Route sys_statfs through the /proc intercept

src/runtime/procemu.c

@@ -2801,7 +2801,7 @@ int proc_intercept_open(const guest_t *g,
      * "self" symlink. The DIR* created from this allows getdents64 to enumerate
      * /proc like a real procfs. Cleaned up via atexit.
      */
-    if (!strcmp(path, "/proc")) {
+    if (!strcmp(path, "/proc") || !strcmp(path, "/proc/")) {
         const char *dir = ensure_proc_tmpdir(g);
         if (!dir)
             return -1;
@@ -3436,6 +3436,39 @@ int proc_intercept_open(const guest_t *g,
     return PROC_NOT_INTERCEPTED;
 }
 
+int proc_intercept_statfs(const char *path, struct statfs *out)
+{
+    /* /proc/<our_pid>[/...] -> /proc/self[...], same alias as
+     * proc_intercept_open/proc_intercept_stat.
+     */
+    char alias[LINUX_PATH_MAX];
+    int aliased = proc_alias_self(path, alias, sizeof(alias));
+    if (aliased < 0)
+        return -1;
+    if (aliased > 0)
+        return proc_intercept_statfs(alias, out);
+
+    /* /proc/self/fd and /proc/self/fdinfo are the only /proc nodes whose open
+     * path allocates per-call scratch state: proc_open_fd_scratch mkdtemp's a
+     * fresh directory and creates one placeholder file per live guest fd on
+     * every single open, purely so getdents has something to enumerate. A
+     * plain statfs() never enumerates the directory, so paying for that
+     * allocation just to fstatfs() the result and immediately discard it is
+     * wasted work -- and once PROC_SCRATCH_DIRS_MAX untracked opens
+     * accumulate, the atexit cleanup no longer knows about further dirs and
+     * they leak in /tmp permanently.
+     *
+     * Every scratch dir (like every other synthetic /proc node) is created
+     * under /tmp, so statfs("/tmp") reports identical filesystem info without
+     * allocating anything.
+     */
+    if (strcmp(path, "/proc/self/fd") && strcmp(path, "/proc/self/fd/") &&
+        strcmp(path, "/proc/self/fdinfo") && strcmp(path, "/proc/self/fdinfo/"))
+        return PROC_NOT_INTERCEPTED;
+
+    return statfs("/tmp", out) < 0 ? -1 : 0;
+}
+
 int proc_intercept_stat(const char *path, struct stat *st)
 {
     /* Intercept stat for /proc paths emulated via proc_intercept_open. Without

src/runtime/procemu.h

@@ -13,6 +13,7 @@
 
 #include <stddef.h>
 #include <stdint.h>
+#include <sys/mount.h>
 #include <sys/stat.h>
 #include <sys/uio.h>
 #include "core/guest.h"
@@ -44,6 +45,16 @@ int proc_intercept_readlink(const char *path, char *buf, size_t bufsiz);
  */
 int proc_intercept_stat(const char *path, struct stat *mac_st);
 
+/* Intercept statfs for /proc paths that would otherwise route through
+ * proc_intercept_open purely to obtain filesystem-level info, forcing
+ * allocation of scratch state (e.g. /proc/self/fd's per-open directory of
+ * placeholder files, one per live guest fd) that statfs() has no use for.
+ * Returns 0 if the filesystem info was synthesized (mac_st filled), or -2
+ * (PROC_NOT_INTERCEPTED) if the path is not intercepted (fall through to the
+ * normal open/fstatfs/close path).
+ */
+int proc_intercept_statfs(const char *path, struct statfs *mac_st);
+
 /* Intercept writes to synthetic proc files that need stateful behavior.
  * Returns 1 if handled (with *written_out set), 0 if not intercepted, or -1 on
  * error with errno set.

src/syscall/fs-stat.c

@@ -21,6 +21,7 @@
 #include "syscall/internal.h"
 #include "syscall/path.h"
 #include "syscall/proc.h"
+#include "utils.h"
 
 static uint64_t mac_to_linux_dev(dev_t dev)
 {
@@ -355,6 +356,17 @@ int64_t sys_newfstatat(guest_t *g,
     return 0;
 }
 
+/* True when path (already translated/normalized by path_translate_at) names
+ * /proc itself or something under it. Boundary-checked so "/procfoo" does not
+ * false-positive the way a bare strncmp(path, "/proc", 5) would.
+ */
+static bool statfs_path_is_proc(const char *path)
+{
+    if (!path || path[0] != '/')
+        return false;
+    return !strncmp(path, "/proc", 5) && (path[5] == '\0' || path[5] == '/');
+}
+
 int64_t sys_statfs(guest_t *g, uint64_t path_gva, uint64_t buf_gva)
 {
     char path[LINUX_PATH_MAX];
@@ -368,8 +380,36 @@ int64_t sys_statfs(guest_t *g, uint64_t path_gva, uint64_t buf_gva)
         return -LINUX_ENOSYS;
 
