2.6.17-rc3 -- SMP alternatives: switching to UP code
Hi list.
I just noticed this in the 2.6.17-rc3 dmesg:
===
Checking 'hlt' instruction... OK. SMP alternatives: switching to UP code Freeing SMP alternatives: 0k freed
ACPI: setting ELCR to 0400 (from 1608)
===
Should I be seeing this "SMP alternatives" thing on a !CONFIG_SMP kernel? It does say 0k, but something is apparently being done at runtime still. Why?
Rene.
Quoted 11 lines · Rene Herman
Hi list.
I just noticed this in the 2.6.17-rc3 dmesg:
===
Checking 'hlt' instruction... OK.
SMP alternatives: switching to UP code
Freeing SMP alternatives: 0k freed
ACPI: setting ELCR to 0400 (from 1608)
===
Quoted 3 lines
Should I be seeing this "SMP alternatives" thing on a !CONFIG_SMP kernel? It does say 0k, but something is apparently being done at runtime still. Why?
This is part of a recent patch by Gerd Hoffmann:
Quoted 3 lines
Quoted 3 lines
Implement SMP alternatives, i.e. switching at runtime between different code versions for UP and SMP. The code can patch both
SMP->UPandUP->SMP. TheUP->SMPcase is useful for CPU hotplug.
An explicit test of the smp variable (not CONFIG_SMP macro) exists in this code. I missed the discussion about this patch, but it appears from reading the patch notes that the behavior is intentional. More information can be found in the git and lkml archives.
Quoted 1 line
Rene.
Thanks,
Chase
Quoted 11 lines · Rene Herman
Hi list.
I just noticed this in the 2.6.17-rc3 dmesg:
===
Checking 'hlt' instruction... OK.
SMP alternatives: switching to UP code
Freeing SMP alternatives: 0k freed
ACPI: setting ELCR to 0400 (from 1608)
===
Quoted 3 lines
Should I be seeing this "SMP alternatives" thing on a !CONFIG_SMP kernel? It does say 0k, but something is apparently being done at runtime still. Why?
The UP kernel has empty alternatives tables (as you've noticed), thus the code doesn't do anything. Nevertheless it probably makes sense to add a few #ifdef CONFIG_SMP lines to avoid confusing people and safe a few bytes ...
cheers,
Gerd
Signature · 4 lines
Gerd Hoffmann <[email protected]> Erst mal heiraten, ein, zwei Kinder, und wenn alles läuft geh' ich nach drei Jahren mit der Familie an die Börse. http://www.suse.de/~kraxel/julika-dora.jpeg
Quoted 10 lines · Gerd Hoffmann
Quoted 3 lines · Rene Herman
Should I be seeing this "SMP alternatives" thing on a !CONFIG_SMP kernel? It does say 0k, but something is apparently being done at runtime still. Why?
The UP kernel has empty alternatives tables (as you've noticed), thus the code doesn't do anything. Nevertheless it probably makes sense to add a few #ifdef
CONFIG_SMPlines to avoid confusing people and safe a few bytes ...
Okay, thanks. Yes, I agree such would make sense.
Rene.
PATCH 1/1 SMP alternatives: skip with UP kernels.
Quoted 1 line
Okay, thanks. Yes, I agree such would make sense.
Patch below. It simply returns in case the tables are empty and nothing is do to, thus avoids printing the confusing message.
cheers,
Gerd
Signature · 4 lines
Gerd Hoffmann <[email protected]> Erst mal heiraten, ein, zwei Kinder, und wenn alles läuft geh' ich nach drei Jahren mit der Familie an die Börse. http://www.suse.de/~kraxel/julika-dora.jpeg
arch/i386/kernel/alternative.c +3 -0
@@ -349,6 +349,9 @@ void __init alternative_instructions(voi smp_alt_once =1 ;#endif +if (0 == (__smp_alt_end - __smp_alt_begin)) +return ;/* no tables, nothing to patch (UP kernel) */ +if (smp_alt_once) {if (1 ==num_possible_cpus ()) {printk (KERN_INFO "SMP alternatives: switching to UP code\n" );
Re: [patch] SMP alternatives: skip with UP kernels.
