[LibOS,PAL] Emulate file-backed mmap via PAL read/write APIs

Documentation/pal/host-abi.rst

@@ -175,9 +175,6 @@ applications.
 .. doxygenfunction:: PalStreamDelete
    :project: pal
 
-.. doxygenfunction:: PalStreamMap
-   :project: pal
-
 .. doxygenfunction:: PalStreamSetLength
    :project: pal
 
@@ -366,3 +363,6 @@ random bits, to obtain an attestation report and quote, etc.
 
 .. doxygenfunction:: PalGetSpecialKey
    :project: pal
+
+.. doxygenfunction:: PalDeviceMap
+   :project: pal

libos/include/libos_fs.h

@@ -108,21 +108,28 @@ struct libos_fs_ops {
     /*
      * \brief Map file at an address.
      *
-     * \param hdl     File handle.
-     * \param addr    Address of the memory region. Cannot be NULL.
-     * \param size    Size of the memory region.
-     * \param prot    Permissions for the memory region (`PROT_*`).
-     * \param flags   `mmap` flags (`MAP_*`).
-     * \param offset  Offset in file.
-     *
-     * Maps the file at given address. This might involve mapping directly (`PalStreamMap`), or
+     * \param hdl                  File handle.
+     * \param addr                 Address of the memory region. Cannot be NULL.
+     * \param size                 Size of the memory region.
+     * \param prot                 Permissions for the memory region (`PROT_*`).
+     * \param flags                `mmap` flags (`MAP_*`).
+     * \param offset               Offset in file.
+     * \param[out] out_valid_size  Valid size (i.e. backed by file).
+     *
+     * Maps the file at given address. This might involve mapping directly (`PalDeviceMap`), or
      * mapping anonymous memory (`PalVirtualMemoryAlloc`) and writing data.
      *
+     * The contents of the mapping are initialized using `size` bytes starting at `offset` offset in
+     * the file. For a file size that is not a multiple of the page size, the remaining bytes on the
+     * last page are zeroed. Pages that are not backed by file contents are inaccessible
+     * (effectively they have PROT_NONE permissions). This function returns the valid size (i.e. the
+     * pages backed by file contents) in `out_valid_size`.
+     *
      * `addr`, `offset` and `size` must be alloc-aligned (see `IS_ALLOC_ALIGNED*` macros in
      * `libos_internal.h`).
      */
     int (*mmap)(struct libos_handle* hdl, void* addr, size_t size, int prot, int flags,
-                uint64_t offset);
+                uint64_t offset, size_t* out_valid_size);
 
     /*
      * \brief Write back mapped memory to file.
@@ -968,7 +975,7 @@ file_off_t generic_inode_seek(struct libos_handle* hdl, file_off_t offset, int o
 int generic_inode_poll(struct libos_handle* hdl, int in_events, int* out_events);
 
 int generic_emulated_mmap(struct libos_handle* hdl, void* addr, size_t size, int prot, int flags,
-                          uint64_t offset);
+                          uint64_t offset, size_t* valid_size);
 int generic_emulated_msync(struct libos_handle* hdl, void* addr, size_t size, int prot, int flags,
                            uint64_t offset);
 int generic_truncate(struct libos_handle* hdl, file_off_t size);

libos/include/libos_handle.h

@@ -303,8 +303,6 @@ int init_exec_handle(const char* const* argv, char*** out_new_argv);
 
 int open_executable(struct libos_handle* hdl, const char* path);
 
-int get_file_size(struct libos_handle* file, uint64_t* size);
-
 ssize_t do_handle_read(struct libos_handle* hdl, void* buf, size_t count);
 ssize_t do_handle_write(struct libos_handle* hdl, const void* buf, size_t count);
 

libos/include/libos_vma.h

@@ -26,6 +26,7 @@
 struct libos_vma_info {
     void* addr;
     size_t length;
+    size_t valid_length; // memory accesses beyond valid_length result in SIGBUS/EFAULT
     int prot;  // memory protection flags: PROT_*
     int flags; // MAP_* and VMA_*
     struct libos_handle* file;
@@ -99,6 +100,9 @@ int bkeep_mmap_any(size_t length, int prot, int flags, struct libos_handle* file
 int bkeep_mmap_any_aslr(size_t length, int prot, int flags, struct libos_handle* file,
                         uint64_t offset, const char* comment, void** ret_val_ptr);
 
+/* Looks up VMA that starts at `begin_addr` and if found, updates `vma->valid_length`. */
+int bkeep_vma_update_valid_length(void* begin_addr, size_t valid_length);
+
 /* Looking up VMA that contains `addr`. If one is found, returns its description in `vma_info`.
  * This function increases ref-count of `vma_info->file` by one (if it is not NULL). */
 int lookup_vma(void* addr, struct libos_vma_info* vma_info);
@@ -133,6 +137,9 @@ int msync_handle(struct libos_handle* hdl);
 /* Reload file mappings of `hdl` */
 int reload_mmaped_from_file_handle(struct libos_handle* hdl);
 
+/* Refresh page protections of file mappings of `hdl` when the file size has changed */
+int prot_refresh_mmaped_from_file_handle(struct libos_handle* hdl, size_t file_size);
+
 void debug_print_all_vmas(void);
 
 /* Returns the peak amount of memory usage */

libos/src/bookkeep/libos_handle.c

@@ -548,27 +548,6 @@ void put_handle(struct libos_handle* hdl) {
     }
 }
 
-int get_file_size(struct libos_handle* hdl, uint64_t* size) {
-    if (!hdl->fs || !hdl->fs->fs_ops)
-        return -EINVAL;
-
-    if (hdl->fs->fs_ops->hstat) {
-        struct stat stat;
-        int ret = hdl->fs->fs_ops->hstat(hdl, &stat);
-        if (ret < 0) {
-            return ret;
-        }
-        if (stat.st_size < 0) {
-            return -EINVAL;
-        }
-        *size = (uint64_t)stat.st_size;
-        return 0;
-    }
-
-    *size = 0;
-    return 0;
-}
-
 static struct libos_handle_map* get_new_handle_map(uint32_t size) {
     struct libos_handle_map* handle_map = calloc(1, sizeof(struct libos_handle_map));
 

libos/src/bookkeep/libos_signal.c

@@ -361,12 +361,8 @@ static void memfault_upcall(bool is_in_pal, uintptr_t addr, PAL_CONTEXT* context
         struct libos_handle* file = vma_info.file;
         if (file && file->type == TYPE_CHROOT) {
             /* If the mapping exceeds end of a file then return a SIGBUS. */
-            lock(&file->inode->lock);
-            file_off_t size = file->inode->size;
-            unlock(&file->inode->lock);
-
-            uintptr_t eof_in_vma = (uintptr_t)vma_info.addr + (size - vma_info.file_offset);
-            if (addr > eof_in_vma) {
+            uintptr_t eof_in_vma = (uintptr_t)vma_info.addr + vma_info.valid_length;
+            if (addr >= eof_in_vma) {
                 info.si_signo = SIGBUS;
                 info.si_code = BUS_ADRERR;
             } else {