[deliverable/linux.git] / arch / x86 / mm / kmmio.c

/* Support for MMIO probes.
 * Benfit many code from kprobes
 * (C) 2002 Louis Zhuang <louis.zhuang@intel.com>.
 *     2007 Alexander Eichner
 *     2008 Pekka Paalanen <pq@iki.fi>
 */

#include <linux/list.h>
#include <linux/spinlock.h>
#include <linux/hash.h>
#include <linux/init.h>
#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/uaccess.h>
#include <linux/ptrace.h>
#include <linux/preempt.h>
#include <linux/percpu.h>
#include <linux/kdebug.h>
#include <linux/mutex.h>
#include <linux/io.h>
#include <asm/cacheflush.h>
#include <asm/tlbflush.h>
#include <linux/errno.h>
#include <asm/debugreg.h>
#include <linux/mmiotrace.h>

#define KMMIO_PAGE_HASH_BITS 4
#define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)

struct kmmio_fault_page {
	struct list_head list;
	struct kmmio_fault_page *release_next;
	unsigned long page; /* location of the fault page */

	/*
	 * Number of times this page has been registered as a part
	 * of a probe. If zero, page is disarmed and this may be freed.
	 * Used only by writers (RCU).
	 */
	int count;
};

struct kmmio_delayed_release {
	struct rcu_head rcu;
	struct kmmio_fault_page *release_list;
};

struct kmmio_context {
	struct kmmio_fault_page *fpage;
	struct kmmio_probe *probe;
	unsigned long saved_flags;
	unsigned long addr;
	int active;
};

static DEFINE_SPINLOCK(kmmio_lock);

/* Protected by kmmio_lock */
unsigned int kmmio_count;

/* Read-protected by RCU, write-protected by kmmio_lock. */
static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
static LIST_HEAD(kmmio_probes);

static struct list_head *kmmio_page_list(unsigned long page)
{
	return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)];
}

/* Accessed per-cpu */
static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);

/*
 * this is basically a dynamic stabbing problem:
 * Could use the existing prio tree code or
 * Possible better implementations:
 * The Interval Skip List: A Data Structure for Finding All Intervals That
 * Overlap a Point (might be simple)
 * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
 */
/* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
{
	struct kmmio_probe *p;
	list_for_each_entry_rcu(p, &kmmio_probes, list) {
		if (addr >= p->addr && addr <= (p->addr + p->len))
			return p;
	}
	return NULL;
}

/* You must be holding RCU read lock. */
static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page)
{
	struct list_head *head;
	struct kmmio_fault_page *p;

	page &= PAGE_MASK;
	head = kmmio_page_list(page);
	list_for_each_entry_rcu(p, head, list) {
		if (p->page == page)
			return p;
	}
	return NULL;
}

static void set_page_present(unsigned long addr, bool present,
							unsigned int *pglevel)
{
	pteval_t pteval;
	pmdval_t pmdval;
	unsigned int level;
	pmd_t *pmd;
	pte_t *pte = lookup_address(addr, &level);

	if (!pte) {
		pr_err("kmmio: no pte for page 0x%08lx\n", addr);
		return;
	}

	if (pglevel)
		*pglevel = level;

	switch (level) {
	case PG_LEVEL_2M:
		pmd = (pmd_t *)pte;
		pmdval = pmd_val(*pmd) & ~_PAGE_PRESENT;
		if (present)
			pmdval |= _PAGE_PRESENT;
		set_pmd(pmd, __pmd(pmdval));
		break;

	case PG_LEVEL_4K:
		pteval = pte_val(*pte) & ~_PAGE_PRESENT;
		if (present)
			pteval |= _PAGE_PRESENT;
		set_pte_atomic(pte, __pte(pteval));
		break;

	default:
		pr_err("kmmio: unexpected page level 0x%x.\n", level);
		return;
	}

	__flush_tlb_one(addr);
}

/** Mark the given page as not present. Access to it will trigger a fault. */
static void arm_kmmio_fault_page(unsigned long page, unsigned int *pglevel)
{
	set_page_present(page & PAGE_MASK, false, pglevel);
}

/** Mark the given page as present. */
static void disarm_kmmio_fault_page(unsigned long page, unsigned int *pglevel)
{
	set_page_present(page & PAGE_MASK, true, pglevel);
}

/*
 * This is being called from do_page_fault().
 *
 * We may be in an interrupt or a critical section. Also prefecthing may
 * trigger a page fault. We may be in the middle of process switch.
 * We cannot take any locks, because we could be executing especially
 * within a kmmio critical section.
 *
 * Local interrupts are disabled, so preemption cannot happen.
 * Do not enable interrupts, do not sleep, and watch out for other CPUs.
 */
/*
 * Interrupts are disabled on entry as trap3 is an interrupt gate
 * and they remain disabled thorough out this function.
 */
int kmmio_handler(struct pt_regs *regs, unsigned long addr)
{
	struct kmmio_context *ctx;
	struct kmmio_fault_page *faultpage;
	int ret = 0; /* default to fault not handled */

