1	// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2	/*
3	* Copyright(c) 2020 Cornelis Networks, Inc.
4	* Copyright(c) 2015-2020 Intel Corporation.
5	*/
6
7	#include <linux/poll.h>
8	#include <linux/cdev.h>
9	#include <linux/vmalloc.h>
10	#include <linux/io.h>
11	#include <linux/sched/mm.h>
12	#include <linux/bitmap.h>
13
14	#include <rdma/ib.h>
15
16	#include "hfi.h"
17	#include "pio.h"
18	#include "device.h"
19	#include "common.h"
20	#include "trace.h"
21	#include "mmu_rb.h"
22	#include "user_sdma.h"
23	#include "user_exp_rcv.h"
24	#include "aspm.h"
25
26	#undef pr_fmt
27	#define pr_fmt(fmt) DRIVER_NAME ": " fmt
28
29	#define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
30
31	/*
32	* File operation functions
33	*/
34	static int hfi1_file_open(struct inode *inode, struct file *fp);
35	static int hfi1_file_close(struct inode *inode, struct file *fp);
36	static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from);
37	static __poll_t hfi1_poll(struct file *fp, struct poll_table_struct *pt);
38	static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma);
39
40	static u64 kvirt_to_phys(void *addr);
41	static int assign_ctxt(struct hfi1_filedata *fd, unsigned long arg, u32 len);
42	static void init_subctxts(struct hfi1_ctxtdata *uctxt,
43	const struct hfi1_user_info *uinfo);
44	static int init_user_ctxt(struct hfi1_filedata *fd,
45	struct hfi1_ctxtdata *uctxt);
46	static void user_init(struct hfi1_ctxtdata *uctxt);
47	static int get_ctxt_info(struct hfi1_filedata *fd, unsigned long arg, u32 len);
48	static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len);
49	static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
50	u32 len);
51	static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
52	u32 len);
53	static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
54	u32 len);
55	static int setup_base_ctxt(struct hfi1_filedata *fd,
56	struct hfi1_ctxtdata *uctxt);
57	static int setup_subctxt(struct hfi1_ctxtdata *uctxt);
58
59	static int find_sub_ctxt(struct hfi1_filedata *fd,
60	const struct hfi1_user_info *uinfo);
61	static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
62	struct hfi1_user_info *uinfo,
63	struct hfi1_ctxtdata **cd);
64	static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt);
65	static __poll_t poll_urgent(struct file *fp, struct poll_table_struct *pt);
66	static __poll_t poll_next(struct file *fp, struct poll_table_struct *pt);
67	static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
68	unsigned long arg);
69	static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg);
70	static int ctxt_reset(struct hfi1_ctxtdata *uctxt);
71	static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
72	unsigned long arg);
73	static vm_fault_t vma_fault(struct vm_fault *vmf);
74	static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
75	unsigned long arg);
76
77	static const struct file_operations hfi1_file_ops = {
78	.owner = THIS_MODULE,
79	.write_iter = hfi1_write_iter,
80	.open = hfi1_file_open,
81	.release = hfi1_file_close,
82	.unlocked_ioctl = hfi1_file_ioctl,
83	.poll = hfi1_poll,
84	.mmap = hfi1_file_mmap,
85	.llseek = noop_llseek,
86	};
87
88	static const struct vm_operations_struct vm_ops = {
89	.fault = vma_fault,
90	};
91
92	/*
93	* Types of memories mapped into user processes' space
94	*/
95	enum mmap_types {
96	PIO_BUFS = 1,
97	PIO_BUFS_SOP,
98	PIO_CRED,
99	RCV_HDRQ,
100	RCV_EGRBUF,
101	UREGS,
102	EVENTS,
103	STATUS,
104	RTAIL,
105	SUBCTXT_UREGS,
106	SUBCTXT_RCV_HDRQ,
107	SUBCTXT_EGRBUF,
108	SDMA_COMP
109	};
110
111	/*
112	* Masks and offsets defining the mmap tokens
113	*/
114	#define HFI1_MMAP_OFFSET_MASK 0xfffULL
115	#define HFI1_MMAP_OFFSET_SHIFT 0
116	#define HFI1_MMAP_SUBCTXT_MASK 0xfULL
117	#define HFI1_MMAP_SUBCTXT_SHIFT 12
118	#define HFI1_MMAP_CTXT_MASK 0xffULL
119	#define HFI1_MMAP_CTXT_SHIFT 16
120	#define HFI1_MMAP_TYPE_MASK 0xfULL
121	#define HFI1_MMAP_TYPE_SHIFT 24
122	#define HFI1_MMAP_MAGIC_MASK 0xffffffffULL
123	#define HFI1_MMAP_MAGIC_SHIFT 32
124
125	#define HFI1_MMAP_MAGIC 0xdabbad00
126
127	#define HFI1_MMAP_TOKEN_SET(field, val) \
128	(((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
129	#define HFI1_MMAP_TOKEN_GET(field, token) \
130	(((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
131	#define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr) \
132	(HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
133	HFI1_MMAP_TOKEN_SET(TYPE, type) | \
134	HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
135	HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
136	HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
137
138	#define dbg(fmt, ...) \
139	pr_info(fmt, ##__VA_ARGS__)
140
141	static inline int is_valid_mmap(u64 token)
142	{
143	return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
144	}
145
146	static int hfi1_file_open(struct inode *inode, struct file *fp)
147	{
148	struct hfi1_filedata *fd;
149	struct hfi1_devdata *dd = container_of(inode->i_cdev,
150	struct hfi1_devdata,
151	user_cdev);
152
153	if (!((dd->flags & HFI1_PRESENT) && dd->kregbase1))
154	return -EINVAL;
155
156	if (!refcount_inc_not_zero(r: &dd->user_refcount))
157	return -ENXIO;
158
159	/* The real work is performed later in assign_ctxt() */
160
161	fd = kzalloc(sizeof(*fd), GFP_KERNEL);
162
163	if (!fd || init_srcu_struct(&fd->pq_srcu))
164	goto nomem;
165	spin_lock_init(&fd->pq_rcu_lock);
166	spin_lock_init(&fd->tid_lock);
167	spin_lock_init(&fd->invalid_lock);
168	fd->rec_cpu_num = -1; /* no cpu affinity by default */
169	fd->dd = dd;
170	fp->private_data = fd;
171	return 0;
172	nomem:
173	kfree(objp: fd);
174	fp->private_data = NULL;
175	if (refcount_dec_and_test(r: &dd->user_refcount))
176	complete(&dd->user_comp);
177	return -ENOMEM;
178	}
179
180	static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
181	unsigned long arg)
182	{
183	struct hfi1_filedata *fd = fp->private_data;
184	struct hfi1_ctxtdata *uctxt = fd->uctxt;
185	int ret = 0;
186	int uval = 0;
187
188	hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
189	if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
