openwrt/target/linux/generic/pending-4.9/305-mips_module_reloc.patch

From: Felix Fietkau <nbd@nbd.name>
Subject: mips: replace -mlong-calls with -mno-long-calls to make function calls faster in kernel modules to achieve this, try to

lede-commit: 3b3d64743ba2a874df9d70cd19e242205b0a788c
Signed-off-by: Felix Fietkau <nbd@nbd.name>
---
 arch/mips/Makefile             |   5 +
 arch/mips/include/asm/module.h |   5 +
 arch/mips/kernel/module.c      | 279 ++++++++++++++++++++++++++++++++++++++++-
 3 files changed, 284 insertions(+), 5 deletions(-)

--- a/arch/mips/Makefile
+++ b/arch/mips/Makefile
@@ -93,8 +93,18 @@ all-$(CONFIG_SYS_SUPPORTS_ZBOOT)+= vmlin
 cflags-y			+= -G 0 -mno-abicalls -fno-pic -pipe -mno-branch-likely
 cflags-y			+= -msoft-float
 LDFLAGS_vmlinux			+= -G 0 -static -n -nostdlib
+ifdef CONFIG_64BIT
 KBUILD_AFLAGS_MODULE		+= -mlong-calls
 KBUILD_CFLAGS_MODULE		+= -mlong-calls
+else
+  ifdef CONFIG_DYNAMIC_FTRACE
+    KBUILD_AFLAGS_MODULE	+= -mlong-calls
+    KBUILD_CFLAGS_MODULE	+= -mlong-calls
+  else
+    KBUILD_AFLAGS_MODULE	+= -mno-long-calls
+    KBUILD_CFLAGS_MODULE	+= -mno-long-calls
+  endif
+endif
 
 ifeq ($(CONFIG_RELOCATABLE),y)
 LDFLAGS_vmlinux			+= --emit-relocs
--- a/arch/mips/include/asm/module.h
+++ b/arch/mips/include/asm/module.h
@@ -11,6 +11,11 @@ struct mod_arch_specific {
 	const struct exception_table_entry *dbe_start;
 	const struct exception_table_entry *dbe_end;
 	struct mips_hi16 *r_mips_hi16_list;
+
+	void *phys_plt_tbl;
+	void *virt_plt_tbl;
+	unsigned int phys_plt_offset;
+	unsigned int virt_plt_offset;
 };
 
 typedef uint8_t Elf64_Byte;		/* Type for a 8-bit quantity.  */
--- a/arch/mips/kernel/module.c
+++ b/arch/mips/kernel/module.c
@@ -44,14 +44,221 @@ struct mips_hi16 {
 static LIST_HEAD(dbe_list);
 static DEFINE_SPINLOCK(dbe_lock);
 
