linux/crypto/async_tx/async_tx.c

645 lines
16 KiB
C
Raw Normal View History

async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
/*
* core routines for the asynchronous memory transfer/transform api
*
* Copyright © 2006, Intel Corporation.
*
* Dan Williams <dan.j.williams@intel.com>
*
* with architecture considerations by:
* Neil Brown <neilb@suse.de>
* Jeff Garzik <jeff@garzik.org>
*
* This program is free software; you can redistribute it and/or modify it
* under the terms and conditions of the GNU General Public License,
* version 2, as published by the Free Software Foundation.
*
* This program is distributed in the hope it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
* more details.
*
* You should have received a copy of the GNU General Public License along with
* this program; if not, write to the Free Software Foundation, Inc.,
* 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
*
*/
#include <linux/rculist.h>
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
#include <linux/kernel.h>
#include <linux/async_tx.h>
#ifdef CONFIG_DMA_ENGINE
static enum dma_state_client
dma_channel_add_remove(struct dma_client *client,
struct dma_chan *chan, enum dma_state state);
static struct dma_client async_tx_dma = {
.event_callback = dma_channel_add_remove,
/* .cap_mask == 0 defaults to all channels */
};
/**
* dma_cap_mask_all - enable iteration over all operation types
*/
static dma_cap_mask_t dma_cap_mask_all;
/**
* chan_ref_percpu - tracks channel allocations per core/opertion
*/
struct chan_ref_percpu {
struct dma_chan_ref *ref;
};
static int channel_table_initialized;
static struct chan_ref_percpu *channel_table[DMA_TX_TYPE_END];
/**
* async_tx_lock - protect modification of async_tx_master_list and serialize
* rebalance operations
*/
static spinlock_t async_tx_lock;
static LIST_HEAD(async_tx_master_list);
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
/* async_tx_issue_pending_all - start all transactions on all channels */
void async_tx_issue_pending_all(void)
{
struct dma_chan_ref *ref;
rcu_read_lock();
list_for_each_entry_rcu(ref, &async_tx_master_list, node)
ref->chan->device->device_issue_pending(ref->chan);
rcu_read_unlock();
}
EXPORT_SYMBOL_GPL(async_tx_issue_pending_all);
/* dma_wait_for_async_tx - spin wait for a transcation to complete
* @tx: transaction to wait on
*/
enum dma_status
dma_wait_for_async_tx(struct dma_async_tx_descriptor *tx)
{
enum dma_status status;
struct dma_async_tx_descriptor *iter;
struct dma_async_tx_descriptor *parent;
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
if (!tx)
return DMA_SUCCESS;
/* poll through the dependency chain, return when tx is complete */
do {
iter = tx;
/* find the root of the unsubmitted dependency chain */
do {
parent = iter->parent;
if (!parent)
break;
else
iter = parent;
} while (parent);
/* there is a small window for ->parent == NULL and
* ->cookie == -EBUSY
*/
while (iter->cookie == -EBUSY)
cpu_relax();
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
status = dma_sync_wait(iter->chan, iter->cookie);
} while (status == DMA_IN_PROGRESS || (iter != tx));
return status;
}
EXPORT_SYMBOL_GPL(dma_wait_for_async_tx);
/* async_tx_run_dependencies - helper routine for dma drivers to process
* (start) dependent operations on their target channel
* @tx: transaction with dependencies
*/
void
async_tx_run_dependencies(struct dma_async_tx_descriptor *tx)
{
struct dma_async_tx_descriptor *next = tx->next;
