Index: linux-2.6.9-5.0.3.EL/drivers/scsi/Kconfig =================================================================== Index: linux-2.6.9/drivers/scsi/Kconfig =================================================================== --- linux-2.6.9.orig/drivers/scsi/Kconfig 2007-07-23 14:19:13.000000000 +0400 +++ linux-2.6.9/drivers/scsi/Kconfig 2007-07-26 14:16:36.000000000 +0400 @@ -61,6 +61,14 @@ config SCSI_DUMP help SCSI dump support +config SD_IOSTATS + bool "Enable SCSI disk I/O stats" + depends on BLK_DEV_SD + default y + ---help--- + This enables SCSI disk I/O stats collection. You must also enable + /proc file system support if you want this feature. + config CHR_DEV_ST tristate "SCSI tape support" depends on SCSI Index: linux-2.6.9/drivers/scsi/scsi_proc.c =================================================================== --- linux-2.6.9.orig/drivers/scsi/scsi_proc.c 2007-03-13 02:47:28.000000000 +0300 +++ linux-2.6.9/drivers/scsi/scsi_proc.c 2007-07-26 14:16:36.000000000 +0400 @@ -38,7 +38,8 @@ /* 4K page size, but our output routines, use some slack for overruns */ #define PROC_BLOCK_SIZE (3*1024) -static struct proc_dir_entry *proc_scsi; +struct proc_dir_entry *proc_scsi; +EXPORT_SYMBOL(proc_scsi); /* Protect sht->present and sht->proc_dir */ static DECLARE_MUTEX(global_host_template_sem); Index: linux-2.6.9/drivers/scsi/sd.c =================================================================== --- linux-2.6.9.orig/drivers/scsi/sd.c 2007-03-13 02:47:27.000000000 +0300 +++ linux-2.6.9/drivers/scsi/sd.c 2007-07-28 14:55:56.000000000 +0400 @@ -63,6 +63,67 @@ #include "scsi_logging.h" +#if (defined(CONFIG_SD_IOSTATS) && defined(CONFIG_PROC_FS)) +# include +# include + +typedef struct { + unsigned long long iostat_size; + unsigned long long iostat_count; +} iostat_counter_t; + +#define IOSTAT_NCOUNTERS 16 +typedef struct { + iostat_counter_t iostat_read_histogram[IOSTAT_NCOUNTERS]; + iostat_counter_t iostat_write_histogram[IOSTAT_NCOUNTERS]; + struct timeval iostat_timeval; + + /* queue depth: how well the pipe is filled up */ + unsigned long long iostat_queue_ticks[IOSTAT_NCOUNTERS]; + unsigned long long iostat_queue_ticks_sum; + unsigned long iostat_queue_depth; + unsigned long iostat_queue_stamp; + + /* seeks: how linear the traffic is */ + unsigned long long iostat_next_sector; + unsigned long long iostat_seek_sectors; + unsigned long long iostat_seeks; + unsigned long long iostat_sectors; + unsigned long long iostat_reqs; + unsigned long iostat_read_reqs; + unsigned long iostat_write_reqs; + + /* process time: how long it takes to process requests */ + unsigned long iostat_rtime[IOSTAT_NCOUNTERS]; + unsigned long iostat_wtime[IOSTAT_NCOUNTERS]; + + /* queue time: how long process spent in elevator's queue */ + unsigned long iostat_rtime_in_queue[IOSTAT_NCOUNTERS]; + unsigned long iostat_wtime_in_queue[IOSTAT_NCOUNTERS]; + + char iostat_name[32]; + + /* must be the last field, as it's used to know size to be memset'ed */ + spinlock_t iostat_lock; +} ____cacheline_aligned_in_smp iostat_stats_t; + +iostat_stats_t **sd_iostats; +struct proc_dir_entry *sd_iostats_procdir; +char sd_iostats_procdir_name[] = "sd_iostats"; + +extern void sd_iostats_init(void); +extern void sd_iostats_init_disk(struct gendisk *); +extern void sd_iostats_fini(void); +void sd_iostats_start_req(struct scsi_cmnd *SCpnt); +void sd_iostats_finish_req(struct scsi_cmnd *SCpnt); +#else +static inline void sd_iostats_init(void) {} +static inline void sd_iostats_init_disk(struct gendisk *disk) {} +static inline void