PM / Hibernate: Improve performance of LZO/plain hibernation, checksum image

Use threads for LZO compression/decompression on hibernate/thaw.
Improve buffering on hibernate/thaw.
Calculate/verify CRC32 of the image pages on hibernate/thaw.

In my testing, this improved write/read speed by a factor of about two.
Signed-off-by: Bojan Smojver <bojan@rexursive.com>
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>

kernel/power/Kconfig

@@ -27,6 +27,7 @@ config HIBERNATION
 	select HIBERNATE_CALLBACKS
 	select LZO_COMPRESS
 	select LZO_DECOMPRESS
+	select CRC32
 	---help---
 	  Enable the suspend to disk (STD) functionality, which is usually
 	  called "hibernation" in user interfaces.  STD checkpoints the

kernel/power/hibernate.c

@@ -657,6 +657,9 @@ int hibernate(void)
 			flags |= SF_PLATFORM_MODE;
 		if (nocompress)
 			flags |= SF_NOCOMPRESS_MODE;
+		else
+		        flags |= SF_CRC32_MODE;
+
 		pr_debug("PM: writing image.\n");
 		error = swsusp_write(flags);
 		swsusp_free();

kernel/power/power.h

@@ -146,6 +146,7 @@ extern int swsusp_swap_in_use(void);
  */
 #define SF_PLATFORM_MODE	1
 #define SF_NOCOMPRESS_MODE	2
+#define SF_CRC32_MODE	        4
 
 /* kernel/power/hibernate.c */
 extern int swsusp_check(void);

kernel/power/swap.c

@@ -27,6 +27,10 @@
 #include <linux/slab.h>
 #include <linux/lzo.h>
 #include <linux/vmalloc.h>
+#include <linux/cpumask.h>
+#include <linux/atomic.h>
+#include <linux/kthread.h>
+#include <linux/crc32.h>
 
 #include "power.h"
 
@@ -43,8 +47,7 @@
  *	allocated and populated one at a time, so we only need one memory
  *	page to set up the entire structure.
  *
- *	During resume we also only need to use one swap_map_page structure
- *	at a time.
+ *	During resume we pick up all swap_map_page structures into a list.
  */
 
 #define MAP_PAGE_ENTRIES	(PAGE_SIZE / sizeof(sector_t) - 1)
@@ -54,6 +57,11 @@ struct swap_map_page {
 	sector_t next_swap;
 };
 
+struct swap_map_page_list {
+	struct swap_map_page *map;
+	struct swap_map_page_list *next;
+};
+
 /**
  *	The swap_map_handle structure is used for handling swap in
  *	a file-alike way
@@ -61,13 +69,18 @@ struct swap_map_page {
 
 struct swap_map_handle {
 	struct swap_map_page *cur;
+	struct swap_map_page_list *maps;
 	sector_t cur_swap;
 	sector_t first_sector;
 	unsigned int k;
+	unsigned long nr_free_pages, written;
+	u32 crc32;
 };
 
 struct swsusp_header {
-	char reserved[PAGE_SIZE - 20 - sizeof(sector_t) - sizeof(int)];
+	char reserved[PAGE_SIZE - 20 - sizeof(sector_t) - sizeof(int) -
+	              sizeof(u32)];
+	u32	crc32;
 	sector_t image;
 	unsigned int flags;	/* Flags to pass to the "boot" kernel */
 	char	orig_sig[10];
@@ -199,6 +212,8 @@ static int mark_swapfiles(struct swap_map_handle *handle, unsigned int flags)
 		memcpy(swsusp_header->sig, HIBERNATE_SIG, 10);
 		swsusp_header->image = handle->first_sector;
 		swsusp_header->flags = flags;
+		if (flags & SF_CRC32_MODE)
+			swsusp_header->crc32 = handle->crc32;
 		error = hib_bio_write_page(swsusp_resume_block,
 					swsusp_header, NULL);
 	} else {
@@ -245,12 +260,20 @@ static int swsusp_swap_check(void)
 static int write_page(void *buf, sector_t offset, struct bio **bio_chain)
 {
 	void *src;
+	int ret;
 
 	if (!offset)
 		return -ENOSPC;
 
