2 * Copyright(c) 2015, 2016 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
47 #include <linux/poll.h>
48 #include <linux/cdev.h>
49 #include <linux/vmalloc.h>
59 #include "user_sdma.h"
60 #include "user_exp_rcv.h"
66 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
68 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
71 * File operation functions
73 static int hfi1_file_open(struct inode
*, struct file
*);
74 static int hfi1_file_close(struct inode
*, struct file
*);
75 static ssize_t
hfi1_file_write(struct file
*, const char __user
*,
77 static ssize_t
hfi1_write_iter(struct kiocb
*, struct iov_iter
*);
78 static unsigned int hfi1_poll(struct file
*, struct poll_table_struct
*);
79 static int hfi1_file_mmap(struct file
*, struct vm_area_struct
*);
81 static u64
kvirt_to_phys(void *);
82 static int assign_ctxt(struct file
*, struct hfi1_user_info
*);
83 static int init_subctxts(struct hfi1_ctxtdata
*, const struct hfi1_user_info
*);
84 static int user_init(struct file
*);
85 static int get_ctxt_info(struct file
*, void __user
*, __u32
);
86 static int get_base_info(struct file
*, void __user
*, __u32
);
87 static int setup_ctxt(struct file
*);
88 static int setup_subctxt(struct hfi1_ctxtdata
*);
89 static int get_user_context(struct file
*, struct hfi1_user_info
*,
91 static int find_shared_ctxt(struct file
*, const struct hfi1_user_info
*);
92 static int allocate_ctxt(struct file
*, struct hfi1_devdata
*,
93 struct hfi1_user_info
*);
94 static unsigned int poll_urgent(struct file
*, struct poll_table_struct
*);
95 static unsigned int poll_next(struct file
*, struct poll_table_struct
*);
96 static int user_event_ack(struct hfi1_ctxtdata
*, int, unsigned long);
97 static int set_ctxt_pkey(struct hfi1_ctxtdata
*, unsigned, u16
);
98 static int manage_rcvq(struct hfi1_ctxtdata
*, unsigned, int);
99 static int vma_fault(struct vm_area_struct
*, struct vm_fault
*);
101 static const struct file_operations hfi1_file_ops
= {
102 .owner
= THIS_MODULE
,
103 .write
= hfi1_file_write
,
104 .write_iter
= hfi1_write_iter
,
105 .open
= hfi1_file_open
,
106 .release
= hfi1_file_close
,
108 .mmap
= hfi1_file_mmap
,
109 .llseek
= noop_llseek
,
112 static struct vm_operations_struct vm_ops
= {
117 * Types of memories mapped into user processes' space
136 * Masks and offsets defining the mmap tokens
138 #define HFI1_MMAP_OFFSET_MASK 0xfffULL
139 #define HFI1_MMAP_OFFSET_SHIFT 0
140 #define HFI1_MMAP_SUBCTXT_MASK 0xfULL
141 #define HFI1_MMAP_SUBCTXT_SHIFT 12
142 #define HFI1_MMAP_CTXT_MASK 0xffULL
143 #define HFI1_MMAP_CTXT_SHIFT 16
144 #define HFI1_MMAP_TYPE_MASK 0xfULL
145 #define HFI1_MMAP_TYPE_SHIFT 24
146 #define HFI1_MMAP_MAGIC_MASK 0xffffffffULL
147 #define HFI1_MMAP_MAGIC_SHIFT 32
149 #define HFI1_MMAP_MAGIC 0xdabbad00
151 #define HFI1_MMAP_TOKEN_SET(field, val) \
152 (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
153 #define HFI1_MMAP_TOKEN_GET(field, token) \
154 (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
155 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr) \
156 (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
157 HFI1_MMAP_TOKEN_SET(TYPE, type) | \
158 HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
159 HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
160 HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
162 #define dbg(fmt, ...) \
163 pr_info(fmt, ##__VA_ARGS__)
165 static inline int is_valid_mmap(u64 token
)
167 return (HFI1_MMAP_TOKEN_GET(MAGIC
, token
) == HFI1_MMAP_MAGIC
);
170 static int hfi1_file_open(struct inode
*inode
, struct file
*fp
)
172 /* The real work is performed later in assign_ctxt() */
173 fp
->private_data
= kzalloc(sizeof(struct hfi1_filedata
), GFP_KERNEL
);
174 if (fp
->private_data
) /* no cpu affinity by default */
175 ((struct hfi1_filedata
*)fp
->private_data
)->rec_cpu_num
= -1;
176 return fp
->private_data
? 0 : -ENOMEM
;
179 static ssize_t
hfi1_file_write(struct file
*fp
, const char __user
*data
,
180 size_t count
, loff_t
*offset
)
182 const struct hfi1_cmd __user
*ucmd
;
183 struct hfi1_filedata
*fd
= fp
->private_data
;
184 struct hfi1_ctxtdata
*uctxt
= fd
->uctxt
;
186 struct hfi1_user_info uinfo
;
187 struct hfi1_tid_info tinfo
;
189 ssize_t consumed
= 0, copy
= 0, ret
= 0;
192 int uctxt_required
= 1;
193 int must_be_root
= 0;
195 /* FIXME: This interface cannot continue out of staging */
196 if (WARN_ON_ONCE(!ib_safe_file_access(fp
)))
199 if (count
< sizeof(cmd
)) {
204 ucmd
= (const struct hfi1_cmd __user
*)data
;
205 if (copy_from_user(&cmd
, ucmd
, sizeof(cmd
))) {
210 consumed
= sizeof(cmd
);
213 case HFI1_CMD_ASSIGN_CTXT
:
214 uctxt_required
= 0; /* assigned user context not required */
215 copy
= sizeof(uinfo
);
218 case HFI1_CMD_SDMA_STATUS_UPD
:
219 case HFI1_CMD_CREDIT_UPD
:
222 case HFI1_CMD_TID_UPDATE
:
223 case HFI1_CMD_TID_FREE
:
224 case HFI1_CMD_TID_INVAL_READ
:
225 copy
= sizeof(tinfo
);
228 case HFI1_CMD_USER_INFO
:
229 case HFI1_CMD_RECV_CTRL
:
230 case HFI1_CMD_POLL_TYPE
:
231 case HFI1_CMD_ACK_EVENT
:
232 case HFI1_CMD_CTXT_INFO
:
233 case HFI1_CMD_SET_PKEY
:
234 case HFI1_CMD_CTXT_RESET
:
238 case HFI1_CMD_EP_INFO
:
239 case HFI1_CMD_EP_ERASE_CHIP
:
240 case HFI1_CMD_EP_ERASE_RANGE
:
241 case HFI1_CMD_EP_READ_RANGE
:
242 case HFI1_CMD_EP_WRITE_RANGE
:
243 uctxt_required
= 0; /* assigned user context not required */
244 must_be_root
= 1; /* validate user */
252 /* If the command comes with user data, copy it. */
254 if (copy_from_user(dest
, (void __user
*)cmd
.addr
, copy
)) {
262 * Make sure there is a uctxt when needed.
