[alsa-devel] [PATCH v2 0/3] ASoC: Introduce the new caching API
This patch series introduces the new caching API. The idea behind this caching interface is that we can provide different means of organizing and accessing the register cache. This is useful for large and sparse register maps, where one can use some kind of compression algorithm to reduce the memory footprint. The caching API is designed in such way to eliminate the need for modifying any existing drivers.
TODO: - Memory usage statistics, to make it easier to select the proper caching technique. - Support for bulk reads/writes. - Overridable caching behaviour from the machine driver. - Optimize certain CPU intensive paths. - Ensure proper locking in init()/exit()/sync().
Dimitris Papastamos (3): ASoC: soc-cache: Add support for flat register caching ASoC: soc-cache: Add support for LZO register caching ASoC: soc-cache: Add support for rbtree based register caching
include/sound/soc.h | 29 ++ sound/soc/Kconfig | 2 + sound/soc/soc-cache.c | 984 +++++++++++++++++++++++++++++++++++++++++++++++-- sound/soc/soc-core.c | 37 +-- 4 files changed, 998 insertions(+), 54 deletions(-)
This patch introduces the new caching API and migrates the old caching interface into the new one. The flat register caching technique does not use compression at all and it is equivalent to the old caching technique. One can still access codec->reg_cache directly but this is not advised as that will not be portable across different caching strategies.
None of the existing drivers need to be changed to adapt to this caching technique. There should be no noticeable overhead associated with using the new caching API.
Signed-off-by: Dimitris Papastamos dp@opensource.wolfsonmicro.com --- include/sound/soc.h | 27 ++++ sound/soc/soc-cache.c | 339 ++++++++++++++++++++++++++++++++++++++++++++----- sound/soc/soc-core.c | 37 ++---- 3 files changed, 349 insertions(+), 54 deletions(-)
diff --git a/include/sound/soc.h b/include/sound/soc.h index 5d39547..275e411 100644 --- a/include/sound/soc.h +++ b/include/sound/soc.h @@ -238,6 +238,7 @@ struct soc_enum; struct snd_soc_ac97_ops; struct snd_soc_jack; struct snd_soc_jack_pin; +struct snd_soc_cache_ops; #include <sound/soc-dapm.h>
#ifdef CONFIG_GPIOLIB @@ -254,6 +255,10 @@ enum snd_soc_control_type { SND_SOC_SPI, };
+enum snd_soc_compress_type { + SND_SOC_NO_COMPRESSION +}; + int snd_soc_register_platform(struct device *dev, struct snd_soc_platform_driver *platform_drv); void snd_soc_unregister_platform(struct device *dev); @@ -265,6 +270,13 @@ int snd_soc_codec_volatile_register(struct snd_soc_codec *codec, int reg); int snd_soc_codec_set_cache_io(struct snd_soc_codec *codec, int addr_bits, int data_bits, enum snd_soc_control_type control); +int snd_soc_cache_sync(struct snd_soc_codec *codec); +int snd_soc_cache_init(struct snd_soc_codec *codec); +int snd_soc_cache_exit(struct snd_soc_codec *codec); +int snd_soc_cache_write(struct snd_soc_codec *codec, + unsigned int reg, unsigned int value); +int snd_soc_cache_read(struct snd_soc_codec *codec, + unsigned int reg, unsigned int *value);
/* Utility functions to get clock rates from various things */ int snd_soc_calc_frame_size(int sample_size, int channels, int tdm_slots); @@ -421,6 +433,18 @@ struct snd_soc_ops { int (*trigger)(struct snd_pcm_substream *, int); };
+/* SoC cache ops */ +struct snd_soc_cache_ops { + enum snd_soc_compress_type id; + int (*init)(struct snd_soc_codec *codec); + int (*exit)(struct snd_soc_codec *codec); + int (*read)(struct snd_soc_codec *codec, unsigned int reg, + unsigned int *value); + int (*write)(struct snd_soc_codec *codec, unsigned int reg, + unsigned int value); + int (*sync)(struct snd_soc_codec *codec); +}; + /* SoC Audio Codec device */ struct snd_soc_codec { const char *name; @@ -450,6 +474,8 @@ struct snd_soc_codec { hw_write_t hw_write; unsigned int (*hw_read)(struct snd_soc_codec *, unsigned int); void *reg_cache; + const struct snd_soc_cache_ops *cache_ops; + struct mutex cache_rw_mutex;
/* dapm */ struct snd_soc_dapm_context dapm; @@ -482,6 +508,7 @@ struct snd_soc_codec_driver { short reg_cache_step; short reg_word_size; const void *reg_cache_default; + enum snd_soc_compress_type compress_type;
/* codec bias level */ int (*set_bias_level)(struct snd_soc_codec *, diff --git a/sound/soc/soc-cache.c b/sound/soc/soc-cache.c index e747998..cbf9694 100644 --- a/sound/soc/soc-cache.c +++ b/sound/soc/soc-cache.c @@ -18,7 +18,8 @@ static unsigned int snd_soc_4_12_read(struct snd_soc_codec *codec, unsigned int reg) { - u16 *cache = codec->reg_cache; + int ret; + unsigned int val;
if (reg >= codec->driver->reg_cache_size || snd_soc_codec_volatile_register(codec, reg)) { @@ -29,13 +30,15 @@ static unsigned int snd_soc_4_12_read(struct snd_soc_codec *codec, return codec->hw_read(codec, reg); }
- return cache[reg]; + ret = snd_soc_cache_read(codec, reg, &val); + if (ret < 0) + return -1; + return val; }
static int snd_soc_4_12_write(struct snd_soc_codec *codec, unsigned int reg, unsigned int value) { - u16 *cache = codec->reg_cache; u8 data[2]; int ret;
@@ -43,8 +46,11 @@ static int snd_soc_4_12_write(struct snd_soc_codec *codec, unsigned int reg, data[1] = value & 0x00ff;
