1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright 2014 IBM Corp.
6 #include <linux/spinlock.h>
7 #include <linux/module.h>
8 #include <linux/export.h>
9 #include <linux/kernel.h>
10 #include <linux/bitmap.h>
11 #include <linux/sched/signal.h>
12 #include <linux/poll.h>
13 #include <linux/pid.h>
16 #include <linux/slab.h>
17 #include <linux/sched/mm.h>
18 #include <linux/mmu_context.h>
19 #include <asm/cputable.h>
20 #include <asm/current.h>
21 #include <asm/copro.h>
26 #define CXL_NUM_MINORS 256 /* Total to reserve */
28 #define CXL_AFU_MINOR_D(afu) (CXL_CARD_MINOR(afu->adapter) + 1 + (3 * afu->slice))
29 #define CXL_AFU_MINOR_M(afu) (CXL_AFU_MINOR_D(afu) + 1)
30 #define CXL_AFU_MINOR_S(afu) (CXL_AFU_MINOR_D(afu) + 2)
31 #define CXL_AFU_MKDEV_D(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_D(afu))
32 #define CXL_AFU_MKDEV_M(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_M(afu))
33 #define CXL_AFU_MKDEV_S(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_S(afu))
35 #define CXL_DEVT_AFU(dev) ((MINOR(dev) % CXL_DEV_MINORS - 1) / 3)
37 #define CXL_DEVT_IS_CARD(dev) (MINOR(dev) % CXL_DEV_MINORS == 0)
41 static struct class *cxl_class
;
43 static int __afu_open(struct inode
*inode
, struct file
*file
, bool master
)
47 struct cxl_context
*ctx
;
48 int adapter_num
= CXL_DEVT_ADAPTER(inode
->i_rdev
);
49 int slice
= CXL_DEVT_AFU(inode
->i_rdev
);
52 pr_devel("afu_open afu%i.%i\n", slice
, adapter_num
);
54 if (!(adapter
= get_cxl_adapter(adapter_num
)))
57 if (slice
> adapter
->slices
)
60 spin_lock(&adapter
->afu_list_lock
);
61 if (!(afu
= adapter
->afu
[slice
])) {
62 spin_unlock(&adapter
->afu_list_lock
);
67 * taking a ref to the afu so that it doesn't go away
68 * for rest of the function. This ref is released before
72 spin_unlock(&adapter
->afu_list_lock
);
74 if (!afu
->current_mode
)
77 if (!cxl_ops
->link_ok(adapter
, afu
)) {
82 if (!(ctx
= cxl_context_alloc())) {
87 rc
= cxl_context_init(ctx
, afu
, master
);
91 cxl_context_set_mapping(ctx
, inode
->i_mapping
);
93 pr_devel("afu_open pe: %i\n", ctx
->pe
);
94 file
->private_data
= ctx
;
96 /* indicate success */
100 /* release the ref taken earlier */
103 put_device(&adapter
->dev
);
107 int afu_open(struct inode
*inode
, struct file
*file
)
109 return __afu_open(inode
, file
, false);
112 static int afu_master_open(struct inode
*inode
, struct file
*file
)
114 return __afu_open(inode
, file
, true);
117 int afu_release(struct inode
*inode
, struct file
*file
)
119 struct cxl_context
*ctx
= file
->private_data
;
121 pr_devel("%s: closing cxl file descriptor. pe: %i\n",
123 cxl_context_detach(ctx
);
127 * Delete the context's mapping pointer, unless it's created by the
128 * kernel API, in which case leave it so it can be freed by reclaim_ctx()
130 if (!ctx
->kernelapi
) {
131 mutex_lock(&ctx
->mapping_lock
);
133 mutex_unlock(&ctx
->mapping_lock
);
137 * At this this point all bottom halfs have finished and we should be
138 * getting no more IRQs from the hardware for this context. Once it's
139 * removed from the IDR (and RCU synchronised) it's safe to free the
142 cxl_context_free(ctx
);
147 static long afu_ioctl_start_work(struct cxl_context
*ctx
,
148 struct cxl_ioctl_start_work __user
*uwork
)
150 struct cxl_ioctl_start_work work
;
154 pr_devel("%s: pe: %i\n", __func__
, ctx
->pe
);
156 /* Do this outside the status_mutex to avoid a circular dependency with
157 * the locking in cxl_mmap_fault() */
158 if (copy_from_user(&work
, uwork
, sizeof(work
)))
161 mutex_lock(&ctx
->status_mutex
);
162 if (ctx
->status
!= OPENED
) {
168 * if any of the reserved fields are set or any of the unused
169 * flags are set it's invalid
171 if (work
.reserved1
|| work
.reserved2
|| work
.reserved3
||
172 work
.reserved4
|| work
.reserved5
||
173 (work
.flags
& ~CXL_START_WORK_ALL
)) {
178 if (!(work
.flags
& CXL_START_WORK_NUM_IRQS
))
179 work
.num_interrupts
= ctx
->afu
->pp_irqs
;
180 else if ((work
.num_interrupts
< ctx
->afu
->pp_irqs
) ||
181 (work
.num_interrupts
> ctx
->afu
->irqs_max
)) {
186 if ((rc
= afu_register_irqs(ctx
, work
.num_interrupts
)))
189 if (work
.flags
& CXL_START_WORK_AMR
)
190 amr
= work
.amr
& mfspr(SPRN_UAMOR
);
192 if (work
.flags
& CXL_START_WORK_TID
)
193 ctx
->assign_tidr
= true;
195 ctx
->mmio_err_ff
= !!(work
.flags
& CXL_START_WORK_ERR_FF
);
198 * Increment the mapped context count for adapter. This also checks
199 * if adapter_context_lock is taken.
