drivers/staging/rdma/hfi1/file_ops.c

   1 /*
   2  *
   3  * This file is provided under a dual BSD/GPLv2 license.  When using or
   4  * redistributing this file, you may do so under either license.
   5  *
   6  * GPL LICENSE SUMMARY
   7  *
   8  * Copyright(c) 2015 Intel Corporation.
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of version 2 of the GNU General Public License as
  12  * published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it will be useful, but
  15  * WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * General Public License for more details.
  18  *
  19  * BSD LICENSE
  20  *
  21  * Copyright(c) 2015 Intel Corporation.
  22  *
  23  * Redistribution and use in source and binary forms, with or without
  24  * modification, are permitted provided that the following conditions
  25  * are met:
  26  *
  27  *  - Redistributions of source code must retain the above copyright
  28  *    notice, this list of conditions and the following disclaimer.
  29  *  - Redistributions in binary form must reproduce the above copyright
  30  *    notice, this list of conditions and the following disclaimer in
  31  *    the documentation and/or other materials provided with the
  32  *    distribution.
  33  *  - Neither the name of Intel Corporation nor the names of its
  34  *    contributors may be used to endorse or promote products derived
  35  *    from this software without specific prior written permission.
  36  *
  37  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  38  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  39  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  40  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  41  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  42  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  43  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  44  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  45  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  46  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  47  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  48  *
  49  */
  50 #include <linux/pci.h>
  51 #include <linux/poll.h>
  52 #include <linux/cdev.h>
  53 #include <linux/swap.h>
  54 #include <linux/vmalloc.h>
  55 #include <linux/highmem.h>
  56 #include <linux/io.h>
  57 #include <linux/jiffies.h>
  58 #include <asm/pgtable.h>
  59 #include <linux/delay.h>
  60 #include <linux/export.h>
  61 #include <linux/module.h>
  62 #include <linux/cred.h>
  63 #include <linux/uio.h>
  64
  65 #include "hfi.h"
  66 #include "pio.h"
  67 #include "device.h"
  68 #include "common.h"
  69 #include "trace.h"
  70 #include "user_sdma.h"
  71 #include "eprom.h"
  72
  73 #undef pr_fmt
  74 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
  75
  76 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
  77
  78 /*
  79  * File operation functions
  80  */
  81 static int hfi1_file_open(struct inode *, struct file *);
  82 static int hfi1_file_close(struct inode *, struct file *);
  83 static ssize_t hfi1_file_write(struct file *, const char __user *,
  84                                size_t, loff_t *);
  85 static ssize_t hfi1_write_iter(struct kiocb *, struct iov_iter *);
  86 static unsigned int hfi1_poll(struct file *, struct poll_table_struct *);
  87 static int hfi1_file_mmap(struct file *, struct vm_area_struct *);
  88
  89 static u64 kvirt_to_phys(void *);
  90 static int assign_ctxt(struct file *, struct hfi1_user_info *);
  91 static int init_subctxts(struct hfi1_ctxtdata *, const struct hfi1_user_info *);
  92 static int user_init(struct file *);
  93 static int get_ctxt_info(struct file *, void __user *, __u32);
  94 static int get_base_info(struct file *, void __user *, __u32);
  95 static int setup_ctxt(struct file *);
  96 static int setup_subctxt(struct hfi1_ctxtdata *);
  97 static int get_user_context(struct file *, struct hfi1_user_info *,
  98                             int, unsigned);
  99 static int find_shared_ctxt(struct file *, const struct hfi1_user_info *);
 100 static int allocate_ctxt(struct file *, struct hfi1_devdata *,
 101                          struct hfi1_user_info *);
 102 static unsigned int poll_urgent(struct file *, struct poll_table_struct *);
 103 static unsigned int poll_next(struct file *, struct poll_table_struct *);
 104 static int user_event_ack(struct hfi1_ctxtdata *, int, unsigned long);
 105 static int set_ctxt_pkey(struct hfi1_ctxtdata *, unsigned, u16);
 106 static int manage_rcvq(struct hfi1_ctxtdata *, unsigned, int);
 107 static int vma_fault(struct vm_area_struct *, struct vm_fault *);
 108 static int exp_tid_setup(struct file *, struct hfi1_tid_info *);
 109 static int exp_tid_free(struct file *, struct hfi1_tid_info *);
 110 static void unlock_exp_tids(struct hfi1_ctxtdata *);
 111
 112 static const struct file_operations hfi1_file_ops = {
 113         .owner = THIS_MODULE,
 114         .write = hfi1_file_write,
 115         .write_iter = hfi1_write_iter,
 116         .open = hfi1_file_open,
 117         .release = hfi1_file_close,
 118         .poll = hfi1_poll,
 119         .mmap = hfi1_file_mmap,
 120         .llseek = noop_llseek,
 121 };
 122
 123 static 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, ((unsigned long)addr & ~PAGE_MASK)))
 172
 173 #define EXP_TID_SET(field, value)                       \
 174         (((value) & EXP_TID_TID##field##_MASK) <<       \
 175          EXP_TID_TID##field##_SHIFT)
 176 #define EXP_TID_CLEAR(tid, field) {                                     \
 177                 (tid) &= ~(EXP_TID_TID##field##_MASK <<                 \
 178                            EXP_TID_TID##field##_SHIFT);                 \
 179                         }
 180 #define EXP_TID_RESET(tid, field, value) do {                           \
 181                 EXP_TID_CLEAR(tid, field);                              \
 182                 (tid) |= EXP_TID_SET(field, value);                     \
 183         } while (0)
 184
 185 #define dbg(fmt, ...)                           \
 186         pr_info(fmt, ##__VA_ARGS__)
 187
 188
 189 static inline int is_valid_mmap(u64 token)
 190 {
 191         return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
 192 }
 193
 194 static int hfi1_file_open(struct inode *inode, struct file *fp)
 195 {
 196         /* The real work is performed later in assign_ctxt() */
 197         fp->private_data = kzalloc(sizeof(struct hfi1_filedata), GFP_KERNEL);
 198         if (fp->private_data) /* no cpu affinity by default */
 199                 ((struct hfi1_filedata *)fp->private_data)->rec_cpu_num = -1;
 200         return fp->private_data ? 0 : -ENOMEM;
 201 }
 202
 203 static ssize_t hfi1_file_write(struct file *fp, const char __user *data,
 204                                size_t count, loff_t *offset)
 205 {
 206         const struct hfi1_cmd __user *ucmd;
 207         struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
 208         struct hfi1_cmd cmd;
 209         struct hfi1_user_info uinfo;
 210         struct hfi1_tid_info tinfo;
 211         ssize_t consumed = 0, copy = 0, ret = 0;
 212         void *dest = NULL;
 213         __u64 user_val = 0;
 214         int uctxt_required = 1;
 215         int must_be_root = 0;
 216
 217         if (count < sizeof(cmd)) {
 218                 ret = -EINVAL;
 219                 goto bail;
 220         }
 221
 222         ucmd = (const struct hfi1_cmd __user *)data;
 223         if (copy_from_user(&cmd, ucmd, sizeof(cmd))) {
 224                 ret = -EFAULT;
 225                 goto bail;
 226         }
 227
 228         consumed = sizeof(cmd);
 229
 230         switch (cmd.type) {
 231         case HFI1_CMD_ASSIGN_CTXT:
 232                 uctxt_required = 0;     /* assigned user context not required */
 233                 copy = sizeof(uinfo);
 234                 dest = &uinfo;
 235                 break;
 236         case HFI1_CMD_SDMA_STATUS_UPD:
 237         case HFI1_CMD_CREDIT_UPD:
 238                 copy = 0;
 239                 break;
 240         case HFI1_CMD_TID_UPDATE:
 241         case HFI1_CMD_TID_FREE:
 242                 copy = sizeof(tinfo);
 243                 dest = &tinfo;
 244                 break;
 245         case HFI1_CMD_USER_INFO:
 246         case HFI1_CMD_RECV_CTRL:
 247         case HFI1_CMD_POLL_TYPE:
 248         case HFI1_CMD_ACK_EVENT:
 249         case HFI1_CMD_CTXT_INFO:
 250         case HFI1_CMD_SET_PKEY:
 251         case HFI1_CMD_CTXT_RESET:
 252                 copy = 0;
 253                 user_val = cmd.addr;
 254                 break;
 255         case HFI1_CMD_EP_INFO:
 256         case HFI1_CMD_EP_ERASE_CHIP:
 257         case HFI1_CMD_EP_ERASE_P0:
 258         case HFI1_CMD_EP_ERASE_P1:
 259         case HFI1_CMD_EP_READ_P0:
 260         case HFI1_CMD_EP_READ_P1:
 261         case HFI1_CMD_EP_WRITE_P0:
 262         case HFI1_CMD_EP_WRITE_P1:
 263                 uctxt_required = 0;     /* assigned user context not required */
 264                 must_be_root = 1;       /* validate user */
 265                 copy = 0;
 266                 break;
 267         default:
 268                 ret = -EINVAL;
 269                 goto bail;
 270         }
 271
 272         /* If the command comes with user data, copy it. */
 273         if (copy) {
 274                 if (copy_from_user(dest, (void __user *)cmd.addr, copy)) {
 275                         ret = -EFAULT;
 276                         goto bail;
 277                 }
 278                 consumed += copy;
 279         }
 280
 281         /*
 282          * Make sure there is a uctxt when needed.