     struct statfs mac_st;
-    if (statfs(tx.host_path, &mac_st) < 0)
+
+    /* /proc has no host-filesystem counterpart under the sysroot: every entry
+     * is synthesized on open (see proc_intercept_open), so a raw statfs() on
+     * the guest path always misses with ENOENT even for "/proc" itself. Try
+     * the lightweight statfs-specific resolver first (it shortcuts nodes whose
+     * open path would otherwise allocate scratch state statfs never uses),
+     * then fall back to the same intercept open/fstatfs/close that
+     * sys_newfstatat uses via proc_intercept_stat.
+     */
+    if (statfs_path_is_proc(tx.intercept_path)) {
+        int intercepted = proc_intercept_statfs(tx.intercept_path, &mac_st);
+        if (intercepted == PROC_NOT_INTERCEPTED) {
+            int host_fd = proc_intercept_open(g, tx.intercept_path, 0, 0);
+            if (host_fd == PROC_NOT_INTERCEPTED) {
+                if (statfs(tx.host_path, &mac_st) < 0)
+                    return linux_errno();
+            } else if (host_fd < 0) {
+                return linux_errno();
+            } else {
+                int rc = fstatfs(host_fd, &mac_st);
+                close_keep_errno(host_fd);
+                if (rc < 0)
+                    return linux_errno();
+            }
+        } else if (intercepted < 0) {
+            return linux_errno();
+        }
+    } else if (statfs(tx.host_path, &mac_st) < 0) {
         return linux_errno();
+    }
 
     linux_statfs_t lin_st;
     translate_statfs(&mac_st, &lin_st);

tests/test-proc.c

@@ -7,10 +7,11 @@
  *
  * Tests: /proc/self/cmdline, /proc/meminfo, /proc/stat, /proc/version,
  *        /proc/filesystems, /proc/mounts, readlink(/proc/self/exe),
+ *        statfs(/proc, /proc/self/cmdline, /proc/self/fd, /proc/self/fdinfo),
  *        /dev/null, /dev/zero, /dev/urandom
  *
  * Syscalls exercised: openat(56), read(63), write(64), readlinkat(78),
- *                     close(57)
+ *                     close(57), statfs(43), fstatfs(44)
  */
 
 #include <stdlib.h>
@@ -21,6 +22,7 @@
 #include <dirent.h>
 #include <errno.h>
 #include <sys/stat.h>
+#include <sys/statfs.h>
 
 #include "test-harness.h"
 #include "test-util.h"
@@ -199,6 +201,85 @@ int main(void)
             FAIL("opendir failed");
     }
 
+    /* statfs("/proc") and friends: /proc has no host-filesystem counterpart
+     * under the sysroot, so this exercises the intercept path (issue #141)
+     * rather than a raw host statfs() on the guest-supplied string.
+     */
+    TEST("statfs /proc");
+    {
+        struct statfs st;
+        if (statfs("/proc", &st) < 0)
+            FAIL("statfs failed");
+        else
+            PASS();
+    }
+
+    TEST("statfs /proc/ (trailing slash)");
+    {
+        struct statfs st;
+        if (statfs("/proc/", &st) < 0)
+            FAIL("statfs failed");
+        else
+            PASS();
+    }
+
+    TEST("statfs /proc/self/cmdline");
+    {
+        struct statfs st;
+        if (statfs("/proc/self/cmdline", &st) < 0)
+            FAIL("statfs failed");
+        else
+            PASS();
+    }
+
+    TEST("statfs matches fstatfs for /proc/self/cmdline");
+    {
+        struct statfs path_st, fd_st;
+        int fd = open("/proc/self/cmdline", O_RDONLY);
+        if (fd < 0) {
+            FAIL("open failed");
+        } else if (statfs("/proc/self/cmdline", &path_st) < 0) {
+            FAIL("statfs failed");
+        } else if (fstatfs(fd, &fd_st) < 0) {
+            FAIL("fstatfs failed");
+        } else {
+            EXPECT_TRUE(path_st.f_type == fd_st.f_type,
+                        "statfs/fstatfs f_type mismatch");
+        }
+        if (fd >= 0)
+            close(fd);
+    }
+
+    /* /proc/self/fd and /proc/self/fdinfo are directories whose *open* path
+     * allocates a fresh scratch directory with one placeholder file per live
+     * guest fd (see proc_open_fd_scratch) -- work a plain statfs() has no use
+     * for. These exercise the statfs-specific bypass (proc_intercept_statfs)
+     * added alongside issue #141's fix, including enough repeats to exceed
+     * PROC_SCRATCH_DIRS_MAX (128). The scratch dirs live under the real host
+     * /tmp, outside anything this guest binary can see, so the loop can only
+     * confirm statfs() keeps succeeding -- it cannot observe, and so cannot
+     * fail on, whether the bypass actually avoids allocating those dirs.
+     */
+    TEST("statfs /proc/self/fd");
+    {
+        struct statfs st;
+        int ok = 1;
+        for (int i = 0; i < 200 && ok; i++)
+            if (statfs("/proc/self/fd", &st) < 0)
+                ok = 0;
+        EXPECT_TRUE(ok, "statfs failed");
+    }
+
+    TEST("statfs /proc/self/fdinfo");
+    {
+        struct statfs st;
+        int ok = 1;
+        for (int i = 0; i < 200 && ok; i++)
+            if (statfs("/proc/self/fdinfo", &st) < 0)
+                ok = 0;
+        EXPECT_TRUE(ok, "statfs failed");
+    }
+
     TEST("chdir /proc preserves guest cwd");
     {
         char cwd[256];