From 5b89dabc048b0bc888f8227658c4a21fc31f13dd Mon Sep 17 00:00:00 2001
From: Gregory Lirent <opensource@lirent.ru>
Date: Mon, 22 Jun 2026 15:17:37 +0300
Subject: [PATCH] fix: pin the vgpu region resident (raise working set, then
 VirtualLock)

The fallback path committed the region but left it demand-zero (not resident). VirtualLock can pin only up to the process minimum working set, and the default quota is far below the 96 MiB region, so a bare lock fails with ERROR_WORKING_SET_QUOTA. Raise the minimum first via SetProcessWorkingSetSize, then VirtualLock; pre-fault every page as a last resort. Neither needs SE_LOCK_MEMORY (that is for large pages / AWE).
---
 src/stream/win32/region.c | 34 ++++++++++++++++++++++++++++++----
 1 file changed, 30 insertions(+), 4 deletions(-)

diff --git a/src/stream/win32/region.c b/src/stream/win32/region.c
index 3930f20..1181402 100644
--- a/src/stream/win32/region.c
+++ b/src/stream/win32/region.c
@@ -106,14 +106,40 @@ int vgpu_region_create(vgpu_region_t* out) {
         }
         uintptr_t addr    = (uintptr_t)p;
         uintptr_t aligned = (addr + VGPU_2MB - 1) & ~(uintptr_t)(VGPU_2MB - 1);
-        if (!VirtualLock((void*)aligned, (SIZE_T)bytes)) {
-            fprintf(stderr, "region: VirtualLock failed (%lu) — pages may not be pinned\n",
-                    GetLastError());
+
+        /* The region must be RESIDENT, not merely committed: the host reads it out
+         * of guest RAM and only PRESENT pages are visible to it — a committed but
+         * demand-zero page has no PTE, so it is unreadable from the host. VirtualLock
+         * pins the pages into the working set, but it can lock at most the process
+         * MINIMUM working set, and the default quota is far below the region size
+         * (so a bare VirtualLock fails with ERROR_WORKING_SET_QUOTA). Raise the
+         * minimum first. NB: VirtualLock / SetProcessWorkingSetSize do NOT need
+         * SE_LOCK_MEMORY — that privilege is only for large pages / AWE. */
+        SIZE_T ws_min = (SIZE_T)(bytes + 64ull * 1024 * 1024);   /* region + headroom */
+        SIZE_T ws_max = ws_min + 128ull * 1024 * 1024;
+        SIZE_T cur_min = 0, cur_max = 0;
+        if (GetProcessWorkingSetSize(GetCurrentProcess(), &cur_min, &cur_max)) {
+            if (cur_min > ws_min) ws_min = cur_min;   /* never shrink an existing quota */
+            if (cur_max > ws_max) ws_max = cur_max;
         }
+        if (!SetProcessWorkingSetSize(GetCurrentProcess(), ws_min, ws_max))
+            fprintf(stderr, "region: SetProcessWorkingSetSize(%llu MiB) failed (%lu)\n",
+                    (unsigned long long)(ws_min / (1024 * 1024)), GetLastError());
+
+        if (!VirtualLock((void*)aligned, (SIZE_T)bytes)) {
+            fprintf(stderr, "region: VirtualLock failed (%lu) — pre-faulting region\n",
+                    GetLastError());
+            /* Last resort: fault every page so it is at least PRESENT now. Without
+             * the lock the trimmer may evict it under pressure, but the raised
+             * minimum working set above makes eviction far less likely. */
+            volatile uint8_t* q = (volatile uint8_t*)aligned;
+            for (uint64_t off = 0; off < bytes; off += 4096u) q[off] = q[off];
+        }
+
         os_base  = p;
         base     = (uint8_t*)aligned;
         os_total = total;
-        fprintf(stderr, "region: fallback VirtualAlloc+VirtualLock %llu MiB, aligned at %p\n",
+        fprintf(stderr, "region: fallback VirtualAlloc+lock %llu MiB, aligned at %p\n",
                 (unsigned long long)(bytes / (1024 * 1024)), (void*)aligned);
     }