From: Jinoh Kang jinoh.kang.kr@gmail.com
The main role of the preloader is to reserve specific virtual memory address ranges used for special purposes on Windows, before Wine could be loaded.
It achieves this goal via the following process:
(1) It eliminates future allocations of addresses in the reserved ranges. Specifically, it issues a series of mmap() calls with PROT_NONE protection to reserve those ranges, so that the OS won't allocate any of the reserved addresses for other users (i.e. Unix system libraries).
(2) It eliminates current references to addresses in the reserved ranges. Specifically, if the vDSO had occupied one of the reserved ranges, the preloader removes it from the auxiliary vector (AT_SYSINFO*).
(3) If (2) is not possible because the address is in use (e.g. current thread stack), it gives up reservation and removes the reserved range from preload_info.
Today, each virtual memory area (VMA) is treated as follows when it overlaps with Wine's reserved address ranges.
- Preloader code/data. Preloader should leave no trace of itself after Wine has been loaded. Thus, no current references to preloader code or data remain. (2) is a no-op in this case. Meanwhile, if any part of the preloader overlaps with a reserved range, that part is overwritten. This could lead to crash if it ever touched any part of code or data that are still being used.
- vDSO/vvar. These are overwritten and ignored completely when they overlap with any reserved range. In other words, both (1) and (2) are performed.
- Stack. Since the stack is always in use, (2) is not possible. Therefore, (3) is performed.
There are a few issues with this approach:
1. Existing VMAs that overlap with any reserved ranges are forcibly overwritten during (1). There is actually no need to overwrite them, since existing VMAs themselves automatically act as reservations by nature (i.e. no future allocations would overlap any existing VMAs). Furthermore, arbitrarily overwriting any memory in use would cause the preloader to crash. The only treatment required for existing VMAs is either (2) or (3), not (1).
2. (1) irrevocably overwrites some useful preexisting VMAs such as vDSO if they overlap with any reserved ranges. Newer versions of Linux kernel supports relocating vDSO, which can be used to move it outside of reserved address ranges instead of discarding it. To do so, however, we first have to allocate a _new_ address for such VMAs before the overlapping address range could be reserved. Notice a chicken-egg problem here:
- If we perform (1) before allocating a new address for vDSO, the vDSO goes away even before we get a chance to relocate it.
- If we allocate a new address for vDSO before performing (1), the new address allocated by the OS might end up overlapping with one of the reserved ranges.
What we need here is a way to mmap()-fill all unallocated regions inside the reserved ranges, *while* still keeping existing VMAs intact. In this way, we can perform (1) foremost to avoid allocating a reserved address for relocated vDSO. After the vDSO is relocated to a safe address, we can perform (1) once more to finalise the reservation.
3. Only the stack receives the special treatment of not being overwritten by PROT_NONE allocation from (1). Theoretically other VMAs that are in use such as the preloader code and data shall receive the equal treatment anyway.
Fix this by reading /proc/self/maps for existing VMAs, and splitting mmap() calls to avoid erasing existing memory mappings.
Note that MAP_FIXED_NOREPLACE is not suitable for this kind of job: it fails entirely if there exist *any* overlapping memory mappings.