 	if (bio_chain) {
 		src = (void *)__get_free_page(__GFP_WAIT | __GFP_HIGH);
+		if (src) {
+			copy_page(src, buf);
+		} else {
+			ret = hib_wait_on_bio_chain(bio_chain); /* Free pages */
+			if (ret)
+				return ret;
+			src = (void *)__get_free_page(__GFP_WAIT | __GFP_HIGH);
 			if (src) {
 				copy_page(src, buf);
 			} else {
@@ -258,6 +281,7 @@ static int write_page(void *buf, sector_t offset, struct bio **bio_chain)
 				bio_chain = NULL;	/* Go synchronous */
 				src = buf;
 			}
+		}
 	} else {
 		src = buf;
 	}
@@ -293,6 +317,8 @@ static int get_swap_writer(struct swap_map_handle *handle)
 		goto err_rel;
 	}
 	handle->k = 0;
+	handle->nr_free_pages = nr_free_pages() >> 1;
+	handle->written = 0;
 	handle->first_sector = handle->cur_swap;
 	return 0;
 err_rel:
@@ -316,20 +342,23 @@ static int swap_write_page(struct swap_map_handle *handle, void *buf,
 		return error;
 	handle->cur->entries[handle->k++] = offset;
 	if (handle->k >= MAP_PAGE_ENTRIES) {
-		error = hib_wait_on_bio_chain(bio_chain);
-		if (error)
-			goto out;
 		offset = alloc_swapdev_block(root_swap);
 		if (!offset)
 			return -ENOSPC;
 		handle->cur->next_swap = offset;
-		error = write_page(handle->cur, handle->cur_swap, NULL);
+		error = write_page(handle->cur, handle->cur_swap, bio_chain);
 		if (error)
 			goto out;
 		clear_page(handle->cur);
 		handle->cur_swap = offset;
 		handle->k = 0;
 	}
+	if (bio_chain && ++handle->written > handle->nr_free_pages) {
+		error = hib_wait_on_bio_chain(bio_chain);
+		if (error)
+			goto out;
+		handle->written = 0;
+	}
  out:
 	return error;
 }
@@ -372,6 +401,13 @@ static int swap_writer_finish(struct swap_map_handle *handle,
 			             LZO_HEADER, PAGE_SIZE)
 #define LZO_CMP_SIZE	(LZO_CMP_PAGES * PAGE_SIZE)
 
+/* Maximum number of threads for compression/decompression. */
+#define LZO_THREADS	3
+
+/* Maximum number of pages for read buffering. */
+#define LZO_READ_PAGES	(MAP_PAGE_ENTRIES * 8)
+
+
 /**
  *	save_image - save the suspend image data
  */
@@ -419,6 +455,92 @@ static int save_image(struct swap_map_handle *handle,
 	return ret;
 }
 
+/**
+ * Structure used for CRC32.
+ */
+struct crc_data {
+	struct task_struct *thr;                  /* thread */
+	atomic_t ready;                           /* ready to start flag */
+	atomic_t stop;                            /* ready to stop flag */
+	unsigned run_threads;                     /* nr current threads */
+	wait_queue_head_t go;                     /* start crc update */
+	wait_queue_head_t done;                   /* crc update done */
+	u32 *crc32;                               /* points to handle's crc32 */
+	size_t *unc_len[LZO_THREADS];             /* uncompressed lengths */
+	unsigned char *unc[LZO_THREADS];          /* uncompressed data */
+};
+
+/**
+ * CRC32 update function that runs in its own thread.
+ */
+static int crc32_threadfn(void *data)
+{
+	struct crc_data *d = data;
+	unsigned i;
+
+	while (1) {
+		wait_event(d->go, atomic_read(&d->ready) ||
+		                  kthread_should_stop());
+		if (kthread_should_stop()) {
+			d->thr = NULL;
+			atomic_set(&d->stop, 1);
+			wake_up(&d->done);
+			break;
+		}
+		atomic_set(&d->ready, 0);
+
+		for (i = 0; i < d->run_threads; i++)
+			*d->crc32 = crc32_le(*d->crc32,
+			                     d->unc[i], *d->unc_len[i]);
+		atomic_set(&d->stop, 1);
+		wake_up(&d->done);
+	}
+	return 0;
+}
+/**
+ * Structure used for LZO data compression.
+ */
+struct cmp_data {
+	struct task_struct *thr;                  /* thread */
+	atomic_t ready;                           /* ready to start flag */
+	atomic_t stop;                            /* ready to stop flag */
+	int ret;                                  /* return code */
+	wait_queue_head_t go;                     /* start compression */
+	wait_queue_head_t done;                   /* compression done */
+	size_t unc_len;                           /* uncompressed length */
+	size_t cmp_len;                           /* compressed length */
+	unsigned char unc[LZO_UNC_SIZE];          /* uncompressed buffer */
+	unsigned char cmp[LZO_CMP_SIZE];          /* compressed buffer */
+	unsigned char wrk[LZO1X_1_MEM_COMPRESS];  /* compression workspace */
+};
+
+/**
+ * Compression function that runs in its own thread.
+ */
+static int lzo_compress_threadfn(void *data)
+{
+	struct cmp_data *d = data;
+
+	while (1) {
+		wait_event(d->go, atomic_read(&d->ready) ||
+		                  kthread_should_stop());
+		if (kthread_should_stop()) {
+			d->thr = NULL;
+			d->ret = -1;
+			atomic_set(&d->stop, 1);
+			wake_up(&d->done);
+			break;
+		}
+		atomic_set(&d->ready, 0);
+
+		d->ret = lzo1x_1_compress(d->unc, d->unc_len,
+		                          d->cmp + LZO_HEADER, &d->cmp_len,
+		                          d->wrk);
+		atomic_set(&d->stop, 1);
+		wake_up(&d->done);
+	}
+	return 0;
+}
 
