kern: put rela in bss (EXPERIMENTAL)

NOTE: This does some really gross things. It saves ~0x4000 of space, but
at the cost of crimes against the linker script. We should consider/evaluate
whether this is worth doing for real.

mesosphere/build_mesosphere.py

@@ -9,6 +9,22 @@ def align_up(val, algn):
     val += algn - 1
     return val - (val % algn)
 
+def find_rela(kernel, dynamic):
+    rela_offset, rela_size = (None, None)
+    while True:
+        dyn_type, dyn_val = up('<QQ', kernel[dynamic:dynamic+0x10])
+        if dyn_type == 0:
+            break
+        elif dyn_type == 7:
+            assert (rela_offset is None)
+            rela_offset = dyn_val
+        elif dyn_type == 8:
+            assert (rela_size is None)
+            rela_size = dyn_val
+        dynamic += 0x10
+    assert (rela_offset is not None)
+    assert (rela_size is not None)
+    return (rela_offset, rela_size)
 
 def main(argc, argv):
     if argc != 4:
@@ -21,8 +37,19 @@ def main(argc, argv):
     kernel_metadata_offset = 4
     assert (kernel_metadata_offset <= len(kernel) - 0x40)
     assert (kernel[kernel_metadata_offset:kernel_metadata_offset + 4] == b'MSS0')
-    kernel_end = up('<I', kernel[kernel_metadata_offset + 0x38:kernel_metadata_offset + 0x3C])[0]
-    assert (kernel_end >= len(kernel))
+    bss_start, bss_end, kernel_end, dynamic = up('<IIII', kernel[kernel_metadata_offset + 0x30:kernel_metadata_offset + 0x40])
+    assert (bss_end >= bss_start)
+    bss_size = bss_end - bss_start
+    assert (bss_end == kernel_end)
+    assert (kernel_end <= len(kernel))
+
+    rela_offset, rela_size = find_rela(kernel, dynamic)
+    assert (rela_size == len(kernel) - kernel_end)
+    assert (bss_start <= rela_offset and rela_offset + rela_size <= bss_end)
+    assert (kernel[bss_start:bss_end] == (b'\x00'* bss_size))
+
+    kernel = kernel[:rela_offset] + kernel[bss_end:] + (b'\x00' * (bss_size - (rela_size + (rela_offset - bss_start))))
+    assert (kernel_end == len(kernel))
 
     embedded_ini = b''
     try:

mesosphere/kernel/kernel.ld

@@ -113,9 +113,6 @@ SECTIONS
 	.gnu_extab        : ONLY_IF_RO { *(.gnu_extab*) } : rodata
 
 	.dynamic           : { *(.dynamic) } :rodata :dyn
-	.dynsym            : { *(.dynsym) } :rodata
-	.dynstr            : { *(.dynstr) } :rodata
-	.rela.dyn          : { *(.rela.*) } :rodata
 	.hash              : { *(.hash) } :rodata
 	.gnu.hash          : { *(.gnu.hash) } :rodata
 	.gnu.version       : { *(.gnu.version) } :rodata
@@ -152,18 +149,21 @@ SECTIONS
 	{
 		*(.data .data.* .gnu.linkonce.d.*)
 		SORT(CONSTRUCTORS)
+        . = ALIGN(8);
 	} :data
 
 	__bss_start__ = .;
-	.bss ALIGN(8) :
-	{
-		*(.dynbss)
-		*(.bss .bss.* .gnu.linkonce.b.*)
-		*(COMMON)
-		. = ALIGN(8);
-	} : data
 
-	. = ALIGN(0x1000);
+    OVERLAY : NOCROSSREFS {
+        .bss {
+            *(.dynbss)
+            *(.bss .bss.* .gnu.linkonce.b.*)
+            *(COMMON)
+            . = ALIGN(0x1000);
+        }
+        .rela.dyn { *(.rela.*) }
+    } :data
+
 
 	__bss_end__ = .;
 
@@ -174,7 +174,7 @@ SECTIONS
 	   ================== */
 
 	/* Discard sections that difficult post-processing */
-	/DISCARD/ : { *(.group .comment .note .interp .fini_array .fini_array.* .text._ZSt23__cmpexch_failure_orderSt12memory_order) }
+	/DISCARD/ : { *(.group .comment .note .interp .fini_array .fini_array.* .dynsym .dynstr) }
 
 	/* Stabs debugging sections. */
 	.stab          0 : { *(.stab) }

mesosphere/kernel_ldr/source/kern_init_loader.cpp

@@ -207,13 +207,15 @@ namespace ams::kern::init::loader {
         /* NOTE: Nintendo does this only on 10.0.0+ */
         init_pt.PhysicallyRandomize(virtual_base_address + rx_offset, bss_end_offset - rx_offset, true);
 
-        /* Clear kernel .bss. */
-        std::memset(GetVoidPointer(virtual_base_address + bss_offset), 0, bss_end_offset - bss_offset);
-
         /* Apply relocations to the kernel. */
         const Elf::Dyn *kernel_dynamic = reinterpret_cast<const Elf::Dyn *>(GetInteger(virtual_base_address) + dynamic_offset);
         Elf::ApplyRelocations(GetInteger(virtual_base_address), kernel_dynamic);
 
+        /* Clear kernel .bss. */
+        /* NOTE: The kernel does this before applying relocations, but we do it after. */
+        /* This allows us to place our relocations in space overlapping with .bss...and thereby reclaim the memory that would otherwise be wasted. */
+        std::memset(GetVoidPointer(virtual_base_address + bss_offset), 0, bss_end_offset - bss_offset);
+
         /* Call the kernel's init array functions. */
         /* NOTE: The kernel does this after reprotecting .rodata, but we do it before. */
         /* This allows our global constructors to edit .rodata, which is valuable for editing the SVC tables to support older firmwares' ABIs. */