Gerd Hoffmann wrote:
[ depending on CONFIG_SMP ]
Quoted 4 lines
Quoted 1 line
Okay, thanks. Yes, I agree such would make sense.
Patch below. It simply returns in case the tables are empty and nothing is do to, thus avoids printing the confusing message.
It does avoid the message but well, that's also _all_ it avoids. Why do you want to do it this way? If I apply the attached, things compile and work fine which seems to confirm that, yes, all this smp_alternatives code is simply dead baggage on !CONFIG_SMP.
Yes, the #ifdef in arch/i386/kernel/module.c is a bit clumsy. Just proof of concept, might be abstracted out a bit better. It does lose most of alternatives.c which would seem to be good thing.
Unless I'm missing some points, I believe this code should not be compiled in for !CONFIG_SMP.
One other point, although probably not a relevant one. Your changes in module_finalize() potentially change behaviour. It used to be:
for (s = sechdrs; s < sechdrs + hdr->e_shnum; s++) {
void *seg;
if (strcmp(".altinstructions", secstrings + s->sh_name))
continue;
seg = (void *)s->sh_addr;
apply_alternatives(seg, seg + s->sh_size);
}which means that any .altinstructions section would be patched. It's now:
for (s = sechdrs; s < sechdrs + hdr->e_shnum; s++) {
...
if (!strcmp(".altinstructions", secstrings + s->sh_name))
alt = s;
...
}
if (alt) {
/* patch .altinstructions */
void *aseg = (void *)alt->sh_addr;
apply_alternatives(aseg, aseg + alt->sh_size);
}which means that only the last such section would. I suppose there's only one such section anyway, but not seeing a break in the original does make me wonder if this was originally done on purpose.
Rene.
arch/i386/kernel/alternative.c +6 -0
@@ -118,6 +118,8 @@ void apply_alternatives(struct alt_instr } } +#ifdef CONFIG_SMP +static void alternatives_smp_save (struct alt_instr *start,struct alt_instr *end) {struct alt_instr *a;
@@ -283,10 +285,13 @@ void alternatives_smp_switch(int smp) spin_unlock_irqrestore(&smp_alt, flags); } +#endif /* CONFIG_SMP */ +void __init alternative_instructions (void ) {apply_alternatives (__alt_instructions, __alt_instructions_end); +#ifdef CONFIG_SMP /* switch to patch-once-at-boottime-only mode and free the * tables in case we know the number of CPUs will never ever * change */
@@ -318,4 +323,5 @@ void __init alternative_instructions(voi _text, _etext); alternatives_smp_switch(0); } +#endif }
arch/i386/kernel/module.c +8 -2
@@ -112,12 +112,14 @@ int module_finalize(const Elf_Ehdr *hdr,char *secstrings = (void *)hdr + sechdrs[hdr->e_shstrndx].sh_offset;for (s = sechdrs; s < sechdrs + hdr->e_shnum; s++) { -if (!strcmp (".text" , secstrings + s->sh_name)) - text = s;if (!strcmp (".altinstructions" , secstrings + s->sh_name)) alt = s; +#ifdef CONFIG_SMP if (!strcmp (".smp_locks" , secstrings + s->sh_name)) locks= s; +if (!strcmp (".text" , secstrings + s->sh_name)) + text = s; +#endif }if (alt) {
@@ -126,16 +128,20 @@ int module_finalize(const Elf_Ehdr *hdr,apply_alternatives (aseg, aseg + alt->sh_size); }if (locks && text) { +#ifdef CONFIG_SMP void *lseg = (void *)locks->sh_addr;void *tseg = (void *)text->sh_addr;alternatives_smp_module_add (me, me->name, lseg, lseg + locks->sh_size, tseg, tseg + text->sh_size); +#endif }return 0 ; }void module_arch_cleanup (struct module *mod) { +#ifdef CONFIG_SMP alternatives_smp_module_del (mod); +#endif }
Re: [patch] SMP alternatives: skip with UP kernels.
