[PATCH v3 0/8] binder: faster page installations

[PATCH v3 0/8] binder: faster page installations

[Carlos Llamas]

The main focus of these patches is to improve the performance of binder
page installations, primarily by reducing contention on the mmap_lock.
The idea is to allow concurrent page insertion by leveraging per-vma
locking and get_user_pages_remote().

Unfortunately, this required reverting the alloc->lock spinlock back
into a mutex in order to serialize with the shrinker. At least until
finding a better solution e.g. support page zapping with a spinlock.
The trade off is still quite worth it though.

Other patches are also included that remove unsafe and redundant things
such as the alloc->vma pointer or the struct binder_lru_page concept.

Note: I'll work on setting up a page fault handler for binder next.
I believe an idea from Alice Ryhl to deferred the page insertions will
make this finally feasible. I only need to figure out a few performance
bits but if/when done most of the manual page insertion code in binder
could be dropped. :)

Changelog:

v3:
 * collect "Reviewed-by" tags from Suren
 * use full commit subject in revert (not only sha1)
 * add "goto unlock" label for vma_lookup() failures
 * address -ENOMEM error override in separate patch
 * squash "remove alloc->vma" patch into alloc->mapped patch
 * pass 'struct binder_alloc *' to  binder_page_lookup() too
 * factor out individual mmget_not_zero()/mmput_async() calls
 * cleanup binder_page_insert() to avoid using goto
 * document only one mapping allowed per binder instance
 * check binder_alloc_is_mapped() in binder_page_lookup()
 * remove no longer need local page_to_free pointer

v2:
https://lore.kernel.org/all/20241107040239.2847143-1-cmllamas@google.com/
 * fix locking order when upgrading from vma lock to mmap lock
 * switch folio_walk_start() for get_user_pages_remote()
 * release vma/mmap locks and mmput() right after vm_insert_page()
 * add binder_page_alloc() helper for binder_install_single_page()

v1:
https://lore.kernel.org/all/20241105200258.2380168-1-cmllamas@google.com/

Cc: Alice Ryhl <aliceryhl@google.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Todd Kjos <tkjos@google.com>
Cc: Viktor Martensson <vmartensson@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

Carlos Llamas (8):
  Revert "binder: switch alloc->mutex to spinlock_t"
  binder: concurrent page installation
  binder: select correct nid for pages in LRU
  binder: remove struct binder_lru_page
  binder: replace alloc->vma with alloc->mapped
  binder: rename alloc->buffer to vm_start
  binder: use per-vma lock in page installation
  binder: propagate vm_insert_page() errors

 drivers/android/binder.c                |   2 +-
 drivers/android/binder_alloc.c          | 313 ++++++++++++++----------
 drivers/android/binder_alloc.h          |  36 +--
 drivers/android/binder_alloc_selftest.c |  18 +-
 drivers/android/binder_trace.h          |   2 +-
 5 files changed, 207 insertions(+), 164 deletions(-)

-- 
2.47.0.277.g8800431eea-goog

[PATCH v3 1/8] Revert "binder: switch alloc->mutex to spinlock_t"

[Carlos Llamas]

This reverts commit 7710e2cca32e7f3958480e8bd44f50e29d0c2509.

In preparation for concurrent page installations, restore the original
alloc->mutex which will serialize zap_page_range_single() against page
installations in subsequent patches (instead of the mmap_sem).

Trivial conflicts with commit 2c10a20f5e84a ("binder_alloc: Fix sleeping
function called from invalid context") were resolved.

Cc: Mukesh Ojha <quic_mojha@quicinc.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Carlos Llamas <cmllamas@google.com>
---
 drivers/android/binder_alloc.c | 46 +++++++++++++++++-----------------
 drivers/android/binder_alloc.h | 10 ++++----
 2 files changed, 28 insertions(+), 28 deletions(-)

diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
index b3acbc4174fb..7241bf4a3ff2 100644
--- a/drivers/android/binder_alloc.c
+++ b/drivers/android/binder_alloc.c
@@ -169,9 +169,9 @@ struct binder_buffer *binder_alloc_prepare_to_free(struct binder_alloc *alloc,
 {
 	struct binder_buffer *buffer;
 
-	spin_lock(&alloc->lock);
+	mutex_lock(&alloc->mutex);
 	buffer = binder_alloc_prepare_to_free_locked(alloc, user_ptr);
-	spin_unlock(&alloc->lock);
+	mutex_unlock(&alloc->mutex);
 	return buffer;
 }
 
@@ -597,10 +597,10 @@ struct binder_buffer *binder_alloc_new_buf(struct binder_alloc *alloc,
 	if (!next)
 		return ERR_PTR(-ENOMEM);
 
-	spin_lock(&alloc->lock);
+	mutex_lock(&alloc->mutex);
 	buffer = binder_alloc_new_buf_locked(alloc, next, size, is_async);
 	if (IS_ERR(buffer)) {
-		spin_unlock(&alloc->lock);
+		mutex_unlock(&alloc->mutex);
 		goto out;
 	}
 
@@ -608,7 +608,7 @@ struct binder_buffer *binder_alloc_new_buf(struct binder_alloc *alloc,
 	buffer->offsets_size = offsets_size;
 	buffer->extra_buffers_size = extra_buffers_size;
 	buffer->pid = current->tgid;
-	spin_unlock(&alloc->lock);
+	mutex_unlock(&alloc->mutex);
 
 	ret = binder_install_buffer_pages(alloc, buffer, size);
 	if (ret) {
@@ -785,17 +785,17 @@ void binder_alloc_free_buf(struct binder_alloc *alloc,
 	 * We could eliminate the call to binder_alloc_clear_buf()
 	 * from binder_alloc_deferred_release() by moving this to
 	 * binder_free_buf_locked(). However, that could
-	 * increase contention for the alloc->lock if clear_on_free
-	 * is used frequently for large buffers. This lock is not
+	 * increase contention for the alloc mutex if clear_on_free
+	 * is used frequently for large buffers. The mutex is not
 	 * needed for correctness here.
 	 */
 	if (buffer->clear_on_free) {
 		binder_alloc_clear_buf(alloc, buffer);
 		buffer->clear_on_free = false;
 	}
-	spin_lock(&alloc->lock);
+	mutex_lock(&alloc->mutex);
 	binder_free_buf_locked(alloc, buffer);
-	spin_unlock(&alloc->lock);
+	mutex_unlock(&alloc->mutex);
 }
 
 /**
@@ -893,7 +893,7 @@ void binder_alloc_deferred_release(struct binder_alloc *alloc)
 	struct binder_buffer *buffer;
 
 	buffers = 0;
-	spin_lock(&alloc->lock);
+	mutex_lock(&alloc->mutex);
 	BUG_ON(alloc->vma);
 
 	while ((n = rb_first(&alloc->allocated_buffers))) {
@@ -940,7 +940,7 @@ void binder_alloc_deferred_release(struct binder_alloc *alloc)
 			page_count++;
 		}
 	}
-	spin_unlock(&alloc->lock);
+	mutex_unlock(&alloc->mutex);
 	kvfree(alloc->pages);
 	if (alloc->mm)
 		mmdrop(alloc->mm);
@@ -964,7 +964,7 @@ void binder_alloc_print_allocated(struct seq_file *m,
 	struct binder_buffer *buffer;
 	struct rb_node *n;
 
-	spin_lock(&alloc->lock);
+	mutex_lock(&alloc->mutex);
 	for (n = rb_first(&alloc->allocated_buffers); n; n = rb_next(n)) {
 		buffer = rb_entry(n, struct binder_buffer, rb_node);
 		seq_printf(m, "  buffer %d: %lx size %zd:%zd:%zd %s\n",
@@ -974,7 +974,7 @@ void binder_alloc_print_allocated(struct seq_file *m,
 			   buffer->extra_buffers_size,
 			   buffer->transaction ? "active" : "delivered");
 	}
-	spin_unlock(&alloc->lock);
+	mutex_unlock(&alloc->mutex);
 }
 
 /**
@@ -991,7 +991,7 @@ void binder_alloc_print_pages(struct seq_file *m,
 	int lru = 0;
 	int free = 0;
 
-	spin_lock(&alloc->lock);
+	mutex_lock(&alloc->mutex);
 	/*
 	 * Make sure the binder_alloc is fully initialized, otherwise we might
 	 * read inconsistent state.
@@ -1007,7 +1007,7 @@ void binder_alloc_print_pages(struct seq_file *m,
 				lru++;
 		}
 	}
-	spin_unlock(&alloc->lock);
+	mutex_unlock(&alloc->mutex);
 	seq_printf(m, "  pages: %d:%d:%d\n", active, lru, free);
 	seq_printf(m, "  pages high watermark: %zu\n", alloc->pages_high);
 }
@@ -1023,10 +1023,10 @@ int binder_alloc_get_allocated_count(struct binder_alloc *alloc)
 	struct rb_node *n;
 	int count = 0;
 
-	spin_lock(&alloc->lock);
+	mutex_lock(&alloc->mutex);
 	for (n = rb_first(&alloc->allocated_buffers); n != NULL; n = rb_next(n))
 		count++;
-	spin_unlock(&alloc->lock);
+	mutex_unlock(&alloc->mutex);
 	return count;
 }
 
@@ -1071,8 +1071,8 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 		goto err_mmget;
 	if (!mmap_read_trylock(mm))
 		goto err_mmap_read_lock_failed;
-	if (!spin_trylock(&alloc->lock))
-		goto err_get_alloc_lock_failed;
+	if (!mutex_trylock(&alloc->mutex))
+		goto err_get_alloc_mutex_failed;
 	if (!page->page_ptr)
 		goto err_page_already_freed;
 
