Re: [RFC,4/5] squashfs: support multiple decompress stream buffer

From: Minchan Kim
Date: Mon Oct 07 2013 - 02:34:17 EST

On Mon, Oct 07, 2013 at 03:09:51PM +0900, Minchan Kim wrote:
> Hello Phillip,
>
> On Mon, Oct 07, 2013 at 04:41:20AM +0100, Phillip Lougher wrote:
> > Hi,
> >
> > This a partial review, based on the stuff I've managed to review
> > so far!
> >
> > 1. This is a substantial performance improvement, which is great
> > stuff!
>
> Thanks.
>
> >
> > But like the "squashfs: remove cache for normal data page" patch
> > it needs to be optional, with the previous behaviour retained as
> > default. Again, without wanting to sound like a broken (vinyl)
>
> Just FYI, I have a plan to drop "squashfs: remove cache for normal
> data page" in next submit as you pointed out it could make regression.
> So my plan is that squashfs_readpage uses the cache but squashfs_readpages
> will not use the cache.
> If you have any concern in my design, please tell me.
>
> > record, this is because as maintainer I get to worry about breaking
> > things for existing users of Squashfs when they upgrade their kernel.
> >
> > I know from consulting experience, many users of Squashfs are "on the
> > edge" of memory and CPU performance, and are using Squashfs to squeeze
> > a bit more performance out of a maxed out system.
> >
> > In these cases, changing Squashfs so it uses more memory and more
> > CPU than previously (and in this patch a lot more memory and CPU as
> > it will try and kick off multiple decompressors per core) is a bit
> > like robbing Peter to pay Paul, Squashfs may take CPU and memory
> > that are needed elsewhere, and used to be available.
> >
> > So, basically, users need to be able to explicitly select this.
>
> Okay.
>
> >
> > 2. The patch breaks the decompressor interface. Compressor option
> > parsing is implemented in the decompressor init() function, which
> > means everytime a new decompressor is dynamically instantiated, we
> > need to read and parse the compression options again and again. This
> > is an unnecessary performance degradation.
> >
> > Compressor option parsing and reading should be split out of init()
> > and into a separate function.
>
> Indeed.
>
> >
> > Compression option parsing and reading is quite obscure, it is a
> > late addition to the filesystem format, and had to be squeezed into
> > the existing format. This means it can be difficult to get it right
> > as the specification exists only in my head.
>
> Hmm, I had a question. Please look at below.
>
> >
> > I'll help you here.
> >
> > Specific comments follow in the patch.
> >
> > Phillip
> >
> >
> >
> > >Now squashfs have used for only one stream buffer for decompression
> > >so it hurts concurrent read performance due to locking lock of getting
> > >stream buffer.
> > >
> > >When file system mount, the number of stream buffer is started from
> > >num_online_cpus() and grows up to num_online_cpus() * 2M / block_size * 2.
> > >The rationale is MM does readahead chunk into 2M unit to prevent too much
> > >memory pin and while one request is waitting, we should request another
> > >chunk. That's why I multiply by 2.
> > >
> > >If it reveals too much memory problem, we can add shrinker routine.
> > >
> > >I did test following as
> > >
> > >Two 1G file dd read
> > >
> > >dd if=test/test1.dat of=/dev/null &
> > >dd if=test/test2.dat of=/dev/null &
> > >
> > >old : 60sec -> new : 30 sec
> > >
> > >Signed-off-by: Minchan Kim <minchan@xxxxxxxxxx>
> > >
> > >---
> > >fs/squashfs/block.c | 9 ++--
> > >fs/squashfs/decompressor.c | 105 ++++++++++++++++++++++++++++++++++++++----
> > >fs/squashfs/decompressor.h | 27 +++++------
> > >fs/squashfs/lzo_wrapper.c | 12 ++---
> > >fs/squashfs/squashfs.h | 3 +-
> > >fs/squashfs/squashfs_fs_sb.h | 7 ++-
> > >fs/squashfs/super.c | 40 ++++++++++++----
> > >fs/squashfs/xz_wrapper.c | 20 ++++----
> > >fs/squashfs/zlib_wrapper.c | 12 ++---
> > >9 files changed, 168 insertions(+), 67 deletions(-)
> > >
> > >diff --git a/fs/squashfs/block.c b/fs/squashfs/block.c
> > >index f33c6ef..d41bac8 100644
> > >--- a/fs/squashfs/block.c
> > >+++ b/fs/squashfs/block.c
> > >@@ -78,14 +78,14 @@ static struct buffer_head *get_block_length(struct super_block *sb,
> > >
> > >
> > >
> > >-int squashfs_decompress_block(struct squashfs_sb_info *msblk, int compressed,
> > >+int squashfs_decompress_block(struct super_block *sb, int compressed,
> > > void **buffer, struct buffer_head **bh, int nr_bh,
> > > int offset, int length, int srclength, int pages)
> > >{
> > > int k = 0;
> > >
> > > if (compressed) {
> > >- length = squashfs_decompress(msblk, buffer, bh, nr_bh,
> > >+ length = squashfs_decompress(sb, buffer, bh, nr_bh,
> > > offset, length, srclength, pages);
> > > if (length < 0)
> > > goto out;
> > >@@ -93,6 +93,7 @@ int squashfs_decompress_block(struct squashfs_sb_info *msblk, int compressed,
> > > /*
> > > * Block is uncompressed.
