Re: [PATCH]: Improve performance of LZO hibernation

From: Bojan Smojver
Date: Wed Sep 21 2011 - 21:39:49 EST

On Mon, 2011-09-19 at 17:53 +1000, Bojan Smojver wrote:
> OK, v2 of the patch is here.

Version 3 follows. Thanks to Rafael for pointing out that there is no
need to enable nonboot cpus - they are already enabled. I also fixed a
potential problem where we could get wedged in an infinite loop on image
load, if the images is not correct.

-----------------------------------
Use threads for LZO compression/decompression on hibernate/thaw.
Improve read buffering on thaw.
v3

Signed-off-by: Bojan Smojver <bojan@xxxxxxxxxxxxx>
---
kernel/power/swap.c | 535 +++++++++++++++++++++++++++++++++++++--------------
1 files changed, 395 insertions(+), 140 deletions(-)

diff --git a/kernel/power/swap.c b/kernel/power/swap.c
index 7c97c3a..1f88d34 100644
--- a/kernel/power/swap.c
+++ b/kernel/power/swap.c
@@ -27,6 +27,9 @@
#include <linux/slab.h>
#include <linux/lzo.h>
#include <linux/vmalloc.h>
+#include <linux/cpu.h>
+#include <linux/atomic.h>
+#include <linux/kthread.h>

#include "power.h"

@@ -372,6 +375,9 @@ 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 2
+
/**
* save_image - save the suspend image data
*/
@@ -419,6 +425,46 @@ static int save_image(struct swap_map_handle *handle,
return ret;
}

+/**
+ * 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 */
+ wait_queue_head_t go; /* start compression */
+ wait_queue_head_t done; /* compression done */
+ int ret; /* return code */
+ 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())
+ 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.
@@ -434,45 +480,74 @@ static int save_image_lzo(struct swap_map_handle *handle,
int ret = 0;
int nr_pages;
int err2;
+ int cpu;
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, thr, cthr, nthr;
+ unsigned char *page = NULL;
+ struct cmp_data *data = NULL;
+
+ /*
+ * We'll limit the number of threads for compression to limit memory
+ * footprint.
+ */
+ nthr = num_online_cpus() - 1;
+ nthr = nthr > LZO_THREADS ? LZO_THREADS : (nthr < 1 ? 1 : nthr);

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) * nthr);
+ if (!data) {
+ printk(KERN_ERR "PM: Failed to allocate LZO data\n");
+ ret = -ENOMEM;
+ goto out_clean;
}
-
- 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;
+ for (thr = 0; thr < nthr; thr++)
+ memset(&data[thr], 0, offsetof(struct cmp_data, go));
+
+ /*
+ * Start the compression threads.
+ */
+ for (thr = 0; thr < nthr; thr++) {
+ init_waitqueue_head(&data[thr].go);
+ init_waitqueue_head(&data[thr].done);
+
+ data[thr].thr = kthread_create(lzo_compress_threadfn,
+ &data[thr],
+ "image_compress/%zu", 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;
+ }
}

- 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;
+ /*
+ * Bind the threads to CPUs and wake them up.
+ */
+ thr = 0;
+ for_each_online_cpu(cpu) {
+ if (cpu == smp_processor_id())
+ continue;
+ kthread_bind(data[thr++].thr, cpu);
+ if (thr >= nthr)
+ break;
}
+ for (thr = 0; thr < nthr; thr++)
+ wake_up_process(data[thr].thr);

printk(KERN_INFO
+ "PM: Using %zu thread(s) for compression.\n"
"PM: Compressing and saving image data (%u pages) ... ",
- nr_to_write);
+ nthr, nr_to_write);
m = nr_to_write / 100;
if (!m)
m = 1;
@@ -480,54 +555,75 @@ static int save_image_lzo(struct swap_map_handle *handle,
bio = NULL;
do_gettimeofday(&start);
for (;;) {
- for (off = 0; off < LZO_UNC_SIZE; off += PAGE_SIZE) {
- ret = snapshot_read_next(snapshot);
- if (ret < 0)
- goto out_finish;
-
- if (!ret)
+ for (thr = 0; thr < nthr; thr++) {
+ for (off = 0; off < LZO_UNC_SIZE; off += PAGE_SIZE) {
+ ret = snapshot_read_next(snapshot);
+ if (ret < 0)
+ goto out_finish;
+
+ if (!ret)
+ break;
+
+ 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);
+ nr_pages++;
+ }
+ if (!off)
break;

- memcpy(unc + off, data_of(*snapshot), PAGE_SIZE);
+ data[thr].unc_len = off;

- if (!(nr_pages % m))
- printk(KERN_CONT "\b\b\b\b%3d%%", nr_pages / m);
- nr_pages++;
+ atomic_set(&data[thr].ready, 1);
+ wake_up(&data[thr].go);
}

