2 * Freescale Hypervisor Management Driver
4 * Copyright (C) 2008-2011 Freescale Semiconductor, Inc.
5 * Author: Timur Tabi <timur@freescale.com>
7 * This file is licensed under the terms of the GNU General Public License
8 * version 2. This program is licensed "as is" without any warranty of any
9 * kind, whether express or implied.
11 * The Freescale hypervisor management driver provides several services to
12 * drivers and applications related to the Freescale hypervisor:
14 * 1. An ioctl interface for querying and managing partitions.
16 * 2. A file interface to reading incoming doorbells.
18 * 3. An interrupt handler for shutting down the partition upon receiving the
19 * shutdown doorbell from a manager partition.
21 * 4. A kernel interface for receiving callbacks when a managed partition
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/init.h>
28 #include <linux/types.h>
29 #include <linux/err.h>
31 #include <linux/miscdevice.h>
33 #include <linux/pagemap.h>
34 #include <linux/slab.h>
35 #include <linux/poll.h>
37 #include <linux/reboot.h>
38 #include <linux/uaccess.h>
39 #include <linux/notifier.h>
40 #include <linux/interrupt.h>
43 #include <asm/fsl_hcalls.h>
45 #include <linux/fsl_hypervisor.h>
47 static BLOCKING_NOTIFIER_HEAD(failover_subscribers);
50 * Ioctl interface for FSL_HV_IOCTL_PARTITION_RESTART
52 * Restart a running partition
54 static long ioctl_restart(struct fsl_hv_ioctl_restart __user *p)
56 struct fsl_hv_ioctl_restart param;
58 /* Get the parameters from the user */
59 if (copy_from_user(¶m, p, sizeof(struct fsl_hv_ioctl_restart)))
62 param.ret = fh_partition_restart(param.partition);
64 if (copy_to_user(&p->ret, ¶m.ret, sizeof(__u32)))
71 * Ioctl interface for FSL_HV_IOCTL_PARTITION_STATUS
73 * Query the status of a partition
75 static long ioctl_status(struct fsl_hv_ioctl_status __user *p)
77 struct fsl_hv_ioctl_status param;
80 /* Get the parameters from the user */
81 if (copy_from_user(¶m, p, sizeof(struct fsl_hv_ioctl_status)))
84 param.ret = fh_partition_get_status(param.partition, &status);
86 param.status = status;
88 if (copy_to_user(p, ¶m, sizeof(struct fsl_hv_ioctl_status)))
95 * Ioctl interface for FSL_HV_IOCTL_PARTITION_START
97 * Start a stopped partition.
99 static long ioctl_start(struct fsl_hv_ioctl_start __user *p)
101 struct fsl_hv_ioctl_start param;
103 /* Get the parameters from the user */
104 if (copy_from_user(¶m, p, sizeof(struct fsl_hv_ioctl_start)))
107 param.ret = fh_partition_start(param.partition, param.entry_point,
110 if (copy_to_user(&p->ret, ¶m.ret, sizeof(__u32)))
117 * Ioctl interface for FSL_HV_IOCTL_PARTITION_STOP
119 * Stop a running partition
121 static long ioctl_stop(struct fsl_hv_ioctl_stop __user *p)
123 struct fsl_hv_ioctl_stop param;
125 /* Get the parameters from the user */
126 if (copy_from_user(¶m, p, sizeof(struct fsl_hv_ioctl_stop)))
129 param.ret = fh_partition_stop(param.partition);
131 if (copy_to_user(&p->ret, ¶m.ret, sizeof(__u32)))
138 * Ioctl interface for FSL_HV_IOCTL_MEMCPY
140 * The FH_MEMCPY hypercall takes an array of address/address/size structures
141 * to represent the data being copied. As a convenience to the user, this
142 * ioctl takes a user-create buffer and a pointer to a guest physically
143 * contiguous buffer in the remote partition, and creates the
144 * address/address/size array for the hypercall.