if (!snd_soc_codec_volatile_register(codec, reg) && - reg < codec->driver->reg_cache_size) - cache[reg] = value; + reg < codec->driver->reg_cache_size) { + ret = snd_soc_cache_write(codec, reg, value); + if (ret < 0) + return -1; + }
if (codec->cache_only) { codec->cache_sync = 1; @@ -93,7 +99,8 @@ static int snd_soc_4_12_spi_write(void *control_data, const char *data, static unsigned int snd_soc_7_9_read(struct snd_soc_codec *codec, unsigned int reg) { - u16 *cache = codec->reg_cache; + int ret; + unsigned int val;
if (reg >= codec->driver->reg_cache_size || snd_soc_codec_volatile_register(codec, reg)) { @@ -104,13 +111,15 @@ static unsigned int snd_soc_7_9_read(struct snd_soc_codec *codec, return codec->hw_read(codec, reg); }
- return cache[reg]; + ret = snd_soc_cache_read(codec, reg, &val); + if (ret < 0) + return -1; + return val; }
static int snd_soc_7_9_write(struct snd_soc_codec *codec, unsigned int reg, unsigned int value) { - u16 *cache = codec->reg_cache; u8 data[2]; int ret;
@@ -118,8 +127,11 @@ static int snd_soc_7_9_write(struct snd_soc_codec *codec, unsigned int reg, data[1] = value & 0x00ff;
if (!snd_soc_codec_volatile_register(codec, reg) && - reg < codec->driver->reg_cache_size) - cache[reg] = value; + reg < codec->driver->reg_cache_size) { + ret = snd_soc_cache_write(codec, reg, value); + if (ret < 0) + return -1; + }
if (codec->cache_only) { codec->cache_sync = 1; @@ -168,16 +180,19 @@ static int snd_soc_7_9_spi_write(void *control_data, const char *data, static int snd_soc_8_8_write(struct snd_soc_codec *codec, unsigned int reg, unsigned int value) { - u8 *cache = codec->reg_cache; u8 data[2]; + int ret;
reg &= 0xff; data[0] = reg; data[1] = value & 0xff;
if (!snd_soc_codec_volatile_register(codec, reg) && - reg < codec->driver->reg_cache_size) - cache[reg] = value; + reg < codec->driver->reg_cache_size) { + ret = snd_soc_cache_write(codec, reg, value); + if (ret < 0) + return -1; + }
if (codec->cache_only) { codec->cache_sync = 1; @@ -193,7 +208,8 @@ static int snd_soc_8_8_write(struct snd_soc_codec *codec, unsigned int reg, static unsigned int snd_soc_8_8_read(struct snd_soc_codec *codec, unsigned int reg) { - u8 *cache = codec->reg_cache; + int ret; + unsigned int val;
reg &= 0xff; if (reg >= codec->driver->reg_cache_size || @@ -205,7 +221,10 @@ static unsigned int snd_soc_8_8_read(struct snd_soc_codec *codec, return codec->hw_read(codec, reg); }
- return cache[reg]; + ret = snd_soc_cache_read(codec, reg, &val); + if (ret < 0) + return -1; + return val; }
#if defined(CONFIG_SPI_MASTER) @@ -241,16 +260,19 @@ static int snd_soc_8_8_spi_write(void *control_data, const char *data, static int snd_soc_8_16_write(struct snd_soc_codec *codec, unsigned int reg, unsigned int value) { - u16 *reg_cache = codec->reg_cache; u8 data[3]; + int ret;
data[0] = reg; data[1] = (value >> 8) & 0xff; data[2] = value & 0xff;
if (!snd_soc_codec_volatile_register(codec, reg) && - reg < codec->driver->reg_cache_size) - reg_cache[reg] = value; + reg < codec->driver->reg_cache_size) { + ret = snd_soc_cache_write(codec, reg, value); + if (ret < 0) + return -1; + }
if (codec->cache_only) { codec->cache_sync = 1; @@ -266,7 +288,8 @@ static int snd_soc_8_16_write(struct snd_soc_codec *codec, unsigned int reg, static unsigned int snd_soc_8_16_read(struct snd_soc_codec *codec, unsigned int reg) { - u16 *cache = codec->reg_cache; + int ret; + unsigned int val;
if (reg >= codec->driver->reg_cache_size || snd_soc_codec_volatile_register(codec, reg)) { @@ -275,9 +298,12 @@ static unsigned int snd_soc_8_16_read(struct snd_soc_codec *codec,
BUG_ON(!codec->hw_read); return codec->hw_read(codec, reg); - } else { - return cache[reg]; } + + ret = snd_soc_cache_read(codec, reg, &val); + if (ret < 0) + return -1; + return val; }
#if defined(CONFIG_SPI_MASTER) @@ -416,7 +442,8 @@ static unsigned int snd_soc_16_8_read_i2c(struct snd_soc_codec *codec, static unsigned int snd_soc_16_8_read(struct snd_soc_codec *codec, unsigned int reg) { - u8 *cache = codec->reg_cache; + int ret; + unsigned int val;
reg &= 0xff; if (reg >= codec->driver->reg_cache_size || @@ -428,13 +455,15 @@ static unsigned int snd_soc_16_8_read(struct snd_soc_codec *codec, return codec->hw_read(codec, reg); }
- return cache[reg]; + ret = snd_soc_cache_read(codec, reg, &val); + if (ret < 0) + return -1; + return val; }
static int snd_soc_16_8_write(struct snd_soc_codec *codec, unsigned int reg, unsigned int value) { - u8 *cache = codec->reg_cache; u8 data[3]; int ret;
@@ -444,8 +473,11 @@ static int snd_soc_16_8_write(struct snd_soc_codec *codec, unsigned int reg,
reg &= 0xff; if (!snd_soc_codec_volatile_register(codec, reg) && - reg < codec->driver->reg_cache_size) - cache[reg] = value; + reg < codec->driver->reg_cache_size) { + ret = snd_soc_cache_write(codec, reg, value); + if (ret < 0) + return -1; + }
if (codec->cache_only) { codec->cache_sync = 1; @@ -529,7 +561,8 @@ static unsigned int snd_soc_16_16_read_i2c(struct snd_soc_codec *codec, static unsigned int snd_soc_16_16_read(struct snd_soc_codec *codec, unsigned int reg) { - u16 *cache = codec->reg_cache; + int ret; + unsigned int val;
if (reg >= codec->driver->reg_cache_size || snd_soc_codec_volatile_register(codec, reg)) { @@ -540,13 +573,16 @@ static unsigned int snd_soc_16_16_read(struct snd_soc_codec *codec, return codec->hw_read(codec, reg); }
