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 "hfi.h"
  53 #include "pio.h"
  54 #include "device.h"
  55 #include "common.h"
  56 #include "trace.h"
  57 #include "user_sdma.h"
  58 #include "user_exp_rcv.h"
  59 #include "eprom.h"
  60 #include "aspm.h"
  61 #include "mmu_rb.h"
  62
  63 #undef pr_fmt
  64 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
  65
  66 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
  67
  68 /*
  69  * File operation functions
  70  */
  71 static int hfi1_file_open(struct inode *, struct file *);
  72 static int hfi1_file_close(struct inode *, struct file *);
  73 static ssize_t hfi1_file_write(struct file *, const char __user *,
  74                                size_t, loff_t *);
  75 static ssize_t hfi1_write_iter(struct kiocb *, struct iov_iter *);
  76 static unsigned int hfi1_poll(struct file *, struct poll_table_struct *);
  77 static int hfi1_file_mmap(struct file *, struct vm_area_struct *);
  78
  79 static u64 kvirt_to_phys(void *);
  80 static int assign_ctxt(struct file *, struct hfi1_user_info *);
  81 static int init_subctxts(struct hfi1_ctxtdata *, const struct hfi1_user_info *);
  82 static int user_init(struct file *);
  83 static int get_ctxt_info(struct file *, void __user *, __u32);
  84 static int get_base_info(struct file *, void __user *, __u32);
  85 static int setup_ctxt(struct file *);
  86 static int setup_subctxt(struct hfi1_ctxtdata *);
  87 static int get_user_context(struct file *, struct hfi1_user_info *,
  88                             int, unsigned);
  89 static int find_shared_ctxt(struct file *, const struct hfi1_user_info *);
  90 static int allocate_ctxt(struct file *, struct hfi1_devdata *,
  91                          struct hfi1_user_info *);
  92 static unsigned int poll_urgent(struct file *, struct poll_table_struct *);
  93 static unsigned int poll_next(struct file *, struct poll_table_struct *);
  94 static int user_event_ack(struct hfi1_ctxtdata *, int, unsigned long);
  95 static int set_ctxt_pkey(struct hfi1_ctxtdata *, unsigned, u16);
  96 static int manage_rcvq(struct hfi1_ctxtdata *, unsigned, int);
  97 static int vma_fault(struct vm_area_struct *, struct vm_fault *);
  98
  99 static const struct file_operations hfi1_file_ops = {
 100         .owner = THIS_MODULE,
 101         .write = hfi1_file_write,
 102         .write_iter = hfi1_write_iter,
 103         .open = hfi1_file_open,
 104         .release = hfi1_file_close,
 105         .poll = hfi1_poll,
 106         .mmap = hfi1_file_mmap,
 107         .llseek = noop_llseek,
 108 };
 109
 110 static struct vm_operations_struct vm_ops = {
 111         .fault = vma_fault,
 112 };
 113
 114 /*
 115  * Types of memories mapped into user processes' space
 116  */
 117 enum mmap_types {
 118         PIO_BUFS = 1,
 119         PIO_BUFS_SOP,
 120         PIO_CRED,
 121         RCV_HDRQ,
 122         RCV_EGRBUF,
 123         UREGS,
 124         EVENTS,
 125         STATUS,
 126         RTAIL,
 127         SUBCTXT_UREGS,
 128         SUBCTXT_RCV_HDRQ,
 129         SUBCTXT_EGRBUF,
 130         SDMA_COMP
 131 };
 132
 133 /*
 134  * Masks and offsets defining the mmap tokens
 135  */
 136 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
 137 #define HFI1_MMAP_OFFSET_SHIFT  0
 138 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
 139 #define HFI1_MMAP_SUBCTXT_SHIFT 12
 140 #define HFI1_MMAP_CTXT_MASK     0xffULL
 141 #define HFI1_MMAP_CTXT_SHIFT    16
 142 #define HFI1_MMAP_TYPE_MASK     0xfULL
 143 #define HFI1_MMAP_TYPE_SHIFT    24
 144 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
 145 #define HFI1_MMAP_MAGIC_SHIFT   32
 146
 147 #define HFI1_MMAP_MAGIC         0xdabbad00
 148
 149 #define HFI1_MMAP_TOKEN_SET(field, val) \
 150         (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
 151 #define HFI1_MMAP_TOKEN_GET(field, token) \
 152         (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
 153 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
 154         (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
 155         HFI1_MMAP_TOKEN_SET(TYPE, type) | \
 156         HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
 157         HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
 158         HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
 159
 160 #define dbg(fmt, ...)                           \
 161         pr_info(fmt, ##__VA_ARGS__)
 162
 163 static inline int is_valid_mmap(u64 token)
 164 {
 165         return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
 166 }
 167
 168 static int hfi1_file_open(struct inode *inode, struct file *fp)
 169 {
 170         /* The real work is performed later in assign_ctxt() */
 171         fp->private_data = kzalloc(sizeof(struct hfi1_filedata), GFP_KERNEL);
 172         if (fp->private_data) /* no cpu affinity by default */
 173                 ((struct hfi1_filedata *)fp->private_data)->rec_cpu_num = -1;
 174         return fp->private_data ? 0 : -ENOMEM;
 175 }
 176
 177 static ssize_t hfi1_file_write(struct file *fp, const char __user *data,
 178                                size_t count, loff_t *offset)
 179 {
 180         const struct hfi1_cmd __user *ucmd;
 181         struct hfi1_filedata *fd = fp->private_data;
 182         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 183         struct hfi1_cmd cmd;
 184         struct hfi1_user_info uinfo;
 185         struct hfi1_tid_info tinfo;
 186         unsigned long addr;
 187         ssize_t consumed = 0, copy = 0, ret = 0;
 188         void *dest = NULL;
 189         __u64 user_val = 0;
 190         int uctxt_required = 1;
 191         int must_be_root = 0;
 192
 193         if (count < sizeof(cmd)) {
 194                 ret = -EINVAL;
 195                 goto bail;
 196         }
 197
 198         ucmd = (const struct hfi1_cmd __user *)data;
 199         if (copy_from_user(&cmd, ucmd, sizeof(cmd))) {
 200                 ret = -EFAULT;
 201                 goto bail;
 202         }
 203
 204         consumed = sizeof(cmd);
 205
 206         switch (cmd.type) {
 207         case HFI1_CMD_ASSIGN_CTXT:
 208                 uctxt_required = 0;     /* assigned user context not required */
 209                 copy = sizeof(uinfo);
 210                 dest = &uinfo;
 211                 break;
 212         case HFI1_CMD_SDMA_STATUS_UPD:
 213         case HFI1_CMD_CREDIT_UPD:
 214                 copy = 0;
 215                 break;
 216         case HFI1_CMD_TID_UPDATE:
 217         case HFI1_CMD_TID_FREE:
 218         case HFI1_CMD_TID_INVAL_READ:
 219                 copy = sizeof(tinfo);
 220                 dest = &tinfo;
 221                 break;
 222         case HFI1_CMD_USER_INFO:
 223         case HFI1_CMD_RECV_CTRL:
 224         case HFI1_CMD_POLL_TYPE:
 225         case HFI1_CMD_ACK_EVENT:
 226         case HFI1_CMD_CTXT_INFO:
 227         case HFI1_CMD_SET_PKEY:
 228         case HFI1_CMD_CTXT_RESET:
 229                 copy = 0;
 230                 user_val = cmd.addr;
 231                 break;
 232         case HFI1_CMD_EP_INFO:
 233         case HFI1_CMD_EP_ERASE_CHIP:
 234         case HFI1_CMD_EP_ERASE_RANGE:
 235         case HFI1_CMD_EP_READ_RANGE:
 236         case HFI1_CMD_EP_WRITE_RANGE:
 237                 uctxt_required = 0;     /* assigned user context not required */
 238                 must_be_root = 1;       /* validate user */
 239                 copy = 0;
 240                 break;
 241         default:
 242                 ret = -EINVAL;
 243                 goto bail;
 244         }
 245
 246         /* If the command comes with user data, copy it. */
 247         if (copy) {
 248                 if (copy_from_user(dest, (void __user *)cmd.addr, copy)) {
 249                         ret = -EFAULT;
 250                         goto bail;
 251                 }
 252                 consumed += copy;
 253         }
 254
 255         /*
 256          * Make sure there is a uctxt when needed.