264 if (uctxt_required
&& !uctxt
) {
269 /* only root can do these operations */
270 if (must_be_root
&& !capable(CAP_SYS_ADMIN
)) {
276 case HFI1_CMD_ASSIGN_CTXT
:
277 ret
= assign_ctxt(fp
, &uinfo
);
280 ret
= setup_ctxt(fp
);
285 case HFI1_CMD_CTXT_INFO
:
286 ret
= get_ctxt_info(fp
, (void __user
*)(unsigned long)
289 case HFI1_CMD_USER_INFO
:
290 ret
= get_base_info(fp
, (void __user
*)(unsigned long)
293 case HFI1_CMD_SDMA_STATUS_UPD
:
295 case HFI1_CMD_CREDIT_UPD
:
296 if (uctxt
&& uctxt
->sc
)
297 sc_return_credits(uctxt
->sc
);
299 case HFI1_CMD_TID_UPDATE
:
300 ret
= hfi1_user_exp_rcv_setup(fp
, &tinfo
);
303 * Copy the number of tidlist entries we used
304 * and the length of the buffer we registered.
305 * These fields are adjacent in the structure so
306 * we can copy them at the same time.
308 addr
= (unsigned long)cmd
.addr
+
309 offsetof(struct hfi1_tid_info
, tidcnt
);
310 if (copy_to_user((void __user
*)addr
, &tinfo
.tidcnt
,
311 sizeof(tinfo
.tidcnt
) +
312 sizeof(tinfo
.length
)))
316 case HFI1_CMD_TID_INVAL_READ
:
317 ret
= hfi1_user_exp_rcv_invalid(fp
, &tinfo
);
320 addr
= (unsigned long)cmd
.addr
+
321 offsetof(struct hfi1_tid_info
, tidcnt
);
322 if (copy_to_user((void __user
*)addr
, &tinfo
.tidcnt
,
323 sizeof(tinfo
.tidcnt
)))
326 case HFI1_CMD_TID_FREE
:
327 ret
= hfi1_user_exp_rcv_clear(fp
, &tinfo
);
330 addr
= (unsigned long)cmd
.addr
+
331 offsetof(struct hfi1_tid_info
, tidcnt
);
332 if (copy_to_user((void __user
*)addr
, &tinfo
.tidcnt
,
333 sizeof(tinfo
.tidcnt
)))
336 case HFI1_CMD_RECV_CTRL
:
337 ret
= manage_rcvq(uctxt
, fd
->subctxt
, (int)user_val
);
339 case HFI1_CMD_POLL_TYPE
:
340 uctxt
->poll_type
= (typeof(uctxt
->poll_type
))user_val
;
342 case HFI1_CMD_ACK_EVENT
:
343 ret
= user_event_ack(uctxt
, fd
->subctxt
, user_val
);
345 case HFI1_CMD_SET_PKEY
:
346 if (HFI1_CAP_IS_USET(PKEY_CHECK
))
347 ret
= set_ctxt_pkey(uctxt
, fd
->subctxt
, user_val
);
351 case HFI1_CMD_CTXT_RESET
: {
352 struct send_context
*sc
;
353 struct hfi1_devdata
*dd
;
355 if (!uctxt
|| !uctxt
->dd
|| !uctxt
->sc
) {
360 * There is no protection here. User level has to
361 * guarantee that no one will be writing to the send
362 * context while it is being re-initialized.
363 * If user level breaks that guarantee, it will break
364 * it's own context and no one else's.
369 * Wait until the interrupt handler has marked the
370 * context as halted or frozen. Report error if we time
373 wait_event_interruptible_timeout(
374 sc
->halt_wait
, (sc
->flags
& SCF_HALTED
),
375 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT
));
376 if (!(sc
->flags
& SCF_HALTED
)) {
381 * If the send context was halted due to a Freeze,
382 * wait until the device has been "unfrozen" before
383 * resetting the context.
385 if (sc
->flags
& SCF_FROZEN
) {
386 wait_event_interruptible_timeout(
388 !(ACCESS_ONCE(dd
->flags
) & HFI1_FROZEN
),
389 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT
));
390 if (dd
->flags
& HFI1_FROZEN
) {
394 if (dd
->flags
& HFI1_FORCED_FREEZE
) {
396 * Don't allow context reset if we are into
404 hfi1_rcvctrl(dd
, HFI1_RCVCTRL_CTXT_ENB
,
407 ret
= sc_restart(sc
);
410 sc_return_credits(sc
);
413 case HFI1_CMD_EP_INFO
:
414 case HFI1_CMD_EP_ERASE_CHIP
:
415 case HFI1_CMD_EP_ERASE_RANGE
:
416 case HFI1_CMD_EP_READ_RANGE
:
417 case HFI1_CMD_EP_WRITE_RANGE
:
418 ret
= handle_eprom_command(fp
, &cmd
);
428 static ssize_t
hfi1_write_iter(struct kiocb
*kiocb
, struct iov_iter
*from
)
430 struct hfi1_filedata
*fd
= kiocb
->ki_filp
->private_data
;
431 struct hfi1_user_sdma_pkt_q
*pq
= fd
->pq
;
432 struct hfi1_user_sdma_comp_q
*cq
= fd
->cq
;
433 int ret
= 0, done
= 0, reqs
= 0;
434 unsigned long dim
= from
->nr_segs
;
441 if (!iter_is_iovec(from
) || !dim
) {
446 hfi1_cdbg(SDMA
, "SDMA request from %u:%u (%lu)",
447 fd
->uctxt
->ctxt
, fd
->subctxt
, dim
);
449 if (atomic_read(&pq
->n_reqs
) == pq
->n_max_reqs
) {
455 unsigned long count
= 0;
457 ret
= hfi1_user_sdma_process_request(
458 kiocb
->ki_filp
, (struct iovec
*)(from
->iov
+ done
),
467 return ret
? ret
: reqs
;
470 static int hfi1_file_mmap(struct file
*fp
, struct vm_area_struct
*vma
)
472 struct hfi1_filedata
*fd
= fp
->private_data
;
473 struct hfi1_ctxtdata
*uctxt
= fd
->uctxt
;
474 struct hfi1_devdata
*dd
;
475 unsigned long flags
, pfn
;
476 u64 token
= vma
->vm_pgoff
<< PAGE_SHIFT
,
478 u8 subctxt
, mapio
= 0, vmf
= 0, type
;
483 if (!is_valid_mmap(token
) || !uctxt
||
484 !(vma
->vm_flags
& VM_SHARED
)) {
489 ctxt
= HFI1_MMAP_TOKEN_GET(CTXT
, token
);
490 subctxt
= HFI1_MMAP_TOKEN_GET(SUBCTXT
, token
);
491 type
= HFI1_MMAP_TOKEN_GET(TYPE
, token
);
492 if (ctxt
!= uctxt
->ctxt
|| subctxt
!= fd
->subctxt
) {
497 flags
= vma
->vm_flags
;
502 memaddr
= ((dd
->physaddr
+ TXE_PIO_SEND
) +
504 (uctxt
->sc
->hw_context
* BIT(16))) +
505 /* 64K PIO space / ctxt */
506 (type
== PIO_BUFS_SOP
?
