[PATCH] Block queue IO tracing support (blktrace) as of 2006-03-23
authorJens Axboe <axboe@suse.de>
Thu, 23 Mar 2006 19:00:26 +0000 (20:00 +0100)
committerJens Axboe <axboe@suse.de>
Thu, 23 Mar 2006 19:00:26 +0000 (20:00 +0100)
Signed-off-by: Jens Axboe <axboe@suse.de>
17 files changed:
block/Kconfig
block/Makefile
block/blktrace.c [new file with mode: 0644]
block/elevator.c
block/ioctl.c
block/ll_rw_blk.c
drivers/block/cciss.c
drivers/md/dm.c
fs/bio.c
fs/compat_ioctl.c
include/linux/blkdev.h
include/linux/blktrace_api.h [new file with mode: 0644]
include/linux/compat_ioctl.h
include/linux/fs.h
include/linux/sched.h
kernel/fork.c
mm/highmem.c

index 377f6dd20e1799afe57f9881e3cd8ac222697f8e..96783645092dc106cd4f4b300958aeae5b684e2a 100644 (file)
@@ -11,4 +11,16 @@ config LBD
          your machine, or if you want to have a raid or loopback device
          bigger than 2TB.  Otherwise say N.
 
+config BLK_DEV_IO_TRACE
+       bool "Support for tracing block io actions"
+       select RELAY
+       select DEBUG_FS
+       help
+         Say Y here, if you want to be able to trace the block layer actions
+         on a given queue. Tracing allows you to see any traffic happening
+         on a block device queue. For more information (and the user space
+         support tools needed), fetch the blktrace app from:
+
+         git://brick.kernel.dk/data/git/blktrace.git
+
 source block/Kconfig.iosched
index 7e4f93e2b44e9a87503a90bce18c246245cd040a..c05de0e0037fe6f14f9e6f508bece1271568bb36 100644 (file)
@@ -8,3 +8,5 @@ obj-$(CONFIG_IOSCHED_NOOP)      += noop-iosched.o
 obj-$(CONFIG_IOSCHED_AS)       += as-iosched.o
 obj-$(CONFIG_IOSCHED_DEADLINE) += deadline-iosched.o
 obj-$(CONFIG_IOSCHED_CFQ)      += cfq-iosched.o
+
+obj-$(CONFIG_BLK_DEV_IO_TRACE) += blktrace.o
diff --git a/block/blktrace.c b/block/blktrace.c
new file mode 100644 (file)
index 0000000..36f3a17
--- /dev/null
@@ -0,0 +1,538 @@
+/*
+ * Copyright (C) 2006 Jens Axboe <axboe@suse.de>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that 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/config.h>
+#include <linux/kernel.h>
+#include <linux/blkdev.h>
+#include <linux/blktrace_api.h>
+#include <linux/percpu.h>
+#include <linux/init.h>
+#include <linux/mutex.h>
+#include <linux/debugfs.h>
+#include <asm/uaccess.h>
+
+static DEFINE_PER_CPU(unsigned long long, blk_trace_cpu_offset) = { 0, };
+static unsigned int blktrace_seq __read_mostly = 1;
+
+/*
+ * Send out a notify for this process, if we haven't done so since a trace
+ * started
+ */
+static void trace_note_tsk(struct blk_trace *bt, struct task_struct *tsk)
+{
+       struct blk_io_trace *t;
+
+       t = relay_reserve(bt->rchan, sizeof(*t) + sizeof(tsk->comm));
+       if (t) {
+               t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
+               t->device = bt->dev;
+               t->action = BLK_TC_ACT(BLK_TC_NOTIFY);
+               t->pid = tsk->pid;
+               t->cpu = smp_processor_id();
+               t->pdu_len = sizeof(tsk->comm);
+               memcpy((void *) t + sizeof(*t), tsk->comm, t->pdu_len);
+               tsk->btrace_seq = blktrace_seq;
+       }
+}
+
+static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
+                        pid_t pid)
+{
+       if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
+               return 1;
+       if (sector < bt->start_lba || sector > bt->end_lba)
+               return 1;
+       if (bt->pid && pid != bt->pid)
+               return 1;
+
+       return 0;
+}
+
+/*
+ * Data direction bit lookup
+ */
+static u32 ddir_act[2] __read_mostly = { BLK_TC_ACT(BLK_TC_READ), BLK_TC_ACT(BLK_TC_WRITE) };
+
+/*
+ * Bio action bits of interest
+ */
+static u32 bio_act[3] __read_mostly = { 0, BLK_TC_ACT(BLK_TC_BARRIER), BLK_TC_ACT(BLK_TC_SYNC) };
+
+/*
+ * More could be added as needed, taking care to increment the decrementer
+ * to get correct indexing
+ */
+#define trace_barrier_bit(rw)  \
+       (((rw) & (1 << BIO_RW_BARRIER)) >> (BIO_RW_BARRIER - 0))
+#define trace_sync_bit(rw)     \
+       (((rw) & (1 << BIO_RW_SYNC)) >> (BIO_RW_SYNC - 1))
+
+/*
+ * The worker for the various blk_add_trace*() types. Fills out a
+ * blk_io_trace structure and places it in a per-cpu subbuffer.
+ */
+void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
+                    int rw, u32 what, int error, int pdu_len, void *pdu_data)
+{
+       struct task_struct *tsk = current;
+       struct blk_io_trace *t;
+       unsigned long flags;
+       unsigned long *sequence;
+       pid_t pid;
+       int cpu;
+
+       if (unlikely(bt->trace_state != Blktrace_running))
+               return;
+
+       what |= ddir_act[rw & WRITE];
+       what |= bio_act[trace_barrier_bit(rw)];
+       what |= bio_act[trace_sync_bit(rw)];
+
+       pid = tsk->pid;
+       if (unlikely(act_log_check(bt, what, sector, pid)))
+               return;
+
+       /*
+        * A word about the locking here - we disable interrupts to reserve
+        * some space in the relay per-cpu buffer, to prevent an irq
+        * from coming in and stepping on our toes. Once reserved, it's
+        * enough to get preemption disabled to prevent read of this data
+        * before we are through filling it. get_cpu()/put_cpu() does this
+        * for us
+        */
+       local_irq_save(flags);
+
+       if (unlikely(tsk->btrace_seq != blktrace_seq))
+               trace_note_tsk(bt, tsk);
+
+       t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
+       if (t) {
+               cpu = smp_processor_id();
+               sequence = per_cpu_ptr(bt->sequence, cpu);
+
+               t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
+               t->sequence = ++(*sequence);
+               t->time = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
+               t->sector = sector;
+               t->bytes = bytes;
+               t->action = what;
+               t->pid = pid;
+               t->device = bt->dev;
+               t->cpu = cpu;
+               t->error = error;
+               t->pdu_len = pdu_len;
+
+               if (pdu_len)
+                       memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
+       }
+
+       local_irq_restore(flags);
+}
+
+EXPORT_SYMBOL_GPL(__blk_add_trace);
+
+static struct dentry *blk_tree_root;
+static struct mutex blk_tree_mutex;
+static unsigned int root_users;
+
+static inline void blk_remove_root(void)
+{
+       if (blk_tree_root) {
+               debugfs_remove(blk_tree_root);
+               blk_tree_root = NULL;
+       }
+}
+
+static void blk_remove_tree(struct dentry *dir)
+{
+       mutex_lock(&blk_tree_mutex);
+       debugfs_remove(dir);
+       if (--root_users == 0)
+               blk_remove_root();
+       mutex_unlock(&blk_tree_mutex);
+}
+
+static struct dentry *blk_create_tree(const char *blk_name)
+{
+       struct dentry *dir = NULL;
+
+       mutex_lock(&blk_tree_mutex);
+
+       if (!blk_tree_root) {
+               blk_tree_root = debugfs_create_dir("block", NULL);
+               if (!blk_tree_root)
+                       goto err;
+       }
+
+       dir = debugfs_create_dir(blk_name, blk_tree_root);
+       if (dir)
+               root_users++;
+       else
+               blk_remove_root();
+
+err:
+       mutex_unlock(&blk_tree_mutex);
+       return dir;
+}
+
+static void blk_trace_cleanup(struct blk_trace *bt)
+{
+       relay_close(bt->rchan);
+       debugfs_remove(bt->dropped_file);
+       blk_remove_tree(bt->dir);
+       free_percpu(bt->sequence);
+       kfree(bt);
+}
+
+static int blk_trace_remove(request_queue_t *q)
+{
+       struct blk_trace *bt;
+
+       bt = xchg(&q->blk_trace, NULL);
+       if (!bt)
+               return -EINVAL;
+
+       if (bt->trace_state == Blktrace_setup ||
+           bt->trace_state == Blktrace_stopped)
+               blk_trace_cleanup(bt);
+
+       return 0;
+}
+
+static int blk_dropped_open(struct inode *inode, struct file *filp)
+{
+       filp->private_data = inode->u.generic_ip;
+
+       return 0;
+}
+
+static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
+                               size_t count, loff_t *ppos)
+{
+       struct blk_trace *bt = filp->private_data;
+       char buf[16];
+
+       snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
+
+       return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
+}
+
+static struct file_operations blk_dropped_fops = {
+       .owner =        THIS_MODULE,
