[RFC PATCH 11/16] DM: Try to avoid temporary buffer allocation to hold compressed data.

From: Ram Pai
Date: Mon Aug 15 2016 - 13:37:58 EST

Next message: Ram Pai: "[RFC PATCH 12/16] DM: release unneeded buffer as soon as possible."
Previous message: Ram Pai: "[RFC PATCH 09/16] DM: Delay allocation of decompression buffer during read."
In reply to: Ram Pai: "[RFC PATCH 09/16] DM: Delay allocation of decompression buffer during read."
Next in thread: Ram Pai: "[RFC PATCH 12/16] DM: release unneeded buffer as soon as possible."
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

The read path
a) allocates a temporary buffer to hold the compressed data.
b) reads the compressed data into the temporary buffer.
c) decompresses the compressed data into the caller's bio-buffer.

We know that the caller's bio-buffer will be atleast as large as the compressed
data. So we could save a buffer allocation in step (a) if we use caller's
bio-buffer to read in the compressed data from the disk. NOTE: this is not
possible all the time, but is possible if the destination within the bio-buffer
falls within the same segment.

This saves us from holding the temporary buffer across read boundaries. Which
is a huge advantage especially if we are operating under acute memory
starvation; mostly when this block device is used as a swap device.

Signed-off-by: Ram Pai <linuxram@xxxxxxxxxx>
---
drivers/md/dm-inplace-compress.c | 185 ++++++++++++++++++-------------------
drivers/md/dm-inplace-compress.h | 1 +
2 files changed, 91 insertions(+), 95 deletions(-)

diff --git a/drivers/md/dm-inplace-compress.c b/drivers/md/dm-inplace-compress.c
index f6b95e3..bc1cf70 100644
--- a/drivers/md/dm-inplace-compress.c
+++ b/drivers/md/dm-inplace-compress.c
@@ -810,10 +810,15 @@ static void dm_icomp_release_comp_buffer(struct dm_icomp_io_range *io)
{
if (!io->comp_data)
return;
- dm_icomp_kfree(io->comp_data, io->comp_len);
+
+ if (io->comp_kmap)
+ kunmap(io->comp_data);
+ else
+ dm_icomp_kfree(io->comp_data, io->comp_len);

io->comp_data = NULL;
io->comp_len = 0;
+ io->comp_kmap = false;
}

static void dm_icomp_free_io_range(struct dm_icomp_io_range *io)
@@ -857,6 +862,39 @@ static void dm_icomp_put_req(struct dm_icomp_req *req)
kmem_cache_free(dm_icomp_req_cachep, req);
}

+static void dm_icomp_bio_copy(struct bio *bio, off_t bio_off, void *buf,
+ ssize_t len, bool to_buf)
+{
+ struct bio_vec bv;
+ struct bvec_iter iter;
+ off_t buf_off = 0;
+ ssize_t size;
+ void *addr;
+
+ WARN_ON(bio_off + len > (bio_sectors(bio) << 9));
+
+ bio_for_each_segment(bv, bio, iter) {
+ int length = bv.bv_len;
+
+ if (bio_off >= length) {
+ bio_off -= length;
+ continue;
+ }
+ addr = kmap_atomic(bv.bv_page);
+ size = min_t(ssize_t, len, length - bio_off);
+ if (to_buf)
+ memcpy(buf + buf_off, addr + bio_off + bv.bv_offset,
+ size);
+ else
+ memcpy(addr + bio_off + bv.bv_offset, buf + buf_off,
+ size);
+ kunmap_atomic(addr);
+ bio_off = 0;
+ buf_off += size;
+ len -= size;
+ }
+}
+
static void dm_icomp_io_range_done(unsigned long error, void *context)
{
struct dm_icomp_io_range *io = context;
@@ -886,7 +924,7 @@ static inline int dm_icomp_compressor_maxlen(struct dm_icomp_info *info,
* comp_data
*/
static struct dm_icomp_io_range *dm_icomp_create_io_range(
- struct dm_icomp_req *req, int comp_len)
+ struct dm_icomp_req *req)
{
struct dm_icomp_io_range *io;

@@ -894,120 +932,67 @@ static struct dm_icomp_io_range *dm_icomp_create_io_range(
if (!io)
return NULL;