-#ifdef MODULE_START
+/*
+ * Get the potential max trampolines size required of the init and
+ * non-init sections. Only used if we cannot find enough contiguous
+ * physically mapped memory to put the module into.
+ */
+static unsigned int
+get_plt_size(const Elf_Ehdr *hdr, const Elf_Shdr *sechdrs,
+             const char *secstrings, unsigned int symindex, bool is_init)
+{
+	unsigned long ret = 0;
+	unsigned int i, j;
+	Elf_Sym *syms;
+
+	/* Everything marked ALLOC (this includes the exported symbols) */
+	for (i = 1; i < hdr->e_shnum; ++i) {
+		unsigned int info = sechdrs[i].sh_info;
+
+		if (sechdrs[i].sh_type != SHT_REL
+		    && sechdrs[i].sh_type != SHT_RELA)
+			continue;
+
+		/* Not a valid relocation section? */
+		if (info >= hdr->e_shnum)
+			continue;
+
+		/* Don't bother with non-allocated sections */
+		if (!(sechdrs[info].sh_flags & SHF_ALLOC))
+			continue;
+
+		/* If it's called *.init*, and we're not init, we're
+                   not interested */
+		if ((strstr(secstrings + sechdrs[i].sh_name, ".init") != 0)
+		    != is_init)
+			continue;
+
+		syms = (Elf_Sym *) sechdrs[symindex].sh_addr;
+		if (sechdrs[i].sh_type == SHT_REL) {
+			Elf_Mips_Rel *rel = (void *) sechdrs[i].sh_addr;
+			unsigned int size = sechdrs[i].sh_size / sizeof(*rel);
+
+			for (j = 0; j < size; ++j) {
+				Elf_Sym *sym;
+
+				if (ELF_MIPS_R_TYPE(rel[j]) != R_MIPS_26)
+					continue;
+
+				sym = syms + ELF_MIPS_R_SYM(rel[j]);
+				if (!is_init && sym->st_shndx != SHN_UNDEF)
+					continue;
+
+				ret += 4 * sizeof(int);
+			}
+		} else {
+			Elf_Mips_Rela *rela = (void *) sechdrs[i].sh_addr;
+			unsigned int size = sechdrs[i].sh_size / sizeof(*rela);
+
+			for (j = 0; j < size; ++j) {
+				Elf_Sym *sym;
+
+				if (ELF_MIPS_R_TYPE(rela[j]) != R_MIPS_26)
+					continue;
+
+				sym = syms + ELF_MIPS_R_SYM(rela[j]);
+				if (!is_init && sym->st_shndx != SHN_UNDEF)
+					continue;
+
+				ret += 4 * sizeof(int);
+			}
+		}
+	}
+
+	return ret;
+}
+
+#ifndef MODULE_START
+static void *alloc_phys(unsigned long size)
+{
+	unsigned order;
+	struct page *page;
+	struct page *p;
+
+	size = PAGE_ALIGN(size);
+	order = get_order(size);
+
+	page = alloc_pages(GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN |
+			__GFP_THISNODE, order);
+	if (!page)
+		return NULL;
+
+	split_page(page, order);
+
+	/* mark all pages except for the last one */
+	for (p = page; p + 1 < page + (size >> PAGE_SHIFT); ++p)
+		set_bit(PG_owner_priv_1, &p->flags);
+
+	for (p = page + (size >> PAGE_SHIFT); p < page + (1 << order); ++p)
+		__free_page(p);
+
+	return page_address(page);
+}
+#endif
+
+static void free_phys(void *ptr)
+{
+	struct page *page;
+	bool free;
+
+	page = virt_to_page(ptr);
+	do {
+		free = test_and_clear_bit(PG_owner_priv_1, &page->flags);
+		__free_page(page);
+		page++;
+	} while (free);
+}
+
+
 void *module_alloc(unsigned long size)
 {
+#ifdef MODULE_START
 	return __vmalloc_node_range(size, 1, MODULE_START, MODULE_END,
 				GFP_KERNEL, PAGE_KERNEL, 0, NUMA_NO_NODE,
 				__builtin_return_address(0));
+#else
+	void *ptr;
+
+	if (size == 0)
+		return NULL;
+
+	ptr = alloc_phys(size);
+
+	/* If we failed to allocate physically contiguous memory,
+	 * fall back to regular vmalloc. The module loader code will
+	 * create jump tables to handle long jumps */
+	if (!ptr)
+		return vmalloc(size);
+
+	return ptr;
+#endif
 }
+
+static inline bool is_phys_addr(void *ptr)
+{
+#ifdef CONFIG_64BIT
+	return (KSEGX((unsigned long)ptr) == CKSEG0);
+#else
+	return (KSEGX(ptr) == KSEG0);
 #endif
+}
+
+/* Free memory returned from module_alloc */
+void module_memfree(void *module_region)
+{
+	if (is_phys_addr(module_region))
+		free_phys(module_region);
+	else
+		vfree(module_region);
+}
+
+static void *__module_alloc(int size, bool phys)
+{
+	void *ptr;
+
+	if (phys)
+		ptr = kmalloc(size, GFP_KERNEL);
+	else
+		ptr = vmalloc(size);
+	return ptr;
+}
+
+static void __module_free(void *ptr)
+{
+	if (is_phys_addr(ptr))
+		kfree(ptr);
+	else
+		vfree(ptr);
+}
+
+int module_frob_arch_sections(Elf_Ehdr *hdr, Elf_Shdr *sechdrs,
+			      char *secstrings, struct module *mod)
+{
+	unsigned int symindex = 0;
+	unsigned int core_size, init_size;
+	int i;
+
+	mod->arch.phys_plt_offset = 0;
+	mod->arch.virt_plt_offset = 0;
+	mod->arch.phys_plt_tbl = NULL;
+	mod->arch.virt_plt_tbl = NULL;
+
+	if (IS_ENABLED(CONFIG_64BIT))
+		return 0;
+
+	for (i = 1; i < hdr->e_shnum; i++)
+		if (sechdrs[i].sh_type == SHT_SYMTAB)
+			symindex = i;
+
+	core_size = get_plt_size(hdr, sechdrs, secstrings, symindex, false);
+	init_size = get_plt_size(hdr, sechdrs, secstrings, symindex, true);
+
+	if ((core_size + init_size) == 0)
+		return 0;
+
+	mod->arch.phys_plt_tbl = __module_alloc(core_size + init_size, 1);
+	if (!mod->arch.phys_plt_tbl)
+		return -ENOMEM;
+
+	mod->arch.virt_plt_tbl = __module_alloc(core_size + init_size, 0);
+	if (!mod->arch.virt_plt_tbl) {
+		__module_free(mod->arch.phys_plt_tbl);
+		mod->arch.phys_plt_tbl = NULL;
+		return -ENOMEM;
+	}
+
+	return 0;
+}
 
