Skip to content

L
linux-elphel
Commit 0e242816 authored by Mikhail Karpenko's avatar Mikhail Karpenko
Browse files
Options

Move IRQ processing to tasklet, use single flag for disk access

parent a8fbc767
Show whitespace changes
Inline Side-by-side
Showing with 130 additions and 81 deletions
src/drivers/ata/ahci_elphel.c
View file @ 0e242816
......@@ -70,9 +70,9 @@ static int init_buffers(struct device *dev, struct frame_buffers *buffs);
static void init_vectors(struct frame_buffers *buffs, struct fvec *chunks);
static void deinit_buffers(struct device *dev, struct frame_buffers *buffs);
static inline struct elphel_ahci_priv *dev_get_dpriv(struct device *dev);
static void finish_cmd(struct device *dev, struct elphel_ahci_priv *dpriv);
static void finish_rec(struct device *dev, struct elphel_ahci_priv *dpriv, struct ata_port *port);
static int process_cmd(struct device *dev, struct elphel_ahci_priv *dpriv, struct ata_port *port);
static void finish_cmd(struct elphel_ahci_priv *dpriv);
static void finish_rec(struct elphel_ahci_priv *dpriv);
static int process_cmd(struct elphel_ahci_priv *dpriv);
//static void start_cmd(struct device *dev, struct elphel_ahci_priv *dpriv, struct ata_port *port);
static inline size_t get_size_from(const struct fvec *vects, int index, size_t offset, int all);
static inline void vectmov(struct fvec *vec, size_t len);
......@@ -80,6 +80,8 @@ static inline void vectsplit(struct fvec *vect, struct fvec *parts, size_t *n_el
int move_tail(struct elphel_ahci_priv *dpriv, int lock);
int move_head(struct elphel_ahci_priv *dpriv);
size_t get_prev_slot(struct elphel_ahci_priv *dpriv);
int is_cmdq_empty(const struct elphel_ahci_priv *dpriv);
void process_queue(unsigned long data);
static ssize_t set_load_flag(struct device *dev, struct device_attribute *attr,
const char *buff, size_t buff_sz)
......@@ -149,28 +151,47 @@ static irqreturn_t elphel_irq_handler(int irq, void * dev_instance)
writel(host_irq_stat, hpriv->mmio + HOST_IRQ_STAT);
handled = IRQ_HANDLED;
tasklet_schedule(&dpriv->bh);
} else {
/* pass handling to AHCI level and then decide if the resource should be freed */
handled = ahci_single_irq_intr(irq, dev_instance);
spin_lock_irqsave(&dpriv->flags_lock, irq_flags);
if (is_cmdq_empty(dpriv)) {
dpriv->flags &= ~DISK_BUSY;
// printk(KERN_DEBUG ">>> irq: flags = %u\n", dpriv->flags);
} else {
// printk(KERN_DEBUG ">>> irq schedule tasklet: flags = %u\n", dpriv->flags);
tasklet_schedule(&dpriv->bh);
}
spin_unlock_irqrestore(&dpriv->flags_lock, irq_flags);
}
return handled;
}
void process_queue(unsigned long data)
{
unsigned long irq_flags;
struct elphel_ahci_priv *dpriv = (struct elphel_ahci_priv *)data;
struct ata_host *host = dev_get_drvdata(dpriv->dev);
struct ata_port *port = host->ports[DEFAULT_PORT_NUM];
if (process_cmd(host->dev, dpriv, host->ports[DEFAULT_PORT_NUM]) == 0) {
finish_cmd(host->dev, dpriv);
if (process_cmd(dpriv) == 0) {
finish_cmd(dpriv);
if (move_head(dpriv) != -1) {
process_cmd(host->dev, dpriv, host->ports[DEFAULT_PORT_NUM]);
process_cmd(dpriv);
} else {
if (dpriv->flags & DELAYED_FINISH) {
dpriv->flags &= ~DELAYED_FINISH;
finish_rec(host->dev, dpriv, port);
finish_rec(dpriv);
} else {
/* all commands have been processed; flag reset here does not need spin lock protection */
dpriv->flags &= ~INTERNAL_CMD;
}
/* all commands have been processed */
spin_lock_irqsave(&dpriv->flags_lock, irq_flags);
dpriv->flags &= ~DISK_BUSY;
spin_unlock_irqrestore(&dpriv->flags_lock, irq_flags);
}
}
} else {
/* pass handling to AHCI level; flag reset here does not need spin lock protection */
dpriv->flags &= ~NATIVE_CMD;
handled = ahci_single_irq_intr(irq, dev_instance);
}
return handled;
}
// What about port_stop and freeing/unmapping ?