507 (TXE_PIO_SIZE
/ 2) : 0); /* sop? */
509 * Map only the amount allocated to the context, not the
510 * entire available context's PIO space.
512 memlen
= PAGE_ALIGN(uctxt
->sc
->credits
* PIO_BLOCK_SIZE
);
513 flags
&= ~VM_MAYREAD
;
514 flags
|= VM_DONTCOPY
| VM_DONTEXPAND
;
515 vma
->vm_page_prot
= pgprot_writecombine(vma
->vm_page_prot
);
519 if (flags
& VM_WRITE
) {
524 * The credit return location for this context could be on the
525 * second or third page allocated for credit returns (if number
526 * of enabled contexts > 64 and 128 respectively).
528 memaddr
= dd
->cr_base
[uctxt
->numa_id
].pa
+
529 (((u64
)uctxt
->sc
->hw_free
-
530 (u64
)dd
->cr_base
[uctxt
->numa_id
].va
) & PAGE_MASK
);
532 flags
&= ~VM_MAYWRITE
;
533 flags
|= VM_DONTCOPY
| VM_DONTEXPAND
;
535 * The driver has already allocated memory for credit
536 * returns and programmed it into the chip. Has that
537 * memory been flagged as non-cached?
539 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
543 memaddr
= uctxt
->rcvhdrq_phys
;
544 memlen
= uctxt
->rcvhdrq_size
;
550 * The RcvEgr buffer need to be handled differently
551 * as multiple non-contiguous pages need to be mapped
552 * into the user process.
554 memlen
= uctxt
->egrbufs
.size
;
555 if ((vma
->vm_end
- vma
->vm_start
) != memlen
) {
556 dd_dev_err(dd
, "Eager buffer map size invalid (%lu != %lu)\n",
557 (vma
->vm_end
- vma
->vm_start
), memlen
);
561 if (vma
->vm_flags
& VM_WRITE
) {
565 vma
->vm_flags
&= ~VM_MAYWRITE
;
566 addr
= vma
->vm_start
;
567 for (i
= 0 ; i
< uctxt
->egrbufs
.numbufs
; i
++) {
568 ret
= remap_pfn_range(
570 uctxt
->egrbufs
.buffers
[i
].phys
>> PAGE_SHIFT
,
571 uctxt
->egrbufs
.buffers
[i
].len
,
575 addr
+= uctxt
->egrbufs
.buffers
[i
].len
;
582 * Map only the page that contains this context's user
585 memaddr
= (unsigned long)
586 (dd
->physaddr
+ RXE_PER_CONTEXT_USER
)
587 + (uctxt
->ctxt
* RXE_PER_CONTEXT_SIZE
);
589 * TidFlow table is on the same page as the rest of the
593 flags
|= VM_DONTCOPY
| VM_DONTEXPAND
;
594 vma
->vm_page_prot
= pgprot_noncached(vma
->vm_page_prot
);
599 * Use the page where this context's flags are. User level
600 * knows where it's own bitmap is within the page.
602 memaddr
= (unsigned long)(dd
->events
+
603 ((uctxt
->ctxt
- dd
->first_user_ctxt
) *
604 HFI1_MAX_SHARED_CTXTS
)) & PAGE_MASK
;
607 * v3.7 removes VM_RESERVED but the effect is kept by
610 flags
|= VM_IO
| VM_DONTEXPAND
;
614 memaddr
= kvirt_to_phys((void *)dd
->status
);
616 flags
|= VM_IO
| VM_DONTEXPAND
;
619 if (!HFI1_CAP_IS_USET(DMA_RTAIL
)) {
621 * If the memory allocation failed, the context alloc
622 * also would have failed, so we would never get here
627 if (flags
& VM_WRITE
) {
631 memaddr
= uctxt
->rcvhdrqtailaddr_phys
;
633 flags
&= ~VM_MAYWRITE
;
636 memaddr
= (u64
)uctxt
->subctxt_uregbase
;
638 flags
|= VM_IO
| VM_DONTEXPAND
;
641 case SUBCTXT_RCV_HDRQ
:
642 memaddr
= (u64
)uctxt
->subctxt_rcvhdr_base
;
643 memlen
= uctxt
->rcvhdrq_size
* uctxt
->subctxt_cnt
;
644 flags
|= VM_IO
| VM_DONTEXPAND
;
648 memaddr
= (u64
)uctxt
->subctxt_rcvegrbuf
;
649 memlen
= uctxt
->egrbufs
.size
* uctxt
->subctxt_cnt
;
650 flags
|= VM_IO
| VM_DONTEXPAND
;
651 flags
&= ~VM_MAYWRITE
;
655 struct hfi1_user_sdma_comp_q
*cq
= fd
->cq
;
661 memaddr
= (u64
)cq
->comps
;
662 memlen
= PAGE_ALIGN(sizeof(*cq
->comps
) * cq
->nentries
);
663 flags
|= VM_IO
| VM_DONTEXPAND
;
672 if ((vma
->vm_end
- vma
->vm_start
) != memlen
) {
673 hfi1_cdbg(PROC
, "%u:%u Memory size mismatch %lu:%lu",
674 uctxt
->ctxt
, fd
->subctxt
,
675 (vma
->vm_end
- vma
->vm_start
), memlen
);
680 vma
->vm_flags
= flags
;
682 "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
683 ctxt
, subctxt
, type
, mapio
, vmf
, memaddr
, memlen
,
684 vma
->vm_end
- vma
->vm_start
, vma
->vm_flags
);
685 pfn
= (unsigned long)(memaddr
>> PAGE_SHIFT
);
688 vma
->vm_ops
= &vm_ops
;
691 ret
= io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, memlen
,
694 ret
= remap_pfn_range(vma
, vma
->vm_start
, pfn
, memlen
,
702 * Local (non-chip) user memory is not mapped right away but as it is
703 * accessed by the user-level code.
705 static int vma_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
709 page
= vmalloc_to_page((void *)(vmf
->pgoff
<< PAGE_SHIFT
));
711 return VM_FAULT_SIGBUS
;
719 static unsigned int hfi1_poll(struct file
*fp
, struct poll_table_struct
*pt
)
721 struct hfi1_ctxtdata
*uctxt
;
724 uctxt
= ((struct hfi1_filedata
*)fp
->private_data
)->uctxt
;
727 else if (uctxt
->poll_type
== HFI1_POLL_TYPE_URGENT
)
728 pollflag
= poll_urgent(fp
, pt
);
729 else if (uctxt
->poll_type
== HFI1_POLL_TYPE_ANYRCV
)
730 pollflag
= poll_next(fp
, pt
);
737 static int hfi1_file_close(struct inode
*inode
, struct file
*fp
)
739 struct hfi1_filedata
*fdata
= fp
->private_data
;
740 struct hfi1_ctxtdata
*uctxt
= fdata
->uctxt
;
741 struct hfi1_devdata
*dd
;
742 unsigned long flags
, *ev
;
744 fp
->private_data
= NULL
;
749 hfi1_cdbg(PROC
, "freeing ctxt %u:%u", uctxt
->ctxt
, fdata
->subctxt
);
751 mutex_lock(&hfi1_mutex
);
754 /* drain user sdma queue */
755 hfi1_user_sdma_free_queues(fdata
);
757 /* release the cpu */
758 hfi1_put_proc_affinity(dd
, fdata
->rec_cpu_num
);
761 * Clear any left over, unhandled events so the next process that
762 * gets this context doesn't get confused.
