This is the second attempt to fix this issue, as a lesson from the last time it doesn't try to use an async monitor command. So with this patchset, in qxl mode, a screendump monitor command will complete before the file is written to disk. This is much better then a hang. To fix it does require asynchronous monitor command completion.
Changes from previous RFC (v4 to match my internal patches, it's actually the second): - dropped the DisplayAllocator - dropping flip (Gerd Hoffman suggestion) - dropping needless call to console_select (from Gerd) - use a bh for the bug fix itself, previously called dpy_update and wrote to vga->ds.surface->data from spice server thread, a no no. - last patch uses a bh for the ppm_save as well, same problem as one in the previous round. I'm going to work on a solution to that as discussed on the list (either a QAPI equivalent of the old async monitor or a new really async command with completion event that libvirt/autotest/etc would have to use), but didn't want to wait for that with the already prolonged wait from the last RFC. Please review, Alon Levy (8): sdl: remove NULL check, g_malloc0 can't fail qxl: drop qxl_spice_update_area_async definition qxl: screen_dump in vga: do a single ppm_save qxl: require spice >= 0.8.2 qxl: remove flipped qxl: introduce QXLCookie qxl: make qxl_render_update async qxl-render: call ppm_save on bh Gerd Hoffman (1): console: don't call console_select unnecessarily configure | 2 +- console.c | 6 +- hw/qxl-render.c | 228 ++++++++++++++++++++++++++++++++++++++-------------- hw/qxl.c | 189 ++++++++++++++++++++++++++++++------------- hw/qxl.h | 28 ++++--- ui/sdl.c | 4 - ui/spice-display.c | 34 +++++---- ui/spice-display.h | 22 +++++ 8 files changed, 359 insertions(+), 154 deletions(-) -- 1.7.9.1