+       .open =         blk_dropped_open,
+       .read =         blk_dropped_read,
+};
+
+/*
+ * Keep track of how many times we encountered a full subbuffer, to aid
+ * the user space app in telling how many lost events there were.
+ */
+static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
+                                    void *prev_subbuf, size_t prev_padding)
+{
+       struct blk_trace *bt;
+
+       if (!relay_buf_full(buf))
+               return 1;
+
+       bt = buf->chan->private_data;
+       atomic_inc(&bt->dropped);
+       return 0;
+}
+
+static int blk_remove_buf_file_callback(struct dentry *dentry)
+{
+       debugfs_remove(dentry);
+       return 0;
+}
+
+static struct dentry *blk_create_buf_file_callback(const char *filename,
+                                                  struct dentry *parent,
+                                                  int mode,
+                                                  struct rchan_buf *buf,
+                                                  int *is_global)
+{
+       return debugfs_create_file(filename, mode, parent, buf,
+                                       &relay_file_operations);
+}
+
+static struct rchan_callbacks blk_relay_callbacks = {
+       .subbuf_start           = blk_subbuf_start_callback,
+       .create_buf_file        = blk_create_buf_file_callback,
+       .remove_buf_file        = blk_remove_buf_file_callback,
+};
+
+/*
+ * Setup everything required to start tracing
+ */
+static int blk_trace_setup(request_queue_t *q, struct block_device *bdev,
+                          char __user *arg)
+{
+       struct blk_user_trace_setup buts;
+       struct blk_trace *old_bt, *bt = NULL;
+       struct dentry *dir = NULL;
+       char b[BDEVNAME_SIZE];
+       int ret, i;
+
+       if (copy_from_user(&buts, arg, sizeof(buts)))
+               return -EFAULT;
+
+       if (!buts.buf_size || !buts.buf_nr)
+               return -EINVAL;
+
+       strcpy(buts.name, bdevname(bdev, b));
+
+       /*
+        * some device names have larger paths - convert the slashes
+        * to underscores for this to work as expected
+        */
+       for (i = 0; i < strlen(buts.name); i++)
+               if (buts.name[i] == '/')
+                       buts.name[i] = '_';
+
+       if (copy_to_user(arg, &buts, sizeof(buts)))
+               return -EFAULT;
+
+       ret = -ENOMEM;
+       bt = kzalloc(sizeof(*bt), GFP_KERNEL);
+       if (!bt)
+               goto err;
+
+       bt->sequence = alloc_percpu(unsigned long);
+       if (!bt->sequence)
+               goto err;
+
+       ret = -ENOENT;
+       dir = blk_create_tree(buts.name);
+       if (!dir)
+               goto err;
+
+       bt->dir = dir;
+       bt->dev = bdev->bd_dev;
+       atomic_set(&bt->dropped, 0);
+
+       ret = -EIO;
+       bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, &blk_dropped_fops);
+       if (!bt->dropped_file)
+               goto err;
+
+       bt->rchan = relay_open("trace", dir, buts.buf_size, buts.buf_nr, &blk_relay_callbacks);
+       if (!bt->rchan)
+               goto err;
+       bt->rchan->private_data = bt;
+
+       bt->act_mask = buts.act_mask;
+       if (!bt->act_mask)
+               bt->act_mask = (u16) -1;
+
+       bt->start_lba = buts.start_lba;
+       bt->end_lba = buts.end_lba;
+       if (!bt->end_lba)
+               bt->end_lba = -1ULL;
+
+       bt->pid = buts.pid;
+       bt->trace_state = Blktrace_setup;
+
+       ret = -EBUSY;
+       old_bt = xchg(&q->blk_trace, bt);
+       if (old_bt) {
+               (void) xchg(&q->blk_trace, old_bt);
+               goto err;
+       }
+
+       return 0;
+err:
+       if (dir)
+               blk_remove_tree(dir);
+       if (bt) {
+               if (bt->dropped_file)
+                       debugfs_remove(bt->dropped_file);
+               if (bt->sequence)
+                       free_percpu(bt->sequence);
+               if (bt->rchan)
+                       relay_close(bt->rchan);
+               kfree(bt);
+       }
+       return ret;
+}
+
+static int blk_trace_startstop(request_queue_t *q, int start)
+{
+       struct blk_trace *bt;
+       int ret;
+
+       if ((bt = q->blk_trace) == NULL)
+               return -EINVAL;
+
+       /*
+        * For starting a trace, we can transition from a setup or stopped
+        * trace. For stopping a trace, the state must be running
+        */
+       ret = -EINVAL;
+       if (start) {
+               if (bt->trace_state == Blktrace_setup ||
+                   bt->trace_state == Blktrace_stopped) {
+                       blktrace_seq++;
+                       smp_mb();
+                       bt->trace_state = Blktrace_running;
+                       ret = 0;
+               }
+       } else {
+               if (bt->trace_state == Blktrace_running) {
+                       bt->trace_state = Blktrace_stopped;
+                       relay_flush(bt->rchan);
+                       ret = 0;
+               }
+       }
+
+       return ret;
+}
+
+/**
+ * blk_trace_ioctl: - handle the ioctls associated with tracing
+ * @bdev:      the block device
+ * @cmd:       the ioctl cmd
+ * @arg:       the argument data, if any
+ *
+ **/
+int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
+{
+       request_queue_t *q;
+       int ret, start = 0;
+
+       q = bdev_get_queue(bdev);
+       if (!q)
+               return -ENXIO;
+
+       mutex_lock(&bdev->bd_mutex);
+
+       switch (cmd) {
+       case BLKTRACESETUP:
+               ret = blk_trace_setup(q, bdev, arg);
+               break;
+       case BLKTRACESTART:
+               start = 1;
+       case BLKTRACESTOP:
+               ret = blk_trace_startstop(q, start);
+               break;
+       case BLKTRACETEARDOWN:
+               ret = blk_trace_remove(q);
+               break;
+       default:
+               ret = -ENOTTY;
+               break;
+       }
+
+       mutex_unlock(&bdev->bd_mutex);
+       return ret;
+}
+
+/**
+ * blk_trace_shutdown: - stop and cleanup trace structures
+ * @q:    the request queue associated with the device
+ *
+ **/
+void blk_trace_shutdown(request_queue_t *q)
+{
+       blk_trace_startstop(q, 0);
+       blk_trace_remove(q);
+}
+
+/*
+ * Average offset over two calls to sched_clock() with a gettimeofday()
+ * in the middle
+ */
+static void blk_check_time(unsigned long long *t)
+{
+       unsigned long long a, b;
+       struct timeval tv;
+
+       a = sched_clock();
+       do_gettimeofday(&tv);
+       b = sched_clock();
+
+       *t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
+       *t -= (a + b) / 2;
+}
+
+static void blk_trace_check_cpu_time(void *data)
+{
+       unsigned long long *t;
+       int cpu = get_cpu();
+
+       t = &per_cpu(blk_trace_cpu_offset, cpu);
+
+       /*
+        * Just call it twice, hopefully the second call will be cache hot
+        * and a little more precise
+        */
+       blk_check_time(t);
+       blk_check_time(t);
+
+       put_cpu();
+}
+
+/*
+ * Call blk_trace_check_cpu_time() on each CPU to calibrate our inter-CPU
+ * timings
+ */
+static void blk_trace_calibrate_offsets(void)
+{
+       unsigned long flags;
+
+       smp_call_function(blk_trace_check_cpu_time, NULL, 1, 1);
+       local_irq_save(flags);
+       blk_trace_check_cpu_time(NULL);
+       local_irq_restore(flags);
+}
+
+static void blk_trace_set_ht_offsets(void)
+{
+#if defined(CONFIG_SCHED_SMT)
+       int cpu, i;
+
+       /*
+        * now make sure HT siblings have the same time offset
+        */
+       preempt_disable();
+       for_each_online_cpu(cpu) {
+               unsigned long long *cpu_off, *sibling_off;
+
+               for_each_cpu_mask(i, cpu_sibling_map[cpu]) {
+                       if (i == cpu)
+                               continue;
+
+                       cpu_off = &per_cpu(blk_trace_cpu_offset, cpu);
+                       sibling_off = &per_cpu(blk_trace_cpu_offset, i);
+                       *sibling_off = *cpu_off;
+               }
+       }
+       preempt_enable();
+#endif
+}
+
+static __init int blk_trace_init(void)
+{
+       mutex_init(&blk_tree_mutex);
+       blk_trace_calibrate_offsets();
+       blk_trace_set_ht_offsets();
+
+       return 0;
+}
+
+module_init(blk_trace_init);
+
index db3d0d8296a05c908f5dd38100593217c8c7d114..5e558c4689a4f561117cf89b5c4867d045699ad8 100644 (file)
@@ -33,6 +33,7 @@
 #include <linux/init.h>
 #include <linux/compiler.h>
 #include <linux/delay.h>
+#include <linux/blktrace_api.h>
 