 257          */
 258         if (uctxt_required && !uctxt) {
 259                 ret = -EINVAL;
 260                 goto bail;
 261         }
 262
 263         /* only root can do these operations */
 264         if (must_be_root && !capable(CAP_SYS_ADMIN)) {
 265                 ret = -EPERM;
 266                 goto bail;
 267         }
 268
 269         switch (cmd.type) {
 270         case HFI1_CMD_ASSIGN_CTXT:
 271                 ret = assign_ctxt(fp, &uinfo);
 272                 if (ret < 0)
 273                         goto bail;
 274                 ret = setup_ctxt(fp);
 275                 if (ret)
 276                         goto bail;
 277                 ret = user_init(fp);
 278                 break;
 279         case HFI1_CMD_CTXT_INFO:
 280                 ret = get_ctxt_info(fp, (void __user *)(unsigned long)
 281                                     user_val, cmd.len);
 282                 break;
 283         case HFI1_CMD_USER_INFO:
 284                 ret = get_base_info(fp, (void __user *)(unsigned long)
 285                                     user_val, cmd.len);
 286                 break;
 287         case HFI1_CMD_SDMA_STATUS_UPD:
 288                 break;
 289         case HFI1_CMD_CREDIT_UPD:
 290                 if (uctxt && uctxt->sc)
 291                         sc_return_credits(uctxt->sc);
 292                 break;
 293         case HFI1_CMD_TID_UPDATE:
 294                 ret = hfi1_user_exp_rcv_setup(fp, &tinfo);
 295                 if (!ret) {
 296                         /*
 297                          * Copy the number of tidlist entries we used
 298                          * and the length of the buffer we registered.
 299                          * These fields are adjacent in the structure so
 300                          * we can copy them at the same time.
 301                          */
 302                         addr = (unsigned long)cmd.addr +
 303                                 offsetof(struct hfi1_tid_info, tidcnt);
 304                         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 305                                          sizeof(tinfo.tidcnt) +
 306                                          sizeof(tinfo.length)))
 307                                 ret = -EFAULT;
 308                 }
 309                 break;
 310         case HFI1_CMD_TID_INVAL_READ:
 311                 ret = hfi1_user_exp_rcv_invalid(fp, &tinfo);
 312                 if (ret)
 313                         break;
 314                 addr = (unsigned long)cmd.addr +
 315                         offsetof(struct hfi1_tid_info, tidcnt);
 316                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 317                                  sizeof(tinfo.tidcnt)))
 318                         ret = -EFAULT;
 319                 break;
 320         case HFI1_CMD_TID_FREE:
 321                 ret = hfi1_user_exp_rcv_clear(fp, &tinfo);
 322                 if (ret)
 323                         break;
 324                 addr = (unsigned long)cmd.addr +
 325                         offsetof(struct hfi1_tid_info, tidcnt);
 326                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
 327                                  sizeof(tinfo.tidcnt)))
 328                         ret = -EFAULT;
 329                 break;
 330         case HFI1_CMD_RECV_CTRL:
 331                 ret = manage_rcvq(uctxt, fd->subctxt, (int)user_val);
 332                 break;
 333         case HFI1_CMD_POLL_TYPE:
 334                 uctxt->poll_type = (typeof(uctxt->poll_type))user_val;
 335                 break;
 336         case HFI1_CMD_ACK_EVENT:
 337                 ret = user_event_ack(uctxt, fd->subctxt, user_val);
 338                 break;
 339         case HFI1_CMD_SET_PKEY:
 340                 if (HFI1_CAP_IS_USET(PKEY_CHECK))
 341                         ret = set_ctxt_pkey(uctxt, fd->subctxt, user_val);
 342                 else
 343                         ret = -EPERM;
 344                 break;
 345         case HFI1_CMD_CTXT_RESET: {
 346                 struct send_context *sc;
 347                 struct hfi1_devdata *dd;
 348
 349                 if (!uctxt || !uctxt->dd || !uctxt->sc) {
 350                         ret = -EINVAL;
 351                         break;
 352                 }
 353                 /*
 354                  * There is no protection here. User level has to
 355                  * guarantee that no one will be writing to the send
 356                  * context while it is being re-initialized.
 357                  * If user level breaks that guarantee, it will break
 358                  * it's own context and no one else's.
 359                  */
 360                 dd = uctxt->dd;
 361                 sc = uctxt->sc;
 362                 /*
 363                  * Wait until the interrupt handler has marked the
 364                  * context as halted or frozen. Report error if we time
 365                  * out.
 366                  */
 367                 wait_event_interruptible_timeout(
 368                         sc->halt_wait, (sc->flags & SCF_HALTED),
 369                         msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 370                 if (!(sc->flags & SCF_HALTED)) {
 371                         ret = -ENOLCK;
 372                         break;
 373                 }
 374                 /*
 375                  * If the send context was halted due to a Freeze,
 376                  * wait until the device has been "unfrozen" before
 377                  * resetting the context.
