xfs: stop using kmalloc in xfs_buf_get_noaddr

Currently xfs_buf_get_noaddr allocates memory using kmem_alloc which can end up either in kmalloc or vmalloc and assigns it to the buffer. This patch changes it to allocate individual pages and if there is more then one maps it into kernel virtual space using vmap.

This means the minimum buffer allocation is PAGE_SIZE now. For two of the three caller (log buffers, log recovery) that is perfectly fine, because they always allocate buffers that are a power of two of the page size anyway. For xfs_zero_remaining_bytes the minimum allocation goes up from blocksize to pagesize and thus there is a potential waste of memory for blocksize < pagesize allocations, which is unfortunate but not directly solveable when block drivers expect reference countable pages. To fix this waste xfs_zero_remaining_bytes could be rewritten to zero more than a single block at a time, which sounds like a good idea in general.

Signed-off-by Christoph Hellwig <[email protected]>

Index: linux-2.6/fs/xfs/linux-2.6/xfs_buf.c
===================================================================

fs/xfs/linux-2.6/xfs_buf.c +22 -29

@@ -314,7 +314,7 @@
 
 	ASSERT(list_empty(&bp->b_hash_list));
 
-	if (bp->b_flags & _XBF_PAGE_CACHE) {
+	if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
 		uint		i;
 
 		if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))

@@ -323,18 +323,11 @@
 		for (i = 0; i < bp->b_page_count; i++) {
 			struct page	*page = bp->b_pages[i];
 
-			ASSERT(!PagePrivate(page));
+			if (bp->b_flags & _XBF_PAGE_CACHE)
+				ASSERT(!PagePrivate(page));
 			page_cache_release(page);
 		}
 		_xfs_buf_free_pages(bp);
-	} else if (bp->b_flags & _XBF_KMEM_ALLOC) {
-		 /*
-		  * XXX(hch): bp->b_count_desired might be incorrect (see
-		  * xfs_buf_associate_memory for details), but fortunately
-		  * the Linux version of kmem_free ignores the len argument..
-		  */
-		kmem_free(bp->b_addr, bp->b_count_desired);
-		_xfs_buf_free_pages(bp);
 	}
 
 	xfs_buf_deallocate(bp);

@@ -764,41 +757,41 @@
 	size_t			len,
 	xfs_buftarg_t		*target)
 {
-	size_t			malloc_len = len;
+	unsigned long		page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
+	int			error, i;
 	xfs_buf_t		*bp;
-	void			*data;
-	int			error;
 
 	bp = xfs_buf_allocate(0);
 	if (unlikely(bp == NULL))
 		goto fail;
 	_xfs_buf_initialize(bp, target, 0, len, 0);
 
- try_again:
-	data = kmem_alloc(malloc_len, KM_SLEEP | KM_MAYFAIL | KM_LARGE);
-	if (unlikely(data == NULL))
+	error = _xfs_buf_get_pages(bp, page_count, 0);
+	if (error)
 		goto fail_free_buf;
 
-	/* check whether alignment matches.. */
-	if ((__psunsigned_t)data !=
-	    ((__psunsigned_t)data & ~target->bt_smask)) {
-		/* .. else double the size and try again */
-		kmem_free(data, malloc_len);
-		malloc_len <<= 1;
-		goto try_again;
-	}
-
-	error = xfs_buf_associate_memory(bp, data, len);
-	if (error)
+	for (i = 0; i < page_count; i++) {
+		bp->b_pages[i] = alloc_page(GFP_KERNEL);
+		if (!bp->b_pages[i])
+			goto fail_free_mem;
+	}
+	bp->b_flags |= _XBF_PAGES;
+
+	error = _xfs_buf_map_pages(bp, XBF_MAPPED);
+	if (unlikely(error)) {
+		printk(KERN_WARNING "%s: failed to map pages\n",
+				__FUNCTION__);
 		goto fail_free_mem;
-	bp->b_flags |= _XBF_KMEM_ALLOC;
+	}
 
 	xfs_buf_unlock(bp);
 
