On Wed, Jun 29, 2022 at 04:56:20PM +0200, Laszlo Ersek wrote:

On 06/29/22 13:03, Richard W.M. Jones wrote:

...

This will be used in a subsequent commit. At the moment the preferred block size for all sources / destinations is simply calculated and stored. --- copy/file-ops.c | 9 ++++++++- copy/main.c | 9 ++++++--- copy/nbd-ops.c | 9 +++++++++ copy/nbdcopy.h | 4 +++- copy/null-ops.c | 1 + copy/pipe-ops.c | 1 + 6 files changed, 28 insertions(+), 5 deletions(-)

diff --git a/copy/file-ops.c b/copy/file-ops.c index ab3787545c..3b901bcdb8 100644 --- a/copy/file-ops.c +++ b/copy/file-ops.c @@ -241,7 +241,8 @@ seek_hole_supported (int fd)

struct rw * file_create (const char *name, int fd, - off_t st_size, bool is_block, direction d) + off_t st_size, blksize_t st_blksize, + bool is_block, direction d) { struct rw_file *rwf = calloc (1, sizeof *rwf); if (rwf == NULL) { perror ("calloc"); exit (EXIT_FAILURE); } @@ -262,6 +263,11 @@ file_create (const char *name, int fd, perror ("lseek"); exit (EXIT_FAILURE); } + rwf->rw.preferred = 4096; +#ifdef BLKIOOPT + if (ioctl (fd, BLKIOOPT, &rwf->rw.preferred)) + fprintf (stderr, "warning: cannot get optimal I/O size: %s: %m", name); +#endif

...

From my reading of the kernel, this ioctl stores an "unsigned int", which (in practice) will not set all bytes in the "preferred" uint64_t field.

Additionally, if the ioctl fails, I'm not sure if we have a guarantee that the destination will be left unchanged. I suggest:

#ifdef BLKIOOPT unsigned int blkioopt;

if (ioctl (fd, BLKIOOPT, &blkioopt)) rwf->rw.preferred = blkioopt; else { fprintf (stderr, "warning: cannot get optimal I/O size: %s: %m", name); rwf->rw.preferred = 4096; } #else rwf->rw.preferred = 4096; #endif

Ouch yes, and there's a similar bug in existing code too :-(

... however, I also suggest doing this *elsewhere* (not here). I think this function should take a plain uint64_t "preferred" param, and simply assign that to "rwf->rw.preferred", independently of "is_block". (There is one call site only, so moving the logic out there does not lead to code duplication.) I'll explain the reason below.

...

rwf->seek_hole_supported = seek_hole_supported (fd); rwf->sector_size = 4096; #ifdef BLKSSZGET @@ -282,6 +288,7 @@ file_create (const char *name, int fd, else { /* Regular file. */ rwf->rw.size = st_size; + rwf->rw.preferred = st_blksize; rwf->seek_hole_supported = seek_hole_supported (fd); /* Possible efficient zero methods for regular file. */ #ifdef FALLOC_FL_PUNCH_HOLE diff --git a/copy/main.c b/copy/main.c index cc379e98bc..9b551de664 100644 --- a/copy/main.c +++ b/copy/main.c @@ -513,7 +513,9 @@ open_local (const char *filename, direction d) exit (EXIT_FAILURE); } if (S_ISBLK (stat.st_mode) || S_ISREG (stat.st_mode)) - return file_create (filename, fd, stat.st_size, S_ISBLK (stat.st_mode), d); + return file_create (filename, fd, + stat.st_size, stat.st_blksize, S_ISBLK (stat.st_mode), + d); else { /* Probably stdin/stdout, a pipe or a socket. */ synchronous = true; /* Force synchronous mode for pipes. */

I couldn't find any info on whether fstat() fills in "stat.st_blksize" at all for block devices. I understand that with "is_block" set, the potential garbage will be ignored in file_create() anyway, but passing garbage *itself* is not great.

So I suggest to split

S_ISBLK (stat.st_mode) || S_ISREG (stat.st_mode)

to two separate branche. Pass stat.st_blksize (as "uint64_t preferred") on the S_ISREG branch. Do the following on the S_ISBLK branch:

unsigned int blkioopt;

#ifdef BLKIOOPT if (ioctl (fd, BLKIOOPT, &blkioopt) == -1) { fprintf (stderr, "warning: cannot get optimal I/O size: %s: %m", filename); blkioopt = 4096; } #else blkioopt = 4096; #endif

return file_create (filename, fd, stat.st_size, blkioopt, false, d);

Yup, good idea. The existing code has "evolved" and was never very well structured in the first place ...

...