 378                  */
 379                 if (sc->flags & SCF_FROZEN) {
 380                         wait_event_interruptible_timeout(
 381                                 dd->event_queue,
 382                                 !(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
 383                                 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
 384                         if (dd->flags & HFI1_FROZEN) {
 385                                 ret = -ENOLCK;
 386                                 break;
 387                         }
 388                         if (dd->flags & HFI1_FORCED_FREEZE) {
 389                                 /*
 390                                  * Don't allow context reset if we are into
 391                                  * forced freeze
 392                                  */
 393                                 ret = -ENODEV;
 394                                 break;
 395                         }
 396                         sc_disable(sc);
 397                         ret = sc_enable(sc);
 398                         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
 399                                      uctxt->ctxt);
 400                 } else {
 401                         ret = sc_restart(sc);
 402                 }
 403                 if (!ret)
 404                         sc_return_credits(sc);
 405                 break;
 406         }
 407         case HFI1_CMD_EP_INFO:
 408         case HFI1_CMD_EP_ERASE_CHIP:
 409         case HFI1_CMD_EP_ERASE_RANGE:
 410         case HFI1_CMD_EP_READ_RANGE:
 411         case HFI1_CMD_EP_WRITE_RANGE:
 412                 ret = handle_eprom_command(fp, &cmd);
 413                 break;
 414         }
 415
 416         if (ret >= 0)
 417                 ret = consumed;
 418 bail:
 419         return ret;
 420 }
 421
 422 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
 423 {
 424         struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
 425         struct hfi1_user_sdma_pkt_q *pq = fd->pq;
 426         struct hfi1_user_sdma_comp_q *cq = fd->cq;
 427         int ret = 0, done = 0, reqs = 0;
 428         unsigned long dim = from->nr_segs;
 429
 430         if (!cq || !pq) {
 431                 ret = -EIO;
 432                 goto done;
 433         }
 434
 435         if (!iter_is_iovec(from) || !dim) {
 436                 ret = -EINVAL;
 437                 goto done;
 438         }
 439
 440         hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
 441                   fd->uctxt->ctxt, fd->subctxt, dim);
 442
 443         if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
 444                 ret = -ENOSPC;
 445                 goto done;
 446         }
 447
 448         while (dim) {
 449                 unsigned long count = 0;
 450
 451                 ret = hfi1_user_sdma_process_request(
 452                         kiocb->ki_filp, (struct iovec *)(from->iov + done),
 453                         dim, &count);
 454                 if (ret)
 455                         goto done;
 456                 dim -= count;
 457                 done += count;
 458                 reqs++;
 459         }
 460 done:
 461         return ret ? ret : reqs;
 462 }
 463
 464 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
 465 {
 466         struct hfi1_filedata *fd = fp->private_data;
 467         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 468         struct hfi1_devdata *dd;
 469         unsigned long flags, pfn;
 470         u64 token = vma->vm_pgoff << PAGE_SHIFT,
 471                 memaddr = 0;
 472         u8 subctxt, mapio = 0, vmf = 0, type;
 473         ssize_t memlen = 0;
 474         int ret = 0;
 475         u16 ctxt;
 476
 477         if (!is_valid_mmap(token) || !uctxt ||
 478             !(vma->vm_flags & VM_SHARED)) {
 479                 ret = -EINVAL;
 480                 goto done;
 481         }
 482         dd = uctxt->dd;
 483         ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
 484         subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
 485         type = HFI1_MMAP_TOKEN_GET(TYPE, token);
 486         if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
 487                 ret = -EINVAL;
 488                 goto done;
 489         }
 490
 491         flags = vma->vm_flags;
 492
 493         switch (type) {
 494         case PIO_BUFS:
 495         case PIO_BUFS_SOP:
 496                 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
 497                                 /* chip pio base */
 498                            (uctxt->sc->hw_context * BIT(16))) +
 499                                 /* 64K PIO space / ctxt */
 500                         (type == PIO_BUFS_SOP ?
 501                                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
 502                 /*
 503                  * Map only the amount allocated to the context, not the
 504                  * entire available context's PIO space.
 505                  */
 506                 memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
 507                 flags &= ~VM_MAYREAD;
 508                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 509                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 510                 mapio = 1;
 511                 break;
 512         case PIO_CRED:
 513                 if (flags & VM_WRITE) {
 514                         ret = -EPERM;
 515                         goto done;
 516                 }
 517                 /*
 518                  * The credit return location for this context could be on the
 519                  * second or third page allocated for credit returns (if number
 520                  * of enabled contexts > 64 and 128 respectively).
 521                  */
 522                 memaddr = dd->cr_base[uctxt->numa_id].pa +
 523                         (((u64)uctxt->sc->hw_free -
 524                           (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
 525                 memlen = PAGE_SIZE;
 526                 flags &= ~VM_MAYWRITE;
 527                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 528                 /*
 529                  * The driver has already allocated memory for credit
 530                  * returns and programmed it into the chip. Has that
 531                  * memory been flagged as non-cached?
 532                  */
 533                 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
 534                 mapio = 1;
 535                 break;
 536         case RCV_HDRQ:
 537                 memaddr = uctxt->rcvhdrq_phys;
 538                 memlen = uctxt->rcvhdrq_size;
 539                 break;
 540         case RCV_EGRBUF: {
 541                 unsigned long addr;
 542                 int i;
 543                 /*
 544                  * The RcvEgr buffer need to be handled differently
 545                  * as multiple non-contiguous pages need to be mapped
 546                  * into the user process.
 547                  */
 548                 memlen = uctxt->egrbufs.size;
 549                 if ((vma->vm_end - vma->vm_start) != memlen) {
 550                         dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
 551                                    (vma->vm_end - vma->vm_start), memlen);
 552                         ret = -EINVAL;
 553                         goto done;
 554                 }
 555                 if (vma->vm_flags & VM_WRITE) {
 556                         ret = -EPERM;
 557                         goto done;
 558                 }
 559                 vma->vm_flags &= ~VM_MAYWRITE;
 560                 addr = vma->vm_start;
 561                 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
 562                         ret = remap_pfn_range(
 563                                 vma, addr,
 564                                 uctxt->egrbufs.buffers[i].phys >> PAGE_SHIFT,
 565                                 uctxt->egrbufs.buffers[i].len,
 566                                 vma->vm_page_prot);
 567                         if (ret < 0)
 568                                 goto done;
 569                         addr += uctxt->egrbufs.buffers[i].len;
 570                 }
 571                 ret = 0;
 572                 goto done;
 573         }
 574         case UREGS:
 575                 /*
 576                  * Map only the page that contains this context's user
 577                  * registers.
 578                  */
 579                 memaddr = (unsigned long)
 580                         (dd->physaddr + RXE_PER_CONTEXT_USER)
 581                         + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
 582                 /*
 583                  * TidFlow table is on the same page as the rest of the
 584                  * user registers.
 585                  */
 586                 memlen = PAGE_SIZE;
 587                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 588                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 589                 mapio = 1;
 590                 break;
 591         case EVENTS:
 592                 /*
 593                  * Use the page where this context's flags are. User level
 594                  * knows where it's own bitmap is within the page.
 595                  */
 596                 memaddr = (unsigned long)(dd->events +
 597                                           ((uctxt->ctxt - dd->first_user_ctxt) *
 598                                            HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
 599                 memlen = PAGE_SIZE;
 600                 /*
 601                  * v3.7 removes VM_RESERVED but the effect is kept by
 602                  * using VM_IO.
 603                  */
 604                 flags |= VM_IO | VM_DONTEXPAND;
 605                 vmf = 1;
 606                 break;
 607         case STATUS:
 608                 memaddr = kvirt_to_phys((void *)dd->status);
 609                 memlen = PAGE_SIZE;
 610                 flags |= VM_IO | VM_DONTEXPAND;
 611                 break;
 612         case RTAIL:
 613                 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
 614                         /*
 615                          * If the memory allocation failed, the context alloc
 616                          * also would have failed, so we would never get here
 617                          */
 618                         ret = -EINVAL;
 619                         goto done;
 620                 }
 621                 if (flags & VM_WRITE) {
 622                         ret = -EPERM;
 623                         goto done;
 624                 }
 625                 memaddr = uctxt->rcvhdrqtailaddr_phys;
 626                 memlen = PAGE_SIZE;
 627                 flags &= ~VM_MAYWRITE;
 628                 break;
 629         case SUBCTXT_UREGS:
 630                 memaddr = (u64)uctxt->subctxt_uregbase;
 631                 memlen = PAGE_SIZE;
 632                 flags |= VM_IO | VM_DONTEXPAND;
 633                 vmf = 1;
 634                 break;
 635         case SUBCTXT_RCV_HDRQ:
 636                 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
 637                 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
 638                 flags |= VM_IO | VM_DONTEXPAND;
 639                 vmf = 1;
 640                 break;
 641         case SUBCTXT_EGRBUF:
 642                 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
 643                 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
 644                 flags |= VM_IO | VM_DONTEXPAND;
 645                 flags &= ~VM_MAYWRITE;
 646                 vmf = 1;
 647                 break;
 648         case SDMA_COMP: {
 649                 struct hfi1_user_sdma_comp_q *cq = fd->cq;
 650
 651                 if (!cq) {
 652                         ret = -EFAULT;
 653                         goto done;
 654                 }
 655                 memaddr = (u64)cq->comps;
 656                 memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
 657                 flags |= VM_IO | VM_DONTEXPAND;
 658                 vmf = 1;
 659                 break;
 660         }
 661         default:
 662                 ret = -EINVAL;
 663                 break;
 664         }
 665
 666         if ((vma->vm_end - vma->vm_start) != memlen) {
 667                 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
 668                           uctxt->ctxt, fd->subctxt,
 669                           (vma->vm_end - vma->vm_start), memlen);
 670                 ret = -EINVAL;
 671                 goto done;
 672         }
 673
 674         vma->vm_flags = flags;
 675         hfi1_cdbg(PROC,
 676                   "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
 677                     ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
 678                     vma->vm_end - vma->vm_start, vma->vm_flags);
 679         pfn = (unsigned long)(memaddr >> PAGE_SHIFT);
 680         if (vmf) {
 681                 vma->vm_pgoff = pfn;
 682                 vma->vm_ops = &vm_ops;
 683                 ret = 0;
 684         } else if (mapio) {
 685                 ret = io_remap_pfn_range(vma, vma->vm_start, pfn, memlen,
 686                                          vma->vm_page_prot);
 687         } else {
 688                 ret = remap_pfn_range(vma, vma->vm_start, pfn, memlen,
 689                                       vma->vm_page_prot);
 690         }
 691 done:
 692         return ret;
 693 }
 694
 695 /*
 696  * Local (non-chip) user memory is not mapped right away but as it is
 697  * accessed by the user-level code.
