On Mon, Feb 24, 2025 at 12:38:22AM +0000, Michael Kelley wrote:
> From: Saurabh Singh Sengar <ssengar@xxxxxxxxxxxxxxxxxxx> Sent: Sunday, February 23, 2025 6:10 AM
> >
> > On Sat, Feb 22, 2025 at 08:16:53PM +0000, Michael Kelley wrote:
> > > From: Saurabh Singh Sengar <ssengar@xxxxxxxxxxxxxxxxxxx> Sent: Saturday, February 22, 2025 9:27 AM
> > > >
>
> [anip]
>
> > > >
> > > > I had considered moving the entire `hvfb_putmem()` function to `destroy`,
> > > > but I was hesitant for two reasons:
> > > >
> > > > 1. I wasn’t aware of any scenario where this would be useful. However,
> > > > your explanation has convinced me that it is necessary.
> > > > 2. `hvfb_release_phymem()` relies on the `hdev` pointer, which requires
> > > > multiple `container_of` operations to derive it from the `info` pointer.
> > > > I was unsure if the complexity was justified, but it seems worthwhile now.
> > > >
> > > > I will move `hvfb_putmem()` to the `destroy` function in V2, and I hope this
> > > > will address all the cases you mentioned.
> > > >
> > >
> > > Yes, that's what I expect needs to happen, though I haven't looked at the
> > > details of making sure all the needed data structures are still around. Like
> > > you, I just had this sense that hvfb_putmem() might need to be moved as
> > > well, so I tried to produce a failure scenario to prove it, which turned out
> > > to be easy.
> > >
> > > Michael
> >
> > I will add this in V2 as well. But I have found an another issue which is
> > not very frequent.
> >
> >
> > [ 176.562153] ------------[ cut here ]------------
> > [ 176.562159] fb0: fb_WARN_ON_ONCE(pageref->page != page)
> > [ 176.562176] WARNING: CPU: 50 PID: 1522 at drivers/video/fbdev/core/fb_defio.c:67
> > fb_deferred_io_mkwrite+0x215/0x280
> >
> > <snip>
> >
> > [ 176.562258] Call Trace:
> > [ 176.562260] <TASK>
> > [ 176.562263] ? show_regs+0x6c/0x80
> > [ 176.562269] ? __warn+0x8d/0x150
> > [ 176.562273] ? fb_deferred_io_mkwrite+0x215/0x280
> > [ 176.562275] ? report_bug+0x182/0x1b0
> > [ 176.562280] ? handle_bug+0x133/0x1a0
> > [ 176.562283] ? exc_invalid_op+0x18/0x80
> > [ 176.562284] ? asm_exc_invalid_op+0x1b/0x20
> > [ 176.562289] ? fb_deferred_io_mkwrite+0x215/0x280
> > [ 176.562291] ? fb_deferred_io_mkwrite+0x215/0x280
> > [ 176.562293] do_page_mkwrite+0x4d/0xb0
> > [ 176.562296] do_wp_page+0xe8/0xd50
> > [ 176.562300] ? ___pte_offset_map+0x1c/0x1b0
> > [ 176.562304] __handle_mm_fault+0xbe1/0x10e0
> > [ 176.562307] handle_mm_fault+0x17f/0x2e0
> > [ 176.562309] do_user_addr_fault+0x2d1/0x8d0
> > [ 176.562314] exc_page_fault+0x85/0x1e0
> > [ 176.562318] asm_exc_page_fault+0x27/0x30
> >
> > Looks this is because driver is unbind still Xorg is trying to write
> > to memory which is causing some page faults. I have confirmed PID 1522
> > is of Xorg. I think this is because we need to cancel the framebuffer
> > deferred work after flushing it.
>
> Does this new issue occur even after moving hvfb_putmem()
> into the destroy() function?
Unfortunately yes :(
> I'm hoping it doesn't. I've
> looked at the fb_deferred_io code, and can't quite figure out
> how that deferred I/O work is supposed to get cancelled. Or
> maybe it's just not supposed to get started again after the flush.
>
I want to understand why cancel_delayed_work_sync was introduce in
hvfb_suspend and not the flush. Following commit introduced it.
382a462217572 ('video: hyperv_fb: Fix hibernation for the deferred IO feature')
But I agree this need more analysis.
> If the new issue still happens, that seems like more of a flaw
> in the fb deferred I/O mechanism not shutting itself down
> properly.
>
As the repro rate is quite low, this will take some effort to get this
fixed. Shall we take this in a separate patch later ?
> Michael
>
<snip>