190	cmd != HFI1_IOCTL_GET_VERS &&
191	!uctxt)
192	return -EINVAL;
193
194	switch (cmd) {
195	case HFI1_IOCTL_ASSIGN_CTXT:
196	ret = assign_ctxt(fd, arg, _IOC_SIZE(cmd));
197	break;
198
199	case HFI1_IOCTL_CTXT_INFO:
200	ret = get_ctxt_info(fd, arg, _IOC_SIZE(cmd));
201	break;
202
203	case HFI1_IOCTL_USER_INFO:
204	ret = get_base_info(fd, arg, _IOC_SIZE(cmd));
205	break;
206
207	case HFI1_IOCTL_CREDIT_UPD:
208	if (uctxt)
209	sc_return_credits(sc: uctxt->sc);
210	break;
211
212	case HFI1_IOCTL_TID_UPDATE:
213	ret = user_exp_rcv_setup(fd, arg, _IOC_SIZE(cmd));
214	break;
215
216	case HFI1_IOCTL_TID_FREE:
217	ret = user_exp_rcv_clear(fd, arg, _IOC_SIZE(cmd));
218	break;
219
220	case HFI1_IOCTL_TID_INVAL_READ:
221	ret = user_exp_rcv_invalid(fd, arg, _IOC_SIZE(cmd));
222	break;
223
224	case HFI1_IOCTL_RECV_CTRL:
225	ret = manage_rcvq(uctxt, subctxt: fd->subctxt, arg);
226	break;
227
228	case HFI1_IOCTL_POLL_TYPE:
229	if (get_user(uval, (int __user *)arg))
230	return -EFAULT;
231	uctxt->poll_type = (typeof(uctxt->poll_type))uval;
232	break;
233
234	case HFI1_IOCTL_ACK_EVENT:
235	ret = user_event_ack(uctxt, subctxt: fd->subctxt, arg);
236	break;
237
238	case HFI1_IOCTL_SET_PKEY:
239	ret = set_ctxt_pkey(uctxt, arg);
240	break;
241
242	case HFI1_IOCTL_CTXT_RESET:
243	ret = ctxt_reset(uctxt);
244	break;
245
246	case HFI1_IOCTL_GET_VERS:
247	uval = HFI1_USER_SWVERSION;
248	if (put_user(uval, (int __user *)arg))
249	return -EFAULT;
250	break;
251
252	default:
253	return -EINVAL;
254	}
255
256	return ret;
257	}
258
259	static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
260	{
261	struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
262	struct hfi1_user_sdma_pkt_q *pq;
263	struct hfi1_user_sdma_comp_q *cq = fd->cq;
264	int done = 0, reqs = 0;
265	unsigned long dim = from->nr_segs;
266	int idx;
267
268	if (!HFI1_CAP_IS_KSET(SDMA))
269	return -EINVAL;
270	if (!user_backed_iter(i: from))
271	return -EINVAL;
272	idx = srcu_read_lock(ssp: &fd->pq_srcu);
273	pq = srcu_dereference(fd->pq, &fd->pq_srcu);
274	if (!cq || !pq) {
275	srcu_read_unlock(ssp: &fd->pq_srcu, idx);
276	return -EIO;
277	}
278
279	trace_hfi1_sdma_request(dd: fd->dd, ctxt: fd->uctxt->ctxt, subctxt: fd->subctxt, dim);
280
281	if (atomic_read(v: &pq->n_reqs) == pq->n_max_reqs) {
282	srcu_read_unlock(ssp: &fd->pq_srcu, idx);
283	return -ENOSPC;
284	}
285
286	while (dim) {
287	const struct iovec *iov = iter_iov(iter: from);
288	int ret;
289	unsigned long count = 0;
290
291	ret = hfi1_user_sdma_process_request(
292	fd, iovec: (struct iovec *)(iov + done),
293	dim, count: &count);
294	if (ret) {
295	reqs = ret;
296	break;
297	}
298	dim -= count;
299	done += count;
300	reqs++;
301	}
302
303	srcu_read_unlock(ssp: &fd->pq_srcu, idx);
304	return reqs;
305	}
306
307	static inline void mmap_cdbg(u16 ctxt, u8 subctxt, u8 type, u8 mapio, u8 vmf,
308	u64 memaddr, void *memvirt, dma_addr_t memdma,
309	ssize_t memlen, struct vm_area_struct *vma)
310	{
311	hfi1_cdbg(PROC,
312	"%u:%u type:%u io/vf/dma:%d/%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx",
313	ctxt, subctxt, type, mapio, vmf, !!memdma,
314	memaddr ?: (u64)memvirt, memlen,
315	vma->vm_end - vma->vm_start, vma->vm_flags);
316	}
317
318	static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
319	{
320	struct hfi1_filedata *fd = fp->private_data;
321	struct hfi1_ctxtdata *uctxt = fd->uctxt;
322	struct hfi1_devdata *dd;
323	unsigned long flags;
324	u64 token = vma->vm_pgoff << PAGE_SHIFT,
325	memaddr = 0;
326	void *memvirt = NULL;
327	dma_addr_t memdma = 0;
328	u8 subctxt, mapio = 0, vmf = 0, type;
329	ssize_t memlen = 0;
330	int ret = 0;
331	u16 ctxt;
332
333	if (!is_valid_mmap(token) || !uctxt ||
334	!(vma->vm_flags & VM_SHARED)) {
335	ret = -EINVAL;
336	goto done;
337	}
338	dd = uctxt->dd;
339	ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
340	subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
341	type = HFI1_MMAP_TOKEN_GET(TYPE, token);
342	if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
343	ret = -EINVAL;
344	goto done;
345	}
346
347	/*
348	* vm_pgoff is used as a buffer selector cookie. Always mmap from
349	* the beginning.
350	*/
351	vma->vm_pgoff = 0;
352	flags = vma->vm_flags;
353
354	switch (type) {
355	case PIO_BUFS:
356	case PIO_BUFS_SOP:
357	memaddr = ((dd->physaddr + TXE_PIO_SEND) +
358	/* chip pio base */
359	(uctxt->sc->hw_context * BIT(16))) +
360	/* 64K PIO space / ctxt */
361	(type == PIO_BUFS_SOP ?
362	(TXE_PIO_SIZE / 2) : 0); /* sop? */
363	/*
364	* Map only the amount allocated to the context, not the
365	* entire available context's PIO space.
366	*/
367	memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
368	flags &= ~VM_MAYREAD;
369	flags |= VM_DONTCOPY | VM_DONTEXPAND;
370	vma->vm_page_prot = pgprot_writecombine(prot: vma->vm_page_prot);
371	mapio = 1;
372	break;
373	case PIO_CRED: {
374	u64 cr_page_offset;
375	if (flags & VM_WRITE) {
376	ret = -EPERM;
377	goto done;
378	}
379	/*
380	* The credit return location for this context could be on the
381	* second or third page allocated for credit returns (if number
382	* of enabled contexts > 64 and 128 respectively).
383	*/
384	cr_page_offset = ((u64)uctxt->sc->hw_free -
385	(u64)dd->cr_base[uctxt->numa_id].va) &
386	PAGE_MASK;
387	memvirt = dd->cr_base[uctxt->numa_id].va + cr_page_offset;
388	memdma = dd->cr_base[uctxt->numa_id].dma + cr_page_offset;
389	memlen = PAGE_SIZE;
390	flags &= ~VM_MAYWRITE;
391	flags |= VM_DONTCOPY | VM_DONTEXPAND;
392	/*
393	* The driver has already allocated memory for credit
394	* returns and programmed it into the chip. Has that
395	* memory been flagged as non-cached?
396	*/
397	/* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
398	break;
399	}
400	case RCV_HDRQ:
401	memlen = rcvhdrq_size(rcd: uctxt);
402	memvirt = uctxt->rcvhdrq;
403	memdma = uctxt->rcvhdrq_dma;
404	break;
405	case RCV_EGRBUF: {
406	unsigned long vm_start_save;
407	unsigned long vm_end_save;
408	int i;
409	/*
410	* The RcvEgr buffer need to be handled differently
411	* as multiple non-contiguous pages need to be mapped
412	* into the user process.