 698  */
 699 static int vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 700 {
 701         struct page *page;
 702
 703         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 704         if (!page)
 705                 return VM_FAULT_SIGBUS;
 706
 707         get_page(page);
 708         vmf->page = page;
 709
 710         return 0;
 711 }
 712
 713 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
 714 {
 715         struct hfi1_ctxtdata *uctxt;
 716         unsigned pollflag;
 717
 718         uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
 719         if (!uctxt)
 720                 pollflag = POLLERR;
 721         else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
 722                 pollflag = poll_urgent(fp, pt);
 723         else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
 724                 pollflag = poll_next(fp, pt);
 725         else /* invalid */
 726                 pollflag = POLLERR;
 727
 728         return pollflag;
 729 }
 730
 731 static int hfi1_file_close(struct inode *inode, struct file *fp)
 732 {
 733         struct hfi1_filedata *fdata = fp->private_data;
 734         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 735         struct hfi1_devdata *dd;
 736         unsigned long flags, *ev;
 737
 738         fp->private_data = NULL;
 739
 740         if (!uctxt)
 741                 goto done;
 742
 743         hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
 744         dd = uctxt->dd;
 745         mutex_lock(&hfi1_mutex);
 746
 747         flush_wc();
 748         /* drain user sdma queue */
 749         hfi1_user_sdma_free_queues(fdata);
 750
 751         /* release the cpu */
 752         hfi1_put_proc_affinity(dd, fdata->rec_cpu_num);
 753
 754         /*
 755          * Clear any left over, unhandled events so the next process that
 756          * gets this context doesn't get confused.
 757          */
 758         ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
 759                            HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
 760         *ev = 0;
 761
 762         if (--uctxt->cnt) {
 763                 uctxt->active_slaves &= ~(1 << fdata->subctxt);
 764                 uctxt->subpid[fdata->subctxt] = 0;
 765                 mutex_unlock(&hfi1_mutex);
 766                 goto done;
 767         }
 768
 769         spin_lock_irqsave(&dd->uctxt_lock, flags);
 770         /*
 771          * Disable receive context and interrupt available, reset all
 772          * RcvCtxtCtrl bits to default values.
 773          */
 774         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
 775                      HFI1_RCVCTRL_TIDFLOW_DIS |
 776                      HFI1_RCVCTRL_INTRAVAIL_DIS |
 777                      HFI1_RCVCTRL_TAILUPD_DIS |
 778                      HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
 779                      HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
 780                      HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
 781         /* Clear the context's J_KEY */
 782         hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
 783         /*
 784          * Reset context integrity checks to default.
 785          * (writes to CSRs probably belong in chip.c)
 786          */
 787         write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
 788                         hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
 789         sc_disable(uctxt->sc);
 790         uctxt->pid = 0;
 791         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 792
 793         dd->rcd[uctxt->ctxt] = NULL;
 794         uctxt->rcvwait_to = 0;
 795         uctxt->piowait_to = 0;
 796         uctxt->rcvnowait = 0;
 797         uctxt->pionowait = 0;
 798         uctxt->event_flags = 0;
 799
 800         hfi1_user_exp_rcv_free(fdata);
 801         hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
 802
 803         hfi1_stats.sps_ctxts--;
 804         if (++dd->freectxts == dd->num_user_contexts)
 805                 aspm_enable_all(dd);
 806         mutex_unlock(&hfi1_mutex);
 807         hfi1_free_ctxtdata(dd, uctxt);
 808 done:
 809         kfree(fdata);
 810         return 0;
 811 }
 812
 813 /*
 814  * Convert kernel *virtual* addresses to physical addresses.
 815  * This is used to vmalloc'ed addresses.
 816  */
 817 static u64 kvirt_to_phys(void *addr)
 818 {
 819         struct page *page;
 820         u64 paddr = 0;
 821
 822         page = vmalloc_to_page(addr);
 823         if (page)
 824                 paddr = page_to_pfn(page) << PAGE_SHIFT;
 825
 826         return paddr;
 827 }
 828
 829 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
 830 {
 831         int i_minor, ret = 0;
 832         unsigned swmajor, swminor, alg = HFI1_ALG_ACROSS;
 833
 834         swmajor = uinfo->userversion >> 16;
 835         if (swmajor != HFI1_USER_SWMAJOR) {
 836                 ret = -ENODEV;
 837                 goto done;
 838         }
 839
 840         swminor = uinfo->userversion & 0xffff;
 841
 842         if (uinfo->hfi1_alg < HFI1_ALG_COUNT)
 843                 alg = uinfo->hfi1_alg;
 844
 845         mutex_lock(&hfi1_mutex);
 846         /* First, lets check if we need to setup a shared context? */
 847         if (uinfo->subctxt_cnt) {
 848                 struct hfi1_filedata *fd = fp->private_data;
 849
 850                 ret = find_shared_ctxt(fp, uinfo);
 851                 if (ret < 0)
 852                         goto done_unlock;
 853                 if (ret)
 854                         fd->rec_cpu_num = hfi1_get_proc_affinity(
 855                                 fd->uctxt->dd, fd->uctxt->numa_id);
 856         }
 857
 858         /*
 859          * We execute the following block if we couldn't find a
 860          * shared context or if context sharing is not required.