@@ -1091,7 +1091,7 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 	trace_binder_unmap_kernel_end(alloc, index);
 
 	list_lru_isolate(lru, item);
-	spin_unlock(&alloc->lock);
+	mutex_unlock(&alloc->mutex);
 	spin_unlock(lock);
 
 	if (vma) {
@@ -1111,8 +1111,8 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 
 err_invalid_vma:
 err_page_already_freed:
-	spin_unlock(&alloc->lock);
-err_get_alloc_lock_failed:
+	mutex_unlock(&alloc->mutex);
+err_get_alloc_mutex_failed:
 	mmap_read_unlock(mm);
 err_mmap_read_lock_failed:
 	mmput_async(mm);
@@ -1147,7 +1147,7 @@ void binder_alloc_init(struct binder_alloc *alloc)
 	alloc->pid = current->group_leader->pid;
 	alloc->mm = current->mm;
 	mmgrab(alloc->mm);
-	spin_lock_init(&alloc->lock);
+	mutex_init(&alloc->mutex);
 	INIT_LIST_HEAD(&alloc->buffers);
 }
 
diff --git a/drivers/android/binder_alloc.h b/drivers/android/binder_alloc.h
index 70387234477e..a5181916942e 100644
--- a/drivers/android/binder_alloc.h
+++ b/drivers/android/binder_alloc.h
@@ -9,7 +9,7 @@
 #include <linux/rbtree.h>
 #include <linux/list.h>
 #include <linux/mm.h>
-#include <linux/spinlock.h>
+#include <linux/rtmutex.h>
 #include <linux/vmalloc.h>
 #include <linux/slab.h>
 #include <linux/list_lru.h>
@@ -72,7 +72,7 @@ struct binder_lru_page {
 
 /**
  * struct binder_alloc - per-binder proc state for binder allocator
- * @lock:               protects binder_alloc fields
+ * @mutex:              protects binder_alloc fields
  * @vma:                vm_area_struct passed to mmap_handler
  *                      (invariant after mmap)
  * @mm:                 copy of task->mm (invariant after open)
@@ -96,7 +96,7 @@ struct binder_lru_page {
  * struct binder_buffer objects used to track the user buffers
  */
 struct binder_alloc {
-	spinlock_t lock;
+	struct mutex mutex;
 	struct vm_area_struct *vma;
 	struct mm_struct *mm;
 	unsigned long buffer;
@@ -153,9 +153,9 @@ binder_alloc_get_free_async_space(struct binder_alloc *alloc)
 {
 	size_t free_async_space;
 
-	spin_lock(&alloc->lock);
+	mutex_lock(&alloc->mutex);
 	free_async_space = alloc->free_async_space;
-	spin_unlock(&alloc->lock);
+	mutex_unlock(&alloc->mutex);
 	return free_async_space;
 }
 
-- 
2.47.0.277.g8800431eea-goog

[PATCH v3 2/8] binder: concurrent page installation

[Carlos Llamas]

Allow multiple callers to install pages simultaneously by downgrading
the mmap_sem to non-exclusive mode. Races to the same PTE are handled
using get_user_pages_remote() to retrieve the already installed page.
This method significantly reduces contention in the mmap semaphore.

To ensure safety, vma_lookup() is used (instead of alloc->vma) to avoid
operating on an isolated VMA. In addition, zap_page_range_single() is
called under the alloc->mutex to avoid racing with the shrinker.

Many thanks to Barry Song who posted a similar approach [1].

Link: https://lore.kernel.org/all/20240902225009.34576-1-21cnbao@gmail.com/ [1]
Cc: David Hildenbrand <david@redhat.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Carlos Llamas <cmllamas@google.com>
---
 drivers/android/binder_alloc.c | 64 +++++++++++++++++++++-------------
 1 file changed, 40 insertions(+), 24 deletions(-)

diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
index 7241bf4a3ff2..2ab520c285b3 100644
--- a/drivers/android/binder_alloc.c
+++ b/drivers/android/binder_alloc.c
@@ -221,26 +221,14 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 				      struct binder_lru_page *lru_page,
 				      unsigned long addr)
 {
+	struct vm_area_struct *vma;
 	struct page *page;
-	int ret = 0;
+	long npages;
+	int ret;
 
 	if (!mmget_not_zero(alloc->mm))
 		return -ESRCH;
 
-	/*
-	 * Protected with mmap_sem in write mode as multiple tasks
-	 * might race to install the same page.
-	 */
-	mmap_write_lock(alloc->mm);
-	if (binder_get_installed_page(lru_page))
-		goto out;
-
-	if (!alloc->vma) {
-		pr_err("%d: %s failed, no vma\n", alloc->pid, __func__);
-		ret = -ESRCH;
-		goto out;
-	}
-
 	page = alloc_page(GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
 	if (!page) {
 		pr_err("%d: failed to allocate page\n", alloc->pid);
@@ -248,19 +236,47 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 		goto out;
 	}
 
-	ret = vm_insert_page(alloc->vma, addr, page);
-	if (ret) {
+	mmap_read_lock(alloc->mm);
+	vma = vma_lookup(alloc->mm, addr);
+	if (!vma || vma != alloc->vma) {
+		__free_page(page);
+		pr_err("%d: %s failed, no vma\n", alloc->pid, __func__);
+		ret = -ESRCH;
+		goto unlock;
+	}
+
+	ret = vm_insert_page(vma, addr, page);
+	switch (ret) {
+	case -EBUSY:
+		/*
+		 * EBUSY is ok. Someone installed the pte first but the
+		 * lru_page->page_ptr has not been updated yet. Discard
+		 * our page and look up the one already installed.
+		 */
+		ret = 0;
+		__free_page(page);
+		npages = get_user_pages_remote(alloc->mm, addr, 1, 0, &page, NULL);
+		if (npages <= 0) {
+			pr_err("%d: failed to find page at offset %lx\n",
+			       alloc->pid, addr - alloc->buffer);
+			ret = -ESRCH;
+			break;
+		}
+		fallthrough;
+	case 0:
+		/* Mark page installation complete and safe to use */
+		binder_set_installed_page(lru_page, page);
+		break;
+	default:
+		__free_page(page);
 		pr_err("%d: %s failed to insert page at offset %lx with %d\n",
 		       alloc->pid, __func__, addr - alloc->buffer, ret);
-		__free_page(page);
 		ret = -ENOMEM;
-		goto out;
+		break;
 	}
-
-	/* Mark page installation complete and safe to use */
-	binder_set_installed_page(lru_page, page);
+unlock:
+	mmap_read_unlock(alloc->mm);
 out:
-	mmap_write_unlock(alloc->mm);
 	mmput_async(alloc->mm);
 	return ret;
 }
@@ -1091,7 +1107,6 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 	trace_binder_unmap_kernel_end(alloc, index);
 
 	list_lru_isolate(lru, item);
-	mutex_unlock(&alloc->mutex);
 	spin_unlock(lock);
 
 	if (vma) {
@@ -1102,6 +1117,7 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 		trace_binder_unmap_user_end(alloc, index);
 	}
 
+	mutex_unlock(&alloc->mutex);
 	mmap_read_unlock(mm);
 	mmput_async(mm);
 	__free_page(page_to_free);
-- 
2.47.0.277.g8800431eea-goog

Re: [PATCH v3 2/8] binder: concurrent page installation

[Liam R. Howlett]

* Carlos Llamas <cmllamas@google.com> [241108 14:11]:
> Allow multiple callers to install pages simultaneously by downgrading
> the mmap_sem to non-exclusive mode.

Since we actually allow downgrading of a lock, it would be more clear to
say that you are changing from a read lock to a write lock.  I was
searching for the lock downgrade in this patch :)

>Races to the same PTE are handled
> using get_user_pages_remote() to retrieve the already installed page.
> This method significantly reduces contention in the mmap semaphore.
> 
> To ensure safety, vma_lookup() is used (instead of alloc->vma) to avoid
> operating on an isolated VMA. In addition, zap_page_range_single() is
> called under the alloc->mutex to avoid racing with the shrinker.
> 
> Many thanks to Barry Song who posted a similar approach [1].
> 
> Link: https://lore.kernel.org/all/20240902225009.34576-1-21cnbao@gmail.com/ [1]
> Cc: David Hildenbrand <david@redhat.com>
> Cc: Barry Song <v-songbaohua@oppo.com>
> Cc: Suren Baghdasaryan <surenb@google.com>
> Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
> Reviewed-by: Suren Baghdasaryan <surenb@google.com>
> Signed-off-by: Carlos Llamas <cmllamas@google.com>
> ---
>  drivers/android/binder_alloc.c | 64 +++++++++++++++++++++-------------
>  1 file changed, 40 insertions(+), 24 deletions(-)
> 
> diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
> index 7241bf4a3ff2..2ab520c285b3 100644
> --- a/drivers/android/binder_alloc.c
> +++ b/drivers/android/binder_alloc.c
> @@ -221,26 +221,14 @@ static int binder_install_single_page(struct binder_alloc *alloc,
>  				      struct binder_lru_page *lru_page,
>  				      unsigned long addr)
>  {
> +	struct vm_area_struct *vma;
>  	struct page *page;
> -	int ret = 0;
> +	long npages;
> +	int ret;
>  
>  	if (!mmget_not_zero(alloc->mm))
>  		return -ESRCH;
>  
> -	/*
> -	 * Protected with mmap_sem in write mode as multiple tasks
> -	 * might race to install the same page.
> -	 */
> -	mmap_write_lock(alloc->mm);
> -	if (binder_get_installed_page(lru_page))
> -		goto out;
> -
> -	if (!alloc->vma) {
> -		pr_err("%d: %s failed, no vma\n", alloc->pid, __func__);
> -		ret = -ESRCH;
> -		goto out;
> -	}
> -
>  	page = alloc_page(GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
>  	if (!page) {
>  		pr_err("%d: failed to allocate page\n", alloc->pid);
> @@ -248,19 +236,47 @@ static int binder_install_single_page(struct binder_alloc *alloc,
>  		goto out;
>  	}
>  
> -	ret = vm_insert_page(alloc->vma, addr, page);
> -	if (ret) {
> +	mmap_read_lock(alloc->mm);

Ah, I debate bring this up, but you can do this without the mmap read
lock.  You can use rcu and the per-vma locking like in the page fault
path.  If you look at how that is done using the lock_vma_under_rcu()
(mm/memory.c) then you can see that pages are installed today without
the mmap locking (on some architectures - which includes x86_64 and
arm64).

userfaultfd had to do something like this as well, and created some
abstraction to facilitate either mmap or rcu, based on the arch support.
Going to rcu really helped performance there [1].  There was also a
chance of the vma being missed, so it is checked again under the mmap
read lock, but realistically that never happens and exists to ensure
correctness.