 283          */
 284         if (uctxt_required && !uctxt) {
 285                 ret = -EINVAL;
 286                 goto bail;
 287         }
 288
 289         /* only root can do these operations */
 290         if (must_be_root && !capable(CAP_SYS_ADMIN)) {
 291                 ret = -EPERM;
 292                 goto bail;
 293         }
 294
 295         switch (cmd.type) {
 296         case HFI1_CMD_ASSIGN_CTXT:
 297                 ret = assign_ctxt(fp, &uinfo);
 298                 if (ret < 0)
 299                         goto bail;
 300                 ret = setup_ctxt(fp);
 301                 if (ret)
 302                         goto bail;
 303                 ret = user_init(fp);
 304                 break;
 305         case HFI1_CMD_CTXT_INFO:
 306                 ret = get_ctxt_info(fp, (void __user *)(unsigned long)
 307                                     user_val, cmd.len);
 308                 break;
 309         case HFI1_CMD_USER_INFO:
 310                 ret = get_base_info(fp, (void __user *)(unsigned long)
 311                                     user_val, cmd.len);
 312                 break;
 313         case HFI1_CMD_SDMA_STATUS_UPD:
 314                 break;
 315         case HFI1_CMD_CREDIT_UPD:
 316                 if (uctxt && uctxt->sc)
 317                         sc_return_credits(uctxt->sc);
 318                 break;
 319         case HFI1_CMD_TID_UPDATE:
 320                 ret = exp_tid_setup(fp, &tinfo);
 321                 if (!ret) {
 322                         unsigned long addr;
 323                         /*
 324                          * Copy the number of tidlist entries we used
 325                          * and the length of the buffer we registered.
 326                          * These fields are adjacent in the structure so
 327                          * we can copy them at the same time.
 328                          */
 329                         addr = (unsigned long)cmd.addr +
 330                                 offsetof(struct hfi1_tid_info, tidcnt);
 331                         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 332                                          sizeof(tinfo.tidcnt) +
 333                                          sizeof(tinfo.length)))
 334                                 ret = -EFAULT;
 335                 }
 336                 break;
 337         case HFI1_CMD_TID_FREE:
 338                 ret = exp_tid_free(fp, &tinfo);
 339                 break;
 340         case HFI1_CMD_RECV_CTRL:
 341                 ret = manage_rcvq(uctxt, subctxt_fp(fp), (int)user_val);
 342                 break;
 343         case HFI1_CMD_POLL_TYPE:
 344                 uctxt->poll_type = (typeof(uctxt->poll_type))user_val;
 345                 break;
 346         case HFI1_CMD_ACK_EVENT:
 347                 ret = user_event_ack(uctxt, subctxt_fp(fp), user_val);
 348                 break;
 349         case HFI1_CMD_SET_PKEY:
 350                 if (HFI1_CAP_IS_USET(PKEY_CHECK))
 351                         ret = set_ctxt_pkey(uctxt, subctxt_fp(fp), user_val);
 352                 else
 353                         ret = -EPERM;
 354                 break;
 355         case HFI1_CMD_CTXT_RESET: {
 356                 struct send_context *sc;
 357                 struct hfi1_devdata *dd;
 358
 359                 if (!uctxt || !uctxt->dd || !uctxt->sc) {
 360                         ret = -EINVAL;
 361                         break;
 362                 }
 363                 /*
 364                  * There is no protection here. User level has to
 365                  * guarantee that no one will be writing to the send
 366                  * context while it is being re-initialized.
 367                  * If user level breaks that guarantee, it will break
 368                  * it's own context and no one else's.
 369                  */
 370                 dd = uctxt->dd;
 371                 sc = uctxt->sc;
 372                 /*
 373                  * Wait until the interrupt handler has marked the
 374                  * context as halted or frozen. Report error if we time
 375                  * out.
 376                  */
 377                 wait_event_interruptible_timeout(
 378                         sc->halt_wait, (sc->flags & SCF_HALTED),
 379                         msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 380                 if (!(sc->flags & SCF_HALTED)) {
 381                         ret = -ENOLCK;
 382                         break;
 383                 }
 384                 /*
 385                  * If the send context was halted due to a Freeze,
 386                  * wait until the device has been "unfrozen" before
 387                  * resetting the context.
 388                  */
 389                 if (sc->flags & SCF_FROZEN) {
 390                         wait_event_interruptible_timeout(
 391                                 dd->event_queue,
 392                                 !(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
 393                                 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 394                         if (dd->flags & HFI1_FROZEN) {
 395                                 ret = -ENOLCK;
 396                                 break;
 397                         }
 398                         if (dd->flags & HFI1_FORCED_FREEZE) {
 399                                 /* Don't allow context reset if we are into
 400                                  * forced freeze */
 401                                 ret = -ENODEV;
 402                                 break;
 403                         }
 404                         sc_disable(sc);
 405                         ret = sc_enable(sc);
 406                         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
 407                                      uctxt->ctxt);
 408                 } else
 409                         ret = sc_restart(sc);
 410                 if (!ret)
 411                         sc_return_credits(sc);
 412                 break;
 413         }
 414         case HFI1_CMD_EP_INFO:
 415         case HFI1_CMD_EP_ERASE_CHIP:
 416         case HFI1_CMD_EP_ERASE_P0:
 417         case HFI1_CMD_EP_ERASE_P1:
 418         case HFI1_CMD_EP_READ_P0:
 419         case HFI1_CMD_EP_READ_P1:
 420         case HFI1_CMD_EP_WRITE_P0:
 421         case HFI1_CMD_EP_WRITE_P1:
 422                 ret = handle_eprom_command(&cmd);
 423                 break;
 424         }
 425
 426         if (ret >= 0)
 427                 ret = consumed;
 428 bail:
 429         return ret;
 430 }
 431
 432 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
 433 {
 434         struct hfi1_user_sdma_pkt_q *pq;
 435         struct hfi1_user_sdma_comp_q *cq;
 436         int ret = 0, done = 0, reqs = 0;
 437         unsigned long dim = from->nr_segs;
 438
 439         if (!user_sdma_comp_fp(kiocb->ki_filp) ||
 440             !user_sdma_pkt_fp(kiocb->ki_filp)) {
 441                 ret = -EIO;
 442                 goto done;
 443         }
 444
 445         if (!iter_is_iovec(from) || !dim) {
 446                 ret = -EINVAL;
 447                 goto done;
 448         }
 449
 450         hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
 451                   ctxt_fp(kiocb->ki_filp)->ctxt, subctxt_fp(kiocb->ki_filp),
 452                   dim);
 453         pq = user_sdma_pkt_fp(kiocb->ki_filp);
 454         cq = user_sdma_comp_fp(kiocb->ki_filp);
 455
 456         if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
 457                 ret = -ENOSPC;
 458                 goto done;
 459         }
 460
 461         while (dim) {
 462                 unsigned long count = 0;
 463
 464                 ret = hfi1_user_sdma_process_request(
 465                         kiocb->ki_filp, (struct iovec *)(from->iov + done),
 466                         dim, &count);
 467                 if (ret)
 468                         goto done;
 469                 dim -= count;
 470                 done += count;
 471                 reqs++;
 472         }
 473 done:
 474         return ret ? ret : reqs;
 475 }
 476
 477 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
 478 {
 479         struct hfi1_ctxtdata *uctxt;
 480         struct hfi1_devdata *dd;
 481         unsigned long flags, pfn;
 482         u64 token = vma->vm_pgoff << PAGE_SHIFT,
 483                 memaddr = 0;
 484         u8 subctxt, mapio = 0, vmf = 0, type;
 485         ssize_t memlen = 0;
 486         int ret = 0;
 487         u16 ctxt;
 488
 489         uctxt = ctxt_fp(fp);
 490         if (!is_valid_mmap(token) || !uctxt ||
 491             !(vma->vm_flags & VM_SHARED)) {
 492                 ret = -EINVAL;
 493                 goto done;
 494         }
 495         dd = uctxt->dd;
 496         ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
 497         subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
 498         type = HFI1_MMAP_TOKEN_GET(TYPE, token);
 499         if (ctxt != uctxt->ctxt || subctxt != subctxt_fp(fp)) {
 500                 ret = -EINVAL;
 501                 goto done;
 502         }
 503
 504         flags = vma->vm_flags;
 505
 506         switch (type) {
 507         case PIO_BUFS:
 508         case PIO_BUFS_SOP:
 509                 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
 510                                 /* chip pio base */
 511                            (uctxt->sc->hw_context * (1 << 16))) +
 512                                 /* 64K PIO space / ctxt */
 513                         (type == PIO_BUFS_SOP ?
 514                                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
 515                 /*
 516                  * Map only the amount allocated to the context, not the
 517                  * entire available context's PIO space.
 518                  */
 519                 memlen = ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE,
 520                                PAGE_SIZE);
 521                 flags &= ~VM_MAYREAD;
 522                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 523                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 524                 mapio = 1;
 525                 break;
 526         case PIO_CRED:
 527                 if (flags & VM_WRITE) {
 528                         ret = -EPERM;
 529                         goto done;
 530                 }
 531                 /*
 532                  * The credit return location for this context could be on the
 533                  * second or third page allocated for credit returns (if number
 534                  * of enabled contexts > 64 and 128 respectively).
 535                  */
 536                 memaddr = dd->cr_base[uctxt->numa_id].pa +
 537                         (((u64)uctxt->sc->hw_free -
 538                           (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
 539                 memlen = PAGE_SIZE;
 540                 flags &= ~VM_MAYWRITE;
 541                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 542                 /*
 543                  * The driver has already allocated memory for credit
 544                  * returns and programmed it into the chip. Has that
 545                  * memory been flagged as non-cached?
