Skip to contentSkip to conversation

qcow2: Fix the calculation of the maximum L2 cache size

PATCH Applied Alberto Garcia 3 participants 7 messages Aug 16, 2019 – Aug 18, 2019 Block layer

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 qcow2: Fix the calculation of the maximum L2 cache size

The size of the qcow2 L2 cache defaults to 32 MB, which can be easily larger than the maximum amount of L2 metadata that the image can have. For example: with 64 KB clusters the user would need a qcow2 image with a virtual size of 256 GB in order to have 32 MB of L2 metadata.

Because of that, since commit b749562d9822d14ef69c9eaa5f85903010b86c30 we forbid the L2 cache to become larger than the maximum amount of L2 metadata for the image, calculated using this formula:

    uint64_t max_l2_cache = virtual_disk_size / (s->cluster_size / 8);

The problem with this formula is that the result should be rounded up to the cluster size because an L2 table on disk always takes one full cluster.

For example, a 1280 MB qcow2 image with 64 KB clusters needs exactly 160 KB of L2 metadata, but we need 192 KB on disk (3 clusters) even if the last 32 KB of those are not going to be used.

However QEMU rounds the numbers down and only creates 2 cache tables (128 KB), which is not enough for the image.

A quick test doing 4KB random writes on a 1280 MB image gives me around 500 IOPS, while with the correct cache size I get 16K IOPS.

---
 block/qcow2.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
block/qcow2.c +5 -1
index 039bdc2f7e..865839682c 100644
@@ -826,7 +826,11 @@ static void read_cache_sizes(BlockDriverState *bs, QemuOpts *opts,
     bool l2_cache_entry_size_set;
     int min_refcount_cache = MIN_REFCOUNT_CACHE_SIZE * s->cluster_size;
     uint64_t virtual_disk_size = bs->total_sectors * BDRV_SECTOR_SIZE;
-    uint64_t max_l2_cache = virtual_disk_size / (s->cluster_size / 8);
+    uint64_t max_l2_entries = DIV_ROUND_UP(virtual_disk_size, s->cluster_size);
+    /* An L2 table is always one cluster in size so the max cache size
+     * should be a multiple of the cluster size. */
+    uint64_t max_l2_cache = ROUND_UP(max_l2_entries * sizeof(uint64_t),
+                                     s->cluster_size);
 
     combined_cache_size_set = qemu_opt_get(opts, QCOW2_OPT_CACHE_SIZE);
     l2_cache_size_set = qemu_opt_get(opts, QCOW2_OPT_L2_CACHE_SIZE);
Signature · 1 line
2.20.1
#1 lore raw

Cc qemu-stable

This bug means that under certain conditions it's impossible to create a cache large enough for the image, resulting in reduced I/O performance.

Quoted 49 lines · Alberto Garcia

The size of the qcow2 L2 cache defaults to 32 MB, which can be easily larger than the maximum amount of L2 metadata that the image can have. For example: with 64 KB clusters the user would need a qcow2 image with a virtual size of 256 GB in order to have 32 MB of L2 metadata.

Because of that, since commit b749562d9822d14ef69c9eaa5f85903010b86c30 we forbid the L2 cache to become larger than the maximum amount of L2 metadata for the image, calculated using this formula:

    uint64_t max_l2_cache = virtual_disk_size / (s->cluster_size / 8);

The problem with this formula is that the result should be rounded up to the cluster size because an L2 table on disk always takes one full cluster.

For example, a 1280 MB qcow2 image with 64 KB clusters needs exactly 160 KB of L2 metadata, but we need 192 KB on disk (3 clusters) even if the last 32 KB of those are not going to be used.

However QEMU rounds the numbers down and only creates 2 cache tables (128 KB), which is not enough for the image.

A quick test doing 4KB random writes on a 1280 MB image gives me around 500 IOPS, while with the correct cache size I get 16K IOPS.

