Re: [osv-dev] [PATCH] libc: implement posix_madvise
On Tuesday, December 21, 2021 at 5:57:52 PM UTC-5 Nadav Har'El wrote: > On Wed, Dec 22, 2021 at 12:42 AM Waldek Kozaczuk > wrote: > >> The OSV_LIBC_API macro makes sure that during the compilation phase the >> posix_madvise symbol is left as GLOBAL DEFAULT in the libc/mman.o where all >> other symbols are left hidden. The presence of posix_madvise in those 4 >> symbols files is used to feed the version script that is used by the linker >> to eliminate all other symbols but the ones listed in the sum of the >> symbols files for given architecture - aarch64 or x64. The reason we have >> given symbol occur almost always in 4 of those files [(x64 + aarch64) times >> 2] is because most symbols are also present in ld-musl-x86_64.so.1 and our >> dynamic linker advertises this library as well. And each of the symbol >> files was originally generated using scripts/extract_symbols.sh which >> identifies symbols to be exported against a version of kernel.elf built >> with all symbols exported and given Linux library file from the host. At >> the end of the day, it does not matter which file the given symbol is part >> of as we collapse them all into a single version script. >> >> The OSV_LIBC_API used during the compilation phase and the symbol files >> used by the linker may be seen as a duplicate mechanism and frankly, the >> version script would have been enough to export (keep) symbols we want to. >> However OSV_LIBC_API + -fvisibility-hidden "*lets the optimizer produce >> better code*" (see https://gcc.gnu.org/wiki/Visibility) and this >> mechanism but itself is also not enough because there are still some >> symbols left (C++ typeinfo and other symbols, possibly originating from >> libstdc++.a) what we need to eliminate using the version script and linker. >> > > Thanks for the explanation. > > So, my next question would be if we can automatically produce the linker > scripts, in one of two ways: > 1. Recreate them every time on the user's build machine like you said you > build them for the first time, or > This would be sensitive to the host. In the older version of Linux with older libc library, some newer libc symbols would not be present. I thought about it before and I came to the conclusion that maintaining these symbols file manually is better. These symbols lists just need to be supersets ideally across all distributions. 2. Create at least parts of it from automatic grepping of OSV_LIBC_API et > al. in the source files. > That would be difficult to achieve as we could only do that against OSv native code with OSV_LIBC_API in it. Musl sources from musl/ use different scheme - compile everything without -fvisiblity-hidden and annotate all symbols to be hidden with their hidden macro. > > The goal is that you won't need to remember to add a symbols in 6 > different places for it to work. > Keep in mind that we'll note really ave comprehensive tests for all > symbols in all cases, so if I forget to add new_function() to libdl, or to > aarch64, I might not notice this in my tests. > > If the answer is no, then my question would be if we could at least not > duplicate the linker scripts. Can't x86 and aarch have the same list of > exported functions? And if two libraries contain the same symbols, could we > somehow automatically copy symbols between them? > At least we can improve it by keeping common lists (for x64 and aacth64) list - > 99% is the same. We can also add more ways to automatically verify symbols do not get lost, etc. The scipts/extract_symbols.sh has a '--verify" option that can be used from makefile to run a check to verify we are not losing any symbols from the list files. We can have another check that verifies that symbols with OSV_LIBC_API are also present in those lists. I will try to come up with some improvements. > > >> On Tuesday, December 21, 2021 at 4:45:42 PM UTC-5 Nadav Har'El wrote: >> >>> Can you please remind me why we need both OSV_LIBC_API and putting the >>> function name in 5 different files? >>> Isn't there a big risk that later when all this export stuff is less >>> fresh in our mind, we forget one of these 6 places and create a mess? >>> >>> -- >>> Nadav Har'El >>> n...@scylladb.com >>> >>> >>> On Tue, Dec 21, 2021 at 8:26 PM Waldemar Kozaczuk >>> wrote: >>> Some apps/runtimes like dotnet call posix_madvise which we do not implement. This patch adds a simple implementation of it based on madvise with a difference that only POSIX_MADV_DONTNEED is supported. On top of this as required posix_madvise() returns an error without setting errno. Signed-off-by: Waldemar Kozaczuk --- .../aarch64/osv_ld-musl-aarch64.so.1.symbols | 1 + exported_symbols/aarch64/osv_libc.so.6.symbols | 1 + .../x64/osv_ld-musl-x86_64.so.1.symbols| 1 + exported_symbols/x64/osv_libc.so.6.symbols | 1 + libc/mman.cc
Re: [osv-dev] [PATCH] libc: implement posix_madvise