 861          */
 862         if (!ret) {
 863                 i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
 864                 ret = get_user_context(fp, uinfo, i_minor - 1, alg);
 865         }
 866 done_unlock:
 867         mutex_unlock(&hfi1_mutex);
 868 done:
 869         return ret;
 870 }
 871
 872 /* return true if the device available for general use */
 873 static int usable_device(struct hfi1_devdata *dd)
 874 {
 875         struct hfi1_pportdata *ppd = dd->pport;
 876
 877         return driver_lstate(ppd) == IB_PORT_ACTIVE;
 878 }
 879
 880 static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
 881                             int devno, unsigned alg)
 882 {
 883         struct hfi1_devdata *dd = NULL;
 884         int ret = 0, devmax, npresent, nup, dev;
 885
 886         devmax = hfi1_count_units(&npresent, &nup);
 887         if (!npresent) {
 888                 ret = -ENXIO;
 889                 goto done;
 890         }
 891         if (!nup) {
 892                 ret = -ENETDOWN;
 893                 goto done;
 894         }
 895         if (devno >= 0) {
 896                 dd = hfi1_lookup(devno);
 897                 if (!dd)
 898                         ret = -ENODEV;
 899                 else if (!dd->freectxts)
 900                         ret = -EBUSY;
 901         } else {
 902                 struct hfi1_devdata *pdd;
 903
 904                 if (alg == HFI1_ALG_ACROSS) {
 905                         unsigned free = 0U;
 906
 907                         for (dev = 0; dev < devmax; dev++) {
 908                                 pdd = hfi1_lookup(dev);
 909                                 if (!pdd)
 910                                         continue;
 911                                 if (!usable_device(pdd))
 912                                         continue;
 913                                 if (pdd->freectxts &&
 914                                     pdd->freectxts > free) {
 915                                         dd = pdd;
 916                                         free = pdd->freectxts;
 917                                 }
 918                         }
 919                 } else {
 920                         for (dev = 0; dev < devmax; dev++) {
 921                                 pdd = hfi1_lookup(dev);
 922                                 if (!pdd)
 923                                         continue;
 924                                 if (!usable_device(pdd))
 925                                         continue;
 926                                 if (pdd->freectxts) {
 927                                         dd = pdd;
 928                                         break;
 929                                 }
 930                         }
 931                 }
 932                 if (!dd)
 933                         ret = -EBUSY;
 934         }
 935 done:
 936         return ret ? ret : allocate_ctxt(fp, dd, uinfo);
 937 }
 938
 939 static int find_shared_ctxt(struct file *fp,
 940                             const struct hfi1_user_info *uinfo)
 941 {
 942         int devmax, ndev, i;
 943         int ret = 0;
 944         struct hfi1_filedata *fd = fp->private_data;
 945
 946         devmax = hfi1_count_units(NULL, NULL);
 947
 948         for (ndev = 0; ndev < devmax; ndev++) {
 949                 struct hfi1_devdata *dd = hfi1_lookup(ndev);
 950
 951                 if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
 952                         continue;
 953                 for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
 954                         struct hfi1_ctxtdata *uctxt = dd->rcd[i];
 955
 956                         /* Skip ctxts which are not yet open */
 957                         if (!uctxt || !uctxt->cnt)
 958                                 continue;
 959                         /* Skip ctxt if it doesn't match the requested one */
 960                         if (memcmp(uctxt->uuid, uinfo->uuid,
 961                                    sizeof(uctxt->uuid)) ||
 962                             uctxt->jkey != generate_jkey(current_uid()) ||
 963                             uctxt->subctxt_id != uinfo->subctxt_id ||
 964                             uctxt->subctxt_cnt != uinfo->subctxt_cnt)
 965                                 continue;
 966
 967                         /* Verify the sharing process matches the master */
 968                         if (uctxt->userversion != uinfo->userversion ||
 969                             uctxt->cnt >= uctxt->subctxt_cnt) {
 970                                 ret = -EINVAL;
 971                                 goto done;
 972                         }
 973                         fd->uctxt = uctxt;
 974                         fd->subctxt  = uctxt->cnt++;
 975                         uctxt->subpid[fd->subctxt] = current->pid;
 976                         uctxt->active_slaves |= 1 << fd->subctxt;
 977                         ret = 1;
 978                         goto done;
 979                 }
 980         }
 981
 982 done:
 983         return ret;
 984 }
 985
 986 static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
 987                          struct hfi1_user_info *uinfo)
 988 {
 989         struct hfi1_filedata *fd = fp->private_data;
 990         struct hfi1_ctxtdata *uctxt;
 991         unsigned ctxt;
 992         int ret, numa;
 993
 994         if (dd->flags & HFI1_FROZEN) {
 995                 /*
 996                  * Pick an error that is unique from all other errors
 997                  * that are returned so the user process knows that
 998                  * it tried to allocate while the SPC was frozen.  It
 999                  * it should be able to retry with success in a short
1000                  * while.
1001                  */
1002                 return -EIO;
1003         }
1004
1005         for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
1006                 if (!dd->rcd[ctxt])
1007                         break;
1008
1009         if (ctxt == dd->num_rcv_contexts)
1010                 return -EBUSY;
1011
1012         fd->rec_cpu_num = hfi1_get_proc_affinity(dd, -1);
1013         if (fd->rec_cpu_num != -1)
1014                 numa = cpu_to_node(fd->rec_cpu_num);
1015         else
1016                 numa = numa_node_id();
1017         uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
1018         if (!uctxt) {
1019                 dd_dev_err(dd,
1020                            "Unable to allocate ctxtdata memory, failing open\n");
1021                 return -ENOMEM;
1022         }
1023         hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
1024                   uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
1025                   uctxt->numa_id);
1026
1027         /*
1028          * Allocate and enable a PIO send context.
1029          */
1030         uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
1031                              uctxt->dd->node);
1032         if (!uctxt->sc)
1033                 return -ENOMEM;
1034
1035         hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
1036                   uctxt->sc->hw_context);
1037         ret = sc_enable(uctxt->sc);
1038         if (ret)
1039                 return ret;
1040         /*
1041          * Setup shared context resources if the user-level has requested
1042          * shared contexts and this is the 'master' process.
1043          * This has to be done here so the rest of the sub-contexts find the
1044          * proper master.
1045          */
1046         if (uinfo->subctxt_cnt && !fd->subctxt) {
1047                 ret = init_subctxts(uctxt, uinfo);
1048                 /*
1049                  * On error, we don't need to disable and de-allocate the
1050                  * send context because it will be done during file close
1051                  */
1052                 if (ret)
1053                         return ret;
1054         }
1055         uctxt->userversion = uinfo->userversion;
1056         uctxt->pid = current->pid;
1057         uctxt->flags = HFI1_CAP_UGET(MASK);
1058         init_waitqueue_head(&uctxt->wait);
1059         strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1060         memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1061         uctxt->jkey = generate_jkey(current_uid());
1062         INIT_LIST_HEAD(&uctxt->sdma_queues);
1063         spin_lock_init(&uctxt->sdma_qlock);
1064         hfi1_stats.sps_ctxts++;
1065         /*
1066          * Disable ASPM when there are open user/PSM contexts to avoid
1067          * issues with ASPM L1 exit latency
1068          */
1069         if (dd->freectxts-- == dd->num_user_contexts)
1070                 aspm_disable_all(dd);
1071         fd->uctxt = uctxt;
1072
1073         return 0;
1074 }
1075
1076 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1077                          const struct hfi1_user_info *uinfo)
1078 {
1079         unsigned num_subctxts;
1080
1081         num_subctxts = uinfo->subctxt_cnt;
1082         if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1083                 return -EINVAL;
1084
1085         uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1086         uctxt->subctxt_id = uinfo->subctxt_id;
1087         uctxt->active_slaves = 1;
1088         uctxt->redirect_seq_cnt = 1;
1089         set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1090
1091         return 0;
1092 }
1093
1094 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1095 {
1096         int ret = 0;
1097         unsigned num_subctxts = uctxt->subctxt_cnt;
1098
1099         uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1100         if (!uctxt->subctxt_uregbase) {
1101                 ret = -ENOMEM;
1102                 goto bail;
1103         }
1104         /* We can take the size of the RcvHdr Queue from the master */
1105         uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1106                                                   num_subctxts);
1107         if (!uctxt->subctxt_rcvhdr_base) {
1108                 ret = -ENOMEM;
1109                 goto bail_ureg;
1110         }
1111
1112         uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1113                                                 num_subctxts);
1114         if (!uctxt->subctxt_rcvegrbuf) {
1115                 ret = -ENOMEM;
1116                 goto bail_rhdr;
1117         }
1118         goto bail;
1119 bail_rhdr:
1120         vfree(uctxt->subctxt_rcvhdr_base);
1121 bail_ureg:
1122         vfree(uctxt->subctxt_uregbase);
1123         uctxt->subctxt_uregbase = NULL;
1124 bail:
1125         return ret;
1126 }
1127
1128 static int user_init(struct file *fp)
1129 {
1130         int ret;
1131         unsigned int rcvctrl_ops = 0;
1132         struct hfi1_filedata *fd = fp->private_data;
1133         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1134
1135         /* make sure that the context has already been setup */
1136         if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags)) {
1137                 ret = -EFAULT;
1138                 goto done;
1139         }
1140
1141         /*
1142          * Subctxts don't need to initialize anything since master
1143          * has done it.
