target/i386: Fixes to the check missing features routine
PATCH v2 1/1 target/i386: Fixes to the check missing features routine
The x86_cpu_class_check_missing_features() returns a list of unavailable features compared to the host CPU. Currently it may return empty strings for unnamed features as well as duplicated names.
For example, the qmp "query-cpu-definitions" below shows one empty string and repeated "mpx" entries:
(...)
{"execute": "query-cpu-definitions"}
(...)
{
"name": "Cascadelake-Server",
"typename": "Cascadelake-Server-x86_64-cpu",
"unavailable-features": [
"hle",
"rtm",
"mpx",
"avx512f",
"avx512dq",
"rdseed",
"adx",
"smap",
"clflushopt",
"clwb",
"intel-pt",
"avx512cd",
"avx512bw",
"avx512vl",
"pku",
"",
"avx512vnni",
"spec-ctrl",
"ssbd",
"3dnowprefetch",
"xsavec",
"xgetbv1",
"mpx",
"mpx",
"avx512f",
"avx512f",
"avx512f",
"pku"
],(...)
---
v2:
* Fixed typos. [eblake]
* Removed unwanted manual test case. [ccarrara, ehabkost]
* Not passing 'accel=kvm' on test's VM. [ehabkost]
* Removed unneeded g_strdup() call. [ehabkost]
* Formatted comment according to QEMU's coding style. [ehabkost]
v1: https://www.mail-archive.com/[email protected]/msg579404.html
--- target/i386/cpu.c | 11 ++++++++- tests/acceptance/cpu_definitions.py | 35 +++++++++++++++++++++++++++++ 2 files changed, 45 insertions(+), 1 deletion(-) create mode 100644 tests/acceptance/cpu_definitions.py
target/i386/cpu.c +10 -1
@@ -3615,19 +3615,28 @@ static void x86_cpu_class_check_missing_features(X86CPUClass *xcc,x86_cpu_filter_features (xc); +/* Auxiliary dictionary to avoid duplicate entries in the list. */ +QDict *unique_feats_dict =qdict_new (); +for (w =0 ; w <FEATURE_WORDS ; w++) {uint32_t filtered = xc->filtered_features[w];int i;for (i =0 ; i <32 ; i++) {if (filtered & (1UL << i)) { +const char *fname =x86_cpu_feature_name (w, i); +if (!fname ||qdict_haskey (unique_feats_dict, fname)) { +continue ; + }strList *new =g_new0 (strList,1 ); - new->value =g_strdup (x86_cpu_feature_name (w, i)); + new->value =g_strdup (fname); *next = new; next = &new->next; +qdict_put_null (unique_feats_dict, new->value); } } } +g_free (unique_feats_dict);object_unref (OBJECT (xc)); }
tests/acceptance/cpu_definitions.py new file +35 -0
@@ -0,0 +1,35 @@ +# CPU definitions tests. +# +# Copyright (c) 2018 Red Hat, Inc. +# +# Author: +# Wainer dos Santos Moschetta <[email protected]> +# +# This work is licensed under the terms of the GNU GPL, version 2 or +# later. See the COPYING file in the top-level directory. + +from avocado import skip +from avocado_qemu import Test + + +class CPUDefinitions(Test): + """ + Tests for the CPU definitions. + + :avocado: enable + :avocado: tags=x86_64 + """ + def test_unavailable_features(self): + self.vm.add_args("-machine", "q35") + self.vm.launch() + cpu_definitions = self.vm.command('query-cpu-definitions') + self.assertTrue(len(cpu_definitions) > 0) + for cpu_model in cpu_definitions: + name = cpu_model.get('name') + unavailable_features = cpu_model.get('unavailable-features') + + self.assertNotIn("", unavailable_features, + name + " has unamed feature") + self.assertEqual(len(unavailable_features), + len(set(unavailable_features)), + name + " has duplicate feature")
Signature · 1 line
2.19.1
Quoted 5 lines · Wainer dos Santos Moschetta
The
x86_cpu_class_check_missing_features()returns a list of unavailable features compared to the host CPU. Currently it may return empty strings for unnamed features as well as duplicated names.
Quoted 7 lines
For example, the qmp "query-cpu-definitions" below shows one empty string and repeated "mpx" entries:
Signed-off-by Wainer dos Santos Moschetta <[email protected]>Reviewed-by Eduardo Habkost <[email protected]>Reviewed-by Eric Blake <[email protected]>
Careful. While I spotted typos in v1,...
Quoted 4 lines
Reviewed-by Caio Carrara <[email protected]>---
v2:
* Fixed typos. [eblake]
...and you indeed addressed them, me pointing out typos does not imply that I reviewed the patch for correctness. In fact, I specifically did NOT give my R-by: tag to v1, because I'm not (yet?) familiar enough with the tests/acceptance/ framework to state that I have fully reviewed the patch for correctness; instead, I'm comfortable relying on the reviews of others (and I am again intentionally not giving R-by to v2).
