Merge tag 'arm-plt-optimizations-for-v4.9' of...

Merge tag 'arm-plt-optimizations-for-v4.9' of git://git.linaro.org/people/ard.biesheuvel/linux-arm into devel-stable

This series of 4 patches optimizes the ARM PLT generation code that
is invoked at module load time, to get rid of the O(n^2) algorithm
that results in pathological load times of 10 seconds or more for
large modules on certain STB platforms

arch/arm/include/asm/module.h

@@ -23,10 +23,8 @@ struct mod_arch_specific {
 	struct unwind_table *unwind[ARM_SEC_MAX];
 #endif
 #ifdef CONFIG_ARM_MODULE_PLTS
-	struct elf32_shdr   *core_plt;
-	struct elf32_shdr   *init_plt;
-	int		    core_plt_count;
-	int		    init_plt_count;
+	struct elf32_shdr   *plt;
+	int		    plt_count;
 #endif
 };
 

arch/arm/kernel/module-plts.c

@@ -9,6 +9,7 @@
 #include <linux/elf.h>
 #include <linux/kernel.h>
 #include <linux/module.h>
+#include <linux/sort.h>
 
 #include <asm/cache.h>
 #include <asm/opcodes.h>
@@ -30,154 +31,198 @@ struct plt_entries {
 	u32	lit[PLT_ENT_COUNT];
 };
 
-static bool in_init(const struct module *mod, u32 addr)
+u32 get_module_plt(struct module *mod, unsigned long loc, Elf32_Addr val)
 {
-	return addr - (u32)mod->init_layout.base < mod->init_layout.size;
+	struct plt_entries *plt = (struct plt_entries *)mod->arch.plt->sh_addr;
+	int idx = 0;
+
+	/*
+	 * Look for an existing entry pointing to 'val'. Given that the
+	 * relocations are sorted, this will be the last entry we allocated.
+	 * (if one exists).
+	 */
+	if (mod->arch.plt_count > 0) {
+		plt += (mod->arch.plt_count - 1) / PLT_ENT_COUNT;
+		idx = (mod->arch.plt_count - 1) % PLT_ENT_COUNT;
+
+		if (plt->lit[idx] == val)
+			return (u32)&plt->ldr[idx];
+
+		idx = (idx + 1) % PLT_ENT_COUNT;
+		if (!idx)
+			plt++;
+	}
+
+	mod->arch.plt_count++;
+	BUG_ON(mod->arch.plt_count * PLT_ENT_SIZE > mod->arch.plt->sh_size);
+
+	if (!idx)
+		/* Populate a new set of entries */
+		*plt = (struct plt_entries){
+			{ [0 ... PLT_ENT_COUNT - 1] = PLT_ENT_LDR, },
+			{ val, }
+		};
+	else
+		plt->lit[idx] = val;
+
+	return (u32)&plt->ldr[idx];
 }
 
-u32 get_module_plt(struct module *mod, unsigned long loc, Elf32_Addr val)
+#define cmp_3way(a,b)	((a) < (b) ? -1 : (a) > (b))
+
+static int cmp_rel(const void *a, const void *b)
 {
-	struct plt_entries *plt, *plt_end;
-	int c, *count;
-
-	if (in_init(mod, loc)) {
-		plt = (void *)mod->arch.init_plt->sh_addr;
-		plt_end = (void *)plt + mod->arch.init_plt->sh_size;
-		count = &mod->arch.init_plt_count;
-	} else {
-		plt = (void *)mod->arch.core_plt->sh_addr;
-		plt_end = (void *)plt + mod->arch.core_plt->sh_size;
-		count = &mod->arch.core_plt_count;
-	}
+	const Elf32_Rel *x = a, *y = b;
+	int i;
 
