diff mbox series

[1/7] brd: use XArray instead of radix-tree to index backing pages

Message ID 20230614114637.89759-2-hare@suse.de (mailing list archive)
State New, archived
Headers show
Series RFC: high-order folio support for I/O | expand

Commit Message

Hannes Reinecke June 14, 2023, 11:46 a.m. UTC
From: Pankaj Raghav <p.raghav@samsung.com>

XArray was introduced to hold large array of pointers with a simple API.
XArray API also provides array semantics which simplifies the way we store
and access the backing pages, and the code becomes significantly easier
to understand.

No performance difference was noticed between the two implementation
using fio with direct=1 [1].

[1] Performance in KIOPS:

          |  radix-tree |    XArray  |   Diff
          |             |            |
write     |    315      |     313    |   -0.6%
randwrite |    286      |     290    |   +1.3%
read      |    330      |     335    |   +1.5%
randread  |    309      |     312    |   +0.9%

Signed-off-by: Pankaj Raghav <p.raghav@samsung.com>
---
 drivers/block/brd.c | 93 ++++++++++++---------------------------------
 1 file changed, 24 insertions(+), 69 deletions(-)

Comments

Matthew Wilcox June 14, 2023, 12:45 p.m. UTC | #1
On Wed, Jun 14, 2023 at 01:46:31PM +0200, Hannes Reinecke wrote:
>  static void brd_free_pages(struct brd_device *brd)
>  {
> -	unsigned long pos = 0;
> -	struct page *pages[FREE_BATCH];
> -	int nr_pages;
> -
> -	do {
> -		int i;
> -
> -		nr_pages = radix_tree_gang_lookup(&brd->brd_pages,
> -				(void **)pages, pos, FREE_BATCH);
> -
> -		for (i = 0; i < nr_pages; i++) {
> -			void *ret;
> -
> -			BUG_ON(pages[i]->index < pos);
> -			pos = pages[i]->index;
> -			ret = radix_tree_delete(&brd->brd_pages, pos);
> -			BUG_ON(!ret || ret != pages[i]);
> -			__free_page(pages[i]);
> -		}
> -
> -		pos++;
> +	struct page *page;
> +	pgoff_t idx;
>  
> -		/*
> -		 * It takes 3.4 seconds to remove 80GiB ramdisk.
> -		 * So, we need cond_resched to avoid stalling the CPU.
> -		 */
> -		cond_resched();
> +	xa_for_each(&brd->brd_pages, idx, page) {
> +		__free_page(page);
> +		cond_resched_rcu();

This should be a regular cond_resched().  The body of the loop is run
without the RCU read lock held.  Surprised none of the bots have noticed
an unlock-underflow.  Perhaps they don't test brd ;-)

With that fixed,

Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Pankaj Raghav June 14, 2023, 12:50 p.m. UTC | #2
>>  
>> -		/*
>> -		 * It takes 3.4 seconds to remove 80GiB ramdisk.
>> -		 * So, we need cond_resched to avoid stalling the CPU.
>> -		 */
>> -		cond_resched();
>> +	xa_for_each(&brd->brd_pages, idx, page) {
>> +		__free_page(page);
>> +		cond_resched_rcu();
> 
> This should be a regular cond_resched().  The body of the loop is run
> without the RCU read lock held.  Surprised none of the bots have noticed
> an unlock-underflow.  Perhaps they don't test brd ;-)
> 
> With that fixed,
> 
> Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>

This patch is already queued up for 6.5 in Jens's tree.
I will send this as a fix soon. Thanks.
Hannes Reinecke June 14, 2023, 1:03 p.m. UTC | #3
On 6/14/23 14:50, Pankaj Raghav wrote:
>>>   
>>> -		/*
>>> -		 * It takes 3.4 seconds to remove 80GiB ramdisk.
>>> -		 * So, we need cond_resched to avoid stalling the CPU.
>>> -		 */
>>> -		cond_resched();
>>> +	xa_for_each(&brd->brd_pages, idx, page) {
>>> +		__free_page(page);
>>> +		cond_resched_rcu();
>>
>> This should be a regular cond_resched().  The body of the loop is run
>> without the RCU read lock held.  Surprised none of the bots have noticed
>> an unlock-underflow.  Perhaps they don't test brd ;-)
>>
>> With that fixed,
>>
>> Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
> 
> This patch is already queued up for 6.5 in Jens's tree.
> I will send this as a fix soon. Thanks.

Ah. Hence. I've been running off akpms mm-unstable branch, which doesn't 
have that patch (yet).

Cheers,

Hannes
diff mbox series

Patch

diff --git a/drivers/block/brd.c b/drivers/block/brd.c
index bcad9b926b0c..2f71376afc71 100644
--- a/drivers/block/brd.c
+++ b/drivers/block/brd.c
@@ -19,7 +19,7 @@ 
 #include <linux/highmem.h>
 #include <linux/mutex.h>
 #include <linux/pagemap.h>
-#include <linux/radix-tree.h>
+#include <linux/xarray.h>
 #include <linux/fs.h>
 #include <linux/slab.h>
 #include <linux/backing-dev.h>
@@ -28,7 +28,7 @@ 
 #include <linux/uaccess.h>
 
