GNU Linux-libre 5.10.215-gnu1
[releases.git] / drivers / infiniband / hw / hfi1 / file_ops.c
1 /*
2  * Copyright(c) 2020 Cornelis Networks, Inc.
3  * Copyright(c) 2015-2020 Intel Corporation.
4  *
5  * This file is provided under a dual BSD/GPLv2 license.  When using or
6  * redistributing this file, you may do so under either license.
7  *
8  * GPL LICENSE SUMMARY
9  *
10  * This program is free software; you can redistribute it and/or modify
11  * it under the terms of version 2 of the GNU General Public License as
12  * published by the Free Software Foundation.
13  *
14  * This program is distributed in the hope that it will be useful, but
15  * WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
17  * General Public License for more details.
18  *
19  * BSD LICENSE
20  *
21  * Redistribution and use in source and binary forms, with or without
22  * modification, are permitted provided that the following conditions
23  * are met:
24  *
25  *  - Redistributions of source code must retain the above copyright
26  *    notice, this list of conditions and the following disclaimer.
27  *  - Redistributions in binary form must reproduce the above copyright
28  *    notice, this list of conditions and the following disclaimer in
29  *    the documentation and/or other materials provided with the
30  *    distribution.
31  *  - Neither the name of Intel Corporation nor the names of its
32  *    contributors may be used to endorse or promote products derived
33  *    from this software without specific prior written permission.
34  *
35  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
36  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
37  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
38  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
39  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
40  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
41  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
42  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
43  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
44  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
45  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
46  *
47  */
48 #include <linux/poll.h>
49 #include <linux/cdev.h>
50 #include <linux/vmalloc.h>
51 #include <linux/io.h>
52 #include <linux/sched/mm.h>
53 #include <linux/bitmap.h>
54
55 #include <rdma/ib.h>
56
57 #include "hfi.h"
58 #include "pio.h"
59 #include "device.h"
60 #include "common.h"
61 #include "trace.h"
62 #include "mmu_rb.h"
63 #include "user_sdma.h"
64 #include "user_exp_rcv.h"
65 #include "aspm.h"
66
67 #undef pr_fmt
68 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
69
70 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
71
72 /*
73  * File operation functions
74  */
75 static int hfi1_file_open(struct inode *inode, struct file *fp);
76 static int hfi1_file_close(struct inode *inode, struct file *fp);
77 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from);
78 static __poll_t hfi1_poll(struct file *fp, struct poll_table_struct *pt);
79 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma);
80
81 static u64 kvirt_to_phys(void *addr);
82 static int assign_ctxt(struct hfi1_filedata *fd, unsigned long arg, u32 len);
83 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
84                           const struct hfi1_user_info *uinfo);
85 static int init_user_ctxt(struct hfi1_filedata *fd,
86                           struct hfi1_ctxtdata *uctxt);
87 static void user_init(struct hfi1_ctxtdata *uctxt);
88 static int get_ctxt_info(struct hfi1_filedata *fd, unsigned long arg, u32 len);
89 static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len);
90 static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
91                               u32 len);
92 static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
93                               u32 len);
94 static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
95                                 u32 len);
96 static int setup_base_ctxt(struct hfi1_filedata *fd,
97                            struct hfi1_ctxtdata *uctxt);
98 static int setup_subctxt(struct hfi1_ctxtdata *uctxt);
99
100 static int find_sub_ctxt(struct hfi1_filedata *fd,
101                          const struct hfi1_user_info *uinfo);
102 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
103                          struct hfi1_user_info *uinfo,
104                          struct hfi1_ctxtdata **cd);
105 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt);
106 static __poll_t poll_urgent(struct file *fp, struct poll_table_struct *pt);
107 static __poll_t poll_next(struct file *fp, struct poll_table_struct *pt);
108 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
109                           unsigned long arg);
110 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg);
111 static int ctxt_reset(struct hfi1_ctxtdata *uctxt);
112 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
113                        unsigned long arg);
114 static vm_fault_t vma_fault(struct vm_fault *vmf);
115 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
116                             unsigned long arg);
117
118 static const struct file_operations hfi1_file_ops = {
119         .owner = THIS_MODULE,
120         .write_iter = hfi1_write_iter,
121         .open = hfi1_file_open,
122         .release = hfi1_file_close,
123         .unlocked_ioctl = hfi1_file_ioctl,
124         .poll = hfi1_poll,
125         .mmap = hfi1_file_mmap,
126         .llseek = noop_llseek,
127 };
128
129 static const struct vm_operations_struct vm_ops = {
130         .fault = vma_fault,
131 };
132
133 /*
134  * Types of memories mapped into user processes' space
135  */
136 enum mmap_types {
137         PIO_BUFS = 1,
138         PIO_BUFS_SOP,
139         PIO_CRED,
140         RCV_HDRQ,
141         RCV_EGRBUF,
142         UREGS,
143         EVENTS,
144         STATUS,
145         RTAIL,
146         SUBCTXT_UREGS,
147         SUBCTXT_RCV_HDRQ,
148         SUBCTXT_EGRBUF,
149         SDMA_COMP
150 };
151
152 /*
153  * Masks and offsets defining the mmap tokens
154  */
155 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
156 #define HFI1_MMAP_OFFSET_SHIFT  0
157 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
158 #define HFI1_MMAP_SUBCTXT_SHIFT 12
159 #define HFI1_MMAP_CTXT_MASK     0xffULL
160 #define HFI1_MMAP_CTXT_SHIFT    16
161 #define HFI1_MMAP_TYPE_MASK     0xfULL
162 #define HFI1_MMAP_TYPE_SHIFT    24
163 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
164 #define HFI1_MMAP_MAGIC_SHIFT   32
165
166 #define HFI1_MMAP_MAGIC         0xdabbad00
167
168 #define HFI1_MMAP_TOKEN_SET(field, val) \
169         (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
170 #define HFI1_MMAP_TOKEN_GET(field, token) \
171         (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
172 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
173         (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
174         HFI1_MMAP_TOKEN_SET(TYPE, type) | \
175         HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
176         HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
177         HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
178
179 #define dbg(fmt, ...)                           \
180         pr_info(fmt, ##__VA_ARGS__)
181
182 static inline int is_valid_mmap(u64 token)
183 {
184         return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
185 }
186
187 static int hfi1_file_open(struct inode *inode, struct file *fp)
188 {
189         struct hfi1_filedata *fd;
190         struct hfi1_devdata *dd = container_of(inode->i_cdev,
191                                                struct hfi1_devdata,
192                                                user_cdev);
193
194         if (!((dd->flags & HFI1_PRESENT) && dd->kregbase1))
195                 return -EINVAL;
196
197         if (!atomic_inc_not_zero(&dd->user_refcount))
198                 return -ENXIO;
199
200         /* The real work is performed later in assign_ctxt() */
201
