vgsp: USERD rejection memory + re-arm (kick tracking continuous at 142-153 FPS; r938) (SESSION 153)

1 files changed, 72 insertions, 0 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index c550843..7f3a4b2 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -624,6 +624,25 @@ struct svc_channel {
 	unsigned long gr_ram_userd_latch_at;	/* jiffies at the latch */
 	bool	gr_ram_userd_confirmed;	/* +1/+2 step observed */
 	/*
+	 * SESSION 153-CONT-5: rejected-latch memory. A speculative latch the
+	 * pump rejects (frozen or jumped PUT) used to exhaust the RAM scan
+	 * budget permanently: r936's kick channel latched a FROZEN lookalike
+	 * (put=7 get=1 on a driver struct page), the pump rejected it twice,
+	 * the 20-slice scan budget retired, and the channel went dark with no
+	 * re-latch path - kick retires (and their publish writes) stopped for
+	 * the rest of the session. Remember the rejected PA and its PUT: the
+	 * scan skips that exact (pa, put) pair silently, re-adopts the block
+	 * the moment its PUT moves (content change = new evidence), and each
+	 * rejection re-arms the scan budget so one stale window can never kill
+	 * the pipeline. The re-arm count is bounded (64: the r937 rate of
+	 * ~4 rejections/min makes that >15 min of continuous retire activity
+	 * on a client that only ever wrote one kick, while a truly dead
+	 * channel still stops sweeping in bounded time).
+	 */
+	u64	gr_ram_userd_reject_pa;		/* PA of the rejected latch */
+	u32	gr_ram_userd_reject_put;	/* its PUT at rejection time */
+	u8	gr_ram_userd_rejects;		/* rejections seen (bound) */
+	/*
 	 * SESSION 154 (M-GL1.2): content anchor for a client push whose
 	 * translation went stale (r913: a validated pin kept answering for
 	 * a reused page and the wrong-VAS walk answer was implausible while
@@ -18556,6 +18575,18 @@ static bool svc_gr_ram_userd_scan(struct nv_vgpu_device *vgpu,
 					continue;
 			}
 			/*
+			 * SESSION 153-CONT-5: a previously rejected block at
+			 * the same PUT is a frozen lookalike - do not spend a
+			 * (re-armed) scan budget re-latching it, or the budget
+			 * loops on the same stale window forever. A DIFFERENT
+			 * PUT is content change (the block moved): adopt it
+			 * again so a real control block that appears late is
+			 * still picked up.
+			 */
+			if ((gpa + o) == ch->gr_ram_userd_reject_pa &&
+			    hp == ch->gr_ram_userd_reject_put)
+				continue;
+			/*
 			 * SESSION 154 (M-GL1.2): a candidate that passed the
 			 * shape gate, the entry validation, the page-density
 			 * gate and the put cap is latched SPECULATIVELY. The
@@ -18607,6 +18638,12 @@ static bool svc_gr_ram_userd_scan(struct nv_vgpu_device *vgpu,
 				ch->gr_ram_userd_passes;
 			svc.chan[i].gr_ram_userd_logged =
 				ch->gr_ram_userd_logged;
+			svc.chan[i].gr_ram_userd_reject_pa =
+				ch->gr_ram_userd_reject_pa;
+			svc.chan[i].gr_ram_userd_reject_put =
+				ch->gr_ram_userd_reject_put;
+			svc.chan[i].gr_ram_userd_rejects =
+				ch->gr_ram_userd_rejects;
 			svc.chan[i].gr_ram_userd_watch_n =
 				ch->gr_ram_userd_watch_n;
 			for (wj = 0; wj < SVC_GR_USERD_WATCH; wj++) {
@@ -29523,6 +29560,24 @@ static void svc_pump_channels(struct nv_vgpu_device *vgpu)
 					ch->gr_ram_userd_latch_put,
 					put != ch->gr_ram_userd_latch_put ?
 						"" : " no advance in 15 s");
+				/*
+				 * SESSION 153-CONT-5: one stale window must not
+				 * kill the pipeline. Remember the rejected
+				 * (pa, put) so the scan never re-latches the same
+				 * frozen block, re-arm the scan budget (bounded)
+				 * so a REAL control block that appears later can
+				 * still be adopted, and only then drop the USERD.
+				 * r936: the rejection retired the 20-slice budget
+				 * and the kick channel never re-latched anything.
+				 */
+				ch->gr_ram_userd_reject_pa = ch->userd_pa;
+				ch->gr_ram_userd_reject_put =
+					ch->gr_ram_userd_latch_put;
+				if (ch->gr_ram_userd_rejects < 64) {
+					ch->gr_ram_userd_rejects++;
+					ch->gr_ram_userd_passes = 0;
+					ch->gr_ram_userd_logged = false;
+				}
 				ch->userd_ok = false;
 				ch->userd_pa = 0;
 				ch->userd_as = 0;
@@ -29544,6 +29599,16 @@ static void svc_pump_channels(struct nv_vgpu_device *vgpu)
 						svc.chan[li].userd_ok = false;
 						svc.chan[li].userd_pa = 0;
 						svc.chan[li].userd_as = 0;
+						svc.chan[li].gr_ram_userd_passes =
+							ch->gr_ram_userd_passes;
+						svc.chan[li].gr_ram_userd_logged =
+							ch->gr_ram_userd_logged;
+						svc.chan[li].gr_ram_userd_reject_pa =
+							ch->gr_ram_userd_reject_pa;
+						svc.chan[li].gr_ram_userd_reject_put =
+							ch->gr_ram_userd_reject_put;
+						svc.chan[li].gr_ram_userd_rejects =
+							ch->gr_ram_userd_rejects;
 						break;
 					}
 				spin_unlock_irqrestore(&svc.chan_lock, lflags);
@@ -32221,6 +32286,13 @@ static void svc_pump_channels(struct nv_vgpu_device *vgpu)
 					ch->gr_ram_userd_latch_at;
 				svc.chan[k].gr_ram_userd_confirmed =
 					ch->gr_ram_userd_confirmed;
+				/* SESSION 153-CONT-5: rejected-latch memory. */
+				svc.chan[k].gr_ram_userd_reject_pa =
+					ch->gr_ram_userd_reject_pa;
+				svc.chan[k].gr_ram_userd_reject_put =
+					ch->gr_ram_userd_reject_put;
+				svc.chan[k].gr_ram_userd_rejects =
+					ch->gr_ram_userd_rejects;
 				svc.chan[k].gr_anchor_cursor = ch->gr_anchor_cursor;
 				svc.chan[k].gr_anchor_va = ch->gr_anchor_va;
 				svc.chan[k].gr_anchor_sweeps = ch->gr_anchor_sweeps;