On Wed, Dec 22, 2021 at 12:42 AM Waldek Kozaczuk wrote: > The OSV_LIBC_API macro makes sure that during the compilation phase the > posix_madvise symbol is left as GLOBAL DEFAULT in the libc/mman.o where all > other symbols are left hidden. The presence of posix_madvise in those 4 > symbols files is used to feed the version script that is used by the linker > to eliminate all other symbols but the ones listed in the sum of the > symbols files for given architecture - aarch64 or x64. The reason we have > given symbol occur almost always in 4 of those files [(x64 + aarch64) times > 2] is because most symbols are also present in ld-musl-x86_64.so.1 and our > dynamic linker advertises this library as well. And each of the symbol > files was originally generated using scripts/extract_symbols.sh which > identifies symbols to be exported against a version of kernel.elf built > with all symbols exported and given Linux library file from the host. At > the end of the day, it does not matter which file the given symbol is part > of as we collapse them all into a single version script. > > The OSV_LIBC_API used during the compilation phase and the symbol files > used by the linker may be seen as a duplicate mechanism and frankly, the > version script would have been enough to export (keep) symbols we want to. > However OSV_LIBC_API + -fvisibility-hidden "*lets the optimizer produce > better code*" (see https://gcc.gnu.org/wiki/Visibility) and this > mechanism but itself is also not enough because there are still some > symbols left (C++ typeinfo and other symbols, possibly originating from > libstdc++.a) what we need to eliminate using the version script and linker. > Thanks for the explanation. So, my next question would be if we can automatically produce the linker scripts, in one of two ways: 1. Recreate them every time on the user's build machine like you said you build them for the first time, or 2. Create at least parts of it from automatic grepping of OSV_LIBC_API et al. in the source files. The goal is that you won't need to remember to add a symbols in 6 different places for it to work. Keep in mind that we'll note really ave comprehensive tests for all symbols in all cases, so if I forget to add new_function() to libdl, or to aarch64, I might not notice this in my tests. If the answer is no, then my question would be if we could at least not duplicate the linker scripts. Can't x86 and aarch have the same list of exported functions? And if two libraries contain the same symbols, could we somehow automatically copy symbols between them? > On Tuesday, December 21, 2021 at 4:45:42 PM UTC-5 Nadav Har'El wrote: > >> Can you please remind me why we need both OSV_LIBC_API and putting the >> function name in 5 different files? >> Isn't there a big risk that later when all this export stuff is less >> fresh in our mind, we forget one of these 6 places and create a mess? >> >> -- >> Nadav Har'El >> n...@scylladb.com >> >> >> On Tue, Dec 21, 2021 at 8:26 PM Waldemar Kozaczuk >> wrote: >> >>> Some apps/runtimes like dotnet call posix_madvise >>> which we do not implement. This patch adds a simple >>> implementation of it based on madvise with a difference >>> that only POSIX_MADV_DONTNEED is supported. On top of >>> this as required posix_madvise() returns an error without >>> setting errno. >>> >>> Signed-off-by: Waldemar Kozaczuk >>> --- >>> .../aarch64/osv_ld-musl-aarch64.so.1.symbols | 1 + >>> exported_symbols/aarch64/osv_libc.so.6.symbols | 1 + >>> .../x64/osv_ld-musl-x86_64.so.1.symbols| 1 + >>> exported_symbols/x64/osv_libc.so.6.symbols | 1 + >>> libc/mman.cc | 14 ++ >>> 5 files changed, 18 insertions(+) >>> >>> diff --git a/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols >>> b/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols >>> index c463131a..4c284fba 100644 >>> --- a/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols >>> +++ b/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols >>> @@ -739,6 +739,7 @@ posix_fadvise >>> posix_fadvise64 >>> posix_fallocate >>> posix_fallocate64 >>> +posix_madvise >>> posix_memalign >>> pow >>> pow10 >>> diff --git a/exported_symbols/aarch64/osv_libc.so.6.symbols >>> b/exported_symbols/aarch64/osv_libc.so.6.symbols >>> index 000191b7..ea0bfcdc 100644 >>> --- a/exported_symbols/aarch64/osv_libc.so.6.symbols >>> +++ b/exported_symbols/aarch64/osv_libc.so.6.symbols >>> @@ -667,6 +667,7 @@ posix_fadvise >>> posix_fadvise64 >>> posix_fallocate >>> posix_fallocate64 >>> +posix_madvise >>> posix_memalign >>> ppoll >>> prctl >>> diff --git a/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols >>> b/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols >>> index b3f87859..d88e98ed 100644 >>> --- a/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols >>> +++ b/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols >>> @@ -721,6 +721,7 @@
Re: [osv-dev] [PATCH] libc: implement posix_madvise
The OSV_LIBC_API macro makes sure that during the compilation phase the
posix_madvise symbol is left as GLOBAL DEFAULT in the libc/mman.o where all
other symbols are left hidden. The presence of posix_madvise in those 4
symbols files is used to feed the version script that is used by the linker
to eliminate all other symbols but the ones listed in the sum of the
symbols files for given architecture - aarch64 or x64. The reason we have
given symbol occur almost always in 4 of those files [(x64 + aarch64) times
2] is because most symbols are also present in ld-musl-x86_64.so.1 and our
dynamic linker advertises this library as well. And each of the symbol
files was originally generated using scripts/extract_symbols.sh which
identifies symbols to be exported against a version of kernel.elf built
with all symbols exported and given Linux library file from the host. At
the end of the day, it does not matter which file the given symbol is part
of as we collapse them all into a single version script.
