uprobes: Pass probed vaddr to arch_uprobe_analyze_insn()
[cascardo/linux.git] / kernel / events / uprobes.c
index e56e56a..b52376d 100644 (file)
 #include <linux/rmap.h>                /* anon_vma_prepare */
 #include <linux/mmu_notifier.h>        /* set_pte_at_notify */
 #include <linux/swap.h>                /* try_to_free_swap */
+#include <linux/ptrace.h>      /* user_enable_single_step */
+#include <linux/kdebug.h>      /* notifier mechanism */
 
 #include <linux/uprobes.h>
 
+#define UINSNS_PER_PAGE                        (PAGE_SIZE/UPROBE_XOL_SLOT_BYTES)
+#define MAX_UPROBE_XOL_SLOTS           UINSNS_PER_PAGE
+
 static struct rb_root uprobes_tree = RB_ROOT;
 
 static DEFINE_SPINLOCK(uprobes_treelock);      /* serialize rbtree access */
@@ -154,6 +159,11 @@ static int __replace_page(struct vm_area_struct *vma, struct page *page, struct
        get_page(kpage);
        page_add_new_anon_rmap(kpage, vma, addr);
 
+       if (!PageAnon(page)) {
+               dec_mm_counter(mm, MM_FILEPAGES);
+               inc_mm_counter(mm, MM_ANONPAGES);
+       }
+
        flush_cache_page(vma, addr, pte_pfn(*ptep));
        ptep_clear_flush(vma, addr, ptep);
        set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
@@ -301,7 +311,7 @@ static int read_opcode(struct mm_struct *mm, unsigned long vaddr, uprobe_opcode_
        void *vaddr_new;
        int ret;
 
-       ret = get_user_pages(NULL, mm, vaddr, 1, 0, 0, &page, NULL);
+       ret = get_user_pages(NULL, mm, vaddr, 1, 0, 1, &page, NULL);
        if (ret <= 0)
                return ret;
 
@@ -322,10 +332,20 @@ static int is_swbp_at_addr(struct mm_struct *mm, unsigned long vaddr)
        uprobe_opcode_t opcode;
        int result;
 
+       if (current->mm == mm) {
+               pagefault_disable();
+               result = __copy_from_user_inatomic(&opcode, (void __user*)vaddr,
+                                                               sizeof(opcode));
+               pagefault_enable();
+
+               if (likely(result == 0))
+                       goto out;
+       }
+
        result = read_opcode(mm, vaddr, &opcode);
        if (result)
                return result;
-
+out:
        if (is_swbp_insn(&opcode))
                return 1;
 
@@ -486,6 +506,9 @@ static struct uprobe *insert_uprobe(struct uprobe *uprobe)
        u = __insert_uprobe(uprobe);
        spin_unlock_irqrestore(&uprobes_treelock, flags);
 
+       /* For now assume that the instruction need not be single-stepped */
+       uprobe->flags |= UPROBE_SKIP_SSTEP;
+
        return u;
 }
 
@@ -523,6 +546,21 @@ static struct uprobe *alloc_uprobe(struct inode *inode, loff_t offset)
        return uprobe;
 }
 
+static void handler_chain(struct uprobe *uprobe, struct pt_regs *regs)
+{
+       struct uprobe_consumer *uc;
+
+       if (!(uprobe->flags & UPROBE_RUN_HANDLER))
+               return;
+
+       down_read(&uprobe->consumer_rwsem);
+       for (uc = uprobe->consumers; uc; uc = uc->next) {
+               if (!uc->filter || uc->filter(uc, current))
+                       uc->handler(uc, regs);
+       }
+       up_read(&uprobe->consumer_rwsem);
+}
+
 /* Returns the previous consumer */
 static struct uprobe_consumer *
 consumer_add(struct uprobe *uprobe, struct uprobe_consumer *uc)
@@ -618,6 +656,29 @@ copy_insn(struct uprobe *uprobe, struct vm_area_struct *vma, unsigned long addr)
        return __copy_insn(mapping, vma, uprobe->arch.insn, bytes, uprobe->offset);
 }
 
