalloc_to_page(addr);
 855         if (page)
 856                 paddr = page_to_pfn(page) << PAGE_SHIFT;
 857
 858         return paddr;
 859 }
 860
 861 /**
 862  * complete_subctxt
 863  * @fd: valid filedata pointer
 864  *
 865  * Sub-context info can only be set up after the base context
 866  * has been completed.  This is indicated by the clearing of the
 867  * HFI1_CTXT_BASE_UINIT bit.
 868  *
 869  * Wait for the bit to be cleared, and then complete the subcontext
 870  * initialization.
 871  *
 872  */
 873 static int complete_subctxt(struct hfi1_filedata *fd)
 874 {
 875         int ret;
 876         unsigned long flags;
 877
 878         /*
 879          * sub-context info can only be set up after the base context
 880          * has been completed.
 881          */
 882         ret = wait_event_interruptible(
 883                 fd->uctxt->wait,
 884                 !test_bit(HFI1_CTXT_BASE_UNINIT, &fd->uctxt->event_flags));
 885
 886         if (test_bit(HFI1_CTXT_BASE_FAILED, &fd->uctxt->event_flags))
 887                 ret = -ENOMEM;
 888
 889         /* Finish the sub-context init */
 890         if (!ret) {
 891                 fd->rec_cpu_num = hfi1_get_proc_affinity(fd->uctxt->numa_id);
 892                 ret = init_user_ctxt(fd, fd->uctxt);
 893         }
 894
 895         if (ret) {
 896                 spin_lock_irqsave(&fd->dd->uctxt_lock, flags);
 897                 __clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
 898                 spin_unlock_irqrestore(&fd->dd->uctxt_lock, flags);
 899                 hfi1_rcd_put(fd->uctxt);
 900                 fd->uctxt = NULL;
 901         }
 902
 903         return ret;
 904 }
 905
 906 static int assign_ctxt(struct hfi1_filedata *fd, struct hfi1_user_info *uinfo)
 907 {
 908         int ret;
 909         unsigned int swmajor, swminor;
 910         struct hfi1_ctxtdata *uctxt = NULL;
 911
 912         swmajor = uinfo->userversion >> 16;
 913         if (swmajor != HFI1_USER_SWMAJOR)
 914                 return -ENODEV;
 915
 916         if (uinfo->subctxt_cnt > HFI1_MAX_SHARED_CTXTS)
 917                 return -EINVAL;
 918
 919         swminor = uinfo->userversion & 0xffff;
 920
 921         /*
 922          * Acquire the mutex to protect against multiple creations of what
 923          * could be a shared base context.
 924          */
 925         mutex_lock(&hfi1_mutex);
 926         /*
 927          * Get a sub context if available  (fd->uctxt will be set).
 928          * ret < 0 error, 0 no context, 1 sub-context found
 929          */
 930         ret = find_sub_ctxt(fd, uinfo);
 931
 932         /*
 933          * Allocate a base context if context sharing is not required or a
 934          * sub context wasn't found.
 935          */
 936         if (!ret)
 937                 ret = allocate_ctxt(fd, fd->dd, uinfo, &uctxt);
 938
 939         mutex_unlock(&hfi1_mutex);
 940
 941         /* Depending on the context type, finish the appropriate init */
 942         switch (ret) {
 943         case 0:
 944                 ret = setup_base_ctxt(fd, uctxt);
 945                 if (ret)
 946                         deallocate_ctxt(uctxt);
 947                 break;
 948         case 1:
 949                 ret = complete_subctxt(fd);
 950                 break;
 951         default:
 952                 break;
 953         }
 954
 955         return ret;
 956 }
 957
 958 /**
 959  * match_ctxt
 960  * @fd: valid filedata pointer
 961  * @uinfo: user info to compare base context with
 962  * @uctxt: context to compare uinfo to.
 963  *
 964  * Compare the given context with the given information to see if it
 965  * can be used for a sub context.