 /**
  * save_image_lzo - Save the suspend image data compressed with LZO.
@@ -437,42 +559,93 @@ static int save_image_lzo(struct swap_map_handle *handle,
 	struct bio *bio;
 	struct timeval start;
 	struct timeval stop;
-	size_t off, unc_len, cmp_len;
-	unsigned char *unc, *cmp, *wrk, *page;
+	size_t off;
+	unsigned thr, run_threads, nr_threads;
+	unsigned char *page = NULL;
+	struct cmp_data *data = NULL;
+	struct crc_data *crc = NULL;
+
+	/*
+	 * We'll limit the number of threads for compression to limit memory
+	 * footprint.
+	 */
+	nr_threads = num_online_cpus() - 1;
+	nr_threads = clamp_val(nr_threads, 1, LZO_THREADS);
 
 	page = (void *)__get_free_page(__GFP_WAIT | __GFP_HIGH);
 	if (!page) {
 		printk(KERN_ERR "PM: Failed to allocate LZO page\n");
-		return -ENOMEM;
+		ret = -ENOMEM;
+		goto out_clean;
 	}
 
-	wrk = vmalloc(LZO1X_1_MEM_COMPRESS);
-	if (!wrk) {
-		printk(KERN_ERR "PM: Failed to allocate LZO workspace\n");
-		free_page((unsigned long)page);
-		return -ENOMEM;
+	data = vmalloc(sizeof(*data) * nr_threads);
+	if (!data) {
+		printk(KERN_ERR "PM: Failed to allocate LZO data\n");
+		ret = -ENOMEM;
+		goto out_clean;
 	}
+	for (thr = 0; thr < nr_threads; thr++)
+		memset(&data[thr], 0, offsetof(struct cmp_data, go));
 
-	unc = vmalloc(LZO_UNC_SIZE);
-	if (!unc) {
-		printk(KERN_ERR "PM: Failed to allocate LZO uncompressed\n");
-		vfree(wrk);
-		free_page((unsigned long)page);
-		return -ENOMEM;
+	crc = kmalloc(sizeof(*crc), GFP_KERNEL);
+	if (!crc) {
+		printk(KERN_ERR "PM: Failed to allocate crc\n");
+		ret = -ENOMEM;
+		goto out_clean;
 	}
+	memset(crc, 0, offsetof(struct crc_data, go));
 
-	cmp = vmalloc(LZO_CMP_SIZE);
-	if (!cmp) {
-		printk(KERN_ERR "PM: Failed to allocate LZO compressed\n");
-		vfree(unc);
-		vfree(wrk);
-		free_page((unsigned long)page);
-		return -ENOMEM;
+	/*
+	 * Start the compression threads.
+	 */
+	for (thr = 0; thr < nr_threads; thr++) {
+		init_waitqueue_head(&data[thr].go);
+		init_waitqueue_head(&data[thr].done);
+
+		data[thr].thr = kthread_run(lzo_compress_threadfn,
+		                            &data[thr],
+		                            "image_compress/%u", thr);
+		if (IS_ERR(data[thr].thr)) {
+			data[thr].thr = NULL;
+			printk(KERN_ERR
+			       "PM: Cannot start compression threads\n");
+			ret = -ENOMEM;
+			goto out_clean;
+		}
+	}
+
+	/*
+	 * Adjust number of free pages after all allocations have been done.
+	 * We don't want to run out of pages when writing.
+	 */
+	handle->nr_free_pages = nr_free_pages() >> 1;
+
+	/*
+	 * Start the CRC32 thread.
+	 */
+	init_waitqueue_head(&crc->go);
+	init_waitqueue_head(&crc->done);
+
+	handle->crc32 = 0;
+	crc->crc32 = &handle->crc32;
+	for (thr = 0; thr < nr_threads; thr++) {
+		crc->unc[thr] = data[thr].unc;
+		crc->unc_len[thr] = &data[thr].unc_len;
+	}
+
+	crc->thr = kthread_run(crc32_threadfn, crc, "image_crc32");
+	if (IS_ERR(crc->thr)) {
+		crc->thr = NULL;
+		printk(KERN_ERR "PM: Cannot start CRC32 thread\n");
+		ret = -ENOMEM;
+		goto out_clean;
 	}
 
 	printk(KERN_INFO
+		"PM: Using %u thread(s) for compression.\n"
 		"PM: Compressing and saving image data (%u pages) ...     ",
-		nr_to_write);
+		nr_threads, nr_to_write);
 	m = nr_to_write / 100;
 	if (!m)
 		m = 1;
@@ -480,6 +653,7 @@ static int save_image_lzo(struct swap_map_handle *handle,
 	bio = NULL;
 	do_gettimeofday(&start);
 	for (;;) {
+		for (thr = 0; thr < nr_threads; thr++) {
 			for (off = 0; off < LZO_UNC_SIZE; off += PAGE_SIZE) {
 				ret = snapshot_read_next(snapshot);
 				if (ret < 0)
@@ -488,42 +662,65 @@ static int save_image_lzo(struct swap_map_handle *handle,
 				if (!ret)
 					break;
 