- if (!off)
+ if (!thr)
break;

- unc_len = off;
- ret = lzo1x_1_compress(unc, unc_len,
- cmp + LZO_HEADER, &cmp_len, wrk);
- if (ret < 0) {
- printk(KERN_ERR "PM: LZO compression failed\n");
- break;
- }
+ for (cthr = thr, thr = 0; thr < cthr; thr++) {
+ wait_event(data[thr].done,
+ atomic_read(&data[thr].stop));
+ atomic_set(&data[thr].stop, 0);

- if (unlikely(!cmp_len ||
- cmp_len > lzo1x_worst_compress(unc_len))) {
- printk(KERN_ERR "PM: Invalid LZO compressed length\n");
- ret = -1;
- break;
- }
+ ret = data[thr].ret;

- *(size_t *)cmp = 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.
- */
- for (off = 0; off < LZO_HEADER + cmp_len; off += PAGE_SIZE) {
- memcpy(page, cmp + off, PAGE_SIZE);
+ if (ret < 0) {
+ printk(KERN_ERR "PM: LZO compression failed\n");
+ goto out_finish;
+ }

- ret = swap_write_page(handle, page, &bio);
- if (ret)
+ 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;
goto out_finish;
+ }
+
+ *(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.
+ */
+ 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)
+ goto out_finish;
+ }
}
}

@@ -541,11 +637,15 @@ out_finish:
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:
+ for (thr = 0; thr < nthr; thr++) {
+ if (data[thr].thr) {
+ kthread_stop(data[thr].thr);
+ wake_up(&data[thr].go);
+ }
+ }
+ if (data) vfree(data);
+ if (page) free_page((unsigned long)page);

return ret;
}
@@ -743,6 +843,46 @@ static int load_image(struct swap_map_handle *handle,
}

/**
+ * 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 */
+ wait_queue_head_t go; /* start compression */
+ wait_queue_head_t done; /* compression done */
+ int ret; /* return code */
+ 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())
+ 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;
+}
+
+/**
* load_image_lzo - Load compressed image data and decompress them with LZO.
* @handle: Swap map handle to use for loading data.
* @snapshot: Image to copy uncompressed data into.
@@ -754,49 +894,90 @@ static int load_image_lzo(struct swap_map_handle *handle,
{
unsigned int m;
int error = 0;
+ int cpu, 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");
-
- while (i)
- free_page((unsigned long)page[--i]);
+ size_t i, off, thr, cthr, nthr;
+ size_t ring = 0, pg = 0, npages, have = 0, want, need, asked = 0;
+ unsigned char **page = NULL;
+ struct dec_data *data = NULL;
+
+ /*
+ * We'll limit the number of threads for decompression to limit memory
+ * footprint.
+ */
+ nthr = num_online_cpus() - 1;
+ nthr = nthr > LZO_THREADS ? LZO_THREADS : (nthr < 1 ? 1 : nthr);
+
+ page = vmalloc(sizeof(*page) * MAP_PAGE_ENTRIES);
+ if (!page) {
+ printk(KERN_ERR "PM: Failed to allocate LZO page\n");
+ error = -ENOMEM;
+ goto out_clean;
+ }
+ memset(page, 0, sizeof(*page) * MAP_PAGE_ENTRIES);

- return -ENOMEM;
+ data = vmalloc(sizeof(*data) * nthr);
+ if (!data) {
+ printk(KERN_ERR "PM: Failed to allocate LZO data\n");
+ error = -ENOMEM;
+ goto out_clean;
+ }
+ for (thr = 0; thr < nthr; thr++)
+ memset(&data[thr], 0, offsetof(struct cmp_data, go));
+
+ /*
+ * Start the decompression threads.
+ */
+ for (thr = 0; thr < nthr; thr++) {
+ init_waitqueue_head(&data[thr].go);
+ init_waitqueue_head(&data[thr].done);
+
+ data[thr].thr = kthread_create(lzo_decompress_threadfn,
+ &data[thr],
+ "image_decompress/%zu", thr);
+ if (IS_ERR(data[thr].thr)) {
+ data[thr].thr = NULL;
+ printk(KERN_ERR
+ "PM: Cannot start decompression threads\n");
+ error = -ENOMEM;
+ goto out_clean;
}
}

