diff options
| author | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-09 22:11:03 -0400 |
|---|---|---|
| committer | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-09 22:11:03 -0400 |
| commit | 3d4777f614e31f3c3a4f2ce49b37054a10fd1fd3 (patch) | |
| tree | 1f465e13d15a47cd062dc558d5477f5db0f89936 | |
| parent | 4c1266e620423a0870c68e2228dbdcf6b2597d29 (diff) | |
| download | open-iov-3d4777f614e31f3c3a4f2ce49b37054a10fd1fd3.tar.gz open-iov-3d4777f614e31f3c3a4f2ce49b37054a10fd1fd3.zip | |
vgsp-sema: rewind last_get on a fresh channel's second put drop (ISSUE-0002)
r73 c3 rank2: discovery read the previous session's residual put (67); the
pump consumed the live entries 0..1, the r48 resync handled the first drop,
but the guest's counter then settled at put=1 while last_get=2 and
userd_wait_fresh was already consumed. The loop condition last_get != put
then marched ~1023 slots the producer had not written, holding each stale
entry for the full CE grace (one fake release per 8 s) - the client looked
wedged and the exec timed out.
When a drop leaves the pump ahead of the producer on a channel with fewer
than half a ring of submissions, rewind last_get to put instead of marching.
A genuine ring wrap needs a full ring of submissions, so this cannot skip a
live entry.
| -rw-r--r-- | kernel/gsp_sema_service.c | 27 |
1 files changed, 27 insertions, 0 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index fa76162..e972732 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -13130,6 +13130,33 @@ static void svc_pump_channels(struct nv_vgpu_device *vgpu) if (ch->userd_wait_fresh) { ch->last_get = 0; ch->userd_wait_fresh = false; + } else if (ch->engine_type && ch->gpfifo_entries && + ch->last_get > put && + ch->submissions < ch->gpfifo_entries / 2) { + /* + * ISSUE-0002 (r73): a SECOND (or later) drop + * while the channel is still fresh. The + * discovery-time put was residual (the previous + * session's counter, e.g. 67) and the pump had + * already consumed the live entries 0..k; the + * guest's counter then settled at a value below + * the pump's read position (r73 rank2: put=1 + * with last_get=2 after the r48 resync). The + * loop condition last_get != put then marches + * forward through ~1023 slots the producer has + * not written in this epoch, holding each + * unreadable stale entry for the full CE grace + * (one fake release per 8 s) - the client + * appears wedged. Never stay ahead of the + * producer: rewind to its position. A genuine + * ring wrap cannot be fresh (it needs a full + * ring of submissions), so this cannot skip a + * live entry. + */ + pr_info_ratelimited("sema-chan: fresh put drop h=0x%x rank=%u put=%u last=%u -> rewind\n", + ch->h_object, ch->ce_index, put, + (u32)ch->last_get); + ch->last_get = put; } bitmap_zero(ch->slot_done_epoch, SVC_GPFIFO_MAX_BITS); } |