146 static long ioctl_memcpy(struct fsl_hv_ioctl_memcpy __user *p)
148 struct fsl_hv_ioctl_memcpy param;
150 struct page **pages = NULL;
151 void *sg_list_unaligned = NULL;
152 struct fh_sg_list *sg_list = NULL;
154 unsigned int num_pages;
155 unsigned long lb_offset; /* Offset within a page of the local buffer */
159 int num_pinned; /* return value from get_user_pages() */
160 phys_addr_t remote_paddr; /* The next address in the remote buffer */
161 uint32_t count; /* The number of bytes left to copy */
163 /* Get the parameters from the user */
164 if (copy_from_user(¶m, p, sizeof(struct fsl_hv_ioctl_memcpy)))
168 * One partition must be local, the other must be remote. In other
169 * words, if source and target are both -1, or are both not -1, then
172 if ((param.source == -1) == (param.target == -1))
176 * The array of pages returned by get_user_pages() covers only
177 * page-aligned memory. Since the user buffer is probably not
178 * page-aligned, we need to handle the discrepancy.
180 * We calculate the offset within a page of the S/G list, and make
181 * adjustments accordingly. This will result in a page list that looks
184 * ---- <-- first page starts before the buffer
205 * | | <-- last page ends after the buffer
208 * The distance between the start of the first page and the start of the
209 * buffer is lb_offset. The hashed (///) areas are the parts of the
210 * page list that contain the actual buffer.
212 * The advantage of this approach is that the number of pages is
213 * equal to the number of entries in the S/G list that we give to the
216 lb_offset = param.local_vaddr & (PAGE_SIZE - 1);
217 num_pages = (param.count + lb_offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
219 /* Allocate the buffers we need */
222 * 'pages' is an array of struct page pointers that's initialized by
225 pages = kzalloc(num_pages * sizeof(struct page *), GFP_KERNEL);
227 pr_debug("fsl-hv: could not allocate page list\n");
232 * sg_list is the list of fh_sg_list objects that we pass to the
235 sg_list_unaligned = kmalloc(num_pages * sizeof(struct fh_sg_list) +
236 sizeof(struct fh_sg_list) - 1, GFP_KERNEL);
237 if (!sg_list_unaligned) {
238 pr_debug("fsl-hv: could not allocate S/G list\n");
242 sg_list = PTR_ALIGN(sg_list_unaligned, sizeof(struct fh_sg_list));
244 /* Get the physical addresses of the source buffer */
245 down_read(¤t->mm->mmap_sem);
246 num_pinned = get_user_pages(current, current->mm,
247 param.local_vaddr - lb_offset, num_pages,
248 (param.source == -1) ? READ : WRITE,
250 up_read(¤t->mm->mmap_sem);
252 if (num_pinned != num_pages) {
253 /* get_user_pages() failed */
254 pr_debug("fsl-hv: could not lock source buffer\n");
255 ret = (num_pinned < 0) ? num_pinned : -EFAULT;
260 * Build the fh_sg_list[] array. The first page is special
261 * because it's misaligned.