 #include <asm/uaccess.h>
 
@@ -333,6 +334,8 @@ void elv_insert(request_queue_t *q, struct request *rq, int where)
        struct list_head *pos;
        unsigned ordseq;
 
+       blk_add_trace_rq(q, rq, BLK_TA_INSERT);
+
        rq->q = q;
 
        switch (where) {
@@ -499,6 +502,7 @@ struct request *elv_next_request(request_queue_t *q)
                         * not be passed by new incoming requests
                         */
                        rq->flags |= REQ_STARTED;
+                       blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
                }
 
                if (!q->boundary_rq || q->boundary_rq == rq) {
index 35fdb7dc6512028be4a3d7e1e4012aaadf1b4ddf..9cfa2e1ecb24c2568e020222026271f68631deb7 100644 (file)
@@ -5,6 +5,7 @@
 #include <linux/backing-dev.h>
 #include <linux/buffer_head.h>
 #include <linux/smp_lock.h>
+#include <linux/blktrace_api.h>
 #include <asm/uaccess.h>
 
 static int blkpg_ioctl(struct block_device *bdev, struct blkpg_ioctl_arg __user *arg)
@@ -189,6 +190,11 @@ static int blkdev_locked_ioctl(struct file *file, struct block_device *bdev,
                return put_ulong(arg, bdev->bd_inode->i_size >> 9);
        case BLKGETSIZE64:
                return put_u64(arg, bdev->bd_inode->i_size);
+       case BLKTRACESTART:
+       case BLKTRACESTOP:
+       case BLKTRACESETUP:
+       case BLKTRACETEARDOWN:
+               return blk_trace_ioctl(bdev, cmd, (char __user *) arg);
        }
        return -ENOIOCTLCMD;
 }
index 6c793b196aa918da946d7378667f8a08dd7c4005..062067fa7ead0340ee3f5da91f82af80ec097e82 100644 (file)
@@ -28,6 +28,7 @@
 #include <linux/writeback.h>
 #include <linux/interrupt.h>
 #include <linux/cpu.h>
+#include <linux/blktrace_api.h>
 
 /*
  * for max sense size
@@ -1556,8 +1557,10 @@ void blk_plug_device(request_queue_t *q)
        if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
                return;
 
-       if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
+       if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
                mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
+               blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
+       }
 }
 
 EXPORT_SYMBOL(blk_plug_device);
@@ -1621,14 +1624,21 @@ static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
        /*
         * devices don't necessarily have an ->unplug_fn defined
         */
-       if (q->unplug_fn)
+       if (q->unplug_fn) {
+               blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
+                                       q->rq.count[READ] + q->rq.count[WRITE]);
+
                q->unplug_fn(q);
+       }
 }
 
 static void blk_unplug_work(void *data)
 {
        request_queue_t *q = data;
 
+       blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
+                               q->rq.count[READ] + q->rq.count[WRITE]);
+
        q->unplug_fn(q);
 }
 
@@ -1636,6 +1646,9 @@ static void blk_unplug_timeout(unsigned long data)
 {
        request_queue_t *q = (request_queue_t *)data;
 
+       blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
+                               q->rq.count[READ] + q->rq.count[WRITE]);
+
        kblockd_schedule_work(&q->unplug_work);
 }
 
@@ -1753,6 +1766,9 @@ static void blk_release_queue(struct kobject *kobj)
        if (q->queue_tags)
                __blk_queue_free_tags(q);
 
+       if (q->blk_trace)
+               blk_trace_shutdown(q);
+
        kmem_cache_free(requestq_cachep, q);
 }
 
@@ -2129,6 +2145,8 @@ rq_starved:
        
        rq_init(q, rq);
        rq->rl = rl;
+
+       blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
 out:
        return rq;
 }
@@ -2157,6 +2175,8 @@ static struct request *get_request_wait(request_queue_t *q, int rw,
                if (!rq) {
                        struct io_context *ioc;
 