201 rc
= cxl_adapter_context_get(ctx
->afu
->adapter
);
203 afu_release_irqs(ctx
, ctx
);
208 * We grab the PID here and not in the file open to allow for the case
209 * where a process (master, some daemon, etc) has opened the chardev on
210 * behalf of another process, so the AFU's mm gets bound to the process
211 * that performs this ioctl and not the process that opened the file.
212 * Also we grab the PID of the group leader so that if the task that
213 * has performed the attach operation exits the mm context of the
214 * process is still accessible.
216 ctx
->pid
= get_task_pid(current
, PIDTYPE_PID
);
218 /* acquire a reference to the task's mm */
219 ctx
->mm
= get_task_mm(current
);
221 /* ensure this mm_struct can't be freed */
222 cxl_context_mm_count_get(ctx
);
225 /* decrement the use count from above */
227 /* make TLBIs for this context global */
228 mm_context_add_copro(ctx
->mm
);
232 * Increment driver use count. Enables global TLBIs for hash
233 * and callbacks to handle the segment table
238 * A barrier is needed to make sure all TLBIs are global
239 * before we attach and the context starts being used by the
242 * Needed after mm_context_add_copro() for radix and
243 * cxl_ctx_get() for hash/p8.
245 * The barrier should really be mb(), since it involves a
246 * device. However, it's only useful when we have local
247 * vs. global TLBIs, i.e SMP=y. So keep smp_mb().
251 trace_cxl_attach(ctx
, work
.work_element_descriptor
, work
.num_interrupts
, amr
);
253 if ((rc
= cxl_ops
->attach_process(ctx
, false, work
.work_element_descriptor
,
255 afu_release_irqs(ctx
, ctx
);
256 cxl_adapter_context_put(ctx
->afu
->adapter
);
260 cxl_context_mm_count_put(ctx
);
262 mm_context_remove_copro(ctx
->mm
);
267 if (work
.flags
& CXL_START_WORK_TID
) {
268 work
.tid
= ctx
->tidr
;
269 if (copy_to_user(uwork
, &work
, sizeof(work
)))
273 ctx
->status
= STARTED
;
276 mutex_unlock(&ctx
->status_mutex
);
280 static long afu_ioctl_process_element(struct cxl_context
*ctx
,
283 pr_devel("%s: pe: %i\n", __func__
, ctx
->pe
);
285 if (copy_to_user(upe
, &ctx
->external_pe
, sizeof(__u32
)))
291 static long afu_ioctl_get_afu_id(struct cxl_context
*ctx
,
292 struct cxl_afu_id __user
*upafuid
)
294 struct cxl_afu_id afuid
= { 0 };
296 afuid
.card_id
= ctx
->afu
->adapter
->adapter_num
;
297 afuid
.afu_offset
= ctx
->afu
->slice
;
298 afuid
.afu_mode
= ctx
->afu
->current_mode
;
300 /* set the flag bit in case the afu is a slave */
301 if (ctx
->afu
->current_mode
== CXL_MODE_DIRECTED
&& !ctx
->master
)
302 afuid
.flags
|= CXL_AFUID_FLAG_SLAVE
;
304 if (copy_to_user(upafuid
, &afuid
, sizeof(afuid
)))
310 long afu_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
312 struct cxl_context
*ctx
= file
->private_data
;
314 if (ctx
->status
== CLOSED
)
317 if (!cxl_ops
->link_ok(ctx
->afu
->adapter
, ctx
->afu
))
320 pr_devel("afu_ioctl\n");
322 case CXL_IOCTL_START_WORK
:
323 return afu_ioctl_start_work(ctx
, (struct cxl_ioctl_start_work __user
*)arg
);
