BALATON Zoltan <bala...@eik.bme.hu> writes: > On Tue, 14 Mar 2017, Mark Cave-Ayland wrote: >> On 14/03/17 10:00, Alex Bennée wrote: >>> Mark Cave-Ayland <mark.cave-ayl...@ilande.co.uk> writes: >>> >>>> I've recently noticed some video artifacts appearing in the form of >>>> horizontal lines whilst testing OpenBIOS boot on some qemu-system-ppc >>>> images (see https://www.ilande.co.uk/tmp/qemu/macos9-stripe.png for an >>>> example) which I've finally bisected down to this commit. >>> >>> Interesting. Is the video hardware in this case a simple framebuffer or >>> is there some sort of GPU involved? >> >> For the mac99 machine it's just the normal QEMU PCI std-vga card, so >> nothing special. Having run tests across SPARC which uses its own >> framebuffer, there have been no obvious artifacts that I've spotted >> there to date. >> >>>> The horizontal lines tend to appear at different positions with >>>> different lengths, although it seems to be more common that a complete >>>> scanline is affected. Reproducing the issue is fairly easy with a MacOS >>>> 9.2.1 ISO and the command line below: >>>> >>>> ./qemu-system-ppc -cdrom MacOS921.iso -boot d -m 512 -M mac99 >>>> >>>> Could it be that this patch is still missing a lock somewhere which >>>> causes these artifacts to appear? >>> >>> It depends. If the hardware is accessed via MMIO then the BQL is taken >>> automatically (unless the area is explicitly marked as doing its own >>> locking see: mr->global_locking). >>> >>> Is the artefact temporary or permanent? Could it be a change in >>> synchronisation between the emulator updating the framebuffer and >>> whatever backend updating its display? >> >> AFAICT they are temporary in that they appear randomly on the display, >> but the next time that particular area of the display is updated by the >> guest then they tend to either change/disappear. > > I've also noticed artifacts when testing the SM501 changes with > MorphOS on recent QEMU so it may not be related to just the std vga. > It's shown as bands (larger number of adjacent scanlines) not updating > when the client first wrote to the framebuffer but then they > disappeared during the next refresh. I'm guessing it may be related to > dirty checking of the framebuffer memory which is used to decide when > a scan line needs update?
Interesting. I guess if the corrupted scan lines are in blocks of PAGE_SIZE that might make sense. Commit (b0706b716769494f321a0d2bfd9fa9893992f995) made tlb_reset_dirty_range update the SoftMMU addr_write entry atomically. Now I don't see how that could race in single threaded TCG mode but it could explain things. I notice that tlb_set_dirty/tlb_set_dirty1 should maybe do the same. The currently assume they are only called from the CPU's context. If you enable #define DEBUG_TLB in cputlb.c does the assert fire? > > Regards, > BALATON Zoltan -- Alex Bennée
