[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/rculist.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>
668a6c36	9	#include <linux/rculist.h>
8b7d89d0 PP	10	#include <linux/spinlock.h>
	11	#include <linux/hash.h>
	12	#include <linux/init.h>
	13	#include <linux/module.h>
8b7d89d0	14	#include <linux/kernel.h>
8b7d89d0 PP	15	#include <linux/uaccess.h>
	16	#include <linux/ptrace.h>
	17	#include <linux/preempt.h>
f5136380	18	#include <linux/percpu.h>
0fd0e3da	19	#include <linux/kdebug.h>
d61fc448	20	#include <linux/mutex.h>
970e6fa0	21	#include <linux/io.h>
8b7d89d0	22	#include <asm/cacheflush.h>
8b7d89d0	23	#include <asm/tlbflush.h>
970e6fa0	24	#include <linux/errno.h>
13829537	25	#include <asm/debugreg.h>
0fd0e3da	26	#include <linux/mmiotrace.h>
8b7d89d0	27
8b7d89d0 PP	28	#define KMMIO_PAGE_HASH_BITS 4
	29	#define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)
	30
0fd0e3da PP	31	struct kmmio_fault_page {
	32	struct list_head list;
	33	struct kmmio_fault_page *release_next;
	34	unsigned long page; /* location of the fault page */
	35
	36	/*
	37	* Number of times this page has been registered as a part
	38	* of a probe. If zero, page is disarmed and this may be freed.
	39	* Used only by writers (RCU).
	40	*/
	41	int count;
	42	};
	43
	44	struct kmmio_delayed_release {
	45	struct rcu_head rcu;
	46	struct kmmio_fault_page *release_list;
	47	};
	48
8b7d89d0 PP	49	struct kmmio_context {
	50	struct kmmio_fault_page *fpage;
	51	struct kmmio_probe *probe;
	52	unsigned long saved_flags;
0fd0e3da	53	unsigned long addr;
8b7d89d0 PP	54	int active;
	55	};
	56
8b7d89d0 PP	57	static DEFINE_SPINLOCK(kmmio_lock);
8b7d89d0 PP	58
13829537	59	/* Protected by kmmio_lock */
8b7d89d0	60	unsigned int kmmio_count;
0fd0e3da PP	61
0fd0e3da PP	62	/* Read-protected by RCU, write-protected by kmmio_lock. */
8b7d89d0 PP	63	static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
	64	static LIST_HEAD(kmmio_probes);
	65
0fd0e3da PP	66	static struct list_head *kmmio_page_list(unsigned long page)
	67	{
	68	return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)];
	69	}
	70
f5136380 PP	71	/* Accessed per-cpu */
f5136380 PP	72	static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
8b7d89d0	73
8b7d89d0 PP	74	/*
	75	* this is basically a dynamic stabbing problem:
	76	* Could use the existing prio tree code or
	77	* Possible better implementations:
	78	* The Interval Skip List: A Data Structure for Finding All Intervals That
	79	* Overlap a Point (might be simple)
	80	* Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
	81	*/
0fd0e3da	82	/* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
8b7d89d0 PP	83	static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
	84	{
	85	struct kmmio_probe *p;
0fd0e3da	86	list_for_each_entry_rcu(p, &kmmio_probes, list) {
8b7d89d0 PP	87	if (addr >= p->addr && addr <= (p->addr + p->len))
	88	return p;
	89	}
	90	return NULL;
	91	}
	92
0fd0e3da	93	/* You must be holding RCU read lock. */
8b7d89d0 PP	94	static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page)
8b7d89d0 PP	95	{
0fd0e3da PP	96	struct list_head *head;
0fd0e3da PP	97	struct kmmio_fault_page *p;
8b7d89d0 PP	98
8b7d89d0 PP	99	page &= PAGE_MASK;
0fd0e3da PP	100	head = kmmio_page_list(page);
0fd0e3da PP	101	list_for_each_entry_rcu(p, head, list) {
8b7d89d0 PP	102	if (p->page == page)
	103	return p;
	104	}
8b7d89d0 PP	105	return NULL;
	106	}
	107
790e2a29 PP	108	static void set_page_present(unsigned long addr, bool present,
790e2a29 PP	109	unsigned int *pglevel)
