block: fix I/O throttling accounting blind spot
authorStefan Hajnoczi <stefanha@redhat.com>
Fri, 5 Apr 2013 09:32:19 +0000 (11:32 +0200)
committerKevin Wolf <kwolf@redhat.com>
Fri, 5 Apr 2013 16:58:05 +0000 (18:58 +0200)
I/O throttling relies on bdrv_acct_done() which is called when a request
completes.  This leaves a blind spot since we only charge for completed
requests, not submitted requests.

For example, if there is 1 operation remaining in this time slice the
guest could submit 3 operations and they will all be submitted
successfully since they don't actually get accounted for until they
complete.

Originally we probably thought this is okay since the requests will be
accounted when the time slice is extended.  In practice it causes
fluctuations since the guest can exceed its I/O limit and it will be
punished for this later on.

Account for I/O upon submission so that I/O limits are enforced
properly.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Tested-By: Benoit Canet <benoit@irqsave.net>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
block.c
include/block/block_int.h

diff --git a/block.c b/block.c
index 0ae2e93982f6dcc3cd94e7e56eefb901227891ec..25976b556aa9919536769260710970f523bd3491 100644 (file)
--- a/block.c
+++ b/block.c
@@ -141,7 +141,6 @@ void bdrv_io_limits_disable(BlockDriverState *bs)
     bs->slice_start = 0;
     bs->slice_end   = 0;
     bs->slice_time  = 0;
-    memset(&bs->io_base, 0, sizeof(bs->io_base));
 }
 
 static void bdrv_block_timer(void *opaque)
@@ -1436,8 +1435,8 @@ static void bdrv_move_feature_fields(BlockDriverState *bs_dest,
     bs_dest->slice_time         = bs_src->slice_time;
     bs_dest->slice_start        = bs_src->slice_start;
     bs_dest->slice_end          = bs_src->slice_end;
+    bs_dest->slice_submitted    = bs_src->slice_submitted;
     bs_dest->io_limits          = bs_src->io_limits;
-    bs_dest->io_base            = bs_src->io_base;
     bs_dest->throttled_reqs     = bs_src->throttled_reqs;
     bs_dest->block_timer        = bs_src->block_timer;
     bs_dest->io_limits_enabled  = bs_src->io_limits_enabled;
@@ -3768,9 +3767,9 @@ static bool bdrv_exceed_bps_limits(BlockDriverState *bs, int nb_sectors,
     slice_time = bs->slice_end - bs->slice_start;
     slice_time /= (NANOSECONDS_PER_SECOND);
     bytes_limit = bps_limit * slice_time;
-    bytes_base  = bs->nr_bytes[is_write] - bs->io_base.bytes[is_write];
+    bytes_base  = bs->slice_submitted.bytes[is_write];
     if (bs->io_limits.bps[BLOCK_IO_LIMIT_TOTAL]) {
-        bytes_base += bs->nr_bytes[!is_write] - bs->io_base.bytes[!is_write];
+        bytes_base += bs->slice_submitted.bytes[!is_write];
     }
 
     /* bytes_base: the bytes of data which have been read/written; and
@@ -3828,9 +3827,9 @@ static bool bdrv_exceed_iops_limits(BlockDriverState *bs, bool is_write,
     slice_time = bs->slice_end - bs->slice_start;
     slice_time /= (NANOSECONDS_PER_SECOND);
     ios_limit  = iops_limit * slice_time;
-    ios_base   = bs->nr_ops[is_write] - bs->io_base.ios[is_write];
+    ios_base   = bs->slice_submitted.ios[is_write];
     if (bs->io_limits.iops[BLOCK_IO_LIMIT_TOTAL]) {
-        ios_base += bs->nr_ops[!is_write] - bs->io_base.ios[!is_write];
+        ios_base += bs->slice_submitted.ios[!is_write];
     }
 
     if (ios_base + 1 <= ios_limit) {
@@ -3875,11 +3874,7 @@ static bool bdrv_exceed_io_limits(BlockDriverState *bs, int nb_sectors,
         bs->slice_start = now;
         bs->slice_end   = now + bs->slice_time;
 
-        bs->io_base.bytes[is_write]  = bs->nr_bytes[is_write];
-        bs->io_base.bytes[!is_write] = bs->nr_bytes[!is_write];
-
-        bs->io_base.ios[is_write]    = bs->nr_ops[is_write];
-        bs->io_base.ios[!is_write]   = bs->nr_ops[!is_write];
+        memset(&bs->slice_submitted, 0, sizeof(bs->slice_submitted));
     }
 
     elapsed_time  = now - bs->slice_start;
@@ -3907,6 +3902,10 @@ static bool bdrv_exceed_io_limits(BlockDriverState *bs, int nb_sectors,
         *wait = 0;
     }
 
+    bs->slice_submitted.bytes[is_write] += (int64_t)nb_sectors *
+                                           BDRV_SECTOR_SIZE;
+    bs->slice_submitted.ios[is_write]++;
+
     return false;
 }
 
index 0986a2d6ac93cec45d66146cfa9faa2abc104634..83941d867232e3027dd1d398396f20a0399dfb91 100644 (file)
@@ -256,7 +256,7 @@ struct BlockDriverState {
     int64_t slice_start;
     int64_t slice_end;
     BlockIOLimit io_limits;
-    BlockIOBaseValue  io_base;
+    BlockIOBaseValue slice_submitted;
     CoQueue      throttled_reqs;
     QEMUTimer    *block_timer;
     bool         io_limits_enabled;