764 ev
= dd
->events
+ ((uctxt
->ctxt
- dd
->first_user_ctxt
) *
765 HFI1_MAX_SHARED_CTXTS
) + fdata
->subctxt
;
769 uctxt
->active_slaves
&= ~(1 << fdata
->subctxt
);
770 uctxt
->subpid
[fdata
->subctxt
] = 0;
771 mutex_unlock(&hfi1_mutex
);
775 spin_lock_irqsave(&dd
->uctxt_lock
, flags
);
777 * Disable receive context and interrupt available, reset all
778 * RcvCtxtCtrl bits to default values.
780 hfi1_rcvctrl(dd
, HFI1_RCVCTRL_CTXT_DIS
|
781 HFI1_RCVCTRL_TIDFLOW_DIS
|
782 HFI1_RCVCTRL_INTRAVAIL_DIS
|
783 HFI1_RCVCTRL_TAILUPD_DIS
|
784 HFI1_RCVCTRL_ONE_PKT_EGR_DIS
|
785 HFI1_RCVCTRL_NO_RHQ_DROP_DIS
|
786 HFI1_RCVCTRL_NO_EGR_DROP_DIS
, uctxt
->ctxt
);
787 /* Clear the context's J_KEY */
788 hfi1_clear_ctxt_jkey(dd
, uctxt
->ctxt
);
790 * Reset context integrity checks to default.
791 * (writes to CSRs probably belong in chip.c)
793 write_kctxt_csr(dd
, uctxt
->sc
->hw_context
, SEND_CTXT_CHECK_ENABLE
,
794 hfi1_pkt_default_send_ctxt_mask(dd
, uctxt
->sc
->type
));
795 sc_disable(uctxt
->sc
);
797 spin_unlock_irqrestore(&dd
->uctxt_lock
, flags
);
799 dd
->rcd
[uctxt
->ctxt
] = NULL
;
801 hfi1_user_exp_rcv_free(fdata
);
802 hfi1_clear_ctxt_pkey(dd
, uctxt
->ctxt
);
804 uctxt
->rcvwait_to
= 0;
805 uctxt
->piowait_to
= 0;
806 uctxt
->rcvnowait
= 0;
807 uctxt
->pionowait
= 0;
808 uctxt
->event_flags
= 0;
810 hfi1_stats
.sps_ctxts
--;
811 if (++dd
->freectxts
== dd
->num_user_contexts
)
813 mutex_unlock(&hfi1_mutex
);
814 hfi1_free_ctxtdata(dd
, uctxt
);
821 * Convert kernel *virtual* addresses to physical addresses.
822 * This is used to vmalloc'ed addresses.
824 static u64
kvirt_to_phys(void *addr
)
829 page
= vmalloc_to_page(addr
);
831 paddr
= page_to_pfn(page
) << PAGE_SHIFT
;
836 static int assign_ctxt(struct file
*fp
, struct hfi1_user_info
*uinfo
)
838 int i_minor
, ret
= 0;
839 unsigned swmajor
, swminor
, alg
= HFI1_ALG_ACROSS
;
841 swmajor
= uinfo
->userversion
>> 16;
842 if (swmajor
!= HFI1_USER_SWMAJOR
) {
847 swminor
= uinfo
->userversion
& 0xffff;
849 if (uinfo
->hfi1_alg
< HFI1_ALG_COUNT
)
850 alg
= uinfo
->hfi1_alg
;
852 mutex_lock(&hfi1_mutex
);
853 /* First, lets check if we need to setup a shared context? */
854 if (uinfo
->subctxt_cnt
) {
855 struct hfi1_filedata
*fd
= fp
->private_data
;
857 ret
= find_shared_ctxt(fp
, uinfo
);
861 fd
->rec_cpu_num
= hfi1_get_proc_affinity(
862 fd
->uctxt
->dd
, fd
->uctxt
->numa_id
);
866 * We execute the following block if we couldn't find a
867 * shared context or if context sharing is not required.
870 i_minor
= iminor(file_inode(fp
)) - HFI1_USER_MINOR_BASE
;
871 ret
= get_user_context(fp
, uinfo
, i_minor
- 1, alg
);
874 mutex_unlock(&hfi1_mutex
);
879 /* return true if the device available for general use */
880 static int usable_device(struct hfi1_devdata
*dd
)
882 struct hfi1_pportdata
*ppd
= dd
->pport
;
884 return driver_lstate(ppd
) == IB_PORT_ACTIVE
;
887 static int get_user_context(struct file
*fp
, struct hfi1_user_info
*uinfo
,
888 int devno
, unsigned alg
)
890 struct hfi1_devdata
*dd
= NULL
;
891 int ret
= 0, devmax
, npresent
, nup
, dev
;
893 devmax
= hfi1_count_units(&npresent
, &nup
);
903 dd
= hfi1_lookup(devno
);
906 else if (!dd
->freectxts
)
909 struct hfi1_devdata
*pdd
;
911 if (alg
== HFI1_ALG_ACROSS
) {
914 for (dev
= 0; dev
< devmax
; dev
++) {
915 pdd
= hfi1_lookup(dev
);
918 if (!usable_device(pdd
))
920 if (pdd
->freectxts
&&
921 pdd
->freectxts
> free
) {
923 free
= pdd
->freectxts
;
927 for (dev
= 0; dev
< devmax
; dev
++) {
928 pdd
= hfi1_lookup(dev
);
931 if (!usable_device(pdd
))
933 if (pdd
->freectxts
) {
943 return ret
? ret
: allocate_ctxt(fp
, dd
, uinfo
);
946 static int find_shared_ctxt(struct file
*fp
,
947 const struct hfi1_user_info
*uinfo
)
951 struct hfi1_filedata
*fd
= fp
->private_data
;
953 devmax
= hfi1_count_units(NULL
, NULL
);
955 for (ndev
= 0; ndev
< devmax
; ndev
++) {
956 struct hfi1_devdata
*dd
= hfi1_lookup(ndev
);
958 if (!(dd
&& (dd
->flags
& HFI1_PRESENT
) && dd
->kregbase
))
960 for (i
= dd
->first_user_ctxt
; i
< dd
->num_rcv_contexts
; i
++) {
961 struct hfi1_ctxtdata
*uctxt
= dd
->rcd
[i
];
963 /* Skip ctxts which are not yet open */
964 if (!uctxt
|| !uctxt
->cnt
)
966 /* Skip ctxt if it doesn't match the requested one */
967 if (memcmp(uctxt
->uuid
, uinfo
->uuid
,
968 sizeof(uctxt
->uuid
)) ||
969 uctxt
->jkey
!= generate_jkey(current_uid()) ||
970 uctxt
->subctxt_id
!= uinfo
->subctxt_id
||
971 uctxt
->subctxt_cnt
!= uinfo
->subctxt_cnt
)
974 /* Verify the sharing process matches the master */
975 if (uctxt
->userversion
!= uinfo
->userversion
||
976 uctxt
->cnt
>= uctxt
->subctxt_cnt
) {
981 fd
->subctxt
= uctxt
->cnt
++;
982 uctxt
->subpid
[fd
->subctxt
] = current
->pid
;
983 uctxt
->active_slaves
|= 1 << fd
->subctxt
;
993 static int allocate_ctxt(struct file
*fp
, struct hfi1_devdata
*dd
,
994 struct hfi1_user_info
*uinfo
)
996 struct hfi1_filedata
*fd
= fp
->private_data
;
997 struct hfi1_ctxtdata
*uctxt
;
1001 if (dd
->flags
& HFI1_FROZEN
) {
1003 * Pick an error that is unique from all other errors
1004 * that are returned so the user process knows that
1005 * it tried to allocate while the SPC was frozen. It
1006 * it should be able to retry with success in a short
1012 for (ctxt
= dd
->first_user_ctxt
; ctxt
< dd
->num_rcv_contexts
; ctxt
++)
1016 if (ctxt
== dd
->num_rcv_contexts
)
1019 fd
->rec_cpu_num
= hfi1_get_proc_affinity(dd
, -1);
1020 if (fd
->rec_cpu_num
!= -1)
1021 numa
= cpu_to_node(fd
->rec_cpu_num
);
1023 numa
= numa_node_id();
1024 uctxt
= hfi1_create_ctxtdata(dd
->pport
, ctxt
, numa
);
1027 "Unable to allocate ctxtdata memory, failing open\n");
1030 hfi1_cdbg(PROC
, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
1031 uctxt
->ctxt
, fd
->subctxt
, current
->pid
, fd
->rec_cpu_num
,
1035 * Allocate and enable a PIO send context.