 966  */
 967 static int match_ctxt(struct hfi1_filedata *fd,
 968                       const struct hfi1_user_info *uinfo,
 969                       struct hfi1_ctxtdata *uctxt)
 970 {
 971         struct hfi1_devdata *dd = fd->dd;
 972         unsigned long flags;
 973         u16 subctxt;
 974
 975         /* Skip dynamically allocated kernel contexts */
 976         if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
 977                 return 0;
 978
 979         /* Skip ctxt if it doesn't match the requested one */
 980         if (memcmp(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid)) ||
 981             uctxt->jkey != generate_jkey(current_uid()) ||
 982             uctxt->subctxt_id != uinfo->subctxt_id ||
 983             uctxt->subctxt_cnt != uinfo->subctxt_cnt)
 984                 return 0;
 985
 986         /* Verify the sharing process matches the base */
 987         if (uctxt->userversion != uinfo->userversion)
 988                 return -EINVAL;
 989
 990         /* Find an unused sub context */
 991         spin_lock_irqsave(&dd->uctxt_lock, flags);
 992         if (bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
 993                 /* context is being closed, do not use */
 994                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 995                 return 0;
 996         }
 997
 998         subctxt = find_first_zero_bit(uctxt->in_use_ctxts,
 999                                       HFI1_MAX_SHARED_CTXTS);
1000         if (subctxt >= uctxt->subctxt_cnt) {
1001                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1002                 return -EBUSY;
1003         }
1004
1005         fd->subctxt = subctxt;
1006         __set_bit(fd->subctxt, uctxt->in_use_ctxts);
1007         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1008
1009         fd->uctxt = uctxt;
1010         hfi1_rcd_get(uctxt);
1011
1012         return 1;
1013 }
1014
1015 /**
1016  * find_sub_ctxt
1017  * @fd: valid filedata pointer
1018  * @uinfo: matching info to use to find a possible context to share.
1019  *
1020  * The hfi1_mutex must be held when this function is called.  It is
1021  * necessary to ensure serialized creation of shared contexts.
1022  *
1023  * Return:
1024  *    0      No sub-context found
1025  *    1      Subcontext found and allocated
1026  *    errno  EINVAL (incorrect parameters)
1027  *           EBUSY (all sub contexts in use)
1028  */
1029 static int find_sub_ctxt(struct hfi1_filedata *fd,
1030                          const struct hfi1_user_info *uinfo)
1031 {
1032         struct hfi1_ctxtdata *uctxt;
1033         struct hfi1_devdata *dd = fd->dd;
1034         u16 i;
1035         int ret;
1036
1037         if (!uinfo->subctxt_cnt)
1038                 return 0;
1039
1040         for (i = dd->first_dyn_alloc_ctxt; i < dd->num_rcv_contexts; i++) {
1041                 uctxt = hfi1_rcd_get_by_index(dd, i);
1042                 if (uctxt) {
1043                         ret = match_ctxt(fd, uinfo, uctxt);
1044                         hfi1_rcd_put(uctxt);
1045                         /* value of != 0 will return */
1046                         if (ret)
1047                                 return ret;
1048                 }
1049         }
1050
1051         return 0;
1052 }
1053
1054 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
1055                          struct hfi1_user_info *uinfo,
1056                          struct hfi1_ctxtdata **rcd)
1057 {
1058         struct hfi1_ctxtdata *uctxt;
1059         int ret, numa;
1060
1061         if (dd->flags & HFI1_FROZEN) {
1062                 /*
1063                  * Pick an error that is unique from all other errors
1064                  * that are returned so the user process knows that
1065                  * it tried to allocate while the SPC was frozen.  It
1066                  * it should be able to retry with success in a short
1067                  * while.
1068                  */
1069                 return -EIO;
1070         }
1071
1072         if (!dd->freectxts)
1073                 return -EBUSY;
1074
1075         /*
1076          * If we don't have a NUMA node requested, preference is towards
1077          * device NUMA node.
1078          */
1079         fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
1080         if (fd->rec_cpu_num != -1)
1081                 numa = cpu_to_node(fd->rec_cpu_num);
1082         else
1083                 numa = numa_node_id();
1084         ret = hfi1_create_ctxtdata(dd->pport, numa, &uctxt);
1085         if (ret < 0) {
1086                 dd_dev_err(dd, "user ctxtdata allocation failed\n");
1087                 return ret;
1088         }
1089         hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
1090                   uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
1091                   uctxt->numa_id);
1092
1093         /*
1094          * Allocate and enable a PIO send context.
1095          */
1096         uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize, dd->node);
1097         if (!uctxt->sc) {
1098                 ret = -ENOMEM;
1099                 goto ctxdata_free;
1100         }
1101         hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
1102                   uctxt->sc->hw_context);
1103         ret = sc_enable(uctxt->sc);
1104         if (ret)
1105                 goto ctxdata_free;
1106
1107         /*
1108          * Setup sub context information if the user-level has requested
1109          * sub contexts.
