From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 2155) id B5A323858403; Mon, 25 Mar 2024 10:35:04 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org B5A323858403 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=sourceware.org; s=default; t=1711362904; bh=qRH85IUUHXmmvuUkE6rCrxwVDoJ+WQlEymC9BjPhsDE=; h=From:To:Subject:Date:From; b=jXy207IOv5LO7ICzFP81tppJ4SqEvrGZN5fdqqUKeI1NGoC8lYY5O+U3E6pvabohG Ahtmr01iDkx6RAEwTMEOyCbbTN+QXH59zm8ZIipMno+vU0HPxXpt7yPiPwpMX8kAvr fMlC17INNP6bTHhEaodBtpqyA4MrRCcgQwu4gebg= Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable From: Corinna Vinschen To: newlib-cvs@sourceware.org Subject: [newlib-cygwin/main] amdgcn: Implement proper locks X-Act-Checkin: newlib-cygwin X-Git-Author: Andrew Stubbs X-Git-Refname: refs/heads/main X-Git-Oldrev: 9761c6cca1dcc8463055876e0ff14b84ba0efb89 X-Git-Newrev: 7dd4eb1db9e1b1b9f14ef5b743705156e5f370e1 Message-Id: <20240325103504.B5A323858403@sourceware.org> Date: Mon, 25 Mar 2024 10:35:04 +0000 (GMT) List-Id: https://sourceware.org/git/gitweb.cgi?p=3Dnewlib-cygwin.git;h=3D7dd4eb1db9e= 1b1b9f14ef5b743705156e5f370e1 commit 7dd4eb1db9e1b1b9f14ef5b743705156e5f370e1 Author: Andrew Stubbs AuthorDate: Fri Mar 22 14:53:30 2024 +0000 Commit: Corinna Vinschen CommitDate: Mon Mar 25 11:32:52 2024 +0100 amdgcn: Implement proper locks =20 This should prevent printf output from multiple threads getting garbled. =20 I don't know why IO ever worked properly -- probably it was always a bit broken -- but the GFX11 devices have a different cache architecture and trying to print from many threads at once corrupted the FILE buffers. Diff: --- newlib/Makefile.in | 24 +++- newlib/libc/sys/amdgcn/Makefile.inc | 3 +- newlib/libc/sys/amdgcn/include/sys/lock.h | 39 +++++++ newlib/libc/sys/amdgcn/lock.c | 187 ++++++++++++++++++++++++++= ++++ 4 files changed, 250 insertions(+), 3 deletions(-) diff --git a/newlib/Makefile.in b/newlib/Makefile.in index 8c71dbab3e61..b45c42c365d7 100644 --- a/newlib/Makefile.in +++ b/newlib/Makefile.in @@ -560,7 +560,8 @@ check_PROGRAMS =3D =20 @HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@am__append_43 =3D \ @HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/close.c libc/sys/amdgcn/fs= tat.c libc/sys/amdgcn/isatty.c libc/sys/amdgcn/lseek.c libc/sys/amdgcn/read= .c libc/sys/amdgcn/write.c \ -@HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/fcntl.c libc/sys/amdgcn/ge= tpid.c libc/sys/amdgcn/kill.c libc/sys/amdgcn/open.c libc/sys/amdgcn/raise.= c libc/sys/amdgcn/stat.c libc/sys/amdgcn/unlink.c +@HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/fcntl.c libc/sys/amdgcn/ge= tpid.c libc/sys/amdgcn/kill.c libc/sys/amdgcn/open.c libc/sys/amdgcn/raise.= c libc/sys/amdgcn/stat.c \ +@HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/unlink.c libc/sys/amdgcn/l= ock.c =20 @HAVE_LIBC_SYS_ARM_DIR_TRUE@am__append_44 =3D libc/sys/arm/access.c libc/s= ys/arm/aeabi_atexit.c libc/sys/arm/sysconf.c @HAVE_LIBC_SYS_ARM_DIR_TRUE@@MAY_SUPPLY_SYSCALLS_TRUE@am__append_45 =3D li= bc/sys/arm/libcfunc.c libc/sys/arm/trap.S libc/sys/arm/syscalls.c @@ -1656,7 +1657,8 @@ am__objects_51 =3D libc/ssp/libc_a-chk_fail.