+                       blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
+
                        __generic_unplug_device(q);
                        spin_unlock_irq(q->queue_lock);
                        io_schedule();
@@ -2210,6 +2230,8 @@ EXPORT_SYMBOL(blk_get_request);
  */
 void blk_requeue_request(request_queue_t *q, struct request *rq)
 {
+       blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
+
        if (blk_rq_tagged(rq))
                blk_queue_end_tag(q, rq);
 
@@ -2844,6 +2866,8 @@ static int __make_request(request_queue_t *q, struct bio *bio)
                        if (!q->back_merge_fn(q, req, bio))
                                break;
 
+                       blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
+
                        req->biotail->bi_next = bio;
                        req->biotail = bio;
                        req->nr_sectors = req->hard_nr_sectors += nr_sectors;
@@ -2859,6 +2883,8 @@ static int __make_request(request_queue_t *q, struct bio *bio)
                        if (!q->front_merge_fn(q, req, bio))
                                break;
 
+                       blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
+
                        bio->bi_next = req->bio;
                        req->bio = bio;
 
@@ -2976,6 +3002,7 @@ void generic_make_request(struct bio *bio)
        request_queue_t *q;
        sector_t maxsector;
        int ret, nr_sectors = bio_sectors(bio);
+       dev_t old_dev;
 
        might_sleep();
        /* Test device or partition size, when known. */
@@ -3002,6 +3029,8 @@ void generic_make_request(struct bio *bio)
         * NOTE: we don't repeat the blk_size check for each new device.
         * Stacking drivers are expected to know what they are doing.
         */
+       maxsector = -1;
+       old_dev = 0;
        do {
                char b[BDEVNAME_SIZE];
 
@@ -3034,6 +3063,15 @@ end_io:
                 */
                blk_partition_remap(bio);
 
+               if (maxsector != -1)
+                       blk_add_trace_remap(q, bio, old_dev, bio->bi_sector, 
+                                           maxsector);
+
+               blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
+
+               maxsector = bio->bi_sector;
+               old_dev = bio->bi_bdev->bd_dev;
+
                ret = q->make_request_fn(q, bio);
        } while (ret);
 }
@@ -3153,6 +3191,8 @@ static int __end_that_request_first(struct request *req, int uptodate,
        int total_bytes, bio_nbytes, error, next_idx = 0;
        struct bio *bio;
 
+       blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
+
        /*
         * extend uptodate bool to allow < 0 value to be direct io error
         */
index e29b8926f80ed08e1c618e236c8186720f0f1ea9..1f2890989b56086f7b175475bfb2cf3a6a6d48c8 100644 (file)
@@ -38,6 +38,7 @@
 #include <linux/hdreg.h>
 #include <linux/spinlock.h>
 #include <linux/compat.h>
+#include <linux/blktrace_api.h>
 #include <asm/uaccess.h>
 #include <asm/io.h>
 
@@ -2331,6 +2332,7 @@ static inline void complete_command( ctlr_info_t *h, CommandList_struct *cmd,
 
        cmd->rq->completion_data = cmd;
        cmd->rq->errors = status;
+       blk_add_trace_rq(cmd->rq->q, cmd->rq, BLK_TA_COMPLETE);
        blk_complete_request(cmd->rq);
 }
 
index 26b08ee425c74d6b0ce9ecee09a5d817a1e51841..8c82373f7ff31f24b8c0439e17c00d5a8873f30d 100644 (file)
@@ -17,6 +17,7 @@
 #include <linux/mempool.h>
 #include <linux/slab.h>
 #include <linux/idr.h>
+#include <linux/blktrace_api.h>
 
 static const char *_name = DM_NAME;
 
@@ -334,6 +335,8 @@ static void dec_pending(struct dm_io *io, int error)
                        /* nudge anyone waiting on suspend queue */
                        wake_up(&io->md->wait);
 
+               blk_add_trace_bio(io->md->queue, io->bio, BLK_TA_COMPLETE);
+
                bio_endio(io->bio, io->bio->bi_size, io->error);
                free_io(io->md, io);
        }
@@ -392,6 +395,7 @@ static void __map_bio(struct dm_target *ti, struct bio *clone,
                      struct target_io *tio)
 {
        int r;
+       sector_t sector;
 
        /*
         * Sanity checks.
@@ -407,10 +411,17 @@ static void __map_bio(struct dm_target *ti, struct bio *clone,
         * this io.
         */
        atomic_inc(&tio->io->io_count);
+       sector = clone->bi_sector;
        r = ti->type->map(ti, clone, &tio->info);
-       if (r > 0)
+       if (r > 0) {
                /* the bio has been remapped so dispatch it */
+
+               blk_add_trace_remap(bdev_get_queue(clone->bi_bdev), clone, 
+                                   tio->io->bio->bi_bdev->bd_dev, sector, 
+                                   clone->bi_sector);
+
                generic_make_request(clone);
+       }
 
        else if (r < 0) {
                /* error the io and bail out */
index 8f1d2e815c96b696a630ba154ae9c438e3fa3fb2..0a8c59cb68f527fd98152d12a10ef1eb2e7f049c 100644 (file)
--- a/fs/bio.c
+++ b/fs/bio.c
@@ -25,6 +25,7 @@
 #include <linux/module.h>
 #include <linux/mempool.h>
 #include <linux/workqueue.h>
+#include <linux/blktrace_api.h>
 #include <scsi/sg.h>           /* for struct sg_iovec */
 
 #define BIO_POOL_SIZE 256
@@ -1095,6 +1096,9 @@ struct bio_pair *bio_split(struct bio *bi, mempool_t *pool, int first_sectors)
        if (!bp)
                return bp;
 