sd_iostats_fini(void) {} +static inline void sd_iostats_start_req(struct scsi_cmnd *SCpnt) {} +static inline void sd_iostats_finish_req(struct scsi_cmnd *SCpnt) {} +#endif + /* * More than enough for everybody ;) The huge number of majors * is a leftover from 16bit dev_t days, we don't really need that @@ -76,6 +137,7 @@ */ #define SD_MAX_DISKS (((26 * 26) + 26 + 1) * 26) +#define SD_STATS 256 /* * Time out in seconds for disks and Magneto-opticals (which are slower). */ @@ -278,6 +340,8 @@ static int sd_init_command(struct scsi_c SCSI_LOG_HLQUEUE(2, printk("%s : block=%llu\n", disk->disk_name, (unsigned long long)block)); + sd_iostats_start_req(SCpnt); + /* * If we have a 1K hardware sectorsize, prevent access to single * 512 byte sectors. In theory we could handle this - in fact @@ -474,6 +538,7 @@ static int sd_open(struct inode *inode, scsi_set_medium_removal(sdev, SCSI_REMOVAL_PREVENT); } + sd_iostats_init_disk(disk); return 0; error_out: @@ -849,6 +914,9 @@ static void sd_rw_intr(struct scsi_cmnd break; } } + + sd_iostats_finish_req(SCpnt); + /* * This calls the generic completion function, now that we know * how many actual sectors finished, and how many sectors we need @@ -1575,6 +1643,481 @@ static void sd_shutdown(struct device *d sd_sync_cache(sdp); } +#if (defined(CONFIG_SD_IOSTATS) && defined(CONFIG_PROC_FS)) +static int +sd_iostats_seq_show(struct seq_file *seq, void *v) +{ + struct timeval now; + struct gendisk *disk = seq->private; + iostat_stats_t *stats; + unsigned long long read_len; + unsigned long long read_len_tot; + unsigned long read_num; + unsigned long read_num_tot; + unsigned long long write_len; + unsigned long long write_len_tot; + unsigned long write_num; + unsigned long write_num_tot; + int i; + int maxi; + + if (sd_iostats == NULL) { + printk(KERN_ERR "sd_iostats_seq_show: NULL stats array\n"); + BUG(); + } + + stats = sd_iostats[scsi_disk(disk)->index]; + if (stats == NULL) { + printk(KERN_ERR "sd_iostats_seq_show: NULL stats entry\n"); + BUG(); + } + + do_gettimeofday(&now); + now.tv_sec -= stats->iostat_timeval.tv_sec; + now.tv_usec -= stats->iostat_timeval.tv_usec; + if (now.tv_usec < 0) { + now.tv_usec += 1000000; + now.tv_sec--; + } + + /* this sampling races with updates */ + seq_printf(seq, "index: %lu snapshot_time: %lu.%06lu\n", + (unsigned long) scsi_disk(disk)->index, + now.tv_sec, now.tv_usec); + + for (i = IOSTAT_NCOUNTERS - 1; i > 0; i--) + if (stats->iostat_read_histogram[i].iostat_count != 0 || + stats->iostat_write_histogram[i].iostat_count != 0) + break; + maxi = i; + + seq_printf(seq, "%8s %8s %12s %8s %12s\n", "size", + "reads", "total", "writes", "total"); + + read_len_tot = write_len_tot = 0; + read_num_tot = write_num_tot = 0; + for (i = 0; i <= maxi; i++) { + read_len = stats->iostat_read_histogram[i].iostat_size; + read_len_tot += read_len; + read_num = stats->iostat_read_histogram[i].iostat_count; + read_num_tot += read_num; + + write_len = stats->iostat_write_histogram[i].iostat_size; + write_len_tot += write_len; + write_num = stats->iostat_write_histogram[i].iostat_count; + write_num_tot += write_num; + + seq_printf (seq, "%8d %8lu %12llu %8lu %12llu\n", + 512<iostat_queue_ticks[i]; + if (ticks == 0) + continue; + percent = stats->iostat_queue_ticks[i] * 100; + do_div(percent, stats->iostat_queue_ticks_sum); + seq_printf(seq, "%8d %8llu %8llu\n", i, ticks, percent); + } + + if (stats->iostat_reqs != 0) { + unsigned long long aveseek = 0, percent = 0; + + if (stats->iostat_seeks) { + aveseek = stats->iostat_seek_sectors; + do_div(aveseek, stats->iostat_seeks); + percent = stats->iostat_seeks * 100; + do_div(percent, stats->iostat_reqs); + } + + seq_printf(seq, "\n%llu sectors in %llu reqs: %llu seek(s) over " + "%llu sectors in ave, %llu%% of all reqs\n", + stats->iostat_sectors, stats->iostat_reqs, + stats->iostat_seeks, aveseek, percent); + } + + seq_printf(seq, "\n%16s %8s %8s %8s %8s\n", "process time", "reads", + "%%", "writes", "%%"); + for (i = 0; i < IOSTAT_NCOUNTERS; i++) { + unsigned long read_percent = 0, write_percent = 0; + if (stats->iostat_wtime[i] == 0 && + stats->iostat_rtime[i] == 0) + continue; + if (stats->iostat_read_reqs) + read_percent = stats->iostat_rtime[i] * 100 / + stats->iostat_read_reqs; + if (stats->iostat_write_reqs) + write_percent = stats->iostat_wtime[i] * 100 / + stats->iostat_write_reqs; + seq_printf(seq, "%16u %8lu %8lu %8lu %8lu\n", + jiffies_to_msecs(((1UL << i) >> 1) << 1), + stats->iostat_rtime[i], read_percent, + stats->iostat_wtime[i], write_percent); + } + + seq_printf(seq, "\n%16s %8s %8s %8s %8s\n", "time in queue", "reads", + "%%", "writes", "%%"); + for (i = 0; i < IOSTAT_NCOUNTERS; i++) { + unsigned long read_percent = 0, write_percent = 0; + if (stats->iostat_wtime_in_queue[i] == 0 && + stats->iostat_rtime_in_queue[i] == 0) + continue; + if (stats->iostat_read_reqs) + read_percent = stats->iostat_rtime_in_queue[i] * 100 / + stats->iostat_read_reqs; + if (stats->iostat_write_reqs) + write_percent = stats->iostat_wtime_in_queue[i] * 100 / + stats->iostat_write_reqs; + seq_printf(seq, "%16u %8lu %8lu %8lu %8lu\n", + jiffies_to_msecs(((1UL << i) >> 1) << 1), + stats->iostat_rtime_in_queue[i], + read_percent, + stats->iostat_wtime_in_queue[i], + write_percent); + } + + return 0; +} + +static void * +sd_iostats_seq_start(struct seq_file *p, loff_t *pos) +{ + return (*pos == 0) ? (void *)1 : NULL; +} + +static void * +sd_iostats_seq_next(struct seq_file *p, void *v, loff_t *pos) +{ + ++*pos; + return NULL; +} + +static void +sd_iostats_seq_stop(struct seq_file *p, void *v) +{ +} + +static struct seq_operations sd_iostats_seqops = { + .start = sd_iostats_seq_start, + .stop = sd_iostats_seq_stop, + .next = sd_iostats_seq_next, + .show = sd_iostats_seq_show, +}; + +static int +sd_iostats_seq_open (struct inode *inode, struct file *file) +{ + int rc; + + rc = seq_open(file, &sd_iostats_seqops); + if (rc != 0) + return rc; + + ((struct seq_file *)file->private_data)->private = PDE(inode)->data; + return 0; +} + +static ssize_t +sd_iostats_seq_write(struct file *file, const char *buffer, + size_t len, loff_t *off) +{ + struct seq_file *seq = file->private_data; + struct gendisk *disk = seq->private; + iostat_stats_t *stats = sd_iostats[scsi_disk(disk)->index]; + unsigned long flags; + unsigned long qdepth; + + + spin_lock_irqsave (&stats->iostat_lock, flags); + qdepth = stats->iostat_queue_depth; + memset (stats, 0, offsetof(iostat_stats_t, iostat_lock)); + do_gettimeofday(&stats->iostat_timeval); + stats->iostat_queue_stamp = jiffies; + stats->iostat_queue_depth = qdepth; + spin_unlock_irqrestore (&stats->iostat_lock, flags); + + return len; +} + +static struct file_operations sd_iostats_proc_fops = { + .owner = THIS_MODULE, + .open = sd_iostats_seq_open, + .read = seq_read, + .write = sd_iostats_seq_write, + .llseek = seq_lseek, + .release = seq_release, +}; + +extern struct proc_dir_entry *proc_scsi; + +void +sd_iostats_init(void) +{ + int i; + + sd_iostats = kmalloc(SD_STATS * sizeof(iostat_stats_t *), GFP_KERNEL); + if (sd_iostats == NULL) { + printk(KERN_WARNING "Can't keep sd iostats: " + "ENOMEM