x1000: bootloader: fix Linux self-extracting kernel boot

Basically, there's longstanding bug in Linux with self-extracting
kernels on MIPS which just happened to manifest now on the M3K as
a hang on boot. The fix is applied to the M3K and Q1 since they
both use this type of kernel image.

Change-Id: I17d2bad6eebd677cd6d2e0bf146450c71fcf1229

firmware/linuxboot.c

@@ -216,3 +216,99 @@ ssize_t uimage_fd_reader(void* buf, size_t size, void* ctx)
     int fd = (intptr_t)ctx;
     return read(fd, buf, size);
 }
+
+/* Linux's self-extracting kernels are broken on MIPS. The decompressor stub
+ * doesn't flush caches after extracting the kernel code which can cause the
+ * boot to fail horribly. This has been true since at least 2009 and at the
+ * time of writing (2022) it's *still* broken.
+ *
+ * The FiiO M3K and Shanling Q1 both have broken kernels of this type, so we
+ * work around this by replacing the direct call to the kernel entry point with
+ * a thunk that adds the necessary cache flush.
+ */
+uint32_t mips_linux_stub_get_entry(void** code_start, size_t code_size)
+{
+    /* The jump to the kernel entry point looks like this:
+     *
+     *   move a0, s0
+     *   move a1, s1
+     *   move a2, s2
+     *   move a3, s3
+     *   ...
+     *   la   k0, KERNEL_ENTRY
+     *   jr   k0
+     *   --- or in kernels since 2021: ---
+     *   la   t9, KERNEL_ENTRY
+     *   jalr t9
+     *
+     * We're trying to identify this code and decode the kernel entry
+     * point address, and return a suitable address where we can patch
+     * in a call to our thunk.
+     */
+
+    /* We should only need to scan within the first 128 bytes
+     * but do up to 256 just in case. */
+    uint32_t* start = *code_start;
+    uint32_t* end = start + (MIN(code_size, 256) + 3) / 4;
+
+    /* Scan for the "move aN, sN" sequence */
+    uint32_t* move_instr = start;
+    for(move_instr += 4; move_instr < end; ++move_instr) {
+        if(move_instr[-4] == 0x02002021 &&  /* move a0, s0 */
+           move_instr[-3] == 0x02202821 &&  /* move a1, s1 */
+           move_instr[-2] == 0x02403021 &&  /* move a2, s2 */
+           move_instr[-1] == 0x02603821)    /* move a3, s3 */
+            break;
+    }
+
+    if(move_instr == end)
+        return 0;
+
+    /* Now search forward for the next jr/jalr instruction */
+    int jreg = 0;
+    uint32_t* jump_instr = move_instr;
+    for(; jump_instr != end; ++jump_instr) {
+        if((jump_instr[0] & 0xfc1ff83f) == 0xf809 ||
+           (jump_instr[0] & 0xfc00003f) == 0x8) {
+            /* jalr rN */
+            jreg = (jump_instr[0] >> 21) & 0x1f;
+            break;
+        }
+    }
+
+    /* Need room here for 4 instructions. Assume everything between the
+     * moves and the jump is safe to overwrite; otherwise, we'll need to
+     * take a different approach.
+     *
+     * Count +1 instruction for the branch delay slot and another +1 because
+     * "move_instr" points to the instruction following the last move. */
+    if(jump_instr - move_instr + 2 < 4)
+        return 0;
+    if(!jreg)
+        return 0;
+
+    /* Now scan from the end of the move sequence until the jump instruction
+     * and try to reconstruct the entry address. We check for lui/ori/addiu. */
+    const uint32_t lui_mask   = 0xffff0000;
+    const uint32_t lui        = 0x3c000000 | (jreg << 16);
+    const uint32_t ori_mask   = 0xffff0000;
+    const uint32_t ori        = 0x34000000 | (jreg << 21) | (jreg << 16);
+    const uint32_t addiu_mask = 0xffff0000;
+    const uint32_t addiu      = 0x24000000 | (jreg << 21) | (jreg << 16);
+
+    /* Can use any initial value here */
+    uint32_t jreg_val = 0xdeadbeef;
+
+    for(uint32_t* instr = move_instr; instr != jump_instr; ++instr) {
+        if((instr[0] & lui_mask) == lui)
+            jreg_val = (instr[0] & 0xffff) << 16;
+        else if((instr[0] & ori_mask) == ori)
+            jreg_val |= instr[0] & 0xffff;
+        else if((instr[0] & addiu_mask) == addiu)
+            jreg_val += instr[0] & 0xffff;
+    }
+
+    /* Success! Probably! */
+    *code_start = move_instr;
+    return jreg_val;
+}