......@@ -275,7 +296,9 @@ static int elphel_drv_probe(struct platform_device *pdev)
if (!dpriv)
return -ENOMEM;
dpriv->dev = dev;
spin_lock_init(&dpriv->flags_lock);
tasklet_init(&dpriv->bh, process_queue, (unsigned long)dpriv);
for (i = 0; i < MAX_CMD_SLOTS; i++) {
ret = init_buffers(dev, &dpriv->fbuffs[i]);
......@@ -328,6 +351,7 @@ static int elphel_drv_remove(struct platform_device *pdev)
struct elphel_ahci_priv *dpriv = dev_get_dpriv(&pdev->dev);
dev_info(&pdev->dev, "removing Elphel AHCI driver");
tasklet_kill(&dpriv->bh);
for (i = 0; i < MAX_CMD_SLOTS; i++)
deinit_buffers(&pdev->dev, &dpriv->fbuffs[i]);
sysfs_remove_group(&pdev->dev.kobj, &dev_attr_root_group);
......@@ -750,7 +774,7 @@ static inline unsigned char *vectrpos(struct fvec *vec, size_t offset)
{
return (unsigned char *)vec->iov_base + (vec->iov_len - offset);
}
static void align_frame(struct device *dev, struct elphel_ahci_priv *dpriv)
static void align_frame(struct elphel_ahci_priv *dpriv)
{
int i;
int is_delayed = 0;
......@@ -759,6 +783,7 @@ static void align_frame(struct device *dev, struct elphel_ahci_priv *dpriv)
size_t cmd_slot = dpriv->tail_ptr;
size_t prev_slot = get_prev_slot(dpriv);
size_t max_len = dpriv->fbuffs[cmd_slot].common_buff.iov_len;
struct device *dev = dpriv->dev;
struct frame_buffers *fbuffs = &dpriv->fbuffs[cmd_slot];
struct fvec *chunks = &dpriv->data_chunks[cmd_slot];
struct fvec *cbuff = &chunks[CHUNK_COMMON];
......@@ -1120,13 +1145,20 @@ static inline struct elphel_ahci_priv *dev_get_dpriv(struct device *dev)
}
/** Process command and return the number of S/G entries mapped */
static int process_cmd(struct device *dev, struct elphel_ahci_priv *dpriv, struct ata_port *port)
static int process_cmd(struct elphel_ahci_priv *dpriv)
{
int num;
struct fvec *cbuff;
struct ata_host *host = dev_get_drvdata(dpriv->dev);
struct ata_port *port = host->ports[DEFAULT_PORT_NUM];
size_t max_sz = (MAX_LBA_COUNT + 1) * PHY_BLOCK_SIZE;
size_t rem_sz = get_size_from(&dpriv->data_chunks[dpriv->head_ptr], dpriv->curr_data_chunk, dpriv->curr_data_offset, EXCLUDE_REM);
if (dpriv->flags & PROC_CMD)
dpriv->lba_ptr.lba_write += dpriv->lba_ptr.wr_count;
dpriv->flags |= PROC_CMD;
printk(KERN_DEBUG ">>> rem_sz = %u, head pos = %u (curr_data_chunk = %d, curr_data_offset = %u)\n", rem_sz, dpriv->head_ptr, dpriv->curr_data_chunk, dpriv->curr_data_offset);
/* define ATA command to use for current transaction */
if ((dpriv->lba_ptr.lba_write & ~ADDR_MASK_28_BIT) || rem_sz > max_sz) {
dpriv->curr_cmd = ATA_CMD_WRITE_EXT;
......@@ -1136,9 +1168,6 @@ static int process_cmd(struct device *dev, struct elphel_ahci_priv *dpriv, struc
dpriv->max_data_sz = (MAX_LBA_COUNT + 1) * PHY_BLOCK_SIZE;
}
if (dpriv->flags & PROC_CMD)
dpriv->lba_ptr.lba_write += dpriv->lba_ptr.wr_count;
dpriv->flags |= PROC_CMD;
dpriv->sg_elems = map_vectors(dpriv);
if (dpriv->sg_elems != 0) {
dump_sg_list(dpriv->sgl, dpriv->sg_elems);
......@@ -1149,14 +1178,14 @@ static int process_cmd(struct device *dev, struct elphel_ahci_priv *dpriv, struc
dpriv->lba_ptr.lba_write = dpriv->lba_ptr.lba_start;
}
cbuff = &dpriv->fbuffs[dpriv->head_ptr].common_buff;
dma_sync_single_for_device(dev, cbuff->iov_dma, cbuff->iov_len, DMA_TO_DEVICE);
dma_sync_single_for_device(dpriv->dev, cbuff->iov_dma, cbuff->iov_len, DMA_TO_DEVICE);
printk(KERN_DEBUG ">>> time before issuing command: %u\n", get_rtc_usec());