+/*
+ * How mm->uprobes_state.count gets updated
+ * uprobe_mmap() increments the count if
+ *     - it successfully adds a breakpoint.
+ *     - it cannot add a breakpoint, but sees that there is a underlying
+ *       breakpoint (via a is_swbp_at_addr()).
+ *
+ * uprobe_munmap() decrements the count if
+ *     - it sees a underlying breakpoint, (via is_swbp_at_addr)
+ *       (Subsequent uprobe_unregister wouldnt find the breakpoint
+ *       unless a uprobe_mmap kicks in, since the old vma would be
+ *       dropped just after uprobe_munmap.)
+ *
+ * uprobe_register increments the count if:
+ *     - it successfully adds a breakpoint.
+ *
+ * uprobe_unregister decrements the count if:
+ *     - it sees a underlying breakpoint and removes successfully.
+ *       (via is_swbp_at_addr)
+ *       (Subsequent uprobe_munmap wouldnt find the breakpoint
+ *       since there is no underlying breakpoint after the
+ *       breakpoint removal.)
+ */
 static int
 install_breakpoint(struct uprobe *uprobe, struct mm_struct *mm,
                        struct vm_area_struct *vma, loff_t vaddr)
@@ -645,13 +706,25 @@ install_breakpoint(struct uprobe *uprobe, struct mm_struct *mm,
                if (is_swbp_insn((uprobe_opcode_t *)uprobe->arch.insn))
                        return -EEXIST;
 
-               ret = arch_uprobes_analyze_insn(&uprobe->arch, mm);
+               ret = arch_uprobe_analyze_insn(&uprobe->arch, mm, addr);
                if (ret)
                        return ret;
 
                uprobe->flags |= UPROBE_COPY_INSN;
        }
+
+       /*
+        * Ideally, should be updating the probe count after the breakpoint
+        * has been successfully inserted. However a thread could hit the
+        * breakpoint we just inserted even before the probe count is
+        * incremented. If this is the first breakpoint placed, breakpoint
+        * notifier might ignore uprobes and pass the trap to the thread.
+        * Hence increment before and decrement on failure.
+        */
+       atomic_inc(&mm->uprobes_state.count);
        ret = set_swbp(&uprobe->arch, mm, addr);
+       if (ret)
+               atomic_dec(&mm->uprobes_state.count);
 
        return ret;
 }
@@ -659,9 +732,15 @@ install_breakpoint(struct uprobe *uprobe, struct mm_struct *mm,
 static void
 remove_breakpoint(struct uprobe *uprobe, struct mm_struct *mm, loff_t vaddr)
 {
-       set_orig_insn(&uprobe->arch, mm, (unsigned long)vaddr, true);
+       if (!set_orig_insn(&uprobe->arch, mm, (unsigned long)vaddr, true))
+               atomic_dec(&mm->uprobes_state.count);
 }
 
+/*
+ * There could be threads that have already hit the breakpoint. They
+ * will recheck the current insn and restart if find_uprobe() fails.
+ * See find_active_uprobe().
+ */
 static void delete_uprobe(struct uprobe *uprobe)
 {
        unsigned long flags;
@@ -767,12 +846,12 @@ static int register_for_each_vma(struct uprobe *uprobe, bool is_register)
                }
 
                mm = vi->mm;
-               down_read(&mm->mmap_sem);
+               down_write(&mm->mmap_sem);
                vma = find_vma(mm, (unsigned long)vi->vaddr);
                if (!vma || !valid_vma(vma, is_register)) {
                        list_del(&vi->probe_list);
                        kfree(vi);
-                       up_read(&mm->mmap_sem);
+                       up_write(&mm->mmap_sem);
                        mmput(mm);
                        continue;
                }
@@ -781,7 +860,7 @@ static int register_for_each_vma(struct uprobe *uprobe, bool is_register)
                                                vaddr != vi->vaddr) {
                        list_del(&vi->probe_list);
                        kfree(vi);
-                       up_read(&mm->mmap_sem);
+                       up_write(&mm->mmap_sem);
                        mmput(mm);
                        continue;
                }
@@ -791,7 +870,7 @@ static int register_for_each_vma(struct uprobe *uprobe, bool is_register)
                else
                        remove_breakpoint(uprobe, mm, vi->vaddr);
 