413	*/
414	memlen = uctxt->egrbufs.size;
415	if ((vma->vm_end - vma->vm_start) != memlen) {
416	dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
417	(vma->vm_end - vma->vm_start), memlen);
418	ret = -EINVAL;
419	goto done;
420	}
421	if (vma->vm_flags & VM_WRITE) {
422	ret = -EPERM;
423	goto done;
424	}
425	vm_flags_clear(vma, VM_MAYWRITE);
426	/*
427	* Mmap multiple separate allocations into a single vma. From
428	* here, dma_mmap_coherent() calls dma_direct_mmap(), which
429	* requires the mmap to exactly fill the vma starting at
430	* vma_start. Adjust the vma start and end for each eager
431	* buffer segment mapped. Restore the originals when done.
432	*/
433	vm_start_save = vma->vm_start;
434	vm_end_save = vma->vm_end;
435	vma->vm_end = vma->vm_start;
436	for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
437	memlen = uctxt->egrbufs.buffers[i].len;
438	memvirt = uctxt->egrbufs.buffers[i].addr;
439	memdma = uctxt->egrbufs.buffers[i].dma;
440	vma->vm_end += memlen;
441	mmap_cdbg(ctxt, subctxt, type, mapio, vmf, memaddr,
442	memvirt, memdma, memlen, vma);
443	ret = dma_mmap_coherent(&dd->pcidev->dev, vma,
444	memvirt, memdma, memlen);
445	if (ret < 0) {
446	vma->vm_start = vm_start_save;
447	vma->vm_end = vm_end_save;
448	goto done;
449	}
450	vma->vm_start += memlen;
451	}
452	vma->vm_start = vm_start_save;
453	vma->vm_end = vm_end_save;
454	ret = 0;
455	goto done;
456	}
457	case UREGS:
458	/*
459	* Map only the page that contains this context's user
460	* registers.
461	*/
462	memaddr = (unsigned long)
463	(dd->physaddr + RXE_PER_CONTEXT_USER)
464	+ (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
465	/*
466	* TidFlow table is on the same page as the rest of the
467	* user registers.
468	*/
469	memlen = PAGE_SIZE;
470	flags |= VM_DONTCOPY | VM_DONTEXPAND;
471	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
472	mapio = 1;
473	break;
474	case EVENTS:
475	/*
476	* Use the page where this context's flags are. User level
477	* knows where it's own bitmap is within the page.
478	*/
479	memaddr = (unsigned long)
480	(dd->events + uctxt_offset(uctxt)) & PAGE_MASK;
481	memlen = PAGE_SIZE;
482	/*
483	* v3.7 removes VM_RESERVED but the effect is kept by
484	* using VM_IO.
485	*/
486	flags |= VM_IO | VM_DONTEXPAND;
487	vmf = 1;
488	break;
489	case STATUS:
490	if (flags & VM_WRITE) {
491	ret = -EPERM;
492	goto done;
493	}
494	memaddr = kvirt_to_phys(addr: (void *)dd->status);
495	memlen = PAGE_SIZE;
496	flags |= VM_IO | VM_DONTEXPAND;
497	break;
498	case RTAIL:
499	if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
500	/*
501	* If the memory allocation failed, the context alloc
502	* also would have failed, so we would never get here
503	*/
504	ret = -EINVAL;
505	goto done;
506	}
507	if ((flags & VM_WRITE) || !hfi1_rcvhdrtail_kvaddr(rcd: uctxt)) {
508	ret = -EPERM;
509	goto done;
510	}
511	memlen = PAGE_SIZE;
512	memvirt = (void *)hfi1_rcvhdrtail_kvaddr(rcd: uctxt);
513	memdma = uctxt->rcvhdrqtailaddr_dma;
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 = rcvhdrq_size(rcd: uctxt) * 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	vm_flags_reset(vma, flags);
562	mmap_cdbg(ctxt, subctxt, type, mapio, vmf, memaddr, memvirt, memdma,
563	memlen, vma);
564	if (vmf) {
565	vma->vm_pgoff = PFN_DOWN(memaddr);
566	vma->vm_ops = &vm_ops;
567	ret = 0;
568	} else if (memdma) {
569	ret = dma_mmap_coherent(&dd->pcidev->dev, vma,
570	memvirt, memdma, memlen);
571	} else if (mapio) {
572	ret = io_remap_pfn_range(vma, addr: vma->vm_start,
573	PFN_DOWN(memaddr),
574	size: memlen,
575	orig_prot: vma->vm_page_prot);
576	} else if (memvirt) {
577	ret = remap_pfn_range(vma, addr: vma->vm_start,
578	PFN_DOWN(__pa(memvirt)),
579	size: memlen,
580	pgprot: vma->vm_page_prot);
581	} else {
582	ret = remap_pfn_range(vma, addr: vma->vm_start,
583	PFN_DOWN(memaddr),
584	size: memlen,
585	pgprot: vma->vm_page_prot);
586	}
587	done:
588	return ret;
589	}
590
591	/*
592	* Local (non-chip) user memory is not mapped right away but as it is
593	* accessed by the user-level code.
594	*/
595	static vm_fault_t vma_fault(struct vm_fault *vmf)
596	{
597	struct page *page;
598
599	page = vmalloc_to_page(addr: (void *)(vmf->pgoff << PAGE_SHIFT));
600	if (!page)
601	return VM_FAULT_SIGBUS;
602
603	get_page(page);
604	vmf->page = page;
605
606	return 0;
607	}
608
609	static __poll_t hfi1_poll(struct file *fp, struct poll_table_struct *pt)
610	{
611	struct hfi1_ctxtdata *uctxt;
612	__poll_t pollflag;
613
614	uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
615	if (!uctxt)
616	pollflag = EPOLLERR;
617	else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
618	pollflag = poll_urgent(fp, pt);
619	else if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
620	pollflag = poll_next(fp, pt);
621	else /* invalid */
622	pollflag = EPOLLERR;
623
624	return pollflag;
625	}
626
627	static int hfi1_file_close(struct inode *inode, struct file *fp)
628	{
629	struct hfi1_filedata *fdata = fp->private_data;
630	struct hfi1_ctxtdata *uctxt = fdata->uctxt;
631	struct hfi1_devdata *dd = container_of(inode->i_cdev,
632	struct hfi1_devdata,
633	user_cdev);
634	unsigned long flags, *ev;
635
636	fp->private_data = NULL;
637
638	if (!uctxt)
639	goto done;
640
641	hfi1_cdbg(PROC, "closing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
642
643	flush_wc();
644	/* drain user sdma queue */
645	hfi1_user_sdma_free_queues(fd: fdata, uctxt);
646
647	/* release the cpu */
648	hfi1_put_proc_affinity(cpu: fdata->rec_cpu_num);
649
650	/* clean up rcv side */
651	hfi1_user_exp_rcv_free(fd: fdata);
652
653	/*
654	* fdata->uctxt is used in the above cleanup. It is not ready to be
655	* removed until here.
656	*/
657	fdata->uctxt = NULL;
658	hfi1_rcd_put(rcd: uctxt);
659
660	/*
661	* Clear any left over, unhandled events so the next process that
662	* gets this context doesn't get confused.