	/*
	 * Preemption is now disabled to prevent process switch during
	 * single stepping. We can only handle one active kmmio trace
	 * per cpu, so ensure that we finish it before something else
	 * gets to run. We also hold the RCU read lock over single
	 * stepping to avoid looking up the probe and kmmio_fault_page
	 * again.
	 */
	preempt_disable();
	rcu_read_lock();

	faultpage = get_kmmio_fault_page(addr);
	if (!faultpage) {
		/*
		 * Either this page fault is not caused by kmmio, or
		 * another CPU just pulled the kmmio probe from under
		 * our feet. The latter case should not be possible.
		 */
		goto no_kmmio;
	}

	ctx = &get_cpu_var(kmmio_ctx);
	if (ctx->active) {
		disarm_kmmio_fault_page(faultpage->page, NULL);
		if (addr == ctx->addr) {
			/*
			 * On SMP we sometimes get recursive probe hits on the
			 * same address. Context is already saved, fall out.
			 */
			pr_debug("kmmio: duplicate probe hit on CPU %d, for "
						"address 0x%08lx.\n",
						smp_processor_id(), addr);
			ret = 1;
			goto no_kmmio_ctx;
		}
		/*
		 * Prevent overwriting already in-flight context.
		 * This should not happen, let's hope disarming at least
		 * prevents a panic.
		 */
		pr_emerg("kmmio: recursive probe hit on CPU %d, "
					"for address 0x%08lx. Ignoring.\n",
					smp_processor_id(), addr);
		pr_emerg("kmmio: previous hit was at 0x%08lx.\n",
					ctx->addr);
		goto no_kmmio_ctx;
	}
	ctx->active++;

	ctx->fpage = faultpage;
	ctx->probe = get_kmmio_probe(addr);
	ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
	ctx->addr = addr;

	if (ctx->probe && ctx->probe->pre_handler)
		ctx->probe->pre_handler(ctx->probe, regs, addr);

	/*
	 * Enable single-stepping and disable interrupts for the faulting
	 * context. Local interrupts must not get enabled during stepping.
	 */
	regs->flags |= X86_EFLAGS_TF;
	regs->flags &= ~X86_EFLAGS_IF;

	/* Now we set present bit in PTE and single step. */
	disarm_kmmio_fault_page(ctx->fpage->page, NULL);

	/*
	 * If another cpu accesses the same page while we are stepping,
	 * the access will not be caught. It will simply succeed and the
	 * only downside is we lose the event. If this becomes a problem,
	 * the user should drop to single cpu before tracing.
	 */

	put_cpu_var(kmmio_ctx);
	return 1; /* fault handled */

no_kmmio_ctx:
	put_cpu_var(kmmio_ctx);
no_kmmio:
	rcu_read_unlock();
	preempt_enable_no_resched();
	return ret;
}

/*
 * Interrupts are disabled on entry as trap1 is an interrupt gate
 * and they remain disabled thorough out this function.
 * This must always get called as the pair to kmmio_handler().
 */
static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
{
	int ret = 0;
	struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);

	if (!ctx->active) {
		pr_debug("kmmio: spurious debug trap on CPU %d.\n",
							smp_processor_id());
		goto out;
	}

	if (ctx->probe && ctx->probe->post_handler)
		ctx->probe->post_handler(ctx->probe, condition, regs);

	arm_kmmio_fault_page(ctx->fpage->page, NULL);

	regs->flags &= ~X86_EFLAGS_TF;
	regs->flags |= ctx->saved_flags;

	/* These were acquired in kmmio_handler(). */
	ctx->active--;
	BUG_ON(ctx->active);
	rcu_read_unlock();
	preempt_enable_no_resched();

	/*
	 * if somebody else is singlestepping across a probe point, flags
	 * will have TF set, in which case, continue the remaining processing
	 * of do_debug, as if this is not a probe hit.
	 */
	if (!(regs->flags & X86_EFLAGS_TF))
		ret = 1;
out:
	put_cpu_var(kmmio_ctx);
	return ret;
}

/* You must be holding kmmio_lock. */
static int add_kmmio_fault_page(unsigned long page)
{
	struct kmmio_fault_page *f;

	page &= PAGE_MASK;
	f = get_kmmio_fault_page(page);
	if (f) {
		if (!f->count)
			arm_kmmio_fault_page(f->page, NULL);
		f->count++;
		return 0;
	}

	f = kmalloc(sizeof(*f), GFP_ATOMIC);
	if (!f)
		return -1;

	f->count = 1;
	f->page = page;
	list_add_rcu(&f->list, kmmio_page_list(f->page));

	arm_kmmio_fault_page(f->page, NULL);

	return 0;
}

/* You must be holding kmmio_lock. */
static void release_kmmio_fault_page(unsigned long page,
				struct kmmio_fault_page **release_list)
{
	struct kmmio_fault_page *f;

	page &= PAGE_MASK;
	f = get_kmmio_fault_page(page);
	if (!f)
		return;

	f->count--;
	BUG_ON(f->count < 0);
	if (!f->count) {
		disarm_kmmio_fault_page(f->page, NULL);
		f->release_next = *release_list;
		*release_list = f;
	}
}

/*
 * With page-unaligned ioremaps, one or two armed pages may contain
 * addresses from outside the intended mapping. Events for these addresses
 * are currently silently dropped. The events may result only from programming
 * mistakes by accessing addresses before the beginning or past the end of a
 * mapping.
 */
int register_kmmio_probe(struct kmmio_probe *p)
{
	unsigned long flags;
	int ret = 0;
	unsigned long size = 0;
	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);

	spin_lock_irqsave(&kmmio_lock, flags);
	if (get_kmmio_probe(p->addr)) {
		ret = -EEXIST;
		goto out;
	}
	kmmio_count++;
	list_add_rcu(&p->list, &kmmio_probes);
	while (size < size_lim) {
		if (add_kmmio_fault_page(p->addr + size))
			pr_err("kmmio: Unable to set page fault.\n");
		size += PAGE_SIZE;
	}
out:
	spin_unlock_irqrestore(&kmmio_lock, flags);
	/*
	 * XXX: What should I do here?
	 * Here was a call to global_flush_tlb(), but it does not exist
	 * anymore. It seems it's not needed after all.
	 */
	return ret;
}
EXPORT_SYMBOL(register_kmmio_probe);

static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
{
	struct kmmio_delayed_release *dr = container_of(
						head,
						struct kmmio_delayed_release,
						rcu);
	struct kmmio_fault_page *p = dr->release_list;
	while (p) {
		struct kmmio_fault_page *next = p->release_next;
		BUG_ON(p->count);
		kfree(p);
		p = next;
	}
	kfree(dr);
}

static void remove_kmmio_fault_pages(struct rcu_head *head)
{
	struct kmmio_delayed_release *dr = container_of(
						head,
						struct kmmio_delayed_release,
						rcu);
	struct kmmio_fault_page *p = dr->release_list;
	struct kmmio_fault_page **prevp = &dr->release_list;
	unsigned long flags;
	spin_lock_irqsave(&kmmio_lock, flags);
	while (p) {
		if (!p->count)
			list_del_rcu(&p->list);
		else
			*prevp = p->release_next;
		prevp = &p->release_next;
		p = p->release_next;
	}
	spin_unlock_irqrestore(&kmmio_lock, flags);
	/* This is the real RCU destroy call. */
	call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
}

/*
 * Remove a kmmio probe. You have to synchronize_rcu() before you can be
 * sure that the callbacks will not be called anymore. Only after that
 * you may actually release your struct kmmio_probe.
 *
 * Unregistering a kmmio fault page has three steps:
 * 1. release_kmmio_fault_page()
 *    Disarm the page, wait a grace period to let all faults finish.
 * 2. remove_kmmio_fault_pages()
 *    Remove the pages from kmmio_page_table.
 * 3. rcu_free_kmmio_fault_pages()
 *    Actally free the kmmio_fault_page structs as with RCU.
 */
void unregister_kmmio_probe(struct kmmio_probe *p)
{
	unsigned long flags;
	unsigned long size = 0;
	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
	struct kmmio_fault_page *release_list = NULL;
	struct kmmio_delayed_release *drelease;

	spin_lock_irqsave(&kmmio_lock, flags);
	while (size < size_lim) {
		release_kmmio_fault_page(p->addr + size, &release_list);
		size += PAGE_SIZE;
	}
	list_del_rcu(&p->list);
	kmmio_count--;
	spin_unlock_irqrestore(&kmmio_lock, flags);

	drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
	if (!drelease) {
		pr_crit("kmmio: leaking kmmio_fault_page objects.\n");
		return;
	}
	drelease->release_list = release_list;

	/*
	 * This is not really RCU here. We have just disarmed a set of
	 * pages so that they cannot trigger page faults anymore. However,
	 * we cannot remove the pages from kmmio_page_table,
	 * because a probe hit might be in flight on another CPU. The
	 * pages are collected into a list, and they will be removed from
	 * kmmio_page_table when it is certain that no probe hit related to
	 * these pages can be in flight. RCU grace period sounds like a
	 * good choice.
	 *
	 * If we removed the pages too early, kmmio page fault handler might
	 * not find the respective kmmio_fault_page and determine it's not
	 * a kmmio fault, when it actually is. This would lead to madness.
	 */
	call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
}
EXPORT_SYMBOL(unregister_kmmio_probe);

static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
								void *args)
{
	struct die_args *arg = args;

	if (val == DIE_DEBUG && (arg->err & DR_STEP))
		if (post_kmmio_handler(arg->err, arg->regs) == 1)
			return NOTIFY_STOP;

	return NOTIFY_DONE;
}

static struct notifier_block nb_die = {
	.notifier_call = kmmio_die_notifier
};

static int __init init_kmmio(void)
{
	int i;
	for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
		INIT_LIST_HEAD(&kmmio_page_table[i]);
	return register_die_notifier(&nb_die);
}
fs_initcall(init_kmmio); /* should be before device_initcall() */
Commit	Line	Data
8b7d89d0 PP	1	/* Support for MMIO probes.
	2	* Benfit many code from kprobes
	3	* (C) 2002 Louis Zhuang <louis.zhuang@intel.com>.
	4	* 2007 Alexander Eichner
	5	* 2008 Pekka Paalanen <pq@iki.fi>
	6	*/
	7
0fd0e3da	8	#include <linux/list.h>
8b7d89d0 PP	9	#include <linux/spinlock.h>
	10	#include <linux/hash.h>
	11	#include <linux/init.h>
	12	#include <linux/module.h>
8b7d89d0	13	#include <linux/kernel.h>
8b7d89d0 PP	14	#include <linux/uaccess.h>
	15	#include <linux/ptrace.h>
	16	#include <linux/preempt.h>
f5136380	17	#include <linux/percpu.h>
0fd0e3da	18	#include <linux/kdebug.h>
d61fc448	19	#include <linux/mutex.h>
970e6fa0	20	#include <linux/io.h>
8b7d89d0	21	#include <asm/cacheflush.h>
8b7d89d0	22	#include <asm/tlbflush.h>
970e6fa0	23	#include <linux/errno.h>
13829537	24	#include <asm/debugreg.h>
0fd0e3da	25	#include <linux/mmiotrace.h>
8b7d89d0	26
8b7d89d0 PP	27	#define KMMIO_PAGE_HASH_BITS 4
	28	#define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)
	29
0fd0e3da PP	30	struct kmmio_fault_page {
	31	struct list_head list;
	32	struct kmmio_fault_page *release_next;
	33	unsigned long page; /* location of the fault page */
	34
	35	/*
	36	* Number of times this page has been registered as a part
	37	* of a probe. If zero, page is disarmed and this may be freed.
	38	* Used only by writers (RCU).
	39	*/
	40	int count;
	41	};
	42
	43	struct kmmio_delayed_release {
	44	struct rcu_head rcu;
	45	struct kmmio_fault_page *release_list;
	46	};
	47
8b7d89d0 PP	48	struct kmmio_context {
	49	struct kmmio_fault_page *fpage;
	50	struct kmmio_probe *probe;
	51	unsigned long saved_flags;
0fd0e3da	52	unsigned long addr;
8b7d89d0 PP	53	int active;
	54	};
	55
8b7d89d0 PP	56	static DEFINE_SPINLOCK(kmmio_lock);
8b7d89d0 PP	57
13829537	58	/* Protected by kmmio_lock */
8b7d89d0	59	unsigned int kmmio_count;
0fd0e3da PP	60
0fd0e3da PP	61	/* Read-protected by RCU, write-protected by kmmio_lock. */
8b7d89d0 PP	62	static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
	63	static LIST_HEAD(kmmio_probes);
	64
0fd0e3da PP	65	static struct list_head *kmmio_page_list(unsigned long page)
	66	{
	67	return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)];
	68	}
	69
f5136380 PP	70	/* Accessed per-cpu */
f5136380 PP	71	static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
8b7d89d0	72
8b7d89d0 PP	73	/*
	74	* this is basically a dynamic stabbing problem:
	75	* Could use the existing prio tree code or
	76	* Possible better implementations:
	77	* The Interval Skip List: A Data Structure for Finding All Intervals That
	78	* Overlap a Point (might be simple)
	79	* Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
	80	*/
0fd0e3da	81	/* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
8b7d89d0 PP	82	static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
	83	{
	84	struct kmmio_probe *p;
0fd0e3da	85	list_for_each_entry_rcu(p, &kmmio_probes, list) {
8b7d89d0 PP	86	if (addr >= p->addr && addr <= (p->addr + p->len))
	87	return p;
	88	}
	89	return NULL;
	90	}
	91
0fd0e3da	92	/* You must be holding RCU read lock. */
8b7d89d0 PP	93	static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page)
8b7d89d0 PP	94	{
0fd0e3da PP	95	struct list_head *head;
0fd0e3da PP	96	struct kmmio_fault_page *p;
8b7d89d0 PP	97
8b7d89d0 PP	98	page &= PAGE_MASK;
0fd0e3da PP	99	head = kmmio_page_list(page);
0fd0e3da PP	100	list_for_each_entry_rcu(p, head, list) {
8b7d89d0 PP	101	if (p->page == page)
	102	return p;
	103	}
8b7d89d0 PP	104	return NULL;
	105	}
	106
790e2a29 PP	107	static void set_page_present(unsigned long addr, bool present,
790e2a29 PP	108	unsigned int *pglevel)
8b7d89d0	109	{
13829537 PP	110	pteval_t pteval;
13829537 PP	111	pmdval_t pmdval;
790e2a29	112	unsigned int level;
13829537 PP	113	pmd_t *pmd;
13829537 PP	114	pte_t *pte = lookup_address(addr, &level);
8b7d89d0	115
75bb8835	116	if (!pte) {
13829537	117	pr_err("kmmio: no pte for page 0x%08lx\n", addr);
75bb8835 PP	118	return;
	119	}
	120
13829537 PP	121	if (pglevel)
	122	*pglevel = level;
	123
	124	switch (level) {
	125	case PG_LEVEL_2M:
	126	pmd = (pmd_t *)pte;
	127	pmdval = pmd_val(*pmd) & ~_PAGE_PRESENT;
	128	if (present)
	129	pmdval \|= _PAGE_PRESENT;
	130	set_pmd(pmd, __pmd(pmdval));
	131	break;
	132
	133	case PG_LEVEL_4K:
	134	pteval = pte_val(*pte) & ~_PAGE_PRESENT;
	135	if (present)
	136	pteval \|= _PAGE_PRESENT;
	137	set_pte_atomic(pte, __pte(pteval));
	138	break;
	139
	140	default:
	141	pr_err("kmmio: unexpected page level 0x%x.\n", level);
	142	return;
8b7d89d0 PP	143	}
8b7d89d0 PP	144
13829537 PP	145	__flush_tlb_one(addr);
13829537 PP	146	}
75bb8835	147
13829537	148	/** Mark the given page as not present. Access to it will trigger a fault. */
790e2a29	149	static void arm_kmmio_fault_page(unsigned long page, unsigned int *pglevel)
13829537	150	{
790e2a29	151	set_page_present(page & PAGE_MASK, false, pglevel);
8b7d89d0 PP	152	}
8b7d89d0 PP	153
0fd0e3da	154	/** Mark the given page as present. */
790e2a29	155	static void disarm_kmmio_fault_page(unsigned long page, unsigned int *pglevel)
8b7d89d0	156	{
790e2a29	157	set_page_present(page & PAGE_MASK, true, pglevel);
8b7d89d0 PP	158	}
8b7d89d0 PP	159
0fd0e3da PP	160	/*
	161	* This is being called from do_page_fault().
	162	*
	163	* We may be in an interrupt or a critical section. Also prefecthing may
	164	* trigger a page fault. We may be in the middle of process switch.
	165	* We cannot take any locks, because we could be executing especially
	166	* within a kmmio critical section.
	167	*
	168	* Local interrupts are disabled, so preemption cannot happen.
	169	* Do not enable interrupts, do not sleep, and watch out for other CPUs.
	170	*/
8b7d89d0 PP	171	/*
	172	* Interrupts are disabled on entry as trap3 is an interrupt gate
	173	* and they remain disabled thorough out this function.
	174	*/
0fd0e3da	175	int kmmio_handler(struct pt_regs *regs, unsigned long addr)
8b7d89d0	176	{
0fd0e3da PP	177	struct kmmio_context *ctx;
0fd0e3da PP	178	struct kmmio_fault_page *faultpage;
13829537	179	int ret = 0; /* default to fault not handled */
8b7d89d0 PP	180
	181	/*
	182	* Preemption is now disabled to prevent process switch during
	183	* single stepping. We can only handle one active kmmio trace