-               up_read(&mm->mmap_sem);
+               up_write(&mm->mmap_sem);
                mmput(mm);
                if (is_register) {
                        if (ret && ret == -EEXIST)
@@ -974,7 +1053,7 @@ int uprobe_mmap(struct vm_area_struct *vma)
        struct list_head tmp_list;
        struct uprobe *uprobe, *u;
        struct inode *inode;
-       int ret;
+       int ret, count;
 
        if (!atomic_read(&uprobe_events) || !valid_vma(vma, true))
                return 0;
@@ -988,6 +1067,7 @@ int uprobe_mmap(struct vm_area_struct *vma)
        build_probe_list(inode, &tmp_list);
 
        ret = 0;
+       count = 0;
 
        list_for_each_entry_safe(uprobe, u, &tmp_list, pending_list) {
                loff_t vaddr;
@@ -995,21 +1075,599 @@ int uprobe_mmap(struct vm_area_struct *vma)
                list_del(&uprobe->pending_list);
                if (!ret) {
                        vaddr = vma_address(vma, uprobe->offset);
-                       if (vaddr >= vma->vm_start && vaddr < vma->vm_end) {
-                               ret = install_breakpoint(uprobe, vma->vm_mm, vma, vaddr);
-                               /* Ignore double add: */
-                               if (ret == -EEXIST)
-                                       ret = 0;
+
+                       if (vaddr < vma->vm_start || vaddr >= vma->vm_end) {
+                               put_uprobe(uprobe);
+                               continue;
                        }
+
+                       ret = install_breakpoint(uprobe, vma->vm_mm, vma, vaddr);
+
+                       /* Ignore double add: */
+                       if (ret == -EEXIST) {
+                               ret = 0;
+
+                               if (!is_swbp_at_addr(vma->vm_mm, vaddr))
+                                       continue;
+
+                               /*
+                                * Unable to insert a breakpoint, but
+                                * breakpoint lies underneath. Increment the
+                                * probe count.
+                                */
+                               atomic_inc(&vma->vm_mm->uprobes_state.count);
+                       }
+
+                       if (!ret)
+                               count++;
                }
                put_uprobe(uprobe);
        }
 
        mutex_unlock(uprobes_mmap_hash(inode));
 
+       if (ret)
+               atomic_sub(count, &vma->vm_mm->uprobes_state.count);
+
        return ret;
 }
 