+       blk_add_trace_pdu_int(bdev_get_queue(bi->bi_bdev), BLK_TA_SPLIT, bi,
+                               bi->bi_sector + first_sectors);
+
        BUG_ON(bi->bi_vcnt != 1);
        BUG_ON(bi->bi_idx != 0);
        atomic_set(&bp->cnt, 3);
index c666769a875d2fc0e80927caf6db038b036bc42b..7c031f00fd798ede5c3edf885c7d8a6fef51250b 100644 (file)
@@ -72,6 +72,7 @@
 #include <linux/i2c-dev.h>
 #include <linux/wireless.h>
 #include <linux/atalk.h>
+#include <linux/blktrace_api.h>
 
 #include <net/sock.h>          /* siocdevprivate_ioctl */
 #include <net/bluetooth/bluetooth.h>
index 56bb6a4e15f3b2afe80bb4dee0026f6fc1518852..c179966f1a2fea92faa912ffbf8133478ee9b30d 100644 (file)
@@ -22,6 +22,7 @@ typedef struct request_queue request_queue_t;
 struct elevator_queue;
 typedef struct elevator_queue elevator_t;
 struct request_pm_state;
+struct blk_trace;
 
 #define BLKDEV_MIN_RQ  4
 #define BLKDEV_MAX_RQ  128     /* Default maximum */
@@ -416,6 +417,8 @@ struct request_queue
        unsigned int            sg_reserved_size;
        int                     node;
 