---
 block/qcow2.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
block/qcow2.c +5 -1
index 039bdc2f7e..865839682c 100644
@@ -826,7 +826,11 @@ static void read_cache_sizes(BlockDriverState *bs, QemuOpts *opts,
     bool l2_cache_entry_size_set;
     int min_refcount_cache = MIN_REFCOUNT_CACHE_SIZE * s->cluster_size;
     uint64_t virtual_disk_size = bs->total_sectors * BDRV_SECTOR_SIZE;
-    uint64_t max_l2_cache = virtual_disk_size / (s->cluster_size / 8);
+    uint64_t max_l2_entries = DIV_ROUND_UP(virtual_disk_size, s->cluster_size);
+    /* An L2 table is always one cluster in size so the max cache size
+     * should be a multiple of the cluster size. */
+    uint64_t max_l2_cache = ROUND_UP(max_l2_entries * sizeof(uint64_t),
+                                     s->cluster_size);
 
     combined_cache_size_set = qemu_opt_get(opts, QCOW2_OPT_CACHE_SIZE);
     l2_cache_size_set = qemu_opt_get(opts, QCOW2_OPT_L2_CACHE_SIZE);
Signature · 1 line
2.20.1
Applied
#2 lore raw
In reply to Alberto Garcia

Am 16.08.2019 um 14:17 hat Alberto Garcia geschrieben:

Quoted 19 lines

The size of the qcow2 L2 cache defaults to 32 MB, which can be easily larger than the maximum amount of L2 metadata that the image can have. For example: with 64 KB clusters the user would need a qcow2 image with a virtual size of 256 GB in order to have 32 MB of L2 metadata.

Because of that, since commit b749562d9822d14ef69c9eaa5f85903010b86c30 we forbid the L2 cache to become larger than the maximum amount of L2 metadata for the image, calculated using this formula:

    uint64_t max_l2_cache = virtual_disk_size / (s->cluster_size / 8);

The problem with this formula is that the result should be rounded up to the cluster size because an L2 table on disk always takes one full cluster.

For example, a 1280 MB qcow2 image with 64 KB clusters needs exactly 160 KB of L2 metadata, but we need 192 KB on disk (3 clusters) even if the last 32 KB of those are not going to be used.

Quoted 7 lines

However QEMU rounds the numbers down and only creates 2 cache tables (128 KB), which is not enough for the image.

A quick test doing 4KB random writes on a 1280 MB image gives me around 500 IOPS, while with the correct cache size I get 16K IOPS.

Hm, this is bad. :-(

The requirement so that this bug doesn't affect the user seems to be that the image size is a multiple of 64k * 8k = 512 MB. Which means that users are probably often lucky enough in practice.

I'll Cc: qemu-stable anyway.

Thanks, applied to the block branch.

Kevin

#3 lore raw
Quoted 3 lines · Kevin Wolf

The requirement so that this bug doesn't affect the user seems to be that the image size is a multiple of 64k * 8k = 512 MB. Which means that users are probably often lucky enough in practice.

Or rather: cluster_size^2 / 8, which, if my numbers are right:

|--------------+-------------|
| Cluster size | Multiple of |
|--------------+-------------|
|         4 KB |        2 MB |
|         8 KB |        8 MB |
|        16 KB |       32 MB |
|        32 KB |      128 MB |
|        64 KB |      512 MB |
|       128 KB |        2 GB |
|       256 KB |        8 GB |
|       512 KB |       32 GB |
|      1024 KB |      128 GB |
|      2048 KB |      512 GB |
|--------------+-------------|

It get trickier with larger clusters, but if you have a larger cluster size you probably have a very large image anyway, so yes I also think that users are probably lucky enough in practice.

(also, the number of cache tables is always >= 2, so if the image size is less than twice those numbers then it's also safe)

And yes, the odd value on the 512KB row on that we discussed last month is due to this same bug:

https://lists.gnu.org/archive/html/qemu-block/2019-07/msg00496.html

Berto

#4 lore raw

Am 16.08.2019 um 15:30 hat Alberto Garcia geschrieben:

Quoted 22 lines · Kevin Wolf
Quoted 3 lines

The requirement so that this bug doesn't affect the user seems to be that the image size is a multiple of 64k * 8k = 512 MB. Which means that users are probably often lucky enough in practice.