- unc = vmalloc(LZO_UNC_SIZE);
- if (!unc) {
- printk(KERN_ERR "PM: Failed to allocate LZO uncompressed\n");
-
- for (i = 0; i < LZO_CMP_PAGES; i++)
- free_page((unsigned long)page[i]);
-
- return -ENOMEM;
+ for (i = 0; i < MAP_PAGE_ENTRIES; i++) {
+ page[i] = (void *)__get_free_page(__GFP_WAIT | __GFP_HIGH);
+ if (!page[i]) {
+ if (i < LZO_CMP_PAGES) {
+ printk(KERN_ERR
+ "PM: Failed to allocate LZO pages\n");
+ error = -ENOMEM;
+ goto out_clean;
+ }
+ }
}
-
- cmp = vmalloc(LZO_CMP_SIZE);
- if (!cmp) {
- printk(KERN_ERR "PM: Failed to allocate LZO compressed\n");
-
- vfree(unc);
- for (i = 0; i < LZO_CMP_PAGES; i++)
- free_page((unsigned long)page[i]);
-
- return -ENOMEM;
+ want = npages = i;
+
+ /*
+ * Bind the threads to CPUs and wake them up.
+ */
+ thr = 0;
+ for_each_online_cpu(cpu) {
+ if (cpu == smp_processor_id())
+ continue;
+ kthread_bind(data[thr++].thr, cpu);
+ if (thr >= nthr)
+ break;
}
+ for (thr = 0; thr < nthr; thr++)
+ wake_up_process(data[thr].thr);

printk(KERN_INFO
+ "PM: Using %zu thread(s) for decompression.\n"
"PM: Loading and decompressing image data (%u pages) ... ",
- nr_to_read);
+ nthr, nr_to_read);
m = nr_to_read / 100;
if (!m)
m = 1;
@@ -808,61 +989,128 @@ static int load_image_lzo(struct swap_map_handle *handle,
if (error <= 0)
goto out_finish;

- for (;;) {
- error = swap_read_page(handle, page[0], NULL); /* sync */
- if (error)
- break;
-
- 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;
- break;
+ for(;;) {
+ for (i = 0; !eof && i < want; i++) {
+ error = swap_read_page(handle, page[ring], &bio);
+ if (error) {
+ /*
+ * 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 >= npages)
+ ring = 0;
}
+ asked += i;
+ want -= i;
+
+ /*
+ * We are out of data, wait for some more.
+ */
+ if (!have) {
+ if (!asked)
+ break;

- for (off = PAGE_SIZE, i = 1;
- off < LZO_HEADER + cmp_len; off += PAGE_SIZE, i++) {
- error = swap_read_page(handle, page[i], &bio);
+ error = hib_wait_on_bio_chain(&bio);
if (error)
goto out_finish;
+ have += asked;
+ asked = 0;
+ if (eof)
+ eof = 2;
}

- error = hib_wait_on_bio_chain(&bio); /* need all data now */
- if (error)
- goto out_finish;
+ for (thr = 0; have && thr < nthr; 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");
+ error = -1;
+ goto out_finish;
+ }

- for (off = 0, i = 0;
- off < LZO_HEADER + cmp_len; off += PAGE_SIZE, i++) {
- memcpy(cmp + off, page[i], PAGE_SIZE);
- }
+ need = DIV_ROUND_UP(data[thr].cmp_len + LZO_HEADER,
+ PAGE_SIZE);
+ if (need > have) {
+ if (eof > 1) {
+ error = -1;
+ goto out_finish;
+ }
+ break;
+ }

- 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;
+ 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 >= npages)
+ pg = 0;
+ }
+
+ 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) {
+ error = hib_wait_on_bio_chain(&bio);
+ if (error)
+ 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 (cthr = thr, thr = 0; thr < cthr; thr++) {
+ wait_event(data[thr].done,
+ atomic_read(&data[thr].stop));
+ atomic_set(&data[thr].stop, 0);
+
+ error = data[thr].ret;

- if (!(nr_pages % m))
- printk("\b\b\b\b%3d%%", nr_pages / m);
- nr_pages++;
+ if (error < 0) {
+ printk(KERN_ERR
+ "PM: LZO decompression failed\n");
+ goto out_finish;
+ }

- error = snapshot_write_next(snapshot);
- if (error <= 0)
+ 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");
+ error = -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)
+ goto out_finish;
+ }
}
}

@@ -876,11 +1124,18 @@ out_finish:
} else
printk("\n");
swsusp_show_speed(&start, &stop, nr_to_read, "Read");
-
- vfree(cmp);
- vfree(unc);
- for (i = 0; i < LZO_CMP_PAGES; i++)
- free_page((unsigned long)page[i]);
+out_clean:
+ for (i = 0; i < npages; i++)
+ if (page[i])
+ free_page((unsigned long)page[i]);
+ for (thr = 0; thr < nthr; thr++) {
+ if (data[thr].thr) {
+ kthread_stop(data[thr].thr);
+ wake_up(&data[thr].go);
+ }
+ }
+ if (data) vfree(data);
+ if (page) vfree(page);

return error;
}
-----------------------------------

--
Bojan

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/