1144          */
1145         if (fd->subctxt) {
1146                 ret = wait_event_interruptible(uctxt->wait, !test_bit(
1147                                                HFI1_CTXT_MASTER_UNINIT,
1148                                                &uctxt->event_flags));
1149                 goto expected;
1150         }
1151
1152         /* initialize poll variables... */
1153         uctxt->urgent = 0;
1154         uctxt->urgent_poll = 0;
1155
1156         /*
1157          * Now enable the ctxt for receive.
1158          * For chips that are set to DMA the tail register to memory
1159          * when they change (and when the update bit transitions from
1160          * 0 to 1.  So for those chips, we turn it off and then back on.
1161          * This will (very briefly) affect any other open ctxts, but the
1162          * duration is very short, and therefore isn't an issue.  We
1163          * explicitly set the in-memory tail copy to 0 beforehand, so we
1164          * don't have to wait to be sure the DMA update has happened
1165          * (chip resets head/tail to 0 on transition to enable).
1166          */
1167         if (uctxt->rcvhdrtail_kvaddr)
1168                 clear_rcvhdrtail(uctxt);
1169
1170         /* Setup J_KEY before enabling the context */
1171         hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1172
1173         rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1174         if (HFI1_CAP_KGET_MASK(uctxt->flags, HDRSUPP))
1175                 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1176         /*
1177          * Ignore the bit in the flags for now until proper
1178          * support for multiple packet per rcv array entry is
1179          * added.
1180          */
1181         if (!HFI1_CAP_KGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1182                 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1183         if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1184                 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1185         if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1186                 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1187         /*
1188          * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1189          * We can't rely on the correct value to be set from prior
1190          * uses of the chip or ctxt. Therefore, add the rcvctrl op
1191          * for both cases.
1192          */
1193         if (HFI1_CAP_KGET_MASK(uctxt->flags, DMA_RTAIL))
1194                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1195         else
1196                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1197         hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1198
1199         /* Notify any waiting slaves */
1200         if (uctxt->subctxt_cnt) {
1201                 clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1202                 wake_up(&uctxt->wait);
1203         }
1204
1205 expected:
1206         /*
1207          * Expected receive has to be setup for all processes (including
1208          * shared contexts). However, it has to be done after the master
1209          * context has been fully configured as it depends on the
1210          * eager/expected split of the RcvArray entries.
1211          * Setting it up here ensures that the subcontexts will be waiting
1212          * (due to the above wait_event_interruptible() until the master
1213          * is setup.
1214          */
1215         ret = hfi1_user_exp_rcv_init(fp);
1216 done:
1217         return ret;
1218 }
1219
1220 static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1221 {
1222         struct hfi1_ctxt_info cinfo;
1223         struct hfi1_filedata *fd = fp->private_data;
1224         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1225         int ret = 0;
1226
1227         memset(&cinfo, 0, sizeof(cinfo));
1228         ret = hfi1_get_base_kinfo(uctxt, &cinfo);
1229         if (ret < 0)
1230                 goto done;
1231         cinfo.num_active = hfi1_count_active_units();
1232         cinfo.unit = uctxt->dd->unit;
1233         cinfo.ctxt = uctxt->ctxt;
1234         cinfo.subctxt = fd->subctxt;
1235         cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1236                                 uctxt->dd->rcv_entries.group_size) +
1237                 uctxt->expected_count;
1238         cinfo.credits = uctxt->sc->credits;
1239         cinfo.numa_node = uctxt->numa_id;
1240         cinfo.rec_cpu = fd->rec_cpu_num;
1241         cinfo.send_ctxt = uctxt->sc->hw_context;
1242
1243         cinfo.egrtids = uctxt->egrbufs.alloced;
1244         cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1245         cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1246         cinfo.sdma_ring_size = fd->cq->nentries;
1247         cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1248
1249         trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1250         if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1251                 ret = -EFAULT;
1252 done:
1253         return ret;
1254 }
1255
1256 static int setup_ctxt(struct file *fp)
1257 {
1258         struct hfi1_filedata *fd = fp->private_data;
1259         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1260         struct hfi1_devdata *dd = uctxt->dd;
1261         int ret = 0;
1262
1263         /*
1264          * Context should be set up only once (including allocation and
1265          * programming of eager buffers. This is done if context sharing
1266          * is not requested or by the master process.
1267          */
1268         if (!uctxt->subctxt_cnt || !fd->subctxt) {
1269                 ret = hfi1_init_ctxt(uctxt->sc);
1270                 if (ret)
1271                         goto done;
1272
1273                 /* Now allocate the RcvHdr queue and eager buffers. */
1274                 ret = hfi1_create_rcvhdrq(dd, uctxt);
1275                 if (ret)
1276                         goto done;
1277                 ret = hfi1_setup_eagerbufs(uctxt);
1278                 if (ret)
1279                         goto done;
1280                 if (uctxt->subctxt_cnt && !fd->subctxt) {
1281                         ret = setup_subctxt(uctxt);
1282                         if (ret)
1283                                 goto done;
1284                 }
1285         }
1286         ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1287         if (ret)
1288                 goto done;
1289
1290         set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1291 done:
1292         return ret;
1293 }
1294
1295 static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1296 {
1297         struct hfi1_base_info binfo;
1298         struct hfi1_filedata *fd = fp->private_data;
1299         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1300         struct hfi1_devdata *dd = uctxt->dd;
1301         ssize_t sz;
1302         unsigned offset;
1303         int ret = 0;
1304
1305         trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1306
1307         memset(&binfo, 0, sizeof(binfo));
1308         binfo.hw_version = dd->revision;
1309         binfo.sw_version = HFI1_KERN_SWVERSION;
1310         binfo.bthqp = kdeth_qp;
1311         binfo.jkey = uctxt->jkey;
1312         /*
1313          * If more than 64 contexts are enabled the allocated credit
1314          * return will span two or three contiguous pages. Since we only
1315          * map the page containing the context's credit return address,
1316          * we need to calculate the offset in the proper page.
