drivers/staging/rdma/hfi1/file_ops.c

   1 /*
   2  * Copyright(c) 2015, 2016 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47 #include <linux/poll.h>
  48 #include <linux/cdev.h>
  49 #include <linux/vmalloc.h>
  50 #include <linux/io.h>
  51
  52 #include <rdma/ib.h>
  53
  54 #include "hfi.h"
  55 #include "pio.h"
  56 #include "device.h"
  57 #include "common.h"
  58 #include "trace.h"
  59 #include "user_sdma.h"
  60 #include "user_exp_rcv.h"
  61 #include "eprom.h"
  62 #include "aspm.h"
  63 #include "mmu_rb.h"
  64
  65 #undef pr_fmt
  66 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
  67
  68 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
  69
  70 /*
  71  * File operation functions
  72  */
  73 static int hfi1_file_open(struct inode *, struct file *);
  74 static int hfi1_file_close(struct inode *, struct file *);
  75 static ssize_t hfi1_file_write(struct file *, const char __user *,
  76                                size_t, loff_t *);
  77 static ssize_t hfi1_write_iter(struct kiocb *, struct iov_iter *);
  78 static unsigned int hfi1_poll(struct file *, struct poll_table_struct *);
  79 static int hfi1_file_mmap(struct file *, struct vm_area_struct *);
  80
  81 static u64 kvirt_to_phys(void *);
  82 static int assign_ctxt(struct file *, struct hfi1_user_info *);
  83 static int init_subctxts(struct hfi1_ctxtdata *, const struct hfi1_user_info *);
  84 static int user_init(struct file *);
  85 static int get_ctxt_info(struct file *, void __user *, __u32);
  86 static int get_base_info(struct file *, void __user *, __u32);
  87 static int setup_ctxt(struct file *);
  88 static int setup_subctxt(struct hfi1_ctxtdata *);
  89 static int get_user_context(struct file *, struct hfi1_user_info *, int);
  90 static int find_shared_ctxt(struct file *, const struct hfi1_user_info *);
  91 static int allocate_ctxt(struct file *, struct hfi1_devdata *,
  92                          struct hfi1_user_info *);
  93 static unsigned int poll_urgent(struct file *, struct poll_table_struct *);
  94 static unsigned int poll_next(struct file *, struct poll_table_struct *);
  95 static int user_event_ack(struct hfi1_ctxtdata *, int, unsigned long);
  96 static int set_ctxt_pkey(struct hfi1_ctxtdata *, unsigned, u16);
  97 static int manage_rcvq(struct hfi1_ctxtdata *, unsigned, int);
  98 static int vma_fault(struct vm_area_struct *, struct vm_fault *);
  99
 100 static const struct file_operations hfi1_file_ops = {
 101         .owner = THIS_MODULE,
 102         .write = hfi1_file_write,
 103         .write_iter = hfi1_write_iter,
 104         .open = hfi1_file_open,
 105         .release = hfi1_file_close,
 106         .poll = hfi1_poll,
 107         .mmap = hfi1_file_mmap,
 108         .llseek = noop_llseek,
 109 };
 110
 111 static struct vm_operations_struct vm_ops = {
 112         .fault = vma_fault,
 113 };
 114
 115 /*
 116  * Types of memories mapped into user processes' space
 117  */
 118 enum mmap_types {
 119         PIO_BUFS = 1,
 120         PIO_BUFS_SOP,
 121         PIO_CRED,
 122         RCV_HDRQ,
 123         RCV_EGRBUF,
 124         UREGS,
 125         EVENTS,
 126         STATUS,
 127         RTAIL,
 128         SUBCTXT_UREGS,
 129         SUBCTXT_RCV_HDRQ,
 130         SUBCTXT_EGRBUF,
 131         SDMA_COMP
 132 };
 133
 134 /*
 135  * Masks and offsets defining the mmap tokens
 136  */
 137 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
 138 #define HFI1_MMAP_OFFSET_SHIFT  0
 139 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
 140 #define HFI1_MMAP_SUBCTXT_SHIFT 12
 141 #define HFI1_MMAP_CTXT_MASK     0xffULL
 142 #define HFI1_MMAP_CTXT_SHIFT    16
 143 #define HFI1_MMAP_TYPE_MASK     0xfULL
 144 #define HFI1_MMAP_TYPE_SHIFT    24
 145 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
 146 #define HFI1_MMAP_MAGIC_SHIFT   32
 147
 148 #define HFI1_MMAP_MAGIC         0xdabbad00
 149
 150 #define HFI1_MMAP_TOKEN_SET(field, val) \
 151         (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
 152 #define HFI1_MMAP_TOKEN_GET(field, token) \
 153         (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
 154 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
 155         (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
 156         HFI1_MMAP_TOKEN_SET(TYPE, type) | \
 157         HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
 158         HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
 159         HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
 160
 161 #define dbg(fmt, ...)                           \
 162         pr_info(fmt, ##__VA_ARGS__)
 163
 164 static inline int is_valid_mmap(u64 token)
 165 {
 166         return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
 167 }
 168
 169 static int hfi1_file_open(struct inode *inode, struct file *fp)
 170 {
 171         /* The real work is performed later in assign_ctxt() */
 172         fp->private_data = kzalloc(sizeof(struct hfi1_filedata), GFP_KERNEL);
 173         if (fp->private_data) /* no cpu affinity by default */
 174                 ((struct hfi1_filedata *)fp->private_data)->rec_cpu_num = -1;
 175         return fp->private_data ? 0 : -ENOMEM;
 176 }
 177
 178 static ssize_t hfi1_file_write(struct file *fp, const char __user *data,
 179                                size_t count, loff_t *offset)
 180 {
 181         const struct hfi1_cmd __user *ucmd;
 182         struct hfi1_filedata *fd = fp->private_data;
 183         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 184         struct hfi1_cmd cmd;
 185         struct hfi1_user_info uinfo;
 186         struct hfi1_tid_info tinfo;
 187         unsigned long addr;
 188         ssize_t consumed = 0, copy = 0, ret = 0;
 189         void *dest = NULL;
 190         __u64 user_val = 0;
 191         int uctxt_required = 1;
 192
 193         /* FIXME: This interface cannot continue out of staging */
 194         if (WARN_ON_ONCE(!ib_safe_file_access(fp)))
 195                 return -EACCES;
 196
 197         if (count < sizeof(cmd)) {
 198                 ret = -EINVAL;
 199                 goto bail;
 200         }
 201
 202         ucmd = (const struct hfi1_cmd __user *)data;
 203         if (copy_from_user(&cmd, ucmd, sizeof(cmd))) {
 204                 ret = -EFAULT;
 205                 goto bail;
 206         }
 207
