Re: [Libguestfs] [PATCH libnbd v4 2/2] copy: Use preferred block size for copying

You're not supposed to read or write NBD servers at a granularity less than the advertised minimum block size. nbdcopy has ignored this requirement, and this is usually fine because the NBD servers we care about support 512-byte sector granularity, and never advertise sizes / extents less granular than sectors (even if it's a bit suboptimal in a few cases).

However there is one new case where we do care: When writing to a compressed qcow2 file, qemu advertises a minimum and preferred block size of 64K, and it really means it. You cannot write blocks smaller than this because of the way qcow2 compression is implemented. This commit attempts to do the least work possible to fix this. The previous multi-thread-copying loop was driven by the extent map received from the source. I have modified the loop so that it iterates over request_size blocks. request_size is set from the command line (--request-size) but will be adjusted upwards if either the source or destination preferred block size is larger. So this will always copy blocks which are at least the preferred block size (except for the very last block of the disk).

While copying these blocks we consult the source extent map. If it contains only zero regions covering the whole block (only_zeroes function) then we can skip straight to zeroing the target (fill_dst_range_with_zeroes), else we do read + write as before. I only modified the multi-thread-copying loop, not the synchronous loop. That should be updated in the same way later.

One side effect of this change is it always makes larger requests, even for regions we know are sparse. This is clear in the copy-sparse.sh and copy-sparse-allocated.sh tests which were previously driven by the 32K sparse map granularity of the source. Without changing these tests, they would make make 256K reads & writes (and also read from areas of the disk even though we know they are sparse). I adjusted these tests to use --request-size=32768 to force the existing behaviour. Note this doesn't attempt to limit the maximum block size when reading or writing. That is for future work. This is a partial fix for https://bugzilla.redhat.com/2047660. Further changes will be required in virt-v2v. Link: https://lists.gnu.org/archive/html/qemu-block/2022-01/threads.html#00729 Link: https://bugzilla.redhat.com/show_bug.cgi?id=2047660 --- TODO | 4 +- copy/Makefile.am | 6 +- copy/copy-file-to-qcow2-compressed.sh | 64 +++++++++++ copy/copy-sparse-allocated.sh | 4 +- copy/copy-sparse.sh | 7 +- copy/main.c | 13 +++ copy/multi-thread-copying.c | 149 +++++++++++++++++++------- copy/nbdcopy.pod | 5 +- 8 files changed, 202 insertions(+), 50 deletions(-) diff --git a/TODO b/TODO index 7c9c15e245..bc38d7045e 100644 --- a/TODO +++ b/TODO @@ -28,7 +28,9 @@ Performance: Chart it over various buffer sizes and threads, as that Examine other fuzzers: https://gitlab.com/akihe/radamsa nbdcopy: - - Minimum/preferred/maximum block size. + - Enforce maximum block size. + - Synchronous loop should be adjusted to take into account + the NBD preferred block size, as was done for multi-thread loop. - Benchmark. - Better page cache usage, see nbdkit-file-plugin options fadvise=sequential cache=none.

+++ b/copy/copy-file-to-qcow2-compressed.sh

+ +# Create a compressed qcow2 file1. +# +# sparse-random files should compress easily because by default each +# block uses repeated bytes. +qemu-img create -f qcow2 $file1 $size +nbdcopy -- [ nbdkit --exit-with-parent sparse-random $size seed=$seed ] \ + [ $QEMU_NBD --image-opts driver=compress,file.driver=qcow2,file.file.driver=file,file.file.filename=$file1 ]

+ +ls -l $file1 + +# Create an uncompressed qcow2 file2 with the same data. +qemu-img create -f qcow2 $file2 $size +nbdcopy -- [ nbdkit --exit-with-parent sparse-random $size seed=$seed ] \ + [ $QEMU_NBD --image-opts driver=qcow2,file.driver=file,file.filename=$file2 ] + +ls -l $file2 + +# file1 < file2 (shows the compression is having some effect). +size1="$( stat -c %s $file1 )" +size2="$( stat -c %s $file2 )" +if [ $size1 -ge $size2 ]; then + echo "$0: qcow2 compression did not make the file smaller" + exit 1 +fi + +# Logical content of the files should be identical. +qemu-img compare -f qcow2 $file1 -F qcow2 $file2

+++ b/copy/main.c @@ -40,6 +40,7 @@ #include "ispowerof2.h" #include "human-size.h" +#include "minmax.h" #include "version.h" #include "nbdcopy.h" @@ -379,10 +380,22 @@ main (int argc, char *argv[]) if (threads < connections) connections = threads; + /* request_size must always be at least as large as the preferred + * size of source & destination. + */ + request_size = MAX (request_size, src->preferred); + request_size = MAX (request_size, dst->preferred); + /* Adapt queue to size to request size if needed. */ if (request_size > queue_size) queue_size = request_size; + /* Sparse size (if using) must not be smaller than the destination + * preferred size, otherwise we end up creating too small requests. + */ + if (sparse_size > 0 && sparse_size < dst->preferred) + sparse_size = dst->preferred; +