Also, when posting a v2, you should include the R-by actually given to v1 only if the patch is roughly the same as the original. Fixing minor issues that a reviewer pointed out, or doing obvious rebasing to changes applied earlier in the series or on upstream git, but where the algorithm of the patch itself did not change, is okay for keeping R-b (so if I _had_ given R-b, and your spelling changes were the only difference, then keeping my R-b would make sense); but where the patch is fundamentally different, such as:
Quoted 2 lines
* Removed unwanted manual test case. [ccarrara, ehabkost]
* Not passing 'accel=kvm' on test's VM. [ehabkost]
then omitting ALL R-by tags, in order to ensure that reviewers check that the new patch is still correct, is a wiser course of action. Yes, this is more of a rule of thumb, and there are cases where keeping or dropping R-b is more of an art form than an exact science; but hopefully this helps you understand how the tag can be useful for iterative reviews.
Signature · 3 lines
Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3266 Virtualization: qemu.org | libvirt.org
Quoted 21 lines · Eric Blake
Quoted 5 lines · Wainer dos Santos Moschetta
The
x86_cpu_class_check_missing_features()returns a list of unavailable features compared to the host CPU. Currently it may return empty strings for unnamed features as well as duplicated names.Quoted 7 lines
For example, the qmp "query-cpu-definitions" below shows one empty string and repeated "mpx" entries:
Signed-off-by Wainer dos Santos Moschetta <[email protected]>Reviewed-by Eduardo Habkost <[email protected]>Reviewed-by Eric Blake <[email protected]>Careful. While I spotted typos in v1,...
Quoted 4 lines
Reviewed-by Caio Carrara <[email protected]>---
v2:
* Fixed typos. [eblake]
Quoted 9 lines
...and you indeed addressed them, me pointing out typos does not imply that I reviewed the patch for correctness. In fact, I specifically did NOT give my R-by: tag to v1, because I'm not (yet?) familiar enough with the tests/acceptance/ framework to state that I have fully reviewed the patch for correctness; instead, I'm comfortable relying on the reviews of others (and I am again intentionally not giving R-by to v2).
Also, when posting a v2, you should include the R-by actually given to v1 only if the patch is roughly the same as the original. Fixing minor issues that a reviewer pointed out, or doing obvious rebasing to changes applied earlier in the series or on upstream git, but where the algorithm of the patch itself did not change, is okay for keeping R-b (so if I _had_ given R-b, and your spelling changes were the only difference, then keeping my R-b would make sense); but where the patch is fundamentally different, such as:
Quoted 2 lines
* Removed unwanted manual test case. [ccarrara, ehabkost]
* Not passing 'accel=kvm' on test's VM. [ehabkost]then omitting ALL R-by tags, in order to ensure that reviewers check that the new patch is still correct, is a wiser course of action. Yes, this is more of a rule of thumb, and there are cases where keeping or dropping R-b is more of an art form than an exact science; but hopefully this helps you understand how the tag can be useful for iterative reviews.
Hi Eric,
Yes, it helped a lot, thanks. And I apologize for my mistake, I'm gonna send a v3 fixing it.
Another doubt that I have: is it advisable to CC everyone that reviewed (with or without R-by) the previous version of my patch?
Thanks!
- Wainer
Quoted 2 lines · Wainer dos Santos Moschetta
Yes, it helped a lot, thanks. And I apologize for my mistake, I'm gonna send a v3 fixing it.
You may want to wait a day or so for any other comments on v2, to minimize the resend churn. A maintainer can fix up tags, particularly when they are aware it is from a newer contributor still learning how things work.
Quoted 2 lines
Another doubt that I have: is it advisable to CC everyone that reviewed (with or without R-by) the previous version of my patch?
CC'ing previous reviewers is generally a reasonable idea, since they are then more likely to double-check that the things they pointed out in the first version are indeed fixed in the respin (and since without the cc, it's a lot easier to miss that a respin is even available on-list for followup review, thanks to the list traffic volume). You might not get a reply from everyone cc'd, but that's not fatal to acceptance of the patch.
Signature · 3 lines
Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3266 Virtualization: qemu.org | libvirt.org
Quoted 23 lines · Wainer dos Santos Moschetta
The
x86_cpu_class_check_missing_features()returns a list of unavailable features compared to the host CPU. Currently it may return empty strings for unnamed features as well as duplicated names.For example, the qmp "query-cpu-definitions" below shows one empty string and repeated "mpx" entries:
(...)