> > > */
> > >+ struct squashfs_sb_info *msblk = sb->s_fs_info;
> > > int bytes, in, avail, pg_offset = 0, page = 0;
> > >
> > > for (bytes = length; k < nr_bh; k++) {
> > >@@ -262,8 +263,8 @@ int squashfs_read_metablock(struct super_block *sb, void **buffer, u64 index,
> > > }
> > > ll_rw_block(READ, b - 1, bh + 1);
> > >
> > >- length = squashfs_decompress_block(msblk, compressed, buffer, bh, b,
> > >- offset, length, srclength, pages);
> > >+ length = squashfs_decompress_block(sb, compressed, buffer, bh,
> > >+ b, offset, length, srclength, pages);
> > > if (length < 0)
> > > goto read_failure;
> > >
> > >diff --git a/fs/squashfs/decompressor.c b/fs/squashfs/decompressor.c
> > >index e47453e..ed35b32 100644
> > >--- a/fs/squashfs/decompressor.c
> > >+++ b/fs/squashfs/decompressor.c
> > >@@ -25,6 +25,8 @@
> > >#include <linux/mutex.h>
> > >#include <linux/slab.h>
> > >#include <linux/buffer_head.h>
> > >+#include <linux/sched.h>
> > >+#include <linux/wait.h>
> > >
> > >#include "squashfs_fs.h"
> > >#include "squashfs_fs_sb.h"
> > >@@ -70,6 +72,80 @@ static const struct squashfs_decompressor *decompressor[] = {
> > > &squashfs_unknown_comp_ops
> > >};
> > >
> > >+void squashfs_decompressor_free(struct squashfs_sb_info *msblk,
> > >+ struct squashfs_decomp_strm *stream)
> > >+{
> > >+ if (msblk->decompressor)
> > >+ msblk->decompressor->free(stream->strm);
> > >+ kfree(stream);
> > >+}
> > >+
> > >+static void *squashfs_get_decomp_strm(struct squashfs_sb_info *msblk)
> > >+{
> > >+ struct squashfs_decomp_strm *strm = NULL;
> > >+ mutex_lock(&msblk->comp_strm_mutex);
> > >+ if (!list_empty(&msblk->strm_list)) {
> > >+ strm = list_entry(msblk->strm_list.next,
> > >+ struct squashfs_decomp_strm, list);
> > >+ list_del(&strm->list);
> > >+ msblk->nr_avail_decomp--;
> > >+ WARN_ON(msblk->nr_avail_decomp < 0);
> > >+ }
> > >+ mutex_unlock(&msblk->comp_strm_mutex);
> > >+ return strm;
> > >+}
> > >+
> > >+static bool full_decomp_strm(struct squashfs_sb_info *msblk)
> > >+{
> > >+ /* MM do readahread 2M unit */
> > >+ int blocks = 2 * 1024 * 1024 / msblk->block_size;
> > >+ return msblk->nr_avail_decomp > (num_online_cpus() * blocks * 2);
> > >+}
> > >+
> > >+static void squashfs_put_decomp_strm(struct squashfs_sb_info *msblk,
> > >+ struct squashfs_decomp_strm *strm)
> > >+{
> > >+ mutex_lock(&msblk->comp_strm_mutex);
> > >+ if (full_decomp_strm(msblk)) {
> > >+ mutex_unlock(&msblk->comp_strm_mutex);
> > >+ squashfs_decompressor_free(msblk, strm);
> > >+ return;
> > >+ }
> > >+
> > >+ list_add(&strm->list, &msblk->strm_list);
> > >+ msblk->nr_avail_decomp++;
> > >+ mutex_unlock(&msblk->comp_strm_mutex);
> > >+ wake_up(&msblk->decomp_wait_queue);
> > >+}
> > >+
> > >+int squashfs_decompress(struct super_block *sb, void **buffer,
> > >+ struct buffer_head **bh, int b, int offset, int length,
> > >+ int srclength, int pages)
> > >+{
> > >+ int ret;
> > >+ struct squashfs_decomp_strm *strm;
> > >+ struct squashfs_sb_info *msblk = sb->s_fs_info;
> > >+ while (1) {
> > >+ strm = squashfs_get_decomp_strm(msblk);
> > >+ if (strm)
> > >+ break;
> > >+
> > >+ if (!full_decomp_strm(msblk)) {
> > >+ strm = squashfs_decompressor_init(sb);
> >
> > here you call squashfs_decompressor_dynamically to instantiate a new
> > decompressor, this needs to read and parse the compression options again.
