Re: [PATCH v1] arm64/module: Optimize module load time by optimizing PLT counting

[Will Deacon]

Hi all,

On Wed, Jun 17, 2020 at 10:17:33AM +0200, Ard Biesheuvel wrote:
> On Tue, 16 Jun 2020 at 23:40, Will Deacon <will@xxxxxxxxxx> wrote:
> > On Fri, Jun 05, 2020 at 03:22:57PM -0700, Saravana Kannan wrote:
> > > This gives significant reduction in module load time for modules with
> > > large number of relocations with no measurable impact on modules with a
> > > small number of relocations. In my test setup with CONFIG_RANDOMIZE_BASE
> > > enabled, the load time for one module went down from 268ms to 100ms.
> > > Another module went down from 143ms to 83ms.
> >
> > Whilst I can see that's a significant relative saving, what proportion of
> > actual boot time are we talking about here? It would be interesting to
> > know if there are bigger potential savings elsewhere.
> >
> 
> Also, 'some module' vs 'some other module' doesn't really say
> anything. Please explain which modules and their sizes.

I suspect these are all out-of-tree modules, but yes, some metadata such as
sizes, nr or relocs etc would be good to have in the commit message.

> > > diff --git a/arch/arm64/kernel/module-plts.c b/arch/arm64/kernel/module-plts.c
> > > index 65b08a74aec6..bf5118b3b828 100644
> > > --- a/arch/arm64/kernel/module-plts.c
> > > +++ b/arch/arm64/kernel/module-plts.c
> > > @@ -253,6 +253,36 @@ static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num,
> > >       return ret;
> > >  }
> > >
> > > +static bool rela_needs_dedup(Elf64_Rela *rela)
> > > +{
> > > +     return ELF64_R_TYPE(rela->r_info) == R_AARCH64_JUMP26
> > > +            || ELF64_R_TYPE(rela->r_info) == R_AARCH64_CALL26;
> > > +}
> >
> 
> Would it help to check the section index here as well? Call/jump
> instructions within a section are never sent through a PLT entry.

(I tried hacking this in below)

> > Does this handle A53 erratum 843419 correctly? I'm worried that we skip
> > the ADRP PLTs there.
> >
> 
> ADRP PLTs cannot be deduplicated, as they incorporate a relative jump
> back to the caller.

Duh yes, thanks. We can't trash the link register here.

> > > +/* Group the CALL26/JUMP26 relas toward the beginning of the array. */
> > > +static int partition_dedup_relas(Elf64_Rela *rela, int numrels)
> > > +{
> > > +     int i = 0, j = numrels - 1;
> > > +     Elf64_Rela t;
> > > +
> > > +     while (i < j) {
> > > +             while (rela_needs_dedup(rela + i) && i < j)
> > > +                     i++;
> > > +             while (!rela_needs_dedup(rela + j) && i < j)
> > > +                     j--;
> > > +             if (i < j) {
> > > +                     t = *(rela + j);
> > > +                     *(rela + j) = *(rela + i);
> > > +                     *(rela + i) = t;
> > > +             }
> > > +     }
> >
> > This is very hard to read and I think some of the 'i < j' comparisons are
> > redundant. Would it make more sense to assign a temporary rather than
> > post-inc/decrement and recheck?
> >
> 
> Agreed.
> 
> Also, what's wrong with [] array indexing?

Saravana, since our stylistic objections are reasonably vague, I tried
to clean this up so you can get an idea of how I'd prefer it to look (can't
speak for Ard). I haven't tried running this, but please feel free to adapt
it. Replacement diff below.

Will

--->8

diff --git a/arch/arm64/kernel/module-plts.c b/arch/arm64/kernel/module-plts.c
index 65b08a74aec6..204290314054 100644
--- a/arch/arm64/kernel/module-plts.c
+++ b/arch/arm64/kernel/module-plts.c
@@ -253,6 +253,39 @@ static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num,
 	return ret;
 }
 
+static bool branch_rela_needs_plt(Elf64_Sym *syms, Elf64_Rela *rela,
+				  Elf64_Word dstidx)
+{
+
+	Elf64_Sym *s = syms + ELF64_R_SYM(rela->r_info);
+
+	if (s->st_shndx == dstidx)
+		return false;
+
+	return ELF64_R_TYPE(rela->r_info) == R_AARCH64_JUMP26 ||
+	       ELF64_R_TYPE(rela->r_info) == R_AARCH64_CALL26;
+}
+
+static int partition_branch_plt_relas(Elf64_Sym *syms, Elf64_Rela *rela,
+				      int numrels, Elf64_Word dstidx)
+{
+	int i = 0, j = numrels - 1;
+
+	if (!IS_ENABLED(CONFIG_RANDOMIZE_BASE))
+		return 0;
+
+	while (i < j) {
+		if (branch_rela_needs_plt(syms, &rela[i], dstidx))
+			i++;
+		else if (branch_rela_needs_plt(syms, &rela[j], dstidx))
+			swap(rela[i], rela[j]);
+		else
+			j--;
+	}
+
+	return i;
+}
+
 int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
 			      char *secstrings, struct module *mod)
 {
@@ -290,7 +323,7 @@ int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
 
 	for (i = 0; i < ehdr->e_shnum; i++) {
 		Elf64_Rela *rels = (void *)ehdr + sechdrs[i].sh_offset;
-		int numrels = sechdrs[i].sh_size / sizeof(Elf64_Rela);
+		int nents, numrels = sechdrs[i].sh_size / sizeof(Elf64_Rela);
 		Elf64_Shdr *dstsec = sechdrs + sechdrs[i].sh_info;
 
 		if (sechdrs[i].sh_type != SHT_RELA)
@@ -300,8 +333,14 @@ int module_frob_arch_sections(Elf_Ehdr *ehdr, Elf_Shdr *sechdrs,
 		if (!(dstsec->sh_flags & SHF_EXECINSTR))
 			continue;
 
-		/* sort by type, symbol index and addend */
-		sort(rels, numrels, sizeof(Elf64_Rela), cmp_rela, NULL);
+		/*
+		 * sort branch relocations requiring a PLT by type, symbol index
+		 * and addend
+		 */
+		nents = partition_branch_plt_relas(syms, rels, numrels,
+						   sechdrs[i].sh_info);
+		if (nents)
+			sort(rels, nents, sizeof(Elf64_Rela), cmp_rela, NULL);
 
 		if (!str_has_prefix(secstrings + dstsec->sh_name, ".init"))
 			core_plts += count_plts(syms, rels, numrels,