 208         consumed = sizeof(cmd);
 209
 210         switch (cmd.type) {
 211         case HFI1_CMD_ASSIGN_CTXT:
 212                 uctxt_required = 0;     /* assigned user context not required */
 213                 copy = sizeof(uinfo);
 214                 dest = &uinfo;
 215                 break;
 216         case HFI1_CMD_SDMA_STATUS_UPD:
 217         case HFI1_CMD_CREDIT_UPD:
 218                 copy = 0;
 219                 break;
 220         case HFI1_CMD_TID_UPDATE:
 221         case HFI1_CMD_TID_FREE:
 222         case HFI1_CMD_TID_INVAL_READ:
 223                 copy = sizeof(tinfo);
 224                 dest = &tinfo;
 225                 break;
 226         case HFI1_CMD_USER_INFO:
 227         case HFI1_CMD_RECV_CTRL:
 228         case HFI1_CMD_POLL_TYPE:
 229         case HFI1_CMD_ACK_EVENT:
 230         case HFI1_CMD_CTXT_INFO:
 231         case HFI1_CMD_SET_PKEY:
 232         case HFI1_CMD_CTXT_RESET:
 233                 copy = 0;
 234                 user_val = cmd.addr;
 235                 break;
 236         default:
 237                 ret = -EINVAL;
 238                 goto bail;
 239         }
 240
 241         /* If the command comes with user data, copy it. */
 242         if (copy) {
 243                 if (copy_from_user(dest, (void __user *)cmd.addr, copy)) {
 244                         ret = -EFAULT;
 245                         goto bail;
 246                 }
 247                 consumed += copy;
 248         }
 249
 250         /*
 251          * Make sure there is a uctxt when needed.
 252          */
 253         if (uctxt_required && !uctxt) {
 254                 ret = -EINVAL;
 255                 goto bail;
 256         }
 257
 258         switch (cmd.type) {
 259         case HFI1_CMD_ASSIGN_CTXT:
 260                 ret = assign_ctxt(fp, &uinfo);
 261                 if (ret < 0)
 262                         goto bail;
 263                 ret = setup_ctxt(fp);
 264                 if (ret)
 265                         goto bail;
 266                 ret = user_init(fp);
 267                 break;
 268         case HFI1_CMD_CTXT_INFO:
 269                 ret = get_ctxt_info(fp, (void __user *)(unsigned long)
 270                                     user_val, cmd.len);
 271                 break;
 272         case HFI1_CMD_USER_INFO:
 273                 ret = get_base_info(fp, (void __user *)(unsigned long)
 274                                     user_val, cmd.len);
 275                 break;
 276         case HFI1_CMD_SDMA_STATUS_UPD:
 277                 break;
 278         case HFI1_CMD_CREDIT_UPD:
 279                 if (uctxt && uctxt->sc)
 280                         sc_return_credits(uctxt->sc);
 281                 break;
 282         case HFI1_CMD_TID_UPDATE:
 283                 ret = hfi1_user_exp_rcv_setup(fp, &tinfo);
 284                 if (!ret) {
 285                         /*
 286                          * Copy the number of tidlist entries we used
 287                          * and the length of the buffer we registered.
 288                          * These fields are adjacent in the structure so
 289                          * we can copy them at the same time.
 290                          */
 291                         addr = (unsigned long)cmd.addr +
 292                                 offsetof(struct hfi1_tid_info, tidcnt);
 293                         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 294                                          sizeof(tinfo.tidcnt) +
 295                                          sizeof(tinfo.length)))
 296                                 ret = -EFAULT;
 297                 }
 298                 break;
 299         case HFI1_CMD_TID_INVAL_READ:
 300                 ret = hfi1_user_exp_rcv_invalid(fp, &tinfo);
 301                 if (ret)
 302                         break;
 303                 addr = (unsigned long)cmd.addr +
 304                         offsetof(struct hfi1_tid_info, tidcnt);
 305                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 306                                  sizeof(tinfo.tidcnt)))
 307                         ret = -EFAULT;
 308                 break;
 309         case HFI1_CMD_TID_FREE:
 310                 ret = hfi1_user_exp_rcv_clear(fp, &tinfo);
 311                 if (ret)
 312                         break;
 313                 addr = (unsigned long)cmd.addr +
 314                         offsetof(struct hfi1_tid_info, tidcnt);
 315                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 316                                  sizeof(tinfo.tidcnt)))
 317                         ret = -EFAULT;
 318                 break;
 319         case HFI1_CMD_RECV_CTRL:
 320                 ret = manage_rcvq(uctxt, fd->subctxt, (int)user_val);
 321                 break;
 322         case HFI1_CMD_POLL_TYPE:
 323                 uctxt->poll_type = (typeof(uctxt->poll_type))user_val;
 324                 break;
 325         case HFI1_CMD_ACK_EVENT:
 326                 ret = user_event_ack(uctxt, fd->subctxt, user_val);
 327                 break;
 328         case HFI1_CMD_SET_PKEY:
 329                 if (HFI1_CAP_IS_USET(PKEY_CHECK))
 330                         ret = set_ctxt_pkey(uctxt, fd->subctxt, user_val);
 331                 else
 332                         ret = -EPERM;
 333                 break;
 334         case HFI1_CMD_CTXT_RESET: {
 335                 struct send_context *sc;
 336                 struct hfi1_devdata *dd;
 337
 338                 if (!uctxt || !uctxt->dd || !uctxt->sc) {
 339                         ret = -EINVAL;
 340                         break;
 341                 }
 342                 /*
 343                  * There is no protection here. User level has to
 344                  * guarantee that no one will be writing to the send
 345                  * context while it is being re-initialized.
 346                  * If user level breaks that guarantee, it will break
 347                  * it's own context and no one else's.
 348                  */
 349                 dd = uctxt->dd;
 350                 sc = uctxt->sc;
 351                 /*
 352                  * Wait until the interrupt handler has marked the
 353                  * context as halted or frozen. Report error if we time
 354                  * out.