> >
> > >+ if (strm)
> > >+ break;
> > >+ }
> > >+
> > >+ wait_event(msblk->decomp_wait_queue, msblk->nr_avail_decomp);
> > >+ continue;
> > >+ }
> > >+
> > >+ ret = msblk->decompressor->decompress(msblk, strm->strm, buffer, bh,
> > >+ b, offset, length, srclength, pages);
> > >+
> > >+ squashfs_put_decomp_strm(msblk, strm);
> > >+ return ret;
> > >+}
> > >
> > >const struct squashfs_decompressor *squashfs_lookup_decompressor(int id)
> > >{
> > >@@ -82,35 +158,48 @@ const struct squashfs_decompressor *squashfs_lookup_decompressor(int id)
> > > return decompressor[i];
> > >}
> > >
> > >-
> > >-void *squashfs_decompressor_init(struct super_block *sb, unsigned short flags)
> > >+struct squashfs_decomp_strm *squashfs_decompressor_init(struct super_block *sb)
> > >{
> > > struct squashfs_sb_info *msblk = sb->s_fs_info;
> > >+ struct squashfs_decomp_strm *decomp_strm = NULL;
> > > void *strm, *buffer = NULL;
> > > int length = 0;
> > >
> > >+ decomp_strm = kmalloc(sizeof(struct squashfs_decomp_strm), GFP_KERNEL);
> > >+ if (!decomp_strm)
> > >+ return ERR_PTR(-ENOMEM);
> > > /*
> > > * Read decompressor specific options from file system if present
> > > */
> > >- if (SQUASHFS_COMP_OPTS(flags)) {
> > >+ if (SQUASHFS_COMP_OPTS(msblk->flags)) {
> > > buffer = kmalloc(PAGE_CACHE_SIZE, GFP_KERNEL);
> > >- if (buffer == NULL)
> > >- return ERR_PTR(-ENOMEM);
> > >+ if (buffer == NULL) {
> > >+ decomp_strm = ERR_PTR(-ENOMEM);
> > >+ goto finished;
> > >+ }
> > >
> > > length = squashfs_read_metablock(sb, &buffer,
> > > sizeof(struct squashfs_super_block), 0, NULL,
> > > PAGE_CACHE_SIZE, 1);
> > >
> >
> > This reads the compressor options each decompressor init(). This should
> > only be done once at mount time.