 546                  */
 547                 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
 548                 mapio = 1;
 549                 break;
 550         case RCV_HDRQ:
 551                 memaddr = uctxt->rcvhdrq_phys;
 552                 memlen = uctxt->rcvhdrq_size;
 553                 break;
 554         case RCV_EGRBUF: {
 555                 unsigned long addr;
 556                 int i;
 557                 /*
 558                  * The RcvEgr buffer need to be handled differently
 559                  * as multiple non-contiguous pages need to be mapped
 560                  * into the user process.
 561                  */
 562                 memlen = uctxt->egrbufs.size;
 563                 if ((vma->vm_end - vma->vm_start) != memlen) {
 564                         dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
 565                                    (vma->vm_end - vma->vm_start), memlen);
 566                         ret = -EINVAL;
 567                         goto done;
 568                 }
 569                 if (vma->vm_flags & VM_WRITE) {
 570                         ret = -EPERM;
 571                         goto done;
 572                 }
 573                 vma->vm_flags &= ~VM_MAYWRITE;
 574                 addr = vma->vm_start;
 575                 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
 576                         ret = remap_pfn_range(
 577                                 vma, addr,
 578                                 uctxt->egrbufs.buffers[i].phys >> PAGE_SHIFT,
 579                                 uctxt->egrbufs.buffers[i].len,
 580                                 vma->vm_page_prot);
 581                         if (ret < 0)
 582                                 goto done;
 583                         addr += uctxt->egrbufs.buffers[i].len;
 584                 }
 585                 ret = 0;
 586                 goto done;
 587         }
 588         case UREGS:
 589                 /*
 590                  * Map only the page that contains this context's user
 591                  * registers.
 592                  */
 593                 memaddr = (unsigned long)
 594                         (dd->physaddr + RXE_PER_CONTEXT_USER)
 595                         + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
 596                 /*
 597                  * TidFlow table is on the same page as the rest of the
 598                  * user registers.
 599                  */
 600                 memlen = PAGE_SIZE;
 601                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 602                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 603                 mapio = 1;
 604                 break;
 605         case EVENTS:
 606                 /*
 607                  * Use the page where this context's flags are. User level
 608                  * knows where it's own bitmap is within the page.
 609                  */
 610                 memaddr = ((unsigned long)dd->events +
 611                            ((uctxt->ctxt - dd->first_user_ctxt) *
 612                             HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
 613                 memlen = PAGE_SIZE;
 614                 /*
 615                  * v3.7 removes VM_RESERVED but the effect is kept by
 616                  * using VM_IO.
 617                  */
 618                 flags |= VM_IO | VM_DONTEXPAND;
 619                 vmf = 1;
 620                 break;
 621         case STATUS:
 622                 memaddr = kvirt_to_phys((void *)dd->status);
 623                 memlen = PAGE_SIZE;
 624                 flags |= VM_IO | VM_DONTEXPAND;
 625                 break;
 626         case RTAIL:
 627                 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
 628                         /*
 629                          * If the memory allocation failed, the context alloc
 630                          * also would have failed, so we would never get here
 631                          */
 632                         ret = -EINVAL;
 633                         goto done;
 634                 }
 635                 if (flags & VM_WRITE) {
 636                         ret = -EPERM;
 637                         goto done;
 638                 }
 639                 memaddr = uctxt->rcvhdrqtailaddr_phys;
 640                 memlen = PAGE_SIZE;
 641                 flags &= ~VM_MAYWRITE;
 642                 break;
 643         case SUBCTXT_UREGS:
 644                 memaddr = (u64)uctxt->subctxt_uregbase;
 645                 memlen = PAGE_SIZE;
 646                 flags |= VM_IO | VM_DONTEXPAND;
 647                 vmf = 1;
 648                 break;
 649         case SUBCTXT_RCV_HDRQ:
 650                 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
 651                 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
 652                 flags |= VM_IO | VM_DONTEXPAND;
 653                 vmf = 1;
 654                 break;
 655         case SUBCTXT_EGRBUF:
 656                 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
 657                 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
 658                 flags |= VM_IO | VM_DONTEXPAND;
 659                 flags &= ~VM_MAYWRITE;
 660                 vmf = 1;
 661                 break;
 662         case SDMA_COMP: {
 663                 struct hfi1_user_sdma_comp_q *cq;
 664
 665                 if (!user_sdma_comp_fp(fp)) {
 666                         ret = -EFAULT;
 667                         goto done;
 668                 }
 669                 cq = user_sdma_comp_fp(fp);
 670                 memaddr = (u64)cq->comps;
 671                 memlen = ALIGN(sizeof(*cq->comps) * cq->nentries, PAGE_SIZE);
 672                 flags |= VM_IO | VM_DONTEXPAND;
 673                 vmf = 1;
 674                 break;
 675         }
 676         default:
 677                 ret = -EINVAL;
 678                 break;
 679         }
 680
 681         if ((vma->vm_end - vma->vm_start) != memlen) {
 682                 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
 683                           uctxt->ctxt, subctxt_fp(fp),
 684                           (vma->vm_end - vma->vm_start), memlen);
 685                 ret = -EINVAL;
 686                 goto done;
 687         }
 688
 689         vma->vm_flags = flags;
 690         dd_dev_info(dd,
 691                     "%s: %u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
 692                     __func__, ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
 693                     vma->vm_end - vma->vm_start, vma->vm_flags);
 694         pfn = (unsigned long)(memaddr >> PAGE_SHIFT);
 695         if (vmf) {
 696                 vma->vm_pgoff = pfn;
 697                 vma->vm_ops = &vm_ops;
 698                 ret = 0;
 699         } else if (mapio) {
 700                 ret = io_remap_pfn_range(vma, vma->vm_start, pfn, memlen,
 701                                          vma->vm_page_prot);
 702         } else {
 703                 ret = remap_pfn_range(vma, vma->vm_start, pfn, memlen,
 704                                       vma->vm_page_prot);
 705         }
 706 done:
 707         return ret;
 708 }
 709
 710 /*
 711  * Local (non-chip) user memory is not mapped right away but as it is
 712  * accessed by the user-level code.
 713  */
 714 static int vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 715 {
 716         struct page *page;
 717
 718         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 719         if (!page)
 720                 return VM_FAULT_SIGBUS;
 721
 722         get_page(page);
 723         vmf->page = page;
 724
 725         return 0;
 726 }
 727
 728 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
 729 {
 730         struct hfi1_ctxtdata *uctxt;
 731         unsigned pollflag;
 732
 733         uctxt = ctxt_fp(fp);
 734         if (!uctxt)
 735                 pollflag = POLLERR;
 736         else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
 737                 pollflag = poll_urgent(fp, pt);
 738         else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
 739                 pollflag = poll_next(fp, pt);
 740         else /* invalid */
 741                 pollflag = POLLERR;
 742
 743         return pollflag;
 744 }
 745
 746 static int hfi1_file_close(struct inode *inode, struct file *fp)
 747 {
 748         struct hfi1_filedata *fdata = fp->private_data;
 749         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 750         struct hfi1_devdata *dd;
 751         unsigned long flags, *ev;
 752
 753         fp->private_data = NULL;
 754
 755         if (!uctxt)
 756                 goto done;
 757
 758         hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
 759         dd = uctxt->dd;
 760         mutex_lock(&hfi1_mutex);
 761
 762         flush_wc();
 763         /* drain user sdma queue */
 764         if (fdata->pq)
 765                 hfi1_user_sdma_free_queues(fdata);
 766
 767         /*
 768          * Clear any left over, unhandled events so the next process that
 769          * gets this context doesn't get confused.
 770          */
 771         ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
 772                            HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
 773         *ev = 0;
 774
 775         if (--uctxt->cnt) {
 776                 uctxt->active_slaves &= ~(1 << fdata->subctxt);
 777                 uctxt->subpid[fdata->subctxt] = 0;
 778                 mutex_unlock(&hfi1_mutex);
 779                 goto done;
 780         }
 781
 782         spin_lock_irqsave(&dd->uctxt_lock, flags);
 783         /*
 784          * Disable receive context and interrupt available, reset all
 785          * RcvCtxtCtrl bits to default values.
 786          */
 787         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
 788                      HFI1_RCVCTRL_TIDFLOW_DIS |
 789                      HFI1_RCVCTRL_INTRAVAIL_DIS |
 790                      HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
 791                      HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
 792                      HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
 793         /* Clear the context's J_KEY */
 794         hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
 795         /*
 796          * Reset context integrity checks to default.
 797          * (writes to CSRs probably belong in chip.c)
 798          */
 799         write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
 800                         hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
 801         sc_disable(uctxt->sc);
 802         uctxt->pid = 0;
 803         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 804
 805         dd->rcd[uctxt->ctxt] = NULL;
 806         uctxt->rcvwait_to = 0;
 807         uctxt->piowait_to = 0;
 808         uctxt->rcvnowait = 0;
 809         uctxt->pionowait = 0;
 810         uctxt->event_flags = 0;
 811
 812         hfi1_clear_tids(uctxt);
 813         hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
 814
 815         if (uctxt->tid_pg_list)
 816                 unlock_exp_tids(uctxt);
 817
 818         hfi1_stats.sps_ctxts--;
 819         dd->freectxts++;
 820         mutex_unlock(&hfi1_mutex);
 821         hfi1_free_ctxtdata(dd, uctxt);
 822 done:
 823         kfree(fdata);
 824         return 0;
 825 }
 826
 827 /*
 828  * Convert kernel *virtual* addresses to physical addresses.
 829  * This is used to vmalloc'ed addresses.