-			memcpy(unc + off, data_of(*snapshot), PAGE_SIZE);
+				memcpy(data[thr].unc + off,
+				       data_of(*snapshot), PAGE_SIZE);
 
 				if (!(nr_pages % m))
-				printk(KERN_CONT "\b\b\b\b%3d%%", nr_pages / m);
+					printk(KERN_CONT "\b\b\b\b%3d%%",
+				               nr_pages / m);
 				nr_pages++;
 			}
-
 			if (!off)
 				break;
 
-		unc_len = off;
-		ret = lzo1x_1_compress(unc, unc_len,
-		                       cmp + LZO_HEADER, &cmp_len, wrk);
+			data[thr].unc_len = off;
+
+			atomic_set(&data[thr].ready, 1);
+			wake_up(&data[thr].go);
+		}
+
+		if (!thr)
+			break;
+
+		crc->run_threads = thr;
+		atomic_set(&crc->ready, 1);
+		wake_up(&crc->go);
+
+		for (run_threads = thr, thr = 0; thr < run_threads; thr++) {
+			wait_event(data[thr].done,
+			           atomic_read(&data[thr].stop));
+			atomic_set(&data[thr].stop, 0);
+
+			ret = data[thr].ret;
+
 			if (ret < 0) {
 				printk(KERN_ERR "PM: LZO compression failed\n");
-			break;
+				goto out_finish;
 			}
 
-		if (unlikely(!cmp_len ||
-		             cmp_len > lzo1x_worst_compress(unc_len))) {
-			printk(KERN_ERR "PM: Invalid LZO compressed length\n");
+			if (unlikely(!data[thr].cmp_len ||
+			             data[thr].cmp_len >
+			             lzo1x_worst_compress(data[thr].unc_len))) {
+				printk(KERN_ERR
+				       "PM: Invalid LZO compressed length\n");
 				ret = -1;
-			break;
+				goto out_finish;
 			}
 
-		*(size_t *)cmp = cmp_len;
+			*(size_t *)data[thr].cmp = data[thr].cmp_len;
 
 			/*
-		 * Given we are writing one page at a time to disk, we copy
-		 * that much from the buffer, although the last bit will likely
-		 * be smaller than full page. This is OK - we saved the length
-		 * of the compressed data, so any garbage at the end will be
-		 * discarded when we read it.
+			 * Given we are writing one page at a time to disk, we
+			 * copy that much from the buffer, although the last
+			 * bit will likely be smaller than full page. This is
+			 * OK - we saved the length of the compressed data, so
+			 * any garbage at the end will be discarded when we
+			 * read it.
 			 */
-		for (off = 0; off < LZO_HEADER + cmp_len; off += PAGE_SIZE) {
-			memcpy(page, cmp + off, PAGE_SIZE);
+			for (off = 0;
+			     off < LZO_HEADER + data[thr].cmp_len;
+			     off += PAGE_SIZE) {
+				memcpy(page, data[thr].cmp + off, PAGE_SIZE);
 
 				ret = swap_write_page(handle, page, &bio);
 				if (ret)
@@ -531,21 +728,34 @@ static int save_image_lzo(struct swap_map_handle *handle,
 			}
 		}
 
+		wait_event(crc->done, atomic_read(&crc->stop));
+		atomic_set(&crc->stop, 0);
+	}
+
 out_finish:
 	err2 = hib_wait_on_bio_chain(&bio);
 	do_gettimeofday(&stop);
 	if (!ret)
 		ret = err2;
-	if (!ret)
+	if (!ret) {
 		printk(KERN_CONT "\b\b\b\bdone\n");
-	else
+	} else {
 		printk(KERN_CONT "\n");
+	}
 	swsusp_show_speed(&start, &stop, nr_to_write, "Wrote");
-
-	vfree(cmp);
-	vfree(unc);
-	vfree(wrk);
-	free_page((unsigned long)page);
+out_clean:
+	if (crc) {
+		if (crc->thr)
+			kthread_stop(crc->thr);
+		kfree(crc);
+	}
+	if (data) {
+		for (thr = 0; thr < nr_threads; thr++)
+			if (data[thr].thr)
+				kthread_stop(data[thr].thr);
+		vfree(data);
+	}
+	if (page) free_page((unsigned long)page);
 
 	return ret;
 }
@@ -625,8 +835,15 @@ int swsusp_write(unsigned int flags)
 
 static void release_swap_reader(struct swap_map_handle *handle)
 {
-	if (handle->cur)
-		free_page((unsigned long)handle->cur);
+	struct swap_map_page_list *tmp;
+
+	while (handle->maps) {
+		if (handle->maps->map)
+			free_page((unsigned long)handle->maps->map);
+		tmp = handle->maps;
+		handle->maps = handle->maps->next;
+		kfree(tmp);
+	}
 	handle->cur = NULL;
 }
 
