drivers/infiniband/hw/hfi1/file_ops.c

   1 /*
   2  * Copyright(c) 2015-2017 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47 #include <linux/poll.h>
  48 #include <linux/cdev.h>
  49 #include <linux/vmalloc.h>
  50 #include <linux/io.h>
  51 #include <linux/sched/mm.h>
  52 #include <linux/bitmap.h>
  53
  54 #include <rdma/ib.h>
  55
  56 #include "hfi.h"
  57 #include "pio.h"
  58 #include "device.h"
  59 #include "common.h"
  60 #include "trace.h"
  61 #include "mmu_rb.h"
  62 #include "user_sdma.h"
  63 #include "user_exp_rcv.h"
  64 #include "aspm.h"
  65
  66 #undef pr_fmt
  67 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
  68
  69 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
  70
  71 /*
  72  * File operation functions
  73  */
  74 static int hfi1_file_open(struct inode *inode, struct file *fp);
  75 static int hfi1_file_close(struct inode *inode, struct file *fp);
  76 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from);
  77 static __poll_t hfi1_poll(struct file *fp, struct poll_table_struct *pt);
  78 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma);
  79
  80 static u64 kvirt_to_phys(void *addr);
  81 static int assign_ctxt(struct hfi1_filedata *fd, unsigned long arg, u32 len);
  82 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
  83                           const struct hfi1_user_info *uinfo);
  84 static int init_user_ctxt(struct hfi1_filedata *fd,
  85                           struct hfi1_ctxtdata *uctxt);
  86 static void user_init(struct hfi1_ctxtdata *uctxt);
  87 static int get_ctxt_info(struct hfi1_filedata *fd, unsigned long arg, u32 len);
  88 static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len);
  89 static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
  90                               u32 len);
  91 static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
  92                               u32 len);
  93 static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
  94                                 u32 len);
  95 static int setup_base_ctxt(struct hfi1_filedata *fd,
  96                            struct hfi1_ctxtdata *uctxt);
  97 static int setup_subctxt(struct hfi1_ctxtdata *uctxt);
  98
  99 static int find_sub_ctxt(struct hfi1_filedata *fd,
 100                          const struct hfi1_user_info *uinfo);
 101 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
 102                          struct hfi1_user_info *uinfo,
 103                          struct hfi1_ctxtdata **cd);
 104 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt);
 105 static __poll_t poll_urgent(struct file *fp, struct poll_table_struct *pt);
 106 static __poll_t poll_next(struct file *fp, struct poll_table_struct *pt);
 107 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
 108                           unsigned long arg);
 109 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg);
 110 static int ctxt_reset(struct hfi1_ctxtdata *uctxt);
 111 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
 112                        unsigned long arg);
 113 static vm_fault_t vma_fault(struct vm_fault *vmf);
 114 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
 115                             unsigned long arg);
 116
 117 static const struct file_operations hfi1_file_ops = {
 118         .owner = THIS_MODULE,
 119         .write_iter = hfi1_write_iter,
 120         .open = hfi1_file_open,
 121         .release = hfi1_file_close,
 122         .unlocked_ioctl = hfi1_file_ioctl,
 123         .poll = hfi1_poll,
 124         .mmap = hfi1_file_mmap,
 125         .llseek = noop_llseek,
 126 };
 127
 128 static const struct vm_operations_struct vm_ops = {
 129         .fault = vma_fault,
 130 };
 131
 132 /*
 133  * Types of memories mapped into user processes' space
 134  */
 135 enum mmap_types {
 136         PIO_BUFS = 1,
 137         PIO_BUFS_SOP,
 138         PIO_CRED,
 139         RCV_HDRQ,
 140         RCV_EGRBUF,
 141         UREGS,
 142         EVENTS,
 143         STATUS,
 144         RTAIL,
 145         SUBCTXT_UREGS,
 146         SUBCTXT_RCV_HDRQ,
 147         SUBCTXT_EGRBUF,
 148         SDMA_COMP
 149 };
 150
 151 /*
 152  * Masks and offsets defining the mmap tokens
 153  */
 154 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
 155 #define HFI1_MMAP_OFFSET_SHIFT  0
 156 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
 157 #define HFI1_MMAP_SUBCTXT_SHIFT 12
 158 #define HFI1_MMAP_CTXT_MASK     0xffULL
 159 #define HFI1_MMAP_CTXT_SHIFT    16
 160 #define HFI1_MMAP_TYPE_MASK     0xfULL
 161 #define HFI1_MMAP_TYPE_SHIFT    24
 162 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
 163 #define HFI1_MMAP_MAGIC_SHIFT   32
 164
 165 #define HFI1_MMAP_MAGIC         0xdabbad00
 166
 167 #define HFI1_MMAP_TOKEN_SET(field, val) \
 168         (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
 169 #define HFI1_MMAP_TOKEN_GET(field, token) \
 170         (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
 171 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
 172         (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
 173         HFI1_MMAP_TOKEN_SET(TYPE, type) | \
 174         HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
 175         HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
 176         HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
 177
 178 #define dbg(fmt, ...)                           \
 179         pr_info(fmt, ##__VA_ARGS__)
 180
 181 static inline int is_valid_mmap(u64 token)
 182 {
 183         return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
 184 }
 185
 186 static int hfi1_file_open(struct inode *inode, struct file *fp)
 187 {
 188         struct hfi1_filedata *fd;
 189         struct hfi1_devdata *dd = container_of(inode->i_cdev,
 190                                                struct hfi1_devdata,
 191                                                user_cdev);
 192
 193         if (!((dd->flags & HFI1_PRESENT) && dd->kregbase1))
 194                 return -EINVAL;
 195
 196         if (!atomic_inc_not_zero(&dd->user_refcount))
 197                 return -ENXIO;
 198
 199         /* The real work is performed later in assign_ctxt() */
 200
 201         fd = kzalloc(sizeof(*fd), GFP_KERNEL);
 202
 203         if (!fd || init_srcu_struct(&fd->pq_srcu))
 204                 goto nomem;
 205         spin_lock_init(&fd->pq_rcu_lock);
 206         spin_lock_init(&fd->tid_lock);
 207         spin_lock_init(&fd->invalid_lock);
 208         fd->rec_cpu_num = -1; /* no cpu affinity by default */
 209         fd->mm = current->mm;
 210         mmgrab(fd->mm);
 211         fd->dd = dd;
 212         kobject_get(&fd->dd->kobj);
 213         fp->private_data = fd;
 214         return 0;
 215 nomem:
 216         kfree(fd);
 217         fp->private_data = NULL;
 218         if (atomic_dec_and_test(&dd->user_refcount))
 219                 complete(&dd->user_comp);
 220         return -ENOMEM;
 221 }
 222
 223 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
 224                             unsigned long arg)
 225 {
 226         struct hfi1_filedata *fd = fp->private_data;
 227         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 228         int ret = 0;
 229         int uval = 0;
 230
 231         hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
 232         if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
 233             cmd != HFI1_IOCTL_GET_VERS &&
 234             !uctxt)
 235                 return -EINVAL;
 236
 237         switch (cmd) {
 238         case HFI1_IOCTL_ASSIGN_CTXT:
 239                 ret = assign_ctxt(fd, arg, _IOC_SIZE(cmd));
