llext: Proper support for Thumb32 relocations

[edersondisouza]

llext: Proper support for Thumb32 relocations

[mathieuchopstm]

Taking the Linux kernel's linker as a reference point - relevant parts being:

• https://github.com/torvalds/linux/blob/a38297e3fb012ddfa7ce0321a7e5a8daeb1872b6/kernel/module/main.c#L1367-L1436
• https://github.com/torvalds/linux/blob/a38297e3fb012ddfa7ce0321a7e5a8daeb1872b6/arch/arm/kernel/module.c#L109-L430

The main difference with Zephyr, as far as I can tell - and ignoring unrelated details - seems to be that Linux completely ignores the ELF_ST_TYPE of symbols; on the other hand, Zephyr only relocates symbols if they have specific values for it, as seen here:

zephyr/subsys/llext/llext.c

Lines 840 to 852 in 8c8e2f6

	} else if (ELF_ST_TYPE(sym.st_info) == STT_SECTION ||
	ELF_ST_TYPE(sym.st_info) == STT_FUNC ||
	ELF_ST_TYPE(sym.st_info) == STT_OBJECT) {
	/* Link address is relative to the start of the section */
	link_addr = (uintptr_t)ext->mem[ldr->sect_map[sym.st_shndx]]
	+ sym.st_value;
	LOG_INF("found section symbol %s addr 0x%lx", name, link_addr);
	} else {
	/* Nothing to relocate here */
	LOG_DBG("not relocated");
	continue;
	}

@edersondisouza could you try if the following patch fixes the issue on your side?

diff --git a/subsys/llext/llext.c b/subsys/llext/llext.c
index c92f49abd60..16748d941b4 100644
--- a/subsys/llext/llext.c
+++ b/subsys/llext/llext.c
@@ -835,18 +835,15 @@ static int llext_link(struct llext_loader *ldr, struct llext *ext, bool do_local
                                } else {
                                        LOG_INF("found symbol %s at 0x%lx", name, link_addr);
                                }
-                       } else if (ELF_ST_TYPE(sym.st_info) == STT_SECTION ||
-                                  ELF_ST_TYPE(sym.st_info) == STT_FUNC ||
-                                  ELF_ST_TYPE(sym.st_info) == STT_OBJECT) {
+                       } else if (IN_RANGE(sym.st_shndx, 0xFF00, 0xFFFF)) {
+                               //0xFF00 => SHN_LORESERVE, 0xFFFF => SHN_HIRESERVE
+                               LOG_DBG("section index in RESERVED range - not relocating");
+                               continue;
+                       } else {
                                /* Link address is relative to the start of the section */
                                link_addr = (uintptr_t)ext->mem[ldr->sect_map[sym.st_shndx]]
                                        + sym.st_value;
 
                                LOG_INF("found section symbol %s addr 0x%lx", name, link_addr);
-                       } else {
-                               /* Nothing to relocate here */
-                               LOG_DBG("not relocated");
-                               continue;
                        }
 
                        LOG_INF("writing relocation symbol %s type %zd sym %zd at addr 0x%lx "

Edit: re-reading the comment thread in EDK PR, I have seen the issue appeared on qemu_cortex_r5, so I tested on my side. Using ./scripts/twister -p qemu_cortex_r5 --testsuite-root tests/subsys/llext -K (-K required because qemu_cortex_r5 is excluded from LLEXT tests):

• Without patch, 8 tests are performed: one is skipped, one passes and the other six fail
• With the patch, one is still skipped but all other tests pass 😄

[edersondisouza]

@mathieuchopstm I've tested it here with the sample and can confirm that your patch does work!!! Will you submit a PR, please?

[mathieuchopstm]

@edersondisouza PR with fix submitted