@@ -634,22 +851,46 @@ static int get_swap_reader(struct swap_map_handle *handle,
 		unsigned int *flags_p)
 {
 	int error;
+	struct swap_map_page_list *tmp, *last;
+	sector_t offset;
 
 	*flags_p = swsusp_header->flags;
 
 	if (!swsusp_header->image) /* how can this happen? */
 		return -EINVAL;
 
-	handle->cur = (struct swap_map_page *)get_zeroed_page(__GFP_WAIT | __GFP_HIGH);
-	if (!handle->cur)
+	handle->cur = NULL;
+	last = handle->maps = NULL;
+	offset = swsusp_header->image;
+	while (offset) {
+		tmp = kmalloc(sizeof(*handle->maps), GFP_KERNEL);
+		if (!tmp) {
+			release_swap_reader(handle);
 			return -ENOMEM;
+		}
+		memset(tmp, 0, sizeof(*tmp));
+		if (!handle->maps)
+			handle->maps = tmp;
+		if (last)
+			last->next = tmp;
+		last = tmp;
+
+		tmp->map = (struct swap_map_page *)
+		           __get_free_page(__GFP_WAIT | __GFP_HIGH);
+		if (!tmp->map) {
+			release_swap_reader(handle);
+			return -ENOMEM;
+		}
 
-	error = hib_bio_read_page(swsusp_header->image, handle->cur, NULL);
+		error = hib_bio_read_page(offset, tmp->map, NULL);
 		if (error) {
 			release_swap_reader(handle);
 			return error;
 		}
+		offset = tmp->map->next_swap;
+	}
 	handle->k = 0;
+	handle->cur = handle->maps->map;
 	return 0;
 }
 
@@ -658,6 +899,7 @@ static int swap_read_page(struct swap_map_handle *handle, void *buf,
 {
 	sector_t offset;
 	int error;
+	struct swap_map_page_list *tmp;
 
 	if (!handle->cur)
 		return -EINVAL;
@@ -668,13 +910,15 @@ static int swap_read_page(struct swap_map_handle *handle, void *buf,
 	if (error)
 		return error;
 	if (++handle->k >= MAP_PAGE_ENTRIES) {
-		error = hib_wait_on_bio_chain(bio_chain);
 		handle->k = 0;
-		offset = handle->cur->next_swap;
-		if (!offset)
+		free_page((unsigned long)handle->maps->map);
+		tmp = handle->maps;
+		handle->maps = handle->maps->next;
+		kfree(tmp);
+		if (!handle->maps)
 			release_swap_reader(handle);
-		else if (!error)
-			error = hib_bio_read_page(offset, handle->cur, NULL);
+		else
+			handle->cur = handle->maps->map;
 	}
 	return error;
 }
@@ -697,7 +941,7 @@ static int load_image(struct swap_map_handle *handle,
                       unsigned int nr_to_read)
 {
 	unsigned int m;
-	int error = 0;
+	int ret = 0;
 	struct timeval start;
 	struct timeval stop;
 	struct bio *bio;
@@ -713,15 +957,15 @@ static int load_image(struct swap_map_handle *handle,
 	bio = NULL;
 	do_gettimeofday(&start);
 	for ( ; ; ) {
-		error = snapshot_write_next(snapshot);
-		if (error <= 0)
+		ret = snapshot_write_next(snapshot);
+		if (ret <= 0)
 			break;
-		error = swap_read_page(handle, data_of(*snapshot), &bio);
-		if (error)
+		ret = swap_read_page(handle, data_of(*snapshot), &bio);
+		if (ret)
 			break;
 		if (snapshot->sync_read)
-			error = hib_wait_on_bio_chain(&bio);
-		if (error)
+			ret = hib_wait_on_bio_chain(&bio);
+		if (ret)
 			break;
 		if (!(nr_pages % m))
 			printk("\b\b\b\b%3d%%", nr_pages / m);
@@ -729,17 +973,61 @@ static int load_image(struct swap_map_handle *handle,
 	}
 	err2 = hib_wait_on_bio_chain(&bio);
 	do_gettimeofday(&stop);
-	if (!error)
-		error = err2;
-	if (!error) {
+	if (!ret)
+		ret = err2;
+	if (!ret) {
 		printk("\b\b\b\bdone\n");
 		snapshot_write_finalize(snapshot);
 		if (!snapshot_image_loaded(snapshot))
-			error = -ENODATA;
+			ret = -ENODATA;
 	} else
 		printk("\n");
 	swsusp_show_speed(&start, &stop, nr_to_read, "Read");
-	return error;
+	return ret;
+}
+
+/**
+ * Structure used for LZO data decompression.
+ */
+struct dec_data {
+	struct task_struct *thr;                  /* thread */
+	atomic_t ready;                           /* ready to start flag */
+	atomic_t stop;                            /* ready to stop flag */
+	int ret;                                  /* return code */
+	wait_queue_head_t go;                     /* start decompression */
+	wait_queue_head_t done;                   /* decompression done */
+	size_t unc_len;                           /* uncompressed length */
+	size_t cmp_len;                           /* compressed length */
+	unsigned char unc[LZO_UNC_SIZE];          /* uncompressed buffer */
+	unsigned char cmp[LZO_CMP_SIZE];          /* compressed buffer */
+};
+
+/**
+ * Deompression function that runs in its own thread.
+ */
+static int lzo_decompress_threadfn(void *data)
+{
+	struct dec_data *d = data;
+
+	while (1) {
+		wait_event(d->go, atomic_read(&d->ready) ||
+		                  kthread_should_stop());
+		if (kthread_should_stop()) {
+			d->thr = NULL;
+			d->ret = -1;
+			atomic_set(&d->stop, 1);
+			wake_up(&d->done);
+			break;
+		}
+		atomic_set(&d->ready, 0);
+
+		d->unc_len = LZO_UNC_SIZE;
+		d->ret = lzo1x_decompress_safe(d->cmp + LZO_HEADER, d->cmp_len,
+		                               d->unc, &d->unc_len);
+		atomic_set(&d->stop, 1);
+		wake_up(&d->done);
+	}
+	return 0;
 }
 