263 if (param.source == -1) {
264 sg_list[0].source = page_to_phys(pages[0]) + lb_offset;
265 sg_list[0].target = param.remote_paddr;
267 sg_list[0].source = param.remote_paddr;
268 sg_list[0].target = page_to_phys(pages[0]) + lb_offset;
270 sg_list[0].size = min_t(uint64_t, param.count, PAGE_SIZE - lb_offset);
272 remote_paddr = param.remote_paddr + sg_list[0].size;
273 count = param.count - sg_list[0].size;
275 for (i = 1; i < num_pages; i++) {
276 if (param.source == -1) {
277 /* local to remote */
278 sg_list[i].source = page_to_phys(pages[i]);
279 sg_list[i].target = remote_paddr;
281 /* remote to local */
282 sg_list[i].source = remote_paddr;
283 sg_list[i].target = page_to_phys(pages[i]);
285 sg_list[i].size = min_t(uint64_t, count, PAGE_SIZE);
287 remote_paddr += sg_list[i].size;
288 count -= sg_list[i].size;
291 param.ret = fh_partition_memcpy(param.source, param.target,
292 virt_to_phys(sg_list), num_pages);
296 for (i = 0; i < num_pages; i++)
301 kfree(sg_list_unaligned);
305 if (copy_to_user(&p->ret, ¶m.ret, sizeof(__u32)))
312 * Ioctl interface for FSL_HV_IOCTL_DOORBELL
316 static long ioctl_doorbell(struct fsl_hv_ioctl_doorbell __user *p)
318 struct fsl_hv_ioctl_doorbell param;
320 /* Get the parameters from the user. */
321 if (copy_from_user(¶m, p, sizeof(struct fsl_hv_ioctl_doorbell)))
324 param.ret = ev_doorbell_send(param.doorbell);
326 if (copy_to_user(&p->ret, ¶m.ret, sizeof(__u32)))
332 static long ioctl_dtprop(struct fsl_hv_ioctl_prop __user *p, int set)
334 struct fsl_hv_ioctl_prop param;
335 char __user *upath, *upropname;
336 void __user *upropval;
337 char *path = NULL, *propname = NULL;
338 void *propval = NULL;
341 /* Get the parameters from the user. */
342 if (copy_from_user(¶m, p, sizeof(struct fsl_hv_ioctl_prop)))
345 upath = (char __user *)(uintptr_t)param.path;
346 upropname = (char __user *)(uintptr_t)param.propname;
347 upropval = (void __user *)(uintptr_t)param.propval;
349 path = strndup_user(upath, FH_DTPROP_MAX_PATHLEN);
355 propname = strndup_user(upropname, FH_DTPROP_MAX_PATHLEN);
356 if (IS_ERR(propname)) {
357 ret = PTR_ERR(propname);
361 if (param.proplen > FH_DTPROP_MAX_PROPLEN) {
366 propval = kmalloc(param.proplen, GFP_KERNEL);
373 if (copy_from_user(propval, upropval, param.proplen)) {
378 param.ret = fh_partition_set_dtprop(param.handle,
380 virt_to_phys(propname),
381 virt_to_phys(propval),
384 param.ret = fh_partition_get_dtprop(param.handle,
386 virt_to_phys(propname),
387 virt_to_phys(propval),
390 if (param.ret == 0) {
391 if (copy_to_user(upropval, propval, param.proplen) ||
392 put_user(param.proplen, &p->proplen)) {
399 if (put_user(param.ret, &p->ret))
411 * Ioctl main entry point
413 static long fsl_hv_ioctl(struct file *file, unsigned int cmd,
414 unsigned long argaddr)
416 void __user *arg = (void __user *)argaddr;
420 case FSL_HV_IOCTL_PARTITION_RESTART:
421 ret = ioctl_restart(arg);
423 case FSL_HV_IOCTL_PARTITION_GET_STATUS:
424 ret = ioctl_status(arg);
426 case FSL_HV_IOCTL_PARTITION_START:
427 ret = ioctl_start(arg);
429 case FSL_HV_IOCTL_PARTITION_STOP:
430 ret = ioctl_stop(arg);
432 case FSL_HV_IOCTL_MEMCPY:
433 ret = ioctl_memcpy(arg);
435 case FSL_HV_IOCTL_DOORBELL:
436 ret = ioctl_doorbell(arg);
438 case FSL_HV_IOCTL_GETPROP:
439 ret = ioctl_dtprop(arg, 0);
441 case FSL_HV_IOCTL_SETPROP:
442 ret = ioctl_dtprop(arg, 1);
445 pr_debug("fsl-hv: bad ioctl dir=%u type=%u cmd=%u size=%u\n",
446 _IOC_DIR(cmd), _IOC_TYPE(cmd), _IOC_NR(cmd),
454 /* Linked list of processes that have us open */
455 static struct list_head db_list;
457 /* spinlock for db_list */
458 static DEFINE_SPINLOCK(db_list_lock);
460 /* The size of the doorbell event queue. This must be a power of two. */
463 /* Returns the next head/tail pointer, wrapping around the queue if necessary */
464 #define nextp(x) (((x) + 1) & (QSIZE - 1))
466 /* Per-open data structure */
467 struct doorbell_queue {
468 struct list_head list;
470 wait_queue_head_t wait;
476 /* Linked list of ISRs that we registered */
477 struct list_head isr_list;
479 /* Per-ISR data structure */
480 struct doorbell_isr {
481 struct list_head list;
483 uint32_t doorbell; /* The doorbell handle */
484 uint32_t partition; /* The partition handle, if used */
488 * Add a doorbell to all of the doorbell queues
490 static void fsl_hv_queue_doorbell(uint32_t doorbell)
492 struct doorbell_queue *dbq;
495 /* Prevent another core from modifying db_list */
496 spin_lock_irqsave(&db_list_lock, flags);
498 list_for_each_entry(dbq, &db_list, list) {
499 if (dbq->head != nextp(dbq->tail)) {
500 dbq->q[dbq->tail] = doorbell;
502 * This memory barrier eliminates the need to grab
503 * the spinlock for dbq.