1037 uctxt
->sc
= sc_alloc(dd
, SC_USER
, uctxt
->rcvhdrqentsize
,
1042 hfi1_cdbg(PROC
, "allocated send context %u(%u)\n", uctxt
->sc
->sw_index
,
1043 uctxt
->sc
->hw_context
);
1044 ret
= sc_enable(uctxt
->sc
);
1048 * Setup shared context resources if the user-level has requested
1049 * shared contexts and this is the 'master' process.
1050 * This has to be done here so the rest of the sub-contexts find the
1053 if (uinfo
->subctxt_cnt
&& !fd
->subctxt
) {
1054 ret
= init_subctxts(uctxt
, uinfo
);
1056 * On error, we don't need to disable and de-allocate the
1057 * send context because it will be done during file close
1062 uctxt
->userversion
= uinfo
->userversion
;
1063 uctxt
->pid
= current
->pid
;
1064 uctxt
->flags
= HFI1_CAP_UGET(MASK
);
1065 init_waitqueue_head(&uctxt
->wait
);
1066 strlcpy(uctxt
->comm
, current
->comm
, sizeof(uctxt
->comm
));
1067 memcpy(uctxt
->uuid
, uinfo
->uuid
, sizeof(uctxt
->uuid
));
1068 uctxt
->jkey
= generate_jkey(current_uid());
1069 INIT_LIST_HEAD(&uctxt
->sdma_queues
);
1070 spin_lock_init(&uctxt
->sdma_qlock
);
1071 hfi1_stats
.sps_ctxts
++;
1073 * Disable ASPM when there are open user/PSM contexts to avoid
1074 * issues with ASPM L1 exit latency
1076 if (dd
->freectxts
-- == dd
->num_user_contexts
)
1077 aspm_disable_all(dd
);
1083 static int init_subctxts(struct hfi1_ctxtdata
*uctxt
,
1084 const struct hfi1_user_info
*uinfo
)
1086 unsigned num_subctxts
;
1088 num_subctxts
= uinfo
->subctxt_cnt
;
1089 if (num_subctxts
> HFI1_MAX_SHARED_CTXTS
)
1092 uctxt
->subctxt_cnt
= uinfo
->subctxt_cnt
;
1093 uctxt
->subctxt_id
= uinfo
->subctxt_id
;
1094 uctxt
->active_slaves
= 1;
1095 uctxt
->redirect_seq_cnt
= 1;
1096 set_bit(HFI1_CTXT_MASTER_UNINIT
, &uctxt
->event_flags
);
1101 static int setup_subctxt(struct hfi1_ctxtdata
*uctxt
)
1104 unsigned num_subctxts
= uctxt
->subctxt_cnt
;
1106 uctxt
->subctxt_uregbase
= vmalloc_user(PAGE_SIZE
);
1107 if (!uctxt
->subctxt_uregbase
) {
1111 /* We can take the size of the RcvHdr Queue from the master */
1112 uctxt
->subctxt_rcvhdr_base
= vmalloc_user(uctxt
->rcvhdrq_size
*
1114 if (!uctxt
->subctxt_rcvhdr_base
) {
1119 uctxt
->subctxt_rcvegrbuf
= vmalloc_user(uctxt
->egrbufs
.size
*
1121 if (!uctxt
->subctxt_rcvegrbuf
) {
1127 vfree(uctxt
->subctxt_rcvhdr_base
);
1129 vfree(uctxt
->subctxt_uregbase
);
1130 uctxt
->subctxt_uregbase
= NULL
;
1135 static int user_init(struct file
*fp
)
1137 unsigned int rcvctrl_ops
= 0;
1138 struct hfi1_filedata
*fd
= fp
->private_data
;
1139 struct hfi1_ctxtdata
*uctxt
= fd
->uctxt
;
1141 /* make sure that the context has already been setup */
1142 if (!test_bit(HFI1_CTXT_SETUP_DONE
, &uctxt
->event_flags
))
1145 /* initialize poll variables... */
1147 uctxt
->urgent_poll
= 0;
1150 * Now enable the ctxt for receive.
1151 * For chips that are set to DMA the tail register to memory
1152 * when they change (and when the update bit transitions from
1153 * 0 to 1. So for those chips, we turn it off and then back on.
1154 * This will (very briefly) affect any other open ctxts, but the
1155 * duration is very short, and therefore isn't an issue. We
1156 * explicitly set the in-memory tail copy to 0 beforehand, so we
1157 * don't have to wait to be sure the DMA update has happened
1158 * (chip resets head/tail to 0 on transition to enable).