663	*/
664	ev = dd->events + uctxt_offset(uctxt) + fdata->subctxt;
665	*ev = 0;
666
667	spin_lock_irqsave(&dd->uctxt_lock, flags);
668	__clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
669	if (!bitmap_empty(src: uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
670	spin_unlock_irqrestore(lock: &dd->uctxt_lock, flags);
671	goto done;
672	}
673	spin_unlock_irqrestore(lock: &dd->uctxt_lock, flags);
674
675	/*
676	* Disable receive context and interrupt available, reset all
677	* RcvCtxtCtrl bits to default values.
678	*/
679	hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
680	HFI1_RCVCTRL_TIDFLOW_DIS |
681	HFI1_RCVCTRL_INTRAVAIL_DIS |
682	HFI1_RCVCTRL_TAILUPD_DIS |
683	HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
684	HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
685	HFI1_RCVCTRL_NO_EGR_DROP_DIS |
686	HFI1_RCVCTRL_URGENT_DIS, rcd: uctxt);
687	/* Clear the context's J_KEY */
688	hfi1_clear_ctxt_jkey(dd, ctxt: uctxt);
689	/*
690	* If a send context is allocated, reset context integrity
691	* checks to default and disable the send context.
692	*/
693	if (uctxt->sc) {
694	sc_disable(sc: uctxt->sc);
695	set_pio_integrity(uctxt->sc);
696	}
697
698	hfi1_free_ctxt_rcv_groups(rcd: uctxt);
699	hfi1_clear_ctxt_pkey(dd, ctxt: uctxt);
700
701	uctxt->event_flags = 0;
702
703	deallocate_ctxt(uctxt);
704	done:
705
706	if (refcount_dec_and_test(r: &dd->user_refcount))
707	complete(&dd->user_comp);
708
709	cleanup_srcu_struct(ssp: &fdata->pq_srcu);
710	kfree(objp: fdata);
711	return 0;
712	}
713
714	/*
715	* Convert kernel *virtual* addresses to physical addresses.
716	* This is used to vmalloc'ed addresses.
717	*/
718	static u64 kvirt_to_phys(void *addr)
719	{
720	struct page *page;
721	u64 paddr = 0;
722
723	page = vmalloc_to_page(addr);
724	if (page)
725	paddr = page_to_pfn(page) << PAGE_SHIFT;
726
727	return paddr;
728	}
729
730	/**
731	* complete_subctxt - complete sub-context info
732	* @fd: valid filedata pointer
733	*
734	* Sub-context info can only be set up after the base context
735	* has been completed. This is indicated by the clearing of the
736	* HFI1_CTXT_BASE_UINIT bit.
737	*
738	* Wait for the bit to be cleared, and then complete the subcontext
739	* initialization.
740	*
741	*/
742	static int complete_subctxt(struct hfi1_filedata *fd)
743	{
744	int ret;
745	unsigned long flags;
746
747	/*
748	* sub-context info can only be set up after the base context
749	* has been completed.
750	*/
751	ret = wait_event_interruptible(
752	fd->uctxt->wait,
753	!test_bit(HFI1_CTXT_BASE_UNINIT, &fd->uctxt->event_flags));
754
755	if (test_bit(HFI1_CTXT_BASE_FAILED, &fd->uctxt->event_flags))
756	ret = -ENOMEM;
757
758	/* Finish the sub-context init */
759	if (!ret) {
760	fd->rec_cpu_num = hfi1_get_proc_affinity(node: fd->uctxt->numa_id);
761	ret = init_user_ctxt(fd, uctxt: fd->uctxt);
762	}
763
764	if (ret) {
765	spin_lock_irqsave(&fd->dd->uctxt_lock, flags);
766	__clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
767	spin_unlock_irqrestore(lock: &fd->dd->uctxt_lock, flags);
768	hfi1_rcd_put(rcd: fd->uctxt);
769	fd->uctxt = NULL;
770	}
771
772	return ret;
773	}
774
775	static int assign_ctxt(struct hfi1_filedata *fd, unsigned long arg, u32 len)
776	{
777	int ret;
778	unsigned int swmajor;
779	struct hfi1_ctxtdata *uctxt = NULL;
780	struct hfi1_user_info uinfo;
781
782	if (fd->uctxt)
783	return -EINVAL;
784
785	if (sizeof(uinfo) != len)
786	return -EINVAL;
787
788	if (copy_from_user(to: &uinfo, from: (void __user *)arg, n: sizeof(uinfo)))
789	return -EFAULT;
790
791	swmajor = uinfo.userversion >> 16;
792	if (swmajor != HFI1_USER_SWMAJOR)
793	return -ENODEV;
794
795	if (uinfo.subctxt_cnt > HFI1_MAX_SHARED_CTXTS)
796	return -EINVAL;
797
798	/*
799	* Acquire the mutex to protect against multiple creations of what
800	* could be a shared base context.
801	*/
802	mutex_lock(&hfi1_mutex);
803	/*
804	* Get a sub context if available (fd->uctxt will be set).
805	* ret < 0 error, 0 no context, 1 sub-context found
806	*/
807	ret = find_sub_ctxt(fd, uinfo: &uinfo);
808
809	/*
810	* Allocate a base context if context sharing is not required or a
811	* sub context wasn't found.
812	*/
813	if (!ret)
814	ret = allocate_ctxt(fd, dd: fd->dd, uinfo: &uinfo, cd: &uctxt);
815
816	mutex_unlock(lock: &hfi1_mutex);
817
818	/* Depending on the context type, finish the appropriate init */
819	switch (ret) {
820	case 0:
821	ret = setup_base_ctxt(fd, uctxt);
822	if (ret)
823	deallocate_ctxt(uctxt);
824	break;
825	case 1:
826	ret = complete_subctxt(fd);
827	break;
828	default:
829	break;
830	}
831
832	return ret;
833	}
834
835	/**
836	* match_ctxt - match context
837	* @fd: valid filedata pointer
838	* @uinfo: user info to compare base context with
839	* @uctxt: context to compare uinfo to.
840	*
841	* Compare the given context with the given information to see if it
842	* can be used for a sub context.
843	*/
844	static int match_ctxt(struct hfi1_filedata *fd,
845	const struct hfi1_user_info *uinfo,
846	struct hfi1_ctxtdata *uctxt)
847	{
848	struct hfi1_devdata *dd = fd->dd;
849	unsigned long flags;
850	u16 subctxt;
851
852	/* Skip dynamically allocated kernel contexts */
853	if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
854	return 0;
855
856	/* Skip ctxt if it doesn't match the requested one */
857	if (memcmp(p: uctxt->uuid, q: uinfo->uuid, size: sizeof(uctxt->uuid)) ||
858	uctxt->jkey != generate_jkey(current_uid()) ||
859	uctxt->subctxt_id != uinfo->subctxt_id ||
860	uctxt->subctxt_cnt != uinfo->subctxt_cnt)
861	return 0;
862
863	/* Verify the sharing process matches the base */
864	if (uctxt->userversion != uinfo->userversion)
865	return -EINVAL;
866
867	/* Find an unused sub context */
868	spin_lock_irqsave(&dd->uctxt_lock, flags);
869	if (bitmap_empty(src: uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
870	/* context is being closed, do not use */
871	spin_unlock_irqrestore(lock: &dd->uctxt_lock, flags);
872	return 0;
873	}
874
875	subctxt = find_first_zero_bit(addr: uctxt->in_use_ctxts,
876	HFI1_MAX_SHARED_CTXTS);
877	if (subctxt >= uctxt->subctxt_cnt) {
878	spin_unlock_irqrestore(lock: &dd->uctxt_lock, flags);
879	return -EBUSY;
880	}
881
882	fd->subctxt = subctxt;
883	__set_bit(fd->subctxt, uctxt->in_use_ctxts);
884	spin_unlock_irqrestore(lock: &dd->uctxt_lock, flags);
885
886	fd->uctxt = uctxt;
887	hfi1_rcd_get(rcd: uctxt);
888
889	return 1;
890	}
891
892	/**
893	* find_sub_ctxt - fund sub-context
894	* @fd: valid filedata pointer
895	* @uinfo: matching info to use to find a possible context to share.