324 case CXL_IOCTL_GET_PROCESS_ELEMENT
:
325 return afu_ioctl_process_element(ctx
, (__u32 __user
*)arg
);
326 case CXL_IOCTL_GET_AFU_ID
:
327 return afu_ioctl_get_afu_id(ctx
, (struct cxl_afu_id __user
*)
333 static long afu_compat_ioctl(struct file
*file
, unsigned int cmd
,
336 return afu_ioctl(file
, cmd
, arg
);
339 int afu_mmap(struct file
*file
, struct vm_area_struct
*vm
)
341 struct cxl_context
*ctx
= file
->private_data
;
343 /* AFU must be started before we can MMIO */
344 if (ctx
->status
!= STARTED
)
347 if (!cxl_ops
->link_ok(ctx
->afu
->adapter
, ctx
->afu
))
350 return cxl_context_iomap(ctx
, vm
);
353 static inline bool ctx_event_pending(struct cxl_context
*ctx
)
355 if (ctx
->pending_irq
|| ctx
->pending_fault
|| ctx
->pending_afu_err
)
358 if (ctx
->afu_driver_ops
&& atomic_read(&ctx
->afu_driver_events
))
364 __poll_t
afu_poll(struct file
*file
, struct poll_table_struct
*poll
)
366 struct cxl_context
*ctx
= file
->private_data
;
371 poll_wait(file
, &ctx
->wq
, poll
);
373 pr_devel("afu_poll wait done pe: %i\n", ctx
->pe
);
375 spin_lock_irqsave(&ctx
->lock
, flags
);
376 if (ctx_event_pending(ctx
))
377 mask
|= EPOLLIN
| EPOLLRDNORM
;
378 else if (ctx
->status
== CLOSED
)
379 /* Only error on closed when there are no futher events pending
382 spin_unlock_irqrestore(&ctx
->lock
, flags
);
384 pr_devel("afu_poll pe: %i returning %#x\n", ctx
->pe
, mask
);
389 static ssize_t
afu_driver_event_copy(struct cxl_context
*ctx
,
391 struct cxl_event
*event
,
392 struct cxl_event_afu_driver_reserved
*pl
)
396 ctx
->afu_driver_ops
->event_delivered(ctx
, pl
, -EINVAL
);
400 /* Check event size */
401 event
->header
.size
+= pl
->data_size
;
402 if (event
->header
.size
> CXL_READ_MIN_SIZE
) {
403 ctx
->afu_driver_ops
->event_delivered(ctx
, pl
, -EINVAL
);
407 /* Copy event header */
408 if (copy_to_user(buf
, event
, sizeof(struct cxl_event_header
))) {
409 ctx
->afu_driver_ops
->event_delivered(ctx
, pl
, -EFAULT
);
413 /* Copy event data */
414 buf
+= sizeof(struct cxl_event_header
);
415 if (copy_to_user(buf
, &pl
->data
, pl
->data_size
)) {
416 ctx
->afu_driver_ops
->event_delivered(ctx
, pl
, -EFAULT
);
420 ctx
->afu_driver_ops
->event_delivered(ctx
, pl
, 0); /* Success */
421 return event
->header
.size
;
424 ssize_t
afu_read(struct file
*file
, char __user
*buf
, size_t count
,
427 struct cxl_context
*ctx
= file
->private_data
;
428 struct cxl_event_afu_driver_reserved
*pl
= NULL
;
429 struct cxl_event event
;
434 if (!cxl_ops
->link_ok(ctx
->afu
->adapter
, ctx
->afu
))
437 if (count
< CXL_READ_MIN_SIZE
)
440 spin_lock_irqsave(&ctx
->lock
, flags
);
443 prepare_to_wait(&ctx
->wq
, &wait
, TASK_INTERRUPTIBLE
);
444 if (ctx_event_pending(ctx
) || (ctx
->status
== CLOSED
))
447 if (!cxl_ops
->link_ok(ctx
->afu
->adapter
, ctx
->afu
)) {
452 if (file
->f_flags
& O_NONBLOCK
) {
457 if (signal_pending(current
)) {
462 spin_unlock_irqrestore(&ctx
->lock
, flags
);
463 pr_devel("afu_read going to sleep...\n");
465 pr_devel("afu_read woken up\n");