$(OBJEXT)= \ @HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/libc_a-open.$(OBJEXT) \ @HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/libc_a-raise.$(OBJEXT) \ @HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/libc_a-stat.$(OBJEXT) \ -@HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/libc_a-unlink.$(OBJEXT) +@HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/libc_a-unlink.$(OBJEXT) \ +@HAVE_LIBC_SYS_AMDGCN_DIR_TRUE@ libc/sys/amdgcn/libc_a-lock.$(OBJEXT) @HAVE_LIBC_SYS_ARM_DIR_TRUE@am__objects_55 =3D libc/sys/arm/libc_a-access.= $(OBJEXT) \ @HAVE_LIBC_SYS_ARM_DIR_TRUE@ libc/sys/arm/libc_a-aeabi_atexit.$(OBJEXT) \ @HAVE_LIBC_SYS_ARM_DIR_TRUE@ libc/sys/arm/libc_a-sysconf.$(OBJEXT) @@ -7268,6 +7270,9 @@ libc/sys/amdgcn/libc_a-stat.$(OBJEXT): \ libc/sys/amdgcn/libc_a-unlink.$(OBJEXT): \ libc/sys/amdgcn/$(am__dirstamp) \ libc/sys/amdgcn/$(DEPDIR)/$(am__dirstamp) +libc/sys/amdgcn/libc_a-lock.$(OBJEXT): \ + libc/sys/amdgcn/$(am__dirstamp) \ + libc/sys/amdgcn/$(DEPDIR)/$(am__dirstamp) libc/sys/arm/$(am__dirstamp): @$(MKDIR_P) libc/sys/arm @: > libc/sys/arm/$(am__dirstamp) @@ -13831,6 +13836,7 @@ distclean-compile: @AMDEP_TRUE@@am__include@ @am__quote@libc/sys/amdgcn/$(DEPDIR)/libc_a-getp= id.Po@am__quote@ @AMDEP_TRUE@@am__include@ @am__quote@libc/sys/amdgcn/$(DEPDIR)/libc_a-isat= ty.Po@am__quote@ @AMDEP_TRUE@@am__include@ @am__quote@libc/sys/amdgcn/$(DEPDIR)/libc_a-kill= .Po@am__quote@ +@AMDEP_TRUE@@am__include@ @am__quote@libc/sys/amdgcn/$(DEPDIR)/libc_a-lock= .Po@am__quote@ @AMDEP_TRUE@@am__include@ @am__quote@libc/sys/amdgcn/$(DEPDIR)/libc_a-lsee= k.Po@am__quote@ @AMDEP_TRUE@@am__include@ @am__quote@libc/sys/amdgcn/$(DEPDIR)/libc_a-open= .Po@am__quote@ @AMDEP_TRUE@@am__include@ @am__quote@libc/sys/amdgcn/$(DEPDIR)/libc_a-rais= e.Po@am__quote@ @@ -30960,6 +30966,20 @@ libc/sys/amdgcn/libc_a-unlink.obj: libc/sys/amdgcn= /unlink.c @AMDEP_TRUE@@am__fastdepCC_FALSE@ DEPDIR=3D$(DEPDIR) $(CCDEPMODE) $(depcom= p) @AMDEPBACKSLASH@ @am__fastdepCC_FALSE@ $(AM_V_CC@am__nodep@)$(CC) $(DEFS) $(DEFAULT_INCLUDE= S) $(INCLUDES) $(libc_a_CPPFLAGS) $(CPPFLAGS) $(libc_a_CFLAGS) $(CFLAGS) -c= -o libc/sys/amdgcn/libc_a-unlink.obj `if test -f 'libc/sys/amdgcn/unlink.c= '; then $(CYGPATH_W) 'libc/sys/amdgcn/unlink.c'; else $(CYGPATH_W) '$(srcdi= r)/libc/sys/amdgcn/unlink.c'; fi` =20 +libc/sys/amdgcn/libc_a-lock.o: libc/sys/amdgcn/lock.c +@am__fastdepCC_TRUE@ $(AM_V_CC)$(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDE= S) $(libc_a_CPPFLAGS) $(CPPFLAGS) $(libc_a_CFLAGS) $(CFLAGS) -MT libc/sys/a= mdgcn/libc_a-lock.o -MD -MP -MF libc/sys/amdgcn/$(DEPDIR)/libc_a-lock.Tpo -= c -o libc/sys/amdgcn/libc_a-lock.o `test -f 'libc/sys/amdgcn/lock.c' || ech= o '$(srcdir)/'`libc/sys/amdgcn/lock.c +@am__fastdepCC_TRUE@ $(AM_V_at)$(am__mv) libc/sys/amdgcn/$(DEPDIR)/libc_a-= lock.Tpo libc/sys/amdgcn/$(DEPDIR)/libc_a-lock.Po +@AMDEP_TRUE@@am__fastdepCC_FALSE@ $(AM_V_CC)source=3D'libc/sys/amdgcn/lock= .c' object=3D'libc/sys/amdgcn/libc_a-lock.o' libtool=3Dno @AMDEPBACKSLASH@ +@AMDEP_TRUE@@am__fastdepCC_FALSE@ DEPDIR=3D$(DEPDIR) $(CCDEPMODE) $(depcom= p) @AMDEPBACKSLASH@ +@am__fastdepCC_FALSE@ $(AM_V_CC@am__nodep@)$(CC) $(DEFS) $(DEFAULT_INCLUDE= S) $(INCLUDES) $(libc_a_CPPFLAGS) $(CPPFLAGS) $(libc_a_CFLAGS) $(CFLAGS) -c= -o libc/sys/amdgcn/libc_a-lock.o `test -f 'libc/sys/amdgcn/lock.c' || echo= '$(srcdir)/'`libc/sys/amdgcn/lock.c + +libc/sys/amdgcn/libc_a-lock.obj: libc/sys/amdgcn/lock.c +@am__fastdepCC_TRUE@ $(AM_V_CC)$(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDE= S) $(libc_a_CPPFLAGS) $(CPPFLAGS) $(libc_a_CFLAGS) $(CFLAGS) -MT libc/sys/a= mdgcn/libc_a-lock.obj -MD -MP -MF libc/sys/amdgcn/$(DEPDIR)/libc_a-lock.Tpo= -c -o libc/sys/amdgcn/libc_a-lock.obj `if test -f 'libc/sys/amdgcn/lock.c'= ; then $(CYGPATH_W) 'libc/sys/amdgcn/lock.c'; else $(CYGPATH_W) '$(srcdir)/= libc/sys/amdgcn/lock.c'; fi` +@am__fastdepCC_TRUE@ $(AM_V_at)$(am__mv) libc/sys/amdgcn/$(DEPDIR)/libc_a-= lock.Tpo libc/sys/amdgcn/$(DEPDIR)/libc_a-lock.Po +@AMDEP_TRUE@@am__fastdepCC_FALSE@ $(AM_V_CC)source=3D'libc/sys/amdgcn/lock= .c' object=3D'libc/sys/amdgcn/libc_a-lock.obj' libtool=3Dno @AMDEPBACKSLASH@ +@AMDEP_TRUE@@am__fastdepCC_FALSE@ DEPDIR=3D$(DEPDIR) $(CCDEPMODE) $(depcom= p) @AMDEPBACKSLASH@ +@am__fastdepCC_FALSE@ $(AM_V_CC@am__nodep@)$(CC) $(DEFS) $(DEFAULT_INCLUDE= S) $(INCLUDES) $(libc_a_CPPFLAGS) $(CPPFLAGS) $(libc_a_CFLAGS) $(CFLAGS) -c= -o libc/sys/amdgcn/libc_a-lock.obj `if test -f 'libc/sys/amdgcn/lock.c'; t= hen $(CYGPATH_W) 'libc/sys/amdgcn/lock.c'; else $(CYGPATH_W) '$(srcdir)/lib= c/sys/amdgcn/lock.c'; fi` + libc/sys/arm/libc_a-access.o: libc/sys/arm/access.c @am__fastdepCC_TRUE@ $(AM_V_CC)$(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDE= S) $(libc_a_CPPFLAGS) $(CPPFLAGS) $(libc_a_CFLAGS) $(CFLAGS) -MT libc/sys/a= rm/libc_a-access.o -MD -MP -MF libc/sys/arm/$(DEPDIR)/libc_a-access.Tpo -c = -o libc/sys/arm/libc_a-access.o `test -f 'libc/sys/arm/access.c' || echo '$= (srcdir)/'`libc/sys/arm/access.c @am__fastdepCC_TRUE@ $(AM_V_at)$(am__mv) libc/sys/arm/$(DEPDIR)/libc_a-acc= ess.Tpo libc/sys/arm/$(DEPDIR)/libc_a-access.Po diff --git a/newlib/libc/sys/amdgcn/Makefile.inc b/newlib/libc/sys/amdgcn/M= akefile.inc index c1570b2ad804..4e540fc24bfe 100644 --- a/newlib/libc/sys/amdgcn/Makefile.inc +++ b/newlib/libc/sys/amdgcn/Makefile.inc @@ -1,3 +1,4 @@ libc_a_SOURCES +=3D \ %D%/close.c %D%/fstat.c %D%/isatty.c %D%/lseek.c %D%/read.c %D%/write.c \ - %D%/fcntl.c %D%/getpid.c %D%/kill.c %D%/open.c %D%/raise.c %D%/stat.c %D%= /unlink.c + %D%/fcntl.c %D%/getpid.c %D%/kill.c %D%/open.c %D%/raise.c %D%/stat.c \ + %D%/unlink.c %D%/lock.c diff --git a/newlib/libc/sys/amdgcn/include/sys/lock.h b/newlib/libc/sys/am= dgcn/include/sys/lock.h new file mode 100644 index 000000000000..0e0e667e5eec --- /dev/null +++ b/newlib/libc/sys/amdgcn/include/sys/lock.h @@ -0,0 +1,39 @@ +#ifndef __SYS_LOCK_H__ +#define __SYS_LOCK_H__ + +#include +#include <_ansi.h> + +typedef unsigned int _LOCK_T; +typedef unsigned int _LOCK_RECURSIVE_T; + +#define __LOCK_INIT(CLASS,LOCK) CLASS _LOCK_T LOCK =3D 0; +#define __LOCK_INIT_RECURSIVE(CLASS,LOCK) __LOCK_INIT(CLASS,LOCK) + +#define __lock_init(LOCK) LOCK =3D 0 +#define __lock_init_recursive(LOCK) LOCK =3D 0 +#define __lock_close(LOCK) ((void)0) +#define __lock_close_recursive(LOCK) ((void) 0) +#define __lock_acquire(LOCK) __gcn_lock_acquire (&LOCK) +#define __lock_acquire_recursive(LOCK) \ + __gcn_lock_acquire_recursive (&LOCK) +#define __lock_try_acquire(LOCK) __gcn_try_lock_acquire (&LOCK) +#define __lock_try_acquire_recursive(LOCK) \ + __gcn_lock_try_acquire_recursive (&LOCK) +#define __lock_release(LOCK) __gcn_lock_release (&LOCK) +#define __lock_release_recursive(LOCK) \ + __gcn_lock_release_recursive (&LOCK) + + +int __gcn_try_lock_acquire (_LOCK_T *lock_ptr); +void __gcn_lock_acquire (_LOCK_T *lock_ptr); +void __gcn_lock_release (_LOCK_T *lock_ptr); +int __gcn_lock_try_acquire_recursive (_LOCK_T *lock_ptr); +void __gcn_lock_acquire_recursive (_LOCK_T *lock_ptr); +void __gcn_lock_release_recursive (_LOCK_T *lock_ptr); + +#ifdef __cplusplus +} +#endif + +#endif /* __SYS_LOCK_H__ */ diff --git a/newlib/libc/sys/amdgcn/lock.c b/newlib/libc/sys/amdgcn/lock.c new file mode 100644 index 000000000000..dcc93cb01621 --- /dev/null +++ b/newlib/libc/sys/amdgcn/lock.c @@ -0,0 +1,187 @@ +/* + * Support file for amdgcn in newlib. + * Copyright (c) 2024 BayLibre. + * + * The authors hereby grant permission to use, copy, modify, distribute, + * and license this software and its documentation for any purpose, provid= ed + * that existing copyright notices are retained in all copies and that this + * notice is included verbatim in any distributions. No written agreement, + * license, or royalty fee is required for any of the authorized uses. + * Modifications to this software may be copyrighted by their authors + * and need not follow the licensing terms described here, provided that + * the new terms are clearly indicated on the first page of each file where + * they apply. + */ + +/* Lock routines for AMD GPU devices. + + The lock is a 32-bit int: + - bits 0-3: wavefront id + - bits 4-23: workgroup id (+1, so never zero) + - bits 24-31: recursive lock count. + + The purpose of the "relaxed" loads and stores being "atomic" here is + mostly just to ensure we punch through the caches consistently. + + Non-recursive locks may be unlocked by any thread. It's an error to + attempt to unlock a recursive lock from the wrong thread. + + The DEBUG statements here use sprintf and write to avoid taking locks + themselves. */ + +#include +#include + +#define DEBUG 0 + +#if DEBUG +extern void write(int, char *, int); +#endif + +static unsigned +__gcn_thread_id () +{ + /* Dim(0) is the workgroup ID; range 0 to maybe thousands. + Dim(1) is the wavefront ID; range 0 to 15. */ + return (((__builtin_gcn_dim_pos (0) + 1) << 4) + + __builtin_gcn_dim_pos (1)); +} + +static int +__gcn_lock_acquire_int (_LOCK_T *lock_ptr, int _try) +{ + int id =3D __gcn_thread_id (); + +#if DEBUG + char buf[1000]; + __builtin_sprintf (buf,"acquire:%p(%d) lock_value:0x%x id:0x%x", lock_p= tr, + _try, *lock_ptr, id); + write (1, buf, __builtin_strlen(buf)); +#endif + + int expected =3D 0; + while (!