	184	* per cpu, so ensure that we finish it before something else
d61fc448 PP	185	* gets to run. We also hold the RCU read lock over single
	186	* stepping to avoid looking up the probe and kmmio_fault_page
	187	* again.
8b7d89d0 PP	188	*/
8b7d89d0 PP	189	preempt_disable();
0fd0e3da	190	rcu_read_lock();
d61fc448	191
0fd0e3da PP	192	faultpage = get_kmmio_fault_page(addr);
	193	if (!faultpage) {
	194	/*
	195	* Either this page fault is not caused by kmmio, or
	196	* another CPU just pulled the kmmio probe from under
13829537	197	* our feet. The latter case should not be possible.
0fd0e3da PP	198	*/
	199	goto no_kmmio;
	200	}
	201
	202	ctx = &get_cpu_var(kmmio_ctx);
8b7d89d0	203	if (ctx->active) {
13829537 PP	204	disarm_kmmio_fault_page(faultpage->page, NULL);
	205	if (addr == ctx->addr) {
	206	/*
	207	* On SMP we sometimes get recursive probe hits on the
	208	* same address. Context is already saved, fall out.
	209	*/
	210	pr_debug("kmmio: duplicate probe hit on CPU %d, for "
	211	"address 0x%08lx.\n",
	212	smp_processor_id(), addr);
	213	ret = 1;
	214	goto no_kmmio_ctx;
	215	}
8b7d89d0	216	/*
0fd0e3da	217	* Prevent overwriting already in-flight context.
13829537 PP	218	* This should not happen, let's hope disarming at least
13829537 PP	219	* prevents a panic.
8b7d89d0	220	*/
0fd0e3da PP	221	pr_emerg("kmmio: recursive probe hit on CPU %d, "
0fd0e3da PP	222	"for address 0x%08lx. Ignoring.\n",
f5136380	223	smp_processor_id(), addr);
13829537 PP	224	pr_emerg("kmmio: previous hit was at 0x%08lx.\n",
13829537 PP	225	ctx->addr);
0fd0e3da	226	goto no_kmmio_ctx;
8b7d89d0 PP	227	}
	228	ctx->active++;
	229
0fd0e3da	230	ctx->fpage = faultpage;
8b7d89d0	231	ctx->probe = get_kmmio_probe(addr);
49023168	232	ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF \| X86_EFLAGS_IF));
0fd0e3da	233	ctx->addr = addr;
8b7d89d0 PP	234
	235	if (ctx->probe && ctx->probe->pre_handler)
	236	ctx->probe->pre_handler(ctx->probe, regs, addr);
	237
d61fc448 PP	238	/*
	239	* Enable single-stepping and disable interrupts for the faulting
	240	* context. Local interrupts must not get enabled during stepping.
	241	*/
49023168 IM	242	regs->flags \|= X86_EFLAGS_TF;
49023168 IM	243	regs->flags &= ~X86_EFLAGS_IF;
8b7d89d0	244
0fd0e3da	245	/* Now we set present bit in PTE and single step. */
8b7d89d0 PP	246	disarm_kmmio_fault_page(ctx->fpage->page, NULL);
8b7d89d0 PP	247
d61fc448 PP	248	/*
	249	* If another cpu accesses the same page while we are stepping,
	250	* the access will not be caught. It will simply succeed and the
	251	* only downside is we lose the event. If this becomes a problem,
	252	* the user should drop to single cpu before tracing.
	253	*/
	254
f5136380	255	put_cpu_var(kmmio_ctx);
13829537	256	return 1; /* fault handled */
8b7d89d0	257
0fd0e3da PP	258	no_kmmio_ctx:
0fd0e3da PP	259	put_cpu_var(kmmio_ctx);
8b7d89d0	260	no_kmmio:
0fd0e3da	261	rcu_read_unlock();
8b7d89d0	262	preempt_enable_no_resched();
13829537	263	return ret;
8b7d89d0 PP	264	}
	265
	266	/*
	267	* Interrupts are disabled on entry as trap1 is an interrupt gate
	268	* and they remain disabled thorough out this function.
0fd0e3da	269	* This must always get called as the pair to kmmio_handler().
8b7d89d0 PP	270	*/
	271	static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
	272	{
f5136380 PP	273	int ret = 0;
f5136380 PP	274	struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
8b7d89d0	275
13829537 PP	276	if (!ctx->active) {
	277	pr_debug("kmmio: spurious debug trap on CPU %d.\n",
	278	smp_processor_id());
f5136380	279	goto out;
13829537	280	}
8b7d89d0 PP	281
	282	if (ctx->probe && ctx->probe->post_handler)
	283	ctx->probe->post_handler(ctx->probe, condition, regs);
	284
d61fc448	285	arm_kmmio_fault_page(ctx->fpage->page, NULL);
