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