506 dbq->tail = nextp(dbq->tail);
507 wake_up_interruptible(&dbq->wait);
511 spin_unlock_irqrestore(&db_list_lock, flags);
515 * Interrupt handler for all doorbells
517 * We use the same interrupt handler for all doorbells. Whenever a doorbell
518 * is rung, and we receive an interrupt, we just put the handle for that
519 * doorbell (passed to us as *data) into all of the queues.
521 static irqreturn_t fsl_hv_isr(int irq, void *data)
523 fsl_hv_queue_doorbell((uintptr_t) data);
529 * State change thread function
531 * The state change notification arrives in an interrupt, but we can't call
532 * blocking_notifier_call_chain() in an interrupt handler. We could call
533 * atomic_notifier_call_chain(), but that would require the clients' call-back
534 * function to run in interrupt context. Since we don't want to impose that
535 * restriction on the clients, we use a threaded IRQ to process the
536 * notification in kernel context.
538 static irqreturn_t fsl_hv_state_change_thread(int irq, void *data)
540 struct doorbell_isr *dbisr = data;
542 blocking_notifier_call_chain(&failover_subscribers, dbisr->partition,
549 * Interrupt handler for state-change doorbells
551 static irqreturn_t fsl_hv_state_change_isr(int irq, void *data)
554 struct doorbell_isr *dbisr = data;
557 /* It's still a doorbell, so add it to all the queues. */
558 fsl_hv_queue_doorbell(dbisr->doorbell);
560 /* Determine the new state, and if it's stopped, notify the clients. */
561 ret = fh_partition_get_status(dbisr->partition, &status);
562 if (!ret && (status == FH_PARTITION_STOPPED))
563 return IRQ_WAKE_THREAD;
569 * Returns a bitmask indicating whether a read will block
571 static unsigned int fsl_hv_poll(struct file *filp, struct poll_table_struct *p)
573 struct doorbell_queue *dbq = filp->private_data;
577 spin_lock_irqsave(&dbq->lock, flags);
579 poll_wait(filp, &dbq->wait, p);
580 mask = (dbq->head == dbq->tail) ? 0 : (POLLIN | POLLRDNORM);
582 spin_unlock_irqrestore(&dbq->lock, flags);
588 * Return the handles for any incoming doorbells
590 * If there are doorbell handles in the queue for this open instance, then
591 * return them to the caller as an array of 32-bit integers. Otherwise,
592 * block until there is at least one handle to return.
594 static ssize_t fsl_hv_read(struct file *filp, char __user *buf, size_t len,
597 struct doorbell_queue *dbq = filp->private_data;
598 uint32_t __user *p = (uint32_t __user *) buf; /* for put_user() */
602 /* Make sure we stop when the user buffer is full. */
603 while (len >= sizeof(uint32_t)) {
604 uint32_t dbell; /* Local copy of doorbell queue data */
606 spin_lock_irqsave(&dbq->lock, flags);
609 * If the queue is empty, then either we're done or we need
610 * to block. If the application specified O_NONBLOCK, then
611 * we return the appropriate error code.