> >
> > > if (length < 0) {
> > >- strm = ERR_PTR(length);
> > >+ decomp_strm = ERR_PTR(length);
> > > goto finished;
> > > }
> > > }
> > >
> >
> >
> > > strm = msblk->decompressor->init(msblk, buffer, length);
> > >+ if (IS_ERR(strm)) {
> > >+ decomp_strm = strm;
> > >+ goto finished;
> > >+ }
> > >
> > >-finished:
> > >+ decomp_strm->strm = strm;
> > > kfree(buffer);
> > >+ return decomp_strm;
> > >
> > >- return strm;
> > >+finished:
> > >+ kfree(decomp_strm);
> > >+ kfree(buffer);
> > >+ return decomp_strm;
> > >}
> > >diff --git a/fs/squashfs/decompressor.h b/fs/squashfs/decompressor.h
> > >index 330073e..207c810 100644
> > >--- a/fs/squashfs/decompressor.h
> > >+++ b/fs/squashfs/decompressor.h
> > >@@ -26,27 +26,24 @@
> > >struct squashfs_decompressor {
> > > void *(*init)(struct squashfs_sb_info *, void *, int);
> > > void (*free)(void *);
> > >- int (*decompress)(struct squashfs_sb_info *, void **,
> > >+ int (*decompress)(struct squashfs_sb_info *, void*, void **,
> > > struct buffer_head **, int, int, int, int, int);
> > > int id;
> > > char *name;
> > > int supported;
> > >};
> > >
> > >-static inline void squashfs_decompressor_free(struct squashfs_sb_info *msblk,
> > >- void *s)
> > >-{
> > >- if (msblk->decompressor)
> > >- msblk->decompressor->free(s);
> > >-}
> > >-
> > >-static inline int squashfs_decompress(struct squashfs_sb_info *msblk,
> > >- void **buffer, struct buffer_head **bh, int b, int offset, int length,
> > >- int srclength, int pages)
> > >-{
> > >- return msblk->decompressor->decompress(msblk, buffer, bh, b, offset,
> > >- length, srclength, pages);
> > >-}
> > >+struct squashfs_decomp_strm {
> > >+ void *strm;
> > >+ struct list_head list;
> > >+};
> > >+
> > >+int squashfs_decompress(struct super_block *sb, void **buffer,
> > >+ struct buffer_head **bh, int b, int offset, int length,
> > >+ int srclength, int pages);
> > >+
> > >+void squashfs_decompressor_free(struct squashfs_sb_info *msblk,
> > >+ struct squashfs_decomp_strm *stream);
> > >
> > >#ifdef CONFIG_SQUASHFS_XZ
> > >extern const struct squashfs_decompressor squashfs_xz_comp_ops;
> > >diff --git a/fs/squashfs/lzo_wrapper.c b/fs/squashfs/lzo_wrapper.c
> > >index 00f4dfc..e1bf135 100644
> > >--- a/fs/squashfs/lzo_wrapper.c
> > >+++ b/fs/squashfs/lzo_wrapper.c
> > >@@ -74,17 +74,15 @@ static void lzo_free(void *strm)
> > >}
> > >
> > >
> > >-static int lzo_uncompress(struct squashfs_sb_info *msblk, void **buffer,
> > >- struct buffer_head **bh, int b, int offset, int length, int srclength,
> > >- int pages)
> > >+static int lzo_uncompress(struct squashfs_sb_info *msblk, void *strm,
> > >+ void **buffer, struct buffer_head **bh, int b, int offset,
> > >+ int length, int srclength, int pages)
> > >{
> > >- struct squashfs_lzo *stream = msblk->stream;
> > >+ struct squashfs_lzo *stream = strm;
> > > void *buff = stream->input;
> > > int avail, i, bytes = length, res;
> > > size_t out_len = srclength;
> > >
> > >- mutex_lock(&msblk->read_data_mutex);
> > >-
> > > for (i = 0; i < b; i++) {
> > > wait_on_buffer(bh[i]);
> > > if (!buffer_uptodate(bh[i]))
> > >@@ -111,7 +109,6 @@ static int lzo_uncompress(struct squashfs_sb_info *msblk, void **buffer,
> > > bytes -= avail;
> > > }
> > >
> > >- mutex_unlock(&msblk->read_data_mutex);
> > > return res;
> > >
> > >block_release:
> > >@@ -119,7 +116,6 @@ block_release:
> > > put_bh(bh[i]);
> > >
> > >failed:
> > >- mutex_unlock(&msblk->read_data_mutex);
> > >
> > > ERROR("lzo decompression failed, data probably corrupt\n");
> > > return -EIO;
> > >diff --git a/fs/squashfs/squashfs.h b/fs/squashfs/squashfs.h
> > >index 405baed..4bb1f90 100644
> > >--- a/fs/squashfs/squashfs.h
> > >+++ b/fs/squashfs/squashfs.h
> > >@@ -52,7 +52,8 @@ extern void *squashfs_read_table(struct super_block *, u64, int);
> > >
> > >/* decompressor.c */
> > >extern const struct squashfs_decompressor *squashfs_lookup_decompressor(int);