 240                 break;
 241
 242         case HFI1_IOCTL_CTXT_INFO:
 243                 ret = get_ctxt_info(fd, arg, _IOC_SIZE(cmd));
 244                 break;
 245
 246         case HFI1_IOCTL_USER_INFO:
 247                 ret = get_base_info(fd, arg, _IOC_SIZE(cmd));
 248                 break;
 249
 250         case HFI1_IOCTL_CREDIT_UPD:
 251                 if (uctxt)
 252                         sc_return_credits(uctxt->sc);
 253                 break;
 254
 255         case HFI1_IOCTL_TID_UPDATE:
 256                 ret = user_exp_rcv_setup(fd, arg, _IOC_SIZE(cmd));
 257                 break;
 258
 259         case HFI1_IOCTL_TID_FREE:
 260                 ret = user_exp_rcv_clear(fd, arg, _IOC_SIZE(cmd));
 261                 break;
 262
 263         case HFI1_IOCTL_TID_INVAL_READ:
 264                 ret = user_exp_rcv_invalid(fd, arg, _IOC_SIZE(cmd));
 265                 break;
 266
 267         case HFI1_IOCTL_RECV_CTRL:
 268                 ret = manage_rcvq(uctxt, fd->subctxt, arg);
 269                 break;
 270
 271         case HFI1_IOCTL_POLL_TYPE:
 272                 if (get_user(uval, (int __user *)arg))
 273                         return -EFAULT;
 274                 uctxt->poll_type = (typeof(uctxt->poll_type))uval;
 275                 break;
 276
 277         case HFI1_IOCTL_ACK_EVENT:
 278                 ret = user_event_ack(uctxt, fd->subctxt, arg);
 279                 break;
 280
 281         case HFI1_IOCTL_SET_PKEY:
 282                 ret = set_ctxt_pkey(uctxt, arg);
 283                 break;
 284
 285         case HFI1_IOCTL_CTXT_RESET:
 286                 ret = ctxt_reset(uctxt);
 287                 break;
 288
 289         case HFI1_IOCTL_GET_VERS:
 290                 uval = HFI1_USER_SWVERSION;
 291                 if (put_user(uval, (int __user *)arg))
 292                         return -EFAULT;
 293                 break;
 294
 295         default:
 296                 return -EINVAL;
 297         }
 298
 299         return ret;
 300 }
 301
 302 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
 303 {
 304         struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
 305         struct hfi1_user_sdma_pkt_q *pq;
 306         struct hfi1_user_sdma_comp_q *cq = fd->cq;
 307         int done = 0, reqs = 0;
 308         unsigned long dim = from->nr_segs;
 309         int idx;
 310
 311         if (!HFI1_CAP_IS_KSET(SDMA))
 312                 return -EINVAL;
 313         idx = srcu_read_lock(&fd->pq_srcu);
 314         pq = srcu_dereference(fd->pq, &fd->pq_srcu);
 315         if (!cq || !pq) {
 316                 srcu_read_unlock(&fd->pq_srcu, idx);
 317                 return -EIO;
 318         }
 319
 320         if (!iter_is_iovec(from) || !dim) {
 321                 srcu_read_unlock(&fd->pq_srcu, idx);
 322                 return -EINVAL;
 323         }
 324
 325         trace_hfi1_sdma_request(fd->dd, fd->uctxt->ctxt, fd->subctxt, dim);
 326
 327         if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
 328                 srcu_read_unlock(&fd->pq_srcu, idx);
 329                 return -ENOSPC;
 330         }
 331
 332         while (dim) {
 333                 int ret;
 334                 unsigned long count = 0;
 335
 336                 ret = hfi1_user_sdma_process_request(
 337                         fd, (struct iovec *)(from->iov + done),
 338                         dim, &count);
 339                 if (ret) {
 340                         reqs = ret;
 341                         break;
 342                 }
 343                 dim -= count;
 344                 done += count;
 345                 reqs++;
 346         }
 347
 348         srcu_read_unlock(&fd->pq_srcu, idx);
 349         return reqs;
 350 }
 351
 352 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
 353 {
 354         struct hfi1_filedata *fd = fp->private_data;
 355         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 356         struct hfi1_devdata *dd;
 357         unsigned long flags;
 358         u64 token = vma->vm_pgoff << PAGE_SHIFT,
 359                 memaddr = 0;
 360         void *memvirt = NULL;
 361         u8 subctxt, mapio = 0, vmf = 0, type;
 362         ssize_t memlen = 0;
 363         int ret = 0;
 364         u16 ctxt;
 365
 366         if (!is_valid_mmap(token) || !uctxt ||
 367             !(vma->vm_flags & VM_SHARED)) {
 368                 ret = -EINVAL;
 369                 goto done;
 370         }
 371         dd = uctxt->dd;
 372         ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
 373         subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
 374         type = HFI1_MMAP_TOKEN_GET(TYPE, token);
 375         if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
 376                 ret = -EINVAL;
 377                 goto done;
 378         }
 379
 380         flags = vma->vm_flags;
 381
 382         switch (type) {
 383         case PIO_BUFS:
 384         case PIO_BUFS_SOP:
 385                 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
 386                                 /* chip pio base */
 387                            (uctxt->sc->hw_context * BIT(16))) +
 388                                 /* 64K PIO space / ctxt */
 389                         (type == PIO_BUFS_SOP ?
 390                                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
 391                 /*
 392                  * Map only the amount allocated to the context, not the
 393                  * entire available context's PIO space.
 394                  */
 395                 memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
 396                 flags &= ~VM_MAYREAD;
 397                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 398                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 399                 mapio = 1;
 400                 break;
 401         case PIO_CRED:
 402                 if (flags & VM_WRITE) {
 403                         ret = -EPERM;
 404                         goto done;
 405                 }
 406                 /*
 407                  * The credit return location for this context could be on the
 408                  * second or third page allocated for credit returns (if number
 409                  * of enabled contexts > 64 and 128 respectively).
 410                  */
 411                 memvirt = dd->cr_base[uctxt->numa_id].va;
 412                 memaddr = virt_to_phys(memvirt) +
 413                         (((u64)uctxt->sc->hw_free -
 414                           (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
 415                 memlen = PAGE_SIZE;
 416                 flags &= ~VM_MAYWRITE;
 417                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 418                 /*
 419                  * The driver has already allocated memory for credit
 420                  * returns and programmed it into the chip. Has that
 421                  * memory been flagged as non-cached?
 422                  */
 423                 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
 424                 mapio = 1;
 425                 break;
 426         case RCV_HDRQ:
 427                 memlen = rcvhdrq_size(uctxt);
 428                 memvirt = uctxt->rcvhdrq;
 429                 break;
 430         case RCV_EGRBUF: {
 431                 unsigned long addr;
 432                 int i;
 433                 /*
 434                  * The RcvEgr buffer need to be handled differently
 435                  * as multiple non-contiguous pages need to be mapped
 436                  * into the user process.