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
struct dma_chan *chan;
if (!next)
return;
tx->next = NULL;
chan = next->chan;
/* keep submitting up until a channel switch is detected
* in that case we will be called again as a result of
* processing the interrupt from async_tx_channel_switch
*/
while (next && next->chan == chan) {
struct dma_async_tx_descriptor *_next;
spin_lock_bh(&next->lock);
next->parent = NULL;
_next = next->next;
next->next = NULL;
spin_unlock_bh(&next->lock);
next->tx_submit(next);
next = _next;
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
}
chan->device->device_issue_pending(chan);
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
}
EXPORT_SYMBOL_GPL(async_tx_run_dependencies);
static void
free_dma_chan_ref(struct rcu_head *rcu)
{
struct dma_chan_ref *ref;
ref = container_of(rcu, struct dma_chan_ref, rcu);
kfree(ref);
}
static void
init_dma_chan_ref(struct dma_chan_ref *ref, struct dma_chan *chan)
{
INIT_LIST_HEAD(&ref->node);
INIT_RCU_HEAD(&ref->rcu);
ref->chan = chan;
atomic_set(&ref->count, 0);
}
/**
* get_chan_ref_by_cap - returns the nth channel of the given capability
* defaults to returning the channel with the desired capability and the
* lowest reference count if the index can not be satisfied
* @cap: capability to match
* @index: nth channel desired, passing -1 has the effect of forcing the
* default return value
*/
static struct dma_chan_ref *
get_chan_ref_by_cap(enum dma_transaction_type cap, int index)
{
struct dma_chan_ref *ret_ref = NULL, *min_ref = NULL, *ref;
rcu_read_lock();
list_for_each_entry_rcu(ref, &async_tx_master_list, node)
if (dma_has_cap(cap, ref->chan->device->cap_mask)) {
if (!min_ref)
min_ref = ref;
else if (atomic_read(&ref->count) <
atomic_read(&min_ref->count))
min_ref = ref;
if (index-- == 0) {
ret_ref = ref;
break;
}
}
rcu_read_unlock();
if (!ret_ref)
ret_ref = min_ref;
if (ret_ref)
atomic_inc(&ret_ref->count);
return ret_ref;
}
/**
* async_tx_rebalance - redistribute the available channels, optimize
* for cpu isolation in the SMP case, and opertaion isolation in the
* uniprocessor case
*/
static void async_tx_rebalance(void)
{
int cpu, cap, cpu_idx = 0;
unsigned long flags;
if (!channel_table_initialized)
return;
spin_lock_irqsave(&async_tx_lock, flags);
/* undo the last distribution */
for_each_dma_cap_mask(cap, dma_cap_mask_all)
for_each_possible_cpu(cpu) {
struct dma_chan_ref *ref =
per_cpu_ptr(channel_table[cap], cpu)->ref;
if (ref) {
atomic_set(&ref->count, 0);
per_cpu_ptr(channel_table[cap], cpu)->ref =
NULL;
}
}
for_each_dma_cap_mask(cap, dma_cap_mask_all)
for_each_online_cpu(cpu) {
struct dma_chan_ref *new;
if (NR_CPUS > 1)
new = get_chan_ref_by_cap(cap, cpu_idx++);
else
new = get_chan_ref_by_cap(cap, -1);
per_cpu_ptr(channel_table[cap], cpu)->ref = new;
}
spin_unlock_irqrestore(&async_tx_lock, flags);
}
static enum dma_state_client
dma_channel_add_remove(struct dma_client *client,
struct dma_chan *chan, enum dma_state state)
{
unsigned long found, flags;
struct dma_chan_ref *master_ref, *ref;
enum dma_state_client ack = DMA_DUP; /* default: take no action */
switch (state) {
case DMA_RESOURCE_AVAILABLE:
found = 0;
rcu_read_lock();
list_for_each_entry_rcu(ref, &async_tx_master_list, node)
if (ref->chan == chan) {
found = 1;
break;
}
rcu_read_unlock();
pr_debug("async_tx: dma resource available [%s]\n",
found ? "old" : "new");
if (!found)
ack = DMA_ACK;
else
break;
/* add the channel to the generic management list */