You also mention the shrinker and using the alloc->mutex, well zapping
pages can also be done under the vma specific lock - if that helps?

Freeing page tables is different though, that needs more locking, but I
don't think this is an issue for you.

[1]. https://lore.kernel.org/all/20240215182756.3448972-1-lokeshgidra@google.com/

> +	vma = vma_lookup(alloc->mm, addr);

Thank you for not saving a pointer anymore.

> +	if (!vma || vma != alloc->vma) {
> +		__free_page(page);
> +		pr_err("%d: %s failed, no vma\n", alloc->pid, __func__);
> +		ret = -ESRCH;
> +		goto unlock;
> +	}
> +
> +	ret = vm_insert_page(vma, addr, page);
> +	switch (ret) {
> +	case -EBUSY:
> +		/*
> +		 * EBUSY is ok. Someone installed the pte first but the
> +		 * lru_page->page_ptr has not been updated yet. Discard
> +		 * our page and look up the one already installed.
> +		 */
> +		ret = 0;
> +		__free_page(page);
> +		npages = get_user_pages_remote(alloc->mm, addr, 1, 0, &page, NULL);
> +		if (npages <= 0) {
> +			pr_err("%d: failed to find page at offset %lx\n",
> +			       alloc->pid, addr - alloc->buffer);
> +			ret = -ESRCH;
> +			break;
> +		}
> +		fallthrough;
> +	case 0:
> +		/* Mark page installation complete and safe to use */
> +		binder_set_installed_page(lru_page, page);
> +		break;
> +	default:
> +		__free_page(page);
>  		pr_err("%d: %s failed to insert page at offset %lx with %d\n",
>  		       alloc->pid, __func__, addr - alloc->buffer, ret);
> -		__free_page(page);
>  		ret = -ENOMEM;
> -		goto out;
> +		break;
>  	}
> -
> -	/* Mark page installation complete and safe to use */
> -	binder_set_installed_page(lru_page, page);
> +unlock:
> +	mmap_read_unlock(alloc->mm);
>  out:
> -	mmap_write_unlock(alloc->mm);
>  	mmput_async(alloc->mm);
>  	return ret;
>  }
> @@ -1091,7 +1107,6 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
>  	trace_binder_unmap_kernel_end(alloc, index);
>  
>  	list_lru_isolate(lru, item);
> -	mutex_unlock(&alloc->mutex);
>  	spin_unlock(lock);
>  
>  	if (vma) {
> @@ -1102,6 +1117,7 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
>  		trace_binder_unmap_user_end(alloc, index);
>  	}
>  
> +	mutex_unlock(&alloc->mutex);
>  	mmap_read_unlock(mm);
>  	mmput_async(mm);
>  	__free_page(page_to_free);
> -- 
> 2.47.0.277.g8800431eea-goog
> 

Re: [PATCH v3 2/8] binder: concurrent page installation

[Carlos Llamas]

On Fri, Nov 08, 2024 at 03:41:41PM -0500, Liam R. Howlett wrote:
> * Carlos Llamas <cmllamas@google.com> [241108 14:11]:
> > Allow multiple callers to install pages simultaneously by downgrading
> > the mmap_sem to non-exclusive mode.
> 
> Since we actually allow downgrading of a lock, it would be more clear to
> say that you are changing from a read lock to a write lock.  I was
> searching for the lock downgrade in this patch :)

ah, I can see how the wording here can be misleading. I'll rephrase that
to avoid confusion.

> > -	ret = vm_insert_page(alloc->vma, addr, page);
> > -	if (ret) {
> > +	mmap_read_lock(alloc->mm);
> 
> Ah, I debate bring this up, but you can do this without the mmap read
> lock.  You can use rcu and the per-vma locking like in the page fault
> path.  If you look at how that is done using the lock_vma_under_rcu()
> (mm/memory.c) then you can see that pages are installed today without
> the mmap locking (on some architectures - which includes x86_64 and
> arm64).

Right, per-vma locking is implemented in patch 7 of this series:
https://lore.kernel.org/all/20241108191057.3288442-8-cmllamas@google.com/

> userfaultfd had to do something like this as well, and created some
> abstraction to facilitate either mmap or rcu, based on the arch support.
> Going to rcu really helped performance there [1].  There was also a
> chance of the vma being missed, so it is checked again under the mmap
> read lock, but realistically that never happens and exists to ensure
> correctness.

hmm, if there are more users of this pattern in the future then perhaps
it might be worth adding a common helper?

> You also mention the shrinker and using the alloc->mutex, well zapping
> pages can also be done under the vma specific lock - if that helps?
> 
> Freeing page tables is different though, that needs more locking, but I
> don't think this is an issue for you.
> 
> [1]. https://lore.kernel.org/all/20240215182756.3448972-1-lokeshgidra@google.com/

ha, I was not aware of this. I'll have a look thanks.

> 
> > +	vma = vma_lookup(alloc->mm, addr);
> 
> Thank you for not saving a pointer anymore.

lol, thanks to you for initiating this effort!

Cheers,
Carlos Llamas

Re: [PATCH v3 2/8] binder: concurrent page installation

[Liam R. Howlett]

* Carlos Llamas <cmllamas@google.com> [241108 17:26]:

...
> > > -	ret = vm_insert_page(alloc->vma, addr, page);
> > > -	if (ret) {
> > > +	mmap_read_lock(alloc->mm);
> > 
> > Ah, I debate bring this up, but you can do this without the mmap read
> > lock.  You can use rcu and the per-vma locking like in the page fault
> > path.  If you look at how that is done using the lock_vma_under_rcu()
> > (mm/memory.c) then you can see that pages are installed today without
> > the mmap locking (on some architectures - which includes x86_64 and
> > arm64).
> 
> Right, per-vma locking is implemented in patch 7 of this series:
> https://lore.kernel.org/all/20241108191057.3288442-8-cmllamas@google.com/
> 

Oh, sorry I missed that.  I had not pulled the entire patch set for
review due to other tasks, but wanted to say something before this went
in.

Is there any performance numbers on how much this helps?

> 
> > userfaultfd had to do something like this as well, and created some
> > abstraction to facilitate either mmap or rcu, based on the arch support.
> > Going to rcu really helped performance there [1].  There was also a
> > chance of the vma being missed, so it is checked again under the mmap
> > read lock, but realistically that never happens and exists to ensure
> > correctness.
> 
> hmm, if there are more users of this pattern in the future then perhaps
> it might be worth adding a common helper?

Yes.  But the userfaultfd is a bit different in that it could need to
deal with a lazy init issue with anon vma, iirc.  Some of the fine
details have prevented common helpers so far.

> 
> > You also mention the shrinker and using the alloc->mutex, well zapping
> > pages can also be done under the vma specific lock - if that helps?
> > 
> > Freeing page tables is different though, that needs more locking, but I
> > don't think this is an issue for you.
> > 
> > [1]. https://lore.kernel.org/all/20240215182756.3448972-1-lokeshgidra@google.com/
> 
> ha, I was not aware of this. I'll have a look thanks.