@@ -528,8 +530,9 @@ print_rw (struct rw *rw, const char *prefix, FILE *fp) char buf[HUMAN_SIZE_LONGEST];

fprintf (fp, "%s: %s \"%s\"\n", prefix, rw->ops->ops_name, rw->name); - fprintf (fp, "%s: size=%" PRIi64 " (%s)\n", - prefix, rw->size, human_size (buf, rw->size, NULL)); + fprintf (fp, "%s: size=%" PRIi64 " (%s), preferred block size=%" PRIu64 "\n", + prefix, rw->size, human_size (buf, rw->size, NULL), + rw->preferred); }

/* Default implementation of rw->ops->get_extents for backends which diff --git a/copy/nbd-ops.c b/copy/nbd-ops.c index 3bc26ba613..6c5d59c578 100644 --- a/copy/nbd-ops.c +++ b/copy/nbd-ops.c @@ -113,11 +113,20 @@ open_one_nbd_handle (struct rw_nbd *rwn) */ if (rwn->handles.len == 0) { rwn->can_zero = nbd_can_zero (nbd) > 0; + rwn->rw.size = nbd_get_size (nbd); if (rwn->rw.size == -1) {

This existent code is fine, as nbd_get_size returns int64_t, and "rwn->rw.size" also has type int64_t...

...

fprintf (stderr, "%s: %s: %s\n", prog, rwn->rw.name, nbd_get_error ()); exit (EXIT_FAILURE); } + + rwn->rw.preferred = nbd_get_block_size (nbd, LIBNBD_SIZE_PREFERRED); + if (rwn->rw.preferred == -1) {

... but this is hard to read (albeit correct in practice).

In case nbd_get_block_size() fails, it will return (int64_t)-1, and we end up assigning "rwn->rw.preferred" UINT64_MAX. In the comparison on the next line, the (int)-1 will in practice be converted to uint64_t, so we do the comparison against UINT64_MAX. Correct, but hard to read.

I'd use a separate "int64_t block_size" local variable for this:

block_size = nbd_get_block_size (...); if (block_size == -1) { fprintf (...); exit (...); } rwn->rw.preferred = block_size == 0 ? 4096 : block_size;

But, replacing -1 with UINT64_MAX, or else (uint64_t)-1, would be informative too.

OK

...

+ fprintf (stderr, "%s: %s: %s\n", prog, rwn->rw.name, nbd_get_error ()); + exit (EXIT_FAILURE); + } + if (rwn->rw.preferred == 0) + rwn->rw.preferred = 4096; }

if (handles_append (&rwn->handles, nbd) == -1) { diff --git a/copy/nbdcopy.h b/copy/nbdcopy.h index 19797dfd66..9b02ddf270 100644 --- a/copy/nbdcopy.h +++ b/copy/nbdcopy.h @@ -43,6 +43,7 @@ struct rw { struct rw_ops *ops; /* Operations. */ const char *name; /* Printable name, for error messages etc. */ int64_t size; /* May be -1 for streams. */ + uint64_t preferred; /* Preferred block size. */ /* Followed by private data for the particular subtype. */ };

@@ -53,7 +54,8 @@ typedef enum { READING, WRITING } direction;

/* Create subtypes. */ extern struct rw *file_create (const char *name, int fd, - off_t st_size, bool is_block, direction d); + off_t st_size, blksize_t st_blksize, + bool is_block, direction d); extern struct rw *nbd_rw_create_uri (const char *name, const char *uri, direction d); extern struct rw *nbd_rw_create_subprocess (const char **argv, size_t argc, diff --git a/copy/null-ops.c b/copy/null-ops.c index 1218a623e2..854e291613 100644 --- a/copy/null-ops.c +++ b/copy/null-ops.c @@ -45,6 +45,7 @@ null_create (const char *name) rw->rw.ops = &null_ops; rw->rw.name = name; rw->rw.size = INT64_MAX; + rw->rw.preferred = 4 * 1024 * 1024; return &rw->rw; }

diff --git a/copy/pipe-ops.c b/copy/pipe-ops.c index 3c8b6c2b39..3815f824fc 100644 --- a/copy/pipe-ops.c +++ b/copy/pipe-ops.c @@ -43,6 +43,7 @@ pipe_create (const char *name, int fd) rwp->rw.ops = &pipe_ops; rwp->rw.name = name; rwp->rw.size = -1; + rwp->rw.preferred = 4096; rwp->fd = fd; return &rwp->rw; }

Thanks, Laszlo

Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-p2v converts physical machines to virtual machines. Boot with a live CD or over the network (PXE) and turn machines into KVM guests. http://libguestfs.org/virt-v2v

On 06/29/22 13:03, Richard W.M. Jones wrote:

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).

Does this mean that, if we have a large hole, we won't skip it on the source / zero it on the destination in one operation (assuming that's what we've been doing, that is), but read through it dutifully in request_size blocks? If that's the case, I'm totally OK with that; it's just exactly the kind of "sacrifice" that I felt would be necessary here!

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.

OK, so we subdivide the request_size-sized block still. I'm curious (and will later look) whether we'll rely on a previously fetched extent map (fetched for the whole block device), or if we get the extent-set that covers the individual request. In the former (more efficient) case, we'll effectively have to build an intersection between the request byte interval, and the intervals in the extent map.

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,

Why? Considering an NBD server where minimum=preferred=maximum block size, what is the difference? (In fact I'd have thought this change causes nbdcopy to use *smaller* requests than before, exactly because now we'll advance in fixed size blocks, not in extent-sized jumps!)

even for regions we know are sparse.

