mm/hmm: Fix error flows in hmm_invalidate_range_start
authorJason Gunthorpe <jgg@mellanox.com>
Fri, 7 Jun 2019 15:10:33 +0000 (12:10 -0300)
committerJason Gunthorpe <jgg@mellanox.com>
Thu, 27 Jun 2019 16:05:02 +0000 (13:05 -0300)
If the trylock on the hmm->mirrors_sem fails the function will return
without decrementing the notifiers that were previously incremented. Since
the caller will not call invalidate_range_end() on EAGAIN this will result
in notifiers becoming permanently incremented and deadlock.

If the sync_cpu_device_pagetables() required blocking the function will
not return EAGAIN even though the device continues to touch the
pages. This is a violation of the mmu notifier contract.

Switch, and rename, the ranges_lock to a spin lock so we can reliably
obtain it without blocking during error unwind.

The error unwind is necessary since the notifiers count must be held
incremented across the call to sync_cpu_device_pagetables() as we cannot
allow the range to become marked valid by a parallel
invalidate_start/end() pair while doing sync_cpu_device_pagetables().

Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Reviewed-by: Ralph Campbell <rcampbell@nvidia.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Philip Yang <Philip.Yang@amd.com>
include/linux/hmm.h
mm/hmm.c

index bf013e9..0fa8ea3 100644 (file)
@@ -86,7 +86,7 @@
 struct hmm {
        struct mm_struct        *mm;
        struct kref             kref;
-       struct mutex            lock;
+       spinlock_t              ranges_lock;
        struct list_head        ranges;
        struct list_head        mirrors;
        struct mmu_notifier     mmu_notifier;
index b224ea6..de35289 100644 (file)
--- a/mm/hmm.c
+++ b/mm/hmm.c
@@ -64,7 +64,7 @@ static struct hmm *hmm_get_or_create(struct mm_struct *mm)
        init_rwsem(&hmm->mirrors_sem);
        hmm->mmu_notifier.ops = NULL;
        INIT_LIST_HEAD(&hmm->ranges);
-       mutex_init(&hmm->lock);
+       spin_lock_init(&hmm->ranges_lock);
        kref_init(&hmm->kref);
        hmm->notifiers = 0;
        hmm->mm = mm;
@@ -144,6 +144,25 @@ static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
        hmm_put(hmm);
 }
 
+static void notifiers_decrement(struct hmm *hmm)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave(&hmm->ranges_lock, flags);
+       hmm->notifiers--;
+       if (!hmm->notifiers) {
+               struct hmm_range *range;
+
+               list_for_each_entry(range, &hmm->ranges, list) {
+                       if (range->valid)
+                               continue;
+                       range->valid = true;
+               }
+               wake_up_all(&hmm->wq);
+       }
+       spin_unlock_irqrestore(&hmm->ranges_lock, flags);
+}
+
 static int hmm_invalidate_range_start(struct mmu_notifier *mn,
                        const struct mmu_notifier_range *nrange)
 {
@@ -151,6 +170,7 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn,
        struct hmm_mirror *mirror;
        struct hmm_update update;
        struct hmm_range *range;
+       unsigned long flags;
        int ret = 0;
 
        if (!kref_get_unless_zero(&hmm->kref))
@@ -161,12 +181,7 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn,
        update.event = HMM_UPDATE_INVALIDATE;
        update.blockable = mmu_notifier_range_blockable(nrange);
 
-       if (mmu_notifier_range_blockable(nrange))
-               mutex_lock(&hmm->lock);
-       else if (!mutex_trylock(&hmm->lock)) {
-               ret = -EAGAIN;
-               goto out;
-       }
+       spin_lock_irqsave(&hmm->ranges_lock, flags);
        hmm->notifiers++;
        list_for_each_entry(range, &hmm->ranges, list) {
                if (update.end < range->start || update.start >= range->end)
@@ -174,7 +189,7 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn,
 
                range->valid = false;
        }
-       mutex_unlock(&hmm->lock);
+       spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 
        if (mmu_notifier_range_blockable(nrange))
                down_read(&hmm->mirrors_sem);
@@ -182,16 +197,23 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn,
                ret = -EAGAIN;
                goto out;
        }
+
        list_for_each_entry(mirror, &hmm->mirrors, list) {
-               int ret;
+               int rc;
 
-               ret = mirror->ops->sync_cpu_device_pagetables(mirror, &update);
-               if (!update.blockable && ret == -EAGAIN)
+               rc = mirror->ops->sync_cpu_device_pagetables(mirror, &update);
+               if (rc) {
+                       if (WARN_ON(update.blockable || rc != -EAGAIN))
+                               continue;
+                       ret = -EAGAIN;
                        break;
+               }
        }
        up_read(&hmm->mirrors_sem);
 
 out:
+       if (ret)
+               notifiers_decrement(hmm);
        hmm_put(hmm);
        return ret;
 }
@@ -204,20 +226,7 @@ static void hmm_invalidate_range_end(struct mmu_notifier *mn,
        if (!kref_get_unless_zero(&hmm->kref))
                return;
 
-       mutex_lock(&hmm->lock);
-       hmm->notifiers--;
-       if (!hmm->notifiers) {
-               struct hmm_range *range;
-
-               list_for_each_entry(range, &hmm->ranges, list) {
-                       if (range->valid)
-                               continue;
-                       range->valid = true;
-               }
-               wake_up_all(&hmm->wq);
-       }
-       mutex_unlock(&hmm->lock);
-
+       notifiers_decrement(hmm);
        hmm_put(hmm);
 }
 
@@ -868,6 +877,7 @@ int hmm_range_register(struct hmm_range *range,
 {
        unsigned long mask = ((1UL << page_shift) - 1UL);
        struct hmm *hmm = mirror->hmm;
+       unsigned long flags;
 
        range->valid = false;
        range->hmm = NULL;
@@ -886,7 +896,7 @@ int hmm_range_register(struct hmm_range *range,
                return -EFAULT;
 
        /* Initialize range to track CPU page table updates. */
-       mutex_lock(&hmm->lock);
+       spin_lock_irqsave(&hmm->ranges_lock, flags);
 
        range->hmm = hmm;
        kref_get(&hmm->kref);
@@ -898,7 +908,7 @@ int hmm_range_register(struct hmm_range *range,
         */
        if (!hmm->notifiers)
                range->valid = true;
-       mutex_unlock(&hmm->lock);
+       spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 
        return 0;
 }
@@ -914,10 +924,11 @@ EXPORT_SYMBOL(hmm_range_register);
 void hmm_range_unregister(struct hmm_range *range)
 {
        struct hmm *hmm = range->hmm;
+       unsigned long flags;
 
-       mutex_lock(&hmm->lock);
+       spin_lock_irqsave(&hmm->ranges_lock, flags);
        list_del_init(&range->list);
-       mutex_unlock(&hmm->lock);
+       spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 
        /* Drop reference taken by hmm_range_register() */
        mmput(hmm->mm);