+       struct blk_trace        *blk_trace;
+
        /*
         * reserved for flush operations
         */
diff --git a/include/linux/blktrace_api.h b/include/linux/blktrace_api.h
new file mode 100644 (file)
index 0000000..b34d3e7
--- /dev/null
@@ -0,0 +1,277 @@
+#ifndef BLKTRACE_H
+#define BLKTRACE_H
+
+#include <linux/config.h>
+#include <linux/blkdev.h>
+#include <linux/relay.h>
+
+/*
+ * Trace categories
+ */
+enum blktrace_cat {
+       BLK_TC_READ     = 1 << 0,       /* reads */
+       BLK_TC_WRITE    = 1 << 1,       /* writes */
+       BLK_TC_BARRIER  = 1 << 2,       /* barrier */
+       BLK_TC_SYNC     = 1 << 3,       /* barrier */
+       BLK_TC_QUEUE    = 1 << 4,       /* queueing/merging */
+       BLK_TC_REQUEUE  = 1 << 5,       /* requeueing */
+       BLK_TC_ISSUE    = 1 << 6,       /* issue */
+       BLK_TC_COMPLETE = 1 << 7,       /* completions */
+       BLK_TC_FS       = 1 << 8,       /* fs requests */
+       BLK_TC_PC       = 1 << 9,       /* pc requests */
+       BLK_TC_NOTIFY   = 1 << 10,      /* special message */
+
+       BLK_TC_END      = 1 << 15,      /* only 16-bits, reminder */
+};
+
+#define BLK_TC_SHIFT           (16)
+#define BLK_TC_ACT(act)                ((act) << BLK_TC_SHIFT)
+
+/*
+ * Basic trace actions
+ */
+enum blktrace_act {
+       __BLK_TA_QUEUE = 1,             /* queued */
+       __BLK_TA_BACKMERGE,             /* back merged to existing rq */
+       __BLK_TA_FRONTMERGE,            /* front merge to existing rq */
+       __BLK_TA_GETRQ,                 /* allocated new request */
+       __BLK_TA_SLEEPRQ,               /* sleeping on rq allocation */
+       __BLK_TA_REQUEUE,               /* request requeued */
+       __BLK_TA_ISSUE,                 /* sent to driver */
+       __BLK_TA_COMPLETE,              /* completed by driver */
+       __BLK_TA_PLUG,                  /* queue was plugged */
+       __BLK_TA_UNPLUG_IO,             /* queue was unplugged by io */
+       __BLK_TA_UNPLUG_TIMER,          /* queue was unplugged by timer */
+       __BLK_TA_INSERT,                /* insert request */
+       __BLK_TA_SPLIT,                 /* bio was split */
+       __BLK_TA_BOUNCE,                /* bio was bounced */
+       __BLK_TA_REMAP,                 /* bio was remapped */
+};
+
+/*
+ * Trace actions in full. Additionally, read or write is masked
+ */
+#define BLK_TA_QUEUE           (__BLK_TA_QUEUE | BLK_TC_ACT(BLK_TC_QUEUE))
+#define BLK_TA_BACKMERGE       (__BLK_TA_BACKMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
+#define BLK_TA_FRONTMERGE      (__BLK_TA_FRONTMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
+#define        BLK_TA_GETRQ            (__BLK_TA_GETRQ | BLK_TC_ACT(BLK_TC_QUEUE))
+#define        BLK_TA_SLEEPRQ          (__BLK_TA_SLEEPRQ | BLK_TC_ACT(BLK_TC_QUEUE))
+#define        BLK_TA_REQUEUE          (__BLK_TA_REQUEUE | BLK_TC_ACT(BLK_TC_REQUEUE))
+#define BLK_TA_ISSUE           (__BLK_TA_ISSUE | BLK_TC_ACT(BLK_TC_ISSUE))
+#define BLK_TA_COMPLETE                (__BLK_TA_COMPLETE| BLK_TC_ACT(BLK_TC_COMPLETE))
+#define BLK_TA_PLUG            (__BLK_TA_PLUG | BLK_TC_ACT(BLK_TC_QUEUE))
+#define BLK_TA_UNPLUG_IO       (__BLK_TA_UNPLUG_IO | BLK_TC_ACT(BLK_TC_QUEUE))
+#define BLK_TA_UNPLUG_TIMER    (__BLK_TA_UNPLUG_TIMER | BLK_TC_ACT(BLK_TC_QUEUE))
+#define BLK_TA_INSERT          (__BLK_TA_INSERT | BLK_TC_ACT(BLK_TC_QUEUE))
+#define BLK_TA_SPLIT           (__BLK_TA_SPLIT)
+#define BLK_TA_BOUNCE          (__BLK_TA_BOUNCE)
+#define BLK_TA_REMAP           (__BLK_TA_REMAP | BLK_TC_ACT(BLK_TC_QUEUE))
+
+#define BLK_IO_TRACE_MAGIC     0x65617400
+#define BLK_IO_TRACE_VERSION   0x07
+
+/*
+ * The trace itself
+ */
+struct blk_io_trace {
+       u32 magic;              /* MAGIC << 8 | version */
+       u32 sequence;           /* event number */
+       u64 time;               /* in microseconds */
+       u64 sector;             /* disk offset */
+       u32 bytes;              /* transfer length */
+       u32 action;             /* what happened */
+       u32 pid;                /* who did it */
+       u32 device;             /* device number */
+       u32 cpu;                /* on what cpu did it happen */
+       u16 error;              /* completion error */
+       u16 pdu_len;            /* length of data after this trace */
+};
+
+/*
+ * The remap event
+ */
+struct blk_io_trace_remap {
+       u32 device;
+       u32 __pad;
+       u64 sector;
+};
+
+enum {
+       Blktrace_setup = 1,
+       Blktrace_running,
+       Blktrace_stopped,
+};
+
+struct blk_trace {
+       int trace_state;
+       struct rchan *rchan;
+       unsigned long *sequence;
+       u16 act_mask;
+       u64 start_lba;
+       u64 end_lba;
+       u32 pid;
+       u32 dev;
+       struct dentry *dir;
+       struct dentry *dropped_file;
+       atomic_t dropped;
+};
+
+/*
+ * User setup structure passed with BLKTRACESTART
+ */
+struct blk_user_trace_setup {
+       char name[BDEVNAME_SIZE];       /* output */
+       u16 act_mask;                   /* input */
+       u32 buf_size;                   /* input */
+       u32 buf_nr;                     /* input */
+       u64 start_lba;
+       u64 end_lba;
+       u32 pid;
+};
+
+#if defined(CONFIG_BLK_DEV_IO_TRACE)
+extern int blk_trace_ioctl(struct block_device *, unsigned, char __user *);
+extern void blk_trace_shutdown(request_queue_t *);
+extern void __blk_add_trace(struct blk_trace *, sector_t, int, int, u32, int, int, void *);
+
+/**