896	*
897	* The hfi1_mutex must be held when this function is called. It is
898	* necessary to ensure serialized creation of shared contexts.
899	*
900	* Return:
901	* 0 No sub-context found
902	* 1 Subcontext found and allocated
903	* errno EINVAL (incorrect parameters)
904	* EBUSY (all sub contexts in use)
905	*/
906	static int find_sub_ctxt(struct hfi1_filedata *fd,
907	const struct hfi1_user_info *uinfo)
908	{
909	struct hfi1_ctxtdata *uctxt;
910	struct hfi1_devdata *dd = fd->dd;
911	u16 i;
912	int ret;
913
914	if (!uinfo->subctxt_cnt)
915	return 0;
916
917	for (i = dd->first_dyn_alloc_ctxt; i < dd->num_rcv_contexts; i++) {
918	uctxt = hfi1_rcd_get_by_index(dd, ctxt: i);
919	if (uctxt) {
920	ret = match_ctxt(fd, uinfo, uctxt);
921	hfi1_rcd_put(rcd: uctxt);
922	/* value of != 0 will return */
923	if (ret)
924	return ret;
925	}
926	}
927
928	return 0;
929	}
930
931	static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
932	struct hfi1_user_info *uinfo,
933	struct hfi1_ctxtdata **rcd)
934	{
935	struct hfi1_ctxtdata *uctxt;
936	int ret, numa;
937
938	if (dd->flags & HFI1_FROZEN) {
939	/*
940	* Pick an error that is unique from all other errors
941	* that are returned so the user process knows that
942	* it tried to allocate while the SPC was frozen. It
943	* it should be able to retry with success in a short
944	* while.
945	*/
946	return -EIO;
947	}
948
949	if (!dd->freectxts)
950	return -EBUSY;
951
952	/*
953	* If we don't have a NUMA node requested, preference is towards
954	* device NUMA node.
955	*/
956	fd->rec_cpu_num = hfi1_get_proc_affinity(node: dd->node);
957	if (fd->rec_cpu_num != -1)
958	numa = cpu_to_node(cpu: fd->rec_cpu_num);
959	else
960	numa = numa_node_id();
961	ret = hfi1_create_ctxtdata(ppd: dd->pport, numa, rcd: &uctxt);
962	if (ret < 0) {
963	dd_dev_err(dd, "user ctxtdata allocation failed\n");
964	return ret;
965	}
966	hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
967	uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
968	uctxt->numa_id);
969
970	/*
971	* Allocate and enable a PIO send context.
972	*/
973	uctxt->sc = sc_alloc(dd, SC_USER, hdrqentsize: uctxt->rcvhdrqentsize, numa: dd->node);
974	if (!uctxt->sc) {
975	ret = -ENOMEM;
976	goto ctxdata_free;
977	}
978	hfi1_cdbg(PROC, "allocated send context %u(%u)", uctxt->sc->sw_index,
979	uctxt->sc->hw_context);
980	ret = sc_enable(sc: uctxt->sc);
981	if (ret)
982	goto ctxdata_free;
983
984	/*
985	* Setup sub context information if the user-level has requested
986	* sub contexts.
987	* This has to be done here so the rest of the sub-contexts find the
988	* proper base context.
989	* NOTE: _set_bit() can be used here because the context creation is
990	* protected by the mutex (rather than the spin_lock), and will be the
991	* very first instance of this context.
992	*/
993	__set_bit(0, uctxt->in_use_ctxts);
994	if (uinfo->subctxt_cnt)
995	init_subctxts(uctxt, uinfo);
996	uctxt->userversion = uinfo->userversion;
997	uctxt->flags = hfi1_cap_mask; /* save current flag state */
998	init_waitqueue_head(&uctxt->wait);
999	strscpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1000	memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1001	uctxt->jkey = generate_jkey(current_uid());
1002	hfi1_stats.sps_ctxts++;
1003	/*
1004	* Disable ASPM when there are open user/PSM contexts to avoid
1005	* issues with ASPM L1 exit latency
1006	*/
1007	if (dd->freectxts-- == dd->num_user_contexts)
1008	aspm_disable_all(dd);
1009
1010	*rcd = uctxt;
1011
1012	return 0;
1013
1014	ctxdata_free:
1015	hfi1_free_ctxt(rcd: uctxt);
1016	return ret;
1017	}
1018
1019	static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt)
1020	{
1021	mutex_lock(&hfi1_mutex);
1022	hfi1_stats.sps_ctxts--;
1023	if (++uctxt->dd->freectxts == uctxt->dd->num_user_contexts)
1024	aspm_enable_all(dd: uctxt->dd);
1025	mutex_unlock(lock: &hfi1_mutex);
1026
1027	hfi1_free_ctxt(rcd: uctxt);
1028	}
1029
1030	static void init_subctxts(struct hfi1_ctxtdata *uctxt,
1031	const struct hfi1_user_info *uinfo)
1032	{
1033	uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1034	uctxt->subctxt_id = uinfo->subctxt_id;
1035	set_bit(HFI1_CTXT_BASE_UNINIT, addr: &uctxt->event_flags);
1036	}
1037
1038	static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1039	{
1040	int ret = 0;
1041	u16 num_subctxts = uctxt->subctxt_cnt;
1042
1043	uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1044	if (!uctxt->subctxt_uregbase)
1045	return -ENOMEM;
1046
1047	/* We can take the size of the RcvHdr Queue from the master */
1048	uctxt->subctxt_rcvhdr_base = vmalloc_user(rcvhdrq_size(uctxt) *
1049	num_subctxts);
1050	if (!uctxt->subctxt_rcvhdr_base) {
1051	ret = -ENOMEM;
1052	goto bail_ureg;
1053	}
1054
1055	uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1056	num_subctxts);
1057	if (!uctxt->subctxt_rcvegrbuf) {
1058	ret = -ENOMEM;
1059	goto bail_rhdr;
1060	}
1061
1062	return 0;
1063
1064	bail_rhdr:
1065	vfree(addr: uctxt->subctxt_rcvhdr_base);
1066	uctxt->subctxt_rcvhdr_base = NULL;
1067	bail_ureg:
1068	vfree(addr: uctxt->subctxt_uregbase);
1069	uctxt->subctxt_uregbase = NULL;
1070
1071	return ret;
1072	}
1073
1074	static void user_init(struct hfi1_ctxtdata *uctxt)
1075	{
1076	unsigned int rcvctrl_ops = 0;
1077
1078	/* initialize poll variables... */
1079	uctxt->urgent = 0;
1080	uctxt->urgent_poll = 0;
1081
1082	/*
1083	* Now enable the ctxt for receive.