1110          * This has to be done here so the rest of the sub-contexts find the
1111          * proper base context.
1112          */
1113         if (uinfo->subctxt_cnt)
1114                 init_subctxts(uctxt, uinfo);
1115         uctxt->userversion = uinfo->userversion;
1116         uctxt->flags = hfi1_cap_mask; /* save current flag state */
1117         init_waitqueue_head(&uctxt->wait);
1118         strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1119         memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1120         uctxt->jkey = generate_jkey(current_uid());
1121         hfi1_stats.sps_ctxts++;
1122         /*
1123          * Disable ASPM when there are open user/PSM contexts to avoid
1124          * issues with ASPM L1 exit latency
1125          */
1126         if (dd->freectxts-- == dd->num_user_contexts)
1127                 aspm_disable_all(dd);
1128
1129         *rcd = uctxt;
1130
1131         return 0;
1132
1133 ctxdata_free:
1134         hfi1_free_ctxt(uctxt);
1135         return ret;
1136 }
1137
1138 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt)
1139 {
1140         mutex_lock(&hfi1_mutex);
1141         hfi1_stats.sps_ctxts--;
1142         if (++uctxt->dd->freectxts == uctxt->dd->num_user_contexts)
1143                 aspm_enable_all(uctxt->dd);
1144         mutex_unlock(&hfi1_mutex);
1145
1146         hfi1_free_ctxt(uctxt);
1147 }
1148
1149 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
1150                           const struct hfi1_user_info *uinfo)
1151 {
1152         uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1153         uctxt->subctxt_id = uinfo->subctxt_id;
1154         set_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1155 }
1156
1157 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1158 {
1159         int ret = 0;
1160         u16 num_subctxts = uctxt->subctxt_cnt;
1161
1162         uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1163         if (!uctxt->subctxt_uregbase)
1164                 return -ENOMEM;
1165
1166         /* We can take the size of the RcvHdr Queue from the master */
1167         uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1168                                                   num_subctxts);
1169         if (!uctxt->subctxt_rcvhdr_base) {
1170                 ret = -ENOMEM;
1171                 goto bail_ureg;
1172         }
1173
1174         uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1175                                                 num_subctxts);
1176         if (!uctxt->subctxt_rcvegrbuf) {
1177                 ret = -ENOMEM;
1178                 goto bail_rhdr;
1179         }
1180
1181         return 0;
1182
1183 bail_rhdr:
1184         vfree(uctxt->subctxt_rcvhdr_base);
1185         uctxt->subctxt_rcvhdr_base = NULL;
1186 bail_ureg:
1187         vfree(uctxt->subctxt_uregbase);
1188         uctxt->subctxt_uregbase = NULL;
1189
1190         return ret;
1191 }
1192
1193 static void user_init(struct hfi1_ctxtdata *uctxt)
1194 {
1195         unsigned int rcvctrl_ops = 0;
1196
1197         /* initialize poll variables... */
1198         uctxt->urgent = 0;
1199         uctxt->urgent_poll = 0;
1200
1201         /*
1202          * Now enable the ctxt for receive.
1203          * For chips that are set to DMA the tail register to memory
1204          * when they change (and when the update bit transitions from
1205          * 0 to 1.  So for those chips, we turn it off and then back on.
1206          * This will (very briefly) affect any other open ctxts, but the
1207          * duration is very short, and therefore isn't an issue.  We
1208          * explicitly set the in-memory tail copy to 0 beforehand, so we
1209          * don't have to wait to be sure the DMA update has happened
1210          * (chip resets head/tail to 0 on transition to enable).
1211          */
1212         if (uctxt->rcvhdrtail_kvaddr)
1213                 clear_rcvhdrtail(uctxt);
1214
1215         /* Setup J_KEY before enabling the context */
1216         hfi1_set_ctxt_jkey(uctxt->dd, uctxt, uctxt->jkey);
1217
1218         rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1219         if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1220                 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1221         /*
1222          * Ignore the bit in the flags for now until proper
1223          * support for multiple packet per rcv array entry is
1224          * added.
1225          */
1226         if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1227                 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1228         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1229                 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1230         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1231                 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1232         /*
1233          * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1234          * We can't rely on the correct value to be set from prior
1235          * uses of the chip or ctxt. Therefore, add the rcvctrl op
1236          * for both cases.