8b7d89d0	286
49023168	287	regs->flags &= ~X86_EFLAGS_TF;
8b7d89d0 PP	288	regs->flags \|= ctx->saved_flags;
	289
	290	/* These were acquired in kmmio_handler(). */
	291	ctx->active--;
0fd0e3da	292	BUG_ON(ctx->active);
d61fc448	293	rcu_read_unlock();
8b7d89d0 PP	294	preempt_enable_no_resched();
	295
	296	/*
	297	* if somebody else is singlestepping across a probe point, flags
	298	* will have TF set, in which case, continue the remaining processing
	299	* of do_debug, as if this is not a probe hit.
	300	*/
49023168	301	if (!(regs->flags & X86_EFLAGS_TF))
f5136380	302	ret = 1;
f5136380 PP	303	out:
	304	put_cpu_var(kmmio_ctx);
	305	return ret;
8b7d89d0 PP	306	}
8b7d89d0 PP	307
0fd0e3da	308	/* You must be holding kmmio_lock. */
8b7d89d0 PP	309	static int add_kmmio_fault_page(unsigned long page)
	310	{
	311	struct kmmio_fault_page *f;
	312
	313	page &= PAGE_MASK;
	314	f = get_kmmio_fault_page(page);
	315	if (f) {
0fd0e3da PP	316	if (!f->count)
0fd0e3da PP	317	arm_kmmio_fault_page(f->page, NULL);
8b7d89d0 PP	318	f->count++;
	319	return 0;
	320	}
	321
	322	f = kmalloc(sizeof(*f), GFP_ATOMIC);
	323	if (!f)
	324	return -1;
	325
	326	f->count = 1;
	327	f->page = page;
0fd0e3da	328	list_add_rcu(&f->list, kmmio_page_list(f->page));
8b7d89d0 PP	329
	330	arm_kmmio_fault_page(f->page, NULL);
	331
	332	return 0;
	333	}
	334
0fd0e3da PP	335	/* You must be holding kmmio_lock. */
	336	static void release_kmmio_fault_page(unsigned long page,
	337	struct kmmio_fault_page **release_list)
8b7d89d0 PP	338	{
	339	struct kmmio_fault_page *f;
	340
	341	page &= PAGE_MASK;
	342	f = get_kmmio_fault_page(page);
	343	if (!f)
	344	return;
	345
	346	f->count--;
0fd0e3da	347	BUG_ON(f->count < 0);
8b7d89d0 PP	348	if (!f->count) {
8b7d89d0 PP	349	disarm_kmmio_fault_page(f->page, NULL);
0fd0e3da PP	350	f->release_next = *release_list;
0fd0e3da PP	351	*release_list = f;
8b7d89d0 PP	352	}
	353	}
	354
87e547fe PP	355	/*
	356	* With page-unaligned ioremaps, one or two armed pages may contain
	357	* addresses from outside the intended mapping. Events for these addresses
	358	* are currently silently dropped. The events may result only from programming
	359	* mistakes by accessing addresses before the beginning or past the end of a
	360	* mapping.
	361	*/
8b7d89d0 PP	362	int register_kmmio_probe(struct kmmio_probe *p)
8b7d89d0 PP	363	{
d61fc448	364	unsigned long flags;
8b7d89d0 PP	365	int ret = 0;
8b7d89d0 PP	366	unsigned long size = 0;
87e547fe	367	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
8b7d89d0	368
d61fc448	369	spin_lock_irqsave(&kmmio_lock, flags);
8b7d89d0 PP	370	if (get_kmmio_probe(p->addr)) {
	371	ret = -EEXIST;
	372	goto out;
	373	}
d61fc448	374	kmmio_count++;
0fd0e3da	375	list_add_rcu(&p->list, &kmmio_probes);
87e547fe	376	while (size < size_lim) {
8b7d89d0	377	if (add_kmmio_fault_page(p->addr + size))
0fd0e3da	378	pr_err("kmmio: Unable to set page fault.\n");
8b7d89d0 PP	379	size += PAGE_SIZE;
8b7d89d0 PP	380	}
8b7d89d0	381	out:
d61fc448	382	spin_unlock_irqrestore(&kmmio_lock, flags);
8b7d89d0 PP	383	/*
	384	* XXX: What should I do here?
	385	* Here was a call to global_flush_tlb(), but it does not exist
0fd0e3da	386	* anymore. It seems it's not needed after all.
8b7d89d0 PP	387	*/
	388	return ret;
	389	}
0fd0e3da	390	EXPORT_SYMBOL(register_kmmio_probe);
8b7d89d0	391
0fd0e3da PP	392	static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
	393	{
	394	struct kmmio_delayed_release *dr = container_of(
	395	head,
	396	struct kmmio_delayed_release,
	397	rcu);
	398	struct kmmio_fault_page *p = dr->release_list;
	399	while (p) {
	400	struct kmmio_fault_page *next = p->release_next;
	401	BUG_ON(p->count);
	402	kfree(p);
	403	p = next;
	404	}
	405	kfree(dr);
	406	}
	407
	408	static void remove_kmmio_fault_pages(struct rcu_head *head)