1317          */
1318         offset = ((u64)uctxt->sc->hw_free -
1319                   (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1320         binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1321                                                 fd->subctxt, offset);
1322         binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1323                                             fd->subctxt,
1324                                             uctxt->sc->base_addr);
1325         binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1326                                                 uctxt->ctxt,
1327                                                 fd->subctxt,
1328                                                 uctxt->sc->base_addr);
1329         binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1330                                                fd->subctxt,
1331                                                uctxt->rcvhdrq);
1332         binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1333                                                fd->subctxt,
1334                                                uctxt->egrbufs.rcvtids[0].phys);
1335         binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1336                                                  fd->subctxt, 0);
1337         /*
1338          * user regs are at
1339          * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1340          */
1341         binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1342                                             fd->subctxt, 0);
1343         offset = offset_in_page((((uctxt->ctxt - dd->first_user_ctxt) *
1344                     HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1345                   sizeof(*dd->events));
1346         binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1347                                               fd->subctxt,
1348                                               offset);
1349         binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1350                                               fd->subctxt,
1351                                               dd->status);
1352         if (HFI1_CAP_IS_USET(DMA_RTAIL))
1353                 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1354                                                        fd->subctxt, 0);
1355         if (uctxt->subctxt_cnt) {
1356                 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1357                                                         uctxt->ctxt,
1358                                                         fd->subctxt, 0);
1359                 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1360                                                          uctxt->ctxt,
1361                                                          fd->subctxt, 0);
1362                 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1363                                                          uctxt->ctxt,
1364                                                          fd->subctxt, 0);
1365         }
1366         sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1367         if (copy_to_user(ubase, &binfo, sz))
1368                 ret = -EFAULT;
1369         return ret;
1370 }
1371
1372 static unsigned int poll_urgent(struct file *fp,
1373                                 struct poll_table_struct *pt)
1374 {
1375         struct hfi1_filedata *fd = fp->private_data;
1376         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1377         struct hfi1_devdata *dd = uctxt->dd;
1378         unsigned pollflag;
1379
1380         poll_wait(fp, &uctxt->wait, pt);
1381
1382         spin_lock_irq(&dd->uctxt_lock);
1383         if (uctxt->urgent != uctxt->urgent_poll) {
1384                 pollflag = POLLIN | POLLRDNORM;
1385                 uctxt->urgent_poll = uctxt->urgent;
1386         } else {
1387                 pollflag = 0;
1388                 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1389         }
1390         spin_unlock_irq(&dd->uctxt_lock);
1391
1392         return pollflag;
1393 }
1394
1395 static unsigned int poll_next(struct file *fp,
1396                               struct poll_table_struct *pt)
1397 {
1398         struct hfi1_filedata *fd = fp->private_data;
1399         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1400         struct hfi1_devdata *dd = uctxt->dd;
1401         unsigned pollflag;
1402
1403         poll_wait(fp, &uctxt->wait, pt);
1404
1405         spin_lock_irq(&dd->uctxt_lock);
1406         if (hdrqempty(uctxt)) {
1407                 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1408                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1409                 pollflag = 0;
1410         } else {
1411                 pollflag = POLLIN | POLLRDNORM;
1412         }
1413         spin_unlock_irq(&dd->uctxt_lock);
1414
1415         return pollflag;
1416 }
1417
1418 /*
1419  * Find all user contexts in use, and set the specified bit in their
1420  * event mask.
1421  * See also find_ctxt() for a similar use, that is specific to send buffers.
1422  */
1423 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1424 {
1425         struct hfi1_ctxtdata *uctxt;
1426         struct hfi1_devdata *dd = ppd->dd;
1427         unsigned ctxt;
1428         int ret = 0;
1429         unsigned long flags;
1430
1431         if (!dd->events) {
1432                 ret = -EINVAL;
1433                 goto done;
1434         }
1435
1436         spin_lock_irqsave(&dd->uctxt_lock, flags);
1437         for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1438              ctxt++) {
1439                 uctxt = dd->rcd[ctxt];
1440                 if (uctxt) {
1441                         unsigned long *evs = dd->events +
1442                                 (uctxt->ctxt - dd->first_user_ctxt) *
1443                                 HFI1_MAX_SHARED_CTXTS;
1444                         int i;
1445                         /*
1446                          * subctxt_cnt is 0 if not shared, so do base
1447                          * separately, first, then remaining subctxt, if any
1448                          */
1449                         set_bit(evtbit, evs);
1450                         for (i = 1; i < uctxt->subctxt_cnt; i++)
1451                                 set_bit(evtbit, evs + i);
1452                 }
1453         }
1454         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1455 done:
1456         return ret;
1457 }
1458
1459 /**
1460  * manage_rcvq - manage a context's receive queue
1461  * @uctxt: the context
1462  * @subctxt: the sub-context
1463  * @start_stop: action to carry out
1464  *
1465  * start_stop == 0 disables receive on the context, for use in queue
1466  * overflow conditions.  start_stop==1 re-enables, to be used to
1467  * re-init the software copy of the head register
1468  */
1469 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1470                        int start_stop)
1471 {
1472         struct hfi1_devdata *dd = uctxt->dd;
1473         unsigned int rcvctrl_op;
1474
1475         if (subctxt)
1476                 goto bail;
1477         /* atomically clear receive enable ctxt. */
1478         if (start_stop) {
1479                 /*
1480                  * On enable, force in-memory copy of the tail register to
1481                  * 0, so that protocol code doesn't have to worry about
1482                  * whether or not the chip has yet updated the in-memory
1483                  * copy or not on return from the system call. The chip
1484                  * always resets it's tail register back to 0 on a
1485                  * transition from disabled to enabled.
1486                  */
1487                 if (uctxt->rcvhdrtail_kvaddr)
1488                         clear_rcvhdrtail(uctxt);
1489                 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1490         } else {
1491                 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1492         }
1493         hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1494         /* always; new head should be equal to new tail; see above */
1495 bail:
1496         return 0;
1497 }
1498
1499 /*
1500  * clear the event notifier events for this context.
1501  * User process then performs actions appropriate to bit having been
1502  * set, if desired, and checks again in future.
1503  */
1504 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1505                           unsigned long events)
1506 {
1507         int i;
1508         struct hfi1_devdata *dd = uctxt->dd;
1509         unsigned long *evs;
1510
1511         if (!dd->events)
1512                 return 0;
1513
1514         evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1515                             HFI1_MAX_SHARED_CTXTS) + subctxt;
1516
1517         for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1518                 if (!test_bit(i, &events))
1519                         continue;
1520                 clear_bit(i, evs);
1521         }
1522         return 0;
1523 }
1524
1525 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1526                          u16 pkey)
1527 {
1528         int ret = -ENOENT, i, intable = 0;
1529         struct hfi1_pportdata *ppd = uctxt->ppd;
1530         struct hfi1_devdata *dd = uctxt->dd;
1531
1532         if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1533                 ret = -EINVAL;
1534                 goto done;
1535         }
1536
1537         for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1538                 if (pkey == ppd->pkeys[i]) {
1539                         intable = 1;
1540                         break;
1541                 }
1542
1543         if (intable)
1544                 ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1545 done:
1546         return ret;
1547 }
1548
1549 static int ui_open(struct inode *inode, struct file *filp)
1550 {
1551         struct hfi1_devdata *dd;
1552
1553         dd = container_of(inode->i_cdev, struct hfi1_devdata, ui_cdev);
1554         filp->private_data = dd; /* for other methods */
1555         return 0;
1556 }
1557
1558 static int ui_release(struct inode *inode, struct file *filp)
1559 {
1560         /* nothing to do */
1561         return 0;
1562 }
1563
1564 static loff_t ui_lseek(struct file *filp, loff_t offset, int whence)
1565 {
1566         struct hfi1_devdata *dd = filp->private_data;
1567
1568         switch (whence) {
1569         case SEEK_SET:
1570                 break;
1571         case SEEK_CUR:
1572                 offset += filp->f_pos;
1573                 break;
1574         case SEEK_END:
1575                 offset = ((dd->kregend - dd->kregbase) + DC8051_DATA_MEM_SIZE) -
1576                         offset;
1577                 break;
1578         default:
1579                 return -EINVAL;
1580         }
1581
1582         if (offset < 0)
1583                 return -EINVAL;
1584
1585         if (offset >= (dd->kregend - dd->kregbase) + DC8051_DATA_MEM_SIZE)
1586                 return -EINVAL;
1587
1588         filp->f_pos = offset;
1589
1590         return filp->f_pos;
1591 }
1592
1593 /* NOTE: assumes unsigned long is 8 bytes */
1594 static ssize_t ui_read(struct file *filp, char __user *buf, size_t count,
1595                        loff_t *f_pos)
1596 {
1597         struct hfi1_devdata *dd = filp->private_data;
1598         void __iomem *base = dd->kregbase;
1599         unsigned long total, csr_off,
1600                 barlen = (dd->kregend - dd->kregbase);
1601         u64 data;
1602
1603         /* only read 8 byte quantities */
1604         if ((count % 8) != 0)
1605                 return -EINVAL;
1606         /* offset must be 8-byte aligned */
1607         if ((*f_pos % 8) != 0)
1608                 return -EINVAL;
1609         /* destination buffer must be 8-byte aligned */
1610         if ((unsigned long)buf % 8 != 0)
1611                 return -EINVAL;
1612         /* must be in range */
1613         if (*f_pos + count > (barlen + DC8051_DATA_MEM_SIZE))
1614                 return -EINVAL;
1615         /* only set the base if we are not starting past the BAR */
1616         if (*f_pos < barlen)
1617                 base += *f_pos;
1618         csr_off = *f_pos;
1619         for (total = 0; total < count; total += 8, csr_off += 8) {
1620                 /* accessing LCB CSRs requires more checks */
1621                 if (is_lcb_offset(csr_off)) {
1622                         if (read_lcb_csr(dd, csr_off, (u64 *)&data))
1623                                 break; /* failed */
1624                 }
1625                 /*
1626                  * Cannot read ASIC GPIO/QSFP* clear and force CSRs without a
1627                  * false parity error.  Avoid the whole issue by not reading
1628                  * them.  These registers are defined as having a read value
1629                  * of 0.