-	/* Look for an existing entry pointing to 'val' */
-	for (c = *count; plt < plt_end; c -= PLT_ENT_COUNT, plt++) {
-		int i;
-
-		if (!c) {
-			/* Populate a new set of entries */
-			*plt = (struct plt_entries){
-				{ [0 ... PLT_ENT_COUNT - 1] = PLT_ENT_LDR, },
-				{ val, }
-			};
-			++*count;
-			return (u32)plt->ldr;
-		}
-		for (i = 0; i < PLT_ENT_COUNT; i++) {
-			if (!plt->lit[i]) {
-				plt->lit[i] = val;
-				++*count;
-			}
-			if (plt->lit[i] == val)
-				return (u32)&plt->ldr[i];
-		}
+	/* sort by type and symbol index */
+	i = cmp_3way(ELF32_R_TYPE(x->r_info), ELF32_R_TYPE(y->r_info));
+	if (i == 0)
+		i = cmp_3way(ELF32_R_SYM(x->r_info), ELF32_R_SYM(y->r_info));
+	return i;
+}
+
+static bool is_zero_addend_relocation(Elf32_Addr base, const Elf32_Rel *rel)
+{
+	u32 *tval = (u32 *)(base + rel->r_offset);
+
+	/*
+	 * Do a bitwise compare on the raw addend rather than fully decoding
+	 * the offset and doing an arithmetic comparison.
+	 * Note that a zero-addend jump/call relocation is encoded taking the
+	 * PC bias into account, i.e., -8 for ARM and -4 for Thumb2.
+	 */
+	switch (ELF32_R_TYPE(rel->r_info)) {
+		u16 upper, lower;
+
+	case R_ARM_THM_CALL:
+	case R_ARM_THM_JUMP24:
+		upper = __mem_to_opcode_thumb16(((u16 *)tval)[0]);
+		lower = __mem_to_opcode_thumb16(((u16 *)tval)[1]);
+
+		return (upper & 0x7ff) == 0x7ff && (lower & 0x2fff) == 0x2ffe;
+
+	case R_ARM_CALL:
+	case R_ARM_PC24:
+	case R_ARM_JUMP24:
+		return (__mem_to_opcode_arm(*tval) & 0xffffff) == 0xfffffe;
 	}
 	BUG();
 }
 
-static int duplicate_rel(Elf32_Addr base, const Elf32_Rel *rel, int num,
-			   u32 mask)
+static bool duplicate_rel(Elf32_Addr base, const Elf32_Rel *rel, int num)
 {
-	u32 *loc1, *loc2;
-	int i;
+	const Elf32_Rel *prev;
 
-	for (i = 0; i < num; i++) {
-		if (rel[i].r_info != rel[num].r_info)
-			continue;
+	/*
+	 * Entries are sorted by type and symbol index. That means that,
+	 * if a duplicate entry exists, it must be in the preceding
+	 * slot.
+	 */
+	if (!num)
+		return false;
 
-		/*
-		 * Identical relocation types against identical symbols can
-		 * still result in different PLT entries if the addend in the
-		 * place is different. So resolve the target of the relocation
-		 * to compare the values.
-		 */
-		loc1 = (u32 *)(base + rel[i].r_offset);
-		loc2 = (u32 *)(base + rel[num].r_offset);
-		if (((*loc1 ^ *loc2) & mask) == 0)
-			return 1;
-	}
-	return 0;
+	prev = rel + num - 1;
+	return cmp_rel(rel + num, prev) == 0 &&
+	       is_zero_addend_relocation(base, prev);
 }
 
 /* Count how many PLT entries we may need */
-static unsigned int count_plts(Elf32_Addr base, const Elf32_Rel *rel, int num)
+static unsigned int count_plts(const Elf32_Sym *syms, Elf32_Addr base,
+			       const Elf32_Rel *rel, int num)
 {
 	unsigned int ret = 0;
+	const Elf32_Sym *s;
 	int i;
 
-	/*
-	 * Sure, this is order(n^2), but it's usually short, and not
-	 * time critical
-	 */
-	for (i = 0; i < num; i++)
+	for (i = 0; i < num; i++) {
 		switch (ELF32_R_TYPE(rel[i].r_info)) {
 		case R_ARM_CALL:
 		case R_ARM_PC24:
 		case R_ARM_JUMP24:
-			if (!duplicate_rel(base, rel, i,
-					   __opcode_to_mem_arm(0x00ffffff)))
-				ret++;
-			break;
-#ifdef CONFIG_THUMB2_KERNEL
 		case R_ARM_THM_CALL:
 		case R_ARM_THM_JUMP24:
-			if (!duplicate_rel(base, rel, i,
-					   __opcode_to_mem_thumb32(0x07ff2fff)))
+			/*
+			 * We only have to consider branch targets that resolve
+			 * to undefined symbols. This is not simply a heuristic,
+			 * it is a fundamental limitation, since the PLT itself
+			 * is part of the module, and needs to be within range
+			 * as well, so modules can never grow beyond that limit.
+			 */
+			s = syms + ELF32_R_SYM(rel[i].r_info);
+			if (s->st_shndx != SHN_UNDEF)
+				break;
+
+			/*
+			 * Jump relocations with non-zero addends against
+			 * undefined symbols are supported by the ELF spec, but
+			 * do not occur in practice (e.g., 'jump n bytes past
+			 * the entry point of undefined function symbol f').
+			 * So we need to support them, but there is no need to
+			 * take them into consideration when trying to optimize
+			 * this code. So let's only check for duplicates when
+			 * the addend is zero.
+			 */
+			if (!is_zero_addend_relocation(base, rel + i) ||
+			    !duplicate_rel(base, rel, i))
 				ret++;
-#endif
 		}
+	}
 	return ret;
 }
 