 int apply_r_mips_none(struct module *me, u32 *location, Elf_Addr v)
 {
@@ -65,8 +272,39 @@ static int apply_r_mips_32_rel(struct mo
 	return 0;
 }
 
+static Elf_Addr add_plt_entry_to(unsigned *plt_offset,
+				 void *start, Elf_Addr v)
+{
+	unsigned *tramp = start + *plt_offset;
+	*plt_offset += 4 * sizeof(int);
+
+	/* adjust carry for addiu */
+	if (v & 0x00008000)
+		v += 0x10000;
+
+	tramp[0] = 0x3c190000 | (v >> 16);      /* lui t9, hi16 */
+	tramp[1] = 0x27390000 | (v & 0xffff);   /* addiu t9, t9, lo16 */
+	tramp[2] = 0x03200008;                  /* jr t9 */
+	tramp[3] = 0x00000000;                  /* nop */
+
+	return (Elf_Addr) tramp;
+}
+
+static Elf_Addr add_plt_entry(struct module *me, void *location, Elf_Addr v)
+{
+	if (is_phys_addr(location))
+		return add_plt_entry_to(&me->arch.phys_plt_offset,
+				me->arch.phys_plt_tbl, v);
+	else
+		return add_plt_entry_to(&me->arch.virt_plt_offset,
+				me->arch.virt_plt_tbl, v);
+
+}
+
 static int apply_r_mips_26_rel(struct module *me, u32 *location, Elf_Addr v)
 {
+	u32 ofs = *location & 0x03ffffff;
+
 	if (v % 4) {
 		pr_err("module %s: dangerous R_MIPS_26 REL relocation\n",
 		       me->name);
@@ -74,13 +312,17 @@ static int apply_r_mips_26_rel(struct mo
 	}
 
 	if ((v & 0xf0000000) != (((unsigned long)location + 4) & 0xf0000000)) {
-		pr_err("module %s: relocation overflow\n",
-		       me->name);
-		return -ENOEXEC;
+		v = add_plt_entry(me, location, v + (ofs << 2));
+		if (!v) {
+			pr_err("module %s: relocation overflow\n",
+			       me->name);
+			return -ENOEXEC;
+		}
+		ofs = 0;
 	}
 
 	*location = (*location & ~0x03ffffff) |
-		    ((*location + (v >> 2)) & 0x03ffffff);
+		    ((ofs + (v >> 2)) & 0x03ffffff);
 
 	return 0;
 }
@@ -349,9 +591,36 @@ int module_finalize(const Elf_Ehdr *hdr,
 		list_add(&me->arch.dbe_list, &dbe_list);
 		spin_unlock_irq(&dbe_lock);
 	}
+
+	/* Get rid of the fixup trampoline if we're running the module
+	 * from physically mapped address space */
+	if (me->arch.phys_plt_offset == 0) {
+		__module_free(me->arch.phys_plt_tbl);
+		me->arch.phys_plt_tbl = NULL;
+	}
+	if (me->arch.virt_plt_offset == 0) {
+		__module_free(me->arch.virt_plt_tbl);
+		me->arch.virt_plt_tbl = NULL;
+	}
+
 	return 0;
 }
 
+void module_arch_freeing_init(struct module *mod)
+{
+	if (mod->state == MODULE_STATE_LIVE)
+		return;
+
+	if (mod->arch.phys_plt_tbl) {
+		__module_free(mod->arch.phys_plt_tbl);
+		mod->arch.phys_plt_tbl = NULL;
+	}
+	if (mod->arch.virt_plt_tbl) {
+		__module_free(mod->arch.virt_plt_tbl);
+		mod->arch.virt_plt_tbl = NULL;
+	}
+}
+
 void module_arch_cleanup(struct module *mod)
 {
 	spin_lock_irq(&dbe_lock);