Quoted 7 lines
Quoted 2 lines
Patch below. It simply returns in case the tables are empty and nothing is do to, thus avoids printing the confusing message.
It does avoid the message but well, that's also _all_ it avoids. Why do you want to do it this way? [ ... patch ... ]
Yes, the #ifdef in
arch/i386/kernel/module.cis a bit clumsy.
Yep, thats why. I wanted avoid exactly that. Having some code need to know that function foobar() is only available with CONFIG_BAZ is set is really ugly ...
The attached patch hides the magic in alternative.h and provides some dummy inline functions for the UP case (gcc should manage to optimize away these calls). No changes in module.c.
cheers,
Gerd
Signature · 4 lines
Gerd Hoffmann <[email protected]> Erst mal heiraten, ein, zwei Kinder, und wenn alles läuft geh' ich nach drei Jahren mit der Familie an die Börse. http://www.suse.de/~kraxel/julika-dora.jpeg
arch/i386/kernel/alternative.c +6 -0
@@ -168,6 +168,8 @@ void apply_alternatives(struct alt_instr } } +#ifdef CONFIG_SMP +static void alternatives_smp_save (struct alt_instr *start,struct alt_instr *end) {struct alt_instr *a;
@@ -328,6 +330,8 @@ void alternatives_smp_switch(int smp) spin_unlock_irqrestore(&smp_alt, flags); } +#endif +void __init alternative_instructions (void ) {if (no_replacement) {
@@ -349,6 +353,7 @@ void __init alternative_instructions(voi smp_alt_once =1 ;#endif +#ifdef CONFIG_SMP if (smp_alt_once) {if (1 ==num_possible_cpus ()) {printk (KERN_INFO "SMP alternatives: switching to UP code\n" );
@@ -370,4 +375,5 @@ void __init alternative_instructions(voi _text, _etext); alternatives_smp_switch(0); } +#endif }
include/asm-i386/alternative.h +8 -0
@@ -17,11 +17,19 @@ struct alt_instr {extern void apply_alternatives (struct alt_instr *start,struct alt_instr *end);struct module ; +#ifdef CONFIG_SMP extern void alternatives_smp_module_add (struct module *mod,char *name,void *locks,void *locks_end,void *text,void *text_end);extern void alternatives_smp_module_del (struct module *mod);extern void alternatives_smp_switch (int smp); +#else +static inline void alternatives_smp_module_add (struct module *mod,char *name, +void *locks,void *locks_end, +void *text,void *text_end) {} +static inline void alternatives_smp_module_del (struct module *mod) {} +static inline void alternatives_smp_switch (int smp) {} +#endif #endif
include/asm-x86_64/alternative.h +9 -0
@@ -17,11 +17,20 @@ struct alt_instr {extern void apply_alternatives (struct alt_instr *start,struct alt_instr *end);struct module ; + +#ifdef CONFIG_SMP extern void alternatives_smp_module_add (struct module *mod,char *name,void *locks,void *locks_end,void *text,void *text_end);extern void alternatives_smp_module_del (struct module *mod);extern void alternatives_smp_switch (int smp); +#else +static inline void alternatives_smp_module_add (struct module *mod,char *name, +void *locks,void *locks_end, +void *text,void *text_end) {} +static inline void alternatives_smp_module_del (struct module *mod) {} +static inline void alternatives_smp_switch (int smp) {} +#endif #endif
Re: [patch] SMP alternatives: skip with UP kernels.
Quoted 9 lines · Gerd Hoffmann
Quoted 1 line
Yes, the #ifdef in
arch/i386/kernel/module.cis a bit clumsy.Yep, thats why. I wanted avoid exactly that. Having some code need to know that function
foobar()is only available withCONFIG_BAZis set is really ugly ...The attached patch hides the magic in
alternative.hand provides some dummy inline functions for the UP case (gcc should manage to optimize away these calls). No changes inmodule.c.
Works for me; it doesn't easily get non-clumsy in module.c I see. Sure you want to keep smp_alt_once outside the #ifdef? Seems to not be doing anything other than being set to 1 for !SMP.
Thanks,
Rene.