{"execute": "query-cpu-definitions"}
(...){ "name": "Cascadelake-Server", "typename": "Cascadelake-Server-x86_64-cpu", "unavailable-features": [ "hle", "rtm", "mpx", "avx512f", "avx512dq", "rdseed", "adx", "smap",
Quoted 8 lines
"clflushopt",
"clwb",
"intel-pt",
"avx512cd",
"avx512bw",
"avx512vl",
"pku",
"",I just noticed one thing: we probably want to find out the cause of this empty entry, instead of ignoring it in the code. Named CPU models must only refer to named CPU features.
I think this is caused by CPUID_7_0_ECX_OSPKE, I will investigate. But this doesn't make your patch incorrect.
Quoted 45 lines
"avx512vnni", "spec-ctrl", "ssbd", "3dnowprefetch", "xsavec", "xgetbv1", "mpx", "mpx", "avx512f", "avx512f", "avx512f", "pku" ],(...)
Signed-off-by Wainer dos Santos Moschetta <[email protected]>Reviewed-by Eduardo Habkost <[email protected]>Reviewed-by Eric Blake <[email protected]>Reviewed-by Caio Carrara <[email protected]>---
v2:
* Fixed typos. [eblake]
* Removed unwanted manual test case. [ccarrara, ehabkost]
* Not passing 'accel=kvm' on test's VM. [ehabkost]
* Removed unneededg_strdup()call. [ehabkost]
* Formatted comment according to QEMU's coding style. [ehabkost]v1: https://www.mail-archive.com/[email protected]/msg579404.html
--- target/i386/cpu.c | 11 ++++++++- tests/acceptance/cpu_definitions.py | 35 +++++++++++++++++++++++++++++ 2 files changed, 45 insertions(+), 1 deletion(-) create mode 100644 tests/acceptance/cpu_definitions.pytarget/i386/cpu.c +3
-0@@ -3615,19 +3615,28 @@ static void x86_cpu_class_check_missing_features(X86CPUClass *xcc,x86_cpu_filter_features (xc); +/* Auxiliary dictionary to avoid duplicate entries in the list. */ +QDict *unique_feats_dict =qdict_new (); +
target/i386/cpu.c static void x86_cpu_class_check_missing_features(X86CPUClass *xcc, · line 3630
Quoted 10 lines
target/i386/cpu.c +4
-0for (w =0 ; w <FEATURE_WORDS ; w++) {uint32_t filtered = xc->filtered_features[w];int i;for (i =0 ; i <32 ; i++) {if (filtered & (1UL << i)) { +const char *fname =x86_cpu_feature_name (w, i); +if (!fname ||qdict_haskey (unique_feats_dict, fname)) { +continue ; + }strList *new =g_new0 (strList,1 );
I like mixed declarations, but unfortunately they are not allowed by CODING_STYLE:
5. Declarations
Mixed declarations (interleaving statements and declarations within blocks) are generally not allowed; declarations should be at the beginning of blocks.
The logic now looks good, though.
Quoted 32 lines
target/i386/cpu.c +3
-1- new->value =g_strdup (x86_cpu_feature_name (w, i)); + new->value =g_strdup (fname); *next = new; next = &new->next; +qdict_put_null (unique_feats_dict, new->value); } } } +g_free (unique_feats_dict);object_unref (OBJECT (xc)); }tests/acceptance/cpu_definitions.py new file +13
-0@@ -0,0 +1,35 @@ +# CPU definitions tests. +# +# Copyright (c) 2018 Red Hat, Inc. +# +# Author: +# Wainer dos Santos Moschetta <[email protected]> +# +# This work is licensed under the terms of the GNU GPL, version 2 or +# later. See the COPYING file in the top-level directory. + +from avocado import skip +from avocado_qemu import Test +
tests/acceptance/cpu_definitions.py line 23
Quoted 10 lines
tests/acceptance/cpu_definitions.py new file +10
-0+ +class CPUDefinitions(Test): + """ + Tests for the CPU definitions. + + :avocado: enable + :avocado: tags=x86_64 + """ + def test_unavailable_features(self): + self.vm.add_args("-machine", "q35")
I thought the explicit -machine option was here only because of the old accel=kvm option, and the whole line would be removed. Why do you need to explicitly ask for a Q35 machine to test this?
tests/acceptance/cpu_definitions.py line 32
Quoted 9 lines
tests/acceptance/cpu_definitions.py new file +9
-0+ self.vm.launch() + cpu_definitions = self.vm.command('query-cpu-definitions') + self.assertTrue(len(cpu_definitions) > 0) + for cpu_model in cpu_definitions: + name = cpu_model.get('name') + unavailable_features = cpu_model.get('unavailable-features') + + self.assertNotIn("", unavailable_features, + name + " has unamed feature")
"unnamed"
Quoted 6 lines
tests/acceptance/cpu_definitions.py new file +3
-0+ self.assertEqual(len(unavailable_features), + len(set(unavailable_features)), + name + " has duplicate feature")Signature · 1 line
2.19.1
Signature · 1 line
Eduardo