master_ref = kmalloc(sizeof(*master_ref), GFP_KERNEL);
if (master_ref) {
/* keep a reference until async_tx is unloaded */
dma_chan_get(chan);
init_dma_chan_ref(master_ref, chan);
spin_lock_irqsave(&async_tx_lock, flags);
list_add_tail_rcu(&master_ref->node,
&async_tx_master_list);
spin_unlock_irqrestore(&async_tx_lock,
flags);
} else {
printk(KERN_WARNING "async_tx: unable to create"
" new master entry in response to"
" a DMA_RESOURCE_ADDED event"
" (-ENOMEM)\n");
return 0;
}
async_tx_rebalance();
break;
case DMA_RESOURCE_REMOVED:
found = 0;
spin_lock_irqsave(&async_tx_lock, flags);
list_for_each_entry(ref, &async_tx_master_list, node)
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
if (ref->chan == chan) {
/* permit backing devices to go away */
dma_chan_put(ref->chan);
list_del_rcu(&ref->node);
call_rcu(&ref->rcu, free_dma_chan_ref);
found = 1;
break;
}
spin_unlock_irqrestore(&async_tx_lock, flags);
pr_debug("async_tx: dma resource removed [%s]\n",
found ? "ours" : "not ours");
if (found)
ack = DMA_ACK;
else
break;
async_tx_rebalance();
break;
case DMA_RESOURCE_SUSPEND:
case DMA_RESOURCE_RESUME:
printk(KERN_WARNING "async_tx: does not support dma channel"
" suspend/resume\n");
break;
default:
BUG();
}
return ack;
}
static int __init
async_tx_init(void)
{
enum dma_transaction_type cap;
spin_lock_init(&async_tx_lock);
bitmap_fill(dma_cap_mask_all.bits, DMA_TX_TYPE_END);
/* an interrupt will never be an explicit operation type.
* clearing this bit prevents allocation to a slot in 'channel_table'
*/
clear_bit(DMA_INTERRUPT, dma_cap_mask_all.bits);
for_each_dma_cap_mask(cap, dma_cap_mask_all) {
channel_table[cap] = alloc_percpu(struct chan_ref_percpu);
if (!channel_table[cap])
goto err;
}
channel_table_initialized = 1;
dma_async_client_register(&async_tx_dma);
dma_async_client_chan_request(&async_tx_dma);
printk(KERN_INFO "async_tx: api initialized (async)\n");
return 0;
err:
printk(KERN_ERR "async_tx: initialization failure\n");
while (--cap >= 0)
free_percpu(channel_table[cap]);
return 1;
}
static void __exit async_tx_exit(void)
{
enum dma_transaction_type cap;
channel_table_initialized = 0;
for_each_dma_cap_mask(cap, dma_cap_mask_all)
if (channel_table[cap])
free_percpu(channel_table[cap]);
dma_async_client_unregister(&async_tx_dma);
}
/**
* __async_tx_find_channel - find a channel to carry out the operation or let
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
* the transaction execute synchronously
* @depend_tx: transaction dependency
* @tx_type: transaction type
*/
struct dma_chan *
__async_tx_find_channel(struct dma_async_tx_descriptor *depend_tx,
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
enum dma_transaction_type tx_type)
{
/* see if we can keep the chain on one channel */
if (depend_tx &&
dma_has_cap(tx_type, depend_tx->chan->device->cap_mask))
return depend_tx->chan;
else if (likely(channel_table_initialized)) {
struct dma_chan_ref *ref;
int cpu = get_cpu();
ref = per_cpu_ptr(channel_table[tx_type], cpu)->ref;
put_cpu();
return ref ? ref->chan : NULL;
} else
return NULL;
}
EXPORT_SYMBOL_GPL(__async_tx_find_channel);
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
#else
static int __init async_tx_init(void)
{
printk(KERN_INFO "async_tx: api initialized (sync-only)\n");
return 0;
}
static void __exit async_tx_exit(void)
{
do { } while (0);
}
#endif
/**
* async_tx_channel_switch - queue an interrupt descriptor with a dependency
* pre-attached.
