64-bit Power PC linux uses 64KiB page size, but qemu-ppc64{,le} uses 4KiB.

modified:   stub/src/amd64-linux.elf-main.c
	modified:   stub/src/include/linux.h
	modified:   stub/src/powerpc64le-linux.elf-fold.S

src/stub/src/amd64-linux.elf-main.c

@@ -144,6 +144,8 @@ unpackExtent(
     f_unfilter *f_unf
 )
 {
+    DPRINTF("unpackExtent xi=(%%p %%p)  xo=(%%p %%p)  f_exp=%%p  f_unf=%%p\\n",
+        xi->size, xi->buf, xo->size, xo->buf, f_exp, f_unf);
     while (xo->size) {
         struct b_info h;
         //   Note: if h.sz_unc == h.sz_cpr then the block was not
@@ -235,6 +237,7 @@ make_hatch_ppc64(
 static void
 upx_bzero(char *p, size_t len)
 {
+    DPRINTF("bzero %%x  %%x\\n", p, len);
     if (len) do {
         *p++= 0;
     } while (--len);
@@ -279,6 +282,9 @@ auxv_up(Elf64_auxv_t *av, unsigned const type, uint64_t const value)
 static Elf64_Addr // returns relocation constant
 xfind_pages(unsigned mflags, Elf64_Phdr const *phdr, int phnum,
     char **const p_brk
+#if defined(__powerpc64__)
+    , size_t const PAGE_MASK
+#endif
 )
 {
     size_t lo= ~0, hi= 0, szlo= 0;
@@ -312,14 +318,21 @@ do_xmap(
     f_expand *const f_exp,
     f_unfilter *const f_unf,
     Elf64_Addr *p_reloc
+#if defined(__powerpc64__)
+    , size_t const PAGE_MASK
+#endif
 )
 {
     Elf64_Phdr const *phdr = (Elf64_Phdr const *)(void const *)(ehdr->e_phoff +
         (char const *)ehdr);
     char *v_brk;
     Elf64_Addr const reloc = xfind_pages(
-        ((ET_DYN!=ehdr->e_type) ? MAP_FIXED : 0), phdr, ehdr->e_phnum, &v_brk);
+        ((ET_DYN!=ehdr->e_type) ? MAP_FIXED : 0), phdr, ehdr->e_phnum, &v_brk
-    DPRINTF("do_xmap reloc=%%p", reloc);
+#if defined(__powerpc64__)
+        , PAGE_MASK
+#endif
+    );
+    DPRINTF("do_xmap reloc=%%p\\n", reloc);
     int j;
     for (j=0; j < ehdr->e_phnum; ++phdr, ++j)
     if (xi && PT_PHDR==phdr->p_type) {
@@ -410,11 +423,12 @@ upx_main(  // returns entry address
     Elf64_Ehdr *const ehdr,  // temp char[sz_ehdr] for decompressing
     Elf64_auxv_t *const av,
     f_expand *const f_exp,
-    f_unfilter *const f_unf,
+    f_unfilter *const f_unf
 #if defined(__x86_64)  //{
-    Elf64_Addr reloc  // In: &Elf64_Ehdr for stub
+    , Elf64_Addr reloc  // In: &Elf64_Ehdr for stub
 #elif defined(__powerpc64__)  //}{
-    Elf64_Addr *p_reloc  // In: &Elf64_Ehdr for stub; Out: 'slide' for PT_INTERP
+    , Elf64_Addr *p_reloc  // In: &Elf64_Ehdr for stub; Out: 'slide' for PT_INTERP
+    , size_t const PAGE_MASK
 #endif  //}
 )
 {
@@ -447,7 +461,11 @@ upx_main(  // returns entry address
     }
 
     // De-compress Ehdr again into actual position, then de-compress the rest.
-    Elf64_Addr entry = do_xmap(ehdr, &xi1, 0, av, f_exp, f_unf, p_reloc);
+    Elf64_Addr entry = do_xmap(ehdr, &xi1, 0, av, f_exp, f_unf, p_reloc
+#if defined(__powerpc64__)
+       , PAGE_MASK
+#endif
+    );
     DPRINTF("upx_main2  entry=%%p  *p_reloc=%%p\\n", entry, *p_reloc);
     auxv_up(av, AT_ENTRY , entry);
 
@@ -466,7 +484,11 @@ ERR_LAB
         }
         // We expect PT_INTERP to be ET_DYN at 0.
         // Thus do_xmap will set *p_reloc = slide.
-        entry = do_xmap(ehdr, 0, fdi, 0, 0, 0, p_reloc);
+        entry = do_xmap(ehdr, 0, fdi, 0, 0, 0, p_reloc
+#if defined(__powerpc64__)
+            , PAGE_MASK
+#endif
+        );
         auxv_up(av, AT_BASE, *p_reloc);  // musl
         close(fdi);
     }

src/stub/src/include/linux.h

@@ -90,16 +90,21 @@ struct timespec {
 
 // misc constants
 
-#if defined(__mips__) || defined(__powerpc64__)
+#if defined(__powerpc64__)  //{
+// __powerpc64__ hardware uses 64KiB in 64-bit mode, 4KiB in 32-bit mode.
+// EXCEPT qemu-ppc64 uses 4KiB in 64-bit mode, too.
+// So PAGE_SIZE is not constant in 64-bit mode of __powerpc64__.
+    // empty: not constant, must use run-time value from AT_PAGESZ
+#elif defined(__mips__)  //}{
 #define PAGE_MASK       (~0ul<<16)   // discards the offset, keeps the page
 #define PAGE_SIZE       ( 1ul<<16)
-#elif defined(__amd64__)
+#elif defined(__amd64__)  //}{
 #define PAGE_MASK       (~0ul<<12)   // discards the offset, keeps the page
 #define PAGE_SIZE       ( 1ul<<12)
-#elif defined(__i386__) || defined(__powerpc__) || defined(__arm__)  || defined(__AARCH64EL__)
+#elif defined(__i386__) || defined(__powerpc__) || defined(__arm__)  || defined(__AARCH64EL__)  //}{
 #define PAGE_MASK       (~0ul<<12)   // discards the offset, keeps the page
 #define PAGE_SIZE       ( 1ul<<12)
-#endif
+#endif  //}
 
 #define SEEK_SET        0
 #define SEEK_CUR        1

src/stub/src/powerpc64le-linux.elf-fold.S

@@ -175,13 +175,15 @@ r_reloc= 32 - 1  # used slot in register save area
         mr  a3,r_auxv  // &Elf64_auxv_t
         mr  a4,r_exp  // &decompress: f_expand
         mr  a5,r_unf  // &f_unf
+        mr  a7,r_PMASK  // page_mask
 #if BIG_ENDIAN  //{  setup indirection for entry to f_exp and f_unf
         std a4,3*NBPW(sp); la a4,3*NBPW(sp)  // use tmp.xlc
         std a5,4*NBPW(sp); la a5,4*NBPW(sp)  // use tmp.ld
 #endif  //}
         call upx_main  // Out: a0= entry
 // entry= upx_main(b_info *a0, total_size a1, Elf_Ehdr *a2, Elf64_auxv_t *a3,
-//      f_exp a4, f_unf a5, p_reloc a6)
+//      f_exp a4, f_unf a5, p_reloc a6, page_mask a7)
+        teq r0,r0
 #if BIG_ENDIAN  //{
         ld r0,SZ_FRAME+OVERHEAD +SZ_FRAME+r_reloc*NBPW(sp)  // reloc for ET_DYN
         ld r2,NBPW(r3); add r2,r2,r0  // toc