[opensuse-kernel] How to solve the delayed grant mapping problem with blkback?
Hello, Recently I got a problem when I use virtual machines in Xen virtualization, description just as follws. Grant table and grant mapping were used for data communication with in Xen front end driver and back end driver. A delay work will be set up to release grant mapping pages in end_block_io_op as bio callbacks, the code as below: static void __gnttab_unmap_refs_async(struct gntab_unmap_queue_data* item) { int ret; int pc; for (pc = 0; pc < item->count; pc++) { if (page_count(item->pages[pc]) > 1) { unsigned long delay = GNTTAB_UNMAP_REFS_DELAY * (item->age + 1); schedule_delayed_work(&item->gnttab_work, msecs_to_jiffies(delay)); return; } } ret = gnttab_unmap_refs(item->unmap_ops, item->kunmap_ops, item->pages, item->count); item->done(ret, item); } Only if the page count of grant mapping page was decrease to one, a response would be made to notify the vm front end. But in the scenario which a network cable was extracted, the TCP transport layer will retransmit the socket data. As a result, the grant mapping pages were occupied by the TCP retransmition which may continued for a long time. And then in blkback driver, the grant unmap delay work will continue until the TCP releases all the pages. As an application of TCP protocol, Is there any method to solve this problem in iscsi driver layer? I have already attempted to modified some parameters provide by iscsi, but it was a futile effort. The socket connection seems not be closed in the network cable extracted scenario. Thanks -- To unsubscribe, e-mail: opensuse-kernel+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-kernel+owner@opensuse.org
On 18.03.17 at 04:53, <xieyingtai@huawei.com> wrote: Hello, Recently I got a problem when I use virtual machines in Xen virtualization, description just as follws. Grant table and grant mapping were used for data communication with in Xen front end driver and back end driver. A delay work will be set up to release grant mapping pages in end_block_io_op as bio callbacks, the code as below: static void __gnttab_unmap_refs_async(struct gntab_unmap_queue_data* item) { int ret; int pc;
for (pc = 0; pc < item->count; pc++) { if (page_count(item->pages[pc]) > 1) { unsigned long delay = GNTTAB_UNMAP_REFS_DELAY * (item->age + 1); schedule_delayed_work(&item->gnttab_work, msecs_to_jiffies(delay)); return; } }
ret = gnttab_unmap_refs(item->unmap_ops, item->kunmap_ops, item->pages, item->count); item->done(ret, item); }
Only if the page count of grant mapping page was decrease to one, a response would be made to notify the vm front end. But in the scenario which a network cable was extracted, the TCP transport layer will retransmit the socket data. As a result, the grant mapping pages were occupied by the TCP retransmition which may continued for a long time. And then in blkback driver, the grant unmap delay work will continue until the TCP releases all the pages. As an application of TCP protocol, Is there any method to solve this problem in iscsi driver layer? I have already attempted to modified some parameters provide by iscsi, but it was a futile effort. The socket connection seems not be closed in the network cable extracted scenario.
I guess the primary option to check here is, just like in the network drivers, whether copying the SKBs at a suitable point in time can be arranged for. But this not being an openSUSE-specific problem (afaict), I think you'd better ask on xen-devel and/or lkml. Jan -- To unsubscribe, e-mail: opensuse-kernel+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-kernel+owner@opensuse.org
participants (2)
-
Jan Beulich -
Xieyingtai