[v4] mk: fix scope of disabling AVX512F support

Message ID 20190107164942.88785-1-ferruh.yigit@intel.com (mailing list archive)
State Superseded, archived
Delegated to: Thomas Monjalon
Headers
Series [v4] mk: fix scope of disabling AVX512F support |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/mellanox-Performance-Testing success Performance Testing PASS
ci/intel-Performance-Testing success Performance Testing PASS
ci/Intel-compilation success Compilation OK

Commit Message

Ferruh Yigit Jan. 7, 2019, 4:49 p.m. UTC
  AVX512 was disabled for GCC because of Bugzilla issue 97 [1],
the GCC defect submitted for the issue [2] highlighted that this is
a known binutils version 2.30 issue.

Narrowed the scope of no-avx512 to the this specific binutils version.

[1]
https://bugs.dpdk.org/show_bug.cgi?id=97

[2]
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88096

Fixes: 8d07c82b239f ("mk: disable gcc AVX512F support")
Cc: stable@dpdk.org

Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
---
Cc: Tom Barbette <barbette@kth.se>
Cc: Thomas Monjalon <thomas@monjalon.net>
Cc: Yongseok Koh <yskoh@mellanox.com>
Cc: Konstantin Ananyev <konstantin.ananyev@intel.com>
Cc: Bruce Richardson <bruce.richardson@intel.com>

v2:
* Added warning message (print only once for eal)
* Moved decision details to compat file, kept execution in cpuflags

v3:
* replace 'ld' with '$(LD)' for cross build
* added meson support

v4:
* Document this limitation in release notes
Cc: Vipin Varghese <vipin.varghese@intel.com>
---
 config/meson.build                       |  8 ++++++
 doc/guides/rel_notes/known_issues.rst    | 32 ++++++++++++++++++++++++
 doc/guides/rel_notes/release_19_02.rst   | 13 ++++++++++
 mk/rte.cpuflags.mk                       |  4 +--
 mk/toolchain/gcc/rte.toolchain-compat.mk | 10 ++++++++
 5 files changed, 65 insertions(+), 2 deletions(-)
  

Comments

Varghese, Vipin Jan. 8, 2019, 3:50 a.m. UTC | #1
diff --git a/doc/guides/rel_notes/known_issues.rst b/doc/guides/rel_notes/known_issues.rst
index a1face9c1..b0057a2bc 100644
--- a/doc/guides/rel_notes/known_issues.rst
+++ b/doc/guides/rel_notes/known_issues.rst
@@ -826,3 +826,35 @@ Kernel crash when hot-unplug igb_uio device while DPDK application is running
 
 **Driver/Module**:
    ``igb_uio`` module.
+
+
+AVX-512 support disabled
+------------------------
+
+**Description**:
+   ``AVX-512`` support has been disabled on some conditions.
+   This shouldn't be confused with ``CONFIG_RTE_ENABLE_AVX512`` config option which is already
+   disabled by default. This config option defines if ``AVX-512`` specific implementations of
+   some file to be used or not. What has been disabled is compiler feature to produce ``AVX-512``
+   instructions from any source code.
+
+   On DPDK v18.11 ``AVX-512`` disabled for all ``GCC`` builds which reported to cause a performance
+   drop.
+
+   On DPDK v19.02 ``AVX-512`` disable scope reduced to ``GCC`` and ``binutils version 2.30`` based
+   on information accured from the GCC community defect.
+
+**Reason**:
+   Generated ``AVX-512`` code cause crash:
+   https://bugs.dpdk.org/show_bug.cgi?id=97
+   https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88096
+
+**Resolution/Workaround**:
+   Update ``binutils`` to newer version than ``2.30``.
+   Use different compiler, like ``clang`` for this case.
+
+**Affected Environment/Platform**:
+    ``GCC`` and ``binutils version 2.30``.
+
+**Driver/Module**:
+    ALL.
diff --git a/doc/guides/rel_notes/release_19_02.rst b/doc/guides/rel_notes/release_19_02.rst
index 39c6ff826..b6322b282 100644
--- a/doc/guides/rel_notes/release_19_02.rst
+++ b/doc/guides/rel_notes/release_19_02.rst
@@ -276,6 +276,19 @@ Known Issues
    Also, make sure to start the actual text at the margin.
    =========================================================
 
+* ``AVX-512`` support has been disabled for ``GCC`` builds when 
+``binutils 2.30``
+  is detected [1] because of a crash [2]. This can affect ``native`` 
+machine type
+  build targets on the platforms that support ``AVX512F`` like ``Intel 
+Skylake``
+  processors, and can cause a possible performance drop. The immediate 
+workaround
+  is to use ``clang`` compiler on these platforms.
+  Initial workaround in DPDK v18.11 was to disable ``AVX-512`` support 
+for ``GCC``
+  completely, but based on information on defect submitted to GCC 
+community [3],
+  issue has been identified as ``binutils 2.30`` issue. Since currently 
+only GCC
+  generates ``AVX-512`` instructions, scope limited to ``GCC`` and 
+``binutils 2.30``
+
+  - [1]: Commit ("mk: fix scope of disabling AVX512F support")
+  - [2]: https://bugs.dpdk.org/show_bug.cgi?id=97
+  - [3]: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88096
 