+/*
+ * Called in context of a munmap of a vma.
+ */
+void uprobe_munmap(struct vm_area_struct *vma, unsigned long start, unsigned long end)
+{
+       struct list_head tmp_list;
+       struct uprobe *uprobe, *u;
+       struct inode *inode;
+
+       if (!atomic_read(&uprobe_events) || !valid_vma(vma, false))
+               return;
+
+       if (!atomic_read(&vma->vm_mm->uprobes_state.count))
+               return;
+
+       inode = vma->vm_file->f_mapping->host;
+       if (!inode)
+               return;
+
+       INIT_LIST_HEAD(&tmp_list);
+       mutex_lock(uprobes_mmap_hash(inode));
+       build_probe_list(inode, &tmp_list);
+
+       list_for_each_entry_safe(uprobe, u, &tmp_list, pending_list) {
+               loff_t vaddr;
+
+               list_del(&uprobe->pending_list);
+               vaddr = vma_address(vma, uprobe->offset);
+
+               if (vaddr >= start && vaddr < end) {
+                       /*
+                        * An unregister could have removed the probe before
+                        * unmap. So check before we decrement the count.
+                        */
+                       if (is_swbp_at_addr(vma->vm_mm, vaddr) == 1)
+                               atomic_dec(&vma->vm_mm->uprobes_state.count);
+               }
+               put_uprobe(uprobe);
+       }
+       mutex_unlock(uprobes_mmap_hash(inode));
+}
+
+/* Slot allocation for XOL */
+static int xol_add_vma(struct xol_area *area)
+{
+       struct mm_struct *mm;
+       int ret;
+
+       area->page = alloc_page(GFP_HIGHUSER);
+       if (!area->page)
+               return -ENOMEM;
+
+       ret = -EALREADY;
+       mm = current->mm;
+
+       down_write(&mm->mmap_sem);
+       if (mm->uprobes_state.xol_area)
+               goto fail;
+
+       ret = -ENOMEM;
+
+       /* Try to map as high as possible, this is only a hint. */
+       area->vaddr = get_unmapped_area(NULL, TASK_SIZE - PAGE_SIZE, PAGE_SIZE, 0, 0);
+       if (area->vaddr & ~PAGE_MASK) {
+               ret = area->vaddr;
+               goto fail;
+       }
+
+       ret = install_special_mapping(mm, area->vaddr, PAGE_SIZE,
+                               VM_EXEC|VM_MAYEXEC|VM_DONTCOPY|VM_IO, &area->page);
+       if (ret)
+               goto fail;
+
+       smp_wmb();      /* pairs with get_xol_area() */
+       mm->uprobes_state.xol_area = area;
+       ret = 0;
+
+fail:
+       up_write(&mm->mmap_sem);
+       if (ret)
+               __free_page(area->page);
+
+       return ret;
+}
+
+static struct xol_area *get_xol_area(struct mm_struct *mm)
+{
+       struct xol_area *area;
+
+       area = mm->uprobes_state.xol_area;
+       smp_read_barrier_depends();     /* pairs with wmb in xol_add_vma() */
+
+       return area;
+}
+
+/*
+ * xol_alloc_area - Allocate process's xol_area.
+ * This area will be used for storing instructions for execution out of
+ * line.
+ *
+ * Returns the allocated area or NULL.
+ */
+static struct xol_area *xol_alloc_area(void)
+{
+       struct xol_area *area;
+
+       area = kzalloc(sizeof(*area), GFP_KERNEL);
+       if (unlikely(!area))
+               return NULL;
+
+       area->bitmap = kzalloc(BITS_TO_LONGS(UINSNS_PER_PAGE) * sizeof(long), GFP_KERNEL);
+
+       if (!area->bitmap)
+               goto fail;
+
+       init_waitqueue_head(&area->wq);
+       if (!xol_add_vma(area))
+               return area;
+
+fail:
+       kfree(area->bitmap);
+       kfree(area);
+
+       return get_xol_area(current->mm);
+}
+
+/*
+ * uprobe_clear_state - Free the area allocated for slots.
+ */
+void uprobe_clear_state(struct mm_struct *mm)
+{
+       struct xol_area *area = mm->uprobes_state.xol_area;
+
+       if (!area)
+               return;
+
+       put_page(area->page);
+       kfree(area->bitmap);
+       kfree(area);
+}
+
+/*
+ * uprobe_reset_state - Free the area allocated for slots.
+ */
+void uprobe_reset_state(struct mm_struct *mm)
+{
+       mm->uprobes_state.xol_area = NULL;
+       atomic_set(&mm->uprobes_state.count, 0);
+}
+
+/*
+ *  - search for a free slot.
+ */
+static unsigned long xol_take_insn_slot(struct xol_area *area)
+{
+       unsigned long slot_addr;
+       int slot_nr;
+
+       do {
+               slot_nr = find_first_zero_bit(area->bitmap, UINSNS_PER_PAGE);
