blktrace support: discard notify messages
[fio.git] / blktrace.c
index 1669fe556bc3324d2b40567e0afdab0f0e359667..5825d29714e13627dd9712699a904f752445cd70 100644 (file)
@@ -8,18 +8,59 @@
 #include "fio.h"
 #include "blktrace_api.h"
 
+#define TRACE_FIFO_SIZE        65536
+
+/*
+ * fifo refill frontend, to avoid reading data in trace sized bites
+ */
+static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
+{
+       char buf[TRACE_FIFO_SIZE];
+       unsigned int total;
+       int ret;
+
+       total = sizeof(buf);
+       if (total > fifo_room(fifo))
+               total = fifo_room(fifo);
+
+       ret = read(fd, buf, total);
+       if (ret < 0) {
+               td_verror(td, errno, "read blktrace file");
+               return -1;
+       }
+
+       if (ret > 0)
+               ret = fifo_put(fifo, buf, ret);
+
+       return ret;
+}
+
+/*
+ * Retrieve 'len' bytes from the fifo, refilling if necessary.
+ */
+static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
+                         void *buf, unsigned int len)
+{
+       if (fifo_len(fifo) < len) {
+               int ret = refill_fifo(td, fifo, fd);
+
+               if (ret < 0)
+                       return ret;
+       }
+
+       return fifo_get(fifo, buf, len);
+}
+
 /*
  * Just discard the pdu by seeking past it.
  */
-static int discard_pdu(int fd, struct blk_io_trace *t)
+static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
+                      struct blk_io_trace *t)
 {
        if (t->pdu_len == 0)
                return 0;
 
-       if (lseek(fd, t->pdu_len, SEEK_CUR) < 0)
-               return errno;
-               
-       return 0;
+       return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
 }
 
 /*
@@ -83,7 +124,8 @@ static void store_ipo(struct thread_data *td, unsigned long long offset,
  * due to internal workings of the block layer.
  */
 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
-                        unsigned long long ttime, unsigned long *ios)
+                        unsigned long long ttime, unsigned long *ios,
+                        unsigned int *bs)
 {
        int rw;
 
@@ -91,14 +133,14 @@ static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
                return;
        if (t->action & BLK_TC_ACT(BLK_TC_PC))
                return;
-
-       /*
-        * should not happen, need to look into that...
-        */
-       if (!t->bytes)
+       if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
                return;
 
        rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
+
+       if (t->bytes > bs[rw])
+               bs[rw] = t->bytes;
+
        ios[rw]++;
        td->o.size += t->bytes;
        store_ipo(td, t->sector, t->bytes, rw, ttime);
@@ -110,9 +152,12 @@ static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
  */
 int load_blktrace(struct thread_data *td, const char *filename)
 {
-       unsigned long long ttime;
+       unsigned long long ttime, delay;
        struct blk_io_trace t;
        unsigned long ios[2];
+       unsigned int cpu;
+       unsigned int rw_bs[2];
+       struct fifo *fifo;
        int fd;
 
        fd = open(filename, O_RDONLY);
@@ -121,59 +166,71 @@ int load_blktrace(struct thread_data *td, const char *filename)
                return 1;
        }
 
+       fifo = fifo_alloc(TRACE_FIFO_SIZE);
+
        td->o.size = 0;
 
-       ios[0] = ios[1] = 0;
+       cpu = 0;
        ttime = 0;
+       ios[0] = ios[1] = 0;
+       rw_bs[0] = rw_bs[1] = 0;
        do {
-               /*
-                * Once this is working fully, I'll add a layer between
-                * here and read to cache trace data. Then we can avoid
-                * doing itsy bitsy reads, but instead pull in a larger
-                * chunk of data at the time.
-                */
-               int ret = read(fd, &t, sizeof(t));
+               int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
 
-               if (ret < 0) {
-                       td_verror(td, errno, "read blktrace file");
+               if (ret < 0)
                        goto err;
-               } else if (!ret) {
+               else if (!ret)
+                       break;
+               else if (ret < (int) sizeof(t)) {
+                       log_err("fio: short fifo get\n");
                        break;
-               } else if (ret != sizeof(t)) {
-                       log_err("fio: short read on blktrace file\n");
-                       goto err;
                }
 
                if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
-                       log_err("fio: bad magic in blktrace data\n");
+                       log_err("fio: bad magic in blktrace data: %x\n", t.magic);
                        goto err;
                }
                if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
                        log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
                        goto err;
                }
-               ret = discard_pdu(fd, &t);
-               if (ret) {
+               ret = discard_pdu(td, fifo, fd, &t);
+               if (ret < 0) {
                        td_verror(td, ret, "blktrace lseek");
                        goto err;
+               } else if (t.pdu_len != ret) {
+                       log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
+                       goto err;
                }
-               if (!ttime)
+               if (!ttime) {
                        ttime = t.time;
-               handle_trace(td, &t, t.time - ttime, ios);
+                       cpu = t.cpu;
+               }
+               delay = 0;
+               if (cpu == t.cpu)
+                       delay = t.time - ttime;
+               handle_trace(td, &t, delay, ios, rw_bs);
                ttime = t.time;
+               cpu = t.cpu;
        } while (1);
 
+       fifo_free(fifo);
        close(fd);
 
        if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
                log_err("fio: found no ios in blktrace data\n");
                return 1;
-       } else if (ios[DDIR_READ] && !ios[DDIR_READ])
+       } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
                td->o.td_ddir = TD_DDIR_READ;
-       else if (!ios[DDIR_READ] && ios[DDIR_WRITE])
+               td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
+       } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
                td->o.td_ddir = TD_DDIR_WRITE;
-       else
+               td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
+       } else {
                td->o.td_ddir = TD_DDIR_RW;
+               td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
+               td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
+       }
 
        /*
         * We need to do direct/raw ios to the device, to avoid getting
@@ -184,5 +241,6 @@ int load_blktrace(struct thread_data *td, const char *filename)
        return 0;
 err:
        close(fd);
+       fifo_free(fifo);
        return 1;
 }