linux/arch/x86/mm/kmmio.c
Steven Rostedt (Google) 3e12758392 x86/mm/kmmio: Remove redundant preempt_disable()
Now that kmmio uses rcu_read_lock_sched_notrace() there's no reason to
call preempt_disable() as the read_lock_sched_notrace() already does that
and is redundant.

This also removes the preempt_enable_no_resched() as the "no_resched()"
portion was bogus as there's no reason to do that.

Link: https://lkml.kernel.org/r/20221212103703.7129cc5d@gandalf.local.home

Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: "x86@kernel.org" <x86@kernel.org>
Cc: Karol Herbst <karolherbst@gmail.com>
Cc: Pekka Paalanen <ppaalanen@gmail.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
2022-12-12 10:54:48 -05:00

633 lines
17 KiB
C

// SPDX-License-Identifier: GPL-2.0
/* Support for MMIO probes.
* Benefit many code from kprobes
* (C) 2002 Louis Zhuang <louis.zhuang@intel.com>.
* 2007 Alexander Eichner
* 2008 Pekka Paalanen <pq@iki.fi>
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/list.h>
#include <linux/rculist.h>
#include <linux/spinlock.h>
#include <linux/hash.h>
#include <linux/export.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 <linux/slab.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 addr; /* the requested address */
pteval_t old_presence; /* page presence prior to arming */
bool armed;
/*
* 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) and post_kmmio_handler().
* Protected by kmmio_lock, when linked into kmmio_page_table.
*/
int count;
bool scheduled_for_release;
};
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;
};
/*
* The kmmio_lock is taken in int3 context, which is treated as NMI context.
* This causes lockdep to complain about it bein in both NMI and normal
* context. Hide it from lockdep, as it should not have any other locks
* taken under it, and this is only enabled for debugging mmio anyway.
*/
static arch_spinlock_t kmmio_lock = __ARCH_SPIN_LOCK_UNLOCKED;
/* 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 addr)
{
unsigned int l;
pte_t *pte = lookup_address(addr, &l);
if (!pte)
return NULL;
addr &= page_level_mask(l);
return &kmmio_page_table[hash_long(addr, 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 addr)
{
struct list_head *head;
struct kmmio_fault_page *f;
unsigned int l;
pte_t *pte = lookup_address(addr, &l);
if (!pte)
return NULL;
addr &= page_level_mask(l);
head = kmmio_page_list(addr);
list_for_each_entry_rcu(f, head, list) {
if (f->addr == addr)
return f;
}
return NULL;
}
static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old)
{
pmd_t new_pmd;
pmdval_t v = pmd_val(*pmd);
if (clear) {
*old = v;
new_pmd = pmd_mkinvalid(*pmd);
} else {
/* Presume this has been called with clear==true previously */
new_pmd = __pmd(*old);
}
set_pmd(pmd, new_pmd);
}
static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old)
{
pteval_t v = pte_val(*pte);
if (clear) {
*old = v;
/* Nothing should care about address */
pte_clear(&init_mm, 0, pte);
} else {
/* Presume this has been called with clear==true previously */
set_pte_atomic(pte, __pte(*old));
}
}
static int clear_page_presence(struct kmmio_fault_page *f, bool clear)
{
unsigned int level;
pte_t *pte = lookup_address(f->addr, &level);
if (!pte) {
pr_err("no pte for addr 0x%08lx\n", f->addr);
return -1;
}
switch (level) {
case PG_LEVEL_2M:
clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence);
break;
case PG_LEVEL_4K:
clear_pte_presence(pte, clear, &f->old_presence);
break;
default:
pr_err("unexpected page level 0x%x.\n", level);
return -1;
}
flush_tlb_one_kernel(f->addr);
return 0;
}
/*
* Mark the given page as not present. Access to it will trigger a fault.
*
* Struct kmmio_fault_page is protected by RCU and kmmio_lock, but the