1237          */
1238         if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1239                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1240         else
1241                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1242         hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt);
1243 }
1244
1245 static int get_ctxt_info(struct hfi1_filedata *fd, void __user *ubase,
1246                          __u32 len)
1247 {
1248         struct hfi1_ctxt_info cinfo;
1249         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1250         int ret = 0;
1251
1252         memset(&cinfo, 0, sizeof(cinfo));
1253         cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1254                                 HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1255                         HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1256                         HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1257         /* adjust flag if this fd is not able to cache */
1258         if (!fd->handler)
1259                 cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1260
1261         cinfo.num_active = hfi1_count_active_units();
1262         cinfo.unit = uctxt->dd->unit;
1263         cinfo.ctxt = uctxt->ctxt;
1264         cinfo.subctxt = fd->subctxt;
1265         cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1266                                 uctxt->dd->rcv_entries.group_size) +
1267                 uctxt->expected_count;
1268         cinfo.credits = uctxt->sc->credits;
1269         cinfo.numa_node = uctxt->numa_id;
1270         cinfo.rec_cpu = fd->rec_cpu_num;
1271         cinfo.send_ctxt = uctxt->sc->hw_context;
1272
1273         cinfo.egrtids = uctxt->egrbufs.alloced;
1274         cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1275         cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1276         cinfo.sdma_ring_size = fd->cq->nentries;
1277         cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1278
1279         trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1280         if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1281                 ret = -EFAULT;
1282
1283         return ret;
1284 }
1285
1286 static int init_user_ctxt(struct hfi1_filedata *fd,
1287                           struct hfi1_ctxtdata *uctxt)
1288 {
1289         int ret;
1290
1291         ret = hfi1_user_sdma_alloc_queues(uctxt, fd);
1292         if (ret)
1293                 return ret;
1294
1295         ret = hfi1_user_exp_rcv_init(fd, uctxt);
1296         if (ret)
1297                 hfi1_user_sdma_free_queues(fd, uctxt);
1298
1299         return ret;
1300 }
1301
1302 static int setup_base_ctxt(struct hfi1_filedata *fd,
1303                            struct hfi1_ctxtdata *uctxt)
1304 {
1305         struct hfi1_devdata *dd = uctxt->dd;
1306         int ret = 0;
1307
1308         hfi1_init_ctxt(uctxt->sc);
1309
1310         /* Now allocate the RcvHdr queue and eager buffers. */
1311         ret = hfi1_create_rcvhdrq(dd, uctxt);
1312         if (ret)
1313                 goto done;
1314
1315         ret = hfi1_setup_eagerbufs(uctxt);
1316         if (ret)
1317                 goto done;
1318
1319         /* If sub-contexts are enabled, do the appropriate setup */
1320         if (uctxt->subctxt_cnt)
1321                 ret = setup_subctxt(uctxt);
1322         if (ret)
1323                 goto done;
1324
1325         ret = hfi1_alloc_ctxt_rcv_groups(uctxt);
1326         if (ret)
1327                 goto done;
1328
1329         ret = init_user_ctxt(fd, uctxt);
1330         if (ret) {
1331                 hfi1_free_ctxt_rcv_groups(uctxt);
1332                 goto done;
1333         }
1334
1335         user_init(uctxt);
1336
1337         /* Now that the context is set up, the fd can get a reference. */
1338         fd->uctxt = uctxt;
1339         hfi1_rcd_get(uctxt);
1340
1341 done:
1342         if (uctxt->subctxt_cnt) {
1343                 /*
1344                  * On error, set the failed bit so sub-contexts will clean up
1345                  * correctly.
1346                  */
1347                 if (ret)
1348                         set_bit(HFI1_CTXT_BASE_FAILED, &uctxt->event_flags);
1349
1350                 /*
1351                  * Base context is done (successfully or not), notify anybody
1352                  * using a sub-context that is waiting for this completion.