 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
 			      char *secstrings, struct module *mod)
 {
-	unsigned long core_plts = 0, init_plts = 0;
+	unsigned long plts = 0;
 	Elf32_Shdr *s, *sechdrs_end = sechdrs + ehdr->e_shnum;
+	Elf32_Sym *syms = NULL;
 
 	/*
 	 * To store the PLTs, we expand the .text section for core module code
-	 * and the .init.text section for initialization code.
+	 * and for initialization code.
 	 */
-	for (s = sechdrs; s < sechdrs_end; ++s)
-		if (strcmp(".core.plt", secstrings + s->sh_name) == 0)
-			mod->arch.core_plt = s;
-		else if (strcmp(".init.plt", secstrings + s->sh_name) == 0)
-			mod->arch.init_plt = s;
-
-	if (!mod->arch.core_plt || !mod->arch.init_plt) {
-		pr_err("%s: sections missing\n", mod->name);
+	for (s = sechdrs; s < sechdrs_end; ++s) {
+		if (strcmp(".plt", secstrings + s->sh_name) == 0)
+			mod->arch.plt = s;
+		else if (s->sh_type == SHT_SYMTAB)
+			syms = (Elf32_Sym *)s->sh_addr;
+	}
+
+	if (!mod->arch.plt) {
+		pr_err("%s: module PLT section missing\n", mod->name);
+		return -ENOEXEC;
+	}
+	if (!syms) {
+		pr_err("%s: module symtab section missing\n", mod->name);
 		return -ENOEXEC;
 	}
 
 	for (s = sechdrs + 1; s < sechdrs_end; ++s) {
-		const Elf32_Rel *rels = (void *)ehdr + s->sh_offset;
+		Elf32_Rel *rels = (void *)ehdr + s->sh_offset;
 		int numrels = s->sh_size / sizeof(Elf32_Rel);
 		Elf32_Shdr *dstsec = sechdrs + s->sh_info;
 
 		if (s->sh_type != SHT_REL)
 			continue;
 
-		if (strstr(secstrings + s->sh_name, ".init"))
-			init_plts += count_plts(dstsec->sh_addr, rels, numrels);
-		else
-			core_plts += count_plts(dstsec->sh_addr, rels, numrels);
+		/* ignore relocations that operate on non-exec sections */
+		if (!(dstsec->sh_flags & SHF_EXECINSTR))
+			continue;
+
+		/* sort by type and symbol index */
+		sort(rels, numrels, sizeof(Elf32_Rel), cmp_rel, NULL);
+
+		plts += count_plts(syms, dstsec->sh_addr, rels, numrels);
 	}
 
-	mod->arch.core_plt->sh_type = SHT_NOBITS;
-	mod->arch.core_plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
-	mod->arch.core_plt->sh_addralign = L1_CACHE_BYTES;
-	mod->arch.core_plt->sh_size = round_up(core_plts * PLT_ENT_SIZE,
-					       sizeof(struct plt_entries));
-	mod->arch.core_plt_count = 0;
-
-	mod->arch.init_plt->sh_type = SHT_NOBITS;
-	mod->arch.init_plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
-	mod->arch.init_plt->sh_addralign = L1_CACHE_BYTES;
-	mod->arch.init_plt->sh_size = round_up(init_plts * PLT_ENT_SIZE,
-					       sizeof(struct plt_entries));
-	mod->arch.init_plt_count = 0;
-	pr_debug("%s: core.plt=%x, init.plt=%x\n", __func__,
-		 mod->arch.core_plt->sh_size, mod->arch.init_plt->sh_size);
+	mod->arch.plt->sh_type = SHT_NOBITS;
+	mod->arch.plt->sh_flags = SHF_EXECINSTR | SHF_ALLOC;
+	mod->arch.plt->sh_addralign = L1_CACHE_BYTES;
+	mod->arch.plt->sh_size = round_up(plts * PLT_ENT_SIZE,
+					  sizeof(struct plt_entries));
+	mod->arch.plt_count = 0;
+
+	pr_debug("%s: plt=%x\n", __func__, mod->arch.plt->sh_size);
 	return 0;
 }

arch/arm/kernel/module.lds

 SECTIONS {
-        .core.plt : { BYTE(0) }
-        .init.plt : { BYTE(0) }
+	.plt : { BYTE(0) }
 }