+ * blk_add_trace_rq - Add a trace for a request oriented action
+ * @q:         queue the io is for
+ * @rq:                the source request
+ * @what:      the action
+ *
+ * Description:
+ *     Records an action against a request. Will log the bio offset + size.
+ *
+ **/
+static inline void blk_add_trace_rq(struct request_queue *q, struct request *rq,
+                                   u32 what)
+{
+       struct blk_trace *bt = q->blk_trace;
+       int rw = rq->flags & 0x07;
+
+       if (likely(!bt))
+               return;
+
+       if (blk_pc_request(rq)) {
+               what |= BLK_TC_ACT(BLK_TC_PC);
+               __blk_add_trace(bt, 0, rq->data_len, rw, what, rq->errors, sizeof(rq->cmd), rq->cmd);
+       } else  {
+               what |= BLK_TC_ACT(BLK_TC_FS);
+               __blk_add_trace(bt, rq->hard_sector, rq->hard_nr_sectors << 9, rw, what, rq->errors, 0, NULL);
+       }
+}
+
+/**
+ * blk_add_trace_bio - Add a trace for a bio oriented action
+ * @q:         queue the io is for
+ * @bio:       the source bio
+ * @what:      the action
+ *
+ * Description:
+ *     Records an action against a bio. Will log the bio offset + size.
+ *
+ **/
+static inline void blk_add_trace_bio(struct request_queue *q, struct bio *bio,
+                                    u32 what)
+{
+       struct blk_trace *bt = q->blk_trace;
+
+       if (likely(!bt))
+               return;
+
+       __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what, !bio_flagged(bio, BIO_UPTODATE), 0, NULL);
+}
+
+/**
+ * blk_add_trace_generic - Add a trace for a generic action
+ * @q:         queue the io is for
+ * @bio:       the source bio
+ * @rw:                the data direction
+ * @what:      the action
+ *
+ * Description:
+ *     Records a simple trace
+ *
+ **/
+static inline void blk_add_trace_generic(struct request_queue *q,
+                                        struct bio *bio, int rw, u32 what)
+{
+       struct blk_trace *bt = q->blk_trace;
+
+       if (likely(!bt))
+               return;
+
+       if (bio)
+               blk_add_trace_bio(q, bio, what);
+       else
+               __blk_add_trace(bt, 0, 0, rw, what, 0, 0, NULL);
+}
+
+/**
+ * blk_add_trace_pdu_int - Add a trace for a bio with an integer payload
+ * @q:         queue the io is for
+ * @what:      the action
+ * @bio:       the source bio
+ * @pdu:       the integer payload
+ *
+ * Description:
+ *     Adds a trace with some integer payload. This might be an unplug
+ *     option given as the action, with the depth at unplug time given
+ *     as the payload
+ *
+ **/
+static inline void blk_add_trace_pdu_int(struct request_queue *q, u32 what,
+                                        struct bio *bio, unsigned int pdu)
+{
+       struct blk_trace *bt = q->blk_trace;
+       u64 rpdu = cpu_to_be64(pdu);
+
+       if (likely(!bt))
+               return;
+
+       if (bio)
+               __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what, !bio_flagged(bio, BIO_UPTODATE), sizeof(rpdu), &rpdu);
+       else
+               __blk_add_trace(bt, 0, 0, 0, what, 0, sizeof(rpdu), &rpdu);
+}
+
+/**
+ * blk_add_trace_remap - Add a trace for a remap operation
+ * @q:         queue the io is for
+ * @bio:       the source bio
+ * @dev:       target device
+ * @from:      source sector
+ * @to:                target sector
+ *
+ * Description:
+ *     Device mapper or raid target sometimes need to split a bio because
+ *     it spans a stripe (or similar). Add a trace for that action.
+ *
+ **/
+static inline void blk_add_trace_remap(struct request_queue *q, struct bio *bio,
+                                      dev_t dev, sector_t from, sector_t to)
+{
+       struct blk_trace *bt = q->blk_trace;
+       struct blk_io_trace_remap r;
+
+       if (likely(!bt))
+               return;
+
+       r.device = cpu_to_be32(dev);
+       r.sector = cpu_to_be64(to);
+
+       __blk_add_trace(bt, from, bio->bi_size, bio->bi_rw, BLK_TA_REMAP, !bio_flagged(bio, BIO_UPTODATE), sizeof(r), &r);
+}
+
+#else /* !CONFIG_BLK_DEV_IO_TRACE */
+#define blk_trace_ioctl(bdev, cmd, arg)                (-ENOTTY)
+#define blk_trace_shutdown(q)                  do { } while (0)
+#define blk_add_trace_rq(q, rq, what)          do { } while (0)
+#define blk_add_trace_bio(q, rq, what)         do { } while (0)
+#define blk_add_trace_generic(q, rq, rw, what) do { } while (0)
+#define blk_add_trace_pdu_int(q, what, bio, pdu)       do { } while (0)
+#define blk_add_trace_remap(q, bio, dev, f, t) do {} while (0)
+#endif /* CONFIG_BLK_DEV_IO_TRACE */
+
+#endif
index ae7dfb790df3b5ebcd2c1e5e0f695c31553c306d..efb518f16bb3a52cd01387de1f487c5ef5aff850 100644 (file)
@@ -97,6 +97,10 @@ COMPATIBLE_IOCTL(BLKRRPART)
 COMPATIBLE_IOCTL(BLKFLSBUF)
 COMPATIBLE_IOCTL(BLKSECTSET)
 COMPATIBLE_IOCTL(BLKSSZGET)
+COMPATIBLE_IOCTL(BLKTRACESTART)
+COMPATIBLE_IOCTL(BLKTRACESTOP)
+COMPATIBLE_IOCTL(BLKTRACESETUP)
+COMPATIBLE_IOCTL(BLKTRACETEARDOWN)
 ULONG_IOCTL(BLKRASET)
 ULONG_IOCTL(BLKFRASET)
 /* RAID */
index f9c9dea636d0ba7be7637c6f1975c669052c58ab..9b34a1b034550738d7cba54c5766401cbf237c1d 100644 (file)
@@ -197,6 +197,10 @@ extern int dir_notify_enable;
 #define BLKBSZGET  _IOR(0x12,112,size_t)
 #define BLKBSZSET  _IOW(0x12,113,size_t)
 #define BLKGETSIZE64 _IOR(0x12,114,size_t)     /* return device size in bytes (u64 *arg) */
+#define BLKTRACESETUP _IOWR(0x12,115,struct blk_user_trace_setup)
+#define BLKTRACESTART _IO(0x12,116)
+#define BLKTRACESTOP _IO(0x12,117)
+#define BLKTRACETEARDOWN _IO(0x12,118)
 