202         fd = kzalloc(sizeof(*fd), GFP_KERNEL);
203
204         if (!fd || init_srcu_struct(&fd->pq_srcu))
205                 goto nomem;
206         spin_lock_init(&fd->pq_rcu_lock);
207         spin_lock_init(&fd->tid_lock);
208         spin_lock_init(&fd->invalid_lock);
209         fd->rec_cpu_num = -1; /* no cpu affinity by default */
210         fd->dd = dd;
211         fp->private_data = fd;
212         return 0;
213 nomem:
214         kfree(fd);
215         fp->private_data = NULL;
216         if (atomic_dec_and_test(&dd->user_refcount))
217                 complete(&dd->user_comp);
218         return -ENOMEM;
219 }
220
221 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
222                             unsigned long arg)
223 {
224         struct hfi1_filedata *fd = fp->private_data;
225         struct hfi1_ctxtdata *uctxt = fd->uctxt;
226         int ret = 0;
227         int uval = 0;
228
229         hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
230         if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
231             cmd != HFI1_IOCTL_GET_VERS &&
232             !uctxt)
233                 return -EINVAL;
234
235         switch (cmd) {
236         case HFI1_IOCTL_ASSIGN_CTXT:
237                 ret = assign_ctxt(fd, arg, _IOC_SIZE(cmd));
238                 break;
239
240         case HFI1_IOCTL_CTXT_INFO:
241                 ret = get_ctxt_info(fd, arg, _IOC_SIZE(cmd));
242                 break;
243
244         case HFI1_IOCTL_USER_INFO:
245                 ret = get_base_info(fd, arg, _IOC_SIZE(cmd));
246                 break;
247
248         case HFI1_IOCTL_CREDIT_UPD:
249                 if (uctxt)
250                         sc_return_credits(uctxt->sc);
251                 break;
252
253         case HFI1_IOCTL_TID_UPDATE:
254                 ret = user_exp_rcv_setup(fd, arg, _IOC_SIZE(cmd));
255                 break;
256
257         case HFI1_IOCTL_TID_FREE:
258                 ret = user_exp_rcv_clear(fd, arg, _IOC_SIZE(cmd));
259                 break;
260
261         case HFI1_IOCTL_TID_INVAL_READ:
262                 ret = user_exp_rcv_invalid(fd, arg, _IOC_SIZE(cmd));
263                 break;
264
265         case HFI1_IOCTL_RECV_CTRL:
266                 ret = manage_rcvq(uctxt, fd->subctxt, arg);
267                 break;
268
269         case HFI1_IOCTL_POLL_TYPE:
270                 if (get_user(uval, (int __user *)arg))
271                         return -EFAULT;
272                 uctxt->poll_type = (typeof(uctxt->poll_type))uval;
273                 break;
274
275         case HFI1_IOCTL_ACK_EVENT:
276                 ret = user_event_ack(uctxt, fd->subctxt, arg);
277                 break;
278
279         case HFI1_IOCTL_SET_PKEY:
280                 ret = set_ctxt_pkey(uctxt, arg);
281                 break;
282
283         case HFI1_IOCTL_CTXT_RESET:
284                 ret = ctxt_reset(uctxt);
285                 break;
286
287         case HFI1_IOCTL_GET_VERS:
288                 uval = HFI1_USER_SWVERSION;
289                 if (put_user(uval, (int __user *)arg))
290                         return -EFAULT;
291                 break;
292
293         default:
294                 return -EINVAL;
295         }
296
297         return ret;
298 }
299
300 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
301 {
302         struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
303         struct hfi1_user_sdma_pkt_q *pq;
304         struct hfi1_user_sdma_comp_q *cq = fd->cq;
305         int done = 0, reqs = 0;
306         unsigned long dim = from->nr_segs;
307         int idx;
308
309         if (!HFI1_CAP_IS_KSET(SDMA))
310                 return -EINVAL;
311         idx = srcu_read_lock(&fd->pq_srcu);
312         pq = srcu_dereference(fd->pq, &fd->pq_srcu);
313         if (!cq || !pq) {
314                 srcu_read_unlock(&fd->pq_srcu, idx);
315                 return -EIO;
316         }
317
318         if (!iter_is_iovec(from) || !dim) {
319                 srcu_read_unlock(&fd->pq_srcu, idx);
320                 return -EINVAL;
321         }
322
323         trace_hfi1_sdma_request(fd->dd, fd->uctxt->ctxt, fd->subctxt, dim);
324
325         if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
326                 srcu_read_unlock(&fd->pq_srcu, idx);
327                 return -ENOSPC;
328         }
329
330         while (dim) {
331                 int ret;
332                 unsigned long count = 0;
333
334                 ret = hfi1_user_sdma_process_request(
335                         fd, (struct iovec *)(from->iov + done),
336                         dim, &count);
337                 if (ret) {
338                         reqs = ret;
339                         break;
340                 }
341                 dim -= count;
342                 done += count;
343                 reqs++;
344         }
345
346         srcu_read_unlock(&fd->pq_srcu, idx);
347         return reqs;
348 }
349
350 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
351 {
352         struct hfi1_filedata *fd = fp->private_data;
353         struct hfi1_ctxtdata *uctxt = fd->uctxt;
354         struct hfi1_devdata *dd;
355         unsigned long flags;
356         u64 token = vma->vm_pgoff << PAGE_SHIFT,
357                 memaddr = 0;
358         void *memvirt = NULL;
359         u8 subctxt, mapio = 0, vmf = 0, type;
360         ssize_t memlen = 0;
361         int ret = 0;
362         u16 ctxt;
363
364         if (!is_valid_mmap(token) || !uctxt ||
365             !(vma->vm_flags & VM_SHARED)) {
366                 ret = -EINVAL;
367                 goto done;
368         }
369         dd = uctxt->dd;
370         ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
371         subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
372         type = HFI1_MMAP_TOKEN_GET(TYPE, token);
373         if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
374                 ret = -EINVAL;
375                 goto done;
376         }
377
378         flags = vma->vm_flags;
379
380         switch (type) {
381         case PIO_BUFS:
382         case PIO_BUFS_SOP:
383                 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
384                                 /* chip pio base */
385                            (uctxt->sc->hw_context * BIT(16))) +
386                                 /* 64K PIO space / ctxt */
387                         (type == PIO_BUFS_SOP ?
388                                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
389                 /*
390                  * Map only the amount allocated to the context, not the
391                  * entire available context's PIO space.
392                  */
393                 memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
394                 flags &= ~VM_MAYREAD;
395                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
396                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
397                 mapio = 1;
398                 break;
399         case PIO_CRED:
400                 if (flags & VM_WRITE) {
401                         ret = -EPERM;
402                         goto done;
403                 }
404                 /*
405                  * The credit return location for this context could be on the
406                  * second or third page allocated for credit returns (if number
407                  * of enabled contexts > 64 and 128 respectively).
408                  */
409                 memvirt = dd->cr_base[uctxt->numa_id].va;
410                 memaddr = virt_to_phys(memvirt) +
411                         (((u64)uctxt->sc->hw_free -
412                           (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
413                 memlen = PAGE_SIZE;
414                 flags &= ~VM_MAYWRITE;
415                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
416                 /*
417                  * The driver has already allocated memory for credit
418                  * returns and programmed it into the chip. Has that
419                  * memory been flagged as non-cached?