> > >-extern void *squashfs_decompressor_init(struct super_block *, unsigned short);
> > >+extern struct squashfs_decomp_strm *squashfs_decompressor_init(
> > >+ struct super_block *);
> > >
> > >/* export.c */
> > >extern __le64 *squashfs_read_inode_lookup_table(struct super_block *, u64, u64,
> > >diff --git a/fs/squashfs/squashfs_fs_sb.h b/fs/squashfs/squashfs_fs_sb.h
> > >index 52934a2..0a209e6 100644
> > >--- a/fs/squashfs/squashfs_fs_sb.h
> > >+++ b/fs/squashfs/squashfs_fs_sb.h
> > >@@ -63,10 +63,10 @@ struct squashfs_sb_info {
> > > __le64 *id_table;
> > > __le64 *fragment_index;
> > > __le64 *xattr_id_table;
> > >- struct mutex read_data_mutex;
> > >+ struct mutex comp_strm_mutex;
> > > struct mutex meta_index_mutex;
> > > struct meta_index *meta_index;
> > >- void *stream;
> > >+ struct list_head strm_list;
> > > __le64 *inode_lookup_table;
> > > u64 inode_table;
> > > u64 directory_table;
> > >@@ -76,5 +76,8 @@ struct squashfs_sb_info {
> > > long long bytes_used;
> > > unsigned int inodes;
> > > int xattr_ids;
> > >+ wait_queue_head_t decomp_wait_queue;
> > >+ int nr_avail_decomp;
> > >+ unsigned short flags;
> > >};
> > >#endif
> > >diff --git a/fs/squashfs/super.c b/fs/squashfs/super.c
> > >index 60553a9..70aa9c4 100644
> > >--- a/fs/squashfs/super.c
> > >+++ b/fs/squashfs/super.c
> > >@@ -84,7 +84,7 @@ static int squashfs_fill_super(struct super_block *sb, void *data, int silent)
> > > unsigned short flags;
> > > unsigned int fragments;
> > > u64 lookup_table_start, xattr_id_table_start, next_table;
> > >- int err;
> > >+ int err, i;
> > >
> > > TRACE("Entered squashfs_fill_superblock\n");
> > >
> > >@@ -98,7 +98,9 @@ static int squashfs_fill_super(struct super_block *sb, void *data, int silent)
> > > msblk->devblksize = sb_min_blocksize(sb, SQUASHFS_DEVBLK_SIZE);
> > > msblk->devblksize_log2 = ffz(~msblk->devblksize);
> > >
> > >- mutex_init(&msblk->read_data_mutex);
> > >+ INIT_LIST_HEAD(&msblk->strm_list);
> > >+ mutex_init(&msblk->comp_strm_mutex);
> > >+ init_waitqueue_head(&msblk->decomp_wait_queue);
> >
> > This should be done via a helper in decompressor.c, i.e. the implementation
> > shouldn't be visible here.
> >
> > When I added the decompressor framework, I should have done this.
> >
> > > mutex_init(&msblk->meta_index_mutex);
> > >
> > > /*
> > >@@ -176,6 +178,7 @@ static int squashfs_fill_super(struct super_block *sb, void *data, int silent)
> > > msblk->directory_table = le64_to_cpu(sblk->directory_table_start);
> > > msblk->inodes = le32_to_cpu(sblk->inodes);
> > > flags = le16_to_cpu(sblk->flags);
> > >+ msblk->flags = flags;
> > >
> >
> > ha, the superblock flags should only be needed at mount time. After
> > mount time there shouldn't be anything in flags we need to look at.
> >
> > You need to do this because flags is needed for the decompressor init
> > function (COMP_OPTS(flags)) which is now called after mount time.
> >
> > Once the compressor options parsing is moved back to being only
> > at mount time, you won't need to store the flags anymore.
>
> Hmm, I'd like to clarify your point further.
> I agree it's unnecessary performance degradation if we read compressor
> option from storage whenever we create new stream buffer.
> But we need it to create new stream buffer(ex, xz_dec_init) dynamically
> so we should keep compressor option into somewhere. It means we should
> keep it to somewhere although we remove flags field from msblk.
> Am I missing something?

I should have been more specific.
My concern is that we could remove the flags from msblk but we need comp_opts
to create new comp buffer dynamically so we should keep comp_opts into somewhere.
So, I think best place to keep it is struct squashfs_sb_info like this.

struct squashfs_sb_info {
const struct squashfs_decompressor *decompressor;
void *comp_opts;
...
};

What do you think about it?
--
Kind regards,
Minchan Kim
--
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/