1160 if (uctxt
->rcvhdrtail_kvaddr
)
1161 clear_rcvhdrtail(uctxt
);
1163 /* Setup J_KEY before enabling the context */
1164 hfi1_set_ctxt_jkey(uctxt
->dd
, uctxt
->ctxt
, uctxt
->jkey
);
1166 rcvctrl_ops
= HFI1_RCVCTRL_CTXT_ENB
;
1167 if (HFI1_CAP_KGET_MASK(uctxt
->flags
, HDRSUPP
))
1168 rcvctrl_ops
|= HFI1_RCVCTRL_TIDFLOW_ENB
;
1170 * Ignore the bit in the flags for now until proper
1171 * support for multiple packet per rcv array entry is
1174 if (!HFI1_CAP_KGET_MASK(uctxt
->flags
, MULTI_PKT_EGR
))
1175 rcvctrl_ops
|= HFI1_RCVCTRL_ONE_PKT_EGR_ENB
;
1176 if (HFI1_CAP_KGET_MASK(uctxt
->flags
, NODROP_EGR_FULL
))
1177 rcvctrl_ops
|= HFI1_RCVCTRL_NO_EGR_DROP_ENB
;
1178 if (HFI1_CAP_KGET_MASK(uctxt
->flags
, NODROP_RHQ_FULL
))
1179 rcvctrl_ops
|= HFI1_RCVCTRL_NO_RHQ_DROP_ENB
;
1181 * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1182 * We can't rely on the correct value to be set from prior
1183 * uses of the chip or ctxt. Therefore, add the rcvctrl op
1186 if (HFI1_CAP_KGET_MASK(uctxt
->flags
, DMA_RTAIL
))
1187 rcvctrl_ops
|= HFI1_RCVCTRL_TAILUPD_ENB
;
1189 rcvctrl_ops
|= HFI1_RCVCTRL_TAILUPD_DIS
;
1190 hfi1_rcvctrl(uctxt
->dd
, rcvctrl_ops
, uctxt
->ctxt
);
1192 /* Notify any waiting slaves */
1193 if (uctxt
->subctxt_cnt
) {
1194 clear_bit(HFI1_CTXT_MASTER_UNINIT
, &uctxt
->event_flags
);
1195 wake_up(&uctxt
->wait
);
1201 static int get_ctxt_info(struct file
*fp
, void __user
*ubase
, __u32 len
)
1203 struct hfi1_ctxt_info cinfo
;
1204 struct hfi1_filedata
*fd
= fp
->private_data
;
1205 struct hfi1_ctxtdata
*uctxt
= fd
->uctxt
;
1208 memset(&cinfo
, 0, sizeof(cinfo
));
1209 ret
= hfi1_get_base_kinfo(uctxt
, &cinfo
);
1212 cinfo
.num_active
= hfi1_count_active_units();
1213 cinfo
.unit
= uctxt
->dd
->unit
;
1214 cinfo
.ctxt
= uctxt
->ctxt
;
1215 cinfo
.subctxt
= fd
->subctxt
;
1216 cinfo
.rcvtids
= roundup(uctxt
->egrbufs
.alloced
,
1217 uctxt
->dd
->rcv_entries
.group_size
) +
1218 uctxt
->expected_count
;
1219 cinfo
.credits
= uctxt
->sc
->credits
;
1220 cinfo
.numa_node
= uctxt
->numa_id
;
1221 cinfo
.rec_cpu
= fd
->rec_cpu_num
;
1222 cinfo
.send_ctxt
= uctxt
->sc
->hw_context
;
1224 cinfo
.egrtids
= uctxt
->egrbufs
.alloced
;
1225 cinfo
.rcvhdrq_cnt
= uctxt
->rcvhdrq_cnt
;
1226 cinfo
.rcvhdrq_entsize
= uctxt
->rcvhdrqentsize
<< 2;
1227 cinfo
.sdma_ring_size
= fd
->cq
->nentries
;
1228 cinfo
.rcvegr_size
= uctxt
->egrbufs
.rcvtid_size
;
1230 trace_hfi1_ctxt_info(uctxt
->dd
, uctxt
->ctxt
, fd
->subctxt
, cinfo
);
1231 if (copy_to_user(ubase
, &cinfo
, sizeof(cinfo
)))
1237 static int setup_ctxt(struct file
*fp
)
1239 struct hfi1_filedata
*fd
= fp
->private_data
;
1240 struct hfi1_ctxtdata
*uctxt
= fd
->uctxt
;
1241 struct hfi1_devdata
*dd
= uctxt
->dd
;
1245 * Context should be set up only once, including allocation and
1246 * programming of eager buffers. This is done if context sharing
1247 * is not requested or by the master process.
1249 if (!uctxt
->subctxt_cnt
|| !fd
->subctxt
) {
1250 ret
= hfi1_init_ctxt(uctxt
->sc
);
1254 /* Now allocate the RcvHdr queue and eager buffers. */
1255 ret
= hfi1_create_rcvhdrq(dd
, uctxt
);
1258 ret
= hfi1_setup_eagerbufs(uctxt
);
1261 if (uctxt
->subctxt_cnt
&& !fd
->subctxt
) {
1262 ret
= setup_subctxt(uctxt
);
1267 ret
= wait_event_interruptible(uctxt
->wait
, !test_bit(
1268 HFI1_CTXT_MASTER_UNINIT
,
1269 &uctxt
->event_flags
));
1274 ret
= hfi1_user_sdma_alloc_queues(uctxt
, fp
);
1278 * Expected receive has to be setup for all processes (including
1279 * shared contexts). However, it has to be done after the master
1280 * context has been fully configured as it depends on the
1281 * eager/expected split of the RcvArray entries.
1282 * Setting it up here ensures that the subcontexts will be waiting
1283 * (due to the above wait_event_interruptible() until the master
1286 ret
= hfi1_user_exp_rcv_init(fp
);
1290 set_bit(HFI1_CTXT_SETUP_DONE
, &uctxt
->event_flags
);
1295 static int get_base_info(struct file
*fp
, void __user
*ubase
, __u32 len
)
1297 struct hfi1_base_info binfo
;
1298 struct hfi1_filedata
*fd
= fp
->private_data
;
1299 struct hfi1_ctxtdata
*uctxt
= fd
->uctxt
;
1300 struct hfi1_devdata
*dd
= uctxt
->dd
;
1305 trace_hfi1_uctxtdata(uctxt
->dd
, uctxt
);
1307 memset(&binfo
, 0, sizeof(binfo
));
1308 binfo
.hw_version
= dd
->revision
;
1309 binfo
.sw_version
= HFI1_KERN_SWVERSION
;
1310 binfo
.bthqp
= kdeth_qp
;
1311 binfo
.jkey
= uctxt
->jkey
;
1313 * If more than 64 contexts are enabled the allocated credit
1314 * return will span two or three contiguous pages. Since we only
1315 * map the page containing the context's credit return address,
1316 * we need to calculate the offset in the proper page.