 437                  */
 438                 memlen = uctxt->egrbufs.size;
 439                 if ((vma->vm_end - vma->vm_start) != memlen) {
 440                         dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
 441                                    (vma->vm_end - vma->vm_start), memlen);
 442                         ret = -EINVAL;
 443                         goto done;
 444                 }
 445                 if (vma->vm_flags & VM_WRITE) {
 446                         ret = -EPERM;
 447                         goto done;
 448                 }
 449                 vma->vm_flags &= ~VM_MAYWRITE;
 450                 addr = vma->vm_start;
 451                 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
 452                         memlen = uctxt->egrbufs.buffers[i].len;
 453                         memvirt = uctxt->egrbufs.buffers[i].addr;
 454                         ret = remap_pfn_range(
 455                                 vma, addr,
 456                                 /*
 457                                  * virt_to_pfn() does the same, but
 458                                  * it's not available on x86_64
 459                                  * when CONFIG_MMU is enabled.
 460                                  */
 461                                 PFN_DOWN(__pa(memvirt)),
 462                                 memlen,
 463                                 vma->vm_page_prot);
 464                         if (ret < 0)
 465                                 goto done;
 466                         addr += memlen;
 467                 }
 468                 ret = 0;
 469                 goto done;
 470         }
 471         case UREGS:
 472                 /*
 473                  * Map only the page that contains this context's user
 474                  * registers.
 475                  */
 476                 memaddr = (unsigned long)
 477                         (dd->physaddr + RXE_PER_CONTEXT_USER)
 478                         + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
 479                 /*
 480                  * TidFlow table is on the same page as the rest of the
 481                  * user registers.
 482                  */
 483                 memlen = PAGE_SIZE;
 484                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
 485                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 486                 mapio = 1;
 487                 break;
 488         case EVENTS:
 489                 /*
 490                  * Use the page where this context's flags are. User level
 491                  * knows where it's own bitmap is within the page.
 492                  */
 493                 memaddr = (unsigned long)
 494                         (dd->events + uctxt_offset(uctxt)) & PAGE_MASK;
 495                 memlen = PAGE_SIZE;
 496                 /*
 497                  * v3.7 removes VM_RESERVED but the effect is kept by
 498                  * using VM_IO.
 499                  */
 500                 flags |= VM_IO | VM_DONTEXPAND;
 501                 vmf = 1;
 502                 break;
 503         case STATUS:
 504                 if (flags & VM_WRITE) {
 505                         ret = -EPERM;
 506                         goto done;
 507                 }
 508                 memaddr = kvirt_to_phys((void *)dd->status);
 509                 memlen = PAGE_SIZE;
 510                 flags |= VM_IO | VM_DONTEXPAND;
 511                 break;
 512         case RTAIL:
 513                 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
 514                         /*
 515                          * If the memory allocation failed, the context alloc
 516                          * also would have failed, so we would never get here
 517                          */
 518                         ret = -EINVAL;
 519                         goto done;
 520                 }
 521                 if ((flags & VM_WRITE) || !uctxt->rcvhdrtail_kvaddr) {
 522                         ret = -EPERM;
 523                         goto done;
 524                 }
 525                 memlen = PAGE_SIZE;
 526                 memvirt = (void *)uctxt->rcvhdrtail_kvaddr;
 527                 flags &= ~VM_MAYWRITE;
 528                 break;
 529         case SUBCTXT_UREGS:
 530                 memaddr = (u64)uctxt->subctxt_uregbase;
 531                 memlen = PAGE_SIZE;
 532                 flags |= VM_IO | VM_DONTEXPAND;
 533                 vmf = 1;
 534                 break;
 535         case SUBCTXT_RCV_HDRQ:
 536                 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
 537                 memlen = rcvhdrq_size(uctxt) * uctxt->subctxt_cnt;
 538                 flags |= VM_IO | VM_DONTEXPAND;
 539                 vmf = 1;
 540                 break;
 541         case SUBCTXT_EGRBUF:
 542                 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
 543                 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
 544                 flags |= VM_IO | VM_DONTEXPAND;
 545                 flags &= ~VM_MAYWRITE;
 546                 vmf = 1;
 547                 break;
 548         case SDMA_COMP: {
 549                 struct hfi1_user_sdma_comp_q *cq = fd->cq;
 550
 551                 if (!cq) {
 552                         ret = -EFAULT;
 553                         goto done;
 554                 }
 555                 memaddr = (u64)cq->comps;
 556                 memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
 557                 flags |= VM_IO | VM_DONTEXPAND;
 558                 vmf = 1;
 559                 break;
 560         }
 561         default:
 562                 ret = -EINVAL;
 563                 break;
 564         }
 565
 566         if ((vma->vm_end - vma->vm_start) != memlen) {
 567                 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
 568                           uctxt->ctxt, fd->subctxt,
 569                           (vma->vm_end - vma->vm_start), memlen);
 570                 ret = -EINVAL;
 571                 goto done;
 572         }
 573
 574         vma->vm_flags = flags;
 575         hfi1_cdbg(PROC,
 576                   "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
 577                     ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
 578                     vma->vm_end - vma->vm_start, vma->vm_flags);
 579         if (vmf) {
 580                 vma->vm_pgoff = PFN_DOWN(memaddr);
 581                 vma->vm_ops = &vm_ops;
 582                 ret = 0;
 583         } else if (mapio) {
 584                 ret = io_remap_pfn_range(vma, vma->vm_start,
 585                                          PFN_DOWN(memaddr),
 586                                          memlen,
 587                                          vma->vm_page_prot);
 588         } else if (memvirt) {
 589                 ret = remap_pfn_range(vma, vma->vm_start,
 590                                       PFN_DOWN(__pa(memvirt)),
 591                                       memlen,
 592                                       vma->vm_page_prot);
 593         } else {
 594                 ret = remap_pfn_range(vma, vma->vm_start,
 595                                       PFN_DOWN(memaddr),
 596                                       memlen,
 597                                       vma->vm_page_prot);
 598         }
 599 done:
 600         return ret;
 601 }
 602
 603 /*
 604  * Local (non-chip) user memory is not mapped right away but as it is
 605  * accessed by the user-level code.
 606  */
 607 static vm_fault_t vma_fault(struct vm_fault *vmf)
 608 {
 609         struct page *page;
 610
 611         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 612         if (!page)
 613                 return VM_FAULT_SIGBUS;
 614
 615         get_page(page);
 616         vmf->page = page;
 617
 618         return 0;
 619 }
 620
 621 static __poll_t hfi1_poll(struct file *fp, struct poll_table_struct *pt)
 622 {
 623         struct hfi1_ctxtdata *uctxt;
 624         __poll_t pollflag;
 625
 626         uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
 627         if (!uctxt)
 628                 pollflag = EPOLLERR;
 629         else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
 630                 pollflag = poll_urgent(fp, pt);
 631         else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
 632                 pollflag = poll_next(fp, pt);
 633         else /* invalid */
 634                 pollflag = EPOLLERR;
 635
 636         return pollflag;
 637 }
 638
 639 static int hfi1_file_close(struct inode *inode, struct file *fp)
 640 {
 641         struct hfi1_filedata *fdata = fp->private_data;
 642         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 643         struct hfi1_devdata *dd = container_of(inode->i_cdev,
 644                                                struct hfi1_devdata,
 645                                                user_cdev);
 646         unsigned long flags, *ev;
 647
 648         fp->private_data = NULL;
 649
 650         if (!uctxt)
 651                 goto done;
 652
 653         hfi1_cdbg(PROC, "closing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
 654
 655         flush_wc();
 656         /* drain user sdma queue */
 657         hfi1_user_sdma_free_queues(fdata, uctxt);
 658
 659         /* release the cpu */
 660         hfi1_put_proc_affinity(fdata->rec_cpu_num);
 661
 662         /* clean up rcv side */
 663         hfi1_user_exp_rcv_free(fdata);
 664
 665         /*
 666          * fdata->uctxt is used in the above cleanup.  It is not ready to be
 667          * removed until here.
