From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:41304) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1XdbV6-0001Qx-TL for qemu-devel@nongnu.org; Mon, 13 Oct 2014 05:01:15 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1XdbUz-0005h9-3Y for qemu-devel@nongnu.org; Mon, 13 Oct 2014 05:01:08 -0400 Received: from mx1.redhat.com ([209.132.183.28]:22300) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1XdbUy-0005h5-Sc for qemu-devel@nongnu.org; Mon, 13 Oct 2014 05:01:01 -0400 Message-ID: <543B94C5.1000706@redhat.com> Date: Mon, 13 Oct 2014 11:00:53 +0200 From: Max Reitz MIME-Version: 1.0 References: <201410091917519618804@sangfor.com>, <201410100954567266628@sangfor.com>, <543A80DA.4090201@redhat.com>, <201410131117118042731@sangfor.com>, <543B73F1.3090907@redhat.com>, <201410131513365349242@sangfor.com>, <543B872C.4090104@redhat.com> <201410131619475305762@sangfor.com> In-Reply-To: <201410131619475305762@sangfor.com> Content-Type: text/plain; charset=utf-8; format=flowed Content-Transfer-Encoding: 7bit Subject: Re: [Qemu-devel] [question] is it possible that big-endian l1 tableoffsetreferencedby other I/O while updating l1 table offset in qcow2_update_snapshot_refcount? List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Zhang Haoyu , Eric Blake , qemu-devel Cc: Kevin Wolf , Stefan Hajnoczi Am 13.10.2014 um 10:19 schrieb Zhang Haoyu: >>>>>>>>> Hi, >>>>>>>>> I encounter a problem that after deleting snapshot, the qcow2 image size is very larger than that it should be displayed by ls command, >>>>>>>>> but the virtual disk size is okay via qemu-img info. >>>>>>>>> I suspect that during updating l1 table offset, other I/O job reference the big-endian l1 table offset (very large value), >>>>>>>>> so the file is truncated to very large. >>>>>>>> Not quite. Rather, all the data that the snapshot used to occupy is >>>>>>>> still consuming holes in the file; the maximum offset of the file is >>>>>>>> still unchanged, even if the file is no longer using as many referenced >>>>>>>> clusters. Recent changes have gone in to sparsify the file when >>>>>>>> possible (punching holes if your kernel and file system is new enough to >>>>>>>> support that), so that it is not consuming the amount of disk space that >>>>>>>> a mere ls reports. But if what you are asking for is a way to compact >>>>>>>> the file back down, then you'll need to submit a patch. The idea of >>>>>>>> having an online defragmenter for qcow2 files has been kicked around >>>>>>>> before, but it is complex enough that no one has attempted a patch yet. >>>>>>> Sorry, I didn't clarify the problem clearly. >>>>>>> In qcow2_update_snapshot_refcount(), below code, >>>>>>> /* Update L1 only if it isn't deleted anyway (addend = -1) */ >>>>>>> if (ret == 0 && addend >= 0 && l1_modified) { >>>>>>> for (i = 0; i < l1_size; i++) { >>>>>>> cpu_to_be64s(&l1_table[i]); >>>>>>> } >>>>>>> >>>>>>> ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2); >>>>>>> >>>>>>> for (i = 0; i < l1_size; i++) { >>>>>>> be64_to_cpus(&l1_table[i]); >>>>>>> } >>>>>>> } >>>>>>> between cpu_to_be64s(&l1_table[i]); and be64_to_cpus(&l1_table[i]);, >>>>>>> is it possible that there is other I/O reference this interim l1 table whose entries contain the be64 l2 table offset? >>>>>>> The be64 l2 table offset maybe a very large value, hundreds of TB is possible, >>>>>>> then the qcow2 file will be truncated to far larger than normal size. >>>>>>> So we'll see the huge size of the qcow2 file by ls -hl, but the size is still normal displayed by qemu-img info. >>>>>>> >>>>>>> If the possibility mentioned above exists, below raw code may fix it, >>>>>>> if (ret == 0 && addend >= 0 && l1_modified) { >>>>>>> tmp_l1_table = g_malloc0(l1_size * sizeof(uint64_t)) >>>>>>> memcpy(tmp_l1_table, l1_table, l1_size * sizeof(uint64_t)); >>>>>>> for (i = 0; i < l1_size; i++) { >>>>>>> cpu_to_be64s(&tmp_l1_table[i]); >>>>>>> } >>>>>>> ret = bdrv_pwrite_sync(bs->file, l1_table_offset, tmp_l1_table, l1_size2); >>>>>>> >>>>>>> free(tmp_l1_table); >>>>>>> } >>>>>> l1_table is already a local variable (local to >>>>>> qcow2_update_snapshot_refcount()), so I can't really imagine how >>>>>> introducing another local buffer should mitigate the problem, if there >>>>>> is any. >>>>>> >>>>> l1_table is not necessarily a local variable to qcow2_update_snapshot_refcount, >>>>> which depends on condition of "if (l1_table_offset != s->l1_table_offset)", >>>>> if the condition not true, l1_table = s->l1_table. >>>> Oh, yes, you're right. Okay, so in theory nothing should happen anyway, >>>> because qcow2 does not have to be reentrant (so s->l1_table will not be >>>> accessed while it's big endian and therefore possibly not in CPU order). >>> Could you detail how qcow2 does not have to be reentrant? >>> In below stack, >>> qcow2_update_snapshot_refcount >>> |- cpu_to_be64s(&l1_table[i]) >>> |- bdrv_pwrite_sync >> This is executed on bs->file, not the qcow2 BDS. >> > Yes, bs->file is passed to bdrv_pwrite_sync here, > but aio_poll(aio_context) will poll all BDS's aio, not only that of bs->file, doesn't it? > Is it possible that there are pending aio which belong to this qcow2 BDS still exist? qcow2 is generally not reentrant, this is secured by locking (BDRVQcowState.lock). As long as one request for a BDS is still running, it will not be interrupted. Max > Thanks, > Zhang Haoyu >> Max >> >>> |-- bdrv_pwrite >>> |--- bdrv_pwritev >>> |---- bdrv_prwv_co >>> |----- aio_poll(aio_context) <== this aio_context is qemu_aio_context >>> |------ aio_dispatch >>> |------- bdrv_co_io_em_complete >>> |-------- qemu_coroutine_enter(co->coroutine, NULL); <== coroutine entry is bdrv_co_do_rw >>> bdrv_co_do_rw will access l1_table to perform I/O operation. >>> >>> Thanks, >>> Zhang Haoyu >>>> But I find it rather ugly to convert the cached L1 table to big endian, >>>> so I'd be fine with the patch you proposed. >>>> >>>> Max