1318 offset
= ((u64
)uctxt
->sc
->hw_free
-
1319 (u64
)dd
->cr_base
[uctxt
->numa_id
].va
) % PAGE_SIZE
;
1320 binfo
.sc_credits_addr
= HFI1_MMAP_TOKEN(PIO_CRED
, uctxt
->ctxt
,
1321 fd
->subctxt
, offset
);
1322 binfo
.pio_bufbase
= HFI1_MMAP_TOKEN(PIO_BUFS
, uctxt
->ctxt
,
1324 uctxt
->sc
->base_addr
);
1325 binfo
.pio_bufbase_sop
= HFI1_MMAP_TOKEN(PIO_BUFS_SOP
,
1328 uctxt
->sc
->base_addr
);
1329 binfo
.rcvhdr_bufbase
= HFI1_MMAP_TOKEN(RCV_HDRQ
, uctxt
->ctxt
,
1332 binfo
.rcvegr_bufbase
= HFI1_MMAP_TOKEN(RCV_EGRBUF
, uctxt
->ctxt
,
1334 uctxt
->egrbufs
.rcvtids
[0].phys
);
1335 binfo
.sdma_comp_bufbase
= HFI1_MMAP_TOKEN(SDMA_COMP
, uctxt
->ctxt
,
1339 * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1341 binfo
.user_regbase
= HFI1_MMAP_TOKEN(UREGS
, uctxt
->ctxt
,
1343 offset
= offset_in_page((((uctxt
->ctxt
- dd
->first_user_ctxt
) *
1344 HFI1_MAX_SHARED_CTXTS
) + fd
->subctxt
) *
1345 sizeof(*dd
->events
));
1346 binfo
.events_bufbase
= HFI1_MMAP_TOKEN(EVENTS
, uctxt
->ctxt
,
1349 binfo
.status_bufbase
= HFI1_MMAP_TOKEN(STATUS
, uctxt
->ctxt
,
1352 if (HFI1_CAP_IS_USET(DMA_RTAIL
))
1353 binfo
.rcvhdrtail_base
= HFI1_MMAP_TOKEN(RTAIL
, uctxt
->ctxt
,
1355 if (uctxt
->subctxt_cnt
) {
1356 binfo
.subctxt_uregbase
= HFI1_MMAP_TOKEN(SUBCTXT_UREGS
,
1359 binfo
.subctxt_rcvhdrbuf
= HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ
,
1362 binfo
.subctxt_rcvegrbuf
= HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF
,
1366 sz
= (len
< sizeof(binfo
)) ? len
: sizeof(binfo
);
1367 if (copy_to_user(ubase
, &binfo
, sz
))
1372 static unsigned int poll_urgent(struct file
*fp
,
1373 struct poll_table_struct
*pt
)
1375 struct hfi1_filedata
*fd
= fp
->private_data
;
1376 struct hfi1_ctxtdata
*uctxt
= fd
->uctxt
;
1377 struct hfi1_devdata
*dd
= uctxt
->dd
;
1380 poll_wait(fp
, &uctxt
->wait
, pt
);
1382 spin_lock_irq(&dd
->uctxt_lock
);
1383 if (uctxt
->urgent
!= uctxt
->urgent_poll
) {
1384 pollflag
= POLLIN
| POLLRDNORM
;
1385 uctxt
->urgent_poll
= uctxt
->urgent
;
1388 set_bit(HFI1_CTXT_WAITING_URG
, &uctxt
->event_flags
);
1390 spin_unlock_irq(&dd
->uctxt_lock
);
1395 static unsigned int poll_next(struct file
*fp
,
1396 struct poll_table_struct
*pt
)
1398 struct hfi1_filedata
*fd
= fp
->private_data
;
1399 struct hfi1_ctxtdata
*uctxt
= fd
->uctxt
;
1400 struct hfi1_devdata
*dd
= uctxt
->dd
;
1403 poll_wait(fp
, &uctxt
->wait
, pt
);
1405 spin_lock_irq(&dd
->uctxt_lock
);
1406 if (hdrqempty(uctxt
)) {
1407 set_bit(HFI1_CTXT_WAITING_RCV
, &uctxt
->event_flags
);
1408 hfi1_rcvctrl(dd
, HFI1_RCVCTRL_INTRAVAIL_ENB
, uctxt
->ctxt
);
1411 pollflag
= POLLIN
| POLLRDNORM
;
1413 spin_unlock_irq(&dd
->uctxt_lock
);
1419 * Find all user contexts in use, and set the specified bit in their
1421 * See also find_ctxt() for a similar use, that is specific to send buffers.
1423 int hfi1_set_uevent_bits(struct hfi1_pportdata
*ppd
, const int evtbit
)
1425 struct hfi1_ctxtdata
*uctxt
;
1426 struct hfi1_devdata
*dd
= ppd
->dd
;
1429 unsigned long flags
;
1436 spin_lock_irqsave(&dd
->uctxt_lock
, flags
);
1437 for (ctxt
= dd
->first_user_ctxt
; ctxt
< dd
->num_rcv_contexts
;
1439 uctxt
= dd
->rcd
[ctxt
];
1441 unsigned long *evs
= dd
->events
+
1442 (uctxt
->ctxt
- dd
->first_user_ctxt
) *
1443 HFI1_MAX_SHARED_CTXTS
;
1446 * subctxt_cnt is 0 if not shared, so do base
1447 * separately, first, then remaining subctxt, if any
1449 set_bit(evtbit
, evs
);
1450 for (i
= 1; i
< uctxt
->subctxt_cnt
; i
++)
1451 set_bit(evtbit
, evs
+ i
);
1454 spin_unlock_irqrestore(&dd
->uctxt_lock
, flags
);
1460 * manage_rcvq - manage a context's receive queue
1461 * @uctxt: the context
1462 * @subctxt: the sub-context
1463 * @start_stop: action to carry out
1465 * start_stop == 0 disables receive on the context, for use in queue
1466 * overflow conditions. start_stop==1 re-enables, to be used to
1467 * re-init the software copy of the head register
1469 static int manage_rcvq(struct hfi1_ctxtdata
*uctxt
, unsigned subctxt
,
1472 struct hfi1_devdata
*dd
= uctxt
->dd
;
1473 unsigned int rcvctrl_op
;
1477 /* atomically clear receive enable ctxt. */
1480 * On enable, force in-memory copy of the tail register to
1481 * 0, so that protocol code doesn't have to worry about
1482 * whether or not the chip has yet updated the in-memory
1483 * copy or not on return from the system call. The chip
1484 * always resets it's tail register back to 0 on a
1485 * transition from disabled to enabled.
1487 if (uctxt
->rcvhdrtail_kvaddr
)
1488 clear_rcvhdrtail(uctxt
);
1489 rcvctrl_op
= HFI1_RCVCTRL_CTXT_ENB
;
1491 rcvctrl_op
= HFI1_RCVCTRL_CTXT_DIS
;
1493 hfi1_rcvctrl(dd
, rcvctrl_op
, uctxt
->ctxt
);
1494 /* always; new head should be equal to new tail; see above */
1500 * clear the event notifier events for this context.
1501 * User process then performs actions appropriate to bit having been
1502 * set, if desired, and checks again in future.