- return cache[reg]; + ret = snd_soc_cache_read(codec, reg, &val); + if (ret < 0) + return -1; + + return val; }
static int snd_soc_16_16_write(struct snd_soc_codec *codec, unsigned int reg, unsigned int value) { - u16 *cache = codec->reg_cache; u8 data[4]; int ret;
@@ -556,8 +592,11 @@ static int snd_soc_16_16_write(struct snd_soc_codec *codec, unsigned int reg, data[3] = value & 0xff;
if (!snd_soc_codec_volatile_register(codec, reg) && - reg < codec->driver->reg_cache_size) - cache[reg] = value; + reg < codec->driver->reg_cache_size) { + ret = snd_soc_cache_write(codec, reg, value); + if (ret < 0) + return -1; + }
if (codec->cache_only) { codec->cache_sync = 1; @@ -718,3 +757,241 @@ int snd_soc_codec_set_cache_io(struct snd_soc_codec *codec, return 0; } EXPORT_SYMBOL_GPL(snd_soc_codec_set_cache_io); + +static int snd_soc_flat_cache_sync(struct snd_soc_codec *codec) +{ + int i; + struct snd_soc_codec_driver *codec_drv; + unsigned int val; + + codec_drv = codec->driver; + for (i = 0; i < codec_drv->reg_cache_size; ++i) { + snd_soc_cache_read(codec, i, &val); + if (codec_drv->reg_cache_default) { + switch (codec_drv->reg_word_size) { + case 1: { + const u8 *cache; + + cache = codec_drv->reg_cache_default; + if (cache[i] == val) + continue; + } + break; + case 2: { + const u16 *cache; + + cache = codec_drv->reg_cache_default; + if (cache[i] == val) + continue; + } + break; + default: + BUG(); + } + } + snd_soc_write(codec, i, val); + dev_dbg(codec->dev, "Synced register %#x, value = %#x\n", + i, val); + } + return 0; +} + +static int snd_soc_flat_cache_write(struct snd_soc_codec *codec, + unsigned int reg, unsigned int value) +{ + switch (codec->driver->reg_word_size) { + case 1: { + u8 *cache; + + cache = codec->reg_cache; + cache[reg] = value; + } + break; + case 2: { + u16 *cache; + + cache = codec->reg_cache; + cache[reg] = value; + } + break; + default: + BUG(); + } + + return 0; +} + +static int snd_soc_flat_cache_read(struct snd_soc_codec *codec, + unsigned int reg, unsigned int *value) +{ + switch (codec->driver->reg_word_size) { + case 1: { + u8 *cache; + + cache = codec->reg_cache; + *value = cache[reg]; + } + break; + case 2: { + u16 *cache; + + cache = codec->reg_cache; + *value = cache[reg]; + } + break; + default: + BUG(); + } + + return 0; +} + +static int snd_soc_flat_cache_exit(struct snd_soc_codec *codec) +{ + if (!codec->reg_cache) + return 0; + kfree(codec->reg_cache); + codec->reg_cache = NULL; + return 0; +} + +static int snd_soc_flat_cache_init(struct snd_soc_codec *codec) +{ + struct snd_soc_codec_driver *codec_drv; + size_t reg_size; + + codec_drv = codec->driver; + reg_size = codec_drv->reg_cache_size * codec_drv->reg_word_size; + + if (codec_drv->reg_cache_default) + codec->reg_cache = kmemdup(codec_drv->reg_cache_default, + reg_size, GFP_KERNEL); + else + codec->reg_cache = kzalloc(reg_size, GFP_KERNEL); + if (!codec->reg_cache) + return -ENOMEM; + + return 0; +} + +/* an array of all supported compression types */ +static const struct snd_soc_cache_ops cache_types[] = { + { + .id = SND_SOC_NO_COMPRESSION, + .init = snd_soc_flat_cache_init, + .exit = snd_soc_flat_cache_exit, + .read = snd_soc_flat_cache_read, + .write = snd_soc_flat_cache_write, + .sync = snd_soc_flat_cache_sync + } +}; + +int snd_soc_cache_init(struct snd_soc_codec *codec) +{ + int i; + + for (i = 0; i < ARRAY_SIZE(cache_types); ++i) + if (cache_types[i].id == codec->driver->compress_type) + break; + if (i == ARRAY_SIZE(cache_types)) { + dev_err(codec->dev, "Could not match compress type: %d\n", + codec->driver->compress_type); + return -EINVAL; + } + + mutex_init(&codec->cache_rw_mutex); + codec->cache_ops = &cache_types[i]; + + if (codec->cache_ops->init) + return codec->cache_ops->init(codec); + return -EINVAL; +} + +/* + * NOTE: keep in mind that this function might be called + * multiple times. + */ +int snd_soc_cache_exit(struct snd_soc_codec *codec) +{ + if (codec->cache_ops && codec->cache_ops->exit) + return codec->cache_ops->exit(codec); + return -EINVAL; +} + +/** + * snd_soc_cache_read: Fetch the value of a given register from the cache. + * + * @codec: CODEC to configure. + * @reg: The register index. + * @value: The value to be returned. + */ +int snd_soc_cache_read(struct snd_soc_codec *codec, + unsigned int reg, unsigned int *value) +{ + int ret; + + mutex_lock(&codec->cache_rw_mutex); + + if (value && codec->cache_ops && codec->cache_ops->read) { + ret = codec->cache_ops->read(codec, reg, value); + mutex_unlock(&codec->cache_rw_mutex); + return ret; + } + + mutex_unlock(&codec->cache_rw_mutex); + return -EINVAL; +} +EXPORT_SYMBOL_GPL(snd_soc_cache_read); + +/** + * snd_soc_cache_write: Set the value of a given register in the cache. + * + * @codec: CODEC to configure. + * @reg: The register index. + * @value: The new register value. + */ +int snd_soc_cache_write(struct snd_soc_codec *codec, + unsigned int reg, unsigned int value) +{ + int ret; + + mutex_lock(&codec->cache_rw_mutex); + + if (codec->cache_ops && codec->cache_ops->write) { + ret = codec->cache_ops->write(codec, reg, value); + mutex_unlock(&codec->cache_rw_mutex); + return ret; + } + + mutex_unlock(&codec->cache_rw_mutex); + return -EINVAL; +} +EXPORT_SYMBOL_GPL(snd_soc_cache_write); + +/** + * snd_soc_cache_sync: Sync the register cache with the hardware. + * + * @codec: CODEC to configure. + * + * Any registers that should not be synced should be marked as + * volatile. In general drivers can choose not to use the provided + * syncing functionality if they so require. + */ +int snd_soc_cache_sync(struct snd_soc_codec *codec) +{ + int ret; + + if (!codec->cache_sync) { + return 0; + } + + if (codec->cache_ops && codec->cache_ops->sync) { + ret = codec->cache_ops->sync(codec); + if (!ret) + codec->cache_sync = 0; + return ret; + } + + return -EINVAL; +} +EXPORT_SYMBOL_GPL(snd_soc_cache_sync); diff --git a/sound/soc/soc-core.c b/sound/soc/soc-core.c index c18ce1d..f9f5f07 100644 --- a/sound/soc/soc-core.c +++ b/sound/soc/soc-core.c @@ -3274,29 +3274,21 @@ int snd_soc_register_codec(struct device *dev, codec->dapm.bias_level = SND_SOC_BIAS_OFF; codec->dapm.dev = dev; codec->dapm.codec = codec; + codec->dev = dev; + codec->driver = codec_drv; + codec->num_dai = num_dai; + mutex_init(&codec->mutex);
/* allocate CODEC register cache */ if (codec_drv->reg_cache_size && codec_drv->reg_word_size) { - - if (codec_drv->reg_cache_default) - codec->reg_cache = kmemdup(codec_drv->reg_cache_default, - codec_drv->reg_cache_size * codec_drv->reg_word_size, GFP_KERNEL); - else - codec->reg_cache = kzalloc(codec_drv->reg_cache_size * - codec_drv->reg_word_size, GFP_KERNEL); - - if (codec->reg_cache == NULL) { - kfree(codec->name); - kfree(codec); - return -ENOMEM; + ret = snd_soc_cache_init(codec); + if (ret < 0) { + dev_err(codec->dev, "Failed to set cache compression type: %d\n", + ret); + goto error_cache; } }
- codec->dev = dev; - codec->driver = codec_drv; - codec->num_dai = num_dai; - mutex_init(&codec->mutex); - for (i = 0; i < num_dai; i++) { fixup_codec_formats(&dai_drv[i].playback); fixup_codec_formats(&dai_drv[i].capture); @@ -3306,7 +3298,7 @@ int snd_soc_register_codec(struct device *dev, if (num_dai) { ret = snd_soc_register_dais(dev, dai_drv, num_dai); if (ret < 0) - goto error; + goto error_dais; }
mutex_lock(&client_mutex); @@ -3317,9 +3309,9 @@ int snd_soc_register_codec(struct device *dev, pr_debug("Registered codec '%s'\n", codec->name); return 0;
-error: - if (codec->reg_cache) - kfree(codec->reg_cache); +error_dais: + snd_soc_cache_exit(codec); +error_cache: kfree(codec->name); kfree(codec); return ret; @@ -3353,8 +3345,7 @@ found:
pr_debug("Unregistered codec '%s'\n", codec->name);
- if (codec->reg_cache) - kfree(codec->reg_cache); + snd_soc_cache_exit(codec); kfree(codec->name); kfree(codec); }
This patch adds support for LZO compression when storing the register cache. The initial register defaults cache is marked as __devinitconst and the only change required for a driver to use LZO compression is to set the compress_type member in codec->driver to SND_SOC_LZO_COMPRESSION.
For a typical device whose register map would normally occupy 25kB or 50kB by using the LZO compression technique, one can get down to ~5-7kB. There might be a performance penalty associated with each individual read/write due to decompressing/compressing the underlying cache, however that should not be noticeable. These memory benefits depend on whether the target architecture can get rid of the memory occupied by the original register defaults cache which is marked as __devinitconst. Nevertheless there will be some memory gain even if the target architecture can't get rid of the original register map, this should be around ~30-32kB instead of 50kB.
Signed-off-by: Dimitris Papastamos dp@opensource.wolfsonmicro.com --- include/sound/soc.h | 3 +- sound/soc/Kconfig | 2 + sound/soc/soc-cache.c | 413 +++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 417 insertions(+), 1 deletions(-)
diff --git a/include/sound/soc.h b/include/sound/soc.h index 275e411..10f5932 100644 --- a/include/sound/soc.h +++ b/include/sound/soc.h @@ -256,7 +256,8 @@ enum snd_soc_control_type { };
enum snd_soc_compress_type { - SND_SOC_NO_COMPRESSION + SND_SOC_NO_COMPRESSION, + SND_SOC_LZO_COMPRESSION };
int snd_soc_register_platform(struct device *dev, diff --git a/sound/soc/Kconfig b/sound/soc/Kconfig index 3e598e7..4562c89 100644 --- a/sound/soc/Kconfig +++ b/sound/soc/Kconfig @@ -4,6 +4,8 @@
menuconfig SND_SOC tristate "ALSA for SoC audio support" + select LZO_COMPRESS + select LZO_DECOMPRESS select SND_PCM select AC97_BUS if SND_SOC_AC97_BUS select SND_JACK if INPUT=y || INPUT=SND diff --git a/sound/soc/soc-cache.c b/sound/soc/soc-cache.c index cbf9694..4581bf1 100644 --- a/sound/soc/soc-cache.c +++ b/sound/soc/soc-cache.c @@ -14,6 +14,8 @@ #include <linux/i2c.h> #include <linux/spi/spi.h> #include <sound/soc.h> +#include <linux/lzo.h> +#include <linux/bitmap.h>
static unsigned int snd_soc_4_12_read(struct snd_soc_codec *codec, unsigned int reg) @@ -758,6 +760,409 @@ int snd_soc_codec_set_cache_io(struct snd_soc_codec *codec, } EXPORT_SYMBOL_GPL(snd_soc_codec_set_cache_io);
