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