Happy to help.
Liam

Re: [PATCH v3 2/8] binder: concurrent page installation

[David Hildenbrand]

On 08.11.24 20:10, Carlos Llamas wrote:
> Allow multiple callers to install pages simultaneously by downgrading
> the mmap_sem to non-exclusive mode. Races to the same PTE are handled
> using get_user_pages_remote() to retrieve the already installed page.
> This method significantly reduces contention in the mmap semaphore.
> 
> To ensure safety, vma_lookup() is used (instead of alloc->vma) to avoid
> operating on an isolated VMA. In addition, zap_page_range_single() is
> called under the alloc->mutex to avoid racing with the shrinker.
> 
> Many thanks to Barry Song who posted a similar approach [1].
> 
> Link: https://lore.kernel.org/all/20240902225009.34576-1-21cnbao@gmail.com/ [1]
> Cc: David Hildenbrand <david@redhat.com>
> Cc: Barry Song <v-songbaohua@oppo.com>
> Cc: Suren Baghdasaryan <surenb@google.com>
> Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
> Reviewed-by: Suren Baghdasaryan <surenb@google.com>
> Signed-off-by: Carlos Llamas <cmllamas@google.com>
> ---
>   drivers/android/binder_alloc.c | 64 +++++++++++++++++++++-------------
>   1 file changed, 40 insertions(+), 24 deletions(-)
> 
> diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
> index 7241bf4a3ff2..2ab520c285b3 100644
> --- a/drivers/android/binder_alloc.c
> +++ b/drivers/android/binder_alloc.c
> @@ -221,26 +221,14 @@ static int binder_install_single_page(struct binder_alloc *alloc,
>   				      struct binder_lru_page *lru_page,
>   				      unsigned long addr)
>   {
> +	struct vm_area_struct *vma;
>   	struct page *page;
> -	int ret = 0;
> +	long npages;
> +	int ret;
>   
>   	if (!mmget_not_zero(alloc->mm))
>   		return -ESRCH;
>   
> -	/*
> -	 * Protected with mmap_sem in write mode as multiple tasks
> -	 * might race to install the same page.
> -	 */
> -	mmap_write_lock(alloc->mm);
> -	if (binder_get_installed_page(lru_page))
> -		goto out;
> -
> -	if (!alloc->vma) {
> -		pr_err("%d: %s failed, no vma\n", alloc->pid, __func__);
> -		ret = -ESRCH;
> -		goto out;
> -	}
> -
>   	page = alloc_page(GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
>   	if (!page) {
>   		pr_err("%d: failed to allocate page\n", alloc->pid);
> @@ -248,19 +236,47 @@ static int binder_install_single_page(struct binder_alloc *alloc,
>   		goto out;
>   	}
>   
> -	ret = vm_insert_page(alloc->vma, addr, page);
> -	if (ret) {
> +	mmap_read_lock(alloc->mm);
> +	vma = vma_lookup(alloc->mm, addr);
> +	if (!vma || vma != alloc->vma) {
> +		__free_page(page);
> +		pr_err("%d: %s failed, no vma\n", alloc->pid, __func__);
> +		ret = -ESRCH;
> +		goto unlock;
> +	}
> +
> +	ret = vm_insert_page(vma, addr, page);
> +	switch (ret) {
> +	case -EBUSY:
> +		/*
> +		 * EBUSY is ok. Someone installed the pte first but the
> +		 * lru_page->page_ptr has not been updated yet. Discard
> +		 * our page and look up the one already installed.
> +		 */
> +		ret = 0;
> +		__free_page(page);
> +		npages = get_user_pages_remote(alloc->mm, addr, 1, 0, &page, NULL);

This will trigger a page fault if we don't find what we expect (are 
races with e.g., MADV_DONTNEED possible?), is that really desired or not 
a problem?

-- 
Cheers,

David / dhildenb

Re: [PATCH v3 2/8] binder: concurrent page installation

[Carlos Llamas]

On Tue, Nov 12, 2024 at 12:10:20PM +0100, David Hildenbrand wrote:
> On 08.11.24 20:10, Carlos Llamas wrote:
> > +	ret = vm_insert_page(vma, addr, page);
> > +	switch (ret) {
> > +	case -EBUSY:
> > +		/*
> > +		 * EBUSY is ok. Someone installed the pte first but the
> > +		 * lru_page->page_ptr has not been updated yet. Discard
> > +		 * our page and look up the one already installed.
> > +		 */
> > +		ret = 0;
> > +		__free_page(page);
> > +		npages = get_user_pages_remote(alloc->mm, addr, 1, 0, &page, NULL);
> 
> This will trigger a page fault if we don't find what we expect (are races
> with e.g., MADV_DONTNEED possible?), is that really desired or not a
> problem?

This is fine. As of now, binder blocks its page faults:

	static vm_fault_t binder_vm_fault(struct vm_fault *vmf)
	{
		return VM_FAULT_SIGBUS;
	}

If we race with something like MADV_DONTNEED then we would just
propagate the -EFAULT error. I could add FOLL_NOFAULT to the gup remote
call to make it evident we don't care though.

--
Carlos Llamas

Re: [PATCH v3 2/8] binder: concurrent page installation

[David Hildenbrand]

On 12.11.24 15:43, Carlos Llamas wrote:
> On Tue, Nov 12, 2024 at 12:10:20PM +0100, David Hildenbrand wrote:
>> On 08.11.24 20:10, Carlos Llamas wrote:
>>> +	ret = vm_insert_page(vma, addr, page);
>>> +	switch (ret) {
>>> +	case -EBUSY:
>>> +		/*
>>> +		 * EBUSY is ok. Someone installed the pte first but the
>>> +		 * lru_page->page_ptr has not been updated yet. Discard
>>> +		 * our page and look up the one already installed.
>>> +		 */
>>> +		ret = 0;
>>> +		__free_page(page);
>>> +		npages = get_user_pages_remote(alloc->mm, addr, 1, 0, &page, NULL);
>>
>> This will trigger a page fault if we don't find what we expect (are races
>> with e.g., MADV_DONTNEED possible?), is that really desired or not a
>> problem?
> 
> This is fine. As of now, binder blocks its page faults:
> 
> 	static vm_fault_t binder_vm_fault(struct vm_fault *vmf)
> 	{
> 		return VM_FAULT_SIGBUS;
> 	}
> 
> If we race with something like MADV_DONTNEED then we would just
> propagate the -EFAULT error. I could add FOLL_NOFAULT to the gup remote
> call to make it evident we don't care though.

Might make it clearer ... or at least adding a comment how this is 
supposed to work. :)

-- 
Cheers,

David / dhildenb

Re: [PATCH v3 2/8] binder: concurrent page installation

[Carlos Llamas]

On Tue, Nov 12, 2024 at 03:53:51PM +0100, David Hildenbrand wrote:
> On 12.11.24 15:43, Carlos Llamas wrote:
> > On Tue, Nov 12, 2024 at 12:10:20PM +0100, David Hildenbrand wrote:
> > > On 08.11.24 20:10, Carlos Llamas wrote:
> > > > +	ret = vm_insert_page(vma, addr, page);
> > > > +	switch (ret) {
> > > > +	case -EBUSY:
> > > > +		/*
> > > > +		 * EBUSY is ok. Someone installed the pte first but the
> > > > +		 * lru_page->page_ptr has not been updated yet. Discard
> > > > +		 * our page and look up the one already installed.
> > > > +		 */
> > > > +		ret = 0;
> > > > +		__free_page(page);
> > > > +		npages = get_user_pages_remote(alloc->mm, addr, 1, 0, &page, NULL);
> > > 
> > > This will trigger a page fault if we don't find what we expect (are races
> > > with e.g., MADV_DONTNEED possible?), is that really desired or not a
> > > problem?
> > 
> > This is fine. As of now, binder blocks its page faults:
> > 
> > 	static vm_fault_t binder_vm_fault(struct vm_fault *vmf)
> > 	{
> > 		return VM_FAULT_SIGBUS;
> > 	}
> > 
> > If we race with something like MADV_DONTNEED then we would just
> > propagate the -EFAULT error. I could add FOLL_NOFAULT to the gup remote
> > call to make it evident we don't care though.
> 
> Might make it clearer ... or at least adding a comment how this is supposed
> to work. :)

Sounds good. I'll add both FOLL_NOFAULT flag and a comment for v4.

Cheers,
Carlos Llamas

[PATCH v3 3/8] binder: select correct nid for pages in LRU

[Carlos Llamas]

The numa node id for binder pages is currently being derived from the
lru entry under struct binder_lru_page. However, this object doesn't
reflect the node id of the struct page items allocated separately.

Instead, select the correct node id from the page itself. This was made
possible since commit 0a97c01cd20b ("list_lru: allow explicit memcg and
NUMA node selection").

Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Carlos Llamas <cmllamas@google.com>
---
 drivers/android/binder_alloc.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
index 2ab520c285b3..08daf0b75b12 100644
--- a/drivers/android/binder_alloc.c
+++ b/drivers/android/binder_alloc.c
@@ -210,7 +210,10 @@ static void binder_lru_freelist_add(struct binder_alloc *alloc,
 
 		trace_binder_free_lru_start(alloc, index);
 
-		ret = list_lru_add_obj(&binder_freelist, &page->lru);
+		ret = list_lru_add(&binder_freelist,
+				   &page->lru,
+				   page_to_nid(page->page_ptr),
+				   NULL);
 		WARN_ON(!ret);
 
 		trace_binder_free_lru_end(alloc, index);
@@ -333,7 +336,10 @@ static void binder_lru_freelist_del(struct binder_alloc *alloc,
 		if (page->page_ptr) {
 			trace_binder_alloc_lru_start(alloc, index);
 
-			on_lru = list_lru_del_obj(&binder_freelist, &page->lru);
+			on_lru = list_lru_del(&binder_freelist,
+					      &page->lru,
+					      page_to_nid(page->page_ptr),
+					      NULL);
 			WARN_ON(!on_lru);
 
 			trace_binder_alloc_lru_end(alloc, index);
@@ -946,8 +952,10 @@ void binder_alloc_deferred_release(struct binder_alloc *alloc)
 			if (!alloc->pages[i].page_ptr)
 				continue;
 
-			on_lru = list_lru_del_obj(&binder_freelist,
-						  &alloc->pages[i].lru);
+			on_lru = list_lru_del(&binder_freelist,
+					      &alloc->pages[i].lru,
+					      page_to_nid(alloc->pages[i].page_ptr),
+					      NULL);
 			binder_alloc_debug(BINDER_DEBUG_BUFFER_ALLOC,
 				     "%s: %d: page %d %s\n",
 				     __func__, alloc->pid, i,
-- 
2.47.0.277.g8800431eea-goog

[PATCH v3 4/8] binder: remove struct binder_lru_page

[Carlos Llamas]

Remove the redundant struct binder_lru_page concept. Instead, let's use
available struct page->lru and page->private members directly to achieve
the same functionality.

This reduces the maximum memory allocated for alloc->pages from 32768
down to 8192 bytes (aarch64). Savings are per binder instance.