+struct snd_soc_lzo_ctx { + void *wmem; + void *dst; + const void *src; + size_t src_len; + size_t dst_len; + size_t decompressed_size; + unsigned long *sync_bmp; + int sync_bmp_nbits; +}; + +#define LZO_BLOCK_NUM 8 +static int snd_soc_lzo_block_count(void) +{ + return LZO_BLOCK_NUM; +} + +static int snd_soc_lzo_prepare(struct snd_soc_lzo_ctx *lzo_ctx) +{ + lzo_ctx->wmem = kmalloc(LZO1X_MEM_COMPRESS, GFP_KERNEL); + if (!lzo_ctx->wmem) + return -ENOMEM; + return 0; +} + +static int snd_soc_lzo_compress(struct snd_soc_lzo_ctx *lzo_ctx) +{ + size_t compress_size; + int ret; + + ret = lzo1x_1_compress(lzo_ctx->src, lzo_ctx->src_len, + lzo_ctx->dst, &compress_size, lzo_ctx->wmem); + if (ret != LZO_E_OK || compress_size > lzo_ctx->dst_len) + return -EINVAL; + lzo_ctx->dst_len = compress_size; + return 0; +} + +static int snd_soc_lzo_decompress(struct snd_soc_lzo_ctx *lzo_ctx) +{ + size_t dst_len; + int ret; + + dst_len = lzo_ctx->dst_len; + ret = lzo1x_decompress_safe(lzo_ctx->src, lzo_ctx->src_len, + lzo_ctx->dst, &dst_len); + if (ret != LZO_E_OK || dst_len != lzo_ctx->dst_len) + return -EINVAL; + return 0; +} + +static int snd_soc_lzo_compress_cache_block(struct snd_soc_codec *codec, + struct snd_soc_lzo_ctx *lzo_ctx) +{ + int ret; + + lzo_ctx->dst_len = lzo1x_worst_compress(PAGE_SIZE); + lzo_ctx->dst = kmalloc(lzo_ctx->dst_len, GFP_KERNEL); + if (!lzo_ctx->dst) { + lzo_ctx->dst_len = 0; + return -ENOMEM; + } + + ret = snd_soc_lzo_compress(lzo_ctx); + if (ret < 0) + return ret; + return 0; +} + +static int snd_soc_lzo_decompress_cache_block(struct snd_soc_codec *codec, + struct snd_soc_lzo_ctx *lzo_ctx) +{ + int ret; + + lzo_ctx->dst_len = lzo_ctx->decompressed_size; + lzo_ctx->dst = kmalloc(lzo_ctx->dst_len, GFP_KERNEL); + if (!lzo_ctx->dst) { + lzo_ctx->dst_len = 0; + return -ENOMEM; + } + + ret = snd_soc_lzo_decompress(lzo_ctx); + if (ret < 0) + return ret; + return 0; +} + +static inline int snd_soc_lzo_get_blkindex(struct snd_soc_codec *codec, + unsigned int reg) +{ + struct snd_soc_codec_driver *codec_drv; + size_t reg_size; + + codec_drv = codec->driver; + reg_size = codec_drv->reg_cache_size * codec_drv->reg_word_size; + return (reg * codec_drv->reg_word_size) / + DIV_ROUND_UP(reg_size, snd_soc_lzo_block_count()); +} + +static inline int snd_soc_lzo_get_blkpos(struct snd_soc_codec *codec, + unsigned int reg) +{ + struct snd_soc_codec_driver *codec_drv; + size_t reg_size; + + codec_drv = codec->driver; + reg_size = codec_drv->reg_cache_size * codec_drv->reg_word_size; + return reg % (DIV_ROUND_UP(reg_size, snd_soc_lzo_block_count()) / + codec_drv->reg_word_size); +} + +static inline int snd_soc_lzo_get_blksize(struct snd_soc_codec *codec) +{ + struct snd_soc_codec_driver *codec_drv; + size_t reg_size; + + codec_drv = codec->driver; + reg_size = codec_drv->reg_cache_size * codec_drv->reg_word_size; + return DIV_ROUND_UP(reg_size, snd_soc_lzo_block_count()); +} + +static int snd_soc_lzo_cache_sync(struct snd_soc_codec *codec) +{ + struct snd_soc_lzo_ctx **lzo_blocks; + unsigned int val; + int i; + + lzo_blocks = codec->reg_cache; + for_each_set_bit(i, lzo_blocks[0]->sync_bmp, lzo_blocks[0]->sync_bmp_nbits) { + snd_soc_cache_read(codec, i, &val); + snd_soc_write(codec, i, val); + dev_dbg(codec->dev, "Synced register %#x, value = %#x\n", + i, val); + } + + return 0; +} + +static int snd_soc_lzo_cache_write(struct snd_soc_codec *codec, + unsigned int reg, unsigned int value) +{ + struct snd_soc_lzo_ctx *lzo_block, **lzo_blocks; + int ret, blkindex, blkpos; + size_t blksize, tmp_dst_len; + void *tmp_dst; + + /* index of the compressed lzo block */ + blkindex = snd_soc_lzo_get_blkindex(codec, reg); + /* register index within the decompressed block */ + blkpos = snd_soc_lzo_get_blkpos(codec, reg); + /* size of the compressed block */ + blksize = snd_soc_lzo_get_blksize(codec); + lzo_blocks = codec->reg_cache; + lzo_block = lzo_blocks[blkindex]; + + /* save the pointer and length of the compressed block */ + tmp_dst = lzo_block->dst; + tmp_dst_len = lzo_block->dst_len; + + /* prepare the source to be the compressed block */ + lzo_block->src = lzo_block->dst; + lzo_block->src_len = lzo_block->dst_len; + + /* decompress the block */ + ret = snd_soc_lzo_decompress_cache_block(codec, lzo_block); + if (ret < 0) { + kfree(lzo_block->dst); + goto out; + } + + /* write the new value to the cache */ + switch (codec->driver->reg_word_size) { + case 1: { + u8 *cache; + cache = lzo_block->dst; + if (cache[blkpos] == value) { + kfree(lzo_block->dst); + goto out; + } + cache[blkpos] = value; + } + break; + case 2: { + u16 *cache; + cache = lzo_block->dst; + if (cache[blkpos] == value) { + kfree(lzo_block->dst); + goto out; + } + cache[blkpos] = value; + } + break; + default: + BUG(); + } + + /* prepare the source to be the decompressed block */ + lzo_block->src = lzo_block->dst; + lzo_block->src_len = lzo_block->dst_len; + + /* compress the block */ + ret = snd_soc_lzo_compress_cache_block(codec, lzo_block); + if (ret < 0) { + kfree(lzo_block->dst); + kfree(lzo_block->src); + goto out; + } + + /* set the bit so we know we have to sync this register */ + set_bit(reg, lzo_block->sync_bmp); + kfree(tmp_dst); + kfree(lzo_block->src); + return 0; +out: + lzo_block->dst = tmp_dst; + lzo_block->dst_len = tmp_dst_len; + return ret; +} + +static int snd_soc_lzo_cache_read(struct snd_soc_codec *codec, + unsigned int reg, unsigned int *value) +{ + struct snd_soc_lzo_ctx *lzo_block, **lzo_blocks; + int