Xen devel - Jan 2012 - [PATCH] qemu-dm: add command to flush buffer cache

Add support for a xenstore dm command to flush qemu''s buffer cache.

qemu will just keep mapping pages and not release them, which causes problems
for the memory pager (since the page is mapped, it won''t get paged
out). When
the pager has trouble finding a page to page out, it asks qemu to flush its
buffer, which releases all the page mappings. This makes it possible to find
pages to swap out agian.

Signed-off-by: Patrick Colp <Patrick.Colp@citrix.com>
Signed-off-by: Olaf Hering <olaf@aepfle.de>

---
 ioemu-remote/xenstore.c |    3 +++
 1 file changed, 3 insertions(+)

--- ioemu-remote/xenstore.c
+++ ioemu-remote/xenstore.c
@@ -927,6 +927,9 @@ static void xenstore_process_dm_command_
         do_pci_add(par);
         free(par);
 #endif
+    } else if (!strncmp(command, "flush-cache", len)) {
+        fprintf(logfile, "dm-command: flush caches\n");
+        qemu_invalidate_map_cache();
     } else {
         fprintf(logfile, "dm-command: unknown
command\"%*s\"\n", len, command);
     }

On Thu, Jan 12, Olaf Hering wrote:
> Add support for a xenstore dm command to flush qemu''s buffer
cache.
Something like this is most likely also needed in mainline qemu.
My search in the qemu.org git web UI did not show any of the commands in
ioemu-remote, so its not clear how to forward port this change. 

Olaf

> --- ioemu-remote/xenstore.c
> +++ ioemu-remote/xenstore.c
> @@ -927,6 +927,9 @@ static void xenstore_process_dm_command_
>          do_pci_add(par);
>          free(par);
>  #endif
> +    } else if (!strncmp(command, "flush-cache", len)) {
> +        fprintf(logfile, "dm-command: flush caches\n");
> +        qemu_invalidate_map_cache();
>      } else {
>          fprintf(logfile, "dm-command: unknown
command\"%*s\"\n", len, command);
>      }

On Thu, 12 Jan 2012, Olaf Hering wrote:
> Add support for a xenstore dm command to flush qemu''s buffer
cache.
> 
> qemu will just keep mapping pages and not release them, which causes
problems
> for the memory pager (since the page is mapped, it won''t get paged
out). When
> the pager has trouble finding a page to page out, it asks qemu to flush its
> buffer, which releases all the page mappings. This makes it possible to
find
> pages to swap out agian.
> 
> Signed-off-by: Patrick Colp <Patrick.Colp@citrix.com>
> Signed-off-by: Olaf Hering <olaf@aepfle.de>
> 
> ---
>  ioemu-remote/xenstore.c |    3 +++
>  1 file changed, 3 insertions(+)
> 
> --- ioemu-remote/xenstore.c
> +++ ioemu-remote/xenstore.c
> @@ -927,6 +927,9 @@ static void xenstore_process_dm_command_
>          do_pci_add(par);
>          free(par);
>  #endif
> +    } else if (!strncmp(command, "flush-cache", len)) {
> +        fprintf(logfile, "dm-command: flush caches\n");
> +        qemu_invalidate_map_cache();
>      } else {
>          fprintf(logfile, "dm-command: unknown
command\"%*s\"\n", len, command);
>      }
I guess it is not possible to send the usual ioreq with type
=IOREQ_TYPE_INVALIDATE because the pager is separate from the hypervisor,
correct?

On Thu, 12 Jan 2012, Olaf Hering wrote:
> On Thu, Jan 12, Olaf Hering wrote:
> 
> > Add support for a xenstore dm command to flush qemu''s buffer
cache.
> 
> Something like this is most likely also needed in mainline qemu.
> My search in the qemu.org git web UI did not show any of the commands in
> ioemu-remote, so its not clear how to forward port this change. 
 
Yes, upstream qemu is going to need this too.
So far we haven''t needed any xenstore command in upstream qemu, because
we have been using qemu''s own rpc mechanism, called qmp.
There is qmp support in libxl already, so it shouldn''t be difficult to
use it, instead of sending the command through xenstore.

On Fri, Jan 13, Stefano Stabellini wrote:
> I guess it is not possible to send the usual ioreq with type =>
IOREQ_TYPE_INVALIDATE because the pager is separate from the hypervisor,
> correct?
Maybe send_invalidate_req() could be triggerd by the pager in some way?
Andres sent a patch to move from domctl to memop. Maybe the paging part
of it could be extended to issue another command which just calls
send_invalidate_req(). If thats possible, then I assume the upstream
qemu does not need changing as well.

Olaf