__atomic_compare_exchange_n (lock_ptr, &expected, id, 0, + __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) + { + /* Lock *not* acquired. */ + if (_try) + return 0; + else + { + asm ("s_sleep 64"); + expected =3D 0; + } + } + +#if DEBUG + __builtin_sprintf (buf,"acquired:%p(%d) lock_value:0x%x id:0x%x", lock_= ptr, + _try, *lock_ptr, id); + write (1, buf, __builtin_strlen(buf)); +#endif + + return 1; +} + +int +__gcn_try_lock_acquire (_LOCK_T *lock_ptr) +{ + return __gcn_lock_acquire_int (lock_ptr, 1); +} + +void +__gcn_lock_acquire (_LOCK_T *lock_ptr) +{ + __gcn_lock_acquire_int (lock_ptr, 0); +} + +static int +__gcn_lock_acquire_recursive_int (_LOCK_T *lock_ptr, int _try) +{ + int id =3D __gcn_thread_id (); + +#if DEBUG + char buf[1000]; + __builtin_sprintf (buf,"acquire recursive:%p(%d) lock_value:0x%x id:0x%= x", + lock_ptr, _try, *lock_ptr, id); + write (1, buf, __builtin_strlen(buf)); +#endif + + unsigned int lock_value =3D __atomic_load_n (lock_ptr, __ATOMIC_RELAXED); + if ((lock_value & 0xffffff) =3D=3D id) + { + /* This thread already holds the lock. + Increment the recursion counter and update the lock. */ + int count =3D lock_value >> 24; + lock_value =3D ((count + 1) << 24) | id; + __atomic_store_n (lock_ptr, lock_value, __ATOMIC_RELAXED); + +#if DEBUG + __builtin_sprintf (buf, + "increment recursive:%p(%d) lock_value:0x%x id:0x%x", + lock_ptr, _try, *lock_ptr, id); + write (1, buf, __builtin_strlen(buf)); +#endif + + return 1; + } + else + return __gcn_lock_acquire_int (lock_ptr, _try); +} + +int +__gcn_lock_try_acquire_recursive (_LOCK_T *lock_ptr) +{ + return __gcn_lock_acquire_recursive_int (lock_ptr, 1); +} + +void +__gcn_lock_acquire_recursive (_LOCK_T *lock_ptr) +{ + __gcn_lock_acquire_recursive_int (lock_ptr, 0); +} + +void +__gcn_lock_release (_LOCK_T *lock_ptr) +{ +#if DEBUG + char buf[1000]; + __builtin_sprintf (buf,"release:%p lock_value:0x%x id:0x%x", lock_ptr, + *lock_ptr, __gcn_thread_id()); + write (1, buf, __builtin_strlen(buf)); +#endif + + __atomic_store_n (lock_ptr, 0, __ATOMIC_RELEASE); +} + +void +__gcn_lock_release_recursive (_LOCK_T *lock_ptr) +{ + int id =3D __gcn_thread_id (); + unsigned int lock_value =3D __atomic_load_n (lock_ptr, __ATOMIC_RELAXED); + +#if DEBUG + char buf[1000]; + __builtin_sprintf (buf, "release recursive:%p lock_value:0x%x id:0x%x", + lock_ptr, lock_value, id); + write (1, buf, __builtin_strlen(buf)); +#endif + + /* It is an error to call this function from the wrong thread. */ + assert ((lock_value & 0xffffff) =3D=3D id); + + /* Decrement or release the lock. */ + int count =3D lock_value >> 24; + if (count > 0) + { + lock_value =3D ((count - 1) << 24) | id; + __atomic_store_n (lock_ptr, lock_value, __ATOMIC_RELAXED); + +#if DEBUG + __builtin_sprintf (buf, "decrement recursive:%p lock_value:0x%x id:= 0x%x", + lock_ptr, *lock_ptr, id); + write (1, buf, __builtin_strlen(buf)); +#endif + } + else + __gcn_lock_release (lock_ptr); +}