 /**
@@ -753,50 +1041,120 @@ static int load_image_lzo(struct swap_map_handle *handle,
                           unsigned int nr_to_read)
 {
 	unsigned int m;
-	int error = 0;
+	int ret = 0;
+	int eof = 0;
 	struct bio *bio;
 	struct timeval start;
 	struct timeval stop;
 	unsigned nr_pages;
-	size_t i, off, unc_len, cmp_len;
-	unsigned char *unc, *cmp, *page[LZO_CMP_PAGES];
-
-	for (i = 0; i < LZO_CMP_PAGES; i++) {
-		page[i] = (void *)__get_free_page(__GFP_WAIT | __GFP_HIGH);
-		if (!page[i]) {
-			printk(KERN_ERR "PM: Failed to allocate LZO page\n");
+	size_t off;
+	unsigned i, thr, run_threads, nr_threads;
+	unsigned ring = 0, pg = 0, ring_size = 0,
+	         have = 0, want, need, asked = 0;
+	unsigned long read_pages;
+	unsigned char **page = NULL;
+	struct dec_data *data = NULL;
+	struct crc_data *crc = NULL;
 
-			while (i)
-				free_page((unsigned long)page[--i]);
+	/*
+	 * We'll limit the number of threads for decompression to limit memory
+	 * footprint.
+	 */
+	nr_threads = num_online_cpus() - 1;
+	nr_threads = clamp_val(nr_threads, 1, LZO_THREADS);
 
-			return -ENOMEM;
+	page = vmalloc(sizeof(*page) * LZO_READ_PAGES);
+	if (!page) {
+		printk(KERN_ERR "PM: Failed to allocate LZO page\n");
+		ret = -ENOMEM;
+		goto out_clean;
 	}
+
+	data = vmalloc(sizeof(*data) * nr_threads);
+	if (!data) {
+		printk(KERN_ERR "PM: Failed to allocate LZO data\n");
+		ret = -ENOMEM;
+		goto out_clean;
 	}
+	for (thr = 0; thr < nr_threads; thr++)
+		memset(&data[thr], 0, offsetof(struct dec_data, go));
 
-	unc = vmalloc(LZO_UNC_SIZE);
-	if (!unc) {
-		printk(KERN_ERR "PM: Failed to allocate LZO uncompressed\n");
+	crc = kmalloc(sizeof(*crc), GFP_KERNEL);
+	if (!crc) {
+		printk(KERN_ERR "PM: Failed to allocate crc\n");
+		ret = -ENOMEM;
+		goto out_clean;
+	}
+	memset(crc, 0, offsetof(struct crc_data, go));
 
-		for (i = 0; i < LZO_CMP_PAGES; i++)
-			free_page((unsigned long)page[i]);
+	/*
+	 * Start the decompression threads.
+	 */
+	for (thr = 0; thr < nr_threads; thr++) {
+		init_waitqueue_head(&data[thr].go);
+		init_waitqueue_head(&data[thr].done);
+
+		data[thr].thr = kthread_run(lzo_decompress_threadfn,
+		                            &data[thr],
+		                            "image_decompress/%u", thr);
+		if (IS_ERR(data[thr].thr)) {
+			data[thr].thr = NULL;
+			printk(KERN_ERR
+			       "PM: Cannot start decompression threads\n");
+			ret = -ENOMEM;
+			goto out_clean;
+		}
+	}
 