* @depend_tx: the operation that must finish before the new operation runs
* @tx: the new operation
*/
static void
async_tx_channel_switch(struct dma_async_tx_descriptor *depend_tx,
struct dma_async_tx_descriptor *tx)
{
struct dma_chan *chan;
struct dma_device *device;
struct dma_async_tx_descriptor *intr_tx = (void *) ~0;
/* first check to see if we can still append to depend_tx */
spin_lock_bh(&depend_tx->lock);
if (depend_tx->parent && depend_tx->chan == tx->chan) {
tx->parent = depend_tx;
depend_tx->next = tx;
intr_tx = NULL;
}
spin_unlock_bh(&depend_tx->lock);
if (!intr_tx)
return;
chan = depend_tx->chan;
device = chan->device;
/* see if we can schedule an interrupt
* otherwise poll for completion
*/
if (dma_has_cap(DMA_INTERRUPT, device->cap_mask))
intr_tx = device->device_prep_dma_interrupt(chan, 0);
else
intr_tx = NULL;
if (intr_tx) {
intr_tx->callback = NULL;
intr_tx->callback_param = NULL;
tx->parent = intr_tx;
/* safe to set ->next outside the lock since we know we are
* not submitted yet
*/
intr_tx->next = tx;
/* check if we need to append */
spin_lock_bh(&depend_tx->lock);
if (depend_tx->parent) {
intr_tx->parent = depend_tx;
depend_tx->next = intr_tx;
async_tx_ack(intr_tx);
intr_tx = NULL;
}
spin_unlock_bh(&depend_tx->lock);
if (intr_tx) {
intr_tx->parent = NULL;
intr_tx->tx_submit(intr_tx);
async_tx_ack(intr_tx);
}
} else {
if (dma_wait_for_async_tx(depend_tx) == DMA_ERROR)
panic("%s: DMA_ERROR waiting for depend_tx\n",
__func__);
tx->tx_submit(tx);
}
}
/**
* submit_disposition - while holding depend_tx->lock we must avoid submitting
* new operations to prevent a circular locking dependency with
* drivers that already hold a channel lock when calling
* async_tx_run_dependencies.
* @ASYNC_TX_SUBMITTED: we were able to append the new operation under the lock
* @ASYNC_TX_CHANNEL_SWITCH: when the lock is dropped schedule a channel switch
* @ASYNC_TX_DIRECT_SUBMIT: when the lock is dropped submit directly
*/
enum submit_disposition {
ASYNC_TX_SUBMITTED,
ASYNC_TX_CHANNEL_SWITCH,
ASYNC_TX_DIRECT_SUBMIT,
};
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
void
async_tx_submit(struct dma_chan *chan, struct dma_async_tx_descriptor *tx,
enum async_tx_flags flags, struct dma_async_tx_descriptor *depend_tx,
dma_async_tx_callback cb_fn, void *cb_param)
{
tx->callback = cb_fn;
tx->callback_param = cb_param;
if (depend_tx) {
enum submit_disposition s;
/* sanity check the dependency chain:
* 1/ if ack is already set then we cannot be sure
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
* we are referring to the correct operation
* 2/ dependencies are 1:1 i.e. two transactions can
* not depend on the same parent
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
*/
BUG_ON(async_tx_test_ack(depend_tx) || depend_tx->next ||
tx->parent);
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
/* the lock prevents async_tx_run_dependencies from missing
* the setting of ->next when ->parent != NULL
*/
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
spin_lock_bh(&depend_tx->lock);
if (depend_tx->parent) {
/* we have a parent so we can not submit directly
* if we are staying on the same channel: append
* else: channel switch
*/
if (depend_tx->chan == chan) {
tx->parent = depend_tx;
depend_tx->next = tx;
s = ASYNC_TX_SUBMITTED;
} else
s = ASYNC_TX_CHANNEL_SWITCH;
} else {
/* we do not have a parent so we may be able to submit
* directly if we are staying on the same channel
*/
if (depend_tx->chan == chan)
s = ASYNC_TX_DIRECT_SUBMIT;
else
s = ASYNC_TX_CHANNEL_SWITCH;
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
}
spin_unlock_bh(&depend_tx->lock);
switch (s) {
case ASYNC_TX_SUBMITTED:
break;
case ASYNC_TX_CHANNEL_SWITCH:
async_tx_channel_switch(depend_tx, tx);
break;
case ASYNC_TX_DIRECT_SUBMIT:
tx->parent = NULL;
tx->tx_submit(tx);
break;
}
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
} else {
tx->parent = NULL;
tx->tx_submit(tx);
}
if (flags & ASYNC_TX_ACK)
async_tx_ack(tx);
if (depend_tx && (flags & ASYNC_TX_DEP_ACK))
async_tx_ack(depend_tx);
}
EXPORT_SYMBOL_GPL(async_tx_submit);
/**
* async_trigger_callback - schedules the callback function to be run after
* any dependent operations have been completed.