 355                  */
 356                 wait_event_interruptible_timeout(
 357                         sc->halt_wait, (sc->flags & SCF_HALTED),
 358                         msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 359                 if (!(sc->flags & SCF_HALTED)) {
 360                         ret = -ENOLCK;
 361                         break;
 362                 }
 363                 /*
 364                  * If the send context was halted due to a Freeze,
 365                  * wait until the device has been "unfrozen" before
 366                  * resetting the context.
 367                  */
 368                 if (sc->flags & SCF_FROZEN) {
 369                         wait_event_interruptible_timeout(
 370                                 dd->event_queue,
 371                                 !(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
 372                                 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 373                         if (dd->flags & HFI1_FROZEN) {
 374                                 ret = -ENOLCK;
 375                                 break;
 376                         }
 377                         if (dd->flags & HFI1_FORCED_FREEZE) {
 378                                 /*
 379                                  * Don't allow context reset if we are into
 380                                  * forced freeze
 381                                  */
 382                                 ret = -ENODEV;
 383                                 break;
 384                         }
 385                         sc_disable(sc);
 386                         ret = sc_enable(sc);
 387                         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
 388                                      uctxt->ctxt);
 389                 } else {
 390                         ret = sc_restart(sc);
 391                 }
 392                 if (!ret)
 393                         sc_return_credits(sc);
 394                 break;
 395         }
 396         }
 397
 398         if (ret >= 0)
 399                 ret = consumed;
 400 bail:
 401         return ret;
 402 }
 403
 404 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
 405 {
 406         struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
 407         struct hfi1_user_sdma_pkt_q *pq = fd->pq;
 408         struct hfi1_user_sdma_comp_q *cq = fd->cq;
 409         int ret = 0, done = 0, reqs = 0;
 410         unsigned long dim = from->nr_segs;
 411
 412         if (!cq || !pq) {
 413                 ret = -EIO;
 414                 goto done;
 415         }
 416
 417         if (!iter_is_iovec(from) || !dim) {
 418                 ret = -EINVAL;
 419                 goto done;
 420         }
 421
 422         hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
 423                   fd->uctxt->ctxt, fd->subctxt, dim);
 424
 425         if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
 426                 ret = -ENOSPC;
 427                 goto done;
 428         }
 429
 430         while (dim) {
 431                 unsigned long count = 0;
 432
 433                 ret = hfi1_user_sdma_process_request(
 434                         kiocb->ki_filp, (struct iovec *)(from->iov + done),
 435                         dim, &count);
 436                 if (ret)
 437                         goto done;
 438                 dim -= count;
 439                 done += count;
 440                 reqs++;
 441         }
 442 done:
 443         return ret ? ret : reqs;
 444 }
 445
 446 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
 447 {
 448         struct hfi1_filedata *fd = fp->private_data;
 449         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 450         struct hfi1_devdata *dd;
 451         unsigned long flags, pfn;
 452         u64 token = vma->vm_pgoff << PAGE_SHIFT,
 453                 memaddr = 0;
 454         u8 subctxt, mapio = 0, vmf = 0, type;
 455         ssize_t memlen = 0;
 456         int ret = 0;
 457         u16 ctxt;
 458
 459         if (!is_valid_mmap(token) || !uctxt ||
 460             !(vma->vm_flags & VM_SHARED)) {
 461                 ret = -EINVAL;
 462                 goto done;
 463         }
 464         dd = uctxt->dd;
 465         ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
 466         subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
 467         type = HFI1_MMAP_TOKEN_GET(TYPE, token);
 468         if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
 469                 ret = -EINVAL;
 470                 goto done;
 471         }
 472
 473         flags = vma->vm_flags;
 474
 475         switch (type) {
 476         case PIO_BUFS:
 477         case PIO_BUFS_SOP:
 478                 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
 479                                 /* chip pio base */
 480                            (uctxt->sc->hw_context * BIT(16))) +
 481                                 /* 64K PIO space / ctxt */
 482                         (type == PIO_BUFS_SOP ?
 483                                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
 484                 /*
 485                  * Map only the amount allocated to the context, not the
 486                  * entire available context's PIO space.
 487                  */
 488                 memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
 489                 flags &= ~VM_MAYREAD;
 490                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 491                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 492                 mapio = 1;
 493                 break;
 494         case PIO_CRED:
 495                 if (flags & VM_WRITE) {
 496                         ret = -EPERM;
 497                         goto done;
 498                 }
 499                 /*
 500                  * The credit return location for this context could be on the
 501                  * second or third page allocated for credit returns (if number
 502                  * of enabled contexts > 64 and 128 respectively).
 503                  */
 504                 memaddr = dd->cr_base[uctxt->numa_id].pa +
 505                         (((u64)uctxt->sc->hw_free -
 506                           (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
 507                 memlen = PAGE_SIZE;
 508                 flags &= ~VM_MAYWRITE;
 509                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 510                 /*
 511                  * The driver has already allocated memory for credit
 512                  * returns and programmed it into the chip. Has that
 513                  * memory been flagged as non-cached?
 514                  */
 515                 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
 516                 mapio = 1;
 517                 break;
 518         case RCV_HDRQ:
 519                 memaddr = uctxt->rcvhdrq_phys;
 520                 memlen = uctxt->rcvhdrq_size;
 521                 break;
 522         case RCV_EGRBUF: {
 523                 unsigned long addr;
 524                 int i;
 525                 /*
 526                  * The RcvEgr buffer need to be handled differently
 527                  * as multiple non-contiguous pages need to be mapped
 528                  * into the user process.
 529                  */
 530                 memlen = uctxt->egrbufs.size;
 531                 if ((vma->vm_end - vma->vm_start) != memlen) {
 532                         dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
 533                                    (vma->vm_end - vma->vm_start), memlen);
 534                         ret = -EINVAL;
 535                         goto done;
 536                 }
 537                 if (vma->vm_flags & VM_WRITE) {
 538                         ret = -EPERM;
 539                         goto done;
 540                 }
 541                 vma->vm_flags &= ~VM_MAYWRITE;
 542                 addr = vma->vm_start;
 543                 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
 544                         ret = remap_pfn_range(
 545                                 vma, addr,
 546                                 uctxt->egrbufs.buffers[i].phys >> PAGE_SHIFT,
 547                                 uctxt->egrbufs.buffers[i].len,
 548                                 vma->vm_page_prot);
 549                         if (ret < 0)
 550                                 goto done;
 551                         addr += uctxt->egrbufs.buffers[i].len;
 552                 }
 553                 ret = 0;
 554                 goto done;
 555         }
 556         case UREGS:
 557                 /*
 558                  * Map only the page that contains this context's user
 559                  * registers.