 830  */
 831 static u64 kvirt_to_phys(void *addr)
 832 {
 833         struct page *page;
 834         u64 paddr = 0;
 835
 836         page = vmalloc_to_page(addr);
 837         if (page)
 838                 paddr = page_to_pfn(page) << PAGE_SHIFT;
 839
 840         return paddr;
 841 }
 842
 843 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
 844 {
 845         int i_minor, ret = 0;
 846         unsigned swmajor, swminor, alg = HFI1_ALG_ACROSS;
 847
 848         swmajor = uinfo->userversion >> 16;
 849         if (swmajor != HFI1_USER_SWMAJOR) {
 850                 ret = -ENODEV;
 851                 goto done;
 852         }
 853
 854         swminor = uinfo->userversion & 0xffff;
 855
 856         if (uinfo->hfi1_alg < HFI1_ALG_COUNT)
 857                 alg = uinfo->hfi1_alg;
 858
 859         mutex_lock(&hfi1_mutex);
 860         /* First, lets check if we need to setup a shared context? */
 861         if (uinfo->subctxt_cnt)
 862                 ret = find_shared_ctxt(fp, uinfo);
 863
 864         /*
 865          * We execute the following block if we couldn't find a
 866          * shared context or if context sharing is not required.
 867          */
 868         if (!ret) {
 869                 i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
 870                 ret = get_user_context(fp, uinfo, i_minor - 1, alg);
 871         }
 872         mutex_unlock(&hfi1_mutex);
 873 done:
 874         return ret;
 875 }
 876
 877 static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
 878                             int devno, unsigned alg)
 879 {
 880         struct hfi1_devdata *dd = NULL;
 881         int ret = 0, devmax, npresent, nup, dev;
 882
 883         devmax = hfi1_count_units(&npresent, &nup);
 884         if (!npresent) {
 885                 ret = -ENXIO;
 886                 goto done;
 887         }
 888         if (!nup) {
 889                 ret = -ENETDOWN;
 890                 goto done;
 891         }
 892         if (devno >= 0) {
 893                 dd = hfi1_lookup(devno);
 894                 if (!dd)
 895                         ret = -ENODEV;
 896                 else if (!dd->freectxts)
 897                         ret = -EBUSY;
 898         } else {
 899                 struct hfi1_devdata *pdd;
 900
 901                 if (alg == HFI1_ALG_ACROSS) {
 902                         unsigned free = 0U;
 903
 904                         for (dev = 0; dev < devmax; dev++) {
 905                                 pdd = hfi1_lookup(dev);
 906                                 if (pdd && pdd->freectxts &&
 907                                     pdd->freectxts > free) {
 908                                         dd = pdd;
 909                                         free = pdd->freectxts;
 910                                 }
 911                         }
 912                 } else {
 913                         for (dev = 0; dev < devmax; dev++) {
 914                                 pdd = hfi1_lookup(dev);
 915                                 if (pdd && pdd->freectxts) {
 916                                         dd = pdd;
 917                                         break;
 918                                 }
 919                         }
 920                 }
 921                 if (!dd)
 922                         ret = -EBUSY;
 923         }
 924 done:
 925         return ret ? ret : allocate_ctxt(fp, dd, uinfo);
 926 }
 927
 928 static int find_shared_ctxt(struct file *fp,
 929                             const struct hfi1_user_info *uinfo)
 930 {
 931         int devmax, ndev, i;
 932         int ret = 0;
 933
 934         devmax = hfi1_count_units(NULL, NULL);
 935
 936         for (ndev = 0; ndev < devmax; ndev++) {
 937                 struct hfi1_devdata *dd = hfi1_lookup(ndev);
 938
 939                 /* device portion of usable() */
 940                 if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
 941                         continue;
 942                 for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
 943                         struct hfi1_ctxtdata *uctxt = dd->rcd[i];
 944
 945                         /* Skip ctxts which are not yet open */
 946                         if (!uctxt || !uctxt->cnt)
 947                                 continue;
 948                         /* Skip ctxt if it doesn't match the requested one */
 949                         if (memcmp(uctxt->uuid, uinfo->uuid,
 950                                    sizeof(uctxt->uuid)) ||
 951                             uctxt->subctxt_id != uinfo->subctxt_id ||
 952                             uctxt->subctxt_cnt != uinfo->subctxt_cnt)
 953                                 continue;
 954
 955                         /* Verify the sharing process matches the master */
 956                         if (uctxt->userversion != uinfo->userversion ||
 957                             uctxt->cnt >= uctxt->subctxt_cnt) {
 958                                 ret = -EINVAL;
 959                                 goto done;
 960                         }
 961                         ctxt_fp(fp) = uctxt;
 962                         subctxt_fp(fp) = uctxt->cnt++;
 963                         uctxt->subpid[subctxt_fp(fp)] = current->pid;
 964                         uctxt->active_slaves |= 1 << subctxt_fp(fp);
 965                         ret = 1;
 966                         goto done;
 967                 }
 968         }
 969
 970 done:
 971         return ret;
 972 }
 973
 974 static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
 975                          struct hfi1_user_info *uinfo)
 976 {
 977         struct hfi1_ctxtdata *uctxt;
 978         unsigned ctxt;
 979         int ret;
 980
 981         if (dd->flags & HFI1_FROZEN) {
 982                 /*
 983                  * Pick an error that is unique from all other errors
 984                  * that are returned so the user process knows that
 985                  * it tried to allocate while the SPC was frozen.  It
 986                  * it should be able to retry with success in a short
 987                  * while.
 988                  */
 989                 return -EIO;
 990         }
 991
 992         for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
 993                 if (!dd->rcd[ctxt])
 994                         break;
 995
 996         if (ctxt == dd->num_rcv_contexts)
 997                 return -EBUSY;
 998
 999         uctxt = hfi1_create_ctxtdata(dd->pport, ctxt);
1000         if (!uctxt) {
1001                 dd_dev_err(dd,
1002                            "Unable to allocate ctxtdata memory, failing open\n");
1003                 return -ENOMEM;
1004         }
1005         /*
1006          * Allocate and enable a PIO send context.
1007          */
1008         uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
1009                              uctxt->numa_id);
1010         if (!uctxt->sc)
1011                 return -ENOMEM;
1012
1013         dbg("allocated send context %u(%u)\n", uctxt->sc->sw_index,
1014                 uctxt->sc->hw_context);
1015         ret = sc_enable(uctxt->sc);
1016         if (ret)
1017                 return ret;
1018         /*
1019          * Setup shared context resources if the user-level has requested
1020          * shared contexts and this is the 'master' process.
1021          * This has to be done here so the rest of the sub-contexts find the
1022          * proper master.
1023          */
1024         if (uinfo->subctxt_cnt && !subctxt_fp(fp)) {
1025                 ret = init_subctxts(uctxt, uinfo);
1026                 /*
1027                  * On error, we don't need to disable and de-allocate the
1028                  * send context because it will be done during file close
1029                  */
1030                 if (ret)
1031                         return ret;
1032         }
1033         uctxt->userversion = uinfo->userversion;
1034         uctxt->pid = current->pid;
1035         uctxt->flags = HFI1_CAP_UGET(MASK);
1036         init_waitqueue_head(&uctxt->wait);
1037         strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1038         memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1039         uctxt->jkey = generate_jkey(current_uid());
1040         INIT_LIST_HEAD(&uctxt->sdma_queues);
1041         spin_lock_init(&uctxt->sdma_qlock);
1042         hfi1_stats.sps_ctxts++;
1043         dd->freectxts--;
1044         ctxt_fp(fp) = uctxt;
1045
1046         return 0;
1047 }
1048
1049 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1050                          const struct hfi1_user_info *uinfo)
1051 {
1052         int ret = 0;
1053         unsigned num_subctxts;
1054
1055         num_subctxts = uinfo->subctxt_cnt;
1056         if (num_subctxts > HFI1_MAX_SHARED_CTXTS) {
1057                 ret = -EINVAL;
1058                 goto bail;
1059         }
1060
1061         uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1062         uctxt->subctxt_id = uinfo->subctxt_id;
1063         uctxt->active_slaves = 1;
1064         uctxt->redirect_seq_cnt = 1;
1065         set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1066 bail:
1067         return ret;
1068 }
1069
1070 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1071 {
1072         int ret = 0;
1073         unsigned num_subctxts = uctxt->subctxt_cnt;
1074
1075         uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1076         if (!uctxt->subctxt_uregbase) {
1077                 ret = -ENOMEM;
1078                 goto bail;
1079         }
1080         /* We can take the size of the RcvHdr Queue from the master */
1081         uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1082                                                   num_subctxts);
1083         if (!uctxt->subctxt_rcvhdr_base) {
1084                 ret = -ENOMEM;
1085                 goto bail_ureg;
1086         }
1087
1088         uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1089                                                 num_subctxts);
1090         if (!uctxt->subctxt_rcvegrbuf) {
1091                 ret = -ENOMEM;
1092                 goto bail_rhdr;
1093         }
1094         goto bail;
1095 bail_rhdr:
1096         vfree(uctxt->subctxt_rcvhdr_base);
1097 bail_ureg:
1098         vfree(uctxt->subctxt_uregbase);
1099         uctxt->subctxt_uregbase = NULL;
1100 bail:
1101         return ret;
1102 }
1103
1104 static int user_init(struct file *fp)
1105 {
1106         int ret;
1107         unsigned int rcvctrl_ops = 0;
1108         struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
1109
1110         /* make sure that the context has already been setup */
1111         if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags)) {
1112                 ret = -EFAULT;
1113                 goto done;
1114         }
1115
1116         /*
1117          * Subctxts don't need to initialize anything since master
1118          * has done it.