613 if (dbq->head == dbq->tail) {
614 spin_unlock_irqrestore(&dbq->lock, flags);
617 if (filp->f_flags & O_NONBLOCK)
619 if (wait_event_interruptible(dbq->wait,
620 dbq->head != dbq->tail))
626 * Even though we have an smp_wmb() in the ISR, the core
627 * might speculatively execute the "dbell = ..." below while
628 * it's evaluating the if-statement above. In that case, the
629 * value put into dbell could be stale if the core accepts the
630 * speculation. To prevent that, we need a read memory barrier
635 /* Copy the data to a temporary local buffer, because
636 * we can't call copy_to_user() from inside a spinlock
638 dbell = dbq->q[dbq->head];
639 dbq->head = nextp(dbq->head);
641 spin_unlock_irqrestore(&dbq->lock, flags);
643 if (put_user(dbell, p))
646 count += sizeof(uint32_t);
647 len -= sizeof(uint32_t);
654 * Open the driver and prepare for reading doorbells.
656 * Every time an application opens the driver, we create a doorbell queue
657 * for that file handle. This queue is used for any incoming doorbells.
659 static int fsl_hv_open(struct inode *inode, struct file *filp)
661 struct doorbell_queue *dbq;
665 dbq = kzalloc(sizeof(struct doorbell_queue), GFP_KERNEL);
667 pr_err("fsl-hv: out of memory\n");
671 spin_lock_init(&dbq->lock);
672 init_waitqueue_head(&dbq->wait);
674 spin_lock_irqsave(&db_list_lock, flags);
675 list_add(&dbq->list, &db_list);
676 spin_unlock_irqrestore(&db_list_lock, flags);
678 filp->private_data = dbq;
686 static int fsl_hv_close(struct inode *inode, struct file *filp)
688 struct doorbell_queue *dbq = filp->private_data;
693 spin_lock_irqsave(&db_list_lock, flags);
694 list_del(&dbq->list);
695 spin_unlock_irqrestore(&db_list_lock, flags);
702 static const struct file_operations fsl_hv_fops = {
703 .owner = THIS_MODULE,
705 .release = fsl_hv_close,
708 .unlocked_ioctl = fsl_hv_ioctl,
709 .compat_ioctl = fsl_hv_ioctl,
712 static struct miscdevice fsl_hv_misc_dev = {
718 static irqreturn_t fsl_hv_shutdown_isr(int irq, void *data)
720 orderly_poweroff(false);
726 * Returns the handle of the parent of the given node
728 * The handle is the value of the 'hv-handle' property
730 static int get_parent_handle(struct device_node *np)
732 struct device_node *parent;
733 const uint32_t *prop;
737 parent = of_get_parent(np);
739 /* It's not really possible for this to fail */
743 * The proper name for the handle property is "hv-handle", but some
744 * older versions of the hypervisor used "reg".
746 prop = of_get_property(parent, "hv-handle", &len);
748 prop = of_get_property(parent, "reg", &len);
750 if (!prop || (len != sizeof(uint32_t))) {
751 /* This can happen only if the node is malformed */
756 handle = be32_to_cpup(prop);
763 * Register a callback for failover events
765 * This function is called by device drivers to register their callback
766 * functions for fail-over events.
768 int fsl_hv_failover_register(struct notifier_block *nb)
770 return blocking_notifier_chain_register(&failover_subscribers, nb);
772 EXPORT_SYMBOL(fsl_hv_failover_register);
775 * Unregister a callback for failover events
777 int fsl_hv_failover_unregister(struct notifier_block *nb)
779 return blocking_notifier_chain_unregister(&failover_subscribers, nb);
781 EXPORT_SYMBOL(fsl_hv_failover_unregister);
784 * Return TRUE if we're running under FSL hypervisor
786 * This function checks to see if we're running under the Freescale
787 * hypervisor, and returns zero if we're not, or non-zero if we are.