466 spin_lock_irqsave(&ctx
->lock
, flags
);
469 finish_wait(&ctx
->wq
, &wait
);
471 memset(&event
, 0, sizeof(event
));
472 event
.header
.process_element
= ctx
->pe
;
473 event
.header
.size
= sizeof(struct cxl_event_header
);
474 if (ctx
->afu_driver_ops
&& atomic_read(&ctx
->afu_driver_events
)) {
475 pr_devel("afu_read delivering AFU driver specific event\n");
476 pl
= ctx
->afu_driver_ops
->fetch_event(ctx
);
477 atomic_dec(&ctx
->afu_driver_events
);
478 event
.header
.type
= CXL_EVENT_AFU_DRIVER
;
479 } else if (ctx
->pending_irq
) {
480 pr_devel("afu_read delivering AFU interrupt\n");
481 event
.header
.size
+= sizeof(struct cxl_event_afu_interrupt
);
482 event
.header
.type
= CXL_EVENT_AFU_INTERRUPT
;
483 event
.irq
.irq
= find_first_bit(ctx
->irq_bitmap
, ctx
->irq_count
) + 1;
484 clear_bit(event
.irq
.irq
- 1, ctx
->irq_bitmap
);
485 if (bitmap_empty(ctx
->irq_bitmap
, ctx
->irq_count
))
486 ctx
->pending_irq
= false;
487 } else if (ctx
->pending_fault
) {
488 pr_devel("afu_read delivering data storage fault\n");
489 event
.header
.size
+= sizeof(struct cxl_event_data_storage
);
490 event
.header
.type
= CXL_EVENT_DATA_STORAGE
;
491 event
.fault
.addr
= ctx
->fault_addr
;
492 event
.fault
.dsisr
= ctx
->fault_dsisr
;
493 ctx
->pending_fault
= false;
494 } else if (ctx
->pending_afu_err
) {
495 pr_devel("afu_read delivering afu error\n");
496 event
.header
.size
+= sizeof(struct cxl_event_afu_error
);
497 event
.header
.type
= CXL_EVENT_AFU_ERROR
;
498 event
.afu_error
.error
= ctx
->afu_err
;
499 ctx
->pending_afu_err
= false;
500 } else if (ctx
->status
== CLOSED
) {
501 pr_devel("afu_read fatal error\n");
502 spin_unlock_irqrestore(&ctx
->lock
, flags
);
505 WARN(1, "afu_read must be buggy\n");
507 spin_unlock_irqrestore(&ctx
->lock
, flags
);
509 if (event
.header
.type
== CXL_EVENT_AFU_DRIVER
)
510 return afu_driver_event_copy(ctx
, buf
, &event
, pl
);
512 if (copy_to_user(buf
, &event
, event
.header
.size
))
514 return event
.header
.size
;
517 finish_wait(&ctx
->wq
, &wait
);
518 spin_unlock_irqrestore(&ctx
->lock
, flags
);
523 * Note: if this is updated, we need to update api.c to patch the new ones in
526 const struct file_operations afu_fops
= {
527 .owner
= THIS_MODULE
,
531 .release
= afu_release
,
532 .unlocked_ioctl
= afu_ioctl
,
533 .compat_ioctl
= afu_compat_ioctl
,
537 static const struct file_operations afu_master_fops
= {
538 .owner
= THIS_MODULE
,
539 .open
= afu_master_open
,
542 .release
= afu_release
,
543 .unlocked_ioctl
= afu_ioctl
,
544 .compat_ioctl
= afu_compat_ioctl
,
549 static char *cxl_devnode(struct device
*dev
, umode_t
*mode
)
551 if (cpu_has_feature(CPU_FTR_HVMODE
) &&
552 CXL_DEVT_IS_CARD(dev
->devt
)) {
554 * These minor numbers will eventually be used to program the
555 * PSL and AFUs once we have dynamic reprogramming support
559 return kasprintf(GFP_KERNEL
, "cxl/%s", dev_name(dev
));
562 extern struct class *cxl_class
;
564 static int cxl_add_chardev(struct cxl_afu
*afu
, dev_t devt
, struct cdev
*cdev
,