- io->comp_data = dm_icomp_kmalloc(comp_len, GFP_NOIO);
- if (!io->comp_data) {
- kmem_cache_free(dm_icomp_io_range_cachep, io);
- return NULL;
- }
-
io->io_req.notify.fn = dm_icomp_io_range_done;
io->io_req.notify.context = io;
io->io_req.client = req->info->io_client;
io->io_req.mem.type = DM_IO_KMEM;
- io->io_req.mem.ptr.addr = io->comp_data;
io->io_req.mem.offset = 0;

io->io_region.bdev = req->info->dev->bdev;
-
- io->comp_len = comp_len;
io->req = req;

- io->decomp_data = NULL;
- io->decomp_real_data = NULL;
- io->decomp_len = 0;
- io->decomp_kmap = false;
- io->decomp_req_len = 0;
+ io->comp_data = io->decomp_data = io->decomp_real_data = NULL;
+ io->comp_len = io->decomp_len = io->decomp_req_len = 0;
+ io->comp_kmap = io->decomp_kmap = false;
return io;
}

static struct dm_icomp_io_range *dm_icomp_create_io_read_range(
struct dm_icomp_req *req, int comp_len, int decomp_len)
{
- struct dm_icomp_io_range *io = dm_icomp_create_io_range(req, comp_len);
-
- if (io) {
- /* note down the requested length for decompress buffer.
- * but dont allocate it yet.
- */
- io->decomp_req_len = decomp_len;
- }
- return io;
-}
+ struct bio *bio = req->bio;
+ sector_t size = bio_sectors(bio)<<9;
+ int segments = bio_segments(bio);
+ void *addr;
+ struct dm_icomp_io_range *io = dm_icomp_create_io_range(req);

-static int dm_icomp_update_io_read_range(struct dm_icomp_io_range *io,
- struct bio *bio, ssize_t bio_off)
-{
- struct bvec_iter iter;
- struct bio_vec bv;
- bool just_use = false;
+ if (!io)
+ return NULL;

- if (io->decomp_len)
- return 0;
+ /* try reusing the bio buffer for compress data. */
+ if (segments == 1) {
+ struct bio_vec bv = bio_iovec(bio);

- /* use a bio buffer long enough to hold the uncompressed data */
- bio_for_each_segment(bv, bio, iter) {
- int avail_len;
- int length = bv.bv_len;
+ WARN_ON(size < comp_len);
+ addr = kmap(bv.bv_page);
+ } else
+ addr = dm_icomp_kmalloc(comp_len, GFP_NOIO);

- if (!just_use && bio_off >= length) {
- bio_off -= length;
- continue;
- }
- avail_len = just_use ? length : length-bio_off;
- if (avail_len >= io->decomp_req_len) {
- io->decomp_real_data = kmap(bv.bv_page);
- io->decomp_data = io->decomp_real_data + bio_off;
- io->decomp_len = io->decomp_req_len = avail_len;
- io->decomp_kmap = true;
- return 0;
- }
- just_use = true;
+ if (!addr) {
+ kmem_cache_free(dm_icomp_io_range_cachep, io);
+ return NULL;
}
+ io->comp_data = io->io_req.mem.ptr.addr = addr;
+ io->comp_len = comp_len;
+ io->comp_kmap = (segments == 1);
+ /* note down the requested length for decompress buffer.
+ * but dont allocate it yet.
+ /
+ io->decomp_req_len = decomp_len;
+ return io;
+}

- /* none available. :( Allocate one */
+static int dm_icomp_update_io_read_range(struct dm_icomp_io_range *io)
+{
io->decomp_data = dm_icomp_kmalloc(io->decomp_req_len, GFP_NOIO);
if (!io->decomp_data)
return 1;
io->decomp_real_data = io->decomp_data;
io->decomp_len = io->decomp_req_len;
io->decomp_kmap = false;
-
return 0;
}

