[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [PATCH v2 13/17] virfdstream: Allow sparse stream vol-download



On 8/20/20 3:42 PM, Peter Krempa wrote:
On Thu, Aug 20, 2020 at 15:31:28 +0200, Michal Privoznik wrote:
On 8/20/20 1:57 PM, Peter Krempa wrote:
On Tue, Jul 07, 2020 at 21:46:31 +0200, Michal Privoznik wrote:
When handling sparse stream, a thread is executed. This thread
runs a read() or write() loop (depending what API is called; in
this case it's virStorageVolDownload() and  this the thread run
read() loop). The read() is handled in virFDStreamThreadDoRead()
which is then data/hole section aware, meaning it uses
virFileInData() to detect data and hole sections and sends
TYPE_DATA or TYPE_HOLE virStream messages accordingly.

However, virFileInData() does not work with block devices. Simply
because block devices don't have data and hole sections. But we
can use new virFileInDataDetectZeroes() which is block device
friendly for that.

Partially resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1852528

Signed-off-by: Michal Privoznik <mprivozn redhat com>
---
   src/util/virfdstream.c | 15 ++++++++++++---
   1 file changed, 12 insertions(+), 3 deletions(-)

IMO this goes against the semantics of the _SPARSE_STREAM flag. A block
device by definition is not sparse, so there are no holes to send.

What you've implemented is a way to sparsify a block device, but that
IMO should not be considered by default when a block device is used.
If a file is not sparse, the previous code doesn't actually transmit
holes either.

If you want to achieve sparsification on the source side of the
transmission, this IMO needs an explicit flag to opt-in and then we
should sparsify also regular files using the same algorithm.


Fair enough. So how about I'll send v3 where:

a) in the first patches I make our stream read/write functions handle block
devices for _SPARSE_STREAM without any zero block detection. Only thing that
will happen is that if the source is a sparse regular file and thus the
stream receiver gets a HOLE packet and it is writing the data into a block
device it will have to emulate the hole by writing block of zeroes. However,
if the stream source is a block device then no HOLE shall ever be sent.

AFAIK I've R-b'd enough patches to fix this portion and provided that
there aren't any merge conflicts you can already commit those.

I'm completely fine with that portion as-is.

Almost :-)
For instance this very patch uses virFileInDataDetectZeroes() to detect zero blocks on block devices. It needs to be changed to always assume data section and some length. The same applies to the next patch 14/17.
But the diff is trivial:

iff --git c/src/util/virfdstream.c w/src/util/virfdstream.c
index 9968cdc623..39514ef555 100644
--- c/src/util/virfdstream.c
+++ w/src/util/virfdstream.c
@@ -440,8 +440,15 @@ virFDStreamThreadDoRead(virFDStreamDataPtr fdst,

     if (sparse && *dataLen == 0) {
         if (isBlock) {
-            if (virFileInDataDetectZeroes(fdin, &inData, &sectionLen) < 0)
-                return -1;
+            /* Block devices are always in data section by definition. The
+ * @sectionLen is slightly more tricky. While we could try and get + * how much bytes is there left until EOF, we can pretend there is + * always X bytes left and let the saferead() below hit EOF (which + * is then handled gracefully anyway). Worst case scenario, this
+             * branch is called more than once.
+             * X was chosen to be 1MiB but it has ho special meaning. */
+            inData = 1;
+            sectionLen = 1 * 1024 * 1024;

And the same for virsh case. Do you want me to resend those two patches?



b) in next patches I'll introduce _DETECT_ZEROES flag (and possibly make it
require _SPARSE_STREAM too) which will handle the case where the stream
source is a block device with zero blocks, at which point it will try to
detect them and be allowed to send HOLE down the stream.

On this topic, I agree that it's a sensible approach for the rest of the
series and it at least unifies the behaviour.

I'm unsure though whether it's worth even doing _DETECT_ZEROES feature
at all though. To me it feels that the users are better off using other
tools rather than re-implementing yet another thing in libvirt.

Alright. Fair enough I guess.


If possible provide some additional justification here.


It was discussed in the bz https://bugzilla.redhat.com/show_bug.cgi?id=1852528 VDSM is doing a thin provisioning and as a part of that they are copying files onto block devices. But for that zero detection shouldn't be needed.

Michal


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]