1353                  */
1354                 clear_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1355                 wake_up(&uctxt->wait);
1356         }
1357
1358         return ret;
1359 }
1360
1361 static int get_base_info(struct hfi1_filedata *fd, void __user *ubase,
1362                          __u32 len)
1363 {
1364         struct hfi1_base_info binfo;
1365         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1366         struct hfi1_devdata *dd = uctxt->dd;
1367         ssize_t sz;
1368         unsigned offset;
1369         int ret = 0;
1370
1371         trace_hfi1_uctxtdata(uctxt->dd, uctxt, fd->subctxt);
1372
1373         memset(&binfo, 0, sizeof(binfo));
1374         binfo.hw_version = dd->revision;
1375         binfo.sw_version = HFI1_KERN_SWVERSION;
1376         binfo.bthqp = kdeth_qp;
1377         binfo.jkey = uctxt->jkey;
1378         /*
1379          * If more than 64 contexts are enabled the allocated credit
1380          * return will span two or three contiguous pages. Since we only
1381          * map the page containing the context's credit return address,
1382          * we need to calculate the offset in the proper page.
1383          */
1384         offset = ((u64)uctxt->sc->hw_free -
1385                   (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1386         binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1387                                                 fd->subctxt, offset);
1388         binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1389                                             fd->subctxt,
1390                                             uctxt->sc->base_addr);
1391         binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1392                                                 uctxt->ctxt,
1393                                                 fd->subctxt,
1394                                                 uctxt->sc->base_addr);
1395         binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1396                                                fd->subctxt,
1397                                                uctxt->rcvhdrq);
1398         binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1399                                                fd->subctxt,
1400                                                uctxt->egrbufs.rcvtids[0].dma);
1401         binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1402                                                  fd->subctxt, 0);
1403         /*
1404          * user regs are at
1405          * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1406          */
1407         binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1408                                             fd->subctxt, 0);
1409         offset = offset_in_page((((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1410                     HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1411                   sizeof(*dd->events));
1412         binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1413                                               fd->subctxt,
1414                                               offset);
1415         binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1416                                               fd->subctxt,
1417                                               dd->status);
1418         if (HFI1_CAP_IS_USET(DMA_RTAIL))
1419                 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1420                                                        fd->subctxt, 0);
1421         if (uctxt->subctxt_cnt) {
1422                 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1423                                                         uctxt->ctxt,
1424                                                         fd->subctxt, 0);
1425                 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1426                                                          uctxt->ctxt,
1427                                                          fd->subctxt, 0);
1428                 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1429                                                          uctxt->ctxt,
1430                                                          fd->subctxt, 0);
1431         }
1432         sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1433         if (copy_to_user(ubase, &binfo, sz))
1434                 ret = -EFAULT;
1435         return ret;
1436 }
1437
1438 static unsigned int poll_urgent(struct file *fp,
1439                                 struct poll_table_struct *pt)
1440 {
1441         struct hfi1_filedata *fd = fp->private_data;
1442         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1443         struct hfi1_devdata *dd = uctxt->dd;
1444         unsigned pollflag;
1445
1446         poll_wait(fp, &uctxt->wait, pt);
1447
1448         spin_lock_irq(&dd->uctxt_lock);
1449         if (uctxt->urgent != uctxt->urgent_poll) {
1450                 pollflag = POLLIN | POLLRDNORM;
1451                 uctxt->urgent_poll = uctxt->urgent;
1452         } else {
1453                 pollflag = 0;
1454                 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1455         }
1456         spin_unlock_irq(&dd->uctxt_lock);
1457
1458         return pollflag;
1459 }
1460
1461 static unsigned int poll_next(struct file *fp,
1462                               struct poll_table_struct *pt)
1463 {
1464         struct hfi1_filedata *fd = fp->private_data;
1465         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1466         struct hfi1_devdata *dd = uctxt->dd;
1467         unsigned pollflag;
1468
1469         poll_wait(fp, &uctxt->wait, pt);
1470
1471         spin_lock_irq(&dd->uctxt_lock);
1472         if (hdrqempty(uctxt)) {
1473                 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1474                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt);
1475                 pollflag = 0;
1476         } else {
1477                 pollflag = POLLIN | POLLRDNORM;
1478         }
1479         spin_unlock_irq(&dd->uctxt_lock);
1480
1481         return pollflag;
1482 }
1483
1484 /*
1485  * Find all user contexts in use, and set the specified bit in their
1486  * event mask.
1487  * See also find_ctxt() for a similar use, that is specific to send buffers.