1119          */
1120         if (subctxt_fp(fp)) {
1121                 ret = wait_event_interruptible(uctxt->wait,
1122                         !test_bit(HFI1_CTXT_MASTER_UNINIT,
1123                         &uctxt->event_flags));
1124                 goto done;
1125         }
1126
1127         /* initialize poll variables... */
1128         uctxt->urgent = 0;
1129         uctxt->urgent_poll = 0;
1130
1131         /*
1132          * Now enable the ctxt for receive.
1133          * For chips that are set to DMA the tail register to memory
1134          * when they change (and when the update bit transitions from
1135          * 0 to 1.  So for those chips, we turn it off and then back on.
1136          * This will (very briefly) affect any other open ctxts, but the
1137          * duration is very short, and therefore isn't an issue.  We
1138          * explicitly set the in-memory tail copy to 0 beforehand, so we
1139          * don't have to wait to be sure the DMA update has happened
1140          * (chip resets head/tail to 0 on transition to enable).
1141          */
1142         if (uctxt->rcvhdrtail_kvaddr)
1143                 clear_rcvhdrtail(uctxt);
1144
1145         /* Setup J_KEY before enabling the context */
1146         hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1147
1148         rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1149         if (HFI1_CAP_KGET_MASK(uctxt->flags, HDRSUPP))
1150                 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1151         /*
1152          * Ignore the bit in the flags for now until proper
1153          * support for multiple packet per rcv array entry is
1154          * added.
1155          */
1156         if (!HFI1_CAP_KGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1157                 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1158         if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1159                 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1160         if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1161                 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1162         if (HFI1_CAP_KGET_MASK(uctxt->flags, DMA_RTAIL))
1163                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1164         hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1165
1166         /* Notify any waiting slaves */
1167         if (uctxt->subctxt_cnt) {
1168                 clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1169                 wake_up(&uctxt->wait);
1170         }
1171         ret = 0;
1172
1173 done:
1174         return ret;
1175 }
1176
1177 static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1178 {
1179         struct hfi1_ctxt_info cinfo;
1180         struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
1181         struct hfi1_filedata *fd = fp->private_data;
1182         int ret = 0;
1183
1184         memset(&cinfo, 0, sizeof(cinfo));
1185         ret = hfi1_get_base_kinfo(uctxt, &cinfo);
1186         if (ret < 0)
1187                 goto done;
1188         cinfo.num_active = hfi1_count_active_units();
1189         cinfo.unit = uctxt->dd->unit;
1190         cinfo.ctxt = uctxt->ctxt;
1191         cinfo.subctxt = subctxt_fp(fp);
1192         cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1193                                 uctxt->dd->rcv_entries.group_size) +
1194                 uctxt->expected_count;
1195         cinfo.credits = uctxt->sc->credits;
1196         cinfo.numa_node = uctxt->numa_id;
1197         cinfo.rec_cpu = fd->rec_cpu_num;
1198         cinfo.send_ctxt = uctxt->sc->hw_context;
1199
1200         cinfo.egrtids = uctxt->egrbufs.alloced;
1201         cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1202         cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1203         cinfo.sdma_ring_size = user_sdma_comp_fp(fp)->nentries;
1204         cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1205
1206         trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, subctxt_fp(fp), cinfo);
1207         if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1208                 ret = -EFAULT;
1209 done:
1210         return ret;
1211 }
1212
1213 static int setup_ctxt(struct file *fp)
1214 {
1215         struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
1216         struct hfi1_devdata *dd = uctxt->dd;
1217         int ret = 0;
1218
1219         /*
1220          * Context should be set up only once (including allocation and
1221          * programming of eager buffers. This is done if context sharing
1222          * is not requested or by the master process.
1223          */
1224         if (!uctxt->subctxt_cnt || !subctxt_fp(fp)) {
1225                 ret = hfi1_init_ctxt(uctxt->sc);
1226                 if (ret)
1227                         goto done;
1228
1229                 /* Now allocate the RcvHdr queue and eager buffers. */
1230                 ret = hfi1_create_rcvhdrq(dd, uctxt);
1231                 if (ret)
1232                         goto done;
1233                 ret = hfi1_setup_eagerbufs(uctxt);
1234                 if (ret)
1235                         goto done;
1236                 if (uctxt->subctxt_cnt && !subctxt_fp(fp)) {
1237                         ret = setup_subctxt(uctxt);
1238                         if (ret)
1239                                 goto done;
1240                 }
1241                 /* Setup Expected Rcv memories */
1242                 uctxt->tid_pg_list = vzalloc(uctxt->expected_count *
1243                                              sizeof(struct page **));
1244                 if (!uctxt->tid_pg_list) {
1245                         ret = -ENOMEM;
1246                         goto done;
1247                 }
1248                 uctxt->physshadow = vzalloc(uctxt->expected_count *
1249                                             sizeof(*uctxt->physshadow));
1250                 if (!uctxt->physshadow) {
1251                         ret = -ENOMEM;
1252                         goto done;
1253                 }
1254                 /* allocate expected TID map and initialize the cursor */
1255                 atomic_set(&uctxt->tidcursor, 0);
1256                 uctxt->numtidgroups = uctxt->expected_count /
1257                         dd->rcv_entries.group_size;
1258                 uctxt->tidmapcnt = uctxt->numtidgroups / BITS_PER_LONG +
1259                         !!(uctxt->numtidgroups % BITS_PER_LONG);
1260                 uctxt->tidusemap = kzalloc_node(uctxt->tidmapcnt *
1261                                                 sizeof(*uctxt->tidusemap),
1262                                                 GFP_KERNEL, uctxt->numa_id);
1263                 if (!uctxt->tidusemap) {
1264                         ret = -ENOMEM;
1265                         goto done;
1266                 }
1267                 /*
1268                  * In case that the number of groups is not a multiple of
1269                  * 64 (the number of groups in a tidusemap element), mark
1270                  * the extra ones as used. This will effectively make them
1271                  * permanently used and should never be assigned. Otherwise,
1272                  * the code which checks how many free groups we have will
1273                  * get completely confused about the state of the bits.
1274                  */
1275                 if (uctxt->numtidgroups % BITS_PER_LONG)
1276                         uctxt->tidusemap[uctxt->tidmapcnt - 1] =
1277                                 ~((1ULL << (uctxt->numtidgroups %
1278                                             BITS_PER_LONG)) - 1);
1279                 trace_hfi1_exp_tid_map(uctxt->ctxt, subctxt_fp(fp), 0,
1280                                        uctxt->tidusemap, uctxt->tidmapcnt);
1281         }
1282         ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1283         if (ret)
1284                 goto done;
1285
1286         set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1287 done:
1288         return ret;
1289 }
1290
1291 static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1292 {
1293         struct hfi1_base_info binfo;
1294         struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
1295         struct hfi1_devdata *dd = uctxt->dd;
1296         ssize_t sz;
1297         unsigned offset;
1298         int ret = 0;
1299
1300         trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1301
1302         memset(&binfo, 0, sizeof(binfo));
1303         binfo.hw_version = dd->revision;
1304         binfo.sw_version = HFI1_KERN_SWVERSION;
1305         binfo.bthqp = kdeth_qp;
1306         binfo.jkey = uctxt->jkey;
1307         /*
1308          * If more than 64 contexts are enabled the allocated credit
1309          * return will span two or three contiguous pages. Since we only
1310          * map the page containing the context's credit return address,
1311          * we need to calculate the offset in the proper page.
1312          */
1313         offset = ((u64)uctxt->sc->hw_free -
1314                   (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1315         binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1316                                                subctxt_fp(fp), offset);
1317         binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1318                                             subctxt_fp(fp),
1319                                             uctxt->sc->base_addr);
1320         binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1321                                                 uctxt->ctxt,
1322                                                 subctxt_fp(fp),
1323                                                 uctxt->sc->base_addr);
1324         binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1325                                                subctxt_fp(fp),
1326                                                uctxt->rcvhdrq);
1327         binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1328                                                subctxt_fp(fp),
1329                                                uctxt->egrbufs.rcvtids[0].phys);
1330         binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1331                                                  subctxt_fp(fp), 0);
1332         /*
1333          * user regs are at
1334          * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1335          */
1336         binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1337                                             subctxt_fp(fp), 0);
1338         offset = ((((uctxt->ctxt - dd->first_user_ctxt) *
1339                     HFI1_MAX_SHARED_CTXTS) + subctxt_fp(fp)) *
1340                   sizeof(*dd->events)) & ~PAGE_MASK;
1341         binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1342                                               subctxt_fp(fp),
1343                                               offset);
1344         binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1345                                               subctxt_fp(fp),
1346                                               dd->status);
1347         if (HFI1_CAP_IS_USET(DMA_RTAIL))
1348                 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1349                                                        subctxt_fp(fp), 0);
1350         if (uctxt->subctxt_cnt) {
1351                 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1352                                                         uctxt->ctxt,
1353                                                         subctxt_fp(fp), 0);
1354                 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1355                                                          uctxt->ctxt,
1356                                                          subctxt_fp(fp), 0);
1357                 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1358                                                          uctxt->ctxt,
1359                                                          subctxt_fp(fp), 0);
1360         }
1361         sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1362         if (copy_to_user(ubase, &binfo, sz))
1363                 ret = -EFAULT;
1364         return ret;
1365 }
1366
1367 static unsigned int poll_urgent(struct file *fp,
1368                                 struct poll_table_struct *pt)
1369 {
1370         struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
1371         struct hfi1_devdata *dd = uctxt->dd;
1372         unsigned pollflag;
1373
1374         poll_wait(fp, &uctxt->wait, pt);
1375
1376         spin_lock_irq(&dd->uctxt_lock);
1377         if (uctxt->urgent != uctxt->urgent_poll) {
1378                 pollflag = POLLIN | POLLRDNORM;
1379                 uctxt->urgent_poll = uctxt->urgent;
1380         } else {
1381                 pollflag = 0;
1382                 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1383         }
1384         spin_unlock_irq(&dd->uctxt_lock);
1385
1386         return pollflag;
1387 }
1388
1389 static unsigned int poll_next(struct file *fp,
1390                               struct poll_table_struct *pt)
1391 {
1392         struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
1393         struct hfi1_devdata *dd = uctxt->dd;
1394         unsigned pollflag;
1395
1396         poll_wait(fp, &uctxt->wait, pt);
1397
1398         spin_lock_irq(&dd->uctxt_lock);
1399         if (hdrqempty(uctxt)) {
1400                 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1401                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1402                 pollflag = 0;
1403         } else
1404                 pollflag = POLLIN | POLLRDNORM;
1405         spin_unlock_irq(&dd->uctxt_lock);
1406
1407         return pollflag;
1408 }
1409
1410 /*
1411  * Find all user contexts in use, and set the specified bit in their
1412  * event mask.