1084	* For chips that are set to DMA the tail register to memory
1085	* when they change (and when the update bit transitions from
1086	* 0 to 1. So for those chips, we turn it off and then back on.
1087	* This will (very briefly) affect any other open ctxts, but the
1088	* duration is very short, and therefore isn't an issue. We
1089	* explicitly set the in-memory tail copy to 0 beforehand, so we
1090	* don't have to wait to be sure the DMA update has happened
1091	* (chip resets head/tail to 0 on transition to enable).
1092	*/
1093	if (hfi1_rcvhdrtail_kvaddr(rcd: uctxt))
1094	clear_rcvhdrtail(rcd: uctxt);
1095
1096	/* Setup J_KEY before enabling the context */
1097	hfi1_set_ctxt_jkey(dd: uctxt->dd, rcd: uctxt, jkey: uctxt->jkey);
1098
1099	rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1100	rcvctrl_ops |= HFI1_RCVCTRL_URGENT_ENB;
1101	if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1102	rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1103	/*
1104	* Ignore the bit in the flags for now until proper
1105	* support for multiple packet per rcv array entry is
1106	* added.
1107	*/
1108	if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1109	rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1110	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1111	rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1112	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1113	rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1114	/*
1115	* The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1116	* We can't rely on the correct value to be set from prior
1117	* uses of the chip or ctxt. Therefore, add the rcvctrl op
1118	* for both cases.
1119	*/
1120	if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1121	rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1122	else
1123	rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1124	hfi1_rcvctrl(dd: uctxt->dd, op: rcvctrl_ops, rcd: uctxt);
1125	}
1126
1127	static int get_ctxt_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1128	{
1129	struct hfi1_ctxt_info cinfo;
1130	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1131
1132	if (sizeof(cinfo) != len)
1133	return -EINVAL;
1134
1135	memset(&cinfo, 0, sizeof(cinfo));
1136	cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1137	HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1138	HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1139	HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1140	/* adjust flag if this fd is not able to cache */
1141	if (!fd->use_mn)
1142	cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1143
1144	cinfo.num_active = hfi1_count_active_units();
1145	cinfo.unit = uctxt->dd->unit;
1146	cinfo.ctxt = uctxt->ctxt;
1147	cinfo.subctxt = fd->subctxt;
1148	cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1149	uctxt->dd->rcv_entries.group_size) +
1150	uctxt->expected_count;
1151	cinfo.credits = uctxt->sc->credits;
1152	cinfo.numa_node = uctxt->numa_id;
1153	cinfo.rec_cpu = fd->rec_cpu_num;
1154	cinfo.send_ctxt = uctxt->sc->hw_context;
1155
1156	cinfo.egrtids = uctxt->egrbufs.alloced;
1157	cinfo.rcvhdrq_cnt = get_hdrq_cnt(rcd: uctxt);
1158	cinfo.rcvhdrq_entsize = get_hdrqentsize(rcd: uctxt) << 2;
1159	cinfo.sdma_ring_size = fd->cq->nentries;
1160	cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1161
1162	trace_hfi1_ctxt_info(dd: uctxt->dd, ctxt: uctxt->ctxt, subctxt: fd->subctxt, cinfo: &cinfo);
1163	if (copy_to_user(to: (void __user *)arg, from: &cinfo, n: len))
1164	return -EFAULT;
1165
1166	return 0;
1167	}
1168
1169	static int init_user_ctxt(struct hfi1_filedata *fd,
1170	struct hfi1_ctxtdata *uctxt)
1171	{
1172	int ret;
1173
1174	ret = hfi1_user_sdma_alloc_queues(uctxt, fd);
1175	if (ret)
1176	return ret;
1177
1178	ret = hfi1_user_exp_rcv_init(fd, uctxt);
1179	if (ret)
1180	hfi1_user_sdma_free_queues(fd, uctxt);
1181
1182	return ret;
1183	}
1184
1185	static int setup_base_ctxt(struct hfi1_filedata *fd,
1186	struct hfi1_ctxtdata *uctxt)
1187	{
1188	struct hfi1_devdata *dd = uctxt->dd;
1189	int ret = 0;
1190
1191	hfi1_init_ctxt(sc: uctxt->sc);
1192
1193	/* Now allocate the RcvHdr queue and eager buffers. */
1194	ret = hfi1_create_rcvhdrq(dd, rcd: uctxt);
1195	if (ret)
1196	goto done;
1197
1198	ret = hfi1_setup_eagerbufs(rcd: uctxt);
1199	if (ret)
1200	goto done;
1201
1202	/* If sub-contexts are enabled, do the appropriate setup */
1203	if (uctxt->subctxt_cnt)
1204	ret = setup_subctxt(uctxt);
1205	if (ret)
1206	goto done;
1207
1208	ret = hfi1_alloc_ctxt_rcv_groups(rcd: uctxt);
1209	if (ret)
1210	goto done;
1211
1212	ret = init_user_ctxt(fd, uctxt);
1213	if (ret) {
1214	hfi1_free_ctxt_rcv_groups(rcd: uctxt);
1215	goto done;
1216	}
1217
1218	user_init(uctxt);
1219
1220	/* Now that the context is set up, the fd can get a reference. */
1221	fd->uctxt = uctxt;
1222	hfi1_rcd_get(rcd: uctxt);
1223
1224	done:
1225	if (uctxt->subctxt_cnt) {
1226	/*
1227	* On error, set the failed bit so sub-contexts will clean up
1228	* correctly.
1229	*/
1230	if (ret)
1231	set_bit(HFI1_CTXT_BASE_FAILED, addr: &uctxt->event_flags);
1232
1233	/*
1234	* Base context is done (successfully or not), notify anybody
1235	* using a sub-context that is waiting for this completion.
1236	*/
1237	clear_bit(HFI1_CTXT_BASE_UNINIT, addr: &uctxt->event_flags);
1238	wake_up(&uctxt->wait);
1239	}
1240
1241	return ret;
1242	}
1243
1244	static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1245	{
1246	struct hfi1_base_info binfo;
1247	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1248	struct hfi1_devdata *dd = uctxt->dd;
1249	unsigned offset;
1250
1251	trace_hfi1_uctxtdata(dd: uctxt->dd, uctxt, subctxt: fd->subctxt);
1252
1253	if (sizeof(binfo) != len)
1254	return -EINVAL;
1255
1256	memset(&binfo, 0, sizeof(binfo));
1257	binfo.hw_version = dd->revision;
1258	binfo.sw_version = HFI1_USER_SWVERSION;
1259	binfo.bthqp = RVT_KDETH_QP_PREFIX;
1260	binfo.jkey = uctxt->jkey;
1261	/*
1262	* If more than 64 contexts are enabled the allocated credit
1263	* return will span two or three contiguous pages. Since we only
1264	* map the page containing the context's credit return address,
1265	* we need to calculate the offset in the proper page.