ret, blkindex, blkpos; + size_t blksize, tmp_dst_len; + void *tmp_dst; + + *value = 0; + /* index of the compressed lzo block */ + blkindex = snd_soc_lzo_get_blkindex(codec, reg); + /* register index within the decompressed block */ + blkpos = snd_soc_lzo_get_blkpos(codec, reg); + /* size of the compressed block */ + blksize = snd_soc_lzo_get_blksize(codec); + lzo_blocks = codec->reg_cache; + lzo_block = lzo_blocks[blkindex]; + + /* save the pointer and length of the compressed block */ + tmp_dst = lzo_block->dst; + tmp_dst_len = lzo_block->dst_len; + + /* prepare the source to be the compressed block */ + lzo_block->src = lzo_block->dst; + lzo_block->src_len = lzo_block->dst_len; + + /* decompress the block */ + ret = snd_soc_lzo_decompress_cache_block(codec, lzo_block); + if (ret >= 0) { + /* fetch the value from the cache */ + switch (codec->driver->reg_word_size) { + case 1: { + u8 *cache; + cache = lzo_block->dst; + *value = cache[blkpos]; + } + break; + case 2: { + u16 *cache; + cache = lzo_block->dst; + *value = cache[blkpos]; + } + break; + default: + BUG(); + } + } + + kfree(lzo_block->dst); + /* restore the pointer and length of the compressed block */ + lzo_block->dst = tmp_dst; + lzo_block->dst_len = tmp_dst_len; + return 0; +} + +static int snd_soc_lzo_cache_exit(struct snd_soc_codec *codec) +{ + struct snd_soc_lzo_ctx **lzo_blocks; + int i, blkcount; + + lzo_blocks = codec->reg_cache; + if (!lzo_blocks) + return 0; + + blkcount = snd_soc_lzo_block_count(); + /* + * the pointer to the bitmap used for syncing the cache + * is shared amongst all lzo_blocks. Ensure it is freed + * only once. + */ + if (lzo_blocks[0]) + kfree(lzo_blocks[0]->sync_bmp); + for (i = 0; i < blkcount; ++i) { + if (lzo_blocks[i]) { + kfree(lzo_blocks[i]->wmem); + kfree(lzo_blocks[i]->dst); + } + /* each lzo_block is a pointer returned by kmalloc or NULL */ + kfree(lzo_blocks[i]); + } + kfree(lzo_blocks); + codec->reg_cache = NULL; + return 0; +} + +static int snd_soc_lzo_cache_init(struct snd_soc_codec *codec) +{ + struct snd_soc_lzo_ctx **lzo_blocks; + size_t reg_size, bmp_size; + struct snd_soc_codec_driver *codec_drv; + int ret, tofree, i, blksize, blkcount; + const char *p, *end; + unsigned long *sync_bmp; + + ret = 0; + codec_drv = codec->driver; + reg_size = codec_drv->reg_cache_size * codec_drv->reg_word_size; + + /* + * If we have not been given a default register cache + * then allocate a dummy zero-ed out region, compress it + * and remember to free it afterwards. + */ + tofree = 0; + if (!codec_drv->reg_cache_default) + tofree = 1; + + if (!codec_drv->reg_cache_default) { + codec_drv->reg_cache_default = kzalloc(reg_size, + GFP_KERNEL); + if (!codec_drv->reg_cache_default) + return -ENOMEM; + } + + blkcount = snd_soc_lzo_block_count(); + codec->reg_cache = kzalloc(blkcount * sizeof *lzo_blocks, + GFP_KERNEL); + if (!codec->reg_cache) { + ret = -ENOMEM; + goto err_tofree; + } + lzo_blocks = codec->reg_cache; + + /* + * allocate a bitmap to be used when syncing the cache with + * the hardware. Each time a register is modified, the corresponding + * bit is set in the bitmap, so we know that we have to sync + * that register. + */ + bmp_size = codec_drv->reg_cache_size; + sync_bmp = kmalloc(BITS_TO_LONGS(bmp_size) * sizeof (long), + GFP_KERNEL); + if (!sync_bmp) { + ret = -ENOMEM; + goto err; + } + bitmap_zero(sync_bmp, reg_size); + + /* allocate the lzo blocks and initialize them */ + for (i = 0; i < blkcount; ++i) { + lzo_blocks[i] = kzalloc(sizeof **lzo_blocks, + GFP_KERNEL); + if (!lzo_blocks[i]) { + kfree(sync_bmp); + ret = -ENOMEM; + goto err; + } + lzo_blocks[i]->sync_bmp = sync_bmp; + lzo_blocks[i]->sync_bmp_nbits = reg_size; + /* alloc the working space for the compressed block */ + ret = snd_soc_lzo_prepare(lzo_blocks[i]); + if (ret < 0) + goto err; + } + + blksize = snd_soc_lzo_get_blksize(codec); + p = codec_drv->reg_cache_default; + end = codec_drv->reg_cache_default + reg_size; + /* compress the register map and fill the lzo blocks */ + for (i = 0; i < blkcount; ++i, p += blksize) { + lzo_blocks[i]->src = p; + if (p + blksize > end) + lzo_blocks[i]->src_len = end - p; + else + lzo_blocks[i]->src_len = blksize; + ret = snd_soc_lzo_compress_cache_block(codec, + lzo_blocks[i]); + if (ret < 0) + goto err; + lzo_blocks[i]->decompressed_size = + lzo_blocks[i]->src_len; + } + + if (tofree) + kfree(codec_drv->reg_cache_default); + return 0; +err: + snd_soc_cache_exit(codec); +err_tofree: + if (tofree) + kfree(codec_drv->reg_cache_default); + return ret; +} + static int snd_soc_flat_cache_sync(struct snd_soc_codec *codec) { int i; @@ -883,6 +1288,14 @@ static const struct snd_soc_cache_ops cache_types[] = { .read = snd_soc_flat_cache_read, .write = snd_soc_flat_cache_write, .sync = snd_soc_flat_cache_sync + }, + { + .id = SND_SOC_LZO_COMPRESSION, + .init = snd_soc_lzo_cache_init, + .exit = snd_soc_lzo_cache_exit, + .read = snd_soc_lzo_cache_read, + .write = snd_soc_lzo_cache_write, + .sync = snd_soc_lzo_cache_sync } };
This patch adds support for rbtree compression when storing the register cache. It does this by not adding any uninitialized registers (those whose value is 0). If any of those registers is written with a nonzero value they get added into the rbtree.
