From 0b2fa98aa5e5dbdac4f5e2b2f67a34cc34dcc6b8 Mon Sep 17 00:00:00 2001 From: Simon Glass Date: Wed, 16 Dec 2020 21:20:06 -0700 Subject: linker_lists: Fix alignment issue The linker script uses alphabetic sorting to group the different linker lists together. Each group has its own struct and potentially its own alignment. But when the linker packs the structs together it cannot ensure that a linker list starts on the expected alignment boundary. For example, if the first list has a struct size of 8 and we place 3 of them in the image, that means that the next struct will start at offset 0x18 from the start of the linker_list section. If the next struct has a size of 16 then it will start at an 8-byte aligned offset, but not a 16-byte aligned offset. With sandbox on x86_64, a reference to a linker list item using ll_entry_get() can force alignment of that particular linker_list item, if it is in the same file as the linker_list item is declared. Consider this example, where struct driver is 0x80 bytes: ll_entry_declare(struct driver, fred, driver) ... void *p = ll_entry_get(struct driver, fred, driver) If these two lines of code are in the same file, then the entry is forced to be aligned at the 'struct driver' alignment, which is 16 bytes. If the second line of code is in a different file, then no action is taken, since the compiler cannot update the alignment of the linker_list item. In the first case, an 8-byte 'fill' region is added: .u_boot_list_2_driver_2_testbus_drv 0x0000000000270018 0x80 test/built-in.o 0x0000000000270018 _u_boot_list_2_driver_2_testbus_drv .u_boot_list_2_driver_2_testfdt1_drv 0x0000000000270098 0x80 test/built-in.o 0x0000000000270098 _u_boot_list_2_driver_2_testfdt1_drv *fill* 0x0000000000270118 0x8 .u_boot_list_2_driver_2_testfdt_drv 0x0000000000270120 0x80 test/built-in.o 0x0000000000270120 _u_boot_list_2_driver_2_testfdt_drv .u_boot_list_2_driver_2_testprobe_drv 0x00000000002701a0 0x80 test/built-in.o 0x00000000002701a0 _u_boot_list_2_driver_2_testprobe_drv With this, the linker_list no-longer works since items after testfdt1_drv are not at the expected address. Ideally we would have a way to tell gcc not to align structs in this way. It is not clear how we could do this, and in any case it would require us to adjust every struct used by the linker_list feature. One possible fix is to force each separate linker_list to start on the largest possible boundary that can be required by the compiler. However that does not seem to work on x86_64, which uses 16-byte alignment in this case but needs 32-byte alignment. So add a Kconfig option to handle this. Set the default value to 4 so as to avoid changing platforms that don't need it. Update the ll_entry_start() accordingly. Signed-off-by: Simon Glass --- doc/api/linker_lists.rst | 59 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 59 insertions(+) (limited to 'doc/api') diff --git a/doc/api/linker_lists.rst b/doc/api/linker_lists.rst index 72f514e0ac..7063fdc831 100644 --- a/doc/api/linker_lists.rst +++ b/doc/api/linker_lists.rst @@ -96,5 +96,64 @@ defined for the whole list and each sub-list: %u_boot_list_2_drivers_2_pci_3 %u_boot_list_2_drivers_3 +Alignment issues +---------------- + +The linker script uses alphabetic sorting to group the different linker +lists together. Each group has its own struct and potentially its own +alignment. But when the linker packs the structs together it cannot ensure +that a linker list starts on the expected alignment boundary. + +For example, if the first list has a struct size of 8 and we place 3 of +them in the image, that means that the next struct will start at offset +0x18 from the start of the linker_list section. If the next struct has +a size of 16 then it will start at an 8-byte aligned offset, but not a +16-byte aligned offset. + +With sandbox on x86_64, a reference to a linker list item using +ll_entry_get() can force alignment of that particular linker_list item, +if it is in the same file as the linker_list item is declared. + +Consider this example, where struct driver is 0x80 bytes:: + + ll_entry_declare(struct driver, fred, driver) + + ... + + void *p = ll_entry_get(struct driver, fred, driver) + +If these two lines of code are in the same file, then the entry is forced +to be aligned at the 'struct driver' alignment, which is 16 bytes. If the +second line of code is in a different file, then no action is taken, since +the compiler cannot update the alignment of the linker_list item. + +In the first case, an 8-byte 'fill' region is added:: + + .u_boot_list_2_driver_2_testbus_drv + 0x0000000000270018 0x80 test/built-in.o + 0x0000000000270018 _u_boot_list_2_driver_2_testbus_drv + .u_boot_list_2_driver_2_testfdt1_drv + 0x0000000000270098 0x80 test/built-in.o + 0x0000000000270098 _u_boot_list_2_driver_2_testfdt1_drv + *fill* 0x0000000000270118 0x8 + .u_boot_list_2_driver_2_testfdt_drv + 0x0000000000270120 0x80 test/built-in.o + 0x0000000000270120 _u_boot_list_2_driver_2_testfdt_drv + .u_boot_list_2_driver_2_testprobe_drv + 0x00000000002701a0 0x80 test/built-in.o + 0x00000000002701a0 _u_boot_list_2_driver_2_testprobe_drv + +With this, the linker_list no-longer works since items after testfdt1_drv +are not at the expected address. + +Ideally we would have a way to tell gcc not to align structs in this way. +It is not clear how we could do this, and in any case it would require us +to adjust every struct used by the linker_list feature. + +The simplest fix seems to be to force each separate linker_list to start +on the largest possible boundary that can be required by the compiler. This +is the purpose of CONFIG_LINKER_LIST_ALIGN + + .. kernel-doc:: include/linker_lists.h :internal: -- cgit v1.2.3