1266	*/
1267	offset = ((u64)uctxt->sc->hw_free -
1268	(u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1269	binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1270	fd->subctxt, offset);
1271	binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1272	fd->subctxt,
1273	uctxt->sc->base_addr);
1274	binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1275	uctxt->ctxt,
1276	fd->subctxt,
1277	uctxt->sc->base_addr);
1278	binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1279	fd->subctxt,
1280	uctxt->rcvhdrq);
1281	binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1282	fd->subctxt,
1283	uctxt->egrbufs.rcvtids[0].dma);
1284	binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1285	fd->subctxt, 0);
1286	/*
1287	* user regs are at
1288	* (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1289	*/
1290	binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1291	fd->subctxt, 0);
1292	offset = offset_in_page((uctxt_offset(uctxt) + fd->subctxt) *
1293	sizeof(*dd->events));
1294	binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1295	fd->subctxt,
1296	offset);
1297	binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1298	fd->subctxt,
1299	dd->status);
1300	if (HFI1_CAP_IS_USET(DMA_RTAIL))
1301	binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1302	fd->subctxt, 0);
1303	if (uctxt->subctxt_cnt) {
1304	binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1305	uctxt->ctxt,
1306	fd->subctxt, 0);
1307	binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1308	uctxt->ctxt,
1309	fd->subctxt, 0);
1310	binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1311	uctxt->ctxt,
1312	fd->subctxt, 0);
1313	}
1314
1315	if (copy_to_user(to: (void __user *)arg, from: &binfo, n: len))
1316	return -EFAULT;
1317
1318	return 0;
1319	}
1320
1321	/**
1322	* user_exp_rcv_setup - Set up the given tid rcv list
1323	* @fd: file data of the current driver instance
1324	* @arg: ioctl argumnent for user space information
1325	* @len: length of data structure associated with ioctl command
1326	*
1327	* Wrapper to validate ioctl information before doing _rcv_setup.
1328	*
1329	*/
1330	static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
1331	u32 len)
1332	{
1333	int ret;
1334	unsigned long addr;
1335	struct hfi1_tid_info tinfo;
1336
1337	if (sizeof(tinfo) != len)
1338	return -EINVAL;
1339
1340	if (copy_from_user(to: &tinfo, from: (void __user *)arg, n: (sizeof(tinfo))))
1341	return -EFAULT;
1342
1343	ret = hfi1_user_exp_rcv_setup(fd, tinfo: &tinfo);
1344	if (!ret) {
1345	/*
1346	* Copy the number of tidlist entries we used
1347	* and the length of the buffer we registered.
1348	*/
1349	addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1350	if (copy_to_user(to: (void __user *)addr, from: &tinfo.tidcnt,
1351	n: sizeof(tinfo.tidcnt)))
1352	ret = -EFAULT;
1353
1354	addr = arg + offsetof(struct hfi1_tid_info, length);
1355	if (!ret && copy_to_user(to: (void __user *)addr, from: &tinfo.length,
1356	n: sizeof(tinfo.length)))
1357	ret = -EFAULT;
1358
1359	if (ret)
1360	hfi1_user_exp_rcv_invalid(fd, tinfo: &tinfo);
1361	}
1362
1363	return ret;
1364	}
1365
1366	/**
1367	* user_exp_rcv_clear - Clear the given tid rcv list
1368	* @fd: file data of the current driver instance
1369	* @arg: ioctl argumnent for user space information
1370	* @len: length of data structure associated with ioctl command
1371	*
1372	* The hfi1_user_exp_rcv_clear() can be called from the error path. Because
1373	* of this, we need to use this wrapper to copy the user space information
1374	* before doing the clear.
1375	*/
1376	static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
1377	u32 len)
1378	{
1379	int ret;
1380	unsigned long addr;
1381	struct hfi1_tid_info tinfo;
1382
1383	if (sizeof(tinfo) != len)
1384	return -EINVAL;
1385
1386	if (copy_from_user(to: &tinfo, from: (void __user *)arg, n: (sizeof(tinfo))))
1387	return -EFAULT;
1388
1389	ret = hfi1_user_exp_rcv_clear(fd, tinfo: &tinfo);
1390	if (!ret) {
1391	addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1392	if (copy_to_user(to: (void __user *)addr, from: &tinfo.tidcnt,
1393	n: sizeof(tinfo.tidcnt)))
1394	return -EFAULT;
1395	}
1396
1397	return ret;
1398	}
1399
1400	/**
1401	* user_exp_rcv_invalid - Invalidate the given tid rcv list
1402	* @fd: file data of the current driver instance
1403	* @arg: ioctl argumnent for user space information
1404	* @len: length of data structure associated with ioctl command
1405	*
1406	* Wrapper to validate ioctl information before doing _rcv_invalid.
1407	*
1408	*/
1409	static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
1410	u32 len)
1411	{
1412	int ret;
1413	unsigned long addr;
1414	struct hfi1_tid_info tinfo;
1415
1416	if (sizeof(tinfo) != len)
1417	return -EINVAL;
1418
1419	if (!fd->invalid_tids)
1420	return -EINVAL;
1421
1422	if (copy_from_user(to: &tinfo, from: (void __user *)arg, n: (sizeof(tinfo))))
1423	return -EFAULT;
1424
1425	ret = hfi1_user_exp_rcv_invalid(fd, tinfo: &tinfo);
1426	if (ret)
1427	return ret;
1428
1429	addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1430	if (copy_to_user(to: (void __user *)addr, from: &tinfo.tidcnt,
1431	n: sizeof(tinfo.tidcnt)))
1432	ret = -EFAULT;
1433
1434	return ret;
1435	}
1436
1437	static __poll_t poll_urgent(struct file *fp,
1438	struct poll_table_struct *pt)
1439	{
1440	struct hfi1_filedata *fd = fp->private_data;
1441	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1442	struct hfi1_devdata *dd = uctxt->dd;
1443	__poll_t pollflag;
1444
1445	poll_wait(filp: fp, wait_address: &uctxt->wait, p: pt);
1446
1447	spin_lock_irq(lock: &dd->uctxt_lock);
1448	if (uctxt->urgent != uctxt->urgent_poll) {
1449	pollflag = EPOLLIN | EPOLLRDNORM;
1450	uctxt->urgent_poll = uctxt->urgent;
1451	} else {
1452	pollflag = 0;
1453	set_bit(HFI1_CTXT_WAITING_URG, addr: &uctxt->event_flags);
1454	}
1455	spin_unlock_irq(lock: &dd->uctxt_lock);
1456
1457	return pollflag;
1458	}
1459
1460	static __poll_t poll_next(struct file *fp,
1461	struct poll_table_struct *pt)
1462	{
1463	struct hfi1_filedata *fd = fp->private_data;
1464	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1465	struct hfi1_devdata *dd = uctxt->dd;
1466	__poll_t pollflag;
1467
1468	poll_wait(filp: fp, wait_address: &uctxt->wait, p: pt);
1469
1470	spin_lock_irq(lock: &dd->uctxt_lock);
1471	if (hdrqempty(rcd: uctxt)) {
1472	set_bit(HFI1_CTXT_WAITING_RCV, addr: &uctxt->event_flags);
1473	hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, rcd: uctxt);
1474	pollflag = 0;
1475	} else {
1476	pollflag = EPOLLIN | EPOLLRDNORM;
1477	}
1478	spin_unlock_irq(lock: &dd->uctxt_lock);
1479
1480	return pollflag;
1481	}
1482
1483	/*
1484	* Find all user contexts in use, and set the specified bit in their
1485	* event mask.
1486	* See also find_ctxt() for a similar use, that is specific to send buffers.