 668          */
 669         fdata->uctxt = NULL;
 670         hfi1_rcd_put(uctxt);
 671
 672         /*
 673          * Clear any left over, unhandled events so the next process that
 674          * gets this context doesn't get confused.
 675          */
 676         ev = dd->events + uctxt_offset(uctxt) + fdata->subctxt;
 677         *ev = 0;
 678
 679         spin_lock_irqsave(&dd->uctxt_lock, flags);
 680         __clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
 681         if (!bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
 682                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 683                 goto done;
 684         }
 685         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 686
 687         /*
 688          * Disable receive context and interrupt available, reset all
 689          * RcvCtxtCtrl bits to default values.
 690          */
 691         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
 692                      HFI1_RCVCTRL_TIDFLOW_DIS |
 693                      HFI1_RCVCTRL_INTRAVAIL_DIS |
 694                      HFI1_RCVCTRL_TAILUPD_DIS |
 695                      HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
 696                      HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
 697                      HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt);
 698         /* Clear the context's J_KEY */
 699         hfi1_clear_ctxt_jkey(dd, uctxt);
 700         /*
 701          * If a send context is allocated, reset context integrity
 702          * checks to default and disable the send context.
 703          */
 704         if (uctxt->sc) {
 705                 sc_disable(uctxt->sc);
 706                 set_pio_integrity(uctxt->sc);
 707         }
 708
 709         hfi1_free_ctxt_rcv_groups(uctxt);
 710         hfi1_clear_ctxt_pkey(dd, uctxt);
 711
 712         uctxt->event_flags = 0;
 713
 714         deallocate_ctxt(uctxt);
 715 done:
 716         mmdrop(fdata->mm);
 717         kobject_put(&dd->kobj);
 718
 719         if (atomic_dec_and_test(&dd->user_refcount))
 720                 complete(&dd->user_comp);
 721
 722         cleanup_srcu_struct(&fdata->pq_srcu);
 723         kfree(fdata);
 724         return 0;
 725 }
 726
 727 /*
 728  * Convert kernel *virtual* addresses to physical addresses.
 729  * This is used to vmalloc'ed addresses.
 730  */
 731 static u64 kvirt_to_phys(void *addr)
 732 {
 733         struct page *page;
 734         u64 paddr = 0;
 735
 736         page = vmalloc_to_page(addr);
 737         if (page)
 738                 paddr = page_to_pfn(page) << PAGE_SHIFT;
 739
 740         return paddr;
 741 }
 742
 743 /**
 744  * complete_subctxt
 745  * @fd: valid filedata pointer
 746  *
 747  * Sub-context info can only be set up after the base context
 748  * has been completed.  This is indicated by the clearing of the
 749  * HFI1_CTXT_BASE_UINIT bit.
 750  *
 751  * Wait for the bit to be cleared, and then complete the subcontext
 752  * initialization.
 753  *
 754  */
 755 static int complete_subctxt(struct hfi1_filedata *fd)
 756 {
 757         int ret;
 758         unsigned long flags;
 759
 760         /*
 761          * sub-context info can only be set up after the base context
 762          * has been completed.
 763          */
 764         ret = wait_event_interruptible(
 765                 fd->uctxt->wait,
 766                 !test_bit(HFI1_CTXT_BASE_UNINIT, &fd->uctxt->event_flags));
 767
 768         if (test_bit(HFI1_CTXT_BASE_FAILED, &fd->uctxt->event_flags))
 769                 ret = -ENOMEM;
 770
 771         /* Finish the sub-context init */
 772         if (!ret) {
 773                 fd->rec_cpu_num = hfi1_get_proc_affinity(fd->uctxt->numa_id);
 774                 ret = init_user_ctxt(fd, fd->uctxt);
 775         }
 776
 777         if (ret) {
 778                 spin_lock_irqsave(&fd->dd->uctxt_lock, flags);
 779                 __clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
 780                 spin_unlock_irqrestore(&fd->dd->uctxt_lock, flags);
 781                 hfi1_rcd_put(fd->uctxt);
 782                 fd->uctxt = NULL;
 783         }
 784
 785         return ret;
 786 }
 787
 788 static int assign_ctxt(struct hfi1_filedata *fd, unsigned long arg, u32 len)
 789 {
 790         int ret;
 791         unsigned int swmajor;
 792         struct hfi1_ctxtdata *uctxt = NULL;
 793         struct hfi1_user_info uinfo;
 794
 795         if (fd->uctxt)
 796                 return -EINVAL;
 797
 798         if (sizeof(uinfo) != len)
 799                 return -EINVAL;
 800
 801         if (copy_from_user(&uinfo, (void __user *)arg, sizeof(uinfo)))
 802                 return -EFAULT;
 803
 804         swmajor = uinfo.userversion >> 16;
 805         if (swmajor != HFI1_USER_SWMAJOR)
 806                 return -ENODEV;
 807
 808         if (uinfo.subctxt_cnt > HFI1_MAX_SHARED_CTXTS)
 809                 return -EINVAL;
 810
 811         /*
 812          * Acquire the mutex to protect against multiple creations of what
 813          * could be a shared base context.
 814          */
 815         mutex_lock(&hfi1_mutex);
 816         /*
 817          * Get a sub context if available  (fd->uctxt will be set).
 818          * ret < 0 error, 0 no context, 1 sub-context found
 819          */
 820         ret = find_sub_ctxt(fd, &uinfo);
 821
 822         /*
 823          * Allocate a base context if context sharing is not required or a
 824          * sub context wasn't found.
 825          */
 826         if (!ret)
 827                 ret = allocate_ctxt(fd, fd->dd, &uinfo, &uctxt);
 828
 829         mutex_unlock(&hfi1_mutex);
 830
 831         /* Depending on the context type, finish the appropriate init */
 832         switch (ret) {
 833         case 0:
 834                 ret = setup_base_ctxt(fd, uctxt);
 835                 if (ret)
 836                         deallocate_ctxt(uctxt);
 837                 break;
 838         case 1:
 839                 ret = complete_subctxt(fd);
 840                 break;
 841         default:
 842                 break;
 843         }
 844
 845         return ret;
 846 }
 847
 848 /**
 849  * match_ctxt
 850  * @fd: valid filedata pointer
 851  * @uinfo: user info to compare base context with
 852  * @uctxt: context to compare uinfo to.