Cc: Matthew Wilcox <willy@infradead.org>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Carlos Llamas <cmllamas@google.com>
---
 drivers/android/binder_alloc.c          | 104 ++++++++++++------------
 drivers/android/binder_alloc.h          |  16 +---
 drivers/android/binder_alloc_selftest.c |  14 ++--
 3 files changed, 63 insertions(+), 71 deletions(-)

diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
index 08daf0b75b12..457fa937aa8f 100644
--- a/drivers/android/binder_alloc.c
+++ b/drivers/android/binder_alloc.c
@@ -176,25 +176,26 @@ struct binder_buffer *binder_alloc_prepare_to_free(struct binder_alloc *alloc,
 }
 
 static inline void
-binder_set_installed_page(struct binder_lru_page *lru_page,
+binder_set_installed_page(struct binder_alloc *alloc,
+			  unsigned long index,
 			  struct page *page)
 {
 	/* Pairs with acquire in binder_get_installed_page() */
-	smp_store_release(&lru_page->page_ptr, page);
+	smp_store_release(&alloc->pages[index], page);
 }
 
 static inline struct page *
-binder_get_installed_page(struct binder_lru_page *lru_page)
+binder_get_installed_page(struct binder_alloc *alloc, unsigned long index)
 {
 	/* Pairs with release in binder_set_installed_page() */
-	return smp_load_acquire(&lru_page->page_ptr);
+	return smp_load_acquire(&alloc->pages[index]);
 }
 
 static void binder_lru_freelist_add(struct binder_alloc *alloc,
 				    unsigned long start, unsigned long end)
 {
-	struct binder_lru_page *page;
 	unsigned long page_addr;
+	struct page *page;
 
 	trace_binder_update_page_range(alloc, false, start, end);
 
@@ -203,16 +204,15 @@ static void binder_lru_freelist_add(struct binder_alloc *alloc,
 		int ret;
 
 		index = (page_addr - alloc->buffer) / PAGE_SIZE;
-		page = &alloc->pages[index];
-
-		if (!binder_get_installed_page(page))
+		page = binder_get_installed_page(alloc, index);
+		if (!page)
 			continue;
 
 		trace_binder_free_lru_start(alloc, index);
 
 		ret = list_lru_add(&binder_freelist,
 				   &page->lru,
-				   page_to_nid(page->page_ptr),
+				   page_to_nid(page),
 				   NULL);
 		WARN_ON(!ret);
 
@@ -220,8 +220,25 @@ static void binder_lru_freelist_add(struct binder_alloc *alloc,
 	}
 }
 
+static struct page *binder_page_alloc(struct binder_alloc *alloc,
+				      unsigned long index,
+				      unsigned long addr)
+{
+	struct page *page;
+
+	page = alloc_page(GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
+	if (!page)
+		return NULL;
+
+	page->private = (unsigned long)alloc;
+	INIT_LIST_HEAD(&page->lru);
+	page->index = index;
+
+	return page;
+}
+
 static int binder_install_single_page(struct binder_alloc *alloc,
-				      struct binder_lru_page *lru_page,
+				      unsigned long index,
 				      unsigned long addr)
 {
 	struct vm_area_struct *vma;
@@ -232,9 +249,8 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 	if (!mmget_not_zero(alloc->mm))
 		return -ESRCH;
 
-	page = alloc_page(GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
+	page = binder_page_alloc(alloc, index, addr);
 	if (!page) {
-		pr_err("%d: failed to allocate page\n", alloc->pid);
 		ret = -ENOMEM;
 		goto out;
 	}
@@ -253,7 +269,7 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 	case -EBUSY:
 		/*
 		 * EBUSY is ok. Someone installed the pte first but the
-		 * lru_page->page_ptr has not been updated yet. Discard
+		 * alloc->pages[index] has not been updated yet. Discard
 		 * our page and look up the one already installed.
 		 */
 		ret = 0;
@@ -268,7 +284,7 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 		fallthrough;
 	case 0:
 		/* Mark page installation complete and safe to use */
-		binder_set_installed_page(lru_page, page);
+		binder_set_installed_page(alloc, index, page);
 		break;
 	default:
 		__free_page(page);
@@ -288,7 +304,6 @@ static int binder_install_buffer_pages(struct binder_alloc *alloc,
 				       struct binder_buffer *buffer,
 				       size_t size)
 {
-	struct binder_lru_page *page;
 	unsigned long start, final;
 	unsigned long page_addr;
 
@@ -300,14 +315,12 @@ static int binder_install_buffer_pages(struct binder_alloc *alloc,
 		int ret;
 
 		index = (page_addr - alloc->buffer) / PAGE_SIZE;
-		page = &alloc->pages[index];
-
-		if (binder_get_installed_page(page))
+		if (binder_get_installed_page(alloc, index))
 			continue;
 
 		trace_binder_alloc_page_start(alloc, index);
 
-		ret = binder_install_single_page(alloc, page, page_addr);
+		ret = binder_install_single_page(alloc, index, page_addr);
 		if (ret)
 			return ret;
 
@@ -321,8 +334,8 @@ static int binder_install_buffer_pages(struct binder_alloc *alloc,
 static void binder_lru_freelist_del(struct binder_alloc *alloc,
 				    unsigned long start, unsigned long end)
 {
-	struct binder_lru_page *page;
 	unsigned long page_addr;
+	struct page *page;
 
 	trace_binder_update_page_range(alloc, true, start, end);
 
@@ -331,14 +344,14 @@ static void binder_lru_freelist_del(struct binder_alloc *alloc,
 		bool on_lru;
 
 		index = (page_addr - alloc->buffer) / PAGE_SIZE;
-		page = &alloc->pages[index];
+		page = binder_get_installed_page(alloc, index);
 
-		if (page->page_ptr) {
+		if (page) {
 			trace_binder_alloc_lru_start(alloc, index);
 
 			on_lru = list_lru_del(&binder_freelist,
 					      &page->lru,
-					      page_to_nid(page->page_ptr),
+					      page_to_nid(page),
 					      NULL);
 			WARN_ON(!on_lru);
 
@@ -759,11 +772,10 @@ static struct page *binder_alloc_get_page(struct binder_alloc *alloc,
 		(buffer->user_data - alloc->buffer);
 	pgoff_t pgoff = buffer_space_offset & ~PAGE_MASK;
 	size_t index = buffer_space_offset >> PAGE_SHIFT;
-	struct binder_lru_page *lru_page;
 
-	lru_page = &alloc->pages[index];
 	*pgoffp = pgoff;
-	return lru_page->page_ptr;
+
+	return binder_get_installed_page(alloc, index);
 }
 
 /**
@@ -838,7 +850,7 @@ int binder_alloc_mmap_handler(struct binder_alloc *alloc,
 {
 	struct binder_buffer *buffer;
 	const char *failure_string;
-	int ret, i;
+	int ret;
 
 	if (unlikely(vma->vm_mm != alloc->mm)) {
 		ret = -EINVAL;
@@ -861,17 +873,12 @@ int binder_alloc_mmap_handler(struct binder_alloc *alloc,
 	alloc->pages = kvcalloc(alloc->buffer_size / PAGE_SIZE,
 				sizeof(alloc->pages[0]),
 				GFP_KERNEL);
-	if (alloc->pages == NULL) {
+	if (!alloc->pages) {
 		ret = -ENOMEM;
 		failure_string = "alloc page array";
 		goto err_alloc_pages_failed;
 	}
 
-	for (i = 0; i < alloc->buffer_size / PAGE_SIZE; i++) {
-		alloc->pages[i].alloc = alloc;
-		INIT_LIST_HEAD(&alloc->pages[i].lru);
-	}
-
 	buffer = kzalloc(sizeof(*buffer), GFP_KERNEL);
 	if (!buffer) {
 		ret = -ENOMEM;
@@ -947,20 +954,22 @@ void binder_alloc_deferred_release(struct binder_alloc *alloc)
 		int i;
 
 		for (i = 0; i < alloc->buffer_size / PAGE_SIZE; i++) {
+			struct page *page;
 			bool on_lru;
 
-			if (!alloc->pages[i].page_ptr)
+			page = binder_get_installed_page(alloc, i);
+			if (!page)
 				continue;
 
 			on_lru = list_lru_del(&binder_freelist,
-					      &alloc->pages[i].lru,
-					      page_to_nid(alloc->pages[i].page_ptr),
+					      &page->lru,
+					      page_to_nid(page),
 					      NULL);
 			binder_alloc_debug(BINDER_DEBUG_BUFFER_ALLOC,
 				     "%s: %d: page %d %s\n",
 				     __func__, alloc->pid, i,
 				     on_lru ? "on lru" : "active");
-			__free_page(alloc->pages[i].page_ptr);
+			__free_page(page);
 			page_count++;
 		}
 	}
@@ -1009,7 +1018,7 @@ void binder_alloc_print_allocated(struct seq_file *m,
 void binder_alloc_print_pages(struct seq_file *m,
 			      struct binder_alloc *alloc)
 {
-	struct binder_lru_page *page;
+	struct page *page;
 	int i;
 	int active = 0;
 	int lru = 0;
@@ -1022,8 +1031,8 @@ void binder_alloc_print_pages(struct seq_file *m,
 	 */
 	if (binder_alloc_get_vma(alloc) != NULL) {
 		for (i = 0; i < alloc->buffer_size / PAGE_SIZE; i++) {
-			page = &alloc->pages[i];
-			if (!page->page_ptr)
+			page = binder_get_installed_page(alloc, i);
+			if (!page)
 				free++;
 			else if (list_empty(&page->lru))
 				active++;
@@ -1083,11 +1092,10 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 				       void *cb_arg)
 	__must_hold(lock)
 {
-	struct binder_lru_page *page = container_of(item, typeof(*page), lru);
-	struct binder_alloc *alloc = page->alloc;
+	struct page *page = container_of(item, typeof(*page), lru);
+	struct binder_alloc *alloc = (struct binder_alloc *)page->private;
 	struct mm_struct *mm = alloc->mm;
 	struct vm_area_struct *vma;
-	struct page *page_to_free;
 	unsigned long page_addr;
 	size_t index;
 