 560                  */
 561                 memaddr = (unsigned long)
 562                         (dd->physaddr + RXE_PER_CONTEXT_USER)
 563                         + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
 564                 /*
 565                  * TidFlow table is on the same page as the rest of the
 566                  * user registers.
 567                  */
 568                 memlen = PAGE_SIZE;
 569                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 570                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 571                 mapio = 1;
 572                 break;
 573         case EVENTS:
 574                 /*
 575                  * Use the page where this context's flags are. User level
 576                  * knows where it's own bitmap is within the page.
 577                  */
 578                 memaddr = (unsigned long)(dd->events +
 579                                           ((uctxt->ctxt - dd->first_user_ctxt) *
 580                                            HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
 581                 memlen = PAGE_SIZE;
 582                 /*
 583                  * v3.7 removes VM_RESERVED but the effect is kept by
 584                  * using VM_IO.
 585                  */
 586                 flags |= VM_IO | VM_DONTEXPAND;
 587                 vmf = 1;
 588                 break;
 589         case STATUS:
 590                 memaddr = kvirt_to_phys((void *)dd->status);
 591                 memlen = PAGE_SIZE;
 592                 flags |= VM_IO | VM_DONTEXPAND;
 593                 break;
 594         case RTAIL:
 595                 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
 596                         /*
 597                          * If the memory allocation failed, the context alloc
 598                          * also would have failed, so we would never get here
 599                          */
 600                         ret = -EINVAL;
 601                         goto done;
 602                 }
 603                 if (flags & VM_WRITE) {
 604                         ret = -EPERM;
 605                         goto done;
 606                 }
 607                 memaddr = uctxt->rcvhdrqtailaddr_phys;
 608                 memlen = PAGE_SIZE;
 609                 flags &= ~VM_MAYWRITE;
 610                 break;
 611         case SUBCTXT_UREGS:
 612                 memaddr = (u64)uctxt->subctxt_uregbase;
 613                 memlen = PAGE_SIZE;
 614                 flags |= VM_IO | VM_DONTEXPAND;
 615                 vmf = 1;
 616                 break;
 617         case SUBCTXT_RCV_HDRQ:
 618                 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
 619                 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
 620                 flags |= VM_IO | VM_DONTEXPAND;
 621                 vmf = 1;
 622                 break;
 623         case SUBCTXT_EGRBUF:
 624                 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
 625                 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
 626                 flags |= VM_IO | VM_DONTEXPAND;
 627                 flags &= ~VM_MAYWRITE;
 628                 vmf = 1;
 629                 break;
 630         case SDMA_COMP: {
 631                 struct hfi1_user_sdma_comp_q *cq = fd->cq;
 632
 633                 if (!cq) {
 634                         ret = -EFAULT;
 635                         goto done;
 636                 }
 637                 memaddr = (u64)cq->comps;
 638                 memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
 639                 flags |= VM_IO | VM_DONTEXPAND;
 640                 vmf = 1;
 641                 break;
 642         }
 643         default:
 644                 ret = -EINVAL;
 645                 break;
 646         }
 647
 648         if ((vma->vm_end - vma->vm_start) != memlen) {
 649                 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
 650                           uctxt->ctxt, fd->subctxt,
 651                           (vma->vm_end - vma->vm_start), memlen);
 652                 ret = -EINVAL;
 653                 goto done;
 654         }
 655
 656         vma->vm_flags = flags;
 657         hfi1_cdbg(PROC,
 658                   "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
 659                     ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
 660                     vma->vm_end - vma->vm_start, vma->vm_flags);
 661         pfn = (unsigned long)(memaddr >> PAGE_SHIFT);
 662         if (vmf) {
 663                 vma->vm_pgoff = pfn;
 664                 vma->vm_ops = &vm_ops;
 665                 ret = 0;
 666         } else if (mapio) {
 667                 ret = io_remap_pfn_range(vma, vma->vm_start, pfn, memlen,
 668                                          vma->vm_page_prot);
 669         } else {
 670                 ret = remap_pfn_range(vma, vma->vm_start, pfn, memlen,
 671                                       vma->vm_page_prot);
 672         }
 673 done:
 674         return ret;
 675 }
 676
 677 /*
 678  * Local (non-chip) user memory is not mapped right away but as it is
 679  * accessed by the user-level code.
 680  */
 681 static int vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 682 {
 683         struct page *page;
 684
 685         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 686         if (!page)
 687                 return VM_FAULT_SIGBUS;
 688
 689         get_page(page);
 690         vmf->page = page;
 691
 692         return 0;
 693 }
 694
 695 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
 696 {
 697         struct hfi1_ctxtdata *uctxt;
 698         unsigned pollflag;
 699
 700         uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
 701         if (!uctxt)
 702                 pollflag = POLLERR;
 703         else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
 704                 pollflag = poll_urgent(fp, pt);
 705         else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
 706                 pollflag = poll_next(fp, pt);
 707         else /* invalid */
 708                 pollflag = POLLERR;
 709
 710         return pollflag;
 711 }
 712
 713 static int hfi1_file_close(struct inode *inode, struct file *fp)
 714 {
 715         struct hfi1_filedata *fdata = fp->private_data;
 716         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 717         struct hfi1_devdata *dd;
 718         unsigned long flags, *ev;
 719
 720         fp->private_data = NULL;
 721
 722         if (!uctxt)
 723                 goto done;
 724
 725         hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
 726         dd = uctxt->dd;
 727         mutex_lock(&hfi1_mutex);
 728
 729         flush_wc();
 730         /* drain user sdma queue */
 731         hfi1_user_sdma_free_queues(fdata);
 732
 733         /* release the cpu */
 734         hfi1_put_proc_affinity(dd, fdata->rec_cpu_num);
 735
 736         /*
 737          * Clear any left over, unhandled events so the next process that
 738          * gets this context doesn't get confused.
