arm64/module: Optimize module load time by optimizing PLT counting
PATCH v2 1/1 arm64/module: Optimize module load time by optimizing PLT counting
When loading a module, module_frob_arch_sections() tries to figure out the number of PLTs that'll be needed to handle all the RELAs. While doing this, it tries to dedupe PLT allocations for multiple R_AARCH64_CALL26 relocations to the same symbol. It does the same for R_AARCH64_JUMP26 relocations.
To make checks for duplicates easier/faster, it sorts the relocation list by type, symbol and addend. That way, to check for a duplicate relocation, it just needs to compare with the previous entry.
However, sorting the entire relocation array is unnecessary and expensive (O(n log n)) because there are a lot of other relocation types that don't need deduping or can't be deduped.
So this commit partitions the array into entries that need deduping and those that don't. And then sorts just the part that needs deduping. And when CONFIG_RANDOMIZE_BASE is disabled, the sorting is skipped entirely because PLTs are not allocated for R_AARCH64_CALL26 and R_AARCH64_JUMP26 if it's disabled.
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, these were the results for a few downstream modules:
Module Size (MB)
wlan 14
video codec 3.8
drm 1.8
IPA 2.5
audio 1.2
gpu 1.8
Without this patch:
Module Number of entries sorted Module load time (ms)
wlan 243739 283
video codec 74029 138
drm 53837 67
IPA 42800 90
audio 21326 27
gpu 20967 32
Total time to load all these module: 637 ms
With this patch:
Module Number of entries sorted Module load time (ms)
wlan 22454 61
video codec 10150 47
drm 13014 40
IPA 8097 63
audio 4606 16
gpu 6527 20
Total time to load all these modules: 247
Time saved during boot for just these 6 modules: 390 ms
---
v1 -> v2:
- Provided more details in the commit text
- Pulled in Will's comments on the coding style
- Pulled in Ard's suggestion about skipping jumps with the same section index (parts of Will's suggested code)
arch/arm64/kernel/module-plts.c | 46 ++++++++++++++++++++++++++++++--- 1 file changed, 43 insertions(+), 3 deletions(-)
arch/arm64/kernel/module-plts.c +43 -3
@@ -253,6 +253,40 @@ 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 ; +} + +/* Group branch PLT relas at the front end of the array. */ +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 +324,7 @@ 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 +334,14 @@ 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,
Signature · 1 line
2.27.0.111.gc72c7da667-goog
Quoted 61 lines · Saravana Kannan
When loading a module,
module_frob_arch_sections()tries to figure out the number of PLTs that'll be needed to handle all the RELAs. While doing this, it tries to dedupe PLT allocations for multipleR_AARCH64_CALL26relocations to the same symbol. It does the same forR_AARCH64_JUMP26relocations.To make checks for duplicates easier/faster, it sorts the relocation list by type, symbol and addend. That way, to check for a duplicate relocation, it just needs to compare with the previous entry.
However, sorting the entire relocation array is unnecessary and expensive (O(n log n)) because there are a lot of other relocation types that don't need deduping or can't be deduped.
So this commit partitions the array into entries that need deduping and those that don't. And then sorts just the part that needs deduping. And when
CONFIG_RANDOMIZE_BASEis disabled, the sorting is skipped entirely because PLTs are not allocated forR_AARCH64_CALL26andR_AARCH64_JUMP26if it's disabled.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_BASEenabled, these were the results for a few downstream modules:Module Size (MB)
wlan 14
video codec 3.8
drm 1.8
IPA 2.5
audio 1.2
gpu 1.8Without this patch:
Module Number of entries sorted Module load time (ms)
wlan 243739 283
video codec 74029 138
drm 53837 67
IPA 42800 90
audio 21326 27
gpu 20967 32Total time to load all these module: 637 ms
With this patch:
Module Number of entries sorted Module load time (ms)
wlan 22454 61
video codec 10150 47
drm 13014 40
IPA 8097 63
audio 4606 16
gpu 6527 20Total time to load all these modules: 247
Time saved during boot for just these 6 modules: 390 ms
Cc Ard Biesheuvel <[email protected]>Signed-off-by Saravana Kannan <[email protected]>---
Quoted 8 lines
v1 -> v2:
- Provided more details in the commit text
- Pulled in Will's comments on the coding style
- Pulled in Ard's suggestion about skipping jumps with the same section index (parts of Will's suggested code)arch/arm64/kernel/module-plts.c | 46 ++++++++++++++++++++++++++++++--- 1 file changed, 43 insertions(+), 3 deletions(-)
Nice, it looks like you were more-or-less able to use my suggestion directly! Commit message looks much better to, so:
Catalin can pick this up when he starts queuing patches for 5.9.