 853  *
 854  * Compare the given context with the given information to see if it
 855  * can be used for a sub context.
 856  */
 857 static int match_ctxt(struct hfi1_filedata *fd,
 858                       const struct hfi1_user_info *uinfo,
 859                       struct hfi1_ctxtdata *uctxt)
 860 {
 861         struct hfi1_devdata *dd = fd->dd;
 862         unsigned long flags;
 863         u16 subctxt;
 864
 865         /* Skip dynamically allocated kernel contexts */
 866         if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
 867                 return 0;
 868
 869         /* Skip ctxt if it doesn't match the requested one */
 870         if (memcmp(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid)) ||
 871             uctxt->jkey != generate_jkey(current_uid()) ||
 872             uctxt->subctxt_id != uinfo->subctxt_id ||
 873             uctxt->subctxt_cnt != uinfo->subctxt_cnt)
 874                 return 0;
 875
 876         /* Verify the sharing process matches the base */
 877         if (uctxt->userversion != uinfo->userversion)
 878                 return -EINVAL;
 879
 880         /* Find an unused sub context */
 881         spin_lock_irqsave(&dd->uctxt_lock, flags);
 882         if (bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
 883                 /* context is being closed, do not use */
 884                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 885                 return 0;
 886         }
 887
 888         subctxt = find_first_zero_bit(uctxt->in_use_ctxts,
 889                                       HFI1_MAX_SHARED_CTXTS);
 890         if (subctxt >= uctxt->subctxt_cnt) {
 891                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 892                 return -EBUSY;
 893         }
 894
 895         fd->subctxt = subctxt;
 896         __set_bit(fd->subctxt, uctxt->in_use_ctxts);
 897         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 898
 899         fd->uctxt = uctxt;
 900         hfi1_rcd_get(uctxt);
 901
 902         return 1;
 903 }
 904
 905 /**
 906  * find_sub_ctxt
 907  * @fd: valid filedata pointer
 908  * @uinfo: matching info to use to find a possible context to share.
 909  *
 910  * The hfi1_mutex must be held when this function is called.  It is
 911  * necessary to ensure serialized creation of shared contexts.
 912  *
 913  * Return:
 914  *    0      No sub-context found
 915  *    1      Subcontext found and allocated
 916  *    errno  EINVAL (incorrect parameters)
 917  *           EBUSY (all sub contexts in use)
 918  */
 919 static int find_sub_ctxt(struct hfi1_filedata *fd,
 920                          const struct hfi1_user_info *uinfo)
 921 {
 922         struct hfi1_ctxtdata *uctxt;
 923         struct hfi1_devdata *dd = fd->dd;
 924         u16 i;
 925         int ret;
 926
 927         if (!uinfo->subctxt_cnt)
 928                 return 0;
 929
 930         for (i = dd->first_dyn_alloc_ctxt; i < dd->num_rcv_contexts; i++) {
 931                 uctxt = hfi1_rcd_get_by_index(dd, i);
 932                 if (uctxt) {
 933                         ret = match_ctxt(fd, uinfo, uctxt);
 934                         hfi1_rcd_put(uctxt);
 935                         /* value of != 0 will return */
 936                         if (ret)
 937                                 return ret;
 938                 }
 939         }
 940
 941         return 0;
 942 }
 943
 944 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
 945                          struct hfi1_user_info *uinfo,
 946                          struct hfi1_ctxtdata **rcd)
 947 {
 948         struct hfi1_ctxtdata *uctxt;
 949         int ret, numa;
 950
 951         if (dd->flags & HFI1_FROZEN) {
 952                 /*
 953                  * Pick an error that is unique from all other errors
 954                  * that are returned so the user process knows that
 955                  * it tried to allocate while the SPC was frozen.  It
 956                  * it should be able to retry with success in a short
 957                  * while.
 958                  */
 959                 return -EIO;
 960         }
 961
 962         if (!dd->freectxts)
 963                 return -EBUSY;
 964
 965         /*
 966          * If we don't have a NUMA node requested, preference is towards
 967          * device NUMA node.
 968          */
 969         fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
 970         if (fd->rec_cpu_num != -1)
 971                 numa = cpu_to_node(fd->rec_cpu_num);
 972         else
 973                 numa = numa_node_id();
 974         ret = hfi1_create_ctxtdata(dd->pport, numa, &uctxt);
 975         if (ret < 0) {
 976                 dd_dev_err(dd, "user ctxtdata allocation failed\n");
 977                 return ret;
 978         }
 979         hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
 980                   uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
 981                   uctxt->numa_id);
 982
 983         /*
 984          * Allocate and enable a PIO send context.
 985          */
 986         uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize, dd->node);
 987         if (!uctxt->sc) {
 988                 ret = -ENOMEM;
 989                 goto ctxdata_free;
 990         }
 991         hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
 992                   uctxt->sc->hw_context);
 993         ret = sc_enable(uctxt->sc);
 994         if (ret)
 995                 goto ctxdata_free;
 996
 997         /*
 998          * Setup sub context information if the user-level has requested
 999          * sub contexts.
1000          * This has to be done here so the rest of the sub-contexts find the
1001          * proper base context.
1002          * NOTE: _set_bit() can be used here because the context creation is
1003          * protected by the mutex (rather than the spin_lock), and will be the
1004          * very first instance of this context.
1005          */
1006         __set_bit(0, uctxt->in_use_ctxts);
1007         if (uinfo->subctxt_cnt)
1008                 init_subctxts(uctxt, uinfo);
1009         uctxt->userversion = uinfo->userversion;
1010         uctxt->flags = hfi1_cap_mask; /* save current flag state */
1011         init_waitqueue_head(&uctxt->wait);
1012         strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1013         memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1014         uctxt->jkey = generate_jkey(current_uid());
1015         hfi1_stats.sps_ctxts++;
1016         /*
1017          * Disable ASPM when there are open user/PSM contexts to avoid
1018          * issues with ASPM L1 exit latency
1019          */
1020         if (dd->freectxts-- == dd->num_user_contexts)
1021                 aspm_disable_all(dd);
1022
1023         *rcd = uctxt;
1024
1025         return 0;
1026
1027 ctxdata_free:
1028         hfi1_free_ctxt(uctxt);
1029         return ret;
1030 }
1031
1032 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt)
1033 {
1034         mutex_lock(&hfi1_mutex);
1035         hfi1_stats.sps_ctxts--;
1036         if (++uctxt->dd->freectxts == uctxt->dd->num_user_contexts)
1037                 aspm_enable_all(uctxt->dd);
1038         mutex_unlock(&hfi1_mutex);
1039
1040         hfi1_free_ctxt(uctxt);
1041 }
1042
1043 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
1044                           const struct hfi1_user_info *uinfo)
1045 {
1046         uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1047         uctxt->subctxt_id = uinfo->subctxt_id;
1048         set_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1049 }
1050
1051 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1052 {
1053         int ret = 0;
1054         u16 num_subctxts = uctxt->subctxt_cnt;
1055
1056         uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1057         if (!uctxt->subctxt_uregbase)
1058                 return -ENOMEM;
1059
1060         /* We can take the size of the RcvHdr Queue from the master */
1061         uctxt->subctxt_rcvhdr_base = vmalloc_user(rcvhdrq_size(uctxt) *
1062                                                   num_subctxts);
1063         if (!uctxt->subctxt_rcvhdr_base) {
1064                 ret = -ENOMEM;
1065                 goto bail_ureg;
1066         }
1067
1068         uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1069                                                 num_subctxts);
1070         if (!uctxt->subctxt_rcvegrbuf) {
1071                 ret = -ENOMEM;
1072                 goto bail_rhdr;
1073         }
1074
1075         return 0;
1076
1077 bail_rhdr:
1078         vfree(uctxt->subctxt_rcvhdr_base);
1079         uctxt->subctxt_rcvhdr_base = NULL;
1080 bail_ureg:
1081         vfree(uctxt->subctxt_uregbase);
1082         uctxt->subctxt_uregbase = NULL;
1083
1084         return ret;
1085 }
1086
1087 static void user_init(struct hfi1_ctxtdata *uctxt)
1088 {
1089         unsigned int rcvctrl_ops = 0;
1090
1091         /* initialize poll variables... */
1092         uctxt->urgent = 0;
1093         uctxt->urgent_poll = 0;
1094
1095         /*
1096          * Now enable the ctxt for receive.