/* Truncate the destination to the same size as the source. Only * has an effect on regular files. */ diff --git a/copy/multi-thread-copying.c b/copy/multi-thread-copying.c index 06cdb8eef6..369b8167a1 100644 --- a/copy/multi-thread-copying.c +++ b/copy/multi-thread-copying.c @@ -166,6 +166,62 @@ decrease_queue_size (struct worker *worker, size_t len) worker->queue_size -= len; } +/* Using the extents map 'exts', check if the region + * [offset..offset+len-1] intersects only with zero extents. + * + * The invariant for '*i' is always an extent which starts before or + * equal to the current offset. + */ +static bool +only_zeroes (const extent_list exts, size_t *i, + uint64_t offset, unsigned len) +{

@@ -177,7 +233,10 @@ worker_thread (void *wp) extent_list exts = empty_vector; while (get_next_offset (&offset, &count)) { - size_t i; + struct command *command; + size_t extent_index; + bool is_zeroing = false; + uint64_t zeroing_start = 0; assert (0 < count && count <= THREAD_WORK_SIZE); if (extents) @@ -185,52 +244,64 @@ worker_thread (void *wp) else default_get_extents (src, w->index, offset, count, &exts); - for (i = 0; i < exts.len; ++i) { - struct command *command; - size_t len; + extent_index = 0; // index into extents array used to optimize only_zeroes + while (count) { + const size_t len = MIN (count, request_size); - if (exts.ptr[i].zero) { + if (only_zeroes (exts, &extent_index, offset, len)) { /* The source is zero so we can proceed directly to skipping, - * fast zeroing, or writing zeroes at the destination. + * fast zeroing, or writing zeroes at the destination. Defer + * zeroing so we can send it as a single large command. */ - command = create_command (exts.ptr[i].offset, exts.ptr[i].length, - true, w); - fill_dst_range_with_zeroes (command); + if (!is_zeroing) { + is_zeroing = true; + zeroing_start = offset; + } } - else /* data */ { - /* As the extent might be larger than permitted for a single - * command, we may have to split this into multiple read - * requests. - */ - while (exts.ptr[i].length > 0) { - len = exts.ptr[i].length; - if (len > request_size) - len = request_size; - - command = create_command (exts.ptr[i].offset, len, - false, w); - - wait_for_request_slots (w); - - /* NOTE: Must increase the queue size after waiting. */ - increase_queue_size (w, len); - - /* Begin the asynch read operation. */ - src->ops->asynch_read (src, command, - (nbd_completion_callback) { - .callback = finished_read, - .user_data = command, - }); - - exts.ptr[i].offset += len; - exts.ptr[i].length -= len; + /* If we were in the middle of deferred zeroing, do it now. */ + if (is_zeroing) { + /* Note that offset-zeroing_start can never exceed + * THREAD_WORK_SIZE, so there is no danger of overflowing + * size_t.

+ */ + command = create_command (zeroing_start, offset-zeroing_start, + true, w); + fill_dst_range_with_zeroes (command); + is_zeroing = false; } + + /* Issue the asynchronous read command. */ + command = create_command (offset, len, false, w); + + wait_for_request_slots (w); + + /* NOTE: Must increase the queue size after waiting. */ + increase_queue_size (w, len); + + /* Begin the asynch read operation. */ + src->ops->asynch_read (src, command, + (nbd_completion_callback) { + .callback = finished_read, + .user_data = command, + });

} - offset += count; - count = 0; - } /* for extents */ + offset += len; + count -= len; + } /* while (count) */ + + /* If we were in the middle of deferred zeroing, do it now. */ + if (is_zeroing) { + /* Note that offset-zeroing_start can never exceed + * THREAD_WORK_SIZE, so there is no danger of overflowing + * size_t. + */ + command = create_command (zeroing_start, offset - zeroing_start, + true, w); + fill_dst_range_with_zeroes (command); + is_zeroing = false; + } } /* Wait for in flight NBD requests to finish. */ diff --git a/copy/nbdcopy.pod b/copy/nbdcopy.pod index 94c713f68f..501d6fb245 100644 --- a/copy/nbdcopy.pod +++ b/copy/nbdcopy.pod @@ -182,8 +182,9 @@ Set the maximum number of requests in flight per NBD connection. =item B<--sparse=>N Detect all zero blocks of size N (bytes) and make them sparse on the -output. You can also turn off sparse detection using S<I<-S 0>>. -The default is 4096 bytes. +output. You can also turn off sparse detection using S<I<-S 0>>. The +default is 4096 bytes, or the destination preferred block size, +whichever is larger.