8b7d89d0	110	{
13829537 PP	111	pteval_t pteval;
13829537 PP	112	pmdval_t pmdval;
790e2a29	113	unsigned int level;
13829537 PP	114	pmd_t *pmd;
13829537 PP	115	pte_t *pte = lookup_address(addr, &level);
8b7d89d0	116
75bb8835	117	if (!pte) {
13829537	118	pr_err("kmmio: no pte for page 0x%08lx\n", addr);
75bb8835 PP	119	return;
	120	}
	121
13829537 PP	122	if (pglevel)
	123	*pglevel = level;
	124
	125	switch (level) {
	126	case PG_LEVEL_2M:
	127	pmd = (pmd_t *)pte;
	128	pmdval = pmd_val(*pmd) & ~_PAGE_PRESENT;
	129	if (present)
	130	pmdval \|= _PAGE_PRESENT;
	131	set_pmd(pmd, __pmd(pmdval));
	132	break;
	133
	134	case PG_LEVEL_4K:
	135	pteval = pte_val(*pte) & ~_PAGE_PRESENT;
	136	if (present)
	137	pteval \|= _PAGE_PRESENT;
	138	set_pte_atomic(pte, __pte(pteval));
	139	break;
	140
	141	default:
	142	pr_err("kmmio: unexpected page level 0x%x.\n", level);
	143	return;
8b7d89d0 PP	144	}
8b7d89d0 PP	145
13829537 PP	146	__flush_tlb_one(addr);
13829537 PP	147	}
75bb8835	148
13829537	149	/** Mark the given page as not present. Access to it will trigger a fault. */
790e2a29	150	static void arm_kmmio_fault_page(unsigned long page, unsigned int *pglevel)
13829537	151	{
790e2a29	152	set_page_present(page & PAGE_MASK, false, pglevel);
8b7d89d0 PP	153	}
8b7d89d0 PP	154
0fd0e3da	155	/** Mark the given page as present. */
790e2a29	156	static void disarm_kmmio_fault_page(unsigned long page, unsigned int *pglevel)
8b7d89d0	157	{
790e2a29	158	set_page_present(page & PAGE_MASK, true, pglevel);
8b7d89d0 PP	159	}
8b7d89d0 PP	160
0fd0e3da PP	161	/*
	162	* This is being called from do_page_fault().
	163	*
	164	* We may be in an interrupt or a critical section. Also prefecthing may
	165	* trigger a page fault. We may be in the middle of process switch.
	166	* We cannot take any locks, because we could be executing especially
	167	* within a kmmio critical section.
	168	*
	169	* Local interrupts are disabled, so preemption cannot happen.
	170	* Do not enable interrupts, do not sleep, and watch out for other CPUs.
	171	*/
8b7d89d0 PP	172	/*
	173	* Interrupts are disabled on entry as trap3 is an interrupt gate
	174	* and they remain disabled thorough out this function.
	175	*/
0fd0e3da	176	int kmmio_handler(struct pt_regs *regs, unsigned long addr)
8b7d89d0	177	{
0fd0e3da PP	178	struct kmmio_context *ctx;
0fd0e3da PP	179	struct kmmio_fault_page *faultpage;
13829537	180	int ret = 0; /* default to fault not handled */
8b7d89d0 PP	181
	182	/*
	183	* Preemption is now disabled to prevent process switch during
	184	* single stepping. We can only handle one active kmmio trace
	185	* per cpu, so ensure that we finish it before something else
d61fc448 PP	186	* gets to run. We also hold the RCU read lock over single
	187	* stepping to avoid looking up the probe and kmmio_fault_page
	188	* again.
8b7d89d0 PP	189	*/
8b7d89d0 PP	190	preempt_disable();
0fd0e3da	191	rcu_read_lock();
d61fc448	192
0fd0e3da PP	193	faultpage = get_kmmio_fault_page(addr);
	194	if (!faultpage) {
	195	/*
	196	* Either this page fault is not caused by kmmio, or
	197	* another CPU just pulled the kmmio probe from under
13829537	198	* our feet. The latter case should not be possible.
0fd0e3da PP	199	*/
	200	goto no_kmmio;
	201	}
	202
	203	ctx = &get_cpu_var(kmmio_ctx);
8b7d89d0	204	if (ctx->active) {
13829537 PP	205	disarm_kmmio_fault_page(faultpage->page, NULL);
	206	if (addr == ctx->addr) {
	207	/*
	208	* On SMP we sometimes get recursive probe hits on the
	209	* same address. Context is already saved, fall out.
	210	*/
	211	pr_debug("kmmio: duplicate probe hit on CPU %d, for "
	212	"address 0x%08lx.\n",
	213	smp_processor_id(), addr);
	214	ret = 1;
	215	goto no_kmmio_ctx;
	216	}
8b7d89d0	217	/*
0fd0e3da	218	* Prevent overwriting already in-flight context.
13829537 PP	219	* This should not happen, let's hope disarming at least
13829537 PP	220	* prevents a panic.
8b7d89d0	221	*/
0fd0e3da PP	222	pr_emerg("kmmio: recursive probe hit on CPU %d, "
0fd0e3da PP	223	"for address 0x%08lx. Ignoring.\n",
f5136380	224	smp_processor_id(), addr);
13829537 PP	225	pr_emerg("kmmio: previous hit was at 0x%08lx.\n",
13829537 PP	226	ctx->addr);
0fd0e3da	227	goto no_kmmio_ctx;
8b7d89d0 PP	228	}
	229	ctx->active++;
	230
0fd0e3da	231	ctx->fpage = faultpage;
8b7d89d0	232	ctx->probe = get_kmmio_probe(addr);
49023168	233	ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF \| X86_EFLAGS_IF));
0fd0e3da	234	ctx->addr = addr;
8b7d89d0 PP	235
	236	if (ctx->probe && ctx->probe->pre_handler)
	237	ctx->probe->pre_handler(ctx->probe, regs, addr);
	238
d61fc448 PP	239	/*
	240	* Enable single-stepping and disable interrupts for the faulting
	241	* context. Local interrupts must not get enabled during stepping.
	242	*/
49023168 IM	243	regs->flags \|= X86_EFLAGS_TF;
49023168 IM	244	regs->flags &= ~X86_EFLAGS_IF;
8b7d89d0	245
0fd0e3da	246	/* Now we set present bit in PTE and single step. */
8b7d89d0 PP	247	disarm_kmmio_fault_page(ctx->fpage->page, NULL);
8b7d89d0 PP	248
d61fc448 PP	249	/*
	250	* If another cpu accesses the same page while we are stepping,
	251	* the access will not be caught. It will simply succeed and the
	252	* only downside is we lose the event. If this becomes a problem,
	253	* the user should drop to single cpu before tracing.
	254	*/
	255
f5136380	256	put_cpu_var(kmmio_ctx);
13829537	257	return 1; /* fault handled */
8b7d89d0	258
0fd0e3da PP	259	no_kmmio_ctx:
0fd0e3da PP	260	put_cpu_var(kmmio_ctx);
8b7d89d0	261	no_kmmio:
0fd0e3da	262	rcu_read_unlock();
8b7d89d0	263	preempt_enable_no_resched();
13829537	264	return ret;
8b7d89d0 PP	265	}
	266
	267	/*
	268	* Interrupts are disabled on entry as trap1 is an interrupt gate
	269	* and they remain disabled thorough out this function.
0fd0e3da	270	* This must always get called as the pair to kmmio_handler().
8b7d89d0 PP	271	*/
	272	static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
	273	{
f5136380 PP	274	int ret = 0;
f5136380 PP	275	struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
8b7d89d0	276
13829537 PP	277	if (!ctx->active) {
	278	pr_debug("kmmio: spurious debug trap on CPU %d.\n",
	279	smp_processor_id());
f5136380	280	goto out;
13829537	281	}
8b7d89d0 PP	282
	283	if (ctx->probe && ctx->probe->post_handler)
	284	ctx->probe->post_handler(ctx->probe, condition, regs);
	285
d61fc448	286	arm_kmmio_fault_page(ctx->fpage->page, NULL);
8b7d89d0	287
49023168	288	regs->flags &= ~X86_EFLAGS_TF;
8b7d89d0 PP	289	regs->flags \|= ctx->saved_flags;
	290
	291	/* These were acquired in kmmio_handler(). */
	292	ctx->active--;
0fd0e3da	293	BUG_ON(ctx->active);
d61fc448	294	rcu_read_unlock();
8b7d89d0 PP	295	preempt_enable_no_resched();
	296
	297	/*
	298	* if somebody else is singlestepping across a probe point, flags
	299	* will have TF set, in which case, continue the remaining processing
	300	* of do_debug, as if this is not a probe hit.