 	XB_TRACE(bp, "no_daddr", data);
 	return bp;
+
  fail_free_mem:
-	kmem_free(data, malloc_len);
+ 	for ( ; i >= 0; i++)
+		__free_page(bp->b_pages[i]);
  fail_free_buf:
 	xfs_buf_free(bp);
  fail:

Index: linux-2.6/fs/xfs/linux-2.6/xfs_buf.h
===================================================================

fs/xfs/linux-2.6/xfs_buf.h +1 -1

@@ -63,7 +63,7 @@
 
 	/* flags used only internally */
 	_XBF_PAGE_CACHE = (1 << 17),/* backed by pagecache		   */
-	_XBF_KMEM_ALLOC = (1 << 18),/* backed by kmem_alloc()		   */
+	_XBF_PAGES = (1 << 18),	    /* backed by refcounted pages	   */
 	_XBF_RUN_QUEUES = (1 << 19),/* run block device task queue	   */
 	_XBF_DELWRI_Q = (1 << 21),   /* buffer on delwri queue		   */
 } xfs_buf_flags_t;

Hi Christoph,
Did you do some testing for recovery when end of the physical log is seen ?

When you will be dealing with striped ICLOG buffers or big sized ICLOGs, header size might range from 512 to 2k. Also, this header
might be split into 2 parts at the end of physical log. Then, you don't have page size buffer. Please verify that XFS_BUF_SETP_PTRs work correctly for those cases.
Same thing is true when data section is split around physical log. You can get one part which is not PAGE sized.

Currently xfs_buf_get_noaddr allocates memory using kmem_alloc which can end up either in kmalloc or vmalloc and assigns it to the buffer. This patch changes it to allocate individual pages and if there is more then one maps it into kernel virtual space using vmap.

This means the minimum buffer allocation is PAGE_SIZE now. For two of the three caller (log buffers, log recovery) that is perfectly fine, because they always allocate buffers that are a power of two of the page size anyway. For xfs_zero_remaining_bytes the minimum allocation goes up from blocksize to pagesize and thus there is a potential waste of memory for blocksize < pagesize allocations, which is unfortunate but not directly solveable when block drivers expect reference countable pages. To fix this waste xfs_zero_remaining_bytes could be rewritten to zero more than a single block at a time, which sounds like a good idea in general.

Signed-off-by Christoph Hellwig <[email protected]>

Index: linux-2.6/fs/xfs/linux-2.6/xfs_buf.c
===================================================================

fs/xfs/linux-2.6/xfs_buf.c +22 -29

@@ -314,7 +314,7 @@
 
 	ASSERT(list_empty(&bp->b_hash_list));
 
-	if (bp->b_flags & _XBF_PAGE_CACHE) {
+	if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
 		uint		i;
 
 		if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))

@@ -323,18 +323,11 @@
 		for (i = 0; i < bp->b_page_count; i++) {
 			struct page	*page = bp->b_pages[i];
 
-			ASSERT(!PagePrivate(page));
+			if (bp->b_flags & _XBF_PAGE_CACHE)
+				ASSERT(!PagePrivate(page));
 			page_cache_release(page);
 		}
 		_xfs_buf_free_pages(bp);
-	} else if (bp->b_flags & _XBF_KMEM_ALLOC) {
-		 /*
-		  * XXX(hch): bp->b_count_desired might be incorrect (see
-		  * xfs_buf_associate_memory for details), but fortunately
-		  * the Linux version of kmem_free ignores the len argument..
-		  */
-		kmem_free(bp->b_addr, bp->b_count_desired);
-		_xfs_buf_free_pages(bp);
 	}
 
 	xfs_buf_deallocate(bp);

@@ -764,41 +757,41 @@
 	size_t			len,
 	xfs_buftarg_t		*target)
 {
-	size_t			malloc_len = len;
+	unsigned long		page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
+	int			error, i;
 	xfs_buf_t		*bp;
-	void			*data;
-	int			error;
 
 	bp = xfs_buf_allocate(0);
 	if (unlikely(bp == NULL))
 		goto fail;
 	_xfs_buf_initialize(bp, target, 0, len, 0);
 