@@ -1097,10 +1105,8 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 		goto err_mmap_read_lock_failed;
 	if (!mutex_trylock(&alloc->mutex))
 		goto err_get_alloc_mutex_failed;
-	if (!page->page_ptr)
-		goto err_page_already_freed;
 
-	index = page - alloc->pages;
+	index = page->index;
 	page_addr = alloc->buffer + index * PAGE_SIZE;
 
 	vma = vma_lookup(mm, page_addr);
@@ -1109,8 +1115,7 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 
 	trace_binder_unmap_kernel_start(alloc, index);
 
-	page_to_free = page->page_ptr;
-	page->page_ptr = NULL;
+	binder_set_installed_page(alloc, index, NULL);
 
 	trace_binder_unmap_kernel_end(alloc, index);
 
@@ -1128,13 +1133,12 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 	mutex_unlock(&alloc->mutex);
 	mmap_read_unlock(mm);
 	mmput_async(mm);
-	__free_page(page_to_free);
+	__free_page(page);
 
 	spin_lock(lock);
 	return LRU_REMOVED_RETRY;
 
 err_invalid_vma:
-err_page_already_freed:
 	mutex_unlock(&alloc->mutex);
 err_get_alloc_mutex_failed:
 	mmap_read_unlock(mm);
diff --git a/drivers/android/binder_alloc.h b/drivers/android/binder_alloc.h
index a5181916942e..5c2473e95494 100644
--- a/drivers/android/binder_alloc.h
+++ b/drivers/android/binder_alloc.h
@@ -58,18 +58,6 @@ struct binder_buffer {
 	int pid;
 };
 
-/**
- * struct binder_lru_page - page object used for binder shrinker
- * @page_ptr: pointer to physical page in mmap'd space
- * @lru:      entry in binder_freelist
- * @alloc:    binder_alloc for a proc
- */
-struct binder_lru_page {
-	struct list_head lru;
-	struct page *page_ptr;
-	struct binder_alloc *alloc;
-};
-
 /**
  * struct binder_alloc - per-binder proc state for binder allocator
  * @mutex:              protects binder_alloc fields
@@ -83,7 +71,7 @@ struct binder_lru_page {
  * @allocated_buffers:  rb tree of allocated buffers sorted by address
  * @free_async_space:   VA space available for async buffers. This is
  *                      initialized at mmap time to 1/2 the full VA space
- * @pages:              array of binder_lru_page
+ * @pages:              array of struct page *
  * @buffer_size:        size of address space specified via mmap
  * @pid:                pid for associated binder_proc (invariant after init)
  * @pages_high:         high watermark of offset in @pages
@@ -104,7 +92,7 @@ struct binder_alloc {
 	struct rb_root free_buffers;
 	struct rb_root allocated_buffers;
 	size_t free_async_space;
-	struct binder_lru_page *pages;
+	struct page **pages;
 	size_t buffer_size;
 	int pid;
 	size_t pages_high;
diff --git a/drivers/android/binder_alloc_selftest.c b/drivers/android/binder_alloc_selftest.c
index 81442fe20a69..c6941b9abad9 100644
--- a/drivers/android/binder_alloc_selftest.c
+++ b/drivers/android/binder_alloc_selftest.c
@@ -105,10 +105,10 @@ static bool check_buffer_pages_allocated(struct binder_alloc *alloc,
 	page_addr = buffer->user_data;
 	for (; page_addr < end; page_addr += PAGE_SIZE) {
 		page_index = (page_addr - alloc->buffer) / PAGE_SIZE;
-		if (!alloc->pages[page_index].page_ptr ||
-		    !list_empty(&alloc->pages[page_index].lru)) {
+		if (!alloc->pages[page_index] ||
+		    !list_empty(&alloc->pages[page_index]->lru)) {
 			pr_err("expect alloc but is %s at page index %d\n",
-			       alloc->pages[page_index].page_ptr ?
+			       alloc->pages[page_index] ?
 			       "lru" : "free", page_index);
 			return false;
 		}
@@ -148,10 +148,10 @@ static void binder_selftest_free_buf(struct binder_alloc *alloc,
 		 * if binder shrinker ran during binder_alloc_free_buf
 		 * calls above.
 		 */
-		if (list_empty(&alloc->pages[i].lru)) {
+		if (list_empty(&alloc->pages[i]->lru)) {
 			pr_err_size_seq(sizes, seq);
 			pr_err("expect lru but is %s at page index %d\n",
-			       alloc->pages[i].page_ptr ? "alloc" : "free", i);
+			       alloc->pages[i] ? "alloc" : "free", i);
 			binder_selftest_failures++;
 		}
 	}
@@ -168,9 +168,9 @@ static void binder_selftest_free_page(struct binder_alloc *alloc)
 	}
 
 	for (i = 0; i < (alloc->buffer_size / PAGE_SIZE); i++) {
-		if (alloc->pages[i].page_ptr) {
+		if (alloc->pages[i]) {
 			pr_err("expect free but is %s at page index %d\n",
-			       list_empty(&alloc->pages[i].lru) ?
+			       list_empty(&alloc->pages[i]->lru) ?
 			       "alloc" : "lru", i);
 			binder_selftest_failures++;
 		}
-- 
2.47.0.277.g8800431eea-goog

[PATCH v3 5/8] binder: replace alloc->vma with alloc->mapped

[Carlos Llamas]

It is unsafe to use alloc->vma outside of the mmap_sem. Instead, add a
new boolean alloc->mapped to save the vma state (mapped or unmmaped) and
use this as a replacement for alloc->vma to validate several paths.

Using the alloc->vma caused several performance and security issues in
the past. Now that it has been replaced with either vm_lookup() or the
alloc->mapped state, we can finally remove it.

Cc: Minchan Kim <minchan@kernel.org>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Carlos Llamas <cmllamas@google.com>
---
 drivers/android/binder_alloc.c          | 43 ++++++++++++-------------
 drivers/android/binder_alloc.h          |  6 ++--
 drivers/android/binder_alloc_selftest.c |  2 +-
 3 files changed, 25 insertions(+), 26 deletions(-)

diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
index 457fa937aa8f..a6697b5a4b2f 100644
--- a/drivers/android/binder_alloc.c
+++ b/drivers/android/binder_alloc.c
@@ -220,6 +220,19 @@ static void binder_lru_freelist_add(struct binder_alloc *alloc,
 	}
 }
 
+static inline
+void binder_alloc_set_mapped(struct binder_alloc *alloc, bool state)
+{
+	/* pairs with smp_load_acquire in binder_alloc_is_mapped() */
+	smp_store_release(&alloc->mapped, state);
+}
+
+static inline bool binder_alloc_is_mapped(struct binder_alloc *alloc)
+{
+	/* pairs with smp_store_release in binder_alloc_set_mapped() */
+	return smp_load_acquire(&alloc->mapped);
+}
+
 static struct page *binder_page_alloc(struct binder_alloc *alloc,
 				      unsigned long index,
 				      unsigned long addr)
@@ -257,7 +270,7 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 
 	mmap_read_lock(alloc->mm);
 	vma = vma_lookup(alloc->mm, addr);
-	if (!vma || vma != alloc->vma) {
+	if (!vma || !binder_alloc_is_mapped(alloc)) {
 		__free_page(page);
 		pr_err("%d: %s failed, no vma\n", alloc->pid, __func__);
 		ret = -ESRCH;
@@ -364,20 +377,6 @@ static void binder_lru_freelist_del(struct binder_alloc *alloc,
 	}
 }
 
-static inline void binder_alloc_set_vma(struct binder_alloc *alloc,
-		struct vm_area_struct *vma)
-{
-	/* pairs with smp_load_acquire in binder_alloc_get_vma() */
-	smp_store_release(&alloc->vma, vma);
-}
-
-static inline struct vm_area_struct *binder_alloc_get_vma(
-		struct binder_alloc *alloc)
-{
-	/* pairs with smp_store_release in binder_alloc_set_vma() */
-	return smp_load_acquire(&alloc->vma);
-}
-
 static void debug_no_space_locked(struct binder_alloc *alloc)
 {
 	size_t largest_alloc_size = 0;
@@ -611,7 +610,7 @@ struct binder_buffer *binder_alloc_new_buf(struct binder_alloc *alloc,
 	int ret;
 
 	/* Check binder_alloc is fully initialized */
-	if (!binder_alloc_get_vma(alloc)) {
+	if (!binder_alloc_is_mapped(alloc)) {
 		binder_alloc_debug(BINDER_DEBUG_USER_ERROR,
 				   "%d: binder_alloc_buf, no vma\n",
 				   alloc->pid);
@@ -893,7 +892,7 @@ int binder_alloc_mmap_handler(struct binder_alloc *alloc,
 	alloc->free_async_space = alloc->buffer_size / 2;
 
 	/* Signal binder_alloc is fully initialized */
-	binder_alloc_set_vma(alloc, vma);
+	binder_alloc_set_mapped(alloc, true);
 
 	return 0;
 
