| Message ID | 20200821141123.28538-1-vsementsov@virtuozzo.com |
|---|---|
| Headers | show |
| Series | preallocate filter | expand |
21.08.2020 17:11, Vladimir Sementsov-Ogievskiy wrote: > Hi all! > > Here is a filter, which does preallocation on write. > > In Virtuozzo we have to deal with some custom distributed storage > solution, where allocation is relatively expensive operation. We have to > workaround it in Qemu, so here is a new filter. I have a problem now with this thing. We need preallocation. But we don't want to explicitly specify it in all the management tools. So it should be inserted by default. It's OK for us to keep this default different from upstream... But there are problems with the implicitly inserted filter (actually iotests fail and I failed to fix them) 1. I have to set bs->inherits_from for filter and it's child by hand after bdrv_replace_node(), otherwise bdrv_check_perm doesn't work. 2. I have to set filter_bs->implicit and teach bdrv_refresh_filename() to ignore implicit filters when it checks for drv->bdrv_file_open, to avoid appearing of json in backing file names 3. And the real design problem, which seems impossible to fix: reopen is broken, just because user is not prepared to the fact that file child is a filter, not a file node and has another options, and don't support options of file-posix. And seems all it (and mostly [3]) shows that implicitly inserting the filter is near to be impossible.. So, what are possible solutions? In virtuozzo7 we have preallocation feature done inside qcow2 driver. This is very uncomfortable: we should to handle each possible over-EOF write to underlying node (to keep data_end in sync to be able to shrink preallocation on close()).. I don't like this way and don't want to port it.. Another option is implementing preallocation inside file-posix driver. Then, instead of BDRV_REQ_NO_WAIT flag I'll need to extend serialising requests API (bdrv_make_request_serialising() is already used in file-posix.c) to dupport no-wait behavior + expanding the serialising request bounds. This option seems feasible, so I'll try this way if no other ideas. Filter is obviously the true way: we use generic block layer for native request serialising, don't need to catch every write in qcow2 driver, don't need to modify any other driver and get a universal thing. But how to insert it implicitly (or at least automatically in some cases) and avoid all the problems?
On 27.08.20 23:08, Vladimir Sementsov-Ogievskiy wrote: > 21.08.2020 17:11, Vladimir Sementsov-Ogievskiy wrote: >> Hi all! >> >> Here is a filter, which does preallocation on write. >> >> In Virtuozzo we have to deal with some custom distributed storage >> solution, where allocation is relatively expensive operation. We have to >> workaround it in Qemu, so here is a new filter. > > I have a problem now with this thing. > > We need preallocation. But we don't want to explicitly specify it in all > the management tools. Why? > So it should be inserted by default. Why? You mean without any option? That seems... Interesting? (Also like a recipe for reports of performance regression in some cases.) > It's OK for > us to keep this default different from upstream... But there are > problems with the implicitly inserted filter (actually iotests fail and > I failed to fix them) I would suspect even if the iotests passed we would end up with a heap of problems that we would only notice at some later point. I thought you too weren’t too fond of the idea of implicit filters. > 1. I have to set bs->inherits_from for filter and it's child by hand > after bdrv_replace_node(), otherwise bdrv_check_perm doesn't work. > > 2. I have to set filter_bs->implicit and teach bdrv_refresh_filename() > to ignore implicit filters when it checks for drv->bdrv_file_open, to > avoid appearing of json in backing file names > > 3. And the real design problem, which seems impossible to fix: reopen is > broken, just because user is not prepared to the fact that file child is > a filter, not a file node and has another options, and don't support > options of file-posix. Well, what should I say. I feel like we have made efforts in the past years to make the block graph fully visible to users and yield the responsibility of managing it to the users, too, so I’m not surprised if a step backwards breaks that. > And seems all it (and mostly [3]) shows that implicitly inserting the > filter is near to be impossible.. > > So, what are possible solutions? > > In virtuozzo7 we have preallocation feature done inside qcow2 driver. > This