[LibOS] Allow madvise(MADV_DONTNEED) on non-writable mappings

Previously, it was not implemented though there is no blocker for this
(only need to temporarily change permissions on the mapping to be
writable). This fixes Node.js v20 workloads.

Signed-off-by: Dmitrii Kuvaiskii <dmitrii.kuvaiskii@intel.com>

Documentation/devel/features.md

@@ -1728,7 +1728,6 @@ applications that would use these flags. In case of SGX backend, `mprotect()` be
 `madvise()` implements only a minimal subset of functionality:
 - `MADV_DONTNEED` is partially supported:
   - resetting writable file-backed mappings is not implemented;
-  - zeroing non-writable mappings is not implemented;
   - all other cases are implemented.
 - `MADV_NORMAL`, `MADV_RANDOM`, `MADV_SEQUENTIAL`, `MADV_WILLNEED`, `MADV_FREE`,
   `MADV_SOFT_OFFLINE`, `MADV_MERGEABLE`, `MADV_UNMERGEABLE`, `MADV_HUGEPAGE`, `MADV_NOHUGEPAGE` are

libos/src/bookkeep/libos_vma.c

@@ -1326,14 +1326,32 @@ static bool madvise_dontneed_visitor(struct libos_vma* vma, void* visitor_arg) {
         return true;
     }
 
-    if (!(vma->prot & PROT_WRITE)) {
-        ctx->error = -ENOSYS; // Zeroing non-writable mappings is not yet implemented.
-        return false;
-    }
-
     uintptr_t zero_start = MAX(ctx->begin, vma->begin);
     uintptr_t zero_end = MIN(ctx->end, vma->end);
+
+    pal_prot_flags_t pal_prot = LINUX_PROT_TO_PAL(vma->prot, vma->flags);
+    pal_prot_flags_t pal_prot_writable = pal_prot | PAL_PROT_WRITE;
+
+    if (pal_prot != pal_prot_writable) {
+        /* make the area writable so that it can be memset-to-zero */
+        int ret = PalVirtualMemoryProtect((void*)zero_start, zero_end - zero_start,
+                                          pal_prot_writable);
+        if (ret < 0) {
+            ctx->error = pal_to_unix_errno(ret);
+            return false;
+        }
+    }
+
     memset((void*)zero_start, 0, zero_end - zero_start);
+
+    if (pal_prot != pal_prot_writable) {
+        /* the area was made writable above; restore the original permissions */
+        int ret = PalVirtualMemoryProtect((void*)zero_start, zero_end - zero_start, pal_prot);
+        if (ret < 0) {
+            log_error("restoring original permissions failed: %s", pal_strerror(ret));
+            BUG();
+        }
+    }
     return true;
 }
 

libos/test/regression/madvise.c

@@ -15,7 +15,7 @@
 
 #define PAGES_CNT 128
 
-int main(void) {
+static void test_madvise_rw(void) {
     size_t page_size = getpagesize();
 
     char* m = (char*)mmap(NULL, PAGES_CNT * page_size,
@@ -53,6 +53,32 @@ int main(void) {
             }
         }
     }
+
+    if (munmap(m, PAGES_CNT * page_size) < 0)
+        err(1, "munmap");
+}
+
+/* this test emulates Node.js v20 behavior */
+static void test_madvise_none(void) {
+    size_t page_size = getpagesize();
+
+    char* m = (char*)mmap(NULL, PAGES_CNT * page_size,
+                          PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS,
+                          -1, 0);
+    if (m == MAP_FAILED)
+        err(1, "mmap()");
+
+    int res = madvise(m, PAGES_CNT * page_size, MADV_DONTNEED);
+    if (res)
+        err(1, "madvise(%p, 0x%zx, MADV_DONTNEED) failed", m, PAGES_CNT * page_size);
+
+    if (munmap(m, PAGES_CNT * page_size) < 0)
+        err(1, "munmap");
+}
+
+int main(void) {
+    test_madvise_rw();
+    test_madvise_none();
     puts("TEST OK");
     return 0;
 }