@@ -923,7 +922,7 @@ void binder_alloc_deferred_release(struct binder_alloc *alloc)
 
 	buffers = 0;
 	mutex_lock(&alloc->mutex);
-	BUG_ON(alloc->vma);
+	BUG_ON(alloc->mapped);
 
 	while ((n = rb_first(&alloc->allocated_buffers))) {
 		buffer = rb_entry(n, struct binder_buffer, rb_node);
@@ -1029,7 +1028,7 @@ void binder_alloc_print_pages(struct seq_file *m,
 	 * Make sure the binder_alloc is fully initialized, otherwise we might
 	 * read inconsistent state.
 	 */
-	if (binder_alloc_get_vma(alloc) != NULL) {
+	if (binder_alloc_is_mapped(alloc)) {
 		for (i = 0; i < alloc->buffer_size / PAGE_SIZE; i++) {
 			page = binder_get_installed_page(alloc, i);
 			if (!page)
@@ -1069,12 +1068,12 @@ int binder_alloc_get_allocated_count(struct binder_alloc *alloc)
  * @alloc: binder_alloc for this proc
  *
  * Called from binder_vma_close() when releasing address space.
- * Clears alloc->vma to prevent new incoming transactions from
+ * Clears alloc->mapped to prevent new incoming transactions from
  * allocating more buffers.
  */
 void binder_alloc_vma_close(struct binder_alloc *alloc)
 {
-	binder_alloc_set_vma(alloc, NULL);
+	binder_alloc_set_mapped(alloc, false);
 }
 
 /**
@@ -1110,7 +1109,7 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 	page_addr = alloc->buffer + index * PAGE_SIZE;
 
 	vma = vma_lookup(mm, page_addr);
-	if (vma && vma != binder_alloc_get_vma(alloc))
+	if (vma && !binder_alloc_is_mapped(alloc))
 		goto err_invalid_vma;
 
 	trace_binder_unmap_kernel_start(alloc, index);
diff --git a/drivers/android/binder_alloc.h b/drivers/android/binder_alloc.h
index 5c2473e95494..a3c043cb8343 100644
--- a/drivers/android/binder_alloc.h
+++ b/drivers/android/binder_alloc.h
@@ -61,8 +61,6 @@ struct binder_buffer {
 /**
  * struct binder_alloc - per-binder proc state for binder allocator
  * @mutex:              protects binder_alloc fields
- * @vma:                vm_area_struct passed to mmap_handler
- *                      (invariant after mmap)
  * @mm:                 copy of task->mm (invariant after open)
  * @buffer:             base of per-proc address space mapped via mmap
  * @buffers:            list of all buffers for this proc
@@ -75,6 +73,8 @@ struct binder_buffer {
  * @buffer_size:        size of address space specified via mmap
  * @pid:                pid for associated binder_proc (invariant after init)
  * @pages_high:         high watermark of offset in @pages
+ * @mapped:             whether the vm area is mapped, each binder instance is
+ *                      allowed a single mapping throughout its lifetime
  * @oneway_spam_detected: %true if oneway spam detection fired, clear that
  * flag once the async buffer has returned to a healthy state
  *
@@ -85,7 +85,6 @@ struct binder_buffer {
  */
 struct binder_alloc {
 	struct mutex mutex;
-	struct vm_area_struct *vma;
 	struct mm_struct *mm;
 	unsigned long buffer;
 	struct list_head buffers;
@@ -96,6 +95,7 @@ struct binder_alloc {
 	size_t buffer_size;
 	int pid;
 	size_t pages_high;
+	bool mapped;
 	bool oneway_spam_detected;
 };
 
diff --git a/drivers/android/binder_alloc_selftest.c b/drivers/android/binder_alloc_selftest.c
index c6941b9abad9..2dda82d0d5e8 100644
--- a/drivers/android/binder_alloc_selftest.c
+++ b/drivers/android/binder_alloc_selftest.c
@@ -291,7 +291,7 @@ void binder_selftest_alloc(struct binder_alloc *alloc)
 	if (!binder_selftest_run)
 		return;
 	mutex_lock(&binder_selftest_lock);
-	if (!binder_selftest_run || !alloc->vma)
+	if (!binder_selftest_run || !alloc->mapped)
 		goto done;
 	pr_info("STARTED\n");
 	binder_selftest_alloc_offset(alloc, end_offset, 0);
-- 
2.47.0.277.g8800431eea-goog

[PATCH v3 6/8] binder: rename alloc->buffer to vm_start

[Carlos Llamas]

The alloc->buffer field in struct binder_alloc stores the starting
address of the mapped vma, rename this field to alloc->vm_start to
better reflect its purpose. It also avoids confusion with the binder
buffer concept, e.g. transaction->buffer.

No functional changes in this patch.

Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Carlos Llamas <cmllamas@google.com>
---
 drivers/android/binder.c                |  2 +-
 drivers/android/binder_alloc.c          | 28 ++++++++++++-------------
 drivers/android/binder_alloc.h          |  4 ++--
 drivers/android/binder_alloc_selftest.c |  2 +-
 drivers/android/binder_trace.h          |  2 +-
 5 files changed, 19 insertions(+), 19 deletions(-)

diff --git a/drivers/android/binder.c b/drivers/android/binder.c
index 978740537a1a..57265cabec43 100644
--- a/drivers/android/binder.c
+++ b/drivers/android/binder.c
@@ -6350,7 +6350,7 @@ static void print_binder_transaction_ilocked(struct seq_file *m,
 		seq_printf(m, " node %d", buffer->target_node->debug_id);
 	seq_printf(m, " size %zd:%zd offset %lx\n",
 		   buffer->data_size, buffer->offsets_size,
-		   proc->alloc.buffer - buffer->user_data);
+		   proc->alloc.vm_start - buffer->user_data);
 }
 
 static void print_binder_work_ilocked(struct seq_file *m,
diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
index a6697b5a4b2f..3716ffd00baf 100644
--- a/drivers/android/binder_alloc.c
+++ b/drivers/android/binder_alloc.c
@@ -61,7 +61,7 @@ static size_t binder_alloc_buffer_size(struct binder_alloc *alloc,
 				       struct binder_buffer *buffer)
 {
 	if (list_is_last(&buffer->entry, &alloc->buffers))
-		return alloc->buffer + alloc->buffer_size - buffer->user_data;
+		return alloc->vm_start + alloc->buffer_size - buffer->user_data;
 	return binder_buffer_next(buffer)->user_data - buffer->user_data;
 }
 
@@ -203,7 +203,7 @@ static void binder_lru_freelist_add(struct binder_alloc *alloc,
 		size_t index;
 		int ret;
 
-		index = (page_addr - alloc->buffer) / PAGE_SIZE;
+		index = (page_addr - alloc->vm_start) / PAGE_SIZE;
 		page = binder_get_installed_page(alloc, index);
 		if (!page)
 			continue;
@@ -290,7 +290,7 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 		npages = get_user_pages_remote(alloc->mm, addr, 1, 0, &page, NULL);
 		if (npages <= 0) {
 			pr_err("%d: failed to find page at offset %lx\n",
-			       alloc->pid, addr - alloc->buffer);
+			       alloc->pid, addr - alloc->vm_start);
 			ret = -ESRCH;
 			break;
 		}
@@ -302,7 +302,7 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 	default:
 		__free_page(page);
 		pr_err("%d: %s failed to insert page at offset %lx with %d\n",
-		       alloc->pid, __func__, addr - alloc->buffer, ret);
+		       alloc->pid, __func__, addr - alloc->vm_start, ret);
 		ret = -ENOMEM;
 		break;
 	}
@@ -327,7 +327,7 @@ static int binder_install_buffer_pages(struct binder_alloc *alloc,
 		unsigned long index;
 		int ret;
 
-		index = (page_addr - alloc->buffer) / PAGE_SIZE;
+		index = (page_addr - alloc->vm_start) / PAGE_SIZE;
 		if (binder_get_installed_page(alloc, index))
 			continue;
 
@@ -356,7 +356,7 @@ static void binder_lru_freelist_del(struct binder_alloc *alloc,
 		unsigned long index;
 		bool on_lru;
 
-		index = (page_addr - alloc->buffer) / PAGE_SIZE;
+		index = (page_addr - alloc->vm_start) / PAGE_SIZE;
 		page = binder_get_installed_page(alloc, index);
 
 		if (page) {
@@ -708,8 +708,8 @@ static void binder_free_buf_locked(struct binder_alloc *alloc,
 	BUG_ON(buffer->free);
 	BUG_ON(size > buffer_size);
 	BUG_ON(buffer->transaction != NULL);
-	BUG_ON(buffer->user_data < alloc->buffer);
-	BUG_ON(buffer->user_data > alloc->buffer + alloc->buffer_size);
+	BUG_ON(buffer->user_data < alloc->vm_start);
+	BUG_ON(buffer->user_data > alloc->vm_start + alloc->buffer_size);
 
 	if (buffer->async_transaction) {
 		alloc->free_async_space += buffer_size;
@@ -768,7 +768,7 @@ static struct page *binder_alloc_get_page(struct binder_alloc *alloc,
 					  pgoff_t *pgoffp)
 {
 	binder_size_t buffer_space_offset = buffer_offset +
-		(buffer->user_data - alloc->buffer);
+		(buffer->user_data - alloc->vm_start);
 	pgoff_t pgoff = buffer_space_offset & ~PAGE_MASK;
 	size_t index = buffer_space_offset >> PAGE_SHIFT;
 
@@ -867,7 +867,7 @@ int binder_alloc_mmap_handler(struct binder_alloc *alloc,
 				   SZ_4M);
 	mutex_unlock(&binder_alloc_mmap_lock);
 