1413  * See also find_ctxt() for a similar use, that is specific to send buffers.
1414  */
1415 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1416 {
1417         struct hfi1_ctxtdata *uctxt;
1418         struct hfi1_devdata *dd = ppd->dd;
1419         unsigned ctxt;
1420         int ret = 0;
1421         unsigned long flags;
1422
1423         if (!dd->events) {
1424                 ret = -EINVAL;
1425                 goto done;
1426         }
1427
1428         spin_lock_irqsave(&dd->uctxt_lock, flags);
1429         for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1430              ctxt++) {
1431                 uctxt = dd->rcd[ctxt];
1432                 if (uctxt) {
1433                         unsigned long *evs = dd->events +
1434                                 (uctxt->ctxt - dd->first_user_ctxt) *
1435                                 HFI1_MAX_SHARED_CTXTS;
1436                         int i;
1437                         /*
1438                          * subctxt_cnt is 0 if not shared, so do base
1439                          * separately, first, then remaining subctxt, if any
1440                          */
1441                         set_bit(evtbit, evs);
1442                         for (i = 1; i < uctxt->subctxt_cnt; i++)
1443                                 set_bit(evtbit, evs + i);
1444                 }
1445         }
1446         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1447 done:
1448         return ret;
1449 }
1450
1451 /**
1452  * manage_rcvq - manage a context's receive queue
1453  * @uctxt: the context
1454  * @subctxt: the sub-context
1455  * @start_stop: action to carry out
1456  *
1457  * start_stop == 0 disables receive on the context, for use in queue
1458  * overflow conditions.  start_stop==1 re-enables, to be used to
1459  * re-init the software copy of the head register
1460  */
1461 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1462                        int start_stop)
1463 {
1464         struct hfi1_devdata *dd = uctxt->dd;
1465         unsigned int rcvctrl_op;
1466
1467         if (subctxt)
1468                 goto bail;
1469         /* atomically clear receive enable ctxt. */
1470         if (start_stop) {
1471                 /*
1472                  * On enable, force in-memory copy of the tail register to
1473                  * 0, so that protocol code doesn't have to worry about
1474                  * whether or not the chip has yet updated the in-memory
1475                  * copy or not on return from the system call. The chip
1476                  * always resets it's tail register back to 0 on a
1477                  * transition from disabled to enabled.
1478                  */
1479                 if (uctxt->rcvhdrtail_kvaddr)
1480                         clear_rcvhdrtail(uctxt);
1481                 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1482         } else
1483                 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1484         hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1485         /* always; new head should be equal to new tail; see above */
1486 bail:
1487         return 0;
1488 }
1489
1490 /*
1491  * clear the event notifier events for this context.
1492  * User process then performs actions appropriate to bit having been
1493  * set, if desired, and checks again in future.
1494  */
1495 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1496                           unsigned long events)
1497 {
1498         int i;
1499         struct hfi1_devdata *dd = uctxt->dd;
1500         unsigned long *evs;
1501
1502         if (!dd->events)
1503                 return 0;
1504
1505         evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1506                             HFI1_MAX_SHARED_CTXTS) + subctxt;
1507
1508         for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1509                 if (!test_bit(i, &events))
1510                         continue;
1511                 clear_bit(i, evs);
1512         }
1513         return 0;
1514 }
1515
1516 #define num_user_pages(vaddr, len)                                      \
1517         (1 + (((((unsigned long)(vaddr) +                               \
1518                  (unsigned long)(len) - 1) & PAGE_MASK) -               \
1519                ((unsigned long)vaddr & PAGE_MASK)) >> PAGE_SHIFT))
1520
1521 /**
1522  * tzcnt - count the number of trailing zeros in a 64bit value
1523  * @value: the value to be examined
1524  *
1525  * Returns the number of trailing least significant zeros in the
1526  * the input value. If the value is zero, return the number of
1527  * bits of the value.
1528  */
1529 static inline u8 tzcnt(u64 value)
1530 {
1531         return value ? __builtin_ctzl(value) : sizeof(value) * 8;
1532 }
1533
1534 static inline unsigned num_free_groups(unsigned long map, u16 *start)
1535 {
1536         unsigned free;
1537         u16 bitidx = *start;
1538
1539         if (bitidx >= BITS_PER_LONG)
1540                 return 0;
1541         /* "Turn off" any bits set before our bit index */
1542         map &= ~((1ULL << bitidx) - 1);
1543         free = tzcnt(map) - bitidx;
1544         while (!free && bitidx < BITS_PER_LONG) {
1545                 /* Zero out the last set bit so we look at the rest */
1546                 map &= ~(1ULL << bitidx);
1547                 /*
1548                  * Account for the previously checked bits and advance
1549                  * the bit index. We don't have to check for bitidx
1550                  * getting bigger than BITS_PER_LONG here as it would
1551                  * mean extra instructions that we don't need. If it
1552                  * did happen, it would push free to a negative value
1553                  * which will break the loop.
1554                  */
1555                 free = tzcnt(map) - ++bitidx;
1556         }
1557         *start = bitidx;
1558         return free;
1559 }
1560
1561 static int exp_tid_setup(struct file *fp, struct hfi1_tid_info *tinfo)
1562 {
1563         int ret = 0;
1564         struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
1565         struct hfi1_devdata *dd = uctxt->dd;
1566         unsigned tid, mapped = 0, npages, ngroups, exp_groups,
1567                 tidpairs = uctxt->expected_count / 2;
1568         struct page **pages;
1569         unsigned long vaddr, tidmap[uctxt->tidmapcnt];
1570         dma_addr_t *phys;
1571         u32 tidlist[tidpairs], pairidx = 0, tidcursor;
1572         u16 useidx, idx, bitidx, tidcnt = 0;
1573
1574         vaddr = tinfo->vaddr;
1575
1576         if (vaddr & ~PAGE_MASK) {
1577                 ret = -EINVAL;
1578                 goto bail;
1579         }
1580
1581         npages = num_user_pages(vaddr, tinfo->length);
1582         if (!npages) {
1583                 ret = -EINVAL;
1584                 goto bail;
1585         }
1586         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
1587                        npages * PAGE_SIZE)) {
1588                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
1589                            (void *)vaddr, npages);
1590                 ret = -EFAULT;
1591                 goto bail;
1592         }
1593
1594         memset(tidmap, 0, sizeof(tidmap[0]) * uctxt->tidmapcnt);
1595         memset(tidlist, 0, sizeof(tidlist[0]) * tidpairs);
1596
1597         exp_groups = uctxt->expected_count / dd->rcv_entries.group_size;
1598         /* which group set do we look at first? */
1599         tidcursor = atomic_read(&uctxt->tidcursor);
1600         useidx = (tidcursor >> 16) & 0xffff;
1601         bitidx = tidcursor & 0xffff;
1602
1603         /*
1604          * Keep going until we've mapped all pages or we've exhausted all
1605          * RcvArray entries.
1606          * This iterates over the number of tidmaps + 1
1607          * (idx <= uctxt->tidmapcnt) so we check the bitmap which we
1608          * started from one more time for any free bits before the
1609          * starting point bit.
1610          */
1611         for (mapped = 0, idx = 0;
1612              mapped < npages && idx <= uctxt->tidmapcnt;) {
1613                 u64 i, offset = 0;
1614                 unsigned free, pinned, pmapped = 0, bits_used;
1615                 u16 grp;
1616
1617                 /*
1618                  * "Reserve" the needed group bits under lock so other
1619                  * processes can't step in the middle of it. Once
1620                  * reserved, we don't need the lock anymore since we
1621                  * are guaranteed the groups.
1622                  */
1623                 spin_lock(&uctxt->exp_lock);
1624                 if (uctxt->tidusemap[useidx] == -1ULL ||
1625                     bitidx >= BITS_PER_LONG) {
1626                         /* no free groups in the set, use the next */
1627                         useidx = (useidx + 1) % uctxt->tidmapcnt;
1628                         idx++;
1629                         bitidx = 0;
1630                         spin_unlock(&uctxt->exp_lock);
1631                         continue;
1632                 }
1633                 ngroups = ((npages - mapped) / dd->rcv_entries.group_size) +
1634                         !!((npages - mapped) % dd->rcv_entries.group_size);
1635
1636                 /*
1637                  * If we've gotten here, the current set of groups does have
1638                  * one or more free groups.
1639                  */
1640                 free = num_free_groups(uctxt->tidusemap[useidx], &bitidx);
1641                 if (!free) {
1642                         /*
1643                          * Despite the check above, free could still come back
1644                          * as 0 because we don't check the entire bitmap but
1645                          * we start from bitidx.