1097          * For chips that are set to DMA the tail register to memory
1098          * when they change (and when the update bit transitions from
1099          * 0 to 1.  So for those chips, we turn it off and then back on.
1100          * This will (very briefly) affect any other open ctxts, but the
1101          * duration is very short, and therefore isn't an issue.  We
1102          * explicitly set the in-memory tail copy to 0 beforehand, so we
1103          * don't have to wait to be sure the DMA update has happened
1104          * (chip resets head/tail to 0 on transition to enable).
1105          */
1106         if (uctxt->rcvhdrtail_kvaddr)
1107                 clear_rcvhdrtail(uctxt);
1108
1109         /* Setup J_KEY before enabling the context */
1110         hfi1_set_ctxt_jkey(uctxt->dd, uctxt, uctxt->jkey);
1111
1112         rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1113         if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1114                 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1115         /*
1116          * Ignore the bit in the flags for now until proper
1117          * support for multiple packet per rcv array entry is
1118          * added.
1119          */
1120         if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1121                 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1122         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1123                 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1124         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1125                 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1126         /*
1127          * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1128          * We can't rely on the correct value to be set from prior
1129          * uses of the chip or ctxt. Therefore, add the rcvctrl op
1130          * for both cases.
1131          */
1132         if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1133                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1134         else
1135                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1136         hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt);
1137 }
1138
1139 static int get_ctxt_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1140 {
1141         struct hfi1_ctxt_info cinfo;
1142         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1143
1144         if (sizeof(cinfo) != len)
1145                 return -EINVAL;
1146
1147         memset(&cinfo, 0, sizeof(cinfo));
1148         cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1149                                 HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1150                         HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1151                         HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1152         /* adjust flag if this fd is not able to cache */
1153         if (!fd->handler)
1154                 cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1155
1156         cinfo.num_active = hfi1_count_active_units();
1157         cinfo.unit = uctxt->dd->unit;
1158         cinfo.ctxt = uctxt->ctxt;
1159         cinfo.subctxt = fd->subctxt;
1160         cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1161                                 uctxt->dd->rcv_entries.group_size) +
1162                 uctxt->expected_count;
1163         cinfo.credits = uctxt->sc->credits;
1164         cinfo.numa_node = uctxt->numa_id;
1165         cinfo.rec_cpu = fd->rec_cpu_num;
1166         cinfo.send_ctxt = uctxt->sc->hw_context;
1167
1168         cinfo.egrtids = uctxt->egrbufs.alloced;
1169         cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1170         cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1171         cinfo.sdma_ring_size = fd->cq->nentries;
1172         cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1173
1174         trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, &cinfo);
1175         if (copy_to_user((void __user *)arg, &cinfo, len))
1176                 return -EFAULT;
1177
1178         return 0;
1179 }
1180
1181 static int init_user_ctxt(struct hfi1_filedata *fd,
1182                           struct hfi1_ctxtdata *uctxt)
1183 {
1184         int ret;
1185
1186         ret = hfi1_user_sdma_alloc_queues(uctxt, fd);
1187         if (ret)
1188                 return ret;
1189
1190         ret = hfi1_user_exp_rcv_init(fd, uctxt);
1191         if (ret)
1192                 hfi1_user_sdma_free_queues(fd, uctxt);
1193
1194         return ret;
1195 }
1196
1197 static int setup_base_ctxt(struct hfi1_filedata *fd,
1198                            struct hfi1_ctxtdata *uctxt)
1199 {
1200         struct hfi1_devdata *dd = uctxt->dd;
1201         int ret = 0;
1202
1203         hfi1_init_ctxt(uctxt->sc);
1204
1205         /* Now allocate the RcvHdr queue and eager buffers. */
1206         ret = hfi1_create_rcvhdrq(dd, uctxt);
1207         if (ret)
1208                 goto done;
1209
1210         ret = hfi1_setup_eagerbufs(uctxt);
1211         if (ret)
1212                 goto done;
1213
1214         /* If sub-contexts are enabled, do the appropriate setup */
1215         if (uctxt->subctxt_cnt)
1216                 ret = setup_subctxt(uctxt);
1217         if (ret)
1218                 goto done;
1219
1220         ret = hfi1_alloc_ctxt_rcv_groups(uctxt);
1221         if (ret)
1222                 goto done;
1223
1224         ret = init_user_ctxt(fd, uctxt);
1225         if (ret) {
1226                 hfi1_free_ctxt_rcv_groups(uctxt);
1227                 goto done;
1228         }
1229
1230         user_init(uctxt);
1231
1232         /* Now that the context is set up, the fd can get a reference. */
1233         fd->uctxt = uctxt;
1234         hfi1_rcd_get(uctxt);
1235
1236 done:
1237         if (uctxt->subctxt_cnt) {
1238                 /*
1239                  * On error, set the failed bit so sub-contexts will clean up
1240                  * correctly.
1241                  */
1242                 if (ret)
1243                         set_bit(HFI1_CTXT_BASE_FAILED, &uctxt->event_flags);
1244
1245                 /*
1246                  * Base context is done (successfully or not), notify anybody
1247                  * using a sub-context that is waiting for this completion.
1248                  */
1249                 clear_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1250                 wake_up(&uctxt->wait);
1251         }
1252
1253         return ret;
1254 }
1255
1256 static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1257 {
1258         struct hfi1_base_info binfo;
1259         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1260         struct hfi1_devdata *dd = uctxt->dd;
1261         unsigned offset;
1262
1263         trace_hfi1_uctxtdata(uctxt->dd, uctxt, fd->subctxt);
1264
1265         if (sizeof(binfo) != len)
1266                 return -EINVAL;
1267
1268         memset(&binfo, 0, sizeof(binfo));
1269         binfo.hw_version = dd->revision;
1270         binfo.sw_version = HFI1_KERN_SWVERSION;
1271         binfo.bthqp = kdeth_qp;
1272         binfo.jkey = uctxt->jkey;
1273         /*
1274          * If more than 64 contexts are enabled the allocated credit
1275          * return will span two or three contiguous pages. Since we only
1276          * map the page containing the context's credit return address,
1277          * we need to calculate the offset in the proper page.