789 * First, it checks if MSR[GS]==1, which means we're running under some
790 * hypervisor. Then it checks if there is a hypervisor node in the device
791 * tree. Currently, that means there needs to be a node in the root called
792 * "hypervisor" and which has a property named "fsl,hv-version".
794 static int has_fsl_hypervisor(void)
796 struct device_node *node;
799 node = of_find_node_by_path("/hypervisor");
803 ret = of_find_property(node, "fsl,hv-version", NULL) != NULL;
811 * Freescale hypervisor management driver init
813 * This function is called when this module is loaded.
815 * Register ourselves as a miscellaneous driver. This will register the
816 * fops structure and create the right sysfs entries for udev.
818 static int __init fsl_hypervisor_init(void)
820 struct device_node *np;
821 struct doorbell_isr *dbisr, *n;
824 pr_info("Freescale hypervisor management driver\n");
826 if (!has_fsl_hypervisor()) {
827 pr_info("fsl-hv: no hypervisor found\n");
831 ret = misc_register(&fsl_hv_misc_dev);
833 pr_err("fsl-hv: cannot register device\n");
837 INIT_LIST_HEAD(&db_list);
838 INIT_LIST_HEAD(&isr_list);
840 for_each_compatible_node(np, NULL, "epapr,hv-receive-doorbell") {
842 const uint32_t *handle;
844 handle = of_get_property(np, "interrupts", NULL);
845 irq = irq_of_parse_and_map(np, 0);
846 if (!handle || (irq == NO_IRQ)) {
847 pr_err("fsl-hv: no 'interrupts' property in %s node\n",
852 dbisr = kzalloc(sizeof(*dbisr), GFP_KERNEL);
857 dbisr->doorbell = be32_to_cpup(handle);
859 if (of_device_is_compatible(np, "fsl,hv-shutdown-doorbell")) {
860 /* The shutdown doorbell gets its own ISR */
861 ret = request_irq(irq, fsl_hv_shutdown_isr, 0,
863 } else if (of_device_is_compatible(np,
864 "fsl,hv-state-change-doorbell")) {
866 * The state change doorbell triggers a notification if
867 * the state of the managed partition changes to
868 * "stopped". We need a separate interrupt handler for
869 * that, and we also need to know the handle of the
870 * target partition, not just the handle of the
873 dbisr->partition = ret = get_parent_handle(np);
875 pr_err("fsl-hv: node %s has missing or "
876 "malformed parent\n", np->full_name);
880 ret = request_threaded_irq(irq, fsl_hv_state_change_isr,
881 fsl_hv_state_change_thread,
884 ret = request_irq(irq, fsl_hv_isr, 0, np->name, dbisr);
887 pr_err("fsl-hv: could not request irq %u for node %s\n",
893 list_add(&dbisr->list, &isr_list);
895 pr_info("fsl-hv: registered handler for doorbell %u\n",
902 list_for_each_entry_safe(dbisr, n, &isr_list, list) {
903 free_irq(dbisr->irq, dbisr);
904 list_del(&dbisr->list);
908 misc_deregister(&fsl_hv_misc_dev);
914 * Freescale hypervisor management driver termination
916 * This function is called when this driver is unloaded.
918 static void __exit fsl_hypervisor_exit(void)
920 struct doorbell_isr *dbisr, *n;
922 list_for_each_entry_safe(dbisr, n, &isr_list, list) {
923 free_irq(dbisr->irq, dbisr);
924 list_del(&dbisr->list);
928 misc_deregister(&fsl_hv_misc_dev);
931 module_init(fsl_hypervisor_init);
932 module_exit(fsl_hypervisor_exit);
934 MODULE_AUTHOR("Timur Tabi <timur@freescale.com>");
935 MODULE_DESCRIPTION("Freescale hypervisor management driver");
936 MODULE_LICENSE("GPL v2");