-		return -ENOMEM;
+	/*
+	 * Start the CRC32 thread.
+	 */
+	init_waitqueue_head(&crc->go);
+	init_waitqueue_head(&crc->done);
+
+	handle->crc32 = 0;
+	crc->crc32 = &handle->crc32;
+	for (thr = 0; thr < nr_threads; thr++) {
+		crc->unc[thr] = data[thr].unc;
+		crc->unc_len[thr] = &data[thr].unc_len;
 	}
 
-	cmp = vmalloc(LZO_CMP_SIZE);
-	if (!cmp) {
-		printk(KERN_ERR "PM: Failed to allocate LZO compressed\n");
+	crc->thr = kthread_run(crc32_threadfn, crc, "image_crc32");
+	if (IS_ERR(crc->thr)) {
+		crc->thr = NULL;
+		printk(KERN_ERR "PM: Cannot start CRC32 thread\n");
+		ret = -ENOMEM;
+		goto out_clean;
+	}
 
-		vfree(unc);
-		for (i = 0; i < LZO_CMP_PAGES; i++)
-			free_page((unsigned long)page[i]);
+	/*
+	 * Adjust number of pages for read buffering, in case we are short.
+	 */
+	read_pages = (nr_free_pages() - snapshot_get_image_size()) >> 1;
+	read_pages = clamp_val(read_pages, LZO_CMP_PAGES, LZO_READ_PAGES);
 
-		return -ENOMEM;
+	for (i = 0; i < read_pages; i++) {
+		page[i] = (void *)__get_free_page(i < LZO_CMP_PAGES ?
+		                                  __GFP_WAIT | __GFP_HIGH :
+		                                  __GFP_WAIT);
+		if (!page[i]) {
+			if (i < LZO_CMP_PAGES) {
+				ring_size = i;
+				printk(KERN_ERR
+				       "PM: Failed to allocate LZO pages\n");
+				ret = -ENOMEM;
+				goto out_clean;
+			} else {
+				break;
+			}
 		}
+	}
+	want = ring_size = i;
 
 	printk(KERN_INFO
+		"PM: Using %u thread(s) for decompression.\n"
 		"PM: Loading and decompressing image data (%u pages) ...     ",
-		nr_to_read);
+		nr_threads, nr_to_read);
 	m = nr_to_read / 100;
 	if (!m)
 		m = 1;
@@ -804,85 +1162,189 @@ static int load_image_lzo(struct swap_map_handle *handle,
 	bio = NULL;
 	do_gettimeofday(&start);
 
-	error = snapshot_write_next(snapshot);
-	if (error <= 0)
+	ret = snapshot_write_next(snapshot);
+	if (ret <= 0)
 		goto out_finish;
 
-	for (;;) {
-		error = swap_read_page(handle, page[0], NULL); /* sync */
-		if (error)
+	for(;;) {
+		for (i = 0; !eof && i < want; i++) {
+			ret = swap_read_page(handle, page[ring], &bio);
+			if (ret) {
+				/*
+				 * On real read error, finish. On end of data,
+				 * set EOF flag and just exit the read loop.
+				 */
+				if (handle->cur &&
+				    handle->cur->entries[handle->k]) {
+					goto out_finish;
+				} else {
+					eof = 1;
 					break;
+				}
+			}
+			if (++ring >= ring_size)
+				ring = 0;
+		}
+		asked += i;
+		want -= i;
 
-		cmp_len = *(size_t *)page[0];
-		if (unlikely(!cmp_len ||
-		             cmp_len > lzo1x_worst_compress(LZO_UNC_SIZE))) {
-			printk(KERN_ERR "PM: Invalid LZO compressed length\n");
-			error = -1;
+		/*
+		 * We are out of data, wait for some more.
+		 */
+		if (!have) {
+			if (!asked)
 				break;
+
+			ret = hib_wait_on_bio_chain(&bio);
+			if (ret)
+				goto out_finish;
+			have += asked;
+			asked = 0;
+			if (eof)
+				eof = 2;
 		}
 
-		for (off = PAGE_SIZE, i = 1;
-		     off < LZO_HEADER + cmp_len; off += PAGE_SIZE, i++) {
-			error = swap_read_page(handle, page[i], &bio);
-			if (error)
+		if (crc->run_threads) {
+			wait_event(crc->done, atomic_read(&crc->stop));
+			atomic_set(&crc->stop, 0);
+			crc->run_threads = 0;
+		}
+
+		for (thr = 0; have && thr < nr_threads; thr++) {
+			data[thr].cmp_len = *(size_t *)page[pg];
+			if (unlikely(!data[thr].cmp_len ||
+			             data[thr].cmp_len >
+			             lzo1x_worst_compress(LZO_UNC_SIZE))) {
+				printk(KERN_ERR
+				       "PM: Invalid LZO compressed length\n");
+				ret = -1;
 				goto out_finish;
 			}
 
-		error = hib_wait_on_bio_chain(&bio); /* need all data now */
-		if (error)
+			need = DIV_ROUND_UP(data[thr].cmp_len + LZO_HEADER,
+			                    PAGE_SIZE);
+			if (need > have) {
+				if (eof > 1) {
+					ret = -1;
 					goto out_finish;
+				}
+				break;
+			}
 