1630                  */
1631                 else if (csr_off == ASIC_GPIO_CLEAR ||
1632                          csr_off == ASIC_GPIO_FORCE ||
1633                          csr_off == ASIC_QSFP1_CLEAR ||
1634                          csr_off == ASIC_QSFP1_FORCE ||
1635                          csr_off == ASIC_QSFP2_CLEAR ||
1636                          csr_off == ASIC_QSFP2_FORCE)
1637                         data = 0;
1638                 else if (csr_off >= barlen) {
1639                         /*
1640                          * read_8051_data can read more than just 8 bytes at
1641                          * a time. However, folding this into the loop and
1642                          * handling the reads in 8 byte increments allows us
1643                          * to smoothly transition from chip memory to 8051
1644                          * memory.
1645                          */
1646                         if (read_8051_data(dd,
1647                                            (u32)(csr_off - barlen),
1648                                            sizeof(data), &data))
1649                                 break; /* failed */
1650                 } else
1651                         data = readq(base + total);
1652                 if (put_user(data, (unsigned long __user *)(buf + total)))
1653                         break;
1654         }
1655         *f_pos += total;
1656         return total;
1657 }
1658
1659 /* NOTE: assumes unsigned long is 8 bytes */
1660 static ssize_t ui_write(struct file *filp, const char __user *buf,
1661                         size_t count, loff_t *f_pos)
1662 {
1663         struct hfi1_devdata *dd = filp->private_data;
1664         void __iomem *base;
1665         unsigned long total, data, csr_off;
1666         int in_lcb;
1667
1668         /* only write 8 byte quantities */
1669         if ((count % 8) != 0)
1670                 return -EINVAL;
1671         /* offset must be 8-byte aligned */
1672         if ((*f_pos % 8) != 0)
1673                 return -EINVAL;
1674         /* source buffer must be 8-byte aligned */
1675         if ((unsigned long)buf % 8 != 0)
1676                 return -EINVAL;
1677         /* must be in range */
1678         if (*f_pos + count > dd->kregend - dd->kregbase)
1679                 return -EINVAL;
1680
1681         base = (void __iomem *)dd->kregbase + *f_pos;
1682         csr_off = *f_pos;
1683         in_lcb = 0;
1684         for (total = 0; total < count; total += 8, csr_off += 8) {
1685                 if (get_user(data, (unsigned long __user *)(buf + total)))
1686                         break;
1687                 /* accessing LCB CSRs requires a special procedure */
1688                 if (is_lcb_offset(csr_off)) {
1689                         if (!in_lcb) {
1690                                 int ret = acquire_lcb_access(dd, 1);
1691
1692                                 if (ret)
1693                                         break;
1694                                 in_lcb = 1;
1695                         }
1696                 } else {
1697                         if (in_lcb) {
1698                                 release_lcb_access(dd, 1);
1699                                 in_lcb = 0;
1700                         }
1701                 }
1702                 writeq(data, base + total);
1703         }
1704         if (in_lcb)
1705                 release_lcb_access(dd, 1);
1706         *f_pos += total;
1707         return total;
1708 }
1709
1710 static const struct file_operations ui_file_ops = {
1711         .owner = THIS_MODULE,
1712         .llseek = ui_lseek,
1713         .read = ui_read,
1714         .write = ui_write,
1715         .open = ui_open,
1716         .release = ui_release,
1717 };
1718
1719 #define UI_OFFSET 192   /* device minor offset for UI devices */
1720 static int create_ui = 1;
1721
1722 static struct cdev wildcard_cdev;
1723 static struct device *wildcard_device;
1724
1725 static atomic_t user_count = ATOMIC_INIT(0);
1726
1727 static void user_remove(struct hfi1_devdata *dd)
1728 {
1729         if (atomic_dec_return(&user_count) == 0)
1730                 hfi1_cdev_cleanup(&wildcard_cdev, &wildcard_device);
1731
1732         hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1733         hfi1_cdev_cleanup(&dd->ui_cdev, &dd->ui_device);
1734 }
1735
1736 static int user_add(struct hfi1_devdata *dd)
1737 {
1738         char name[10];
1739         int ret;
1740
1741         if (atomic_inc_return(&user_count) == 1) {
1742                 ret = hfi1_cdev_init(0, class_name(), &hfi1_file_ops,
1743                                      &wildcard_cdev, &wildcard_device,
1744                                      true);
1745                 if (ret)
1746                         goto done;
1747         }
1748
1749         snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1750         ret = hfi1_cdev_init(dd->unit + 1, name, &hfi1_file_ops,
1751                              &dd->user_cdev, &dd->user_device,
1752                              true);
1753         if (ret)
1754                 goto done;
1755
1756         if (create_ui) {
1757                 snprintf(name, sizeof(name),
1758                          "%s_ui%d", class_name(), dd->unit);
1759                 ret = hfi1_cdev_init(dd->unit + UI_OFFSET, name, &ui_file_ops,
1760                                      &dd->ui_cdev, &dd->ui_device,
1761                                      false);
1762                 if (ret)
1763                         goto done;
1764         }
1765
1766         return 0;
1767 done:
1768         user_remove(dd);
1769         return ret;
1770 }
1771
1772 /*
1773  * Create per-unit files in /dev
1774  */
1775 int hfi1_device_create(struct hfi1_devdata *dd)
1776 {
1777         int r, ret;
1778
1779         r = user_add(dd);
1780         ret = hfi1_diag_add(dd);
1781         if (r && !ret)
1782                 ret = r;
1783         return ret;
1784 }
1785
1786 /*
1787  * Remove per-unit files in /dev
1788  * void, core kernel returns no errors for this stuff
1789  */
1790 void hfi1_device_remove(struct hfi1_devdata *dd)
1791 {
1792         user_remove(dd);
1793         hfi1_diag_remove(dd);
1794 }