is very uncomfortable: we should to handle each possible over-EOF > write to underlying node (to keep data_end in sync to be able to shrink > preallocation on close()).. I don't like this way and don't want to port > it.. > > Another option is implementing preallocation inside file-posix driver. > Then, instead of BDRV_REQ_NO_WAIT flag I'll need to extend serialising > requests API (bdrv_make_request_serialising() is already used in > file-posix.c) to dupport no-wait behavior + expanding the serialising > request bounds. This option seems feasible, so I'll try this way if no > other ideas. Possible, but you haven’t yet explained what the problem with the management layer inserting the preallocation filter is. > Filter is obviously the true way: we use generic block layer for native > request serialising, don't need to catch every write in qcow2 driver, > don't need to modify any other driver and get a universal thing. But how > to insert it implicitly (or at least automatically in some cases) and > avoid all the problems? I don’t understand why inserting it implicitly is important. Max
01.09.2020 18:07, Max Reitz wrote: > On 27.08.20 23:08, Vladimir Sementsov-Ogievskiy wrote: >> 21.08.2020 17:11, Vladimir Sementsov-Ogievskiy wrote: >>> Hi all! >>> >>> Here is a filter, which does preallocation on write. >>> >>> In Virtuozzo we have to deal with some custom distributed storage >>> solution, where allocation is relatively expensive operation. We have to >>> workaround it in Qemu, so here is a new filter. >> >> I have a problem now with this thing. >> >> We need preallocation. But we don't want to explicitly specify it in all >> the management tools. > > Why? > >> So it should be inserted by default. > > Why? You mean without any option? That seems... Interesting? > > (Also like a recipe for reports of performance regression in some cases.) > >> It's OK for >> us to keep this default different from upstream... But there are >> problems with the implicitly inserted filter (actually iotests fail and >> I failed to fix them) > > I would suspect even if the iotests passed we would end up with a heap > of problems that we would only notice at some later point. I thought > you too weren’t too fond of the idea of implicit filters. > >> 1. I have to set bs->inherits_from for filter and it's child by hand >> after bdrv_replace_node(), otherwise bdrv_check_perm doesn't work. >> >> 2. I have to set filter_bs->implicit and teach bdrv_refresh_filename() >> to ignore implicit filters when it checks for drv->bdrv_file_open, to >> avoid appearing of json in backing file names >> >> 3. And the real design problem, which seems impossible to fix: reopen is >> broken, just because user is not prepared to the fact that file child is >> a filter, not a file node and has another options, and don't support >> options of file-posix. > > Well, what should I say. I feel like we have made efforts in the past > years to make the block graph fully visible to users and yield the > responsibility of managing it to the users, too, so I’m not surprised if > a step backwards breaks that. > >> And seems all it (and mostly [3]) shows that implicitly inserting the >> filter is near to be impossible.. >> >> So, what are possible solutions? >> >> In virtuozzo7 we have preallocation feature done inside qcow2 driver. >> This is very uncomfortable: we should to handle each possible over-EOF >> write to underlying node (to keep data_end in sync to be able to shrink >> preallocation on close()).. I don't like this way and don't want to port >> it.. >> >> Another option is implementing preallocation inside file-posix driver. >> Then, instead of BDRV_REQ_NO_WAIT flag I'll need to extend serialising >> requests API (bdrv_make_request_serialising() is already used in >> file-posix.c) to dupport no-wait behavior + expanding the serialising >> request bounds. This option seems feasible, so I'll try this way if no >> other ideas. > > Possible, but you haven’t yet explained what the problem with the > management layer inserting the preallocation filter is. > >> Filter is obviously the true way: we use generic block layer for native >> request serialising, don't need to catch every write in qcow2 driver, >> don't need to modify any other driver and get a universal thing. But how >> to insert it implicitly (or at least automatically in some cases) and >> avoid all the problems? > > I don’t understand why inserting it implicitly is important. > You are right. Thanks for strong point of view, this makes me to revise my own. Now I'm working on v6.