1488  */
1489 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1490 {
1491         struct hfi1_ctxtdata *uctxt;
1492         struct hfi1_devdata *dd = ppd->dd;
1493         u16 ctxt;
1494
1495         if (!dd->events)
1496                 return -EINVAL;
1497
1498         for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
1499              ctxt++) {
1500                 uctxt = hfi1_rcd_get_by_index(dd, ctxt);
1501                 if (uctxt) {
1502                         unsigned long *evs = dd->events +
1503                                 (uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1504                                 HFI1_MAX_SHARED_CTXTS;
1505                         int i;
1506                         /*
1507                          * subctxt_cnt is 0 if not shared, so do base
1508                          * separately, first, then remaining subctxt, if any
1509                          */
1510                         set_bit(evtbit, evs);
1511                         for (i = 1; i < uctxt->subctxt_cnt; i++)
1512                                 set_bit(evtbit, evs + i);
1513                         hfi1_rcd_put(uctxt);
1514                 }
1515         }
1516
1517         return 0;
1518 }
1519
1520 /**
1521  * manage_rcvq - manage a context's receive queue
1522  * @uctxt: the context
1523  * @subctxt: the sub-context
1524  * @start_stop: action to carry out
1525  *
1526  * start_stop == 0 disables receive on the context, for use in queue
1527  * overflow conditions.  start_stop==1 re-enables, to be used to
1528  * re-init the software copy of the head register
1529  */
1530 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1531                        int start_stop)
1532 {
1533         struct hfi1_devdata *dd = uctxt->dd;
1534         unsigned int rcvctrl_op;
1535
1536         if (subctxt)
1537                 goto bail;
1538         /* atomically clear receive enable ctxt. */
1539         if (start_stop) {
1540                 /*
1541                  * On enable, force in-memory copy of the tail register to
1542                  * 0, so that protocol code doesn't have to worry about
1543                  * whether or not the chip has yet updated the in-memory
1544                  * copy or not on return from the system call. The chip
1545                  * always resets it's tail register back to 0 on a
1546                  * transition from disabled to enabled.
1547                  */
1548                 if (uctxt->rcvhdrtail_kvaddr)
1549                         clear_rcvhdrtail(uctxt);
1550                 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1551         } else {
1552                 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1553         }
1554         hfi1_rcvctrl(dd, rcvctrl_op, uctxt);
1555         /* always; new head should be equal to new tail; see above */
1556 bail:
1557         return 0;
1558 }
1559
1560 /*
1561  * clear the event notifier events for this context.
1562  * User process then performs actions appropriate to bit having been
1563  * set, if desired, and checks again in future.
1564  */
1565 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1566                           unsigned long events)
1567 {
1568         int i;
1569         struct hfi1_devdata *dd = uctxt->dd;
1570         unsigned long *evs;
1571
1572         if (!dd->events)
1573                 return 0;
1574
1575         evs = dd->events + ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1576                             HFI1_MAX_SHARED_CTXTS) + subctxt;
1577
1578         for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1579                 if (!test_bit(i, &events))
1580                         continue;
1581                 clear_bit(i, evs);
1582         }
1583         return 0;
1584 }
1585
1586 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, u16 subctxt, u16 pkey)
1587 {
1588         int ret = -ENOENT, i, intable = 0;
1589         struct hfi1_pportdata *ppd = uctxt->ppd;
1590         struct hfi1_devdata *dd = uctxt->dd;
1591
1592         if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1593                 ret = -EINVAL;
1594                 goto done;
1595         }
1596
1597         for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1598                 if (pkey == ppd->pkeys[i]) {
1599                         intable = 1;
1600                         break;
1601                 }
1602
1603         if (intable)
1604                 ret = hfi1_set_ctxt_pkey(dd, uctxt, pkey);
1605 done:
1606         return ret;
1607 }
1608
1609 static void user_remove(struct hfi1_devdata *dd)
1610 {
1611
1612         hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1613 }
1614
1615 static int user_add(struct hfi1_devdata *dd)
1616 {
1617         char name[10];
1618         int ret;
1619
1620         snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1621         ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1622                              &dd->user_cdev, &dd->user_device,
1623                              true, &dd->kobj);
1624         if (ret)
1625                 user_remove(dd);
1626
1627         return ret;
1628 }
1629
1630 /*
1631  * Create per-unit files in /dev
1632  */
1633 int hfi1_device_create(struct hfi1_devdata *dd)
1634 {
1635         return user_add(dd);
1636 }
1637
1638 /*
1639  * Remove per-unit files in /dev
1640  * void, core kernel returns no errors for this stuff
1641  */
1642 void hfi1_device_remove(struct hfi1_devdata *dd)
1643 {
1644         user_remove(dd);
1645 }