1646                          */
1647                         spin_unlock(&uctxt->exp_lock);
1648                         continue;
1649                 }
1650                 bits_used = min(free, ngroups);
1651                 tidmap[useidx] |= ((1ULL << bits_used) - 1) << bitidx;
1652                 uctxt->tidusemap[useidx] |= tidmap[useidx];
1653                 spin_unlock(&uctxt->exp_lock);
1654
1655                 /*
1656                  * At this point, we know where in the map we have free bits.
1657                  * properly offset into the various "shadow" arrays and compute
1658                  * the RcvArray entry index.
1659                  */
1660                 offset = ((useidx * BITS_PER_LONG) + bitidx) *
1661                         dd->rcv_entries.group_size;
1662                 pages = uctxt->tid_pg_list + offset;
1663                 phys = uctxt->physshadow + offset;
1664                 tid = uctxt->expected_base + offset;
1665
1666                 /* Calculate how many pages we can pin based on free bits */
1667                 pinned = min((bits_used * dd->rcv_entries.group_size),
1668                              (npages - mapped));
1669                 /*
1670                  * Now that we know how many free RcvArray entries we have,
1671                  * we can pin that many user pages.
1672                  */
1673                 ret = hfi1_get_user_pages(vaddr + (mapped * PAGE_SIZE),
1674                                           pinned, pages);
1675                 if (ret) {
1676                         /*
1677                          * We can't continue because the pages array won't be
1678                          * initialized. This should never happen,
1679                          * unless perhaps the user has mpin'ed the pages
1680                          * themselves.
1681                          */
1682                         dd_dev_info(dd,
1683                                     "Failed to lock addr %p, %u pages: errno %d\n",
1684                                     (void *) vaddr, pinned, -ret);
1685                         /*
1686                          * Let go of the bits that we reserved since we are not
1687                          * going to use them.
1688                          */
1689                         spin_lock(&uctxt->exp_lock);
1690                         uctxt->tidusemap[useidx] &=
1691                                 ~(((1ULL << bits_used) - 1) << bitidx);
1692                         spin_unlock(&uctxt->exp_lock);
1693                         goto done;
1694                 }
1695                 /*
1696                  * How many groups do we need based on how many pages we have
1697                  * pinned?
1698                  */
1699                 ngroups = (pinned / dd->rcv_entries.group_size) +
1700                         !!(pinned % dd->rcv_entries.group_size);
1701                 /*
1702                  * Keep programming RcvArray entries for all the <ngroups> free
1703                  * groups.
1704                  */
1705                 for (i = 0, grp = 0; grp < ngroups; i++, grp++) {
1706                         unsigned j;
1707                         u32 pair_size = 0, tidsize;
1708                         /*
1709                          * This inner loop will program an entire group or the
1710                          * array of pinned pages (which ever limit is hit
1711                          * first).
1712                          */
1713                         for (j = 0; j < dd->rcv_entries.group_size &&
1714                                      pmapped < pinned; j++, pmapped++, tid++) {
1715                                 tidsize = PAGE_SIZE;
1716                                 phys[pmapped] = hfi1_map_page(dd->pcidev,
1717                                                    pages[pmapped], 0,
1718                                                    tidsize, PCI_DMA_FROMDEVICE);
1719                                 trace_hfi1_exp_rcv_set(uctxt->ctxt,
1720                                                        subctxt_fp(fp),
1721                                                        tid, vaddr,
1722                                                        phys[pmapped],
1723                                                        pages[pmapped]);
1724                                 /*
1725                                  * Each RcvArray entry is programmed with one
1726                                  * page * worth of memory. This will handle
1727                                  * the 8K MTU as well as anything smaller
1728                                  * due to the fact that both entries in the
1729                                  * RcvTidPair are programmed with a page.
1730                                  * PSM currently does not handle anything
1731                                  * bigger than 8K MTU, so should we even worry
1732                                  * about 10K here?
1733                                  */
1734                                 hfi1_put_tid(dd, tid, PT_EXPECTED,
1735                                              phys[pmapped],
1736                                              ilog2(tidsize >> PAGE_SHIFT) + 1);
1737                                 pair_size += tidsize >> PAGE_SHIFT;
1738                                 EXP_TID_RESET(tidlist[pairidx], LEN, pair_size);
1739                                 if (!(tid % 2)) {
1740                                         tidlist[pairidx] |=
1741                                            EXP_TID_SET(IDX,
1742                                                 (tid - uctxt->expected_base)
1743                                                        / 2);
1744                                         tidlist[pairidx] |=
1745                                                 EXP_TID_SET(CTRL, 1);
1746                                         tidcnt++;
1747                                 } else {
1748                                         tidlist[pairidx] |=
1749                                                 EXP_TID_SET(CTRL, 2);
1750                                         pair_size = 0;
1751                                         pairidx++;
1752                                 }
1753                         }
1754                         /*
1755                          * We've programmed the entire group (or as much of the
1756                          * group as we'll use. Now, it's time to push it out...
1757                          */
1758                         flush_wc();
1759                 }
1760                 mapped += pinned;
1761                 atomic_set(&uctxt->tidcursor,
1762                            (((useidx & 0xffffff) << 16) |
1763                             ((bitidx + bits_used) & 0xffffff)));
1764         }
1765         trace_hfi1_exp_tid_map(uctxt->ctxt, subctxt_fp(fp), 0, uctxt->tidusemap,
1766                                uctxt->tidmapcnt);
1767
1768 done:
1769         /* If we've mapped anything, copy relevant info to user */
1770         if (mapped) {
1771                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
1772                                  tidlist, sizeof(tidlist[0]) * tidcnt)) {
1773                         ret = -EFAULT;
1774                         goto done;
1775                 }
1776                 /* copy TID info to user */
1777                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidmap,
1778                                  tidmap, sizeof(tidmap[0]) * uctxt->tidmapcnt))
1779                         ret = -EFAULT;
1780         }
1781 bail:
1782         /*
1783          * Calculate mapped length. New Exp TID protocol does not "unwind" and
1784          * report an error if it can't map the entire buffer. It just reports
1785          * the length that was mapped.
1786          */
1787         tinfo->length = mapped * PAGE_SIZE;
1788         tinfo->tidcnt = tidcnt;
1789         return ret;
1790 }
1791
1792 static int exp_tid_free(struct file *fp, struct hfi1_tid_info *tinfo)
1793 {
1794         struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
1795         struct hfi1_devdata *dd = uctxt->dd;
1796         unsigned long tidmap[uctxt->tidmapcnt];
1797         struct page **pages;
1798         dma_addr_t *phys;
1799         u16 idx, bitidx, tid;
1800         int ret = 0;
1801
1802         if (copy_from_user(&tidmap, (void __user *)(unsigned long)
1803                            tinfo->tidmap,
1804                            sizeof(tidmap[0]) * uctxt->tidmapcnt)) {
1805                 ret = -EFAULT;
1806                 goto done;
1807         }
1808         for (idx = 0; idx < uctxt->tidmapcnt; idx++) {
1809                 unsigned long map;
1810
1811                 bitidx = 0;
1812                 if (!tidmap[idx])
1813                         continue;
1814                 map = tidmap[idx];
1815                 while ((bitidx = tzcnt(map)) < BITS_PER_LONG) {
1816                         int i, pcount = 0;
1817                         struct page *pshadow[dd->rcv_entries.group_size];
1818                         unsigned offset = ((idx * BITS_PER_LONG) + bitidx) *
1819                                 dd->rcv_entries.group_size;
1820
1821                         pages = uctxt->tid_pg_list + offset;
1822                         phys = uctxt->physshadow + offset;
1823                         tid = uctxt->expected_base + offset;
1824                         for (i = 0; i < dd->rcv_entries.group_size;
1825                              i++, tid++) {
1826                                 if (pages[i]) {
1827                                         hfi1_put_tid(dd, tid, PT_INVALID,
1828                                                       0, 0);
1829                                         trace_hfi1_exp_rcv_free(uctxt->ctxt,
1830                                                                 subctxt_fp(fp),
1831                                                                 tid, phys[i],
1832                                                                 pages[i]);
1833                                         pci_unmap_page(dd->pcidev, phys[i],
1834                                               PAGE_SIZE, PCI_DMA_FROMDEVICE);
1835                                         pshadow[pcount] = pages[i];
1836                                         pages[i] = NULL;
1837                                         pcount++;
1838                                         phys[i] = 0;
1839                                 }
1840                         }
1841                         flush_wc();
1842                         hfi1_release_user_pages(pshadow, pcount);
1843                         clear_bit(bitidx, &uctxt->tidusemap[idx]);
1844                         map &= ~(1ULL<<bitidx);
1845                 }
1846         }
1847         trace_hfi1_exp_tid_map(uctxt->ctxt, subctxt_fp(fp), 1, uctxt->tidusemap,