elphel_cmd_issue(port, dpriv->lba_ptr.lba_write, dpriv->lba_ptr.wr_count, dpriv->sgl, dpriv->sg_elems, dpriv->curr_cmd);
}
return dpriv->sg_elems;
}
static void finish_cmd(struct device *dev, struct elphel_ahci_priv *dpriv)
static void finish_cmd(struct elphel_ahci_priv *dpriv)
{
int all;
unsigned long irq_flags;
......@@ -1170,16 +1199,15 @@ static void finish_cmd(struct device *dev, struct elphel_ahci_priv *dpriv)
}
printk(KERN_DEBUG "reset chunks in slot %u, all = %d\n", dpriv->head_ptr, all);
reset_chunks(dpriv->data_chunks[dpriv->head_ptr], all);
/* flag reset here does not need spin lock protection */
dpriv->flags &= ~PROC_CMD;
dpriv->curr_cmd = 0;
dpriv->max_data_sz = 0;
dpriv->curr_data_chunk = 0;
dpriv->curr_data_offset = 0;
dpriv->flags &= ~PROC_CMD;
}
/** Fill free space in REM buffer with 0 and save the reaming data chunk */
static void finish_rec(struct device *dev, struct elphel_ahci_priv *dpriv, struct ata_port *port)
static void finish_rec(struct elphel_ahci_priv *dpriv)
{
size_t stuff_len;
unsigned char *src;
......@@ -1189,17 +1217,17 @@ static void finish_rec(struct device *dev, struct elphel_ahci_priv *dpriv, struc
if (rvect->iov_len == 0)
return;
dev_dbg(dev, "write last chunk of data from slot %u, size: %u\n", dpriv->head_ptr, rvect->iov_len);
dev_dbg(dpriv->dev, "write last chunk of data from slot %u, size: %u\n", dpriv->head_ptr, rvect->iov_len);
stuff_len = PHY_BLOCK_SIZE - rvect->iov_len;
src = vectrpos(rvect, 0);
memset(src, 0, stuff_len);
rvect->iov_len += stuff_len;
dma_sync_single_for_cpu(dev, dpriv->fbuffs[dpriv->head_ptr].common_buff.iov_dma, dpriv->fbuffs[dpriv->head_ptr].common_buff.iov_len, DMA_TO_DEVICE);
dma_sync_single_for_cpu(dpriv->dev, dpriv->fbuffs[dpriv->head_ptr].common_buff.iov_dma, dpriv->fbuffs[dpriv->head_ptr].common_buff.iov_len, DMA_TO_DEVICE);
vectcpy(cvect, rvect->iov_base, rvect->iov_len);
vectshrink(rvect, rvect->iov_len);
dpriv->flags |= LAST_BLOCK;
process_cmd(dev, dpriv, port);
process_cmd(dpriv);
}
int move_tail(struct elphel_ahci_priv *dpriv, int lock)
......@@ -1241,6 +1269,22 @@ int move_head(struct elphel_ahci_priv *dpriv)
}
int is_cmdq_empty(const struct elphel_ahci_priv *dpriv)
{
size_t use_tail;
if (dpriv->flags & LOCK_TAIL) {
/* current command slot is not ready yet, use previous */
use_tail = get_prev_slot(dpriv);
} else {
use_tail = dpriv->tail_ptr;
}
if (dpriv->head_ptr != use_tail)
return 0;
else
return 1;
}
size_t get_prev_slot(struct elphel_ahci_priv *dpriv)
{
size_t slot;
......@@ -1262,9 +1306,8 @@ static ssize_t rawdev_write(struct device *dev, ///<
size_t buff_sz) ///<
{
int i;
bool proceed = false;
unsigned long irq_flags;
struct ata_host *host = dev_get_drvdata(dev);
struct ata_port *port = host->ports[DEFAULT_PORT_NUM];
struct elphel_ahci_priv *dpriv = dev_get_dpriv(dev);
struct frame_data fdata;
size_t rcvd = 0;
......@@ -1279,21 +1322,18 @@ static ssize_t rawdev_write(struct device *dev, ///<
}
memcpy(&fdata, buff, sizeof(struct frame_data));
/* check if we can issue internal command */
printk_ratelimited(KERN_DEBUG "flags on receive cmd: %u\n", dpriv->flags);
/* lock disk resource as soon as possible */
spin_lock_irqsave(&dpriv->flags_lock, irq_flags);
if ((dpriv->flags & NATIVE_CMD) == 0) {
dpriv->flags |= INTERNAL_CMD;
} else {
spin_unlock_irqrestore(&dpriv->flags_lock, irq_flags);