 739          */
 740         ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
 741                            HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
 742         *ev = 0;
 743
 744         if (--uctxt->cnt) {
 745                 uctxt->active_slaves &= ~(1 << fdata->subctxt);
 746                 uctxt->subpid[fdata->subctxt] = 0;
 747                 mutex_unlock(&hfi1_mutex);
 748                 goto done;
 749         }
 750
 751         spin_lock_irqsave(&dd->uctxt_lock, flags);
 752         /*
 753          * Disable receive context and interrupt available, reset all
 754          * RcvCtxtCtrl bits to default values.
 755          */
 756         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
 757                      HFI1_RCVCTRL_TIDFLOW_DIS |
 758                      HFI1_RCVCTRL_INTRAVAIL_DIS |
 759                      HFI1_RCVCTRL_TAILUPD_DIS |
 760                      HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
 761                      HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
 762                      HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
 763         /* Clear the context's J_KEY */
 764         hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
 765         /*
 766          * Reset context integrity checks to default.
 767          * (writes to CSRs probably belong in chip.c)
 768          */
 769         write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
 770                         hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
 771         sc_disable(uctxt->sc);
 772         uctxt->pid = 0;
 773         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 774
 775         dd->rcd[uctxt->ctxt] = NULL;
 776
 777         hfi1_user_exp_rcv_free(fdata);
 778         hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
 779
 780         uctxt->rcvwait_to = 0;
 781         uctxt->piowait_to = 0;
 782         uctxt->rcvnowait = 0;
 783         uctxt->pionowait = 0;
 784         uctxt->event_flags = 0;
 785
 786         hfi1_stats.sps_ctxts--;
 787         if (++dd->freectxts == dd->num_user_contexts)
 788                 aspm_enable_all(dd);
 789         mutex_unlock(&hfi1_mutex);
 790         hfi1_free_ctxtdata(dd, uctxt);
 791 done:
 792         kfree(fdata);
 793         return 0;
 794 }
 795
 796 /*
 797  * Convert kernel *virtual* addresses to physical addresses.
 798  * This is used to vmalloc'ed addresses.
 799  */
 800 static u64 kvirt_to_phys(void *addr)
 801 {
 802         struct page *page;
 803         u64 paddr = 0;
 804
 805         page = vmalloc_to_page(addr);
 806         if (page)
 807                 paddr = page_to_pfn(page) << PAGE_SHIFT;
 808
 809         return paddr;
 810 }
 811
 812 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
 813 {
 814         int i_minor, ret = 0;
 815         unsigned int swmajor, swminor;
 816
 817         swmajor = uinfo->userversion >> 16;
 818         if (swmajor != HFI1_USER_SWMAJOR) {
 819                 ret = -ENODEV;
 820                 goto done;
 821         }
 822
 823         swminor = uinfo->userversion & 0xffff;
 824
 825         mutex_lock(&hfi1_mutex);
 826         /* First, lets check if we need to setup a shared context? */
 827         if (uinfo->subctxt_cnt) {
 828                 struct hfi1_filedata *fd = fp->private_data;
 829
 830                 ret = find_shared_ctxt(fp, uinfo);
 831                 if (ret < 0)
 832                         goto done_unlock;
 833                 if (ret)
 834                         fd->rec_cpu_num = hfi1_get_proc_affinity(
 835                                 fd->uctxt->dd, fd->uctxt->numa_id);
 836         }
 837
 838         /*
 839          * We execute the following block if we couldn't find a
 840          * shared context or if context sharing is not required.
 841          */
 842         if (!ret) {
 843                 i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
 844                 ret = get_user_context(fp, uinfo, i_minor);
 845         }
 846 done_unlock:
 847         mutex_unlock(&hfi1_mutex);
 848 done:
 849         return ret;
 850 }
 851
 852 static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
 853                             int devno)
 854 {
 855         struct hfi1_devdata *dd = NULL;
 856         int devmax, npresent, nup;
 857
 858         devmax = hfi1_count_units(&npresent, &nup);
 859         if (!npresent)
 860                 return -ENXIO;
 861
 862         if (!nup)
 863                 return -ENETDOWN;
 864
 865         dd = hfi1_lookup(devno);
 866         if (!dd)
 867                 return -ENODEV;
 868         else if (!dd->freectxts)
 869                 return -EBUSY;
 870
 871         return allocate_ctxt(fp, dd, uinfo);
 872 }
 873
 874 static int find_shared_ctxt(struct file *fp,
 875                             const struct hfi1_user_info *uinfo)
 876 {
 877         int devmax, ndev, i;
 878         int ret = 0;
 879         struct hfi1_filedata *fd = fp->private_data;
 880
 881         devmax = hfi1_count_units(NULL, NULL);
 882
 883         for (ndev = 0; ndev < devmax; ndev++) {
 884                 struct hfi1_devdata *dd = hfi1_lookup(ndev);
 885
 886                 if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
 887                         continue;
 888                 for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
 889                         struct hfi1_ctxtdata *uctxt = dd->rcd[i];
 890
 891                         /* Skip ctxts which are not yet open */
 892                         if (!uctxt || !uctxt->cnt)
 893                                 continue;
 894                         /* Skip ctxt if it doesn't match the requested one */
 895                         if (memcmp(uctxt->uuid, uinfo->uuid,
 896                                    sizeof(uctxt->uuid)) ||
 897                             uctxt->jkey != generate_jkey(current_uid()) ||
 898                             uctxt->subctxt_id != uinfo->subctxt_id ||
 899                             uctxt->subctxt_cnt != uinfo->subctxt_cnt)
 900                                 continue;
 901
 902                         /* Verify the sharing process matches the master */
 903                         if (uctxt->userversion != uinfo->userversion ||
 904                             uctxt->cnt >= uctxt->subctxt_cnt) {
 905                                 ret = -EINVAL;
 906                                 goto done;
 907                         }
 908                         fd->uctxt = uctxt;
 909                         fd->subctxt  = uctxt->cnt++;
 910                         uctxt->subpid[fd->subctxt] = current->pid;
 911                         uctxt->active_slaves |= 1 << fd->subctxt;
 912                         ret = 1;
 913                         goto done;
 914                 }
 915         }
 916
 917 done:
 918         return ret;
 919 }
 920
 921 static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
 922                          struct hfi1_user_info *uinfo)
 923 {
 924         struct hfi1_filedata *fd = fp->private_data;
 925         struct hfi1_ctxtdata *uctxt;
 926         unsigned ctxt;
 927         int ret, numa;
 928
 929         if (dd->flags & HFI1_FROZEN) {
 930                 /*
 931                  * Pick an error that is unique from all other errors
 932                  * that are returned so the user process knows that
 933                  * it tried to allocate while the SPC was frozen.  It
 934                  * it should be able to retry with success in a short
 935                  * while.