	301	*/
49023168	302	if (!(regs->flags & X86_EFLAGS_TF))
f5136380	303	ret = 1;
f5136380 PP	304	out:
	305	put_cpu_var(kmmio_ctx);
	306	return ret;
8b7d89d0 PP	307	}
8b7d89d0 PP	308
0fd0e3da	309	/* You must be holding kmmio_lock. */
8b7d89d0 PP	310	static int add_kmmio_fault_page(unsigned long page)
	311	{
	312	struct kmmio_fault_page *f;
	313
	314	page &= PAGE_MASK;
	315	f = get_kmmio_fault_page(page);
	316	if (f) {
0fd0e3da PP	317	if (!f->count)
0fd0e3da PP	318	arm_kmmio_fault_page(f->page, NULL);
8b7d89d0 PP	319	f->count++;
	320	return 0;
	321	}
	322
	323	f = kmalloc(sizeof(*f), GFP_ATOMIC);
	324	if (!f)
	325	return -1;
	326
	327	f->count = 1;
	328	f->page = page;
0fd0e3da	329	list_add_rcu(&f->list, kmmio_page_list(f->page));
8b7d89d0 PP	330
	331	arm_kmmio_fault_page(f->page, NULL);
	332
	333	return 0;
	334	}
	335
0fd0e3da PP	336	/* You must be holding kmmio_lock. */
	337	static void release_kmmio_fault_page(unsigned long page,
	338	struct kmmio_fault_page **release_list)
8b7d89d0 PP	339	{
	340	struct kmmio_fault_page *f;
	341
	342	page &= PAGE_MASK;
	343	f = get_kmmio_fault_page(page);
	344	if (!f)
	345	return;
	346
	347	f->count--;
0fd0e3da	348	BUG_ON(f->count < 0);
8b7d89d0 PP	349	if (!f->count) {
8b7d89d0 PP	350	disarm_kmmio_fault_page(f->page, NULL);
0fd0e3da PP	351	f->release_next = *release_list;
0fd0e3da PP	352	*release_list = f;
8b7d89d0 PP	353	}
	354	}
	355
87e547fe PP	356	/*
	357	* With page-unaligned ioremaps, one or two armed pages may contain
	358	* addresses from outside the intended mapping. Events for these addresses
	359	* are currently silently dropped. The events may result only from programming
	360	* mistakes by accessing addresses before the beginning or past the end of a
	361	* mapping.
	362	*/
8b7d89d0 PP	363	int register_kmmio_probe(struct kmmio_probe *p)
8b7d89d0 PP	364	{
d61fc448	365	unsigned long flags;
8b7d89d0 PP	366	int ret = 0;
8b7d89d0 PP	367	unsigned long size = 0;
87e547fe	368	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
8b7d89d0	369
d61fc448	370	spin_lock_irqsave(&kmmio_lock, flags);
8b7d89d0 PP	371	if (get_kmmio_probe(p->addr)) {
	372	ret = -EEXIST;
	373	goto out;
	374	}
d61fc448	375	kmmio_count++;
0fd0e3da	376	list_add_rcu(&p->list, &kmmio_probes);
87e547fe	377	while (size < size_lim) {
8b7d89d0	378	if (add_kmmio_fault_page(p->addr + size))
0fd0e3da	379	pr_err("kmmio: Unable to set page fault.\n");
8b7d89d0 PP	380	size += PAGE_SIZE;
8b7d89d0 PP	381	}
8b7d89d0	382	out:
d61fc448	383	spin_unlock_irqrestore(&kmmio_lock, flags);
8b7d89d0 PP	384	/*
	385	* XXX: What should I do here?
	386	* Here was a call to global_flush_tlb(), but it does not exist
0fd0e3da	387	* anymore. It seems it's not needed after all.