The OSV_LIBC_API used during the compilation phase and the symbol files
used by the linker may be seen as a duplicate mechanism and frankly, the
version script would have been enough to export (keep) symbols we want to.
However OSV_LIBC_API + -fvisibility-hidden "*lets the optimizer produce
better code*" (see https://gcc.gnu.org/wiki/Visibility) and this mechanism
but itself is also not enough because there are still some symbols left
(C++ typeinfo and other symbols, possibly originating from libstdc++.a)
what we need to eliminate using the version script and linker.
On Tuesday, December 21, 2021 at 4:45:42 PM UTC-5 Nadav Har'El wrote:
> Can you please remind me why we need both OSV_LIBC_API and putting the
> function name in 5 different files?
> Isn't there a big risk that later when all this export stuff is less fresh
> in our mind, we forget one of these 6 places and create a mess?
>
> --
> Nadav Har'El
> n...@scylladb.com
>
>
> On Tue, Dec 21, 2021 at 8:26 PM Waldemar Kozaczuk
> wrote:
>
>> Some apps/runtimes like dotnet call posix_madvise
>> which we do not implement. This patch adds a simple
>> implementation of it based on madvise with a difference
>> that only POSIX_MADV_DONTNEED is supported. On top of
>> this as required posix_madvise() returns an error without
>> setting errno.
>>
>> Signed-off-by: Waldemar Kozaczuk
>> ---
>> .../aarch64/osv_ld-musl-aarch64.so.1.symbols | 1 +
>> exported_symbols/aarch64/osv_libc.so.6.symbols | 1 +
>> .../x64/osv_ld-musl-x86_64.so.1.symbols| 1 +
>> exported_symbols/x64/osv_libc.so.6.symbols | 1 +
>> libc/mman.cc | 14 ++
>> 5 files changed, 18 insertions(+)
>>
>> diff --git a/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
>> b/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
>> index c463131a..4c284fba 100644
>> --- a/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
>> +++ b/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
>> @@ -739,6 +739,7 @@ posix_fadvise
>> posix_fadvise64
>> posix_fallocate
>> posix_fallocate64
>> +posix_madvise
>> posix_memalign
>> pow
>> pow10
>> diff --git a/exported_symbols/aarch64/osv_libc.so.6.symbols
>> b/exported_symbols/aarch64/osv_libc.so.6.symbols
>> index 000191b7..ea0bfcdc 100644
>> --- a/exported_symbols/aarch64/osv_libc.so.6.symbols
>> +++ b/exported_symbols/aarch64/osv_libc.so.6.symbols
>> @@ -667,6 +667,7 @@ posix_fadvise
>> posix_fadvise64
>> posix_fallocate
>> posix_fallocate64
>> +posix_madvise
>> posix_memalign
>> ppoll
>> prctl
>> diff --git a/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
>> b/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
>> index b3f87859..d88e98ed 100644
>> --- a/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
>> +++ b/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
>> @@ -721,6 +721,7 @@ posix_fadvise
>> posix_fadvise64
>> posix_fallocate
>> posix_fallocate64
>> +posix_madvise
>> __posix_getopt
>> posix_memalign
>> pow
>> diff --git a/exported_symbols/x64/osv_libc.so.6.symbols
>> b/exported_symbols/x64/osv_libc.so.6.symbols
>> index 6635cabb..07b5368b 100644
>> --- a/exported_symbols/x64/osv_libc.so.6.symbols
>> +++ b/exported_symbols/x64/osv_libc.so.6.symbols
>> @@ -596,6 +596,7 @@ posix_fadvise
>> posix_fadvise64
>> posix_fallocate
>> posix_fallocate64
>> +posix_madvise
>> __posix_getopt
>> posix_memalign
>> ppoll
>> diff --git a/libc/mman.cc b/libc/mman.cc
>> index 9dd6429a..75a94eb0 100644
>> --- a/libc/mman.cc
>> +++ b/libc/mman.cc
>> @@ -257,3 +257,17 @@ void *sbrk(intptr_t increment)
>> errno = ENOMEM;
>> return (void *)-1;
>> }
>> +
>> +static unsigned posix_madvise_to_advise(int advice)
>> +{
>> +if (advice == POSIX_MADV_DONTNEED) {
>> +return mmu::advise_dontneed;
>> +}
>> +return 0;
>> +}
>> +
>> +OSV_LIBC_API
>> +int posix_madvise(void *addr, size_t len, int
Re: [osv-dev] [PATCH] libc: implement posix_madvise
Can you please remind me why we need both OSV_LIBC_API and putting the
function name in 5 different files?