printk_ratelimited(KERN_DEBUG, "system command is in progress, defer Elphel command\n");
// dev_dbg(dev, "system command is in progress, defer Elphel command\n");
return -EAGAIN;
if ((dpriv->flags & DISK_BUSY) == 0) {
dpriv->flags |= DISK_BUSY;
proceed = true;
}
spin_unlock_irqrestore(&dpriv->flags_lock, irq_flags);
if (fdata.cmd == DRV_CMD_FINISH) {
if (!(dpriv->flags & PROC_CMD)) {
finish_rec(dev, dpriv, port);
if ((dpriv->flags & PROC_CMD) == 0 && proceed) {
finish_rec(dpriv);
} else {
dpriv->flags |= DELAYED_FINISH;
}
......@@ -1302,6 +1342,7 @@ static ssize_t rawdev_write(struct device *dev, ///<
if ((move_tail(dpriv, 1) == -1) || dont_process) {
/* we are not ready yet because command queue is full */
printk_ratelimited(KERN_DEBUG "command queue is full\n");
return -EAGAIN;
}
chunks = &dpriv->data_chunks[dpriv->tail_ptr];
......@@ -1313,7 +1354,6 @@ static ssize_t rawdev_write(struct device *dev, ///<
printk(KERN_DEBUG ">>> sensor port: %u\n", fdata.sensor_port);
printk(KERN_DEBUG ">>> cirbuf ptr: %d, cirbuf data len: %d\n", fdata.cirbuf_ptr, fdata.jpeg_len);
printk(KERN_DEBUG ">>> meta_index: %d\n", fdata.meta_index);
printk(KERN_DEBUG ">>> frame will be place to slot %u\n", dpriv->tail_ptr);
rcvd = exif_get_data(fdata.sensor_port, fdata.meta_index, buffs->exif_buff.iov_base, buffs->exif_buff.iov_len);
// rcvd = exif_get_data_tst(fdata.sensor_port, fdata.meta_index, buffs->exif_buff.iov_base, buffs->exif_buff.iov_len, 1);
......@@ -1353,26 +1393,36 @@ static ssize_t rawdev_write(struct device *dev, ///<
// chunks[CHUNK_DATA_1].iov_dma = dma_map_single(dev, chunks[CHUNK_DATA_1].iov_base, chunks[CHUNK_DATA_1].iov_len, DMA_TO_DEVICE);
printk(KERN_DEBUG ">>> unaligned frame dump:\n");
for (i = 0; i < MAX_DATA_CHUNKS; i++) {
printk(KERN_DEBUG ">>>\tslot: %i; len: %u\n", i, chunks[i].iov_len);
}
align_frame(dev, dpriv);
printk(KERN_DEBUG ">>> aligned frame dump:\n");
for (i = 0; i < MAX_DATA_CHUNKS; i++) {
printk(KERN_DEBUG ">>>\tslot: %i; len: %u\n", i, chunks[i].iov_len);
}
if (check_chunks(chunks) != 0) {
dont_process = 1;
return -EINVAL;
}
// printk(KERN_DEBUG ">>> unaligned frame dump:\n");
// for (i = 0; i < MAX_DATA_CHUNKS; i++) {
// printk(KERN_DEBUG ">>>\tslot: %i; len: %u\n", i, chunks[i].iov_len);
// }
align_frame(dpriv);
// printk(KERN_DEBUG ">>> aligned frame dump:\n");
// for (i = 0; i < MAX_DATA_CHUNKS; i++) {
// printk(KERN_DEBUG ">>>\tslot: %i; len: %u\n", i, chunks[i].iov_len);
// }
// if (check_chunks(chunks) != 0) {
// dont_process = 1;
// return -EINVAL;
// }
/* new command slot is ready now and can be unlocked */
dpriv->flags &= ~LOCK_TAIL;
if ((dpriv->flags & PROC_CMD) == 0) {
if (!proceed) {
/* disk may be free by the moment, try to grab it */
spin_lock_irqsave(&dpriv->flags_lock, irq_flags);
if ((dpriv->flags & DISK_BUSY) == 0) {
dpriv->flags |= DISK_BUSY;
proceed = true;
}
spin_unlock_irqrestore(&dpriv->flags_lock, irq_flags);
}
printk(KERN_DEBUG ">>> flags before exec: %u, proceed = %d\n", dpriv->flags, proceed);
if ((dpriv->flags & PROC_CMD) == 0 && proceed) {
if (get_size_from(&dpriv->data_chunks[dpriv->head_ptr], 0, 0, EXCLUDE_REM) == 0)
move_head(dpriv);
process_cmd(dev, dpriv, port);
process_cmd(dpriv);
}
// while (dpriv->flags & PROC_CMD) {
// int sg_elems;
......@@ -1533,13 +1583,14 @@ static int elphel_qc_defer(struct ata_queued_cmd *qc)
return ret;
/* And now check if internal command is in progress */