+               if (slot_nr < UINSNS_PER_PAGE) {
+                       if (!test_and_set_bit(slot_nr, area->bitmap))
+                               break;
+
+                       slot_nr = UINSNS_PER_PAGE;
+                       continue;
+               }
+               wait_event(area->wq, (atomic_read(&area->slot_count) < UINSNS_PER_PAGE));
+       } while (slot_nr >= UINSNS_PER_PAGE);
+
+       slot_addr = area->vaddr + (slot_nr * UPROBE_XOL_SLOT_BYTES);
+       atomic_inc(&area->slot_count);
+
+       return slot_addr;
+}
+
+/*
+ * xol_get_insn_slot - If was not allocated a slot, then
+ * allocate a slot.
+ * Returns the allocated slot address or 0.
+ */
+static unsigned long xol_get_insn_slot(struct uprobe *uprobe, unsigned long slot_addr)
+{
+       struct xol_area *area;
+       unsigned long offset;
+       void *vaddr;
+
+       area = get_xol_area(current->mm);
+       if (!area) {
+               area = xol_alloc_area();
+               if (!area)
+                       return 0;
+       }
+       current->utask->xol_vaddr = xol_take_insn_slot(area);
+
+       /*
+        * Initialize the slot if xol_vaddr points to valid
+        * instruction slot.
+        */
+       if (unlikely(!current->utask->xol_vaddr))
+               return 0;
+
+       current->utask->vaddr = slot_addr;
+       offset = current->utask->xol_vaddr & ~PAGE_MASK;
+       vaddr = kmap_atomic(area->page);
+       memcpy(vaddr + offset, uprobe->arch.insn, MAX_UINSN_BYTES);
+       kunmap_atomic(vaddr);
+
+       return current->utask->xol_vaddr;
+}
+
+/*
+ * xol_free_insn_slot - If slot was earlier allocated by
+ * @xol_get_insn_slot(), make the slot available for
+ * subsequent requests.
+ */
+static void xol_free_insn_slot(struct task_struct *tsk)
+{
+       struct xol_area *area;
+       unsigned long vma_end;
+       unsigned long slot_addr;
+
+       if (!tsk->mm || !tsk->mm->uprobes_state.xol_area || !tsk->utask)
+               return;
+
+       slot_addr = tsk->utask->xol_vaddr;
+
+       if (unlikely(!slot_addr || IS_ERR_VALUE(slot_addr)))
+               return;
+
+       area = tsk->mm->uprobes_state.xol_area;
+       vma_end = area->vaddr + PAGE_SIZE;
+       if (area->vaddr <= slot_addr && slot_addr < vma_end) {
+               unsigned long offset;
+               int slot_nr;
+
+               offset = slot_addr - area->vaddr;
+               slot_nr = offset / UPROBE_XOL_SLOT_BYTES;
+               if (slot_nr >= UINSNS_PER_PAGE)
+                       return;
+
+               clear_bit(slot_nr, area->bitmap);
+               atomic_dec(&area->slot_count);
+               if (waitqueue_active(&area->wq))
+                       wake_up(&area->wq);
+
+               tsk->utask->xol_vaddr = 0;
+       }
+}
+
+/**
+ * uprobe_get_swbp_addr - compute address of swbp given post-swbp regs
+ * @regs: Reflects the saved state of the task after it has hit a breakpoint
+ * instruction.
+ * Return the address of the breakpoint instruction.
+ */
+unsigned long __weak uprobe_get_swbp_addr(struct pt_regs *regs)
+{
+       return instruction_pointer(regs) - UPROBE_SWBP_INSN_SIZE;
+}
+
+/*
+ * Called with no locks held.
+ * Called in context of a exiting or a exec-ing thread.
+ */
+void uprobe_free_utask(struct task_struct *t)
+{
+       struct uprobe_task *utask = t->utask;
+
+       if (!utask)
+               return;
+
+       if (utask->active_uprobe)
+               put_uprobe(utask->active_uprobe);
+
+       xol_free_insn_slot(t);
+       kfree(utask);
+       t->utask = NULL;
+}
+
+/*
+ * Called in context of a new clone/fork from copy_process.
+ */
+void uprobe_copy_process(struct task_struct *t)
+{
+       t->utask = NULL;
+}
+
+/*
+ * Allocate a uprobe_task object for the task.
+ * Called when the thread hits a breakpoint for the first time.
+ *
+ * Returns:
+ * - pointer to new uprobe_task on success
+ * - NULL otherwise
+ */
+static struct uprobe_task *add_utask(void)
+{
+       struct uprobe_task *utask;
+
+       utask = kzalloc(sizeof *utask, GFP_KERNEL);
+       if (unlikely(!utask))
+               return NULL;
+
+       utask->active_uprobe = NULL;
+       current->utask = utask;
+       return utask;
+}
+
+/* Prepare to single-step probed instruction out of line. */
+static int
+pre_ssout(struct uprobe *uprobe, struct pt_regs *regs, unsigned long vaddr)
+{
+       if (xol_get_insn_slot(uprobe, vaddr) && !arch_uprobe_pre_xol(&uprobe->arch, regs))
+               return 0;
+
+       return -EFAULT;
+}
+
+/*
+ * If we are singlestepping, then ensure this thread is not connected to
+ * non-fatal signals until completion of singlestep.  When xol insn itself
+ * triggers the signal,  restart the original insn even if the task is
+ * already SIGKILL'ed (since coredump should report the correct ip).  This
+ * is even more important if the task has a handler for SIGSEGV/etc, The
+ * _same_ instruction should be repeated again after return from the signal
+ * handler, and SSTEP can never finish in this case.
+ */
+bool uprobe_deny_signal(void)
+{
+       struct task_struct *t = current;
+       struct uprobe_task *utask = t->utask;
+
+       if (likely(!utask || !utask->active_uprobe))
+               return false;
+
+       WARN_ON_ONCE(utask->state != UTASK_SSTEP);
+
+       if (signal_pending(t)) {
+               spin_lock_irq(&t->sighand->siglock);
+               clear_tsk_thread_flag(t, TIF_SIGPENDING);
+               spin_unlock_irq(&t->sighand->siglock);
+
+               if (__fatal_signal_pending(t) || arch_uprobe_xol_was_trapped(t)) {
+                       utask->state = UTASK_SSTEP_TRAPPED;
+                       set_tsk_thread_flag(t, TIF_UPROBE);
+                       set_tsk_thread_flag(t, TIF_NOTIFY_RESUME);
+               }
+       }
+
+       return true;
+}
+
+/*
+ * Avoid singlestepping the original instruction if the original instruction
+ * is a NOP or can be emulated.
+ */
+static bool can_skip_sstep(struct uprobe *uprobe, struct pt_regs *regs)
+{
+       if (arch_uprobe_skip_sstep(&uprobe->arch, regs))
+               return true;
+
+       uprobe->flags &= ~UPROBE_SKIP_SSTEP;
+       return false;
+}
+
+static struct uprobe *find_active_uprobe(unsigned long bp_vaddr, int *is_swbp)
+{
+       struct mm_struct *mm = current->mm;
+       struct uprobe *uprobe = NULL;
+       struct vm_area_struct *vma;
+
+       down_read(&mm->mmap_sem);
+       vma = find_vma(mm, bp_vaddr);
+       if (vma && vma->vm_start <= bp_vaddr) {
+               if (valid_vma(vma, false)) {
+                       struct inode *inode;
+                       loff_t offset;
+
+                       inode = vma->vm_file->f_mapping->host;
+                       offset = bp_vaddr - vma->vm_start;
+                       offset += (vma->vm_pgoff << PAGE_SHIFT);
+                       uprobe = find_uprobe(inode, offset);
+               }
+
+               if (!uprobe)
+                       *is_swbp = is_swbp_at_addr(mm, bp_vaddr);
+       } else {
+               *is_swbp = -EFAULT;
+       }
+       up_read(&mm->mmap_sem);
+
+       return uprobe;
+}
+
+/*
+ * Run handler and ask thread to singlestep.
+ * Ensure all non-fatal signals cannot interrupt thread while it singlesteps.
+ */
+static void handle_swbp(struct pt_regs *regs)
+{
+       struct uprobe_task *utask;
+       struct uprobe *uprobe;
+       unsigned long bp_vaddr;
+       int uninitialized_var(is_swbp);
+
+       bp_vaddr = uprobe_get_swbp_addr(regs);
+       uprobe = find_active_uprobe(bp_vaddr, &is_swbp);
+
+       if (!uprobe) {
+               if (is_swbp > 0) {
+                       /* No matching uprobe; signal SIGTRAP. */
+                       send_sig(SIGTRAP, current, 0);
+               } else {
+                       /*
+                        * Either we raced with uprobe_unregister() or we can't
+                        * access this memory. The latter is only possible if
+                        * another thread plays with our ->mm. In both cases
+                        * we can simply restart. If this vma was unmapped we
+                        * can pretend this insn was not executed yet and get
+                        * the (correct) SIGSEGV after restart.
+                        */
+                       instruction_pointer_set(regs, bp_vaddr);
+               }
+               return;
+       }
+
+       utask = current->utask;
+       if (!utask) {
+               utask = add_utask();
+               /* Cannot allocate; re-execute the instruction. */
+               if (!utask)
+                       goto cleanup_ret;