 936                  */
 937                 return -EIO;
 938         }
 939
 940         for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
 941                 if (!dd->rcd[ctxt])
 942                         break;
 943
 944         if (ctxt == dd->num_rcv_contexts)
 945                 return -EBUSY;
 946
 947         fd->rec_cpu_num = hfi1_get_proc_affinity(dd, -1);
 948         if (fd->rec_cpu_num != -1)
 949                 numa = cpu_to_node(fd->rec_cpu_num);
 950         else
 951                 numa = numa_node_id();
 952         uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
 953         if (!uctxt) {
 954                 dd_dev_err(dd,
 955                            "Unable to allocate ctxtdata memory, failing open\n");
 956                 return -ENOMEM;
 957         }
 958         hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
 959                   uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
 960                   uctxt->numa_id);
 961
 962         /*
 963          * Allocate and enable a PIO send context.
 964          */
 965         uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
 966                              uctxt->dd->node);
 967         if (!uctxt->sc)
 968                 return -ENOMEM;
 969
 970         hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
 971                   uctxt->sc->hw_context);
 972         ret = sc_enable(uctxt->sc);
 973         if (ret)
 974                 return ret;
 975         /*
 976          * Setup shared context resources if the user-level has requested
 977          * shared contexts and this is the 'master' process.
 978          * This has to be done here so the rest of the sub-contexts find the
 979          * proper master.
 980          */
 981         if (uinfo->subctxt_cnt && !fd->subctxt) {
 982                 ret = init_subctxts(uctxt, uinfo);
 983                 /*
 984                  * On error, we don't need to disable and de-allocate the
 985                  * send context because it will be done during file close
 986                  */
 987                 if (ret)
 988                         return ret;
 989         }
 990         uctxt->userversion = uinfo->userversion;
 991         uctxt->pid = current->pid;
 992         uctxt->flags = HFI1_CAP_UGET(MASK);
 993         init_waitqueue_head(&uctxt->wait);
 994         strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
 995         memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
 996         uctxt->jkey = generate_jkey(current_uid());
 997         INIT_LIST_HEAD(&uctxt->sdma_queues);
 998         spin_lock_init(&uctxt->sdma_qlock);
 999         hfi1_stats.sps_ctxts++;
1000         /*
1001          * Disable ASPM when there are open user/PSM contexts to avoid
1002          * issues with ASPM L1 exit latency
1003          */
1004         if (dd->freectxts-- == dd->num_user_contexts)
1005                 aspm_disable_all(dd);
1006         fd->uctxt = uctxt;
1007
1008         return 0;
1009 }
1010
1011 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1012                          const struct hfi1_user_info *uinfo)
1013 {
1014         unsigned num_subctxts;
1015
1016         num_subctxts = uinfo->subctxt_cnt;
1017         if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1018                 return -EINVAL;
1019
1020         uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1021         uctxt->subctxt_id = uinfo->subctxt_id;
1022         uctxt->active_slaves = 1;
1023         uctxt->redirect_seq_cnt = 1;
1024         set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1025
1026         return 0;
1027 }
1028
1029 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1030 {
1031         int ret = 0;
1032         unsigned num_subctxts = uctxt->subctxt_cnt;
1033
1034         uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1035         if (!uctxt->subctxt_uregbase) {
1036                 ret = -ENOMEM;
1037                 goto bail;
1038         }
1039         /* We can take the size of the RcvHdr Queue from the master */
1040         uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1041                                                   num_subctxts);
1042         if (!uctxt->subctxt_rcvhdr_base) {
1043                 ret = -ENOMEM;
1044                 goto bail_ureg;
1045         }
1046
1047         uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1048                                                 num_subctxts);
1049         if (!uctxt->subctxt_rcvegrbuf) {
1050                 ret = -ENOMEM;
1051                 goto bail_rhdr;
1052         }
1053         goto bail;
1054 bail_rhdr:
1055         vfree(uctxt->subctxt_rcvhdr_base);
1056 bail_ureg:
1057         vfree(uctxt->subctxt_uregbase);
1058         uctxt->subctxt_uregbase = NULL;
1059 bail:
1060         return ret;
1061 }
1062
1063 static int user_init(struct file *fp)
1064 {
1065         unsigned int rcvctrl_ops = 0;
1066         struct hfi1_filedata *fd = fp->private_data;
1067         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1068
1069         /* make sure that the context has already been setup */
1070         if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags))
1071                 return -EFAULT;
1072
1073         /* initialize poll variables... */
1074         uctxt->urgent = 0;
1075         uctxt->urgent_poll = 0;
1076
1077         /*
1078          * Now enable the ctxt for receive.
1079          * For chips that are set to DMA the tail register to memory
1080          * when they change (and when the update bit transitions from
1081          * 0 to 1.  So for those chips, we turn it off and then back on.
1082          * This will (very briefly) affect any other open ctxts, but the
1083          * duration is very short, and therefore isn't an issue.  We
1084          * explicitly set the in-memory tail copy to 0 beforehand, so we
1085          * don't have to wait to be sure the DMA update has happened
1086          * (chip resets head/tail to 0 on transition to enable).
1087          */
1088         if (uctxt->rcvhdrtail_kvaddr)
1089                 clear_rcvhdrtail(uctxt);
1090
1091         /* Setup J_KEY before enabling the context */
1092         hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1093
1094         rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1095         if (HFI1_CAP_KGET_MASK(uctxt->flags, HDRSUPP))
1096                 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1097         /*
1098          * Ignore the bit in the flags for now until proper
1099          * support for multiple packet per rcv array entry is
1100          * added.
1101          */
1102         if (!HFI1_CAP_KGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1103                 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1104         if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1105                 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1106         if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1107                 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1108         /*
1109          * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1110          * We can't rely on the correct value to be set from prior
1111          * uses of the chip or ctxt. Therefore, add the rcvctrl op
1112          * for both cases.