printk_ratelimited(KERN_DEBUG ">>> %s: cmd queue head pos = %u, tail pos = %u, flags = %u\n", __func__, dpriv->head_ptr, dpriv->tail_ptr, dpriv->flags);
spin_lock_irqsave(&dpriv->flags_lock, irq_flags);
if (dpriv->flags & INTERNAL_CMD) {
if ((dpriv->flags & DISK_BUSY) || is_cmdq_empty(dpriv) == 0) {
ret = ATA_DEFER_LINK;
printk_ratelimited(KERN_DEBUG "defer system command\n");
} else {
printk_ratelimited(KERN_DEBUG "proceed to system command\n");
dpriv->flags |= NATIVE_CMD;
dpriv->flags |= DISK_BUSY;
}
spin_unlock_irqrestore(&dpriv->flags_lock, irq_flags);
......
src/drivers/ata/ahci_elphel.h
View file @ 0e242816
......@@ -24,22 +24,20 @@
#ifndef _AHCI_ELPHEL_EXT
#define _AHCI_ELPHEL_EXT
/** Flag indicating that IRQ should not be processed in ahci_handle_port_interrupt */
/** Flag indicating that IRQ corresponds to internal command and should not be
* processed in ahci_handle_port_interrupt */
#define IRQ_SIMPLE (1 << 0)
/** Flag indicating that driver's internal command is in progress. This flag is set when
* driver performs write on its own and is mutually exclusive with #NATIVE_CMD flag */
#define INTERNAL_CMD (1 << 1)
/** Flag indicating that libahci command is in progress. This flag is set when
* driver performs write from the system calls and is mutually exclusive with #INTERNAL_CMD flag */
#define NATIVE_CMD (1 << 2)
/** Flag indicating that disk is currently busy. Access to this flag should be protected by
* spin locks to prevent race conditions */
#define DISK_BUSY (1 << 1)
/** Processing driver's internal command is in progress */
#define PROC_CMD (1 << 3)
/** Flag indicating that the remaining chunk of data will be recorder */
#define LAST_BLOCK (1 << 4)
#define PROC_CMD (1 << 2)
/** Flag indicating that the remaining chunk of data will be recorded */
#define LAST_BLOCK (1 << 3)
/** Flag indicating that recording should be stopped right after the last chunk of data
* is written */
#define DELAYED_FINISH (1 << 5)
#define LOCK_TAIL (1 << 6)
#define DELAYED_FINISH (1 << 4)
#define LOCK_TAIL (1 << 5)
/** The length of a command FIS in double words */
#define CMD_FIS_LEN 5
/** This is used to get 28-bit address from 64-bit value */
......@@ -125,12 +123,12 @@ struct elphel_ahci_priv {
size_t curr_data_offset; ///< offset of the last byte in a data chunk pointed to by @e curr_data_chunk
size_t head_ptr; ///< pointer to command slot which will be written next
size_t tail_ptr; ///< pointer to next free command slot
spinlock_t flags_lock; ///< controls access to two flags in @e flags variable:
///< #INTERNAL_CMD and #NATIVE_CMD. These flags control access to disk
///< write operations either from the the driver itself or from the system and
///< thus are mutually exclusive. Only flags set operation should be
///> protected by this spin lock. Mutex is not used because these flags are accessed
///> from interrupt context
spinlock_t flags_lock; ///< controls access to #DISK_BUSY flag in @e flags variable.
///< This flag controls access to disk write operations either from
///< the the driver itself or from the system. Mutex is not used
///< because this flag is accessed from interrupt context
struct tasklet_struct bh; ///< command processing tasklet
struct device *dev; ///< pointer to parent device structure
};
#endif /* _AHCI_ELPHEL_EXT */
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment