minor qcow2 compression improvements

Although off_t permits up to 63 bits (8EB) of file offsets, in practice, we're going to hit other limits first. Document some of those limits in the qcow2 spec, and how choice of cluster size can influence some of the limits.

While at it, notice that since we cannot map any virtual cluster to any address higher than 64 PB (56 bits) (due to the current L1/L2 field encoding stopping at bit 55), it makes little sense to require the refcount table to access host offsets beyond that point. Mark the upper bits of the refcount table entries as reserved to match the L1/L2 table, with no ill effects, since it is unlikely that there are any existing images larger than 64PB in the first place, and thus all existing images already have those bits as 0. If 64PB proves to be too small in the future, we could enlarge all three uses of bit 55 into the reserved bits at that time. (For reference, ext4 with 4k blocks caps files at 16PB.)

However, there is one limit that reserved bits don't help with: for compressed clusters, the L2 layout requires an ever-smaller maximum host offset as cluster size gets larger, down to a 512 TB maximum with 2M clusters.

Signed-off-by Eric Blake <[email protected]>

Our code was already checking that we did not attempt to allocate more clusters than what would fit in an INT64 (the physical maximimum if we can access a full off_t's worth of data). But this does not catch smaller limits enforced by various spots in the qcow2 image description: L1 and normal clusters of L2 are documented as having bits 63-56 reserved for other purposes, capping our maximum offset at 64PB (bit 55 is the maximum bit set). And for compressed images with 2M clusters, the cap drops the maximum offset to bit 48, or a maximum offset of 512TB. If we overflow that offset, we would write compressed data into one place, but try to decompress from another, which won't work.

I don't have 512TB handy to prove whether things break if we compress so much data that we overflow that limit, and don't think that iotests can (quickly) test it either. Test 138 comes close (it corrupts an image into thinking something lives at 32PB, which is half the maximum for L1 sizing - although it relies on 512-byte clusters). But that test points out that we will generally hit other limits first (such as running out of memory for the refcount table, or exceeding file system limits like 16TB on ext4, etc), so this is more a theoretical safety valve than something likely to be hit.

Signed-off-by Eric Blake <[email protected]>

Reviewed-by Alberto Garcia <[email protected]>

---
v3: use s->cluster_offset_mask instead of open-coding it [Berto], use MIN() to clamp offset on small cluster size, add spec patch first to justify clamping even on refcount allocations

---
 block/qcow2.h          |  6 ++++++
 block/qcow2-refcount.c | 21 ++++++++++++++-------
 2 files changed, 20 insertions(+), 7 deletions(-)

block/qcow2.h +6 -0

index 1df15a18aa1..a3b9faa9d53 100644
@@ -41,6 +41,12 @@
 #define QCOW_MAX_CRYPT_CLUSTERS 32
 #define QCOW_MAX_SNAPSHOTS 65536
 
+/* Field widths in qcow2 mean normal cluster offsets cannot reach
+ * 64PB; depending on cluster size, compressed clusters can have a
+ * smaller limit (64PB for up to 16k clusters, then ramps down to
+ * 512TB for 2M clusters).  */
+#define QCOW_MAX_CLUSTER_OFFSET ((1ULL << 56) - 1)
+
 /* 8 MB refcount table is enough for 2 PB images at 64k cluster size
  * (128 GB for 512 byte clusters, 2 EB for 2 MB clusters) */
 #define QCOW_MAX_REFTABLE_SIZE 0x800000

block/qcow2-refcount.c +14 -7

index 6bfc11bb48f..fcbea3c9644 100644
@@ -31,7 +31,8 @@
 #include "qemu/bswap.h"
 #include "qemu/cutils.h"
 
-static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size);
+static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size,
+                                    uint64_t max);
 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
                             int64_t offset, int64_t length, uint64_t addend,
                             bool decrease, enum qcow2_discard_type type);

@@ -362,7 +363,8 @@ static int alloc_refcount_block(BlockDriverState *bs,
     }
 
     /* Allocate the refcount block itself and mark it as used */
-    int64_t new_block = alloc_clusters_noref(bs, s->cluster_size);
+    int64_t new_block = alloc_clusters_noref(bs, s->cluster_size,
+                                             QCOW_MAX_CLUSTER_OFFSET);
     if (new_block < 0) {
         return new_block;
     }

@@ -954,7 +956,8 @@ int qcow2_update_cluster_refcount(BlockDriverState *bs,
 
 
 /* return < 0 if error */
-static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size)
+static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size,
+                                    uint64_t max)
 {
     BDRVQcow2State *s = bs->opaque;
     uint64_t i, nb_clusters, refcount;

@@ -979,9 +982,9 @@ retry:
     }
 
     /* Make sure that all offsets in the "allocated" range are representable
-     * in an int64_t */
+     * in the requested max */
     if (s->free_cluster_index > 0 &&
-        s->free_cluster_index - 1 > (INT64_MAX >> s->cluster_bits))
+        s->free_cluster_index - 1 > (max >> s->cluster_bits))
     {
         return -EFBIG;
     }

@@ -1001,7 +1004,7 @@ int64_t qcow2_alloc_clusters(BlockDriverState *bs, uint64_t size)
 
     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
     do {
-        offset = alloc_clusters_noref(bs, size);
+        offset = alloc_clusters_noref(bs, size, QCOW_MAX_CLUSTER_OFFSET);
         if (offset < 0) {
             return offset;
         }

@@ -1083,7 +1086,11 @@ int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
     free_in_cluster = s->cluster_size - offset_into_cluster(s, offset);
     do {
         if (!offset || free_in_cluster < size) {
-            int64_t new_cluster = alloc_clusters_noref(bs, s->cluster_size);
+            int64_t new_cluster;
+
+            new_cluster = alloc_clusters_noref(bs, s->cluster_size,
+                                               MIN(s->cluster_offset_mask,
+                                                   QCOW_MAX_CLUSTER_OFFSET));
             if (new_cluster < 0) {
                 return new_cluster;
             }

Quoted 3 lines · Eric Blake

Our code was already checking that we did not attempt to allocate more clusters than what would fit in an INT64 (the physical maximimum if we can access a full off_t's worth of

Quoted 10 lines

data). But this does not catch smaller limits enforced by various spots in the qcow2 image description: L1 and normal clusters of L2 are documented as having bits 63-56 reserved for other purposes, capping our maximum offset at 64PB (bit 55 is the maximum bit set). And for compressed images with 2M clusters, the cap drops the maximum offset to bit 48, or a maximum offset of 512TB. If we overflow that offset, we would write compressed data into one place, but try to decompress from another, which won't work.

Quoted 10 lines

I don't have 512TB handy to prove whether things break if we compress so much data that we overflow that limit, and don't think that iotests can (quickly) test it either. Test 138 comes close (it corrupts an image into thinking something lives at 32PB, which is half the maximum for L1 sizing - although it relies on 512-byte clusters). But that test points out that we will generally hit other limits first (such as running out of memory for the refcount table, or exceeding file system limits like 16TB on ext4, etc), so this is more a theoretical safety valve than something likely to be hit.

You don't need 512 TB, though, 36 MB is sufficient.

Here's what you do:
(1) Create a 513 TB image with cluster_size=2M,refcount_bits=1
(2) Take a hex editor and enter 16 refblocks into the reftable
(3) Fill all of those refblocks with 1s

(Funny side note: qemu-img check thinks that image is clean because it doesn't check refcounts beyond the image end...)

I've attached a compressed test image (unsurprisingly, it compresses really well).

Before this series:

$ ./qemu-io -c 'write -c 0 2M' test.qcow2
qcow2: Marking image as corrupt: Preventing invalid write on metadata
(overlaps with refcount block); further corruption events will be suppressed
write failed: Input/output error

Aw.

After this series:

$ ./qemu-io -c 'write -c 0 2M' test.qcow2
write failed: Input/output error

(Normal writes just work fine.)

Maybe you want to add a test still -- creating the image is rather quick (well, you have to write 64 MB of 1s, but other than that). The only thing that takes a bit of time is qemu figuring out where the first free cluster is... That takes like 15 seconds here.

And another issue of course is...

$ ls -lhs test.qcow2
42M -rw-r--r--. 1 maxx maxx 513T 25. Apr 16:42 test.qcow2

Yeah, that. Depends on the host file system, of course, whether that is a real issue. O:-)

Here's what you do:
(1) Create a 513 TB image with cluster_size=2M,refcount_bits=1
(2) Take a hex editor and enter 16 refblocks into the reftable
(3) Fill all of those refblocks with 1s

(Funny side note: qemu-img check thinks that image is clean because it doesn't check refcounts beyond the image end...)

I've attached a compressed test image (unsurprisingly, it compresses really well).

Before this series:

$ ./qemu-io -c 'write -c 0 2M' test.qcow2
qcow2: Marking image as corrupt: Preventing invalid write on metadata
(overlaps with refcount block); further corruption events will be suppressed
write failed: Input/output error

Aw.

After this series:

$ ./qemu-io -c 'write -c 0 2M' test.qcow2
write failed: Input/output error

(Normal writes just work fine.)

Maybe you want to add a test still -- creating the image is rather quick (well, you have to write 64 MB of 1s, but other than that). The only thing that takes a bit of time is qemu figuring out where the first free cluster is... That takes like 15 seconds here.

And another issue of course is...

$ ls -lhs test.qcow2
42M -rw-r--r--. 1 maxx maxx 513T 25. Apr 16:42 test.qcow2

Yeah, that. Depends on the host file system, of course, whether that is a real issue. O:-)

When reading a compressed image, we were allocating s->cluster_data to 32*cluster_size + 512 (possibly over 64 megabytes, for an image with 2M clusters). Let's check out the history:

Back when qcow2 was first written, we used s->cluster_data for everything, including copy_sectors() and encryption, where we want to operate on more than one cluster at once. Obviously, at that point, the buffer had to be aligned for other users, even though compression itself doesn't require any alignment (the fact that the compressed data generally starts mid-sector means that aligning our buffer buys us nothing - either the protocol already supports byte-based access into whatever offset we want, or we are already using a bounce buffer to read a full sector, and copying into our destination no longer requires alignment).

But commit 1b9f1491 (v1.1!) changed things to allocate parallel buffers on demand rather than sharing a single buffer, for encryption and COW, leaving compression as the final client of s->cluster_data. That use was still preserved, because if a single compressed cluster is read more than once, we reuse the cache instead of decompressing it a second time (someday, we may come up with better caching to avoid wasting repeated decompressions while still being more parallel, but that is a task for another patch; the XXX comment in qcow2_co_preadv for QCOW2_CLUSTER_COMPRESSED is telling).

Much later, in commit de82815d (v2.2), we noticed that a 64M allocation is prone to failure, so we switched over to a graceful memory allocation error message. Elsewhere in the code, we do g_malloc(2 * cluster_size) without ever checking for failure, but even 4M starts to be large enough that trying to be nice is worth the effort, so we want to keep that aspect.

Then even later, in 3e4c7052 (2.11), we realized that allocating a large buffer up front for every qcow2 image is expensive, and switched to lazy allocation only for images that actually had compressed clusters. But in the process, we never even bothered to check whether what we were allocating still made sense in its new context!

So, it's time to cut back on the waste. A compressed cluster written by qemu will NEVER occupy more than an uncompressed cluster, but based on mid-sector alignment, we may still need to read 1 cluster + 1 sector in order to recover enough bytes for the decompression. But third-party producers of qcow2 may not be as smart, and gzip DOES document that because the compression stream adds metadata, and because of the pigeonhole principle, there are worst case scenarios where attempts to compress will actually inflate an image, by up to 0.015% (or 62 sectors larger for an unfortunate 2M compression).

In fact,
the qcow2 spec permits an all-ones sector count, plus a full sector containing the initial offset, for a maximum read of 2 full clusters; and thanks to the way decompression works, even though such a value is probably too large for the actual compressed data, it really doesn't matter if we read too much (gzip ignores slop, once it has decoded a full cluster). So it's easier to just allocate cluster_data to be as large as we can ever possibly see; even if it still wastes up to 2M on any image created by qcow2, that's still an improvment of 60M less waste than pre-patch.

Signed-off-by Eric Blake <[email protected]>

Reviewed-by Alberto Garcia <[email protected]>

---
v4: fix off-by-one in assertion and commit message [Berto]
v3: tighten allocation [Berto]
v2: actually check allocation failure (previous version meant to use g_malloc, but ended up posted with g_try_malloc without checking); add assertions outside of conditional, improve commit message to better match reality now that qcow2 spec bug has been fixed

---
 block/qcow2-cluster.c | 27 ++++++++++++++++++---------
 block/qcow2.c         |  2 +-
 2 files changed, 19 insertions(+), 10 deletions(-)

block/qcow2-cluster.c +18 -9

index 1dc00ff2110..6e3eb88a37a 100644
@@ -1599,20 +1599,29 @@ int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset)
         sector_offset = coffset & 511;
         csize = nb_csectors * 512 - sector_offset;
 
-        /* Allocate buffers on first decompress operation, most images are
-         * uncompressed and the memory overhead can be avoided.  The buffers
-         * are freed in .bdrv_close().
+        /* Allocate buffers on the first decompress operation; most
+         * images are uncompressed and the memory overhead can be
+         * avoided.  The buffers are freed in .bdrv_close().  qemu
+         * never writes an inflated cluster, and gzip itself never
+         * inflates a problematic cluster by more than 0.015%, but the
+         * qcow2 format allows up to 1 byte short of 2 full clusters
+         * when including the sector containing offset.  gzip ignores
+         * trailing slop, so just allocate that much up front rather
+         * than reject third-party images with overlarge csize.
          */
+        assert(!!s->cluster_data == !!s->cluster_cache);
+        assert(csize <= 2 * s->cluster_size);
         if (!s->cluster_data) {
-            /* one more sector for decompressed data alignment */
-            s->cluster_data = qemu_try_blockalign(bs->file->bs,
-                    QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size + 512);
+            s->cluster_data = g_try_malloc(2 * s->cluster_size);
             if (!s->cluster_data) {
                 return -ENOMEM;
             }
-        }
-        if (!s->cluster_cache) {
-            s->cluster_cache = g_malloc(s->cluster_size);
+            s->cluster_cache = g_try_malloc(s->cluster_size);
+            if (!s->cluster_cache) {
+                g_free(s->cluster_data);
+                s->cluster_data = NULL;
+                return -ENOMEM;
+            }
         }
 
         BLKDBG_EVENT(bs->file, BLKDBG_READ_COMPRESSED);

block/qcow2.c +1 -1

index ef68772acad..a8301371ccc 100644
@@ -2151,7 +2151,7 @@ static void qcow2_close(BlockDriverState *bs)
     g_free(s->image_backing_format);
 
     g_free(s->cluster_cache);
-    qemu_vfree(s->cluster_data);
+    g_free(s->cluster_data);
     qcow2_refcount_close(bs);
     qcow2_free_snapshots(bs);
 }

Quoted 6 lines · Eric Blake

When reading a compressed image, we were allocating s->cluster_data to 32*cluster_size + 512 (possibly over 64 megabytes, for an image with 2M clusters). Let's check out the history:

Quoted 10 lines

So, it's time to cut back on the waste. A compressed cluster written by qemu will NEVER occupy more than an uncompressed cluster, but based on mid-sector alignment, we may still need to read 1 cluster + 1 sector in order to recover enough bytes for the decompression. But third-party producers of qcow2 may not be as smart, and gzip DOES document that because the compression stream adds metadata, and because of the pigeonhole principle, there are worst case scenarios where attempts to compress will actually inflate an image, by up to 0.015% (or 62 sectors larger for an unfortunate 2M compression).

Hm? 2M * 0.00015 < 315 (bytes!), so where does that number come from?

Quoted 12 lines

In fact,
the qcow2 spec permits an all-ones sector count, plus a full sector containing the initial offset, for a maximum read of 2 full clusters; and thanks to the way decompression works, even though such a value is probably too large for the actual compressed data, it really doesn't matter if we read too much (gzip ignores slop, once it has decoded a full cluster). So it's easier to just allocate cluster_data to be as large as we can ever possibly see; even if it still wastes up to 2M on any image created by qcow2, that's still an improvment of 60M less waste than pre-patch.

OK, so from the technical perspective it's irrelevant anyway, but I suppose the number should still be fixed in the commit message.