-	alloc->buffer = vma->vm_start;
+	alloc->vm_start = vma->vm_start;
 
 	alloc->pages = kvcalloc(alloc->buffer_size / PAGE_SIZE,
 				sizeof(alloc->pages[0]),
@@ -885,7 +885,7 @@ int binder_alloc_mmap_handler(struct binder_alloc *alloc,
 		goto err_alloc_buf_struct_failed;
 	}
 
-	buffer->user_data = alloc->buffer;
+	buffer->user_data = alloc->vm_start;
 	list_add(&buffer->entry, &alloc->buffers);
 	buffer->free = 1;
 	binder_insert_free_buffer(alloc, buffer);
@@ -900,7 +900,7 @@ int binder_alloc_mmap_handler(struct binder_alloc *alloc,
 	kvfree(alloc->pages);
 	alloc->pages = NULL;
 err_alloc_pages_failed:
-	alloc->buffer = 0;
+	alloc->vm_start = 0;
 	mutex_lock(&binder_alloc_mmap_lock);
 	alloc->buffer_size = 0;
 err_already_mapped:
@@ -1001,7 +1001,7 @@ void binder_alloc_print_allocated(struct seq_file *m,
 		buffer = rb_entry(n, struct binder_buffer, rb_node);
 		seq_printf(m, "  buffer %d: %lx size %zd:%zd:%zd %s\n",
 			   buffer->debug_id,
-			   buffer->user_data - alloc->buffer,
+			   buffer->user_data - alloc->vm_start,
 			   buffer->data_size, buffer->offsets_size,
 			   buffer->extra_buffers_size,
 			   buffer->transaction ? "active" : "delivered");
@@ -1106,7 +1106,7 @@ enum lru_status binder_alloc_free_page(struct list_head *item,
 		goto err_get_alloc_mutex_failed;
 
 	index = page->index;
-	page_addr = alloc->buffer + index * PAGE_SIZE;
+	page_addr = alloc->vm_start + index * PAGE_SIZE;
 
 	vma = vma_lookup(mm, page_addr);
 	if (vma && !binder_alloc_is_mapped(alloc))
diff --git a/drivers/android/binder_alloc.h b/drivers/android/binder_alloc.h
index a3c043cb8343..088829ce6668 100644
--- a/drivers/android/binder_alloc.h
+++ b/drivers/android/binder_alloc.h
@@ -62,7 +62,7 @@ struct binder_buffer {
  * struct binder_alloc - per-binder proc state for binder allocator
  * @mutex:              protects binder_alloc fields
  * @mm:                 copy of task->mm (invariant after open)
- * @buffer:             base of per-proc address space mapped via mmap
+ * @vm_start:           base of per-proc address space mapped via mmap
  * @buffers:            list of all buffers for this proc
  * @free_buffers:       rb tree of buffers available for allocation
  *                      sorted by size
@@ -86,7 +86,7 @@ struct binder_buffer {
 struct binder_alloc {
 	struct mutex mutex;
 	struct mm_struct *mm;
-	unsigned long buffer;
+	unsigned long vm_start;
 	struct list_head buffers;
 	struct rb_root free_buffers;
 	struct rb_root allocated_buffers;
diff --git a/drivers/android/binder_alloc_selftest.c b/drivers/android/binder_alloc_selftest.c
index 2dda82d0d5e8..d2d086d2c037 100644
--- a/drivers/android/binder_alloc_selftest.c
+++ b/drivers/android/binder_alloc_selftest.c
@@ -104,7 +104,7 @@ static bool check_buffer_pages_allocated(struct binder_alloc *alloc,
 	end = PAGE_ALIGN(buffer->user_data + size);
 	page_addr = buffer->user_data;
 	for (; page_addr < end; page_addr += PAGE_SIZE) {
-		page_index = (page_addr - alloc->buffer) / PAGE_SIZE;
+		page_index = (page_addr - alloc->vm_start) / PAGE_SIZE;
 		if (!alloc->pages[page_index] ||
 		    !list_empty(&alloc->pages[page_index]->lru)) {
 			pr_err("expect alloc but is %s at page index %d\n",
diff --git a/drivers/android/binder_trace.h b/drivers/android/binder_trace.h
index fe38c6fc65d0..16de1b9e72f7 100644
--- a/drivers/android/binder_trace.h
+++ b/drivers/android/binder_trace.h
@@ -328,7 +328,7 @@ TRACE_EVENT(binder_update_page_range,
 	TP_fast_assign(
 		__entry->proc = alloc->pid;
 		__entry->allocate = allocate;
-		__entry->offset = start - alloc->buffer;
+		__entry->offset = start - alloc->vm_start;
 		__entry->size = end - start;
 	),
 	TP_printk("proc=%d allocate=%d offset=%zu size=%zu",
-- 
2.47.0.277.g8800431eea-goog

[PATCH v3 7/8] binder: use per-vma lock in page installation

[Carlos Llamas]

Use per-vma locking for concurrent page installations, this minimizes
contention with unrelated vmas improving performance. The mmap_lock is
still acquired when needed though, e.g. before get_user_pages_remote().

Many thanks to Barry Song who posted a similar approach [1].

Link: https://lore.kernel.org/all/20240902225009.34576-1-21cnbao@gmail.com/ [1]
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Carlos Llamas <cmllamas@google.com>
---
 drivers/android/binder_alloc.c | 61 +++++++++++++++++++++++++---------
 1 file changed, 45 insertions(+), 16 deletions(-)

diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
index 3716ffd00baf..7d2cad9beebb 100644
--- a/drivers/android/binder_alloc.c
+++ b/drivers/android/binder_alloc.c
@@ -233,6 +233,48 @@ static inline bool binder_alloc_is_mapped(struct binder_alloc *alloc)
 	return smp_load_acquire(&alloc->mapped);
 }
 
+static struct page *binder_page_lookup(struct binder_alloc *alloc,
+				       unsigned long addr)
+{
+	struct mm_struct *mm = alloc->mm;
+	struct page *page;
+	long ret = 0;
+
+	mmap_read_lock(mm);
+	if (binder_alloc_is_mapped(alloc))
+		ret = get_user_pages_remote(mm, addr, 1, 0, &page, NULL);
+	mmap_read_unlock(mm);
+
+	return ret > 0 ? page : NULL;
+}
+
+static int binder_page_insert(struct binder_alloc *alloc,
+			      unsigned long addr,
+			      struct page *page)
+{
+	struct mm_struct *mm = alloc->mm;
+	struct vm_area_struct *vma;
+	int ret = -ESRCH;
+
+	/* attempt per-vma lock first */
+	vma = lock_vma_under_rcu(mm, addr);
+	if (vma) {
+		if (binder_alloc_is_mapped(alloc))
+			ret = vm_insert_page(vma, addr, page);
+		vma_end_read(vma);
+		return ret;
+	}
+
+	/* fall back to mmap_lock */
+	mmap_read_lock(mm);
+	vma = vma_lookup(mm, addr);
+	if (vma && binder_alloc_is_mapped(alloc))
+		ret = vm_insert_page(vma, addr, page);
+	mmap_read_unlock(mm);
+
+	return ret;
+}
+
 static struct page *binder_page_alloc(struct binder_alloc *alloc,
 				      unsigned long index,
 				      unsigned long addr)
@@ -254,9 +296,7 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 				      unsigned long index,
 				      unsigned long addr)
 {
-	struct vm_area_struct *vma;
 	struct page *page;
-	long npages;
 	int ret;
 
 	if (!mmget_not_zero(alloc->mm))
@@ -268,16 +308,7 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 		goto out;
 	}
 
-	mmap_read_lock(alloc->mm);
-	vma = vma_lookup(alloc->mm, addr);
-	if (!vma || !binder_alloc_is_mapped(alloc)) {
-		__free_page(page);
-		pr_err("%d: %s failed, no vma\n", alloc->pid, __func__);
-		ret = -ESRCH;
-		goto unlock;
-	}
-
-	ret = vm_insert_page(vma, addr, page);
+	ret = binder_page_insert(alloc, addr, page);
 	switch (ret) {
 	case -EBUSY:
 		/*
@@ -287,8 +318,8 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 		 */
 		ret = 0;
 		__free_page(page);
-		npages = get_user_pages_remote(alloc->mm, addr, 1, 0, &page, NULL);
-		if (npages <= 0) {
+		page = binder_page_lookup(alloc, addr);
+		if (!page) {
 			pr_err("%d: failed to find page at offset %lx\n",
 			       alloc->pid, addr - alloc->vm_start);
 			ret = -ESRCH;
@@ -306,8 +337,6 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 		ret = -ENOMEM;
 		break;
 	}
-unlock:
-	mmap_read_unlock(alloc->mm);
 out:
 	mmput_async(alloc->mm);
 	return ret;
-- 
2.47.0.277.g8800431eea-goog

[PATCH v3 8/8] binder: propagate vm_insert_page() errors

[Carlos Llamas]

Instead of always overriding errors with -ENOMEM, propagate the specific
error code returned by vm_insert_page(). This allows for more accurate
error logs and handling.

Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Carlos Llamas <cmllamas@google.com>
---
 drivers/android/binder_alloc.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
index 7d2cad9beebb..dd15519e321f 100644
--- a/drivers/android/binder_alloc.c
+++ b/drivers/android/binder_alloc.c
@@ -334,7 +334,6 @@ static int binder_install_single_page(struct binder_alloc *alloc,
 		__free_page(page);
 		pr_err("%d: %s failed to insert page at offset %lx with %d\n",
 		       alloc->pid, __func__, addr - alloc->vm_start, ret);
-		ret = -ENOMEM;
 		break;
 	}
 out:
-- 
2.47.0.277.g8800431eea-goog