A sparse region is specifically where I'd expect this change to slow down progress (use smaller requests than what would be permitted by a big hole).

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.

What does this mean? Was the test image (well, "test device") intentionally populated by alternating, 32KB, "data" and "gap" extents?

Without changing these tests, they would make make 256K reads & writes

Why? What advertizes a 256K preferred block size? nbdkit?

(and also read from areas of the disk even though we know they are sparse).

But that's unavoidable, no? (I mean it's inevitable that we now investigate (check for zeroes) the internals of these gaps, at request boundaries, against the extent map; we can still avoid actually reading those bytes.)

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/copy-sparse-allocated.sh | 4 +- copy/copy-sparse.sh | 7 +- copy/main.c | 7 ++ copy/multi-thread-copying.c | 143 +++++++++++++++++++++++++--------- 5 files changed, 119 insertions(+), 46 deletions(-)

I need to understand the design better here, before I can look at the code. Thanks Laszlo

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. diff --git a/copy/copy-sparse-allocated.sh b/copy/copy-sparse-allocated.sh index 203c3b9850..465e347067 100755 --- a/copy/copy-sparse-allocated.sh +++ b/copy/copy-sparse-allocated.sh @@ -17,8 +17,6 @@ # Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

# Adapted from copy-sparse.sh. -# -# This test depends on the nbdkit default sparse block size (32K).

. ../tests/functions.sh

@@ -33,7 +31,7 @@ requires nbdkit eval --version out=copy-sparse-allocated.out cleanup_fn rm -f $out

-$VG nbdcopy --allocated -- \ +$VG nbdcopy --allocated --request-size=32768 -- \ [ nbdkit --exit-with-parent data data=' 1 @1073741823 1 diff --git a/copy/copy-sparse.sh b/copy/copy-sparse.sh index 1a6da868e9..7912a2120a 100755 --- a/copy/copy-sparse.sh +++ b/copy/copy-sparse.sh @@ -16,8 +16,6 @@ # License along with this library; if not, write to the Free Software # Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

-# This test depends on the nbdkit default sparse block size (32K). - . ../tests/functions.sh

set -e @@ -34,8 +32,9 @@ cleanup_fn rm -f $out # Copy from a sparse data disk to an nbdkit-eval-plugin instance which # is logging everything. This allows us to see exactly what nbdcopy # is writing, to ensure it is writing and zeroing the target as -# expected. -$VG nbdcopy -S 0 -- \ +# expected. Force request size to match nbdkit default sparse +# allocator block size (32K). +$VG nbdcopy -S 0 --request-size=32768 -- \ [ nbdkit --exit-with-parent data data=' 1 @1073741823 1 diff --git a/copy/main.c b/copy/main.c index 9b551de664..cbfe5b5436 100644 --- a/copy/main.c +++ b/copy/main.c @@ -40,6 +40,7 @@

#include "ispowerof2.h" #include "human-size.h" +#include "minmax.h" #include "version.h" #include "nbdcopy.h"

@@ -379,6 +380,12 @@ main (int argc, char *argv[]) if (threads < connections) connections = threads;

+ /* request_size must always be larger than preferred size of + * source or 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; diff --git a/copy/multi-thread-copying.c b/copy/multi-thread-copying.c index 06cdb8eef6..7194f4fafd 100644 --- a/copy/multi-thread-copying.c +++ b/copy/multi-thread-copying.c @@ -166,6 +166,58 @@ 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] contains only 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) +{ + size_t j; + + /* Invariant. */ + assert (0 <= *i && *i < exts.len); + assert (exts.ptr[*i].offset <= offset); + + /* Update the variant. Search for the last possible extent in the + * list which is <= offset. + */ + for (j = *i + 1; j < exts.len; ++j) { + if (exts.ptr[j].offset <= offset) + *i = j; + else + break; + } + + /* Check invariant again. */ + assert (0 <= *i && *i < exts.len); + assert (exts.ptr[*i].offset <= offset); + + /* Search forward, look for any non-zero extents covering the region. */ + for (j = *i; j < exts.len; ++j) { + uint64_t start, end; + + /* [start..end] of the current extent. */ + start = exts.ptr[j].offset; + end = exts.ptr[j].offset + exts.ptr[j].length; + + if (end <= offset) + continue; + + if (start >= offset + len) + break; + + /* Non-zero extent covering this region => test failed. */ + if (!exts.ptr[j].zero) + return false; + } + + return true; +} + /* There are 'threads' worker threads, each copying work ranges from * src to dst until there are no more work ranges. */ @@ -177,7 +229,10 @@ worker_thread (void *wp) extent_list exts = empty_vector;

while (get_next_offset (&offset, &count)) { + struct command *command; size_t i; + bool is_zeroing = false; + uint64_t zeroing_start = 0;

assert (0 < count && count <= THREAD_WORK_SIZE); if (extents) @@ -185,52 +240,64 @@ 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; + i = 0; // index into extents array + while (count) { + const size_t len = MIN (count, request_size);

- if (exts.ptr[i].zero) { + if (only_zeroes (exts, &i, 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. */