+       }
+       utask->active_uprobe = uprobe;
+       handler_chain(uprobe, regs);
+       if (uprobe->flags & UPROBE_SKIP_SSTEP && can_skip_sstep(uprobe, regs))
+               goto cleanup_ret;
+
+       utask->state = UTASK_SSTEP;
+       if (!pre_ssout(uprobe, regs, bp_vaddr)) {
+               user_enable_single_step(current);
+               return;
+       }
+
+cleanup_ret:
+       if (utask) {
+               utask->active_uprobe = NULL;
+               utask->state = UTASK_RUNNING;
+       }
+       if (uprobe) {
+               if (!(uprobe->flags & UPROBE_SKIP_SSTEP))
+
+                       /*
+                        * cannot singlestep; cannot skip instruction;
+                        * re-execute the instruction.
+                        */
+                       instruction_pointer_set(regs, bp_vaddr);
+
+               put_uprobe(uprobe);
+       }
+}
+
+/*
+ * Perform required fix-ups and disable singlestep.
+ * Allow pending signals to take effect.
+ */
+static void handle_singlestep(struct uprobe_task *utask, struct pt_regs *regs)
+{
+       struct uprobe *uprobe;
+
+       uprobe = utask->active_uprobe;
+       if (utask->state == UTASK_SSTEP_ACK)
+               arch_uprobe_post_xol(&uprobe->arch, regs);
+       else if (utask->state == UTASK_SSTEP_TRAPPED)
+               arch_uprobe_abort_xol(&uprobe->arch, regs);
+       else
+               WARN_ON_ONCE(1);
+
+       put_uprobe(uprobe);
+       utask->active_uprobe = NULL;
+       utask->state = UTASK_RUNNING;
+       user_disable_single_step(current);
+       xol_free_insn_slot(current);
+
+       spin_lock_irq(&current->sighand->siglock);
+       recalc_sigpending(); /* see uprobe_deny_signal() */
+       spin_unlock_irq(&current->sighand->siglock);
+}
+
+/*
+ * On breakpoint hit, breakpoint notifier sets the TIF_UPROBE flag.  (and on
+ * subsequent probe hits on the thread sets the state to UTASK_BP_HIT) and
+ * allows the thread to return from interrupt.
+ *
+ * On singlestep exception, singlestep notifier sets the TIF_UPROBE flag and
+ * also sets the state to UTASK_SSTEP_ACK and allows the thread to return from
+ * interrupt.
+ *
+ * While returning to userspace, thread notices the TIF_UPROBE flag and calls
+ * uprobe_notify_resume().
+ */
+void uprobe_notify_resume(struct pt_regs *regs)
+{
+       struct uprobe_task *utask;
+
+       utask = current->utask;
+       if (!utask || utask->state == UTASK_BP_HIT)
+               handle_swbp(regs);
+       else
+               handle_singlestep(utask, regs);
+}
+
+/*
+ * uprobe_pre_sstep_notifier gets called from interrupt context as part of
+ * notifier mechanism. Set TIF_UPROBE flag and indicate breakpoint hit.
+ */
+int uprobe_pre_sstep_notifier(struct pt_regs *regs)
+{
+       struct uprobe_task *utask;
+
+       if (!current->mm || !atomic_read(&current->mm->uprobes_state.count))
+               /* task is currently not uprobed */
+               return 0;
+
+       utask = current->utask;
+       if (utask)
+               utask->state = UTASK_BP_HIT;
+
+       set_thread_flag(TIF_UPROBE);
+
+       return 1;
+}
+
+/*
+ * uprobe_post_sstep_notifier gets called in interrupt context as part of notifier
+ * mechanism. Set TIF_UPROBE flag and indicate completion of singlestep.
+ */
+int uprobe_post_sstep_notifier(struct pt_regs *regs)
+{
+       struct uprobe_task *utask = current->utask;
+
+       if (!current->mm || !utask || !utask->active_uprobe)
+               /* task is currently not uprobed */
+               return 0;
+
+       utask->state = UTASK_SSTEP_ACK;
+       set_thread_flag(TIF_UPROBE);
+       return 1;
+}
+
+static struct notifier_block uprobe_exception_nb = {
+       .notifier_call          = arch_uprobe_exception_notify,
+       .priority               = INT_MAX-1,    /* notified after kprobes, kgdb */
+};
+
 static int __init init_uprobes(void)
 {
        int i;
@@ -1018,12 +1676,12 @@ static int __init init_uprobes(void)
                mutex_init(&uprobes_mutex[i]);
                mutex_init(&uprobes_mmap_mutex[i]);
        }
-       return 0;
+
+       return register_die_notifier(&uprobe_exception_nb);
 }
+module_init(init_uprobes);
 
 static void __exit exit_uprobes(void)
 {
 }
-
-module_init(init_uprobes);
 module_exit(exit_uprobes);