Signed-off-by: Jinoh Kang jinoh.kang.kr@gmail.com --- loader/preloader.c | 434 ++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 411 insertions(+), 23 deletions(-)
diff --git a/loader/preloader.c b/loader/preloader.c index 88865587975..785b6153eaf 100644 --- a/loader/preloader.c +++ b/loader/preloader.c @@ -184,6 +184,32 @@ struct preloader_state struct stackarg_info s; };
+/* Buffer for line-buffered I/O read. */ +struct linebuffer +{ + char *base; /* start of the buffer */ + char *limit; /* last byte of the buffer (for NULL terminator) */ + char *head; /* next byte to write to */ + char *tail; /* next byte to read from */ + int truncated; /* line truncated? (if true, skip until next line) */ +}; + +struct vma_area +{ + unsigned long start; + unsigned long end; +}; + +struct vma_area_list +{ + struct vma_area *base; + struct vma_area *list_end; + struct vma_area *alloc_end; +}; + +#define FOREACH_VMA(list, item) \ + for ((item) = (list)->base; (item) != (list)->list_end; (item)++) + /* * The __bb_init_func is an empty function only called when file is * compiled with gcc flags "-fprofile-arcs -ftest-coverage". This @@ -723,6 +749,44 @@ static size_t wld_strlen( const char *str ) return ptr - str; }
+static inline void *wld_memmove( void *dest, const void *src, size_t len ) +{ + unsigned char *destp = dest; + const unsigned char *srcp = src; + + /* Two area overlaps and src precedes dest? + * + * Note: comparing pointers to different objects leads to undefined + * behavior in C; therefore, we cast them to unsigned long for comparison + * (which is implementation-defined instead). This also allows us to rely + * on unsigned overflow on dest < src (forward copy case) in which case the + * LHS exceeds len and makes the condition false. + */ + if ((unsigned long)dest - (unsigned long)src < len) + { + destp += len; + srcp += len; + while (len--) *--destp = *--srcp; + } + else + { + while (len--) *destp++ = *srcp++; + } + + return dest; +} + +static inline void *wld_memchr( const void *mem, int val, size_t len ) +{ + const unsigned char *ptr = mem, *end = (const unsigned char *)ptr + len; + + for (ptr = mem; ptr != end; ptr++) + if (*ptr == (unsigned char)val) + return (void *)ptr; + + return NULL; +} + /* * parse_ul - parse an unsigned long number with given radix * @@ -1516,6 +1580,347 @@ static void set_process_name( int argc, char *argv[] ) for (i = 1; i < argc; i++) argv[i] -= off; }
+/* + * linebuffer_init + * + * Initialise a linebuffer with the given buffer. + */ +static void linebuffer_init( struct linebuffer *lbuf, char *base, size_t len ) +{ + lbuf->base = base; + lbuf->limit = base + (len - 1); /* NULL terminator */ + lbuf->head = base; + lbuf->tail = base; + lbuf->truncated = 0; +} + +/* + * linebuffer_getline + * + * Retrieve a line from the linebuffer. + * If a line is longer than the allocated buffer, then the line is truncated; + * the truncated flag is set to indicate this condition. + */ +static char *linebuffer_getline( struct linebuffer *lbuf ) +{ + char *lnp, *line; + + while ((lnp = wld_memchr( lbuf->tail, '\n', lbuf->head - lbuf->tail ))) + { + /* Consume the current line from the buffer. */ + line = lbuf->tail; + lbuf->tail = lnp + 1; + + if (!lbuf->truncated) + { + *lnp = '\0'; + return line; + } + + /* Remainder of a previously truncated line; ignore it. */ + lbuf->truncated = 0; + } + + if (lbuf->tail == lbuf->base && lbuf->head == lbuf->limit) + { + /* We have not encountered the end of the current line yet; however, + * the buffer is full and cannot be compacted to accept more + * characters. Truncate the line here, and consume it from the buffer. + */ + line = lbuf->tail; + lbuf->tail = lbuf->head; + + /* Ignore any further characters until the start of the next line. */ + lbuf->truncated = 1; + *lbuf->head = '\0'; + return line; + } + + if (lbuf->tail != lbuf->base) + { + /* Compact the buffer. Make room for reading more data by zapping the + * leading gap in the buffer. + */ + wld_memmove( lbuf->base, lbuf->tail, lbuf->head - lbuf->tail); + lbuf->head -= lbuf->tail - lbuf->base; + lbuf->tail = lbuf->base; + } + + return NULL; +} + +/* + * parse_maps_line + * + * Parse an entry from /proc/self/maps file into a vma_area structure. + */ +static int parse_maps_line( struct vma_area *entry, const