- try_again:
-	data = kmem_alloc(malloc_len, KM_SLEEP | KM_MAYFAIL | KM_LARGE);
-	if (unlikely(data == NULL))
+	error = _xfs_buf_get_pages(bp, page_count, 0);
+	if (error)
 		goto fail_free_buf;
 
-	/* check whether alignment matches.. */
-	if ((__psunsigned_t)data !=
-	    ((__psunsigned_t)data & ~target->bt_smask)) {
-		/* .. else double the size and try again */
-		kmem_free(data, malloc_len);
-		malloc_len <<= 1;
-		goto try_again;
-	}
-
-	error = xfs_buf_associate_memory(bp, data, len);
-	if (error)
+	for (i = 0; i < page_count; i++) {
+		bp->b_pages[i] = alloc_page(GFP_KERNEL);
+		if (!bp->b_pages[i])
+			goto fail_free_mem;
+	}
+	bp->b_flags |= _XBF_PAGES;
+
+	error = _xfs_buf_map_pages(bp, XBF_MAPPED);
+	if (unlikely(error)) {
+		printk(KERN_WARNING "%s: failed to map pages\n",
+				__FUNCTION__);
 		goto fail_free_mem;
-	bp->b_flags |= _XBF_KMEM_ALLOC;
+	}
 
 	xfs_buf_unlock(bp);
 
 	XB_TRACE(bp, "no_daddr", data);
 	return bp;
+
 fail_free_mem:
-	kmem_free(data, malloc_len);
+ 	for ( ; i >= 0; i++)
+		__free_page(bp->b_pages[i]);
 fail_free_buf:
 	xfs_buf_free(bp);
 fail:

Index: linux-2.6/fs/xfs/linux-2.6/xfs_buf.h
===================================================================

fs/xfs/linux-2.6/xfs_buf.h +1 -1

@@ -63,7 +63,7 @@
 
 	/* flags used only internally */
 	_XBF_PAGE_CACHE = (1 << 17),/* backed by pagecache		   */
-	_XBF_KMEM_ALLOC = (1 << 18),/* backed by kmem_alloc()		   */
+	_XBF_PAGES = (1 << 18),	    /* backed by refcounted pages	   */
 	_XBF_RUN_QUEUES = (1 << 19),/* run block device task queue	   */
 	_XBF_DELWRI_Q = (1 << 21),   /* buffer on delwri queue		   */
 } xfs_buf_flags_t;

Incore log buffers are not always a power of two of the page size. In particular, when xfs is running over software raid devices, the log buffers are allocated to match the size of a stripe.

However, they are always a multiple of PAGE_SIZE, so we are still safe.

Ed Cashin found a bug in the error handling code for the case where a page allocation fails. Here's the updated version:

Index: linux-2.6/fs/xfs/linux-2.6/xfs_buf.c
===================================================================

fs/xfs/linux-2.6/xfs_buf.c +0 -0

fs/xfs/linux-2.6/xfs_buf.c +12 -1

+	for (i = 0; i < page_count; i++) {
+		bp->b_pages[i] = alloc_page(GFP_KERNEL);
+		if (!bp->b_pages[i])
+			goto fail_free_mem;
+	}
+	bp->b_flags |= _XBF_PAGES;
+
+	error = _xfs_buf_map_pages(bp, XBF_MAPPED);
+	if (unlikely(error)) {
+		printk(KERN_WARNING "%s: failed to map pages\n",
+				__FUNCTION__);
 		goto fail_free_mem;
-	bp->b_flags |= _XBF_KMEM_ALLOC;
+	}
 
 	xfs_buf_unlock(bp);
 

It looks like you might need: for (i--; i >= 0; i--) (or: for (j = 0; j < i; j++) etc.)

Because if the initial alloc_page loop goes to completion then:
i == pagecount
and if alloc_page loop terminates early then
bp->b_pages[i] == NULL
So we have gone 1 too far in both cases and need to start free'ing back one.
Unless I missed something.