* protection is ignored here. RCU read lock is assumed held, so the struct
* will not disappear unexpectedly. Furthermore, the caller must guarantee,
* that double arming the same virtual address (page) cannot occur.
*
* Double disarming on the other hand is allowed, and may occur when a fault
* and mmiotrace shutdown happen simultaneously.
*/
static int arm_kmmio_fault_page(struct kmmio_fault_page *f)
{
int ret;
WARN_ONCE(f->armed, KERN_ERR pr_fmt("kmmio page already armed.\n"));
if (f->armed) {
pr_warn("double-arm: addr 0x%08lx, ref %d, old %d\n",
f->addr, f->count, !!f->old_presence);
}
ret = clear_page_presence(f, true);
WARN_ONCE(ret < 0, KERN_ERR pr_fmt("arming at 0x%08lx failed.\n"),
f->addr);
f->armed = true;
return ret;
}
/** Restore the given page to saved presence state. */
static void disarm_kmmio_fault_page(struct kmmio_fault_page *f)
{
int ret = clear_page_presence(f, false);
WARN_ONCE(ret < 0,
KERN_ERR "kmmio disarming at 0x%08lx failed.\n", f->addr);
f->armed = false;
}
/*
* 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 throughout 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 */
unsigned long page_base = addr;
unsigned int l;
pte_t *pte = lookup_address(addr, &l);
if (!pte)
return -EINVAL;
page_base &= page_level_mask(l);
/*
* Hold the RCU read lock over single stepping to avoid looking
* up the probe and kmmio_fault_page again. The rcu_read_lock_sched()
* also disables preemption and prevents process switch during
* the single stepping. We can only handle one active kmmio trace
* per cpu, so ensure that we finish it before something else
* gets to run.
*/
rcu_read_lock_sched_notrace();
faultpage = get_kmmio_fault_page(page_base);
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 = this_cpu_ptr(&kmmio_ctx);
if (ctx->active) {
if (page_base == ctx->addr) {
/*
* A second fault on the same page means some other
* condition needs handling by do_page_fault(), the
* page really not being present is the most common.
*/
pr_debug("secondary hit for 0x%08lx CPU %d.\n",
addr, smp_processor_id());
if (!faultpage->old_presence)
pr_info("unexpected secondary hit for address 0x%08lx on CPU %d.\n",
addr, smp_processor_id());
} else {
/*
* Prevent overwriting already in-flight context.
* This should not happen, let's hope disarming at
* least prevents a panic.
*/
pr_emerg("recursive probe hit on CPU %d, for address 0x%08lx. Ignoring.\n",
smp_processor_id(), addr);
pr_emerg("previous hit was at 0x%08lx.\n", ctx->addr);
disarm_kmmio_fault_page(faultpage);
}
goto no_kmmio;
}
ctx->active++;
ctx->fpage = faultpage;
ctx->probe = get_kmmio_probe(page_base);
ctx->saved_flags = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
ctx->addr = page_base;
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);
/*
* 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.
*/
return 1; /* fault handled */
no_kmmio:
rcu_read_unlock_sched_notrace();
return ret;
}
/*
* Interrupts are disabled on entry as trap1 is an interrupt gate
* and they remain disabled throughout 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 = this_cpu_ptr(&kmmio_ctx);
if (!ctx->active) {
/*
* debug traps without an active context are due to either
* something external causing them (f.e. using a debugger while
* mmio tracing enabled), or erroneous behaviour
*/
pr_warn("unexpected 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);
/* Prevent racing against release_kmmio_fault_page(). */
arch_spin_lock(&kmmio_lock);
if (ctx->fpage->count)
arm_kmmio_fault_page(ctx->fpage);
arch_spin_unlock(&kmmio_lock);
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_sched_notrace();
/*
* 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:
return ret;
}
/* You must be holding kmmio_lock. */
static int add_kmmio_fault_page(unsigned long addr)
{
struct kmmio_fault_page *f;
f = get_kmmio_fault_page(addr);
if (f) {
if (!f->count)
arm_kmmio_fault_page(f);
f->count++;
return 0;
}
f = kzalloc(sizeof(*f), GFP_ATOMIC);
if (!f)
return -1;
f->count = 1;
f->addr = addr;
if (arm_kmmio_fault_page(f)) {
kfree(f);
return -1;
}
list_add_rcu(&f->list, kmmio_page_list(f->addr));
return 0;
}
/* You must be holding kmmio_lock. */
static void release_kmmio_fault_page(unsigned long addr,
struct kmmio_fault_page **release_list)
{
struct kmmio_fault_page *f;
f = get_kmmio_fault_page(addr);
if (!f)
return;
f->count--;
BUG_ON(f->count < 0);
if (!f->count) {
disarm_kmmio_fault_page(f);
if (!f->scheduled_for_release) {
f->release_next = *release_list;
*release_list = f;
f->scheduled_for_release = true;
}
}
}
/*
* 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;
unsigned long addr = p->addr & PAGE_MASK;
const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
unsigned int l;
pte_t *pte;
local_irq_save(flags);
arch_spin_lock(&kmmio_lock);
if (get_kmmio_probe(addr)) {
ret = -EEXIST;
goto out;
}
pte = lookup_address(addr, &l);
if (!pte) {
ret = -EINVAL;
goto out;
}
kmmio_count++;
list_add_rcu(&p->list, &kmmio_probes);
while (size < size_lim) {
if (add_kmmio_fault_page(addr + size))
pr_err("Unable to set page fault.\n");
size += page_level_size(l);
}
out:
arch_spin_unlock(&kmmio_lock);
local_irq_restore(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 *f = dr->release_list;
while (f) {
struct kmmio_fault_page *next = f->release_next;
BUG_ON(f->count);
kfree(f);
f = 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 *f = dr->release_list;
struct kmmio_fault_page **prevp = &dr->release_list;
unsigned long flags;
local_irq_save(flags);
arch_spin_lock(&kmmio_lock);
while (f) {
if (!f->count) {
list_del_rcu(&f->list);
prevp = &f->release_next;
} else {
*prevp = f->release_next;
f->release_next = NULL;
f->scheduled_for_release = false;
}
f = *prevp;
}
arch_spin_unlock(&kmmio_lock);
local_irq_restore(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()
* Actually free the kmmio_fault_page structs as with RCU.
*/
void unregister_kmmio_probe(struct kmmio_probe *p)
{
unsigned long flags;
unsigned long size = 0;
unsigned long addr = p->addr & PAGE_MASK;
const unsigned long size_lim = p->len + (p->addr & ~PAGE_MASK);
struct kmmio_fault_page *release_list = NULL;
struct kmmio_delayed_release *drelease;
unsigned int l;
pte_t *pte;
pte = lookup_address(addr, &l);
if (!pte)
return;
local_irq_save(flags);
arch_spin_lock(&kmmio_lock);
while (size < size_lim) {
release_kmmio_fault_page(addr + size, &release_list);
size += page_level_size(l);
}
list_del_rcu(&p->list);
kmmio_count--;
arch_spin_unlock(&kmmio_lock);
local_irq_restore(flags);
if (!release_list)
return;
drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
if (!drelease) {
pr_crit("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;
unsigned long* dr6_p = (unsigned long *)ERR_PTR(arg->err);
if (val == DIE_DEBUG && (*dr6_p & DR_STEP))
if (post_kmmio_handler(*dr6_p, arg->regs) == 1) {
/*
* Reset the BS bit in dr6 (pointed by args->err) to
* denote completion of processing
*/
*dr6_p &= ~DR_STEP;
return NOTIFY_STOP;
}
return NOTIFY_DONE;
}
static struct notifier_block nb_die = {
.notifier_call = kmmio_die_notifier
};
int kmmio_init(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);
}
void kmmio_cleanup(void)
{
int i;
unregister_die_notifier(&nb_die);
for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) {
WARN_ONCE(!list_empty(&kmmio_page_table[i]),
KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n");
}
}