 #define BMAP_IOCTL 1           /* obsolete - kept for compatibility */
 #define FIBMAP    _IO(0x00,1)  /* bmap access */
index 62e6314382f091826490f80ef6c839d9587c8694..e60a91d5b36929560fd34895039eca315c166bd3 100644 (file)
@@ -706,6 +706,7 @@ struct task_struct {
        prio_array_t *array;
 
        unsigned short ioprio;
+       unsigned int btrace_seq;
 
        unsigned long sleep_avg;
        unsigned long long timestamp, last_ran;
index c79ae0b19a4922da1ff651f802eb199afe4fe7c0..c21bae8c93b94bd8cfe60129af508a64ddb67c33 100644 (file)
@@ -181,6 +181,7 @@ static struct task_struct *dup_task_struct(struct task_struct *orig)
        /* One for us, one for whoever does the "release_task()" (usually parent) */
        atomic_set(&tsk->usage,2);
        atomic_set(&tsk->fs_excl, 0);
+       tsk->btrace_seq = 0;
        return tsk;
 }
 
index ce2e7e8bbfa7102a78473d1a47c24a41f7788b83..d0ea1eec6a9a023415e3621e0332cdd837130ca3 100644 (file)
@@ -26,6 +26,7 @@
 #include <linux/init.h>
 #include <linux/hash.h>
 #include <linux/highmem.h>
+#include <linux/blktrace_api.h>
 #include <asm/tlbflush.h>
 
 static mempool_t *page_pool, *isa_page_pool;
@@ -483,6 +484,8 @@ void blk_queue_bounce(request_queue_t *q, struct bio **bio_orig)
                pool = isa_page_pool;
        }
 
+       blk_add_trace_bio(q, *bio_orig, BLK_TA_BOUNCE);
+
        /*
         * slow path
         */