Cheers,
Will
Quoted 11 lines · Saravana Kannan
When loading a module,
module_frob_arch_sections()tries to figure out the number of PLTs that'll be needed to handle all the RELAs. While doing this, it tries to dedupe PLT allocations for multipleR_AARCH64_CALL26relocations to the same symbol. It does the same forR_AARCH64_JUMP26relocations.To make checks for duplicates easier/faster, it sorts the relocation list by type, symbol and addend. That way, to check for a duplicate relocation, it just needs to compare with the previous entry.
[...]
Applied to arm64 (for-next/misc), thanks!
[1/1] arm64/module: Optimize module load time by optimizing PLT counting
https://git.kernel.org/arm64/c/d4e0340919fb
Signature · 1 line
Catalin
Quoted 54 lines · Saravana Kannan
When loading a module,
module_frob_arch_sections()tries to figure out the number of PLTs that'll be needed to handle all the RELAs. While doing this, it tries to dedupe PLT allocations for multipleR_AARCH64_CALL26relocations to the same symbol. It does the same forR_AARCH64_JUMP26relocations.To make checks for duplicates easier/faster, it sorts the relocation list by type, symbol and addend. That way, to check for a duplicate relocation, it just needs to compare with the previous entry.
However, sorting the entire relocation array is unnecessary and expensive (O(n log n)) because there are a lot of other relocation types that don't need deduping or can't be deduped.
So this commit partitions the array into entries that need deduping and those that don't. And then sorts just the part that needs deduping. And when
CONFIG_RANDOMIZE_BASEis disabled, the sorting is skipped entirely because PLTs are not allocated forR_AARCH64_CALL26andR_AARCH64_JUMP26if it's disabled.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_BASEenabled, these were the results for a few downstream modules:Module Size (MB) wlan 14 video codec 3.8 drm 1.8 IPA 2.5 audio 1.2 gpu 1.8Without this patch:
Module Number of entries sorted Module load time (ms) wlan 243739 283 video codec 74029 138 drm 53837 67 IPA 42800 90 audio 21326 27 gpu 20967 32Total time to load all these module: 637 ms
With this patch:
Module Number of entries sorted Module load time (ms) wlan 22454 61 video codec 10150 47 drm 13014 40 IPA 8097 63 audio 4606 16 gpu 6527 20
Quoted 5 lines
Total time to load all these modules: 247
Time saved during boot for just these 6 modules: 390 ms
Cc Ard Biesheuvel <[email protected]>
[I am no longer at Linaro so please don't use my @linaro.org address]
Quoted 37 lines
Signed-off-by Saravana Kannan <[email protected]>---
v1 -> v2:
- Provided more details in the commit text
- Pulled in Will's comments on the coding style
- Pulled in Ard's suggestion about skipping jumps with the same section index (parts of Will's suggested code)arch/arm64/kernel/module-plts.c | 46 ++++++++++++++++++++++++++++++--- 1 file changed, 43 insertions(+), 3 deletions(-)arch/arm64/kernel/module-plts.c +17
-0@@ -253,6 +253,40 @@ 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 ; +} + +/* Group branch PLT relas at the front end of the array. */ +static int partition_branch_plt_relas (Elf64_Sym *syms,Elf64_Rela *rela, +int numrels,Elf64_Word dstidx) +{
arch/arm64/kernel/module-plts.c static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num, · line 282
Quoted 10 lines
arch/arm64/kernel/module-plts.c +10
-0+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]);
Nit: would be slightly better to put
swap(rela[i++], rela[j]);
here so the next iteration of the loop will not call branch_rela_needs_plt() on rela[i] redundantly. But the current code is also correct.
Quoted 39 lines
arch/arm64/kernel/module-plts.c +16
-4+else + j--; + } + +return i; +} +int module_frob_arch_sections (Elf_Ehdr *ehdr,Elf_Shdr *sechdrs,char *secstrings,struct module *mod) {@@ -290,7 +324,7 @@ 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 +334,14 @@ 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, --2.27.0.111.gc72c7da667-goog
Quoted 56 lines · Ard Biesheuvel
Quoted 54 lines · Saravana Kannan
When loading a module,
module_frob_arch_sections()tries to figure out the number of PLTs that'll be needed to handle all the RELAs. While doing this, it tries to dedupe PLT allocations for multipleR_AARCH64_CALL26relocations to the same symbol. It does the same forR_AARCH64_JUMP26relocations.To make checks for duplicates easier/faster, it sorts the relocation list by type, symbol and addend. That way, to check for a duplicate relocation, it just needs to compare with the previous entry.
However, sorting the entire relocation array is unnecessary and expensive (O(n log n)) because there are a lot of other relocation types that don't need deduping or can't be deduped.
So this commit partitions the array into entries that need deduping and those that don't. And then sorts just the part that needs deduping. And when
CONFIG_RANDOMIZE_BASEis disabled, the sorting is skipped entirely because PLTs are not allocated forR_AARCH64_CALL26andR_AARCH64_JUMP26if it's disabled.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_BASEenabled, these were the results for a few downstream modules:Module Size (MB) wlan 14 video codec 3.8 drm 1.8 IPA 2.5 audio 1.2 gpu 1.8Without this patch:
Module Number of entries sorted Module load time (ms) wlan 243739 283 video codec 74029 138 drm 53837 67 IPA 42800 90 audio 21326 27 gpu 20967 32Total time to load all these module: 637 ms
With this patch:
Module Number of entries sorted Module load time (ms) wlan 22454 61 video codec 10150 47 drm 13014 40 IPA 8097 63 audio 4606 16 gpu 6527 20
Quoted 7 lines
Quoted 5 lines
Total time to load all these modules: 247
Time saved during boot for just these 6 modules: 390 ms
Cc Ard Biesheuvel <[email protected]>[I am no longer at Linaro so please don't use my @linaro.org address]
Hmm... I'm pretty sure I got this using the get_maintainers script. Maybe update the MAINTAINERS file if you haven't already (I didn't check)?
But if I ever manually add in your email, I'll keep this in mind.
Quoted 50 lines
Quoted 37 lines
Signed-off-by Saravana Kannan <[email protected]>---
v1 -> v2:
- Provided more details in the commit text
- Pulled in Will's comments on the coding style
- Pulled in Ard's suggestion about skipping jumps with the same section index (parts of Will's suggested code)arch/arm64/kernel/module-plts.c | 46 ++++++++++++++++++++++++++++++--- 1 file changed, 43 insertions(+), 3 deletions(-)arch/arm64/kernel/module-plts.c +17
-0@@ -253,6 +253,40 @@ 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 ; +} + +/* Group branch PLT relas at the front end of the array. */ +static int partition_branch_plt_relas (Elf64_Sym *syms,Elf64_Rela *rela, +int numrels,Elf64_Word dstidx) +{arch/arm64/kernel/module-plts.c static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num, · line 282
Quoted 10 lines
arch/arm64/kernel/module-plts.c +10
-0+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]);Nit: would be slightly better to put
Quoted 5 lines
swap(rela[i++], rela[j]);here so the next iteration of the loop will not call
branch_rela_needs_plt()on rela[i] redundantly. But the current code is also correct.
Oh yeah, I noticed that unnecessary repeat of branch_rela_needs_plt() on rela[i] when j had to be decremented, but forgot to handle it after I was done with all the testing.
But I did compare it to the code I had written in v1 that didn't have this extra check for rela[i]. I couldn't find any measurable difference in the module load time. Maybe 1ms for the worst case module, but that could have been just run to run variation.
Anyway, maybe send this as another patch since Catalin has already picked this mine?
Thanks,
Saravana
Will Deacon
maintainerQuoted 42 lines · Saravana Kannan
Quoted 41 lines · Ard Biesheuvel
Quoted 25 lines · Saravana Kannan
arch/arm64/kernel/module-plts.c +17
-0@@ -253,6 +253,40 @@ 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 ; +} + +/* Group branch PLT relas at the front end of the array. */ +static int partition_branch_plt_relas (Elf64_Sym *syms,Elf64_Rela *rela, +int numrels,Elf64_Word dstidx) +{arch/arm64/kernel/module-plts.c static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num, · line 282
Quoted 10 lines
arch/arm64/kernel/module-plts.c +10
-0+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]);Nit: would be slightly better to put
swap(rela[i++], rela[j]);
Quoted 7 lines
Quoted 3 lines
here so the next iteration of the loop will not call
branch_rela_needs_plt()on rela[i] redundantly. But the current code is also correct.Oh yeah, I noticed that unnecessary repeat of
branch_rela_needs_plt()on rela[i] when j had to be decremented, but forgot to handle it after I was done with all the testing.
Yeah, I guess you can decrement j as well, but I just think it makes the logic harder to read and more error-prone if we change it later.
Quoted 7 lines
But I did compare it to the code I had written in v1 that didn't have this extra check for rela[i]. I couldn't find any measurable difference in the module load time. Maybe 1ms for the worst case module, but that could have been just run to run variation.
Anyway, maybe send this as another patch since Catalin has already picked this mine?
I think the queued code is fine, so we don't need to micro-optimise it.
Will
Quoted 48 lines · Will Deacon
Quoted 46 lines · Saravana Kannan
Quoted 44 lines · Ard Biesheuvel
Quoted 25 lines · Saravana Kannan
arch/arm64/kernel/module-plts.c +17
-0@@ -253,6 +253,40 @@ 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 ; +} + +/* Group branch PLT relas at the front end of the array. */ +static int partition_branch_plt_relas (Elf64_Sym *syms,Elf64_Rela *rela, +int numrels,Elf64_Word dstidx) +{arch/arm64/kernel/module-plts.c static unsigned int count_plts(Elf64_Sym *syms, Elf64_Rela *rela, int num, · line 282
Quoted 10 lines
arch/arm64/kernel/module-plts.c +10
-0+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]);Nit: would be slightly better to put
swap(rela[i++], rela[j]);here so the next iteration of the loop will not call
branch_rela_needs_plt()on rela[i] redundantly. But the current code is also correct.
Quoted 7 lines
Quoted 3 lines
Oh yeah, I noticed that unnecessary repeat of
branch_rela_needs_plt()on rela[i] when j had to be decremented, but forgot to handle it after I was done with all the testing.Yeah, I guess you can decrement j as well, but I just think it makes the logic harder to read and more error-prone if we change it later.
Indeed,
swap(rela[i++], rela[j--]);
looks even bettter!
But you're right, it's not a big deal.
Quoted 11 lines
Quoted 7 lines
But I did compare it to the code I had written in v1 that didn't have this extra check for rela[i]. I couldn't find any measurable difference in the module load time. Maybe 1ms for the worst case module, but that could have been just run to run variation.
Anyway, maybe send this as another patch since Catalin has already picked this mine?
I think the queued code is fine, so we don't need to micro-optimise it.
Will