* @flags: ASYNC_TX_ACK, ASYNC_TX_DEP_ACK
* @depend_tx: 'callback' requires the completion of this transaction
* @cb_fn: function to call after depend_tx completes
* @cb_param: parameter to pass to the callback routine
*/
struct dma_async_tx_descriptor *
async_trigger_callback(enum async_tx_flags flags,
struct dma_async_tx_descriptor *depend_tx,
dma_async_tx_callback cb_fn, void *cb_param)
{
struct dma_chan *chan;
struct dma_device *device;
struct dma_async_tx_descriptor *tx;
if (depend_tx) {
chan = depend_tx->chan;
device = chan->device;
/* see if we can schedule an interrupt
* otherwise poll for completion
*/
if (device && !dma_has_cap(DMA_INTERRUPT, device->cap_mask))
device = NULL;
tx = device ? device->device_prep_dma_interrupt(chan, 0) : NULL;
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
} else
tx = NULL;
if (tx) {
pr_debug("%s: (async)\n", __func__);
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param);
} else {
pr_debug("%s: (sync)\n", __func__);
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
/* wait for any prerequisite operations */
async_tx_quiesce(&depend_tx);
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
async_tx_sync_epilog(cb_fn, cb_param);
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
}
return tx;
}
EXPORT_SYMBOL_GPL(async_trigger_callback);
/**
* async_tx_quiesce - ensure tx is complete and freeable upon return
* @tx - transaction to quiesce
*/
void async_tx_quiesce(struct dma_async_tx_descriptor **tx)
{
if (*tx) {
/* if ack is already set then we cannot be sure
* we are referring to the correct operation
*/
BUG_ON(async_tx_test_ack(*tx));
if (dma_wait_for_async_tx(*tx) == DMA_ERROR)
panic("DMA_ERROR waiting for transaction\n");
async_tx_ack(*tx);
*tx = NULL;
}
}
EXPORT_SYMBOL_GPL(async_tx_quiesce);
async_tx: add the async_tx api The async_tx api provides methods for describing a chain of asynchronous bulk memory transfers/transforms with support for inter-transactional dependencies. It is implemented as a dmaengine client that smooths over the details of different hardware offload engine implementations. Code that is written to the api can optimize for asynchronous operation and the api will fit the chain of operations to the available offload resources. I imagine that any piece of ADMA hardware would register with the 'async_*' subsystem, and a call to async_X would be routed as appropriate, or be run in-line. - Neil Brown async_tx exploits the capabilities of struct dma_async_tx_descriptor to provide an api of the following general format: struct dma_async_tx_descriptor * async_<operation>(..., struct dma_async_tx_descriptor *depend_tx, dma_async_tx_callback cb_fn, void *cb_param) { struct dma_chan *chan = async_tx_find_channel(depend_tx, <operation>); struct dma_device *device = chan ? chan->device : NULL; int int_en = cb_fn ? 1 : 0; struct dma_async_tx_descriptor *tx = device ? device->device_prep_dma_<operation>(chan, len, int_en) : NULL; if (tx) { /* run <operation> asynchronously */ ... tx->tx_set_dest(addr, tx, index); ... tx->tx_set_src(addr, tx, index); ... async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); } else { /* run <operation> synchronously */ ... <operation> ... async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param); } return tx; } async_tx_find_channel() returns a capable channel from its pool. The channel pool is organized as a per-cpu array of channel pointers. The async_tx_rebalance() routine is tasked with managing these arrays. In the uniprocessor case async_tx_rebalance() tries to spread responsibility evenly over channels of similar capabilities. For example if there are two copy+xor channels, one will handle copy operations and the other will handle xor. In the SMP case async_tx_rebalance() attempts to spread