Consider a sample device with a large sparse register map. The register indices are between [0, 0x31ff]. An array of 12800 registers is thus created each of which is 2 bytes. This results in a 25kB region. This array normally lives outside soc-core, normally in the driver itself. The original soc-core code would kmemdup this region resulting in 50kB total memory. When using the rbtree compression technique and __devinitconst on the original array the figures are as follows. For this typical device, you might have 100 initialized registers, that is registers that are nonzero by default. We build an rbtree with 100 nodes, each of which is 24 bytes. This results in ~2kB of memory. Assuming that the target arch can freeup the memory used by the initial __devinitconst array, we end up using about ~2kB bytes of actual memory. The memory footprint will increase as uninitialized registers get written and thus new nodes created in the rbtree. In practice, most of those registers are never changed. If the target arch can't freeup the __devinitconst array, we end up using a total of ~27kB. The difference between the rbtree and the LZO caching techniques, is that if using the LZO technique the size of the cache will increase slower as more uninitialized registers get changed.
Signed-off-by: Dimitris Papastamos dp@opensource.wolfsonmicro.com --- include/sound/soc.h | 3 +- sound/soc/soc-cache.c | 232 +++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 234 insertions(+), 1 deletions(-)
diff --git a/include/sound/soc.h b/include/sound/soc.h index 10f5932..af23f42 100644 --- a/include/sound/soc.h +++ b/include/sound/soc.h @@ -257,7 +257,8 @@ enum snd_soc_control_type {
enum snd_soc_compress_type { SND_SOC_NO_COMPRESSION, - SND_SOC_LZO_COMPRESSION + SND_SOC_LZO_COMPRESSION, + SND_SOC_RBTREE_COMPRESSION };
int snd_soc_register_platform(struct device *dev, diff --git a/sound/soc/soc-cache.c b/sound/soc/soc-cache.c index 4581bf1..6c0589e 100644 --- a/sound/soc/soc-cache.c +++ b/sound/soc/soc-cache.c @@ -16,6 +16,7 @@ #include <sound/soc.h> #include <linux/lzo.h> #include <linux/bitmap.h> +#include <linux/rbtree.h>
static unsigned int snd_soc_4_12_read(struct snd_soc_codec *codec, unsigned int reg) @@ -760,6 +761,229 @@ int snd_soc_codec_set_cache_io(struct snd_soc_codec *codec, } EXPORT_SYMBOL_GPL(snd_soc_codec_set_cache_io);
+struct snd_soc_rbtree_node { + struct rb_node node; + unsigned int reg; + unsigned int value; + unsigned int defval; +} __attribute__ ((packed)); + +struct snd_soc_rbtree_ctx { + struct rb_root root; +}; + +static struct snd_soc_rbtree_node *snd_soc_rbtree_lookup( + struct rb_root *root, unsigned int reg) +{ + struct rb_node *node; + struct snd_soc_rbtree_node *rbnode; + + node = root->rb_node; + while (node) { + rbnode = container_of(node, struct snd_soc_rbtree_node, node); + if (rbnode->reg < reg) + node = node->rb_left; + else if (rbnode->reg > reg) + node = node->rb_right; + else + return rbnode; + } + + return NULL; +} + + +static int snd_soc_rbtree_insert(struct rb_root *root, + struct snd_soc_rbtree_node *rbnode) +{ + struct rb_node **new, *parent; + struct snd_soc_rbtree_node *rbnode_tmp; + + parent = NULL; + new = &root->rb_node; + while (*new) { + rbnode_tmp = container_of(*new, struct snd_soc_rbtree_node, + node); + parent = *new; + if (rbnode_tmp->reg < rbnode->reg) + new = &((*new)->rb_left); + else if (rbnode_tmp->reg > rbnode->reg) + new = &((*new)->rb_right); + else + return 0; + } + + /* insert the node into the rbtree */ + rb_link_node(&rbnode->node, parent, new); + rb_insert_color(&rbnode->node, root); + + return 1; +} + +static int snd_soc_rbtree_cache_sync(struct snd_soc_codec *codec) +{ + struct snd_soc_rbtree_ctx *rbtree_ctx; + struct rb_node *node; + struct snd_soc_rbtree_node *rbnode; + unsigned int val; + + rbtree_ctx = codec->reg_cache; + for (node = rb_first(&rbtree_ctx->root); node; node = rb_next(node)) { + rbnode = rb_entry(node, struct snd_soc_rbtree_node, node); + if (rbnode->value == rbnode->defval) + continue; + snd_soc_cache_read(codec, rbnode->reg, &val); + snd_soc_write(codec, rbnode->reg, val); + dev_dbg(codec->dev, "Synced register %#x, value = %#x\n", + rbnode->reg, val); + } + + return 0; +} + +static int snd_soc_rbtree_cache_write(struct snd_soc_codec *codec, + unsigned int reg, unsigned int value) +{ + struct snd_soc_rbtree_ctx *rbtree_ctx; + struct snd_soc_rbtree_node *rbnode; + + rbtree_ctx = codec->reg_cache; + rbnode = snd_soc_rbtree_lookup(&rbtree_ctx->root, reg); + if (rbnode) { + if (rbnode->value == value) + return 0; + rbnode->value = value; + } else { + /* bail out early, no need to create the rbnode yet */ + if (!value) + return 0; + /* + * for uninitialized registers whose value is changed + * from the default zero, create an rbnode and insert + * it into the tree. + */ + rbnode = kzalloc(sizeof *rbnode, GFP_KERNEL); + if (!rbnode) + return -ENOMEM; + rbnode->reg = reg; + rbnode->value = value; + snd_soc_rbtree_insert(&rbtree_ctx->root, rbnode); + } + + return 0; +} + +static int snd_soc_rbtree_cache_read(struct snd_soc_codec *codec, + unsigned int reg, unsigned int *value) +{ + struct snd_soc_rbtree_ctx *rbtree_ctx; + struct snd_soc_rbtree_node *rbnode; + + rbtree_ctx = codec->reg_cache; + rbnode = snd_soc_rbtree_lookup(&rbtree_ctx->root, reg); + if (rbnode) { + *value = rbnode->value; + } else { + /* uninitialized registers