Isn't there a big risk that later when all this export stuff is less fresh
in our mind, we forget one of these 6 places and create a mess?
--
Nadav Har'El
n...@scylladb.com
On Tue, Dec 21, 2021 at 8:26 PM Waldemar Kozaczuk
wrote:
> Some apps/runtimes like dotnet call posix_madvise
> which we do not implement. This patch adds a simple
> implementation of it based on madvise with a difference
> that only POSIX_MADV_DONTNEED is supported. On top of
> this as required posix_madvise() returns an error without
> setting errno.
>
> Signed-off-by: Waldemar Kozaczuk
> ---
> .../aarch64/osv_ld-musl-aarch64.so.1.symbols | 1 +
> exported_symbols/aarch64/osv_libc.so.6.symbols | 1 +
> .../x64/osv_ld-musl-x86_64.so.1.symbols| 1 +
> exported_symbols/x64/osv_libc.so.6.symbols | 1 +
> libc/mman.cc | 14 ++
> 5 files changed, 18 insertions(+)
>
> diff --git a/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
> b/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
> index c463131a..4c284fba 100644
> --- a/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
> +++ b/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
> @@ -739,6 +739,7 @@ posix_fadvise
> posix_fadvise64
> posix_fallocate
> posix_fallocate64
> +posix_madvise
> posix_memalign
> pow
> pow10
> diff --git a/exported_symbols/aarch64/osv_libc.so.6.symbols
> b/exported_symbols/aarch64/osv_libc.so.6.symbols
> index 000191b7..ea0bfcdc 100644
> --- a/exported_symbols/aarch64/osv_libc.so.6.symbols
> +++ b/exported_symbols/aarch64/osv_libc.so.6.symbols
> @@ -667,6 +667,7 @@ posix_fadvise
> posix_fadvise64
> posix_fallocate
> posix_fallocate64
> +posix_madvise
> posix_memalign
> ppoll
> prctl
> diff --git a/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
> b/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
> index b3f87859..d88e98ed 100644
> --- a/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
> +++ b/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
> @@ -721,6 +721,7 @@ posix_fadvise
> posix_fadvise64
> posix_fallocate
> posix_fallocate64
> +posix_madvise
> __posix_getopt
> posix_memalign
> pow
> diff --git a/exported_symbols/x64/osv_libc.so.6.symbols
> b/exported_symbols/x64/osv_libc.so.6.symbols
> index 6635cabb..07b5368b 100644
> --- a/exported_symbols/x64/osv_libc.so.6.symbols
> +++ b/exported_symbols/x64/osv_libc.so.6.symbols
> @@ -596,6 +596,7 @@ posix_fadvise
> posix_fadvise64
> posix_fallocate
> posix_fallocate64
> +posix_madvise
> __posix_getopt
> posix_memalign
> ppoll
> diff --git a/libc/mman.cc b/libc/mman.cc
> index 9dd6429a..75a94eb0 100644
> --- a/libc/mman.cc
> +++ b/libc/mman.cc
> @@ -257,3 +257,17 @@ void *sbrk(intptr_t increment)
> errno = ENOMEM;
> return (void *)-1;
> }
> +
> +static unsigned posix_madvise_to_advise(int advice)
> +{
> +if (advice == POSIX_MADV_DONTNEED) {
> +return mmu::advise_dontneed;
> +}
> +return 0;
> +}
> +
> +OSV_LIBC_API
> +int posix_madvise(void *addr, size_t len, int advice) {
> +auto err = mmu::advise(addr, len, posix_madvise_to_advise(advice));
> +return err.get();
> +}
> --
> 2.31.1
>
> --
> You received this message because you are subscribed to the Google Groups
> "OSv Development" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to osv-dev+unsubscr...@googlegroups.com.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/osv-dev/20211221182614.242226-1-jwkozaczuk%40gmail.com
> .