the operations evenly over the cpus, e.g. cpu0 gets copy channel0 and xor channel0 while cpu1 gets copy channel 1 and xor channel 1. When a dependency is specified async_tx_find_channel defaults to keeping the operation on the same channel. A xor->copy->xor chain will stay on one channel if it supports both operation types, otherwise the transaction will transition between a copy and a xor resource. Currently the raid5 implementation in the MD raid456 driver has been converted to the async_tx api. A driver for the offload engines on the Intel Xscale series of I/O processors, iop-adma, is provided in a later commit. With the iop-adma driver and async_tx, raid456 is able to offload copy, xor, and xor-zero-sum operations to hardware engines. On iop342 tiobench showed higher throughput for sequential writes (20 - 30% improvement) and sequential reads to a degraded array (40 - 55% improvement). For the other cases performance was roughly equal, +/- a few percentage points. On a x86-smp platform the performance of the async_tx implementation (in synchronous mode) was also +/- a few percentage points of the original implementation. According to 'top' on iop342 CPU utilization drops from ~50% to ~15% during a 'resync' while the speed according to /proc/mdstat doubles from ~25 MB/s to ~50 MB/s. The tiobench command line used for testing was: tiobench --size 2048 --block 4096 --block 131072 --dir /mnt/raid --numruns 5 * iop342 had 1GB of memory available Details: * if CONFIG_DMA_ENGINE=n the asynchronous path is compiled away by making async_tx_find_channel a static inline routine that always returns NULL * when a callback is specified for a given transaction an interrupt will fire at operation completion time and the callback will occur in a tasklet. if the the channel does not support interrupts then a live polling wait will be performed * the api is written as a dmaengine client that requests all available channels * In support of dependencies the api implicitly schedules channel-switch interrupts. The interrupt triggers the cleanup tasklet which causes pending operations to be scheduled on the next channel * Xor engines treat an xor destination address differently than a software xor routine. To the software routine the destination address is an implied source, whereas engines treat it as a write-only destination. This patch modifies the xor_blocks routine to take a an explicit destination address to mirror the hardware. Changelog: * fixed a leftover debug print * don't allow callbacks in async_interrupt_cond * fixed xor_block changes * fixed usage of ASYNC_TX_XOR_DROP_DEST * drop dma mapping methods, suggested by Chris Leech * printk warning fixups from Andrew Morton * don't use inline in C files, Adrian Bunk * select the API when MD is enabled * BUG_ON xor source counts <= 1 * implicitly handle hardware concerns like channel switching and interrupts, Neil Brown * remove the per operation type list, and distribute operation capabilities evenly amongst the available channels * simplify async_tx_find_channel to optimize the fast path * introduce the channel_table_initialized flag to prevent early calls to the api * reorganize the code to mimic crypto * include mm.h as not all archs include it in dma-mapping.h * make the Kconfig options non-user visible, Adrian Bunk * move async_tx under crypto since it is meant as 'core' functionality, and the two may share algorithms in the future * move large inline functions into c files * checkpatch.pl fixes * gpl v2 only correction Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 18:10:44 +00:00
module_init(async_tx_init);
module_exit(async_tx_exit);
MODULE_AUTHOR("Intel Corporation");
MODULE_DESCRIPTION("Asynchronous Bulk Memory Transactions API");
MODULE_LICENSE("GPL");