	409	{
	410	struct kmmio_delayed_release *dr = container_of(
	411	head,
	412	struct kmmio_delayed_release,
	413	rcu);
	414	struct kmmio_fault_page *p = dr->release_list;
	415	struct kmmio_fault_page **prevp = &dr->release_list;
	416	unsigned long flags;
	417	spin_lock_irqsave(&kmmio_lock, flags);
	418	while (p) {
	419	if (!p->count)
	420	list_del_rcu(&p->list);
	421	else
	422	*prevp = p->release_next;
	423	prevp = &p->release_next;
	424	p = p->release_next;
	425	}
	426	spin_unlock_irqrestore(&kmmio_lock, flags);
	427	/* This is the real RCU destroy call. */
	428	call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
	429	}
	430
	431	/*
	432	* Remove a kmmio probe. You have to synchronize_rcu() before you can be
d61fc448 PP	433	* sure that the callbacks will not be called anymore. Only after that
d61fc448 PP	434	* you may actually release your struct kmmio_probe.
0fd0e3da PP	435	*
	436	* Unregistering a kmmio fault page has three steps:
	437	* 1. release_kmmio_fault_page()
	438	* Disarm the page, wait a grace period to let all faults finish.
	439	* 2. remove_kmmio_fault_pages()
	440	* Remove the pages from kmmio_page_table.
	441	* 3. rcu_free_kmmio_fault_pages()
	442	* Actally free the kmmio_fault_page structs as with RCU.
	443	*/
8b7d89d0 PP	444	void unregister_kmmio_probe(struct kmmio_probe *p)
8b7d89d0 PP	445	{
d61fc448	446	unsigned long flags;
8b7d89d0	447	unsigned long size = 0;
87e547fe	448	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
0fd0e3da PP	449	struct kmmio_fault_page *release_list = NULL;
0fd0e3da PP	450	struct kmmio_delayed_release *drelease;
8b7d89d0	451
d61fc448	452	spin_lock_irqsave(&kmmio_lock, flags);
87e547fe	453	while (size < size_lim) {
0fd0e3da	454	release_kmmio_fault_page(p->addr + size, &release_list);
8b7d89d0 PP	455	size += PAGE_SIZE;
8b7d89d0 PP	456	}
0fd0e3da	457	list_del_rcu(&p->list);
8b7d89d0	458	kmmio_count--;
d61fc448	459	spin_unlock_irqrestore(&kmmio_lock, flags);
8b7d89d0	460
0fd0e3da PP	461	drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
	462	if (!drelease) {
	463	pr_crit("kmmio: leaking kmmio_fault_page objects.\n");
	464	return;
	465	}
	466	drelease->release_list = release_list;
	467
	468	/*
	469	* This is not really RCU here. We have just disarmed a set of
	470	* pages so that they cannot trigger page faults anymore. However,
	471	* we cannot remove the pages from kmmio_page_table,
	472	* because a probe hit might be in flight on another CPU. The
	473	* pages are collected into a list, and they will be removed from
	474	* kmmio_page_table when it is certain that no probe hit related to
	475	* these pages can be in flight. RCU grace period sounds like a
	476	* good choice.
	477	*
	478	* If we removed the pages too early, kmmio page fault handler might
	479	* not find the respective kmmio_fault_page and determine it's not
	480	* a kmmio fault, when it actually is. This would lead to madness.
	481	*/
	482	call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
8b7d89d0	483	}
0fd0e3da	484	EXPORT_SYMBOL(unregister_kmmio_probe);
8b7d89d0 PP	485
	486	static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
	487	void *args)
	488	{
	489	struct die_args *arg = args;
	490
13829537	491	if (val == DIE_DEBUG && (arg->err & DR_STEP))
8b7d89d0 PP	492	if (post_kmmio_handler(arg->err, arg->regs) == 1)
	493	return NOTIFY_STOP;
	494
	495	return NOTIFY_DONE;
	496	}
13829537 PP	497
	498	static struct notifier_block nb_die = {
	499	.notifier_call = kmmio_die_notifier
	500	};
	501
	502	static int __init init_kmmio(void)
	503	{
	504	int i;
	505	for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
	506	INIT_LIST_HEAD(&kmmio_page_table[i]);
	507	return register_die_notifier(&nb_die);
	508	}
	509	fs_initcall(init_kmmio); /* should be before device_initcall() */