Sharing ACK for documentation and update in release notes

Acked-by: Vipin Varghese <Vipin.varghese@intel.com>
  
Thomas Monjalon Jan. 14, 2019, 11:55 a.m. UTC | #2
Hi,
Few details reviewed below,

07/01/2019 17:49, Ferruh Yigit:
> --- a/doc/guides/rel_notes/known_issues.rst
> +++ b/doc/guides/rel_notes/known_issues.rst
> +AVX-512 support disabled
> +------------------------
> +
> +**Description**:
> +   ``AVX-512`` support has been disabled on some conditions.
> +   This shouldn't be confused with ``CONFIG_RTE_ENABLE_AVX512`` config option which is already
> +   disabled by default. This config option defines if ``AVX-512`` specific implementations of
> +   some file to be used or not. What has been disabled is compiler feature to produce ``AVX-512``
> +   instructions from any source code.
> +
> +   On DPDK v18.11 ``AVX-512`` disabled for all ``GCC`` builds which reported to cause a performance
> +   drop.

*is* disabled

> +
> +   On DPDK v19.02 ``AVX-512`` disable scope reduced to ``GCC`` and ``binutils version 2.30`` based

*is* reduced

> +   on information accured from the GCC community defect.
> +
> +**Reason**:
> +   Generated ``AVX-512`` code cause crash:
> +   https://bugs.dpdk.org/show_bug.cgi?id=97
> +   https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88096
> +
> +**Resolution/Workaround**:
> +   Update ``binutils`` to newer version than ``2.30``.
> +   Use different compiler, like ``clang`` for this case.

These are 2 possible workarounds. Should we say "Or" ?

> +
> +**Affected Environment/Platform**:
> +    ``GCC`` and ``binutils version 2.30``.
> --- a/doc/guides/rel_notes/release_19_02.rst
> +++ b/doc/guides/rel_notes/release_19_02.rst
> @@ -276,6 +276,19 @@ Known Issues
>     Also, make sure to start the actual text at the margin.
>     =========================================================
>  
> +* ``AVX-512`` support has been disabled for ``GCC`` builds when ``binutils 2.30``
> +  is detected [1] because of a crash [2]. This can affect ``native`` machine type
> +  build targets on the platforms that support ``AVX512F`` like ``Intel Skylake``
> +  processors, and can cause a possible performance drop. The immediate workaround
> +  is to use ``clang`` compiler on these platforms.
> +  Initial workaround in DPDK v18.11 was to disable ``AVX-512`` support for ``GCC``
> +  completely, but based on information on defect submitted to GCC community [3],
> +  issue has been identified as ``binutils 2.30`` issue. Since currently only GCC
> +  generates ``AVX-512`` instructions, scope limited to ``GCC`` and ``binutils 2.30``

*the* scope *is* limited

> +
> +  - [1]: Commit ("mk: fix scope of disabling AVX512F support")
> +  - [2]: https://bugs.dpdk.org/show_bug.cgi?id=97
> +  - [3]: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88096
>  

Space missing here

>  Tested Platforms
>  ----------------
> --- a/mk/toolchain/gcc/rte.toolchain-compat.mk
> +++ b/mk/toolchain/gcc/rte.toolchain-compat.mk
> +LD_VERSION = $(shell $(LD) -v)
> +# disable AVX512F support for GCC & binutils 2.30 as a workaround for Bug 97
> +ifneq ($(filter 2.30%,$(LD_VERSION)),)
> +FORCE_DISABLE_AVX512 := y
> +# print warning only once for librte_eal
> +ifneq ($(filter %librte_eal,$(lastword $(CURDIR))),)

Do we need lastword for CURDIR?
  

Patch

diff --git a/config/meson.build b/config/meson.build
index db32499b3..40802fc88 100644
--- a/config/meson.build
+++ b/config/meson.build
@@ -43,6 +43,14 @@  toolchain = cc.get_id()
 dpdk_conf.set_quoted('RTE_TOOLCHAIN', toolchain)
 dpdk_conf.set('RTE_TOOLCHAIN_' + toolchain.to_upper(), 1)
 
+# get binutils version for the workaround of Bug 97
+ldver = run_command('ld', '-v').stdout().strip()
+if ldver.contains('2.30')
+	if cc.has_argument('-mno-avx512f')
+		machine_args += '-mno-avx512f'
+	endif
+endif
+
 add_project_link_arguments('-Wl,--no-as-needed', language: 'c')
 dpdk_extra_ldflags += '-Wl,--no-as-needed'
 
diff --git a/doc/guides/rel_notes/known_issues.rst b/doc/guides/rel_notes/known_issues.rst
index a1face9c1..b0057a2bc 100644
--- a/doc/guides/rel_notes/known_issues.rst
+++ b/doc/guides/rel_notes/known_issues.rst
@@ -826,3 +826,35 @@  Kernel crash when hot-unplug igb_uio device while DPDK application is running
 