-static void dm_icomp_bio_copy(struct bio *bio, off_t bio_off, void *buf,
- ssize_t len, bool to_buf)
-{
- struct bio_vec bv;
- struct bvec_iter iter;
- off_t buf_off = 0;
- ssize_t size;
- void *addr;
-
- WARN_ON(bio_off + len > (bio_sectors(bio) << 9));
-
- bio_for_each_segment(bv, bio, iter) {
- int length = bv.bv_len;
-
- if (bio_off >= length) {
- bio_off -= length;
- continue;
- }
- addr = kmap_atomic(bv.bv_page);
- size = min_t(ssize_t, len, length - bio_off);
- if (to_buf)
- memcpy(buf + buf_off, addr + bio_off + bv.bv_offset,
- size);
- else
- memcpy(addr + bio_off + bv.bv_offset, buf + buf_off,
- size);
- kunmap_atomic(addr);
- bio_off = 0;
- buf_off += size;
- len -= size;
- }
-}
-
static int dm_icomp_mod_to_max_io_range(struct dm_icomp_info *info,
struct dm_icomp_io_range *io)
{
@@ -1030,13 +1015,26 @@ static int dm_icomp_mod_to_max_io_range(struct dm_icomp_info *info,
static struct dm_icomp_io_range *dm_icomp_create_io_write_range(
struct dm_icomp_req *req)
{
- struct dm_icomp_io_range *io;
struct bio *bio = req->bio;
sector_t size = bio_sectors(req->bio)<<9;
int segments = bio_segments(bio);
int comp_len = dm_icomp_compressor_len(req->info, size);
void *addr;
+ struct dm_icomp_io_range *io = dm_icomp_create_io_range(req);
+
+ if (!io)
+ return NULL;
+
+ addr = dm_icomp_kmalloc(comp_len, GFP_NOIO);
+ if (!addr) {
+ kmem_cache_free(dm_icomp_io_range_cachep, io);
+ return NULL;
+ }
+ io->comp_data = io->io_req.mem.ptr.addr = addr;
+ io->comp_len = comp_len;
+ io->comp_kmap = false;

+ /* try reusing the bio buffer for decomp data. */
if (segments == 1) {
struct bio_vec bv = bio_iovec(bio);

@@ -1044,21 +1042,18 @@ static struct dm_icomp_io_range *dm_icomp_create_io_write_range(
} else
addr = dm_icomp_kmalloc(size, GFP_NOIO);

- if (!addr)
- return NULL;
-
- io = dm_icomp_create_io_range(req, comp_len);
- if (!io) {
- (segments == 1) ? kunmap(addr) : dm_icomp_kfree(addr, size);
+ if (!addr) {
+ dm_icomp_kfree(io->comp_data, comp_len);
+ kmem_cache_free(dm_icomp_io_range_cachep, io);
return NULL;
}
-
io->decomp_data = io->decomp_real_data = addr;
io->decomp_len = size;

io->decomp_kmap = (segments == 1);
if (!io->decomp_kmap)
dm_icomp_bio_copy(req->bio, 0, io->decomp_data, size, true);
+
return io;
}

@@ -1176,7 +1171,7 @@ static void dm_icomp_handle_read_decomp(struct dm_icomp_req *req)
src_off = (req->bio->bi_iter.bi_sector -
io->io_region.sector) << 9;

- if (dm_icomp_update_io_read_range(io, req->bio, dst_off)) {
+ if (dm_icomp_update_io_read_range(io)) {
req->result = -EIO;
return;
}
diff --git a/drivers/md/dm-inplace-compress.h b/drivers/md/dm-inplace-compress.h
index e144e96..775ccbf 100644
--- a/drivers/md/dm-inplace-compress.h
+++ b/drivers/md/dm-inplace-compress.h
@@ -120,6 +120,7 @@ struct dm_icomp_io_range {
void *decomp_real_data; /* holds the actual start of the buffer */
unsigned int decomp_req_len;/* originally requested length */
unsigned int decomp_len; /* actual allocated/mapped length */
+ int comp_kmap; /* Is the comp_data kmapped'? */
void *comp_data;
unsigned int comp_len; /* For write, this is estimated */
struct list_head next;
--
1.7.1

Next message: Ram Pai: "[RFC PATCH 12/16] DM: release unneeded buffer as soon as possible."
Previous message: Ram Pai: "[RFC PATCH 09/16] DM: Delay allocation of decompression buffer during read."
In reply to: Ram Pai: "[RFC PATCH 09/16] DM: Delay allocation of decompression buffer during read."
Next in thread: Ram Pai: "[RFC PATCH 12/16] DM: release unneeded buffer as soon as possible."
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]