1278          */
1279         offset = ((u64)uctxt->sc->hw_free -
1280                   (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1281         binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1282                                                 fd->subctxt, offset);
1283         binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1284                                             fd->subctxt,
1285                                             uctxt->sc->base_addr);
1286         binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1287                                                 uctxt->ctxt,
1288                                                 fd->subctxt,
1289                                                 uctxt->sc->base_addr);
1290         binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1291                                                fd->subctxt,
1292                                                uctxt->rcvhdrq);
1293         binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1294                                                fd->subctxt,
1295                                                uctxt->egrbufs.rcvtids[0].dma);
1296         binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1297                                                   fd->subctxt, 0);
1298         /*
1299          * user regs are at
1300          * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1301          */
1302         binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1303                                              fd->subctxt, 0);
1304         offset = offset_in_page((uctxt_offset(uctxt) + fd->subctxt) *
1305                                 sizeof(*dd->events));
1306         binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1307                                                fd->subctxt,
1308                                                offset);
1309         binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1310                                                fd->subctxt,
1311                                                dd->status);
1312         if (HFI1_CAP_IS_USET(DMA_RTAIL))
1313                 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1314                                                         fd->subctxt, 0);
1315         if (uctxt->subctxt_cnt) {
1316                 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1317                                                          uctxt->ctxt,
1318                                                          fd->subctxt, 0);
1319                 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1320                                                           uctxt->ctxt,
1321                                                           fd->subctxt, 0);
1322                 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1323                                                           uctxt->ctxt,
1324                                                           fd->subctxt, 0);
1325         }
1326
1327         if (copy_to_user((void __user *)arg, &binfo, len))
1328                 return -EFAULT;
1329
1330         return 0;
1331 }
1332
1333 /**
1334  * user_exp_rcv_setup - Set up the given tid rcv list
1335  * @fd: file data of the current driver instance
1336  * @arg: ioctl argumnent for user space information
1337  * @len: length of data structure associated with ioctl command
1338  *
1339  * Wrapper to validate ioctl information before doing _rcv_setup.
1340  *
1341  */
1342 static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
1343                               u32 len)
1344 {
1345         int ret;
1346         unsigned long addr;
1347         struct hfi1_tid_info tinfo;
1348
1349         if (sizeof(tinfo) != len)
1350                 return -EINVAL;
1351
1352         if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1353                 return -EFAULT;
1354
1355         ret = hfi1_user_exp_rcv_setup(fd, &tinfo);
1356         if (!ret) {
1357                 /*
1358                  * Copy the number of tidlist entries we used
1359                  * and the length of the buffer we registered.
1360                  */
1361                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1362                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1363                                  sizeof(tinfo.tidcnt)))
1364                         ret = -EFAULT;
1365
1366                 addr = arg + offsetof(struct hfi1_tid_info, length);
1367                 if (!ret && copy_to_user((void __user *)addr, &tinfo.length,
1368                                  sizeof(tinfo.length)))
1369                         ret = -EFAULT;
1370
1371                 if (ret)
1372                         hfi1_user_exp_rcv_invalid(fd, &tinfo);
1373         }
1374
1375         return ret;
1376 }
1377
1378 /**
1379  * user_exp_rcv_clear - Clear the given tid rcv list
1380  * @fd: file data of the current driver instance
1381  * @arg: ioctl argumnent for user space information
1382  * @len: length of data structure associated with ioctl command
1383  *
1384  * The hfi1_user_exp_rcv_clear() can be called from the error path.  Because
1385  * of this, we need to use this wrapper to copy the user space information
1386  * before doing the clear.
1387  */
1388 static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
1389                               u32 len)
1390 {
1391         int ret;
1392         unsigned long addr;
1393         struct hfi1_tid_info tinfo;
1394
1395         if (sizeof(tinfo) != len)
1396                 return -EINVAL;
1397
1398         if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1399                 return -EFAULT;
1400
1401         ret = hfi1_user_exp_rcv_clear(fd, &tinfo);
1402         if (!ret) {
1403                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1404                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1405                                  sizeof(tinfo.tidcnt)))
1406                         return -EFAULT;
1407         }
1408
1409         return ret;
1410 }
1411
1412 /**
1413  * user_exp_rcv_invalid - Invalidate the given tid rcv list
1414  * @fd: file data of the current driver instance
1415  * @arg: ioctl argumnent for user space information
1416  * @len: length of data structure associated with ioctl command
1417  *
1418  * Wrapper to validate ioctl information before doing _rcv_invalid.
1419  *
1420  */
1421 static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
1422                                 u32 len)
1423 {
1424         int ret;
1425         unsigned long addr;
1426         struct hfi1_tid_info tinfo;
1427
1428         if (sizeof(tinfo) != len)
1429                 return -EINVAL;
1430
1431         if (!fd->invalid_tids)
1432                 return -EINVAL;
1433
1434         if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1435                 return -EFAULT;
1436
1437         ret = hfi1_user_exp_rcv_invalid(fd, &tinfo);
1438         if (ret)
1439                 return ret;
1440
1441         addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1442         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1443                          sizeof(tinfo.tidcnt)))
1444                 ret = -EFAULT;
1445
1446         return ret;
1447 }
1448
1449 static __poll_t poll_urgent(struct file *fp,
1450                                 struct poll_table_struct *pt)
1451 {
1452         struct hfi1_filedata *fd = fp->private_data;
1453         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1454         struct hfi1_devdata *dd = uctxt->dd;
1455         __poll_t pollflag;
1456
1457         poll_wait(fp, &uctxt->wait, pt);
1458
1459         spin_lock_irq(&dd->uctxt_lock);
1460         if (uctxt->urgent != uctxt->urgent_poll) {
1461                 pollflag = EPOLLIN | EPOLLRDNORM;
1462                 uctxt->urgent_poll = uctxt->urgent;
1463         } else {
1464                 pollflag = 0;
1465                 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1466         }
1467         spin_unlock_irq(&dd->uctxt_lock);
1468
1469         return pollflag;
1470 }
1471
1472 static __poll_t poll_next(struct file *fp,
1473                               struct poll_table_struct *pt)
1474 {
1475         struct hfi1_filedata *fd = fp->private_data;
1476         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1477         struct hfi1_devdata *dd = uctxt->dd;
1478         __poll_t pollflag;
1479
1480         poll_wait(fp, &uctxt->wait, pt);
1481
1482         spin_lock_irq(&dd->uctxt_lock);
1483         if (hdrqempty(uctxt)) {
1484                 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1485                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt);
1486                 pollflag = 0;
1487         } else {
1488                 pollflag = EPOLLIN | EPOLLRDNORM;
1489         }
1490         spin_unlock_irq(&dd->uctxt_lock);
1491
1492         return pollflag;
1493 }
1494
1495 /*
1496  * Find all user contexts in use, and set the specified bit in their
1497  * event mask.