 **Driver/Module**:
    ``igb_uio`` module.
+
+
+AVX-512 support disabled
+------------------------
+
+**Description**:
+   ``AVX-512`` support has been disabled on some conditions.
+   This shouldn't be confused with ``CONFIG_RTE_ENABLE_AVX512`` config option which is already
+   disabled by default. This config option defines if ``AVX-512`` specific implementations of
+   some file to be used or not. What has been disabled is compiler feature to produce ``AVX-512``
+   instructions from any source code.
+
+   On DPDK v18.11 ``AVX-512`` disabled for all ``GCC`` builds which reported to cause a performance
+   drop.
+
+   On DPDK v19.02 ``AVX-512`` disable scope reduced to ``GCC`` and ``binutils version 2.30`` based
+   on information accured from the GCC community defect.
+
+**Reason**:
+   Generated ``AVX-512`` code cause crash:
+   https://bugs.dpdk.org/show_bug.cgi?id=97
+   https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88096
+
+**Resolution/Workaround**:
+   Update ``binutils`` to newer version than ``2.30``.
+   Use different compiler, like ``clang`` for this case.
+
+**Affected Environment/Platform**:
+    ``GCC`` and ``binutils version 2.30``.
+
+**Driver/Module**:
+    ALL.
diff --git a/doc/guides/rel_notes/release_19_02.rst b/doc/guides/rel_notes/release_19_02.rst
index 39c6ff826..b6322b282 100644
--- a/doc/guides/rel_notes/release_19_02.rst
+++ b/doc/guides/rel_notes/release_19_02.rst
@@ -276,6 +276,19 @@  Known Issues
    Also, make sure to start the actual text at the margin.
    =========================================================
 
+* ``AVX-512`` support has been disabled for ``GCC`` builds when ``binutils 2.30``
+  is detected [1] because of a crash [2]. This can affect ``native`` machine type
+  build targets on the platforms that support ``AVX512F`` like ``Intel Skylake``
+  processors, and can cause a possible performance drop. The immediate workaround
+  is to use ``clang`` compiler on these platforms.
+  Initial workaround in DPDK v18.11 was to disable ``AVX-512`` support for ``GCC``
+  completely, but based on information on defect submitted to GCC community [3],
+  issue has been identified as ``binutils 2.30`` issue. Since currently only GCC
+  generates ``AVX-512`` instructions, scope limited to ``GCC`` and ``binutils 2.30``
+
+  - [1]: Commit ("mk: fix scope of disabling AVX512F support")
+  - [2]: https://bugs.dpdk.org/show_bug.cgi?id=97
+  - [3]: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88096
 
 Tested Platforms
 ----------------
diff --git a/mk/rte.cpuflags.mk b/mk/rte.cpuflags.mk
index c3291b17a..541211c61 100644
--- a/mk/rte.cpuflags.mk
+++ b/mk/rte.cpuflags.mk
@@ -69,8 +69,8 @@  ifneq ($(filter $(AUTO_CPUFLAGS),__AVX512F__),)
 ifeq ($(CONFIG_RTE_ENABLE_AVX512),y)
 CPUFLAGS += AVX512F
 else
-# disable AVX512F support of gcc as a workaround for Bug 97
-ifeq ($(CONFIG_RTE_TOOLCHAIN_GCC),y)
+# disable AVX512F support for GCC & binutils 2.30 as a workaround for Bug 97
+ifeq ($(FORCE_DISABLE_AVX512),y)
 MACHINE_CFLAGS += -mno-avx512f
 endif
 endif
diff --git a/mk/toolchain/gcc/rte.toolchain-compat.mk b/mk/toolchain/gcc/rte.toolchain-compat.mk
index 44904295c..33ea3f03a 100644
--- a/mk/toolchain/gcc/rte.toolchain-compat.mk
+++ b/mk/toolchain/gcc/rte.toolchain-compat.mk
@@ -20,6 +20,16 @@  HOST_GCC_MINOR = $(shell echo __GNUC_MINOR__ | $(HOSTCC) -E -x c - | tail -n 1)
 HOST_GCC_PATCHLEVEL = $(shell echo __GNUC_PATCHLEVEL__ | $(HOSTCC) -E -x c - | tail -n 1)
 HOST_GCC_VERSION = $(HOST_GCC_MAJOR)$(HOST_GCC_MINOR)
 
+LD_VERSION = $(shell $(LD) -v)
+# disable AVX512F support for GCC & binutils 2.30 as a workaround for Bug 97
+ifneq ($(filter 2.30%,$(LD_VERSION)),)
+FORCE_DISABLE_AVX512 := y
+# print warning only once for librte_eal
+ifneq ($(filter %librte_eal,$(lastword $(CURDIR))),)
+$(warning AVX512 support disabled because of ld 2.30. See Bug 97)
+endif
+endif
+
 # if GCC is older than 4.x
 ifeq ($(shell test $(GCC_VERSION) -lt 40 && echo 1), 1)
 	MACHINE_CFLAGS =