char *line ) +{ + struct vma_area item = { 0 }; + char *ptr = (char *)line; + int overflow; + + item.start = parse_ul( ptr, &ptr, 16, &overflow ); + if (overflow) return -1; + if (*ptr != '-') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + + item.end = parse_ul( ptr, &ptr, 16, &overflow ); + if (overflow) item.end = -page_size; + if (*ptr != ' ') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + + if (item.start >= item.end) return -1; + + if (*ptr != 'r' && *ptr != '-') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + if (*ptr != 'w' && *ptr != '-') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + if (*ptr != 'x' && *ptr != '-') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + if (*ptr != 's' && *ptr != 'p') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + if (*ptr != ' ') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + + parse_ul( ptr, &ptr, 16, NULL ); + if (*ptr != ' ') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + + parse_ul( ptr, &ptr, 16, NULL ); + if (*ptr != ':') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + + parse_ul( ptr, &ptr, 16, NULL ); + if (*ptr != ' ') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + + parse_ul( ptr, &ptr, 10, NULL ); + if (*ptr != ' ') fatal_error( "parse error in /proc/self/maps\n" ); + ptr++; + + *entry = item; + return 0; +} + +/* + * lookup_vma_entry + * + * Find the first VMA of which end address is greater than the given address. + */ +static struct vma_area *lookup_vma_entry( const struct vma_area_list *list, unsigned long address ) +{ + const struct vma_area *left = list->base, *right = list->list_end, *mid; + while (left < right) + { + mid = left + (right - left) / 2; + if (mid->end <= address) left = mid + 1; + else right = mid; + } + return (struct vma_area *)left; +} + +/* + * map_reserve_range + * + * Reserve the specified address range. + * If there are any existing VMAs in the range, they are replaced. + */ +static int map_reserve_range( void *addr, size_t size ) +{ + if (addr == (void *)-1 || + wld_mmap( addr, size, PROT_NONE, + MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0) != addr) + return -1; + return 0; +} + +/* + * map_reserve_unmapped_range + * + * Reserve the specified address range excluding already mapped areas. + */ +static int map_reserve_unmapped_range( const struct vma_area_list *list, void *addr, size_t size ) +{ + unsigned long range_start = (unsigned long)addr, + range_end = (unsigned long)addr + size; + const struct vma_area *start, *item; + unsigned long last_addr = range_start; + + start = lookup_vma_entry( list, range_start ); + for (item = start; item != list->list_end && item->start < range_end; item++) + { + if (item->start > last_addr && + map_reserve_range( (void *)last_addr, item->start - last_addr ) < 0) + goto fail; + last_addr = item->end; + } + + if (range_end > last_addr && + map_reserve_range( (void *)last_addr, range_end - last_addr ) < 0) + goto fail; + return 0; + +fail: + while (item != start) + { + item--; + last_addr = item == start ? range_start : item[-1].end; + if (item->start > last_addr) + wld_munmap( (void *)last_addr, item->start - last_addr ); + } + return -1; +} + +/* + * insert_vma_entry + * + * Insert the given VMA into the list. + */ +static void insert_vma_entry( struct vma_area_list *list, const struct vma_area *item ) +{ + struct vma_area *left = list->base, *right = list->list_end, *mid; + + if (left < right) + { + mid = right - 1; /* optimisation: start search from end */ + for (;;) + { + if (mid->end < item->end) left = mid + 1; + else right = mid; + if (left >= right) break; + mid = left + (right - left) / 2; + } + } + wld_memmove(left + 1, left, list->list_end - left); + wld_memmove(left, item, sizeof(*item)); + list->list_end++; + return; +} + +/* + * scan_vma + * + * Parse /proc/self/maps into the given VMA area list. + */ +static void scan_vma( struct vma_area_list *list, size_t *real_count ) +{ + int fd; + size_t n = 0; + ssize_t nread; + struct linebuffer lbuf; + char buffer[80 + PATH_MAX], *line; + struct vma_area item; + + fd = wld_open( "/proc/self/maps", O_RDONLY ); + if (fd == -1) fatal_error( "could not open /proc/self/maps\n" ); + + linebuffer_init(&lbuf, buffer, sizeof(buffer)); + for (;;) + { + nread = wld_read( fd, lbuf.head, lbuf.limit - lbuf.head ); + if (nread < 0) fatal_error( "could not read /proc/self/maps\n" ); + if (nread == 0) break; + lbuf.head += nread; + + while ((line = linebuffer_getline( &lbuf ))) + { + if (parse_maps_line( &item, line ) >= 0) + { + if (list->list_end < list->alloc_end) insert_vma_entry( list, &item ); + n++; + } + } + } + + wld_close(fd); + *real_count = n; +} + +/* + * free_vma_list + * + * Free the buffer in the given VMA list. + */ +static void free_vma_list( struct vma_area_list *list ) +{ + if (list->base) + wld_munmap( list->base, + (unsigned char *)list->alloc_end - (unsigned char *)list->base ); + list->base = NULL; + list->list_end = NULL; + list->alloc_end = NULL; +} + +/* + * alloc_scan_vma + * + * Parse /proc/self/maps into a newly allocated VMA area list. + */ +static void alloc_scan_vma( struct vma_area_list *listp ) +{ + size_t max_count = page_size / sizeof(struct vma_area); + struct vma_area_list vma_list; + + for (;;) + { + vma_list.base = wld_mmap( NULL, sizeof(struct vma_area) * max_count, + PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, + -1, 0 ); + if (vma_list.base == (struct vma_area *)-1) + fatal_error( "could not allocate memory for VMA list\n"); + vma_list.list_end = vma_list.base; + vma_list.alloc_end = vma_list.base + max_count; + + scan_vma( &vma_list, &max_count ); + if (vma_list.list_end - vma_list.base == max_count) + { + wld_memmove(listp, &vma_list, sizeof(*listp)); + break; + } + + free_vma_list( &vma_list ); + } +} + +/* + * map_reserve_preload_ranges + * + * Attempt to reserve memory ranges into preload_info. + * If any preload_info entry overlaps with stack, remove the entry instead of + * reserving. + */ +static void map_reserve_preload_ranges( const struct vma_area_list *vma_list, + const struct stackarg_info *stackinfo ) +{ + size_t i; + unsigned long exclude_start = (unsigned long)stackinfo->stack - 1; + unsigned long exclude_end = (unsigned long)stackinfo->auxv + 1; + + for (i = 0; preload_info[i].size; i++) + { + if (exclude_end > (unsigned long)preload_info[i].addr && + exclude_start <= (unsigned long)preload_info[i].addr + preload_info[i].size - 1) + { + remove_preload_range( i ); + i--; + } + else if (map_reserve_unmapped_range( vma_list, preload_info[i].addr, preload_info[i].size ) < 0) + { + /* don't warn for low 64k */ + if (preload_info[i].addr >= (void *)0x10000 +#ifdef __aarch64__ + && preload_info[i].addr < (void *)0x7fffffffff /* ARM64 address space might end here*/ +#endif + ) + wld_printf( "preloader: Warning: failed to reserve range %p-%p\n", + preload_info[i].addr, (char *)preload_info[i].addr + preload_info[i].size ); + remove_preload_range( i ); + i--; + } + } +} +
/* * wld_start @@ -1532,6 +1937,7 @@ void* wld_start( void **stack ) struct wld_link_map main_binary_map, ld_so_map; struct wine_preload_info **wine_main_preload_info; struct preloader_state state = { 0 }; + struct vma_area_list vma_list = { NULL };
parse_stackargs( &state.s, *stack );
@@ -1560,29 +1966,9 @@ void* wld_start( void **stack ) /* reserve memory that Wine needs */ reserve = stackargs_getenv( &state.s, "WINEPRELOADRESERVE" ); if (reserve) preload_reserve( reserve ); - for (i = 0; preload_info[i].size; i++) - { - if ((char *)state.s.auxv >= (char *)preload_info[i].addr && - (char *)state.s.stack <= (char *)preload_info[i].addr + preload_info[i].size) - { - remove_preload_range( i ); - i--; - } - else if (wld_mmap( preload_info[i].addr, preload_info[i].size, PROT_NONE, - MAP_FIXED | MAP_PRIVATE | MAP_ANON | MAP_NORESERVE, -1, 0 ) == (void *)-1) - { - /* don't warn for low 64k */ - if (preload_info[i].addr >= (void *)0x10000 -#ifdef __aarch64__ - && preload_info[i].addr < (void *)0x7fffffffff /* ARM64 address space might end here*/ -#endif - ) - wld_printf( "preloader: Warning: failed to reserve range %p-%p\n", - preload_info[i].addr, (char *)preload_info[i].addr + preload_info[i].size ); - remove_preload_range( i ); - i--; - } - } + + alloc_scan_vma( &vma_list ); + map_reserve_preload_ranges( &vma_list, &state.s );
/* add an executable page at the top of the address space to defeat * broken no-exec protections that play with the code selector limit */ @@ -1645,6 +2031,8 @@ void* wld_start( void **stack ) } #endif
+ free_vma_list( &vma_list ); + return (void *)ld_so_map.l_entry; }