diff options
| author | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-09 15:45:21 -0400 |
|---|---|---|
| committer | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-09 15:45:21 -0400 |
| commit | 0f3258e8b72e51ea5d2f171590938628160a84b6 (patch) | |
| tree | 81402380ae3325e7302fafa6947e21f69a212dcc /kernel | |
| parent | 523c0839a21a798ca7f92395cde71198d0cf030c (diff) | |
| download | open-iov-0f3258e8b72e51ea5d2f171590938628160a84b6.tar.gz open-iov-0f3258e8b72e51ea5d2f171590938628160a84b6.zip | |
vgsp-sema: CE layout USERD accepts on PUT alone; GET is GPU-owned (ISSUE-0002)
r47 guest truth (channel_info walk in the hung second client): every CE
channel's GP_PUT/GP_GET GPU VAs are ring+0x2008c / ring+0x20088 and the live
rank0 USERD is the layout block at fb 0x90000 (the emulation's offsets were
right all along). The coherence gate required g <= p, but the GET word is
GPU-owned and holds leftovers (session 2 rank0: PUT=1, GET=136 from the
previous session) - so the live block was rejected forever (r44 strict never
accepted; r42/r46 latched the stale registered candidate). Accept the CE
layout block when the PUT word is plausible; the pump tracks ch->last_get
and rewrites GP_GET as it consumes entries.
Diffstat (limited to 'kernel')
| -rw-r--r-- | kernel/gsp_sema_service.c | 28 |
1 files changed, 24 insertions, 4 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index 43a8d62..1637d63 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -7789,10 +7789,30 @@ static int svc_chan_userd(struct nv_vgpu_device *vgpu, struct svc_channel *ch, va >= 0x120000000ULL && va < 0x121000000ULL && len >= 4 && len <= 0x10000; } - if (ring_ok) - rc = svc_userd_probe(vgpu, ring_pa + 0x10000, - ring_ap, ch->gpfifo_entries, - &p, &g); + /* + * ISSUE-0002 (r47): read only the PUT word for the CE layout + * block. The GET word is GPU-owned and its initial content is + * not coherent - session 2 rank0 read PUT=1 / GET=136 (a + * leftover from the previous session) and requiring g <= p + * rejected the LIVE layout block forever (r44 strict never + * accepted it; r42/r46 fell back to the stale registered + * candidate instead). The pump tracks its own position in + * ch->last_get and rewrites GP_GET as it consumes entries, so + * the initial GET value is irrelevant. + */ + if (ring_ok) { + if (!svc_read_ap(vgpu, + ring_pa + 0x10000 + SVC_USERD_GP_PUT, + ring_ap, &p, sizeof(p))) { + (void)svc_read_ap(vgpu, + ring_pa + 0x10000 + + SVC_USERD_GP_GET, + ring_ap, &g, sizeof(g)); + rc = p <= ch->gpfifo_entries ? 0 : -EINVAL; + } else { + rc = -ENOENT; + } + } /* * ISSUE-0002 (r46 diag): name the exact failing check on the * CE discovery decision. Three lines per rank, 36 per session. |