Or rather: cluster_size^2 / 8, which, if my numbers are right:

|--------------+-------------|
| Cluster size | Multiple of |
|--------------+-------------|
|         4 KB |        2 MB |
|         8 KB |        8 MB |
|        16 KB |       32 MB |
|        32 KB |      128 MB |
|        64 KB |      512 MB |
|       128 KB |        2 GB |
|       256 KB |        8 GB |
|       512 KB |       32 GB |
|      1024 KB |      128 GB |
|      2048 KB |      512 GB |
|--------------+-------------|
Quoted 3 lines

It get trickier with larger clusters, but if you have a larger cluster size you probably have a very large image anyway, so yes I also think that users are probably lucky enough in practice.

Yes, I assumed 64k clusters.

The other somewhat popular cluster size is probably 2 MB, where I think images sizes that are not a multiple of 512 GB are rather likely...

Quoted 2 lines

(also, the number of cache tables is always >= 2, so if the image size is less than twice those numbers then it's also safe)

Right. I already corrected my statement to include > 1024 MB in the Red Hat Bugzilla (but still didn't consider other cluster sizes).

Quoted 4 lines

And yes, the odd value on the 512KB row on that we discussed last month is due to this same bug:

https://lists.gnu.org/archive/html/qemu-block/2019-07/msg00496.html

Hm... And suddently it makes sense. :-)

So I assume all of 512k/1024k/2048k actually perform better? Or is the effect neglegible for 1024k/2048k?

Kevin

#5 lore raw
Quoted 9 lines · Kevin Wolf
Quoted 4 lines

And yes, the odd value on the 512KB row on that we discussed last month is due to this same bug:

https://lists.gnu.org/archive/html/qemu-block/2019-07/msg00496.html

Hm... And suddently it makes sense. :-)

So I assume all of 512k/1024k/2048k actually perform better? Or is the effect neglegible for 1024k/2048k?

The 512K case is the only one that performs better, my test image was too small (40 GB) for the other cases.

Berto

#6 lore raw
In reply to Alberto Garcia

Thanks, Berto!

Quoted 48 lines · Alberto Garcia

The size of the qcow2 L2 cache defaults to 32 MB, which can be easily larger than the maximum amount of L2 metadata that the image can have. For example: with 64 KB clusters the user would need a qcow2 image with a virtual size of 256 GB in order to have 32 MB of L2 metadata.

Because of that, since commit b749562d9822d14ef69c9eaa5f85903010b86c30 we forbid the L2 cache to become larger than the maximum amount of L2 metadata for the image, calculated using this formula:

     uint64_t max_l2_cache = virtual_disk_size / (s->cluster_size / 8);

The problem with this formula is that the result should be rounded up to the cluster size because an L2 table on disk always takes one full cluster.

For example, a 1280 MB qcow2 image with 64 KB clusters needs exactly 160 KB of L2 metadata, but we need 192 KB on disk (3 clusters) even if the last 32 KB of those are not going to be used.

However QEMU rounds the numbers down and only creates 2 cache tables (128 KB), which is not enough for the image.

A quick test doing 4KB random writes on a 1280 MB image gives me around 500 IOPS, while with the correct cache size I get 16K IOPS.

---
  block/qcow2.c | 6 +++++-
  1 file changed, 5 insertions(+), 1 deletion(-)
block/qcow2.c +5 -1
index 039bdc2f7e..865839682c 100644
@@ -826,7 +826,11 @@ static void read_cache_sizes(BlockDriverState *bs, QemuOpts *opts,
      bool l2_cache_entry_size_set;
      int min_refcount_cache = MIN_REFCOUNT_CACHE_SIZE * s->cluster_size;
      uint64_t virtual_disk_size = bs->total_sectors * BDRV_SECTOR_SIZE;
-    uint64_t max_l2_cache = virtual_disk_size / (s->cluster_size / 8);
+    uint64_t max_l2_entries = DIV_ROUND_UP(virtual_disk_size, s->cluster_size);
+    /* An L2 table is always one cluster in size so the max cache size
+     * should be a multiple of the cluster size. */
+    uint64_t max_l2_cache = ROUND_UP(max_l2_entries * sizeof(uint64_t),
+                                     s->cluster_size);
 
      combined_cache_size_set = qemu_opt_get(opts, QCOW2_OPT_CACHE_SIZE);
      l2_cache_size_set = qemu_opt_get(opts, QCOW2_OPT_L2_CACHE_SIZE);