-		for (off = 0, i = 0;
-		     off < LZO_HEADER + cmp_len; off += PAGE_SIZE, i++) {
-			memcpy(cmp + off, page[i], PAGE_SIZE);
+			for (off = 0;
+			     off < LZO_HEADER + data[thr].cmp_len;
+			     off += PAGE_SIZE) {
+				memcpy(data[thr].cmp + off,
+				       page[pg], PAGE_SIZE);
+				have--;
+				want++;
+				if (++pg >= ring_size)
+					pg = 0;
 			}
 
-		unc_len = LZO_UNC_SIZE;
-		error = lzo1x_decompress_safe(cmp + LZO_HEADER, cmp_len,
-		                              unc, &unc_len);
-		if (error < 0) {
-			printk(KERN_ERR "PM: LZO decompression failed\n");
-			break;
+			atomic_set(&data[thr].ready, 1);
+			wake_up(&data[thr].go);
 		}
 
-		if (unlikely(!unc_len ||
-		             unc_len > LZO_UNC_SIZE ||
-		             unc_len & (PAGE_SIZE - 1))) {
-			printk(KERN_ERR "PM: Invalid LZO uncompressed length\n");
-			error = -1;
-			break;
+		/*
+		 * Wait for more data while we are decompressing.
+		 */
+		if (have < LZO_CMP_PAGES && asked) {
+			ret = hib_wait_on_bio_chain(&bio);
+			if (ret)
+				goto out_finish;
+			have += asked;
+			asked = 0;
+			if (eof)
+				eof = 2;
 		}
 
-		for (off = 0; off < unc_len; off += PAGE_SIZE) {
-			memcpy(data_of(*snapshot), unc + off, PAGE_SIZE);
+		for (run_threads = thr, thr = 0; thr < run_threads; thr++) {
+			wait_event(data[thr].done,
+			           atomic_read(&data[thr].stop));
+			atomic_set(&data[thr].stop, 0);
+
+			ret = data[thr].ret;
+
+			if (ret < 0) {
+				printk(KERN_ERR
+				       "PM: LZO decompression failed\n");
+				goto out_finish;
+			}
+
+			if (unlikely(!data[thr].unc_len ||
+			             data[thr].unc_len > LZO_UNC_SIZE ||
+			             data[thr].unc_len & (PAGE_SIZE - 1))) {
+				printk(KERN_ERR
+				       "PM: Invalid LZO uncompressed length\n");
+				ret = -1;
+				goto out_finish;
+			}
+
+			for (off = 0;
+			     off < data[thr].unc_len; off += PAGE_SIZE) {
+				memcpy(data_of(*snapshot),
+				       data[thr].unc + off, PAGE_SIZE);
 
 				if (!(nr_pages % m))
 					printk("\b\b\b\b%3d%%", nr_pages / m);
 				nr_pages++;
 
-			error = snapshot_write_next(snapshot);
-			if (error <= 0)
+				ret = snapshot_write_next(snapshot);
+				if (ret <= 0) {
+					crc->run_threads = thr + 1;
+					atomic_set(&crc->ready, 1);
+					wake_up(&crc->go);
 					goto out_finish;
 				}
 			}
+		}
+
+		crc->run_threads = thr;
+		atomic_set(&crc->ready, 1);
+		wake_up(&crc->go);
+	}
 
 out_finish:
+	if (crc->run_threads) {
+		wait_event(crc->done, atomic_read(&crc->stop));
+		atomic_set(&crc->stop, 0);
+	}
 	do_gettimeofday(&stop);
-	if (!error) {
+	if (!ret) {
 		printk("\b\b\b\bdone\n");
 		snapshot_write_finalize(snapshot);
 		if (!snapshot_image_loaded(snapshot))
-			error = -ENODATA;
+			ret = -ENODATA;
+		if (!ret) {
+			if (swsusp_header->flags & SF_CRC32_MODE) {
+				if(handle->crc32 != swsusp_header->crc32) {
+					printk(KERN_ERR
+					       "PM: Invalid image CRC32!\n");
+					ret = -ENODATA;
+				}
+			}
+		}
 	} else
 		printk("\n");
 	swsusp_show_speed(&start, &stop, nr_to_read, "Read");
-
-	vfree(cmp);
-	vfree(unc);
-	for (i = 0; i < LZO_CMP_PAGES; i++)
+out_clean:
+	for (i = 0; i < ring_size; i++)
 		free_page((unsigned long)page[i]);
+	if (crc) {
+		if (crc->thr)
+			kthread_stop(crc->thr);
+		kfree(crc);
+	}
+	if (data) {
+		for (thr = 0; thr < nr_threads; thr++)
+			if (data[thr].thr)
+				kthread_stop(data[thr].thr);
+		vfree(data);
+	}
+	if (page) vfree(page);
 
-	return error;
+	return ret;
 }
 
 /**