Input: synaptics-rmi4 - validate register descriptor structure against its declared size
Subsystems (2)
PATCH 1/2 Input: synaptics-rmi4 - validate register descriptor structure against its declared size
rmi_read_register_desc() trusts three independent device-supplied quantities to be correct:
- struct_size, taken from the presence-register header (up to 65535
via buf[1]|buf[2]<<8 when buf[0]==0),
- num_registers, the popcount of the presence bitmap (up to 255),
- the per-register entries inside struct_buf, each a variable-length {reg_size, subpacket-continuation-bytes...} record.
But nothing checks that num_registers entries actually fit in struct_size bytes, and nothing bounds the subpacket continuation chain, which can cause two different types of overruns if left unchecked.
Fix this all up by properly checking the values every time and aborting if anything is out of range.
--- drivers/input/rmi4/rmi_driver.c | 19 ++++++++++++++++++- 1 file changed, 18 insertions(+), 1 deletion(-)
drivers/input/rmi4/rmi_driver.c +18 -1
@@ -643,16 +643,24 @@ int rmi_read_register_desc(struct rmi_device *d, u16 addr, reg =find_first_bit (rdesc->presense_map,RMI_REG_DESC_PRESENSE_BITS );for (i =0 ; i < rdesc->num_registers; i++) {struct rmi_register_desc_item *item = &rdesc->registers[i]; -int reg_size = struct_buf[offset]; +int reg_size; + +if (offset >= rdesc->struct_size) + goto malformed; + reg_size = struct_buf[offset]; ++offset;if (reg_size ==0 ) { +if (offset +2 > rdesc->struct_size) + goto malformed; reg_size = struct_buf[offset] | (struct_buf[offset +1 ] <<8 ); offset +=2 ; }if (reg_size ==0 ) { +if (offset +4 > rdesc->struct_size) + goto malformed; reg_size = struct_buf[offset] | (struct_buf[offset +1 ] <<8 ) | (struct_buf[offset + 2] << 16) |
@@ -666,6 +674,9 @@ int rmi_read_register_desc(struct rmi_device *d, u16 addr, map_offset =0 ; do { + if (offset >= rdesc->struct_size || + map_offset >= RMI_REG_DESC_SUBPACKET_BITS) + goto malformed; for (b = 0; b < 7; b++) { if (struct_buf[offset] & (0x1 << b)) bitmap_set(item->subpacket_map,
@@ -688,6 +699,12 @@ int rmi_read_register_desc(struct rmi_device *d, u16 addr, free_struct_buff:kfree (struct_buf);return ret; + +malformed: +dev_err (&d->dev, +"register descriptor structure does not match its declared size\n" ); + ret = -EIO ; + goto free_struct_buff; } const struct rmi_register_desc_item *rmi_get_register_desc_item(
Signature · 1 line
2.53.0
PATCH 2/2 Input: synaptics-rmi4 - use u32 for reg_size to avoid sign extension into item->reg_size
rmi_read_register_desc() builds the 4-byte register size from device bytes:
reg_size = struct_buf[offset] | (struct_buf[offset + 1] << 8) | (struct_buf[offset + 2] << 16) | (struct_buf[offset + 3] << 24);
struct_buf is u8 *, so each byte is promoted to int before the shift. A device that supplies a top byte with bit 7 set (e.g. 00 00 00 00 00 00 80 in struct_buf to reach the 4-byte path with offset+3 = 0x80) makes (0x80 << 24) overflow into the int sign bit, and the OR result is negative. reg_size is then assigned to item->reg_size, which is unsigned long, so the negative int sign-extends to a value near ULONG_MAX.
After this, bad things happen when numbers start wrapping and buffers are allocatged based on those numbers, and then accessed based on those buffers assuming to be a sane size (bigger or smaller).
Fix this all up by just properly making reg_size be a u32.
--- drivers/input/rmi4/rmi_driver.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/input/rmi4/rmi_driver.c +1 -1
@@ -643,7 +643,7 @@ int rmi_read_register_desc(struct rmi_device *d, u16 addr, reg =find_first_bit (rdesc->presense_map,RMI_REG_DESC_PRESENSE_BITS );for (i =0 ; i < rdesc->num_registers; i++) {struct rmi_register_desc_item *item = &rdesc->registers[i]; -int reg_size; +u32 reg_size;if (offset >= rdesc->struct_size) goto malformed;
Signature · 1 line
2.53.0
Dmitry Torokhov
maintainerRe: [PATCH 2/2] Input: synaptics-rmi4 - use u32 for reg_size to avoid sign extension into item->reg_size
Quoted 16 lines · Greg Kroah-Hartman
rmi_read_register_desc()builds the 4-byte register size from device bytes:reg_size = struct_buf[offset] | (struct_buf[offset + 1] << 8) | (struct_buf[offset + 2] << 16) | (struct_buf[offset + 3] << 24);
struct_bufis u8 *, so each byte is promoted to int before the shift. A device that supplies a top byte with bit 7 set (e.g. 00 00 00 00 00 00 80 instruct_bufto reach the 4-byte path with offset+3 =0x80) makes (0x80<< 24) overflow into the int sign bit, and the OR result is negative.reg_sizeis then assigned toitem->reg_size, which is unsigned long, so the negative int sign-extends to a value nearULONG_MAX.
Quoted 5 lines
After this, bad things happen when numbers start wrapping and buffers are allocatged based on those numbers, and then accessed based on those buffers assuming to be a sane size (bigger or smaller).
Fix this all up by just properly making
reg_sizebe a u32.
Shouldn't it be replaced with get_unaligned_le32() instead?
Thanks.
Signature · 1 line
Dmitry
Re: [PATCH 2/2] Input: synaptics-rmi4 - use u32 for reg_size to avoid sign extension into item->reg_size
Quoted 17 lines · Dmitry Torokhov
Quoted 16 lines · Greg Kroah-Hartman
rmi_read_register_desc()builds the 4-byte register size from device bytes:reg_size = struct_buf[offset] | (struct_buf[offset + 1] << 8) | (struct_buf[offset + 2] << 16) | (struct_buf[offset + 3] << 24);
struct_bufis u8 *, so each byte is promoted to int before the shift. A device that supplies a top byte with bit 7 set (e.g. 00 00 00 00 00 00 80 instruct_bufto reach the 4-byte path with offset+3 =0x80) makes (0x80<< 24) overflow into the int sign bit, and the OR result is negative.reg_sizeis then assigned toitem->reg_size, which is unsigned long, so the negative int sign-extends to a value nearULONG_MAX.
Quoted 7 lines
Quoted 5 lines
After this, bad things happen when numbers start wrapping and buffers are allocatged based on those numbers, and then accessed based on those buffers assuming to be a sane size (bigger or smaller).
Fix this all up by just properly making
reg_sizebe a u32.Shouldn't it be replaced with
get_unaligned_le32()instead?
Sure, be glad to do that instead, this was just a "simpler" solution :)
I'll respin that after -rc1 is out.
thanks,
greg k-h