vgsp-sema: never seed last_get past a discovery put >= 2 (ISSUE-0002)

A fresh CE channel can legitimately reach the service's first pass with a
live put far above 1: the driver publishes the put once per submitted push
((cpu_put + 1) % entries) and clients stream pushes while the service is
still discovering the channel. r56 measured healthy sessions at put=66/134
within seconds; r54's third-session rank2 discovery read p=67, g=67 (a
LIVE put), the r48 seeding set last_get=67, and the client's first push
(entry 0) was never decoded - the channel_pool_add spin. Always start at
slot 0; a genuinely stale block is still handled by the existing put-drop
resync. Also make the CE put-snapshot diagnostic per-rank (the channel
struct iterated is a local copy, so the old per-channel stamp let the
first rank spend the whole session budget) and raise its cap for a
third-session rank2 to keep logging through a stall.

1 files changed, 36 insertions, 24 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 5fd583a..96c9064 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -4717,6 +4717,8 @@ static u64 svc_rel_va_seen;
 static u32 svc_rel_count;
 static u32 svc_dbg_streams;
 static u32 svc_chan_put_log;
+/* ISSUE-0002 (r55): per-rank put-snapshot throttle (ch is a local copy). */
+static unsigned long svc_put_log_at[32];
 /*
  * ISSUE-0002 (r46 diag): bounded per-session log for the CE USERD discovery
  * decision (svc_chan_userd). Budgets reset in svc_session_flush; 3 attempts
@@ -6767,6 +6769,7 @@ static void svc_session_flush(void)
 	svc.gpfifo_entry_log = 0;
 	svc_dbg_streams = 0;
 	svc_chan_put_log = 0;
+	memset(svc_put_log_at, 0, sizeof(svc_put_log_at));
 	svc_userd_dbg = 0;
 	memset(svc_userd_dbg_ch, 0, sizeof(svc_userd_dbg_ch));
 	svc_userd_sweep_log = 0;
@@ -7866,21 +7869,25 @@ static int svc_chan_userd(struct nv_vgpu_device *vgpu, struct svc_channel *ch,
 			ch->userd_as = ring_ap;
 			ch->userd_ok = true;
 			/*
-			 * ISSUE-0002 (r48): a PUT >= 2 at discovery can only be
-			 * the previous session's residual state - the client
-			 * always publishes PUT=1 for a fresh channel's first
-			 * push and waits for its release before a second. Seed
-			 * last_get past the residual entries so the pump does
-			 * not consume stale slots and desync from the guest's
-			 * counter reset (r48: rank0 accepted PUT=134, the pump
-			 * ran 34 stale entries, then the guest reset PUT=1 and
-			 * the pump wrapped; the live first push was never
-			 * processed). The pump detects the reset as a PUT drop
-			 * and resyncs last_get to slot 0.
+			 * ISSUE-0002 (r55): a PUT >= 2 at discovery is NOT
+			 * necessarily the previous session's residual state.
+			 * UVM publishes the put once per submitted push
+			 * (host_hal->write_gpu_put stores
+			 * (cpu_put + 1) % entries), and a client that streams
+			 * pushes while the service is still discovering the
+			 * channel leaves a LIVE put far above 1 (r56: a healthy
+			 * session's CE puts reached 66/134 within seconds; r54's
+			 * third-session rank2 discovery read p=67, g=67). The
+			 * r48 seeding (last_get = p) then skipped those LIVE
+			 * entries and the client's first push (entry 0) was
+			 * never decoded - the channel_pool_add stall. Always
+			 * start at slot 0; a genuinely stale block is still
+			 * handled: the fresh channel's first publish drops the
+			 * put below the stale put_prev and the pump's drop
+			 * resync rewinds last_get to slot 0.
 			 */
-			ch->userd_wait_fresh = p >= 2;
-			if (ch->userd_wait_fresh)
-				ch->last_get = p;
+			ch->userd_wait_fresh = p >= 1;
+			ch->last_get = 0;
 			*put = p;
 			*get = g;
 			if (ch->ce_index < 32)
@@ -12521,18 +12528,23 @@ static void svc_pump_channels(struct nv_vgpu_device *vgpu)
 		}
 
 		/*
-		 * ISSUE-0002 diag: per-second put snapshots for CE channels,
-		 * bounded per session. Names whether a re-registered session's
-		 * GP_PUT advances at all (the pump's view of the second
-		 * client's submissions).
+		 * ISSUE-0002 diag: put snapshots for CE channels, bounded per
+		 * session. The throttle is indexed by rank (the channel struct
+		 * iterated here is a local copy, so a per-channel stamp never
+		 * persisted and the first channel's burst spent the whole
+		 * budget - r54 had no rank2 line at all). Cap raised to 160 so
+		 * a third-session rank2 keeps logging through its stall.
 		 */
-		if (gsp_sema_dbg && ch->engine_type && svc_chan_put_log < 48 &&
-		    time_after(jiffies, ch->put_log_at)) {
-			ch->put_log_at = jiffies + HZ;
+		if (gsp_sema_dbg && ch->engine_type && ch->ce_index < 32 &&
+		    svc_chan_put_log < 160 &&
+		    time_after(jiffies, svc_put_log_at[ch->ce_index])) {
+			svc_put_log_at[ch->ce_index] = jiffies + HZ;
 			svc_chan_put_log++;
-			pr_info("sema-put: h=0x%x uok=%d upa=0x%llx uas=%u put=%u get=%u last=%u\n",
-				ch->h_object, ch->userd_ok, ch->userd_pa,
-				ch->userd_as, put, get, (u32)ch->last_get);
+			pr_info("sema-put: h=0x%x rank=%u uok=%d upa=0x%llx uas=%u put=%u get=%u last=%u prev=%u fresh=%d\n",
+				ch->h_object, ch->ce_index, ch->userd_ok,
+				ch->userd_pa, ch->userd_as, put, get,
+				(u32)ch->last_get, ch->put_prev,
+				ch->userd_wait_fresh ? 1 : 0);
 		}
 
 		/*