1113          */
1114         if (HFI1_CAP_KGET_MASK(uctxt->flags, DMA_RTAIL))
1115                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1116         else
1117                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1118         hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1119
1120         /* Notify any waiting slaves */
1121         if (uctxt->subctxt_cnt) {
1122                 clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1123                 wake_up(&uctxt->wait);
1124         }
1125
1126         return 0;
1127 }
1128
1129 static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1130 {
1131         struct hfi1_ctxt_info cinfo;
1132         struct hfi1_filedata *fd = fp->private_data;
1133         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1134         int ret = 0;
1135
1136         memset(&cinfo, 0, sizeof(cinfo));
1137         ret = hfi1_get_base_kinfo(uctxt, &cinfo);
1138         if (ret < 0)
1139                 goto done;
1140         cinfo.num_active = hfi1_count_active_units();
1141         cinfo.unit = uctxt->dd->unit;
1142         cinfo.ctxt = uctxt->ctxt;
1143         cinfo.subctxt = fd->subctxt;
1144         cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1145                                 uctxt->dd->rcv_entries.group_size) +
1146                 uctxt->expected_count;
1147         cinfo.credits = uctxt->sc->credits;
1148         cinfo.numa_node = uctxt->numa_id;
1149         cinfo.rec_cpu = fd->rec_cpu_num;
1150         cinfo.send_ctxt = uctxt->sc->hw_context;
1151
1152         cinfo.egrtids = uctxt->egrbufs.alloced;
1153         cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1154         cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1155         cinfo.sdma_ring_size = fd->cq->nentries;
1156         cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1157
1158         trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1159         if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1160                 ret = -EFAULT;
1161 done:
1162         return ret;
1163 }
1164
1165 static int setup_ctxt(struct file *fp)
1166 {
1167         struct hfi1_filedata *fd = fp->private_data;
1168         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1169         struct hfi1_devdata *dd = uctxt->dd;
1170         int ret = 0;
1171
1172         /*
1173          * Context should be set up only once, including allocation and
1174          * programming of eager buffers. This is done if context sharing
1175          * is not requested or by the master process.
1176          */
1177         if (!uctxt->subctxt_cnt || !fd->subctxt) {
1178                 ret = hfi1_init_ctxt(uctxt->sc);
1179                 if (ret)
1180                         goto done;
1181
1182                 /* Now allocate the RcvHdr queue and eager buffers. */
1183                 ret = hfi1_create_rcvhdrq(dd, uctxt);
1184                 if (ret)
1185                         goto done;
1186                 ret = hfi1_setup_eagerbufs(uctxt);
1187                 if (ret)
1188                         goto done;
1189                 if (uctxt->subctxt_cnt && !fd->subctxt) {
1190                         ret = setup_subctxt(uctxt);
1191                         if (ret)
1192                                 goto done;
1193                 }
1194         } else {
1195                 ret = wait_event_interruptible(uctxt->wait, !test_bit(
1196                                                HFI1_CTXT_MASTER_UNINIT,
1197                                                &uctxt->event_flags));
1198                 if (ret)
1199                         goto done;
1200         }
1201
1202         ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1203         if (ret)
1204                 goto done;
1205         /*
1206          * Expected receive has to be setup for all processes (including
1207          * shared contexts). However, it has to be done after the master
1208          * context has been fully configured as it depends on the
1209          * eager/expected split of the RcvArray entries.
1210          * Setting it up here ensures that the subcontexts will be waiting
1211          * (due to the above wait_event_interruptible() until the master
1212          * is setup.
1213          */
1214         ret = hfi1_user_exp_rcv_init(fp);
1215         if (ret)
1216                 goto done;
1217
1218         set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1219 done:
1220         return ret;
1221 }
1222
1223 static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1224 {
1225         struct hfi1_base_info binfo;
1226         struct hfi1_filedata *fd = fp->private_data;
1227         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1228         struct hfi1_devdata *dd = uctxt->dd;
1229         ssize_t sz;
1230         unsigned offset;
1231         int ret = 0;
1232
1233         trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1234
1235         memset(&binfo, 0, sizeof(binfo));
1236         binfo.hw_version = dd->revision;
1237         binfo.sw_version = HFI1_KERN_SWVERSION;
1238         binfo.bthqp = kdeth_qp;
1239         binfo.jkey = uctxt->jkey;
1240         /*
1241          * If more than 64 contexts are enabled the allocated credit
1242          * return will span two or three contiguous pages. Since we only
1243          * map the page containing the context's credit return address,
1244          * we need to calculate the offset in the proper page.
1245          */
1246         offset = ((u64)uctxt->sc->hw_free -
1247                   (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1248         binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1249                                                 fd->subctxt, offset);
1250         binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1251                                             fd->subctxt,
1252                                             uctxt->sc->base_addr);
1253         binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1254                                                 uctxt->ctxt,
1255                                                 fd->subctxt,
1256                                                 uctxt->sc->base_addr);
1257         binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1258                                                fd->subctxt,
1259                                                uctxt->rcvhdrq);
1260         binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1261                                                fd->subctxt,
1262                                                uctxt->egrbufs.rcvtids[0].phys);
1263         binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1264                                                  fd->subctxt, 0);
1265         /*
1266          * user regs are at
1267          * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1268          */
1269         binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1270                                             fd->subctxt, 0);
1271         offset = offset_in_page((((uctxt->ctxt - dd->first_user_ctxt) *
1272                     HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1273                   sizeof(*dd->events));
1274         binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1275                                               fd->subctxt,
1276                                               offset);
1277         binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1278                                               fd->subctxt,
1279                                               dd->status);
1280         if (HFI1_CAP_IS_USET(DMA_RTAIL))
1281                 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1282                                                        fd->subctxt, 0);
1283         if (uctxt->subctxt_cnt) {
1284                 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1285                                                         uctxt->ctxt,
1286                                                         fd->subctxt, 0);
1287                 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1288                                                          uctxt->ctxt,
1289                                                          fd->subctxt, 0);
1290                 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1291                                                          uctxt->ctxt,
1292                                                          fd->subctxt, 0);
1293         }
1294         sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1295         if (copy_to_user(ubase, &binfo, sz))
1296                 ret = -EFAULT;
1297         return ret;
1298 }
1299
1300 static unsigned int poll_urgent(struct file *fp,
1301                                 struct poll_table_struct *pt)
1302 {
1303         struct hfi1_filedata *fd = fp->private_data;
1304         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1305         struct hfi1_devdata *dd = uctxt->dd;
1306         unsigned pollflag;
1307
1308         poll_wait(fp, &uctxt->wait, pt);
1309
1310         spin_lock_irq(&dd->uctxt_lock);
1311         if (uctxt->urgent != uctxt->urgent_poll) {
1312                 pollflag = POLLIN | POLLRDNORM;
1313                 uctxt->urgent_poll = uctxt->urgent;
1314         } else {
1315                 pollflag = 0;
1316                 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1317         }
1318         spin_unlock_irq(&dd->uctxt_lock);
1319
1320         return pollflag;
1321 }
1322
1323 static unsigned int poll_next(struct file *fp,
1324                               struct poll_table_struct *pt)
1325 {
1326         struct hfi1_filedata *fd = fp->private_data;
1327         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1328         struct hfi1_devdata *dd = uctxt->dd;
1329         unsigned pollflag;
1330
1331         poll_wait(fp, &uctxt->wait, pt);
1332
1333         spin_lock_irq(&dd->uctxt_lock);
1334         if (hdrqempty(uctxt)) {
1335                 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1336                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1337                 pollflag = 0;
1338         } else {
1339                 pollflag = POLLIN | POLLRDNORM;
1340         }
1341         spin_unlock_irq(&dd->uctxt_lock);
1342
1343         return pollflag;
1344 }
1345
1346 /*
1347  * Find all user contexts in use, and set the specified bit in their
1348  * event mask.