1487	*/
1488	int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1489	{
1490	struct hfi1_ctxtdata *uctxt;
1491	struct hfi1_devdata *dd = ppd->dd;
1492	u16 ctxt;
1493
1494	if (!dd->events)
1495	return -EINVAL;
1496
1497	for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
1498	ctxt++) {
1499	uctxt = hfi1_rcd_get_by_index(dd, ctxt);
1500	if (uctxt) {
1501	unsigned long *evs;
1502	int i;
1503	/*
1504	* subctxt_cnt is 0 if not shared, so do base
1505	* separately, first, then remaining subctxt, if any
1506	*/
1507	evs = dd->events + uctxt_offset(uctxt);
1508	set_bit(nr: evtbit, addr: evs);
1509	for (i = 1; i < uctxt->subctxt_cnt; i++)
1510	set_bit(nr: evtbit, addr: evs + i);
1511	hfi1_rcd_put(rcd: uctxt);
1512	}
1513	}
1514
1515	return 0;
1516	}
1517
1518	/**
1519	* manage_rcvq - manage a context's receive queue
1520	* @uctxt: the context
1521	* @subctxt: the sub-context
1522	* @arg: start/stop action to carry out
1523	*
1524	* start_stop == 0 disables receive on the context, for use in queue
1525	* overflow conditions. start_stop==1 re-enables, to be used to
1526	* re-init the software copy of the head register
1527	*/
1528	static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1529	unsigned long arg)
1530	{
1531	struct hfi1_devdata *dd = uctxt->dd;
1532	unsigned int rcvctrl_op;
1533	int start_stop;
1534
1535	if (subctxt)
1536	return 0;
1537
1538	if (get_user(start_stop, (int __user *)arg))
1539	return -EFAULT;
1540
1541	/* atomically clear receive enable ctxt. */
1542	if (start_stop) {
1543	/*
1544	* On enable, force in-memory copy of the tail register to
1545	* 0, so that protocol code doesn't have to worry about
1546	* whether or not the chip has yet updated the in-memory
1547	* copy or not on return from the system call. The chip
1548	* always resets it's tail register back to 0 on a
1549	* transition from disabled to enabled.
1550	*/
1551	if (hfi1_rcvhdrtail_kvaddr(rcd: uctxt))
1552	clear_rcvhdrtail(rcd: uctxt);
1553	rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1554	} else {
1555	rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1556	}
1557	hfi1_rcvctrl(dd, op: rcvctrl_op, rcd: uctxt);
1558	/* always; new head should be equal to new tail; see above */
1559
1560	return 0;
1561	}
1562
1563	/*
1564	* clear the event notifier events for this context.
1565	* User process then performs actions appropriate to bit having been
1566	* set, if desired, and checks again in future.
1567	*/
1568	static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1569	unsigned long arg)
1570	{
1571	int i;
1572	struct hfi1_devdata *dd = uctxt->dd;
1573	unsigned long *evs;
1574	unsigned long events;
1575
1576	if (!dd->events)
1577	return 0;
1578
1579	if (get_user(events, (unsigned long __user *)arg))
1580	return -EFAULT;
1581
1582	evs = dd->events + uctxt_offset(uctxt) + subctxt;
1583
1584	for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1585	if (!test_bit(i, &events))
1586	continue;
1587	clear_bit(nr: i, addr: evs);
1588	}
1589	return 0;
1590	}
1591
1592	static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg)
1593	{
1594	int i;
1595	struct hfi1_pportdata *ppd = uctxt->ppd;
1596	struct hfi1_devdata *dd = uctxt->dd;
1597	u16 pkey;
1598
1599	if (!HFI1_CAP_IS_USET(PKEY_CHECK))
1600	return -EPERM;
1601
1602	if (get_user(pkey, (u16 __user *)arg))
1603	return -EFAULT;
1604
1605	if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY)
1606	return -EINVAL;
1607
1608	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1609	if (pkey == ppd->pkeys[i])
1610	return hfi1_set_ctxt_pkey(dd, ctxt: uctxt, pkey);
1611
1612	return -ENOENT;
1613	}
1614
1615	/**
1616	* ctxt_reset - Reset the user context
1617	* @uctxt: valid user context
1618	*/
1619	static int ctxt_reset(struct hfi1_ctxtdata *uctxt)
1620	{
1621	struct send_context *sc;
1622	struct hfi1_devdata *dd;
1623	int ret = 0;
1624
1625	if (!uctxt || !uctxt->dd || !uctxt->sc)
1626	return -EINVAL;
1627
1628	/*
1629	* There is no protection here. User level has to guarantee that
1630	* no one will be writing to the send context while it is being
1631	* re-initialized. If user level breaks that guarantee, it will
1632	* break it's own context and no one else's.
1633	*/
1634	dd = uctxt->dd;
1635	sc = uctxt->sc;
1636
1637	/*
1638	* Wait until the interrupt handler has marked the context as
1639	* halted or frozen. Report error if we time out.
1640	*/
1641	wait_event_interruptible_timeout(
1642	sc->halt_wait, (sc->flags & SCF_HALTED),
1643	msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1644	if (!(sc->flags & SCF_HALTED))
1645	return -ENOLCK;
1646
1647	/*
1648	* If the send context was halted due to a Freeze, wait until the
1649	* device has been "unfrozen" before resetting the context.
1650	*/
1651	if (sc->flags & SCF_FROZEN) {
1652	wait_event_interruptible_timeout(
1653	dd->event_queue,
1654	!(READ_ONCE(dd->flags) & HFI1_FROZEN),
1655	msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1656	if (dd->flags & HFI1_FROZEN)
1657	return -ENOLCK;
1658
1659	if (dd->flags & HFI1_FORCED_FREEZE)
1660	/*
1661	* Don't allow context reset if we are into
1662	* forced freeze
1663	*/
1664	return -ENODEV;
1665
1666	sc_disable(sc);
1667	ret = sc_enable(sc);
1668	hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB, rcd: uctxt);
1669	} else {
1670	ret = sc_restart(sc);
1671	}
1672	if (!ret)
1673	sc_return_credits(sc);
1674
1675	return ret;
1676	}
1677
1678	static void user_remove(struct hfi1_devdata *dd)
1679	{
1680
1681	hfi1_cdev_cleanup(cdev: &dd->user_cdev, devp: &dd->user_device);
1682	}
1683
1684	static int user_add(struct hfi1_devdata *dd)
1685	{
1686	char name[10];
1687	int ret;
1688
1689	snprintf(buf: name, size: sizeof(name), fmt: "%s_%d", class_name(), dd->unit);
1690	ret = hfi1_cdev_init(minor: dd->unit, name, fops: &hfi1_file_ops,
1691	cdev: &dd->user_cdev, devp: &dd->user_device,
1692	user_accessible: true, parent: &dd->verbs_dev.rdi.ibdev.dev.kobj);
1693	if (ret)
1694	user_remove(dd);
1695
1696	return ret;
1697	}
1698
1699	/*
1700	* Create per-unit files in /dev
1701	*/
1702	int hfi1_device_create(struct hfi1_devdata *dd)
1703	{
1704	return user_add(dd);
1705	}
1706
1707	/*
1708	* Remove per-unit files in /dev
1709	* void, core kernel returns no errors for this stuff
1710	*/
1711	void hfi1_device_remove(struct hfi1_devdata *dd)
1712	{
1713	user_remove(dd);
1714	}
1715