1848                                uctxt->tidmapcnt);
1849 done:
1850         return ret;
1851 }
1852
1853 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt)
1854 {
1855         struct hfi1_devdata *dd = uctxt->dd;
1856         unsigned tid;
1857
1858         dd_dev_info(dd, "ctxt %u unlocking any locked expTID pages\n",
1859                     uctxt->ctxt);
1860         for (tid = 0; tid < uctxt->expected_count; tid++) {
1861                 struct page *p = uctxt->tid_pg_list[tid];
1862                 dma_addr_t phys;
1863
1864                 if (!p)
1865                         continue;
1866
1867                 phys = uctxt->physshadow[tid];
1868                 uctxt->physshadow[tid] = 0;
1869                 uctxt->tid_pg_list[tid] = NULL;
1870                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE, PCI_DMA_FROMDEVICE);
1871                 hfi1_release_user_pages(&p, 1);
1872         }
1873 }
1874
1875 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1876                          u16 pkey)
1877 {
1878         int ret = -ENOENT, i, intable = 0;
1879         struct hfi1_pportdata *ppd = uctxt->ppd;
1880         struct hfi1_devdata *dd = uctxt->dd;
1881
1882         if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1883                 ret = -EINVAL;
1884                 goto done;
1885         }
1886
1887         for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1888                 if (pkey == ppd->pkeys[i]) {
1889                         intable = 1;
1890                         break;
1891                 }
1892
1893         if (intable)
1894                 ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1895 done:
1896         return ret;
1897 }
1898
1899 static int ui_open(struct inode *inode, struct file *filp)
1900 {
1901         struct hfi1_devdata *dd;
1902
1903         dd = container_of(inode->i_cdev, struct hfi1_devdata, ui_cdev);
1904         filp->private_data = dd; /* for other methods */
1905         return 0;
1906 }
1907
1908 static int ui_release(struct inode *inode, struct file *filp)
1909 {
1910         /* nothing to do */
1911         return 0;
1912 }
1913
1914 static loff_t ui_lseek(struct file *filp, loff_t offset, int whence)
1915 {
1916         struct hfi1_devdata *dd = filp->private_data;
1917
1918         switch (whence) {
1919         case SEEK_SET:
1920                 break;
1921         case SEEK_CUR:
1922                 offset += filp->f_pos;
1923                 break;
1924         case SEEK_END:
1925                 offset = ((dd->kregend - dd->kregbase) + DC8051_DATA_MEM_SIZE) -
1926                         offset;
1927                 break;
1928         default:
1929                 return -EINVAL;
1930         }
1931
1932         if (offset < 0)
1933                 return -EINVAL;
1934
1935         if (offset >= (dd->kregend - dd->kregbase) + DC8051_DATA_MEM_SIZE)
1936                 return -EINVAL;
1937
1938         filp->f_pos = offset;
1939
1940         return filp->f_pos;
1941 }
1942
1943
1944 /* NOTE: assumes unsigned long is 8 bytes */
1945 static ssize_t ui_read(struct file *filp, char __user *buf, size_t count,
1946                         loff_t *f_pos)
1947 {
1948         struct hfi1_devdata *dd = filp->private_data;
1949         void __iomem *base = dd->kregbase;
1950         unsigned long total, csr_off,
1951                 barlen = (dd->kregend - dd->kregbase);
1952         u64 data;
1953
1954         /* only read 8 byte quantities */
1955         if ((count % 8) != 0)
1956                 return -EINVAL;
1957         /* offset must be 8-byte aligned */
1958         if ((*f_pos % 8) != 0)
1959                 return -EINVAL;
1960         /* destination buffer must be 8-byte aligned */
1961         if ((unsigned long)buf % 8 != 0)
1962                 return -EINVAL;
1963         /* must be in range */
1964         if (*f_pos + count > (barlen + DC8051_DATA_MEM_SIZE))
1965                 return -EINVAL;
1966         /* only set the base if we are not starting past the BAR */
1967         if (*f_pos < barlen)
1968                 base += *f_pos;
1969         csr_off = *f_pos;
1970         for (total = 0; total < count; total += 8, csr_off += 8) {
1971                 /* accessing LCB CSRs requires more checks */
1972                 if (is_lcb_offset(csr_off)) {
1973                         if (read_lcb_csr(dd, csr_off, (u64 *)&data))
1974                                 break; /* failed */
1975                 }
1976                 /*
1977                  * Cannot read ASIC GPIO/QSFP* clear and force CSRs without a
1978                  * false parity error.  Avoid the whole issue by not reading
1979                  * them.  These registers are defined as having a read value
1980                  * of 0.
1981                  */
1982                 else if (csr_off == ASIC_GPIO_CLEAR
1983                                 || csr_off == ASIC_GPIO_FORCE
1984                                 || csr_off == ASIC_QSFP1_CLEAR
1985                                 || csr_off == ASIC_QSFP1_FORCE
1986                                 || csr_off == ASIC_QSFP2_CLEAR
1987                                 || csr_off == ASIC_QSFP2_FORCE)
1988                         data = 0;
1989                 else if (csr_off >= barlen) {
1990                         /*
1991                          * read_8051_data can read more than just 8 bytes at
1992                          * a time. However, folding this into the loop and
1993                          * handling the reads in 8 byte increments allows us
1994                          * to smoothly transition from chip memory to 8051
1995                          * memory.
1996                          */
1997                         if (read_8051_data(dd,
1998                                            (u32)(csr_off - barlen),
1999                                            sizeof(data), &data))
2000                                 break; /* failed */
2001                 } else
2002                         data = readq(base + total);
2003                 if (put_user(data, (unsigned long __user *)(buf + total)))
2004                         break;
2005         }
2006         *f_pos += total;
2007         return total;
2008 }
2009
2010 /* NOTE: assumes unsigned long is 8 bytes */
2011 static ssize_t ui_write(struct file *filp, const char __user *buf,
2012                         size_t count, loff_t *f_pos)
2013 {
2014         struct hfi1_devdata *dd = filp->private_data;
2015         void __iomem *base;
2016         unsigned long total, data, csr_off;
2017         int in_lcb;
2018
2019         /* only write 8 byte quantities */
2020         if ((count % 8) != 0)
2021                 return -EINVAL;
2022         /* offset must be 8-byte aligned */
2023         if ((*f_pos % 8) != 0)
2024                 return -EINVAL;
2025         /* source buffer must be 8-byte aligned */
2026         if ((unsigned long)buf % 8 != 0)
2027                 return -EINVAL;
2028         /* must be in range */
2029         if (*f_pos + count > dd->kregend - dd->kregbase)
2030                 return -EINVAL;
2031
2032         base = (void __iomem *)dd->kregbase + *f_pos;
2033         csr_off = *f_pos;
2034         in_lcb = 0;
2035         for (total = 0; total < count; total += 8, csr_off += 8) {
2036                 if (get_user(data, (unsigned long __user *)(buf + total)))
2037                         break;
2038                 /* accessing LCB CSRs requires a special procedure */
2039                 if (is_lcb_offset(csr_off)) {
2040                         if (!in_lcb) {
2041                                 int ret = acquire_lcb_access(dd, 1);
2042
2043                                 if (ret)
2044                                         break;
2045                                 in_lcb = 1;
2046                         }
2047                 } else {
2048                         if (in_lcb) {
2049                                 release_lcb_access(dd, 1);
2050                                 in_lcb = 0;
2051                         }
2052                 }
2053                 writeq(data, base + total);
2054         }
2055         if (in_lcb)
2056                 release_lcb_access(dd, 1);
2057         *f_pos += total;
2058         return total;
2059 }
2060
2061 static const struct file_operations ui_file_ops = {
2062         .owner = THIS_MODULE,
2063         .llseek = ui_lseek,
2064         .read = ui_read,
2065         .write = ui_write,
2066         .open = ui_open,
2067         .release = ui_release,
2068 };
2069 #define UI_OFFSET 192   /* device minor offset for UI devices */
2070 static int create_ui = 1;
2071
2072 static struct cdev wildcard_cdev;
2073 static struct device *wildcard_device;
2074
2075 static atomic_t user_count = ATOMIC_INIT(0);
2076
2077 static void user_remove(struct hfi1_devdata *dd)
2078 {
2079         if (atomic_dec_return(&user_count) == 0)
2080                 hfi1_cdev_cleanup(&wildcard_cdev, &wildcard_device);
2081
2082         hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
2083         hfi1_cdev_cleanup(&dd->ui_cdev, &dd->ui_device);
2084 }
2085
2086 static int user_add(struct hfi1_devdata *dd)
2087 {
2088         char name[10];
2089         int ret;
2090
2091         if (atomic_inc_return(&user_count) == 1) {
2092                 ret = hfi1_cdev_init(0, class_name(), &hfi1_file_ops,
2093                                      &wildcard_cdev, &wildcard_device,
2094                                      true);
2095                 if (ret)
2096                         goto done;
2097         }
2098
2099         snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
2100         ret = hfi1_cdev_init(dd->unit + 1, name, &hfi1_file_ops,
2101                              &dd->user_cdev, &dd->user_device,
2102                              true);
2103         if (ret)
2104                 goto done;
2105
2106         if (create_ui) {
2107                 snprintf(name, sizeof(name),
2108                          "%s_ui%d", class_name(), dd->unit);
2109                 ret = hfi1_cdev_init(dd->unit + UI_OFFSET, name, &ui_file_ops,
2110                                      &dd->ui_cdev, &dd->ui_device,
2111                                      false);
2112                 if (ret)
2113                         goto done;
2114         }
2115
2116         return 0;
2117 done:
2118         user_remove(dd);
2119         return ret;
2120 }
2121
2122 /*
2123  * Create per-unit files in /dev
2124  */
2125 int hfi1_device_create(struct hfi1_devdata *dd)
2126 {
2127         int r, ret;
2128
2129         r = user_add(dd);
2130         ret = hfi1_diag_add(dd);
2131         if (r && !ret)
2132                 ret = r;
2133         return ret;
2134 }
2135
2136 /*
2137  * Remove per-unit files in /dev
2138  * void, core kernel returns no errors for this stuff
2139  */
2140 void hfi1_device_remove(struct hfi1_devdata *dd)
2141 {
2142         user_remove(dd);
2143         hfi1_diag_remove(dd);
2144 }