1504 static int user_event_ack(struct hfi1_ctxtdata
*uctxt
, int subctxt
,
1505 unsigned long events
)
1508 struct hfi1_devdata
*dd
= uctxt
->dd
;
1514 evs
= dd
->events
+ ((uctxt
->ctxt
- dd
->first_user_ctxt
) *
1515 HFI1_MAX_SHARED_CTXTS
) + subctxt
;
1517 for (i
= 0; i
<= _HFI1_MAX_EVENT_BIT
; i
++) {
1518 if (!test_bit(i
, &events
))
1525 static int set_ctxt_pkey(struct hfi1_ctxtdata
*uctxt
, unsigned subctxt
,
1528 int ret
= -ENOENT
, i
, intable
= 0;
1529 struct hfi1_pportdata
*ppd
= uctxt
->ppd
;
1530 struct hfi1_devdata
*dd
= uctxt
->dd
;
1532 if (pkey
== LIM_MGMT_P_KEY
|| pkey
== FULL_MGMT_P_KEY
) {
1537 for (i
= 0; i
< ARRAY_SIZE(ppd
->pkeys
); i
++)
1538 if (pkey
== ppd
->pkeys
[i
]) {
1544 ret
= hfi1_set_ctxt_pkey(dd
, uctxt
->ctxt
, pkey
);
1549 static int ui_open(struct inode
*inode
, struct file
*filp
)
1551 struct hfi1_devdata
*dd
;
1553 dd
= container_of(inode
->i_cdev
, struct hfi1_devdata
, ui_cdev
);
1554 filp
->private_data
= dd
; /* for other methods */
1558 static int ui_release(struct inode
*inode
, struct file
*filp
)
1564 static loff_t
ui_lseek(struct file
*filp
, loff_t offset
, int whence
)
1566 struct hfi1_devdata
*dd
= filp
->private_data
;
1568 return fixed_size_llseek(filp
, offset
, whence
,
1569 (dd
->kregend
- dd
->kregbase
) + DC8051_DATA_MEM_SIZE
);
1572 /* NOTE: assumes unsigned long is 8 bytes */
1573 static ssize_t
ui_read(struct file
*filp
, char __user
*buf
, size_t count
,
1576 struct hfi1_devdata
*dd
= filp
->private_data
;
1577 void __iomem
*base
= dd
->kregbase
;
1578 unsigned long total
, csr_off
,
1579 barlen
= (dd
->kregend
- dd
->kregbase
);
1582 /* only read 8 byte quantities */
1583 if ((count
% 8) != 0)
1585 /* offset must be 8-byte aligned */
1586 if ((*f_pos
% 8) != 0)
1588 /* destination buffer must be 8-byte aligned */
1589 if ((unsigned long)buf
% 8 != 0)
1591 /* must be in range */
1592 if (*f_pos
+ count
> (barlen
+ DC8051_DATA_MEM_SIZE
))
1594 /* only set the base if we are not starting past the BAR */
1595 if (*f_pos
< barlen
)
1598 for (total
= 0; total
< count
; total
+= 8, csr_off
+= 8) {
1599 /* accessing LCB CSRs requires more checks */
1600 if (is_lcb_offset(csr_off
)) {
1601 if (read_lcb_csr(dd
, csr_off
, (u64
*)&data
))
1605 * Cannot read ASIC GPIO/QSFP* clear and force CSRs without a
1606 * false parity error. Avoid the whole issue by not reading
1607 * them. These registers are defined as having a read value
1610 else if (csr_off
== ASIC_GPIO_CLEAR
||
1611 csr_off
== ASIC_GPIO_FORCE
||
1612 csr_off
== ASIC_QSFP1_CLEAR
||
1613 csr_off
== ASIC_QSFP1_FORCE
||
1614 csr_off
== ASIC_QSFP2_CLEAR
||
1615 csr_off
== ASIC_QSFP2_FORCE
)
1617 else if (csr_off
>= barlen
) {
1619 * read_8051_data can read more than just 8 bytes at
1620 * a time. However, folding this into the loop and
1621 * handling the reads in 8 byte increments allows us
1622 * to smoothly transition from chip memory to 8051
1625 if (read_8051_data(dd
,
1626 (u32
)(csr_off
- barlen
),
1627 sizeof(data
), &data
))
1630 data
= readq(base
+ total
);
1631 if (put_user(data
, (unsigned long __user
*)(buf
+ total
)))
1638 /* NOTE: assumes unsigned long is 8 bytes */
1639 static ssize_t
ui_write(struct file
*filp
, const char __user
*buf
,
1640 size_t count
, loff_t
*f_pos
)
1642 struct hfi1_devdata
*dd
= filp
->private_data
;
1644 unsigned long total
, data
, csr_off
;
1647 /* only write 8 byte quantities */
1648 if ((count
% 8) != 0)
1650 /* offset must be 8-byte aligned */
1651 if ((*f_pos
% 8) != 0)
1653 /* source buffer must be 8-byte aligned */
1654 if ((unsigned long)buf
% 8 != 0)
1656 /* must be in range */
1657 if (*f_pos
+ count
> dd
->kregend
- dd
->kregbase
)
1660 base
= (void __iomem
*)dd
->kregbase
+ *f_pos
;
1663 for (total
= 0; total
< count
; total
+= 8, csr_off
+= 8) {
1664 if (get_user(data
, (unsigned long __user
*)(buf
+ total
)))
1666 /* accessing LCB CSRs requires a special procedure */
1667 if (is_lcb_offset(csr_off
)) {
1669 int ret
= acquire_lcb_access(dd
, 1);
1677 release_lcb_access(dd
, 1);
1681 writeq(data
, base
+ total
);
1684 release_lcb_access(dd
, 1);
1689 static const struct file_operations ui_file_ops
= {
1690 .owner
= THIS_MODULE
,
1695 .release
= ui_release
,
1698 #define UI_OFFSET 192 /* device minor offset for UI devices */
1699 static int create_ui
= 1;
1701 static struct cdev wildcard_cdev
;
1702 static struct device
*wildcard_device
;
1704 static atomic_t user_count
= ATOMIC_INIT(0);
1706 static void user_remove(struct hfi1_devdata
*dd
)
1708 if (atomic_dec_return(&user_count
) == 0)
1709 hfi1_cdev_cleanup(&wildcard_cdev
, &wildcard_device
);
1711 hfi1_cdev_cleanup(&dd
->user_cdev
, &dd
->user_device
);
1712 hfi1_cdev_cleanup(&dd
->ui_cdev
, &dd
->ui_device
);
1715 static int user_add(struct hfi1_devdata
*dd
)
1720 if (atomic_inc_return(&user_count
) == 1) {
1721 ret
= hfi1_cdev_init(0, class_name(), &hfi1_file_ops
,
1722 &wildcard_cdev
, &wildcard_device
,
1728 snprintf(name
, sizeof(name
), "%s_%d", class_name(), dd
->unit
);
1729 ret
= hfi1_cdev_init(dd
->unit
+ 1, name
, &hfi1_file_ops
,
1730 &dd
->user_cdev
, &dd
->user_device
,
1736 snprintf(name
, sizeof(name
),
1737 "%s_ui%d", class_name(), dd
->unit
);
1738 ret
= hfi1_cdev_init(dd
->unit
+ UI_OFFSET
, name
, &ui_file_ops
,
1739 &dd
->ui_cdev
, &dd
->ui_device
,
1752 * Create per-unit files in /dev
1754 int hfi1_device_create(struct hfi1_devdata
*dd
)
1759 ret
= hfi1_diag_add(dd
);
1766 * Remove per-unit files in /dev
1767 * void, core kernel returns no errors for this stuff
1769 void hfi1_device_remove(struct hfi1_devdata
*dd
)
1772 hfi1_diag_remove(dd
);