sema-svc: destination-keyed copy watch (dump page-table writes)

sema_copy_dst=dumps the source bytes of every copy whose destination falls
in the given page, so the 32-byte page-directory updates the guest writes
into vidmem can be read directly and their entry encoding learned.

3 files changed, 45 insertions, 0 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index db52299..018e28c 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -345,6 +345,7 @@ struct sema_svc {
 	u64			lda_log;		/* LAUNCH_DMA inventory budget */
 	u64			copy_watch_log;		/* sema_copy_watch log budget */
 	u64			copy_pat_log;		/* sema_copy_pat log budget */
+	u64			copy_dst_log;		/* sema_copy_dst log budget */
 	u64			push_pat_log;		/* push-content pattern budget */
 	u64			dma_skip_log;		/* unrecognized-DMA budget */
 	u64			grce_log;		/* GRCE inline-write budget */
@@ -2098,6 +2099,32 @@ static int svc_copy(struct nv_vgpu_device *vgpu, u64 dst, u64 src, size_t total,
 		}
 	}
 
+	/*
+	 * Destination-keyed watch: dump the source bytes of every copy whose
+	 * destination lies in gsp_sema_copy_dst's page. This reads the
+	 * page-table entries the guest writes into vidmem directly.
+	 */
+	if (gsp_sema_copy_dst &&
+	    (dst & ~0xfffULL) == (gsp_sema_copy_dst & ~0xfffULL) &&
+	    svc.copy_dst_log < 32) {
+		u8 peek[32];
+		u32 n = total < sizeof(peek) ? (u32)total : sizeof(peek);
+
+		memset(peek, 0, sizeof(peek));
+		if (src_aperture == CE_APERTURE_VID) {
+			u8 *base = vgpu->bars[1].vaddr;
+
+			if (base && src + n <= vgpu->bars[1].size)
+				memcpy(peek, base + src, n);
+		} else {
+			svc_read(vgpu, src, peek, n);
+		}
+		svc.copy_dst_log++;
+		pr_info("sema-copy-dst: dst=0x%llx off=0x%llx total=%zu src=0x%llx sap=%u data=%*phN h=0x%x\n",
+			dst, dst & 0xfffULL, total, src, src_aperture,
+			n, peek, svc.cur_ch);
+	}
+
 	for (o = 0; o < total; ) {
 		size_t chunk = total - o;
 		u32 i;
@@ -3944,6 +3971,13 @@ unsigned gsp_sema_copy_watch;
 u64 gsp_sema_copy_pat;
 
 /*
+ * Data-path watch: log copies whose destination falls in this address's
+ * page, dumping the source bytes. Used to read the page-table entries the
+ * guest writes to vidmem (32-byte PD updates) and learn their encoding.
+ */
+u64 gsp_sema_copy_dst;
+
+/*
  * Grace window (ms) an untranslatable CE entry stays pending before the
  * fallback acknowledges its release and consumes it. The push's work (data
  * copies the client will read back) must run before the release the client
diff --git a/kernel/gsp_sema_service.h b/kernel/gsp_sema_service.h
index 234ec2f..51397b9 100644
--- a/kernel/gsp_sema_service.h
+++ b/kernel/gsp_sema_service.h
@@ -100,6 +100,9 @@ extern unsigned gsp_sema_copy_watch;
 /* Data-path watch: 8-byte value matched in copy src/dst and push bytes. */
 extern u64 gsp_sema_copy_pat;
 
+/* Data-path watch: dump copies landing in this page (source bytes). */
+extern u64 gsp_sema_copy_dst;
+
 /* CE pending-entry grace window in ms before the release fallback fires. */
 extern unsigned gsp_sema_ce_grace_ms;
 
diff --git a/kernel/main.c b/kernel/main.c
index 7fe2d15..d6e6c9d 100644
--- a/kernel/main.c
+++ b/kernel/main.c
@@ -2476,6 +2476,14 @@ module_param_named(sema_copy_pat, gsp_sema_copy_pat, ullong, 0644);
 MODULE_PARM_DESC(sema_copy_pat, "8-byte pattern matched in copy src/dst and pushes (default 0)");
 
 /*
+ * Data-path watch: dump the source bytes of copies whose destination falls
+ * in this address's page (up to 32 samples):
+ *   echo 0x200000 > .../parameters/sema_copy_dst
+ */
+module_param_named(sema_copy_dst, gsp_sema_copy_dst, ullong, 0644);
+MODULE_PARM_DESC(sema_copy_dst, "Dump source bytes of copies landing in this page (default 0)");
+
+/*
  * Grace window (ms) an untranslatable CE entry stays pending before the
  * fallback acknowledges its release and consumes it. Its work must run before
  * the release the client synchronizes on; the translation usually lands within