1349  * See also find_ctxt() for a similar use, that is specific to send buffers.
1350  */
1351 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1352 {
1353         struct hfi1_ctxtdata *uctxt;
1354         struct hfi1_devdata *dd = ppd->dd;
1355         unsigned ctxt;
1356         int ret = 0;
1357         unsigned long flags;
1358
1359         if (!dd->events) {
1360                 ret = -EINVAL;
1361                 goto done;
1362         }
1363
1364         spin_lock_irqsave(&dd->uctxt_lock, flags);
1365         for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1366              ctxt++) {
1367                 uctxt = dd->rcd[ctxt];
1368                 if (uctxt) {
1369                         unsigned long *evs = dd->events +
1370                                 (uctxt->ctxt - dd->first_user_ctxt) *
1371                                 HFI1_MAX_SHARED_CTXTS;
1372                         int i;
1373                         /*
1374                          * subctxt_cnt is 0 if not shared, so do base
1375                          * separately, first, then remaining subctxt, if any
1376                          */
1377                         set_bit(evtbit, evs);
1378                         for (i = 1; i < uctxt->subctxt_cnt; i++)
1379                                 set_bit(evtbit, evs + i);
1380                 }
1381         }
1382         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1383 done:
1384         return ret;
1385 }
1386
1387 /**
1388  * manage_rcvq - manage a context's receive queue
1389  * @uctxt: the context
1390  * @subctxt: the sub-context
1391  * @start_stop: action to carry out
1392  *
1393  * start_stop == 0 disables receive on the context, for use in queue
1394  * overflow conditions.  start_stop==1 re-enables, to be used to
1395  * re-init the software copy of the head register
1396  */
1397 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1398                        int start_stop)
1399 {
1400         struct hfi1_devdata *dd = uctxt->dd;
1401         unsigned int rcvctrl_op;
1402
1403         if (subctxt)
1404                 goto bail;
1405         /* atomically clear receive enable ctxt. */
1406         if (start_stop) {
1407                 /*
1408                  * On enable, force in-memory copy of the tail register to
1409                  * 0, so that protocol code doesn't have to worry about
1410                  * whether or not the chip has yet updated the in-memory
1411                  * copy or not on return from the system call. The chip
1412                  * always resets it's tail register back to 0 on a
1413                  * transition from disabled to enabled.
1414                  */
1415                 if (uctxt->rcvhdrtail_kvaddr)
1416                         clear_rcvhdrtail(uctxt);
1417                 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1418         } else {
1419                 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1420         }
1421         hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1422         /* always; new head should be equal to new tail; see above */
1423 bail:
1424         return 0;
1425 }
1426
1427 /*
1428  * clear the event notifier events for this context.
1429  * User process then performs actions appropriate to bit having been
1430  * set, if desired, and checks again in future.
1431  */
1432 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1433                           unsigned long events)
1434 {
1435         int i;
1436         struct hfi1_devdata *dd = uctxt->dd;
1437         unsigned long *evs;
1438
1439         if (!dd->events)
1440                 return 0;
1441
1442         evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1443                             HFI1_MAX_SHARED_CTXTS) + subctxt;
1444
1445         for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1446                 if (!test_bit(i, &events))
1447                         continue;
1448                 clear_bit(i, evs);
1449         }
1450         return 0;
1451 }
1452
1453 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1454                          u16 pkey)
1455 {
1456         int ret = -ENOENT, i, intable = 0;
1457         struct hfi1_pportdata *ppd = uctxt->ppd;
1458         struct hfi1_devdata *dd = uctxt->dd;
1459
1460         if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1461                 ret = -EINVAL;
1462                 goto done;
1463         }
1464
1465         for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1466                 if (pkey == ppd->pkeys[i]) {
1467                         intable = 1;
1468                         break;
1469                 }
1470
1471         if (intable)
1472                 ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1473 done:
1474         return ret;
1475 }
1476
1477 static void user_remove(struct hfi1_devdata *dd)
1478 {
1479
1480         hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1481 }
1482
1483 static int user_add(struct hfi1_devdata *dd)
1484 {
1485         char name[10];
1486         int ret;
1487
1488         snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1489         ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1490                              &dd->user_cdev, &dd->user_device,
1491                              true);
1492         if (ret)
1493                 user_remove(dd);
1494
1495         return ret;
1496 }
1497
1498 /*
1499  * Create per-unit files in /dev
1500  */
1501 int hfi1_device_create(struct hfi1_devdata *dd)
1502 {
1503         int r, ret;
1504
1505         r = user_add(dd);
1506         ret = hfi1_diag_add(dd);
1507         if (r && !ret)
1508                 ret = r;
1509         return ret;
1510 }
1511
1512 /*
1513  * Remove per-unit files in /dev
1514  * void, core kernel returns no errors for this stuff
1515  */
1516 void hfi1_device_remove(struct hfi1_devdata *dd)
1517 {
1518         user_remove(dd);
1519         hfi1_diag_remove(dd);
1520 }