allocating stats array size %d\n", + SD_STATS * sizeof(iostat_stats_t *)); + return; + } + + for (i = 0; i < SD_STATS; i++) + sd_iostats[i] = NULL; + + if (proc_scsi == NULL) { + printk(KERN_WARNING "No access to sd iostats: " + "proc_scsi is NULL\n"); + return; + } + + sd_iostats_procdir = create_proc_entry(sd_iostats_procdir_name, + S_IFDIR | S_IRUGO | S_IXUGO, + proc_scsi); + if (sd_iostats_procdir == NULL) { + printk(KERN_WARNING "No access to sd iostats: " + "can't create /proc/scsi/%s\n", sd_iostats_procdir_name); + return; + } +} + +void +sd_iostats_init_disk(struct gendisk *disk) +{ + struct proc_dir_entry *pde; + unsigned long flags; + iostat_stats_t *stats; + + if (sd_iostats == NULL || sd_iostats_procdir == NULL) + return; + + if (scsi_disk(disk)->index > SD_STATS) { + printk(KERN_ERR "sd_iostats_init_disk: " + "unexpected disk index %d(%d)\n", + scsi_disk(disk)->index, SD_STATS); + return; + } + + if (sd_iostats[scsi_disk(disk)->index] != NULL) + return; + + stats = kmalloc(sizeof(*stats), GFP_KERNEL); + if (stats == NULL) { + printk(KERN_WARNING "Can't keep %s iostats: " + "ENOMEM allocating stats size %d\n", + disk->disk_name, sizeof(*stats)); + return; + } + + memset (stats, 0, sizeof(*stats)); + do_gettimeofday(&stats->iostat_timeval); + stats->iostat_queue_stamp = jiffies; + spin_lock_init(&stats->iostat_lock); + + + spin_lock_irqsave(&stats->iostat_lock, flags); + + if (sd_iostats[scsi_disk(disk)->index] != NULL) { + spin_unlock_irqrestore(&stats->iostat_lock, flags); + kfree (stats); + return; + } + + sd_iostats[scsi_disk(disk)->index] = stats; + + spin_unlock_irqrestore(&stats->iostat_lock, flags); + + strncpy(stats->iostat_name, disk->disk_name, + sizeof(stats->iostat_name)-1); + + pde = create_proc_entry(stats->iostat_name, S_IRUGO | S_IWUSR, + sd_iostats_procdir); + if (pde == NULL) { + printk(KERN_WARNING "Can't create /proc/scsi/%s/%s\n", + sd_iostats_procdir_name, disk->disk_name); + } else { + pde->proc_fops = &sd_iostats_proc_fops; + pde->data = disk; + } +} + +void sd_iostats_fini(void) +{ + int i; + + if (sd_iostats == NULL) + return; + + for (i = 0; i < SD_STATS; i++) { + if (sd_iostats[i] == NULL) + continue; + if (sd_iostats_procdir != NULL) + remove_proc_entry(sd_iostats[i]->iostat_name, + sd_iostats_procdir); + kfree(sd_iostats[i]); + } + + if (proc_scsi != NULL && sd_iostats_procdir != NULL) + remove_proc_entry(sd_iostats_procdir_name, proc_scsi); + + sd_iostats_procdir = NULL; + kfree(sd_iostats); + sd_iostats = NULL; +} + +void sd_iostats_finish_req(struct scsi_cmnd *SCpnt) +{ + struct request *rq = SCpnt->request; + iostat_stats_t *stats; + unsigned long *tcounter; + int tbucket; + int tmp; + unsigned long irqflags; + int disk, i; + + disk = scsi_disk(rq->rq_disk)->index; + + if (sd_iostats == NULL) + return; + + if (disk < 0 || disk >= SD_STATS) { + printk(KERN_ERR "sd_iostats_bump: unexpected disk index " + "%d([0-%d])\n", disk, SD_STATS); + BUG(); + } + + stats = sd_iostats[disk]; + if (stats == NULL) + return; + + tmp = jiffies - rq->start_time; + for (tbucket = 0; tmp > 1; tbucket++) + tmp >>= 1; + if (tbucket >= IOSTAT_NCOUNTERS) + tbucket = IOSTAT_NCOUNTERS - 1; + //printk("%u ticks in D to %u\n", jiffies - rq->start_time, tbucket); + + tcounter = rq_data_dir(rq) == WRITE ? + &stats->iostat_wtime[tbucket] : &stats->iostat_rtime[tbucket]; + + spin_lock_irqsave(&stats->iostat_lock, irqflags); + + /* update delay stats */ + (*tcounter)++; + + /* update queue depth stats */ + i = stats->iostat_queue_depth; + if (i >= IOSTAT_NCOUNTERS) + i = IOSTAT_NCOUNTERS - 1; + stats->iostat_queue_ticks[i] += jiffies - stats->iostat_queue_stamp; + stats->iostat_queue_ticks_sum += jiffies - stats->iostat_queue_stamp; + stats->iostat_queue_depth--; + + /* update seek stats. XXX: not sure about nr_sectors */ + stats->iostat_sectors += rq->nr_sectors; + stats->iostat_reqs++; + if (rq->sector != stats->iostat_next_sector) { + stats->iostat_seek_sectors += + rq->sector > stats->iostat_next_sector ? + rq->sector - stats->iostat_next_sector : + stats->iostat_next_sector - rq->sector; + stats->iostat_seeks++; + } + stats->iostat_next_sector = rq->sector + rq->nr_sectors; + + stats->iostat_queue_stamp = jiffies; + + spin_unlock_irqrestore(&stats->iostat_lock, irqflags); +} + +void sd_iostats_start_req(struct scsi_cmnd *SCpnt) +{ + struct request *rq = SCpnt->request; + iostat_stats_t *stats; + iostat_counter_t *counter; + int bucket; + int tbucket; + int tmp; + unsigned long irqflags; + int disk, i; + int nsect; + + disk = scsi_disk(rq->rq_disk)->index; + + if (sd_iostats == NULL) + return; + + if (disk < 0 || disk >= SD_STATS) { + printk(KERN_ERR "sd_iostats_bump: unexpected disk index %d([0-%d])\n", + disk, SD_STATS); + BUG(); + } + + stats = sd_iostats[disk]; + if (stats == NULL) + return; + + nsect = SCpnt->request_bufflen >> 9; + for (bucket = 0, tmp = nsect; tmp > 1; bucket++) + tmp >>= 1; + + if (bucket >= IOSTAT_NCOUNTERS) { + printk (KERN_ERR "sd_iostats_bump: nsect %d too big\n", nsect); + BUG(); + } + + counter = rq_data_dir(rq) == WRITE ? + &stats->iostat_write_histogram[bucket] : + &stats->iostat_read_histogram[bucket]; + + tmp = jiffies - rq->start_time; + for (tbucket = 0; tmp > 1; tbucket++) + tmp >>= 1; + if (tbucket >= IOSTAT_NCOUNTERS) + tbucket = IOSTAT_NCOUNTERS - 1; + //printk("%u ticks in Q to %u\n", jiffies - rq->start_time, tbucket); + + /* an ugly hack to know exact processing time. the right + * solution is to add one more field to struct request + * hopefully it will break nothing ... */ + rq->start_time = jiffies; + + spin_lock_irqsave(&stats->iostat_lock, irqflags); + + /* update queue depth stats */ + i = stats->iostat_queue_depth; + if (i >= IOSTAT_NCOUNTERS) + i = IOSTAT_NCOUNTERS - 1; + stats->iostat_queue_ticks[i] += jiffies - stats->iostat_queue_stamp; + stats->iostat_queue_ticks_sum += jiffies - stats->iostat_queue_stamp; + stats->iostat_queue_depth++; + + /* update delay stats */ + if (rq_data_dir(rq) == WRITE) { + stats->iostat_wtime_in_queue[tbucket]++; + stats->iostat_write_reqs++; + } else { + stats->iostat_rtime_in_queue[tbucket]++; + stats->iostat_read_reqs++; + } + + /* update size stats */ + counter->iostat_size += nsect; + counter->iostat_count++; + + stats->iostat_queue_stamp = jiffies; + + spin_unlock_irqrestore(&stats->iostat_lock, irqflags); +} +#endif + /** * init_sd - entry point for this driver (both when built in or when * a module). @@ -1584,6 +2127,7 @@ static void sd_shutdown(struct device *d static int __init init_sd(void) { int majors = 0, i; + int rc = 0; SCSI_LOG_HLQUEUE(3, printk("init_sd: sd driver entry point\n")); @@ -1594,7 +2138,10 @@ static int __init init_sd(void) if (!majors) return -ENODEV; - return scsi_register_driver(&sd_template.gendrv); + rc = scsi_register_driver(&sd_template.gendrv); + if (rc == 0) + sd_iostats_init(); + return rc; } /** @@ -1608,6 +2155,7 @@ static void __exit exit_sd(void) SCSI_LOG_HLQUEUE(3, printk("exit_sd: exiting sd driver\n")); + sd_iostats_fini(); scsi_unregister_driver(&sd_template.gendrv); for (i = 0; i < SD_MAJORS; i++) unregister_blkdev(sd_major(i), "sd");