420                  */
421                 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
422                 mapio = 1;
423                 break;
424         case RCV_HDRQ:
425                 memlen = rcvhdrq_size(uctxt);
426                 memvirt = uctxt->rcvhdrq;
427                 break;
428         case RCV_EGRBUF: {
429                 unsigned long addr;
430                 int i;
431                 /*
432                  * The RcvEgr buffer need to be handled differently
433                  * as multiple non-contiguous pages need to be mapped
434                  * into the user process.
435                  */
436                 memlen = uctxt->egrbufs.size;
437                 if ((vma->vm_end - vma->vm_start) != memlen) {
438                         dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
439                                    (vma->vm_end - vma->vm_start), memlen);
440                         ret = -EINVAL;
441                         goto done;
442                 }
443                 if (vma->vm_flags & VM_WRITE) {
444                         ret = -EPERM;
445                         goto done;
446                 }
447                 vma->vm_flags &= ~VM_MAYWRITE;
448                 addr = vma->vm_start;
449                 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
450                         memlen = uctxt->egrbufs.buffers[i].len;
451                         memvirt = uctxt->egrbufs.buffers[i].addr;
452                         ret = remap_pfn_range(
453                                 vma, addr,
454                                 /*
455                                  * virt_to_pfn() does the same, but
456                                  * it's not available on x86_64
457                                  * when CONFIG_MMU is enabled.
458                                  */
459                                 PFN_DOWN(__pa(memvirt)),
460                                 memlen,
461                                 vma->vm_page_prot);
462                         if (ret < 0)
463                                 goto done;
464                         addr += memlen;
465                 }
466                 ret = 0;
467                 goto done;
468         }
469         case UREGS:
470                 /*
471                  * Map only the page that contains this context's user
472                  * registers.
473                  */
474                 memaddr = (unsigned long)
475                         (dd->physaddr + RXE_PER_CONTEXT_USER)
476                         + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
477                 /*
478                  * TidFlow table is on the same page as the rest of the
479                  * user registers.
480                  */
481                 memlen = PAGE_SIZE;
482                 flags |= VM_DONTCOPY | VM_DONTEXPAND;
483                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
484                 mapio = 1;
485                 break;
486         case EVENTS:
487                 /*
488                  * Use the page where this context's flags are. User level
489                  * knows where it's own bitmap is within the page.
490                  */
491                 memaddr = (unsigned long)
492                         (dd->events + uctxt_offset(uctxt)) & PAGE_MASK;
493                 memlen = PAGE_SIZE;
494                 /*
495                  * v3.7 removes VM_RESERVED but the effect is kept by
496                  * using VM_IO.
497                  */
498                 flags |= VM_IO | VM_DONTEXPAND;
499                 vmf = 1;
500                 break;
501         case STATUS:
502                 if (flags & VM_WRITE) {
503                         ret = -EPERM;
504                         goto done;
505                 }
506                 memaddr = kvirt_to_phys((void *)dd->status);
507                 memlen = PAGE_SIZE;
508                 flags |= VM_IO | VM_DONTEXPAND;
509                 break;
510         case RTAIL:
511                 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
512                         /*
513                          * If the memory allocation failed, the context alloc
514                          * also would have failed, so we would never get here
515                          */
516                         ret = -EINVAL;
517                         goto done;
518                 }
519                 if ((flags & VM_WRITE) || !hfi1_rcvhdrtail_kvaddr(uctxt)) {
520                         ret = -EPERM;
521                         goto done;
522                 }
523                 memlen = PAGE_SIZE;
524                 memvirt = (void *)hfi1_rcvhdrtail_kvaddr(uctxt);
525                 flags &= ~VM_MAYWRITE;
526                 break;
527         case SUBCTXT_UREGS:
528                 memaddr = (u64)uctxt->subctxt_uregbase;
529                 memlen = PAGE_SIZE;
530                 flags |= VM_IO | VM_DONTEXPAND;
531                 vmf = 1;
532                 break;
533         case SUBCTXT_RCV_HDRQ:
534                 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
535                 memlen = rcvhdrq_size(uctxt) * uctxt->subctxt_cnt;
536                 flags |= VM_IO | VM_DONTEXPAND;
537                 vmf = 1;
538                 break;
539         case SUBCTXT_EGRBUF:
540                 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
541                 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
542                 flags |= VM_IO | VM_DONTEXPAND;
543                 flags &= ~VM_MAYWRITE;
544                 vmf = 1;
545                 break;
546         case SDMA_COMP: {
547                 struct hfi1_user_sdma_comp_q *cq = fd->cq;
548
549                 if (!cq) {
550                         ret = -EFAULT;
551                         goto done;
552                 }
553                 memaddr = (u64)cq->comps;
554                 memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
555                 flags |= VM_IO | VM_DONTEXPAND;
556                 vmf = 1;
557                 break;
558         }
559         default:
560                 ret = -EINVAL;
561                 break;
562         }
563
564         if ((vma->vm_end - vma->vm_start) != memlen) {
565                 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
566                           uctxt->ctxt, fd->subctxt,
567                           (vma->vm_end - vma->vm_start), memlen);
568                 ret = -EINVAL;
569                 goto done;
570         }
571
572         vma->vm_flags = flags;
573         hfi1_cdbg(PROC,
574                   "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
575                     ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
576                     vma->vm_end - vma->vm_start, vma->vm_flags);
577         if (vmf) {
578                 vma->vm_pgoff = PFN_DOWN(memaddr);
579                 vma->vm_ops = &vm_ops;
580                 ret = 0;
581         } else if (mapio) {
582                 ret = io_remap_pfn_range(vma, vma->vm_start,
583                                          PFN_DOWN(memaddr),
584                                          memlen,
585                                          vma->vm_page_prot);
586         } else if (memvirt) {
587                 ret = remap_pfn_range(vma, vma->vm_start,
588                                       PFN_DOWN(__pa(memvirt)),
589                                       memlen,
590                                       vma->vm_page_prot);
591         } else {
592                 ret = remap_pfn_range(vma, vma->vm_start,
593                                       PFN_DOWN(memaddr),
594                                       memlen,
595                                       vma->vm_page_prot);
596         }
597 done:
598         return ret;
599 }
600
601 /*
602  * Local (non-chip) user memory is not mapped right away but as it is
603  * accessed by the user-level code.
604  */
605 static vm_fault_t vma_fault(struct vm_fault *vmf)
606 {
607         struct page *page;
608
609         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
610         if (!page)
611                 return VM_FAULT_SIGBUS;
612
613         get_page(page);
614         vmf->page = page;
615
616         return 0;
617 }
618
619 static __poll_t hfi1_poll(struct file *fp, struct poll_table_struct *pt)
620 {
621         struct hfi1_ctxtdata *uctxt;
622         __poll_t pollflag;
623
624         uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
625         if (!uctxt)
626                 pollflag = EPOLLERR;
627         else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
628                 pollflag = poll_urgent(fp, pt);
629         else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
630                 pollflag = poll_next(fp, pt);
631         else /* invalid */
632                 pollflag = EPOLLERR;
633
634         return pollflag;
635 }
636
637 static int hfi1_file_close(struct inode *inode, struct file *fp)
638 {
639         struct hfi1_filedata *fdata = fp->private_data;
640         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
641         struct hfi1_devdata *dd = container_of(inode->i_cdev,
642                                                struct hfi1_devdata,
643                                                user_cdev);
644         unsigned long flags, *ev;
645
646         fp->private_data = NULL;
647
648         if (!uctxt)
649                 goto done;
650
651         hfi1_cdbg(PROC, "closing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
652
653         flush_wc();
654         /* drain user sdma queue */
655         hfi1_user_sdma_free_queues(fdata, uctxt);
656
657         /* release the cpu */
658         hfi1_put_proc_affinity(fdata->rec_cpu_num);
659
660         /* clean up rcv side */
661         hfi1_user_exp_rcv_free(fdata);
662
663         /*
664          * fdata->uctxt is used in the above cleanup.  It is not ready to be
665          * removed until here.
666          */
667         fdata->uctxt = NULL;
668         hfi1_rcd_put(uctxt);
669
670         /*
671          * Clear any left over, unhandled events so the next process that
672          * gets this context doesn't get confused.
673          */
674         ev = dd->events + uctxt_offset(uctxt) + fdata->subctxt;
675         *ev = 0;
676
677         spin_lock_irqsave(&dd->uctxt_lock, flags);
678         __clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
679         if (!bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
680                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
681                 goto done;
682         }
683         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
684
685         /*
686          * Disable receive context and interrupt available, reset all
687          * RcvCtxtCtrl bits to default values.
688          */
689         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
690                      HFI1_RCVCTRL_TIDFLOW_DIS |
691                      HFI1_RCVCTRL_INTRAVAIL_DIS |
692                      HFI1_RCVCTRL_TAILUPD_DIS |
693                      HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
694                      HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
695                      HFI1_RCVCTRL_NO_EGR_DROP_DIS |
696                      HFI1_RCVCTRL_URGENT_DIS, uctxt);
697         /* Clear the context's J_KEY */
698         hfi1_clear_ctxt_jkey(dd, uctxt);
699         /*
700          * If a send context is allocated, reset context integrity
701          * checks to default and disable the send context.
702          */
703         if (uctxt->sc) {
704                 sc_disable(uctxt->sc);
705                 set_pio_integrity(uctxt->sc);
706         }
707
708         hfi1_free_ctxt_rcv_groups(uctxt);
709         hfi1_clear_ctxt_pkey(dd, uctxt);
710
711         uctxt->event_flags = 0;
712
713         deallocate_ctxt(uctxt);
714 done:
715
716         if (atomic_dec_and_test(&dd->user_refcount))
717                 complete(&dd->user_comp);
718
719         cleanup_srcu_struct(&fdata->pq_srcu);
720         kfree(fdata);
721         return 0;
722 }
723
724 /*
725  * Convert kernel *virtual* addresses to physical addresses.
726  * This is used to vmalloc'ed addresses.
727  */
728 static u64 kvirt_to_phys(void *addr)
729 {
730         struct page *page;
731         u64 paddr = 0;
732
733         page = vmalloc_to_page(addr);
734         if (page)
735                 paddr = page_to_pfn(page) << PAGE_SHIFT;
736
737         return paddr;
738 }
739
740 /**
741  * complete_subctxt
742  * @fd: valid filedata pointer
743  *
744  * Sub-context info can only be set up after the base context
745  * has been completed.  This is indicated by the clearing of the
746  * HFI1_CTXT_BASE_UINIT bit.
747  *
748  * Wait for the bit to be cleared, and then complete the subcontext
749  * initialization.
750  *
751  */
752 static int complete_subctxt(struct hfi1_filedata *fd)
753 {
754         int ret;
755         unsigned long flags;
756
757         /*
758          * sub-context info can only be set up after the base context
759          * has been completed.
760          */
761         ret = wait_event_interruptible(
762                 fd->uctxt->wait,
763                 !test_bit(HFI1_CTXT_BASE_UNINIT, &fd->uctxt->event_flags));
764
765         if (test_bit(HFI1_CTXT_BASE_FAILED, &fd->uctxt->event_flags))
766                 ret = -ENOMEM;
767
768         /* Finish the sub-context init */
769         if (!ret) {
770                 fd->rec_cpu_num = hfi1_get_proc_affinity(fd->uctxt->numa_id);
771                 ret = init_user_ctxt(fd, fd->uctxt);
772         }
773
774         if (ret) {
775                 spin_lock_irqsave(&fd->dd->uctxt_lock, flags);
776                 __clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
777                 spin_unlock_irqrestore(&fd->dd->uctxt_lock, flags);
778                 hfi1_rcd_put(fd->uctxt);
779                 fd->uctxt = NULL;
780         }
781
782         return ret;
783 }
784
785 static int assign_ctxt(struct hfi1_filedata *fd, unsigned long arg, u32 len)
786 {
787         int ret;
788         unsigned int swmajor;
789         struct hfi1_ctxtdata *uctxt = NULL;
790         struct hfi1_user_info uinfo;
791
792         if (fd->uctxt)
793                 return -EINVAL;
794
795         if (sizeof(uinfo) != len)
796                 return -EINVAL;
797
798         if (copy_from_user(&uinfo, (void __user *)arg, sizeof(uinfo)))
799                 return -EFAULT;
800
801         swmajor = uinfo.userversion >> 16;
802         if (swmajor != HFI1_USER_SWMAJOR)
803                 return -ENODEV;
804
805         if (uinfo.subctxt_cnt > HFI1_MAX_SHARED_CTXTS)
806                 return -EINVAL;
807
808         /*
809          * Acquire the mutex to protect against multiple creations of what
810          * could be a shared base context.
811          */
812         mutex_lock(&hfi1_mutex);
813         /*
814          * Get a sub context if available  (fd->uctxt will be set).
815          * ret < 0 error, 0 no context, 1 sub-context found
816          */
817         ret = find_sub_ctxt(fd, &uinfo);
818
819         /*
820          * Allocate a base context if context sharing is not required or a
821          * sub context wasn't found.
822          */
823         if (!ret)
824                 ret = allocate_ctxt(fd, fd->dd, &uinfo, &uctxt);
825
826         mutex_unlock(&hfi1_mutex);
827
828         /* Depending on the context type, finish the appropriate init */
829         switch (ret) {
830         case 0:
831                 ret = setup_base_ctxt(fd, uctxt);
832                 if (ret)
833                         deallocate_ctxt(uctxt);
834                 break;
835         case 1:
836                 ret = complete_subctxt(fd);
837                 break;
838         default:
839                 break;
840         }
841
842         return ret;
843 }
844
845 /**
846  * match_ctxt
847  * @fd: valid filedata pointer
848  * @uinfo: user info to compare base context with
849  * @uctxt: context to compare uinfo to.
850  *
851  * Compare the given context with the given information to see if it
852  * can be used for a sub context.
853  */
854 static int match_ctxt(struct hfi1_filedata *fd,
855                       const struct hfi1_user_info *uinfo,
856                       struct hfi1_ctxtdata *uctxt)
857 {
858         struct hfi1_devdata *dd = fd->dd;
859         unsigned long flags;
860         u16 subctxt;
861
862         /* Skip dynamically allocated kernel contexts */
863         if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
864                 return 0;
865
866         /* Skip ctxt if it doesn't match the requested one */
867         if (memcmp(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid)) ||
868             uctxt->jkey != generate_jkey(current_uid()) ||
869             uctxt->subctxt_id != uinfo->subctxt_id ||
870             uctxt->subctxt_cnt != uinfo->subctxt_cnt)
871                 return 0;
872
873         /* Verify the sharing process matches the base */
874         if (uctxt->userversion != uinfo->userversion)
875                 return -EINVAL;
876
877         /* Find an unused sub context */
878         spin_lock_irqsave(&dd->uctxt_lock, flags);
879         if (bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
880                 /* context is being closed, do not use */
881                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
882                 return 0;
883         }
884
885         subctxt = find_first_zero_bit(uctxt->in_use_ctxts,
886                                       HFI1_MAX_SHARED_CTXTS);
887         if (subctxt >= uctxt->subctxt_cnt) {
888                 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
889                 return -EBUSY;
890         }
891
892         fd->subctxt = subctxt;
893         __set_bit(fd->subctxt, uctxt->in_use_ctxts);
894         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
895
896         fd->uctxt = uctxt;
897         hfi1_rcd_get(uctxt);
898
899         return 1;
900 }
901
902 /**
903  * find_sub_ctxt
904  * @fd: valid filedata pointer
905  * @uinfo: matching info to use to find a possible context to share.
906  *
907  * The hfi1_mutex must be held when this function is called.  It is
908  * necessary to ensure serialized creation of shared contexts.
909  *
910  * Return:
911  *    0      No sub-context found
912  *    1      Subcontext found and allocated
913  *    errno  EINVAL (incorrect parameters)
914  *           EBUSY (all sub contexts in use)
915  */
916 static int find_sub_ctxt(struct hfi1_filedata *fd,
917                          const struct hfi1_user_info *uinfo)
918 {
919         struct hfi1_ctxtdata *uctxt;
920         struct hfi1_devdata *dd = fd->dd;
921         u16 i;
922         int ret;
923
924         if (!uinfo->subctxt_cnt)
925                 return 0;
926
927         for (i = dd->first_dyn_alloc_ctxt; i < dd->num_rcv_contexts; i++) {
928                 uctxt = hfi1_rcd_get_by_index(dd, i);
929                 if (uctxt) {
930                         ret = match_ctxt(fd, uinfo, uctxt);
931                         hfi1_rcd_put(uctxt);
932                         /* value of != 0 will return */
933                         if (ret)
934                                 return ret;
935                 }
936         }
937
938         return 0;
939 }
940
941 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
942                          struct hfi1_user_info *uinfo,
943                          struct hfi1_ctxtdata **rcd)
944 {
945         struct hfi1_ctxtdata *uctxt;
946         int ret, numa;
947
948         if (dd->flags & HFI1_FROZEN) {
949                 /*
950                  * Pick an error that is unique from all other errors
951                  * that are returned so the user process knows that
952                  * it tried to allocate while the SPC was frozen.  It
953                  * it should be able to retry with success in a short
954                  * while.
955                  */
956                 return -EIO;
957         }
958
959         if (!dd->freectxts)
960                 return -EBUSY;
961
962         /*
963          * If we don't have a NUMA node requested, preference is towards
964          * device NUMA node.
965          */
966         fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
967         if (fd->rec_cpu_num != -1)
968                 numa = cpu_to_node(fd->rec_cpu_num);
969         else
970                 numa = numa_node_id();
971         ret = hfi1_create_ctxtdata(dd->pport, numa, &uctxt);
972         if (ret < 0) {
973                 dd_dev_err(dd, "user ctxtdata allocation failed\n");
974                 return ret;
975         }
976         hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
977                   uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
978                   uctxt->numa_id);
979
980         /*
981          * Allocate and enable a PIO send context.
982          */
983         uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize, dd->node);
984         if (!uctxt->sc) {
985                 ret = -ENOMEM;
986                 goto ctxdata_free;
987         }
988         hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
989                   uctxt->sc->hw_context);
990         ret = sc_enable(uctxt->sc);
991         if (ret)
992                 goto ctxdata_free;
993
994         /*
995          * Setup sub context information if the user-level has requested
996          * sub contexts.
997          * This has to be done here so the rest of the sub-contexts find the
998          * proper base context.
999          * NOTE: _set_bit() can be used here because the context creation is
1000          * protected by the mutex (rather than the spin_lock), and will be the
1001          * very first instance of this context.
1002          */
1003         __set_bit(0, uctxt->in_use_ctxts);
1004         if (uinfo->subctxt_cnt)
1005                 init_subctxts(uctxt, uinfo);
1006         uctxt->userversion = uinfo->userversion;
1007         uctxt->flags = hfi1_cap_mask; /* save current flag state */
1008         init_waitqueue_head(&uctxt->wait);
1009         strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1010         memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1011         uctxt->jkey = generate_jkey(current_uid());
1012         hfi1_stats.sps_ctxts++;
1013         /*
1014          * Disable ASPM when there are open user/PSM contexts to avoid
1015          * issues with ASPM L1 exit latency
1016          */
1017         if (dd->freectxts-- == dd->num_user_contexts)
1018                 aspm_disable_all(dd);
1019
1020         *rcd = uctxt;
1021
1022         return 0;
1023
1024 ctxdata_free:
1025         hfi1_free_ctxt(uctxt);
1026         return ret;
1027 }
1028
1029 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt)
1030 {
1031         mutex_lock(&hfi1_mutex);
1032         hfi1_stats.sps_ctxts--;
1033         if (++uctxt->dd->freectxts == uctxt->dd->num_user_contexts)
1034                 aspm_enable_all(uctxt->dd);
1035         mutex_unlock(&hfi1_mutex);
1036
1037         hfi1_free_ctxt(uctxt);
1038 }
1039
1040 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
1041                           const struct hfi1_user_info *uinfo)
1042 {
1043         uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1044         uctxt->subctxt_id = uinfo->subctxt_id;
1045         set_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1046 }
1047
1048 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1049 {
1050         int ret = 0;
1051         u16 num_subctxts = uctxt->subctxt_cnt;
1052
1053         uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1054         if (!uctxt->subctxt_uregbase)
1055                 return -ENOMEM;
1056
1057         /* We can take the size of the RcvHdr Queue from the master */
1058         uctxt->subctxt_rcvhdr_base = vmalloc_user(rcvhdrq_size(uctxt) *
1059                                                   num_subctxts);
1060         if (!uctxt->subctxt_rcvhdr_base) {
1061                 ret = -ENOMEM;
1062                 goto bail_ureg;
1063         }
1064
1065         uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1066                                                 num_subctxts);
1067         if (!uctxt->subctxt_rcvegrbuf) {
1068                 ret = -ENOMEM;
1069                 goto bail_rhdr;
1070         }
1071
1072         return 0;
1073
1074 bail_rhdr:
1075         vfree(uctxt->subctxt_rcvhdr_base);
1076         uctxt->subctxt_rcvhdr_base = NULL;
1077 bail_ureg:
1078         vfree(uctxt->subctxt_uregbase);
1079         uctxt->subctxt_uregbase = NULL;
1080
1081         return ret;
1082 }
1083
1084 static void user_init(struct hfi1_ctxtdata *uctxt)
1085 {
1086         unsigned int rcvctrl_ops = 0;
1087
1088         /* initialize poll variables... */
1089         uctxt->urgent = 0;
1090         uctxt->urgent_poll = 0;
1091
1092         /*
1093          * Now enable the ctxt for receive.
1094          * For chips that are set to DMA the tail register to memory
1095          * when they change (and when the update bit transitions from
1096          * 0 to 1.  So for those chips, we turn it off and then back on.
1097          * This will (very briefly) affect any other open ctxts, but the
1098          * duration is very short, and therefore isn't an issue.  We
1099          * explicitly set the in-memory tail copy to 0 beforehand, so we
1100          * don't have to wait to be sure the DMA update has happened
1101          * (chip resets head/tail to 0 on transition to enable).
1102          */
1103         if (hfi1_rcvhdrtail_kvaddr(uctxt))
1104                 clear_rcvhdrtail(uctxt);
1105
1106         /* Setup J_KEY before enabling the context */
1107         hfi1_set_ctxt_jkey(uctxt->dd, uctxt, uctxt->jkey);
1108
1109         rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1110         rcvctrl_ops |= HFI1_RCVCTRL_URGENT_ENB;
1111         if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1112                 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1113         /*
1114          * Ignore the bit in the flags for now until proper
1115          * support for multiple packet per rcv array entry is
1116          * added.
1117          */
1118         if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1119                 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1120         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1121                 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1122         if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1123                 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1124         /*
1125          * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1126          * We can't rely on the correct value to be set from prior
1127          * uses of the chip or ctxt. Therefore, add the rcvctrl op
1128          * for both cases.
1129          */
1130         if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1131                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1132         else
1133                 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1134         hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt);
1135 }
1136
1137 static int get_ctxt_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1138 {
1139         struct hfi1_ctxt_info cinfo;
1140         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1141
1142         if (sizeof(cinfo) != len)
1143                 return -EINVAL;
1144
1145         memset(&cinfo, 0, sizeof(cinfo));
1146         cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1147                                 HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1148                         HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1149                         HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1150         /* adjust flag if this fd is not able to cache */
1151         if (!fd->use_mn)
1152                 cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1153
1154         cinfo.num_active = hfi1_count_active_units();
1155         cinfo.unit = uctxt->dd->unit;
1156         cinfo.ctxt = uctxt->ctxt;
1157         cinfo.subctxt = fd->subctxt;
1158         cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1159                                 uctxt->dd->rcv_entries.group_size) +
1160                 uctxt->expected_count;
1161         cinfo.credits = uctxt->sc->credits;
1162         cinfo.numa_node = uctxt->numa_id;
1163         cinfo.rec_cpu = fd->rec_cpu_num;
1164         cinfo.send_ctxt = uctxt->sc->hw_context;
1165
1166         cinfo.egrtids = uctxt->egrbufs.alloced;
1167         cinfo.rcvhdrq_cnt = get_hdrq_cnt(uctxt);
1168         cinfo.rcvhdrq_entsize = get_hdrqentsize(uctxt) << 2;
1169         cinfo.sdma_ring_size = fd->cq->nentries;
1170         cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1171
1172         trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, &cinfo);
1173         if (copy_to_user((void __user *)arg, &cinfo, len))
1174                 return -EFAULT;
1175
1176         return 0;
1177 }
1178
1179 static int init_user_ctxt(struct hfi1_filedata *fd,
1180                           struct hfi1_ctxtdata *uctxt)
1181 {
1182         int ret;
1183
1184         ret = hfi1_user_sdma_alloc_queues(uctxt, fd);
1185         if (ret)
1186                 return ret;
1187
1188         ret = hfi1_user_exp_rcv_init(fd, uctxt);
1189         if (ret)
1190                 hfi1_user_sdma_free_queues(fd, uctxt);
1191
1192         return ret;
1193 }
1194
1195 static int setup_base_ctxt(struct hfi1_filedata *fd,
1196                            struct hfi1_ctxtdata *uctxt)
1197 {
1198         struct hfi1_devdata *dd = uctxt->dd;
1199         int ret = 0;
1200
1201         hfi1_init_ctxt(uctxt->sc);
1202
1203         /* Now allocate the RcvHdr queue and eager buffers. */
1204         ret = hfi1_create_rcvhdrq(dd, uctxt);
1205         if (ret)
1206                 goto done;
1207
1208         ret = hfi1_setup_eagerbufs(uctxt);
1209         if (ret)
1210                 goto done;
1211
1212         /* If sub-contexts are enabled, do the appropriate setup */
1213         if (uctxt->subctxt_cnt)
1214                 ret = setup_subctxt(uctxt);
1215         if (ret)
1216                 goto done;
1217
1218         ret = hfi1_alloc_ctxt_rcv_groups(uctxt);
1219         if (ret)
1220                 goto done;
1221
1222         ret = init_user_ctxt(fd, uctxt);
1223         if (ret) {
1224                 hfi1_free_ctxt_rcv_groups(uctxt);
1225                 goto done;
1226         }
1227
1228         user_init(uctxt);
1229
1230         /* Now that the context is set up, the fd can get a reference. */
1231         fd->uctxt = uctxt;
1232         hfi1_rcd_get(uctxt);
1233
1234 done:
1235         if (uctxt->subctxt_cnt) {
1236                 /*
1237                  * On error, set the failed bit so sub-contexts will clean up
1238                  * correctly.
1239                  */
1240                 if (ret)
1241                         set_bit(HFI1_CTXT_BASE_FAILED, &uctxt->event_flags);
1242
1243                 /*
1244                  * Base context is done (successfully or not), notify anybody
1245                  * using a sub-context that is waiting for this completion.
1246                  */
1247                 clear_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1248                 wake_up(&uctxt->wait);
1249         }
1250
1251         return ret;
1252 }
1253
1254 static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1255 {
1256         struct hfi1_base_info binfo;
1257         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1258         struct hfi1_devdata *dd = uctxt->dd;
1259         unsigned offset;
1260
1261         trace_hfi1_uctxtdata(uctxt->dd, uctxt, fd->subctxt);
1262
1263         if (sizeof(binfo) != len)
1264                 return -EINVAL;
1265
1266         memset(&binfo, 0, sizeof(binfo));
1267         binfo.hw_version = dd->revision;
1268         binfo.sw_version = HFI1_KERN_SWVERSION;
1269         binfo.bthqp = RVT_KDETH_QP_PREFIX;
1270         binfo.jkey = uctxt->jkey;
1271         /*
1272          * If more than 64 contexts are enabled the allocated credit
1273          * return will span two or three contiguous pages. Since we only
1274          * map the page containing the context's credit return address,
1275          * we need to calculate the offset in the proper page.
1276          */
1277         offset = ((u64)uctxt->sc->hw_free -
1278                   (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1279         binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1280                                                 fd->subctxt, offset);
1281         binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1282                                             fd->subctxt,
1283                                             uctxt->sc->base_addr);
1284         binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1285                                                 uctxt->ctxt,
1286                                                 fd->subctxt,
1287                                                 uctxt->sc->base_addr);
1288         binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1289                                                fd->subctxt,
1290                                                uctxt->rcvhdrq);
1291         binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1292                                                fd->subctxt,
1293                                                uctxt->egrbufs.rcvtids[0].dma);
1294         binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1295                                                   fd->subctxt, 0);
1296         /*
1297          * user regs are at
1298          * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1299          */
1300         binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1301                                              fd->subctxt, 0);
1302         offset = offset_in_page((uctxt_offset(uctxt) + fd->subctxt) *
1303                                 sizeof(*dd->events));
1304         binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1305                                                fd->subctxt,
1306                                                offset);
1307         binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1308                                                fd->subctxt,
1309                                                dd->status);
1310         if (HFI1_CAP_IS_USET(DMA_RTAIL))
1311                 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1312                                                         fd->subctxt, 0);
1313         if (uctxt->subctxt_cnt) {
1314                 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1315                                                          uctxt->ctxt,
1316                                                          fd->subctxt, 0);
1317                 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1318                                                           uctxt->ctxt,
1319                                                           fd->subctxt, 0);
1320                 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1321                                                           uctxt->ctxt,
1322                                                           fd->subctxt, 0);
1323         }
1324
1325         if (copy_to_user((void __user *)arg, &binfo, len))
1326                 return -EFAULT;
1327
1328         return 0;
1329 }
1330
1331 /**
1332  * user_exp_rcv_setup - Set up the given tid rcv list
1333  * @fd: file data of the current driver instance
1334  * @arg: ioctl argumnent for user space information
1335  * @len: length of data structure associated with ioctl command
1336  *
1337  * Wrapper to validate ioctl information before doing _rcv_setup.
1338  *
1339  */
1340 static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
1341                               u32 len)
1342 {
1343         int ret;
1344         unsigned long addr;
1345         struct hfi1_tid_info tinfo;
1346
1347         if (sizeof(tinfo) != len)
1348                 return -EINVAL;
1349
1350         if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1351                 return -EFAULT;
1352
1353         ret = hfi1_user_exp_rcv_setup(fd, &tinfo);
1354         if (!ret) {
1355                 /*
1356                  * Copy the number of tidlist entries we used
1357                  * and the length of the buffer we registered.
1358                  */
1359                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1360                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1361                                  sizeof(tinfo.tidcnt)))
1362                         ret = -EFAULT;
1363
1364                 addr = arg + offsetof(struct hfi1_tid_info, length);
1365                 if (!ret && copy_to_user((void __user *)addr, &tinfo.length,
1366                                  sizeof(tinfo.length)))
1367                         ret = -EFAULT;
1368
1369                 if (ret)
1370                         hfi1_user_exp_rcv_invalid(fd, &tinfo);
1371         }
1372
1373         return ret;
1374 }
1375
1376 /**
1377  * user_exp_rcv_clear - Clear the given tid rcv list
1378  * @fd: file data of the current driver instance
1379  * @arg: ioctl argumnent for user space information
1380  * @len: length of data structure associated with ioctl command
1381  *
1382  * The hfi1_user_exp_rcv_clear() can be called from the error path.  Because
1383  * of this, we need to use this wrapper to copy the user space information
1384  * before doing the clear.
1385  */
1386 static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
1387                               u32 len)
1388 {
1389         int ret;
1390         unsigned long addr;
1391         struct hfi1_tid_info tinfo;
1392
1393         if (sizeof(tinfo) != len)
1394                 return -EINVAL;
1395
1396         if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1397                 return -EFAULT;
1398
1399         ret = hfi1_user_exp_rcv_clear(fd, &tinfo);
1400         if (!ret) {
1401                 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1402                 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1403                                  sizeof(tinfo.tidcnt)))
1404                         return -EFAULT;
1405         }
1406
1407         return ret;
1408 }
1409
1410 /**
1411  * user_exp_rcv_invalid - Invalidate the given tid rcv list
1412  * @fd: file data of the current driver instance
1413  * @arg: ioctl argumnent for user space information
1414  * @len: length of data structure associated with ioctl command
1415  *
1416  * Wrapper to validate ioctl information before doing _rcv_invalid.
1417  *
1418  */
1419 static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
1420                                 u32 len)
1421 {
1422         int ret;
1423         unsigned long addr;
1424         struct hfi1_tid_info tinfo;
1425
1426         if (sizeof(tinfo) != len)
1427                 return -EINVAL;
1428
1429         if (!fd->invalid_tids)
1430                 return -EINVAL;
1431
1432         if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1433                 return -EFAULT;
1434
1435         ret = hfi1_user_exp_rcv_invalid(fd, &tinfo);
1436         if (ret)
1437                 return ret;
1438
1439         addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1440         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1441                          sizeof(tinfo.tidcnt)))
1442                 ret = -EFAULT;
1443
1444         return ret;
1445 }
1446
1447 static __poll_t poll_urgent(struct file *fp,
1448                                 struct poll_table_struct *pt)
1449 {
1450         struct hfi1_filedata *fd = fp->private_data;
1451         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1452         struct hfi1_devdata *dd = uctxt->dd;
1453         __poll_t pollflag;
1454
1455         poll_wait(fp, &uctxt->wait, pt);
1456
1457         spin_lock_irq(&dd->uctxt_lock);
1458         if (uctxt->urgent != uctxt->urgent_poll) {
1459                 pollflag = EPOLLIN | EPOLLRDNORM;
1460                 uctxt->urgent_poll = uctxt->urgent;
1461         } else {
1462                 pollflag = 0;
1463                 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1464         }
1465         spin_unlock_irq(&dd->uctxt_lock);
1466
1467         return pollflag;
1468 }
1469
1470 static __poll_t poll_next(struct file *fp,
1471                               struct poll_table_struct *pt)
1472 {
1473         struct hfi1_filedata *fd = fp->private_data;
1474         struct hfi1_ctxtdata *uctxt = fd->uctxt;
1475         struct hfi1_devdata *dd = uctxt->dd;
1476         __poll_t pollflag;
1477
1478         poll_wait(fp, &uctxt->wait, pt);
1479
1480         spin_lock_irq(&dd->uctxt_lock);
1481         if (hdrqempty(uctxt)) {
1482                 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1483                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt);
1484                 pollflag = 0;
1485         } else {
1486                 pollflag = EPOLLIN | EPOLLRDNORM;
1487         }
1488         spin_unlock_irq(&dd->uctxt_lock);
1489
1490         return pollflag;
1491 }
1492
1493 /*
1494  * Find all user contexts in use, and set the specified bit in their
1495  * event mask.
1496  * See also find_ctxt() for a similar use, that is specific to send buffers.
1497  */
1498 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1499 {
1500         struct hfi1_ctxtdata *uctxt;
1501         struct hfi1_devdata *dd = ppd->dd;
1502         u16 ctxt;
1503
1504         if (!dd->events)
1505                 return -EINVAL;
1506
1507         for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
1508              ctxt++) {
1509                 uctxt = hfi1_rcd_get_by_index(dd, ctxt);
1510                 if (uctxt) {
1511                         unsigned long *evs;
1512                         int i;
1513                         /*
1514                          * subctxt_cnt is 0 if not shared, so do base
1515                          * separately, first, then remaining subctxt, if any
1516                          */
1517                         evs = dd->events + uctxt_offset(uctxt);
1518                         set_bit(evtbit, evs);
1519                         for (i = 1; i < uctxt->subctxt_cnt; i++)
1520                                 set_bit(evtbit, evs + i);
1521                         hfi1_rcd_put(uctxt);
1522                 }
1523         }
1524
1525         return 0;
1526 }
1527
1528 /**
1529  * manage_rcvq - manage a context's receive queue
1530  * @uctxt: the context
1531  * @subctxt: the sub-context
1532  * @start_stop: action to carry out
1533  *
1534  * start_stop == 0 disables receive on the context, for use in queue
1535  * overflow conditions.  start_stop==1 re-enables, to be used to
1536  * re-init the software copy of the head register
1537  */
1538 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1539                        unsigned long arg)
1540 {
1541         struct hfi1_devdata *dd = uctxt->dd;
1542         unsigned int rcvctrl_op;
1543         int start_stop;
1544
1545         if (subctxt)
1546                 return 0;
1547
1548         if (get_user(start_stop, (int __user *)arg))
1549                 return -EFAULT;
1550
1551         /* atomically clear receive enable ctxt. */
1552         if (start_stop) {
1553                 /*
1554                  * On enable, force in-memory copy of the tail register to
1555                  * 0, so that protocol code doesn't have to worry about
1556                  * whether or not the chip has yet updated the in-memory
1557                  * copy or not on return from the system call. The chip
1558                  * always resets it's tail register back to 0 on a
1559                  * transition from disabled to enabled.
1560                  */
1561                 if (hfi1_rcvhdrtail_kvaddr(uctxt))
1562                         clear_rcvhdrtail(uctxt);
1563                 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1564         } else {
1565                 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1566         }
1567         hfi1_rcvctrl(dd, rcvctrl_op, uctxt);
1568         /* always; new head should be equal to new tail; see above */
1569
1570         return 0;
1571 }
1572
1573 /*
1574  * clear the event notifier events for this context.
1575  * User process then performs actions appropriate to bit having been
1576  * set, if desired, and checks again in future.
1577  */
1578 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1579                           unsigned long arg)
1580 {
1581         int i;
1582         struct hfi1_devdata *dd = uctxt->dd;
1583         unsigned long *evs;
1584         unsigned long events;
1585
1586         if (!dd->events)
1587                 return 0;
1588
1589         if (get_user(events, (unsigned long __user *)arg))
1590                 return -EFAULT;
1591
1592         evs = dd->events + uctxt_offset(uctxt) + subctxt;
1593
1594         for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1595                 if (!test_bit(i, &events))
1596                         continue;
1597                 clear_bit(i, evs);
1598         }
1599         return 0;
1600 }
1601
1602 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg)
1603 {
1604         int i;
1605         struct hfi1_pportdata *ppd = uctxt->ppd;
1606         struct hfi1_devdata *dd = uctxt->dd;
1607         u16 pkey;
1608
1609         if (!HFI1_CAP_IS_USET(PKEY_CHECK))
1610                 return -EPERM;
1611
1612         if (get_user(pkey, (u16 __user *)arg))
1613                 return -EFAULT;
1614
1615         if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY)
1616                 return -EINVAL;
1617
1618         for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1619                 if (pkey == ppd->pkeys[i])
1620                         return hfi1_set_ctxt_pkey(dd, uctxt, pkey);
1621
1622         return -ENOENT;
1623 }
1624
1625 /**
1626  * ctxt_reset - Reset the user context
1627  * @uctxt: valid user context
1628  */
1629 static int ctxt_reset(struct hfi1_ctxtdata *uctxt)
1630 {
1631         struct send_context *sc;
1632         struct hfi1_devdata *dd;
1633         int ret = 0;
1634
1635         if (!uctxt || !uctxt->dd || !uctxt->sc)
1636                 return -EINVAL;
1637
1638         /*
1639          * There is no protection here. User level has to guarantee that
1640          * no one will be writing to the send context while it is being
1641          * re-initialized.  If user level breaks that guarantee, it will
1642          * break it's own context and no one else's.
1643          */
1644         dd = uctxt->dd;
1645         sc = uctxt->sc;
1646
1647         /*
1648          * Wait until the interrupt handler has marked the context as
1649          * halted or frozen. Report error if we time out.
1650          */
1651         wait_event_interruptible_timeout(
1652                 sc->halt_wait, (sc->flags & SCF_HALTED),
1653                 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1654         if (!(sc->flags & SCF_HALTED))
1655                 return -ENOLCK;
1656
1657         /*
1658          * If the send context was halted due to a Freeze, wait until the
1659          * device has been "unfrozen" before resetting the context.
1660          */
1661         if (sc->flags & SCF_FROZEN) {
1662                 wait_event_interruptible_timeout(
1663                         dd->event_queue,
1664                         !(READ_ONCE(dd->flags) & HFI1_FROZEN),
1665                         msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1666                 if (dd->flags & HFI1_FROZEN)
1667                         return -ENOLCK;
1668
1669                 if (dd->flags & HFI1_FORCED_FREEZE)
1670                         /*
1671                          * Don't allow context reset if we are into
1672                          * forced freeze
1673                          */
1674                         return -ENODEV;
1675
1676                 sc_disable(sc);
1677                 ret = sc_enable(sc);
1678                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB, uctxt);
1679         } else {
1680                 ret = sc_restart(sc);
1681         }
1682         if (!ret)
1683                 sc_return_credits(sc);
1684
1685         return ret;
1686 }
1687
1688 static void user_remove(struct hfi1_devdata *dd)
1689 {
1690
1691         hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1692 }
1693
1694 static int user_add(struct hfi1_devdata *dd)
1695 {
1696         char name[10];
1697         int ret;
1698
1699         snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1700         ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1701                              &dd->user_cdev, &dd->user_device,
1702                              true, &dd->verbs_dev.rdi.ibdev.dev.kobj);
1703         if (ret)
1704                 user_remove(dd);
1705
1706         return ret;
1707 }
1708
1709 /*
1710  * Create per-unit files in /dev
1711  */
1712 int hfi1_device_create(struct hfi1_devdata *dd)
1713 {
1714         return user_add(dd);
1715 }
1716
1717 /*
1718  * Remove per-unit files in /dev
1719  * void, core kernel returns no errors for this stuff
1720  */
1721 void hfi1_device_remove(struct hfi1_devdata *dd)
1722 {
1723         user_remove(dd);
1724 }