default to 0 */ + *value = 0; + } + + return 0; +} + +static int snd_soc_rbtree_cache_exit(struct snd_soc_codec *codec) +{ + struct rb_node *next; + struct snd_soc_rbtree_ctx *rbtree_ctx; + struct snd_soc_rbtree_node *rbtree_node; + + /* if we've already been called then just return */ + rbtree_ctx = codec->reg_cache; + if (!rbtree_ctx) + return 0; + + /* free up the rbtree */ + next = rb_first(&rbtree_ctx->root); + while (next) { + rbtree_node = rb_entry(next, struct snd_soc_rbtree_node, node); + next = rb_next(&rbtree_node->node); + rb_erase(&rbtree_node->node, &rbtree_ctx->root); + kfree(rbtree_node); + } + + /* release the resources */ + kfree(codec->reg_cache); + codec->reg_cache = NULL; + + return 0; +} + +static int snd_soc_rbtree_cache_init(struct snd_soc_codec *codec) +{ + struct snd_soc_rbtree_ctx *rbtree_ctx; + + codec->reg_cache = kmalloc(sizeof *rbtree_ctx, GFP_KERNEL); + if (!codec->reg_cache) + return -ENOMEM; + + rbtree_ctx = codec->reg_cache; + rbtree_ctx->root = RB_ROOT; + + if (!codec->driver->reg_cache_default) + return 0; + +/* + * populate the rbtree with the initialized registers. All other + * registers will be inserted into the tree when they are first written. + * + * The reasoning behind this, is that we need to step through and + * dereference the cache in u8/u16 increments without sacrificing + * portability. This could also be done using memcpy() but that would + * be slightly more cryptic. + */ +#define snd_soc_rbtree_populate(cache) \ +({ \ + int ret, i; \ + struct snd_soc_rbtree_node *rbtree_node; \ + \ + ret = 0; \ + cache = codec->driver->reg_cache_default; \ + for (i = 0; i < codec->driver->reg_cache_size; ++i) { \ + if (!cache[i]) \ + continue; \ + rbtree_node = kzalloc(sizeof *rbtree_node, GFP_KERNEL); \ + if (!rbtree_node) { \ + ret = -ENOMEM; \ + snd_soc_cache_exit(codec); \ + break; \ + } \ + rbtree_node->reg = i; \ + rbtree_node->value = cache[i]; \ + rbtree_node->defval = cache[i]; \ + snd_soc_rbtree_insert(&rbtree_ctx->root, \ + rbtree_node); \ + } \ + ret; \ +}) + + switch (codec->driver->reg_word_size) { + case 1: { + const u8 *cache; + + return snd_soc_rbtree_populate(cache); + } + case 2: { + const u16 *cache; + + return snd_soc_rbtree_populate(cache); + } + default: + BUG(); + } + + return 0; +} + struct snd_soc_lzo_ctx { void *wmem; void *dst; @@ -1296,6 +1520,14 @@ static const struct snd_soc_cache_ops cache_types[] = { .read = snd_soc_lzo_cache_read, .write = snd_soc_lzo_cache_write, .sync = snd_soc_lzo_cache_sync + }, + { + .id = SND_SOC_RBTREE_COMPRESSION, + .init = snd_soc_rbtree_cache_init, + .exit = snd_soc_rbtree_cache_exit, + .read = snd_soc_rbtree_cache_read, + .write = snd_soc_rbtree_cache_write, + .sync = snd_soc_rbtree_cache_sync } };
On Thu, 2010-11-11 at 10:04 +0000, Dimitris Papastamos wrote:
This patch series introduces the new caching API. The idea behind this caching interface is that we can provide different means of organizing and accessing the register cache. This is useful for large and sparse register maps, where one can use some kind of compression algorithm to reduce the memory footprint. The caching API is designed in such way to eliminate the need for modifying any existing drivers.
TODO:
- Memory usage statistics, to make it easier to select the proper caching
technique.
- Support for bulk reads/writes.
- Overridable caching behaviour from the machine driver.
- Optimize certain CPU intensive paths.
- Ensure proper locking in init()/exit()/sync().
Dimitris Papastamos (3): ASoC: soc-cache: Add support for flat register caching ASoC: soc-cache: Add support for LZO register caching ASoC: soc-cache: Add support for rbtree based register caching
Acked-by: Liam Girdwood lrg@slimlogic.co.uk
On Thu, Nov 11, 2010 at 10:04:56AM +0000, Dimitris Papastamos wrote:
This patch series introduces the new caching API. The idea behind this caching interface is that we can provide different means of organizing and accessing the register cache. This is useful for large and sparse register maps, where one can use some kind of compression algorithm to reduce the memory footprint. The caching API is designed in such way to eliminate the need for modifying any existing drivers.
Applied, thanks.
TODO:
- Memory usage statistics, to make it easier to select the proper caching
technique.
- Support for bulk reads/writes.
- Overridable caching behaviour from the machine driver.
This one in particular would be very good.
On Thu, 2010-11-11 at 16:20 +0000, Mark Brown wrote:
On Thu, Nov 11, 2010 at 10:04:56AM +0000, Dimitris Papastamos wrote:
This patch series introduces the new caching API. The idea behind this caching interface is that we can provide different means of organizing and accessing the register cache. This is useful for large and sparse register maps, where one can use some kind of compression algorithm to reduce the memory footprint. The caching API is designed in such way to eliminate the need for modifying any existing drivers.
Applied, thanks.
TODO:
- Memory usage statistics, to make it easier to select the proper caching
technique.
- Support for bulk reads/writes.
- Overridable caching behaviour from the machine driver.
This one in particular would be very good.
Yes, I have an idea on how to do that. I will send you a patch at some point.
Thanks, Dimitrios
participants (3)
-
Dimitris Papastamos
-
Liam Girdwood
-
Mark Brown