>
--
You received this message because you are subscribed to the Google Groups "OSv
Development" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to osv-dev+unsubscr...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/osv-dev/CANEVyjsy2qaz-LRX4ZjMuRFa848e2rjAbsTs239HObzYiF5Ueg%40mail.gmail.com.
[osv-dev] [PATCH] libc: implement posix_madvise
Some apps/runtimes like dotnet call posix_madvise
which we do not implement. This patch adds a simple
implementation of it based on madvise with a difference
that only POSIX_MADV_DONTNEED is supported. On top of
this as required posix_madvise() returns an error without
setting errno.
Signed-off-by: Waldemar Kozaczuk
---
.../aarch64/osv_ld-musl-aarch64.so.1.symbols | 1 +
exported_symbols/aarch64/osv_libc.so.6.symbols | 1 +
.../x64/osv_ld-musl-x86_64.so.1.symbols| 1 +
exported_symbols/x64/osv_libc.so.6.symbols | 1 +
libc/mman.cc | 14 ++
5 files changed, 18 insertions(+)
diff --git a/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
b/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
index c463131a..4c284fba 100644
--- a/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
+++ b/exported_symbols/aarch64/osv_ld-musl-aarch64.so.1.symbols
@@ -739,6 +739,7 @@ posix_fadvise
posix_fadvise64
posix_fallocate
posix_fallocate64
+posix_madvise
posix_memalign
pow
pow10
diff --git a/exported_symbols/aarch64/osv_libc.so.6.symbols
b/exported_symbols/aarch64/osv_libc.so.6.symbols
index 000191b7..ea0bfcdc 100644
--- a/exported_symbols/aarch64/osv_libc.so.6.symbols
+++ b/exported_symbols/aarch64/osv_libc.so.6.symbols
@@ -667,6 +667,7 @@ posix_fadvise
posix_fadvise64
posix_fallocate
posix_fallocate64
+posix_madvise
posix_memalign
ppoll
prctl
diff --git a/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
b/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
index b3f87859..d88e98ed 100644
--- a/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
+++ b/exported_symbols/x64/osv_ld-musl-x86_64.so.1.symbols
@@ -721,6 +721,7 @@ posix_fadvise
posix_fadvise64
posix_fallocate
posix_fallocate64
+posix_madvise
__posix_getopt
posix_memalign
pow
diff --git a/exported_symbols/x64/osv_libc.so.6.symbols
b/exported_symbols/x64/osv_libc.so.6.symbols
index 6635cabb..07b5368b 100644
--- a/exported_symbols/x64/osv_libc.so.6.symbols
+++ b/exported_symbols/x64/osv_libc.so.6.symbols
@@ -596,6 +596,7 @@ posix_fadvise
posix_fadvise64
posix_fallocate
posix_fallocate64
+posix_madvise
__posix_getopt
posix_memalign
ppoll
diff --git a/libc/mman.cc b/libc/mman.cc
index 9dd6429a..75a94eb0 100644
--- a/libc/mman.cc
+++ b/libc/mman.cc
@@ -257,3 +257,17 @@ void *sbrk(intptr_t increment)
errno = ENOMEM;
return (void *)-1;
}
+
+static unsigned posix_madvise_to_advise(int advice)
+{
+if (advice == POSIX_MADV_DONTNEED) {
+return mmu::advise_dontneed;
+}
+return 0;
+}
+
+OSV_LIBC_API
+int posix_madvise(void *addr, size_t len, int advice) {
+auto err = mmu::advise(addr, len, posix_madvise_to_advise(advice));
+return err.get();
+}
--
2.31.1
--
You received this message because you are subscribed to the Google Groups "OSv
Development" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to osv-dev+unsubscr...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/osv-dev/20211221182614.242226-1-jwkozaczuk%40gmail.com.