 /*
- * Each block ramdisk device has a radix_tree brd_pages of pages that stores
+ * Each block ramdisk device has a xarray brd_pages of pages that stores
  * the pages containing the block device's contents. A brd page's ->index is
  * its offset in PAGE_SIZE units. This is similar to, but in no way connected
  * with, the kernel's pagecache or buffer cache (which sit above our block
@@ -40,11 +40,9 @@  struct brd_device {
 	struct list_head	brd_list;
 
 	/*
-	 * Backing store of pages and lock to protect it. This is the contents
-	 * of the block device.
+	 * Backing store of pages. This is the contents of the block device.
 	 */
-	spinlock_t		brd_lock;
-	struct radix_tree_root	brd_pages;
+	struct xarray	        brd_pages;
 	u64			brd_nr_pages;
 };
 
@@ -56,21 +54,8 @@  static struct page *brd_lookup_page(struct brd_device *brd, sector_t sector)
 	pgoff_t idx;
 	struct page *page;
 
-	/*
-	 * The page lifetime is protected by the fact that we have opened the
-	 * device node -- brd pages will never be deleted under us, so we
-	 * don't need any further locking or refcounting.
-	 *
-	 * This is strictly true for the radix-tree nodes as well (ie. we
-	 * don't actually need the rcu_read_lock()), however that is not a
-	 * documented feature of the radix-tree API so it is better to be
-	 * safe here (we don't have total exclusion from radix tree updates
-	 * here, only deletes).
-	 */
-	rcu_read_lock();
 	idx = sector >> PAGE_SECTORS_SHIFT; /* sector to page index */
-	page = radix_tree_lookup(&brd->brd_pages, idx);
-	rcu_read_unlock();
+	page = xa_load(&brd->brd_pages, idx);
 
 	BUG_ON(page && page->index != idx);
 
@@ -83,7 +68,7 @@  static struct page *brd_lookup_page(struct brd_device *brd, sector_t sector)
 static int brd_insert_page(struct brd_device *brd, sector_t sector, gfp_t gfp)
 {
 	pgoff_t idx;
-	struct page *page;
+	struct page *page, *cur;
 	int ret = 0;
 
 	page = brd_lookup_page(brd, sector);
@@ -94,71 +79,42 @@  static int brd_insert_page(struct brd_device *brd, sector_t sector, gfp_t gfp)
 	if (!page)
 		return -ENOMEM;
 
-	if (radix_tree_maybe_preload(gfp)) {
-		__free_page(page);
-		return -ENOMEM;
-	}
+	xa_lock(&brd->brd_pages);
 
-	spin_lock(&brd->brd_lock);
 	idx = sector >> PAGE_SECTORS_SHIFT;
 	page->index = idx;
-	if (radix_tree_insert(&brd->brd_pages, idx, page)) {
+
+	cur = __xa_cmpxchg(&brd->brd_pages, idx, NULL, page, gfp);
+
+	if (unlikely(cur)) {
 		__free_page(page);
-		page = radix_tree_lookup(&brd->brd_pages, idx);
-		if (!page)
-			ret = -ENOMEM;
-		else if (page->index != idx)
+		ret = xa_err(cur);
+		if (!ret && (cur->index != idx))
 			ret = -EIO;
 	} else {
 		brd->brd_nr_pages++;
 	}
-	spin_unlock(&brd->brd_lock);
 
-	radix_tree_preload_end();
+	xa_unlock(&brd->brd_pages);
+
 	return ret;
 }
 
 /*
- * Free all backing store pages and radix tree. This must only be called when
+ * Free all backing store pages and xarray. This must only be called when
  * there are no other users of the device.
  */
-#define FREE_BATCH 16
 static void brd_free_pages(struct brd_device *brd)
 {
-	unsigned long pos = 0;
-	struct page *pages[FREE_BATCH];
-	int nr_pages;
-
-	do {
-		int i;
-
-		nr_pages = radix_tree_gang_lookup(&brd->brd_pages,
-				(void **)pages, pos, FREE_BATCH);
-
-		for (i = 0; i < nr_pages; i++) {
-			void *ret;
-
-			BUG_ON(pages[i]->index < pos);
-			pos = pages[i]->index;
-			ret = radix_tree_delete(&brd->brd_pages, pos);
-			BUG_ON(!ret || ret != pages[i]);
-			__free_page(pages[i]);
-		}
-
-		pos++;
+	struct page *page;
+	pgoff_t idx;
 
-		/*
-		 * It takes 3.4 seconds to remove 80GiB ramdisk.
-		 * So, we need cond_resched to avoid stalling the CPU.
-		 */
-		cond_resched();
+	xa_for_each(&brd->brd_pages, idx, page) {
+		__free_page(page);
+		cond_resched_rcu();
+	}
 
-		/*
-		 * This assumes radix_tree_gang_lookup always returns as
-		 * many pages as possible. If the radix-tree code changes,
-		 * so will this have to.
-		 */
-	} while (nr_pages == FREE_BATCH);
+	xa_destroy(&brd->brd_pages);
 }
 
 /*
@@ -372,8 +328,7 @@  static int brd_alloc(int i)
 	brd->brd_number		= i;
 	list_add_tail(&brd->brd_list, &brd_devices);
 
-	spin_lock_init(&brd->brd_lock);
-	INIT_RADIX_TREE(&brd->brd_pages, GFP_ATOMIC);
+	xa_init(&brd->brd_pages);
 
 	snprintf(buf, DISK_NAME_LEN, "ram%d", i);
 	if (!IS_ERR_OR_NULL(brd_debugfs_dir))