1498  * See also find_ctxt() for a similar use, that is specific to send buffers.
1499  */
1500 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1501 {
1502         struct hfi1_ctxtdata *uctxt;
1503         struct hfi1_devdata *dd = ppd->dd;
1504         u16 ctxt;
1505
1506         if (!dd->events)
1507                 return -EINVAL;
1508
1509         for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
1510              ctxt++) {
1511                 uctxt = hfi1_rcd_get_by_index(dd, ctxt);
1512                 if (uctxt) {
1513                         unsigned long *evs;
1514                         int i;
1515                         /*
1516                          * subctxt_cnt is 0 if not shared, so do base
1517                          * separately, first, then remaining subctxt, if any
1518                          */
1519                         evs = dd->events + uctxt_offset(uctxt);
1520                         set_bit(evtbit, evs);
1521                         for (i = 1; i < uctxt->subctxt_cnt; i++)
1522                                 set_bit(evtbit, evs + i);
1523                         hfi1_rcd_put(uctxt);
1524                 }
1525         }
1526
1527         return 0;
1528 }
1529
1530 /**
1531  * manage_rcvq - manage a context's receive queue
1532  * @uctxt: the context
1533  * @subctxt: the sub-context
1534  * @start_stop: action to carry out
1535  *
1536  * start_stop == 0 disables receive on the context, for use in queue
1537  * overflow conditions.  start_stop==1 re-enables, to be used to
1538  * re-init the software copy of the head register
1539  */
1540 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1541                        unsigned long arg)
1542 {
1543         struct hfi1_devdata *dd = uctxt->dd;
1544         unsigned int rcvctrl_op;
1545         int start_stop;
1546
1547         if (subctxt)
1548                 return 0;
1549
1550         if (get_user(start_stop, (int __user *)arg))
1551                 return -EFAULT;
1552
1553         /* atomically clear receive enable ctxt. */
1554         if (start_stop) {
1555                 /*
1556                  * On enable, force in-memory copy of the tail register to
1557                  * 0, so that protocol code doesn't have to worry about
1558                  * whether or not the chip has yet updated the in-memory
1559                  * copy or not on return from the system call. The chip
1560                  * always resets it's tail register back to 0 on a
1561                  * transition from disabled to enabled.
1562                  */
1563                 if (uctxt->rcvhdrtail_kvaddr)
1564                         clear_rcvhdrtail(uctxt);
1565                 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1566         } else {
1567                 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1568         }
1569         hfi1_rcvctrl(dd, rcvctrl_op, uctxt);
1570         /* always; new head should be equal to new tail; see above */
1571
1572         return 0;
1573 }
1574
1575 /*
1576  * clear the event notifier events for this context.
1577  * User process then performs actions appropriate to bit having been
1578  * set, if desired, and checks again in future.
1579  */
1580 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1581                           unsigned long arg)
1582 {
1583         int i;
1584         struct hfi1_devdata *dd = uctxt->dd;
1585         unsigned long *evs;
1586         unsigned long events;
1587
1588         if (!dd->events)
1589                 return 0;
1590
1591         if (get_user(events, (unsigned long __user *)arg))
1592                 return -EFAULT;
1593
1594         evs = dd->events + uctxt_offset(uctxt) + subctxt;
1595
1596         for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1597                 if (!test_bit(i, &events))
1598                         continue;
1599                 clear_bit(i, evs);
1600         }
1601         return 0;
1602 }
1603
1604 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg)
1605 {
1606         int i;
1607         struct hfi1_pportdata *ppd = uctxt->ppd;
1608         struct hfi1_devdata *dd = uctxt->dd;
1609         u16 pkey;
1610
1611         if (!HFI1_CAP_IS_USET(PKEY_CHECK))
1612                 return -EPERM;
1613
1614         if (get_user(pkey, (u16 __user *)arg))
1615                 return -EFAULT;
1616
1617         if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY)
1618                 return -EINVAL;
1619
1620         for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1621                 if (pkey == ppd->pkeys[i])
1622                         return hfi1_set_ctxt_pkey(dd, uctxt, pkey);
1623
1624         return -ENOENT;
1625 }
1626
1627 /**
1628  * ctxt_reset - Reset the user context
1629  * @uctxt: valid user context
1630  */
1631 static int ctxt_reset(struct hfi1_ctxtdata *uctxt)
1632 {
1633         struct send_context *sc;
1634         struct hfi1_devdata *dd;
1635         int ret = 0;
1636
1637         if (!uctxt || !uctxt->dd || !uctxt->sc)
1638                 return -EINVAL;
1639
1640         /*
1641          * There is no protection here. User level has to guarantee that
1642          * no one will be writing to the send context while it is being
1643          * re-initialized.  If user level breaks that guarantee, it will
1644          * break it's own context and no one else's.
1645          */
1646         dd = uctxt->dd;
1647         sc = uctxt->sc;
1648
1649         /*
1650          * Wait until the interrupt handler has marked the context as
1651          * halted or frozen. Report error if we time out.
1652          */
1653         wait_event_interruptible_timeout(
1654                 sc->halt_wait, (sc->flags & SCF_HALTED),
1655                 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1656         if (!(sc->flags & SCF_HALTED))
1657                 return -ENOLCK;
1658
1659         /*
1660          * If the send context was halted due to a Freeze, wait until the
1661          * device has been "unfrozen" before resetting the context.
1662          */
1663         if (sc->flags & SCF_FROZEN) {
1664                 wait_event_interruptible_timeout(
1665                         dd->event_queue,
1666                         !(READ_ONCE(dd->flags) & HFI1_FROZEN),
1667                         msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1668                 if (dd->flags & HFI1_FROZEN)
1669                         return -ENOLCK;
1670
1671                 if (dd->flags & HFI1_FORCED_FREEZE)
1672                         /*
1673                          * Don't allow context reset if we are into
1674                          * forced freeze
1675                          */
1676                         return -ENODEV;
1677
1678                 sc_disable(sc);
1679                 ret = sc_enable(sc);
1680                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB, uctxt);
1681         } else {
1682                 ret = sc_restart(sc);
1683         }
1684         if (!ret)
1685                 sc_return_credits(sc);
1686
1687         return ret;
1688 }
1689
1690 static void user_remove(struct hfi1_devdata *dd)
1691 {
1692
1693         hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1694 }
1695
1696 static int user_add(struct hfi1_devdata *dd)
1697 {
1698         char name[10];
1699         int ret;
1700
1701         snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1702         ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1703                              &dd->user_cdev, &dd->user_device,
1704                              true, &dd->kobj);
1705         if (ret)
1706                 user_remove(dd);
1707
1708         return ret;
1709 }
1710
1711 /*
1712  * Create per-unit files in /dev
1713  */
1714 int hfi1_device_create(struct hfi1_devdata *dd)
1715 {
1716         return user_add(dd);
1717 }
1718
1719 /*
1720  * Remove per-unit files in /dev
1721  * void, core kernel returns no errors for this stuff
1722  */
1723 void hfi1_device_remove(struct hfi1_devdata *dd)
1724 {
1725         user_remove(dd);
1726 }