565 struct device
**chardev
, char *postfix
, char *desc
,
566 const struct file_operations
*fops
)
571 cdev_init(cdev
, fops
);
572 if ((rc
= cdev_add(cdev
, devt
, 1))) {
573 dev_err(&afu
->dev
, "Unable to add %s chardev: %i\n", desc
, rc
);
577 dev
= device_create(cxl_class
, &afu
->dev
, devt
, afu
,
578 "afu%i.%i%s", afu
->adapter
->adapter_num
, afu
->slice
, postfix
);
580 dev_err(&afu
->dev
, "Unable to create %s chardev in sysfs: %i\n", desc
, rc
);
593 int cxl_chardev_d_afu_add(struct cxl_afu
*afu
)
595 return cxl_add_chardev(afu
, CXL_AFU_MKDEV_D(afu
), &afu
->afu_cdev_d
,
596 &afu
->chardev_d
, "d", "dedicated",
597 &afu_master_fops
); /* Uses master fops */
600 int cxl_chardev_m_afu_add(struct cxl_afu
*afu
)
602 return cxl_add_chardev(afu
, CXL_AFU_MKDEV_M(afu
), &afu
->afu_cdev_m
,
603 &afu
->chardev_m
, "m", "master",
607 int cxl_chardev_s_afu_add(struct cxl_afu
*afu
)
609 return cxl_add_chardev(afu
, CXL_AFU_MKDEV_S(afu
), &afu
->afu_cdev_s
,
610 &afu
->chardev_s
, "s", "shared",
614 void cxl_chardev_afu_remove(struct cxl_afu
*afu
)
616 if (afu
->chardev_d
) {
617 cdev_del(&afu
->afu_cdev_d
);
618 device_unregister(afu
->chardev_d
);
619 afu
->chardev_d
= NULL
;
621 if (afu
->chardev_m
) {
622 cdev_del(&afu
->afu_cdev_m
);
623 device_unregister(afu
->chardev_m
);
624 afu
->chardev_m
= NULL
;
626 if (afu
->chardev_s
) {
627 cdev_del(&afu
->afu_cdev_s
);
628 device_unregister(afu
->chardev_s
);
629 afu
->chardev_s
= NULL
;
633 int cxl_register_afu(struct cxl_afu
*afu
)
635 afu
->dev
.class = cxl_class
;
637 return device_register(&afu
->dev
);
640 int cxl_register_adapter(struct cxl
*adapter
)
642 adapter
->dev
.class = cxl_class
;
645 * Future: When we support dynamically reprogramming the PSL & AFU we
646 * will expose the interface to do that via a chardev:
647 * adapter->dev.devt = CXL_CARD_MKDEV(adapter);
650 return device_register(&adapter
->dev
);
653 dev_t
cxl_get_dev(void)
658 int __init
cxl_file_init(void)
663 * If these change we really need to update API. Either change some
664 * flags or update API version number CXL_API_VERSION.
666 BUILD_BUG_ON(CXL_API_VERSION
!= 3);
667 BUILD_BUG_ON(sizeof(struct cxl_ioctl_start_work
) != 64);
668 BUILD_BUG_ON(sizeof(struct cxl_event_header
) != 8);
669 BUILD_BUG_ON(sizeof(struct cxl_event_afu_interrupt
) != 8);
670 BUILD_BUG_ON(sizeof(struct cxl_event_data_storage
) != 32);
671 BUILD_BUG_ON(sizeof(struct cxl_event_afu_error
) != 16);
673 if ((rc
= alloc_chrdev_region(&cxl_dev
, 0, CXL_NUM_MINORS
, "cxl"))) {
674 pr_err("Unable to allocate CXL major number: %i\n", rc
);
678 pr_devel("CXL device allocated, MAJOR %i\n", MAJOR(cxl_dev
));
680 cxl_class
= class_create(THIS_MODULE
, "cxl");
681 if (IS_ERR(cxl_class
)) {
682 pr_err("Unable to create CXL class\n");
683 rc
= PTR_ERR(cxl_class
);
686 cxl_class
->devnode
= cxl_devnode
;
691 unregister_chrdev_region(cxl_dev
, CXL_NUM_MINORS
);
695 void cxl_file_exit(void)
697 unregister_chrdev_region(cxl_dev
, CXL_NUM_MINORS
);
698 class_destroy(cxl_class
);