8b7d89d0 PP	388	*/
	389	return ret;
	390	}
0fd0e3da	391	EXPORT_SYMBOL(register_kmmio_probe);
8b7d89d0	392
0fd0e3da PP	393	static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
	394	{
	395	struct kmmio_delayed_release *dr = container_of(
	396	head,
	397	struct kmmio_delayed_release,
	398	rcu);
	399	struct kmmio_fault_page *p = dr->release_list;
	400	while (p) {
	401	struct kmmio_fault_page *next = p->release_next;
	402	BUG_ON(p->count);
	403	kfree(p);
	404	p = next;
	405	}
	406	kfree(dr);
	407	}
	408
	409	static void remove_kmmio_fault_pages(struct rcu_head *head)
	410	{
	411	struct kmmio_delayed_release *dr = container_of(
	412	head,
	413	struct kmmio_delayed_release,
	414	rcu);
	415	struct kmmio_fault_page *p = dr->release_list;
	416	struct kmmio_fault_page **prevp = &dr->release_list;
	417	unsigned long flags;
	418	spin_lock_irqsave(&kmmio_lock, flags);
	419	while (p) {
	420	if (!p->count)
	421	list_del_rcu(&p->list);
	422	else
	423	*prevp = p->release_next;
	424	prevp = &p->release_next;
	425	p = p->release_next;
	426	}
	427	spin_unlock_irqrestore(&kmmio_lock, flags);
	428	/* This is the real RCU destroy call. */
	429	call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
	430	}
	431
	432	/*
	433	* Remove a kmmio probe. You have to synchronize_rcu() before you can be
d61fc448 PP	434	* sure that the callbacks will not be called anymore. Only after that
d61fc448 PP	435	* you may actually release your struct kmmio_probe.
0fd0e3da PP	436	*
	437	* Unregistering a kmmio fault page has three steps:
	438	* 1. release_kmmio_fault_page()
	439	* Disarm the page, wait a grace period to let all faults finish.
	440	* 2. remove_kmmio_fault_pages()
	441	* Remove the pages from kmmio_page_table.
	442	* 3. rcu_free_kmmio_fault_pages()
	443	* Actally free the kmmio_fault_page structs as with RCU.
	444	*/
8b7d89d0 PP	445	void unregister_kmmio_probe(struct kmmio_probe *p)
8b7d89d0 PP	446	{
d61fc448	447	unsigned long flags;
8b7d89d0	448	unsigned long size = 0;
87e547fe	449	const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
0fd0e3da PP	450	struct kmmio_fault_page *release_list = NULL;
0fd0e3da PP	451	struct kmmio_delayed_release *drelease;
8b7d89d0	452
d61fc448	453	spin_lock_irqsave(&kmmio_lock, flags);
87e547fe	454	while (size < size_lim) {
0fd0e3da	455	release_kmmio_fault_page(p->addr + size, &release_list);
8b7d89d0 PP	456	size += PAGE_SIZE;
8b7d89d0 PP	457	}
0fd0e3da	458	list_del_rcu(&p->list);
8b7d89d0	459	kmmio_count--;
d61fc448	460	spin_unlock_irqrestore(&kmmio_lock, flags);
8b7d89d0	461
0fd0e3da PP	462	drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
	463	if (!drelease) {
	464	pr_crit("kmmio: leaking kmmio_fault_page objects.\n");
	465	return;
	466	}
	467	drelease->release_list = release_list;
	468
	469	/*
	470	* This is not really RCU here. We have just disarmed a set of
	471	* pages so that they cannot trigger page faults anymore. However,
	472	* we cannot remove the pages from kmmio_page_table,
	473	* because a probe hit might be in flight on another CPU. The
	474	* pages are collected into a list, and they will be removed from
	475	* kmmio_page_table when it is certain that no probe hit related to
	476	* these pages can be in flight. RCU grace period sounds like a
	477	* good choice.
	478	*
	479	* If we removed the pages too early, kmmio page fault handler might
	480	* not find the respective kmmio_fault_page and determine it's not
	481	* a kmmio fault, when it actually is. This would lead to madness.
	482	*/
	483	call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
8b7d89d0	484	}
0fd0e3da	485	EXPORT_SYMBOL(unregister_kmmio_probe);
8b7d89d0 PP	486
	487	static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
	488	void *args)
	489	{
	490	struct die_args *arg = args;
	491
13829537	492	if (val == DIE_DEBUG && (arg->err & DR_STEP))
8b7d89d0 PP	493	if (post_kmmio_handler(arg->err, arg->regs) == 1)
	494	return NOTIFY_STOP;
	495
	496	return NOTIFY_DONE;
	497	}
13829537 PP	498
	499	static struct notifier_block nb_die = {
	500	.notifier_call = kmmio_die_notifier
	501	};
	502
	503	static int __init init_kmmio(void)
	504	{
	505	int i;
	506	for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
	507	INIT_LIST_HEAD(&kmmio_page_table[i]);
	508	return register_die_notifier(&nb_die);
	509	}
	510	fs_initcall(init_kmmio); /* should be before device_initcall() */