Skip to contentSkip to conversation

Misc fixes and CI improvements

PATCH New Thomas Huth 1 participant 12 messages May 14, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover Misc fixes and CI improvements

Here's a set of accumulated patches that fix the various problems with our CI pipelines, and then update the Gitlab-CI to use Fedora 32 instead of 30. Additionally, the new version of Clang in Fedora 32 is finally also able to compile the kvm-unit-tests (with some small fixes included in this series), so we can now also add a CI test with this compiler, too.

Andrew Jones (1):
Fix out-of-tree builds

Bill Wendling (2):
x86: use a non-negative number in shift
x86: use inline asm to retrieve stack pointer

Mohammed Gamal (1):
x86/access: Fix phys-bits parameter

Paolo Bonzini (1):
x86: avoid multiply defined symbol

Thomas Huth (6):
Fixes for the umip test
Always compile the kvm-unit-tests with -fno-common Fix powerpc issue with the linker from Fedora 32 Update the gitlab-ci to Fedora 32
vmx_tests: Silence warning from Clang Compile the kvm-unit-tests also with Clang

 .gitlab-ci.yml       | 17 +++++++++++++++--
 Makefile             |  2 +-
 configure            |  8 +++-----
 lib/auxinfo.h        |  3 +--
 lib/x86/fault_test.c |  2 +-
 lib/x86/usermode.c   |  2 +-
 powerpc/flat.lds     | 19 ++++++++++++++++---
 x86/Makefile.common  |  1 +
 x86/svm_tests.c      |  2 +-
 x86/umip.c           |  6 ++++--
 x86/unittests.cfg    |  2 +-
 x86/vmx_tests.c      | 10 +++++++---
 12 files changed, 52 insertions(+), 22 deletions(-)
Signature · 1 line
2.18.1
#1 lore raw

PATCH 1/11 x86/access: Fix phys-bits parameter

From: Mohammed Gamal <[email protected]>

Some QEMU versions don't support setting phys-bits argument directly. This causes breakage to Travis CI. Work around the bug by setting host-phys-bits=on

Fixes 1a296ac170f ("x86: access: Add tests for reserved bits of guest physical address")
---
 x86/unittests.cfg | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
x86/unittests.cfg +1 -1
index bf0d02e..d43cac2 100644
@@ -116,7 +116,7 @@ extra_params = -cpu qemu64,+x2apic,+tsc-deadline -append tscdeadline_immed
 [access]
 file = access.flat
 arch = x86_64
-extra_params = -cpu host,phys-bits=36
+extra_params = -cpu host,host-phys-bits=on,phys-bits=36
 
 [smap]
 file = smap.flat
Signature · 1 line
2.18.1
#2 lore raw

PATCH 2/11 Fix out-of-tree builds

In reply to Thomas Huth

From: Andrew Jones <[email protected]>

Since b16df9ee5f3b out-of-tree builds have been broken because we started validating the newly user-configurable $erratatxt file before linking it into the build dir. We fix this not by moving the validation, but by removing the linking and instead using the full path of the $erratatxt file. This allows one to keep that file separate from the src and build dirs.

Fixes b16df9ee5f3b ("arch-run: Add reserved variables to the default environ")
---
 configure | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
configure +3 -5
index 5d2cd90..f9d030f 100755
@@ -17,7 +17,7 @@ environ_default=yes
 u32_long=
 vmm="qemu"
 errata_force=0
-erratatxt="errata.txt"
+erratatxt="$srcdir/errata.txt"
 
 usage() {
     cat <<-EOF
@@ -89,7 +89,8 @@ while [[ "$1" = -* ]]; do
 	    environ_default=no
 	    ;;
 	--erratatxt)
-	    erratatxt="$arg"
+	    erratatxt=
+	    [ "$arg" ] && erratatxt=$(eval realpath "$arg")
 	    ;;
 	--help)
 	    usage
@@ -169,9 +170,6 @@ if test ! -e Makefile; then
 
     echo "linking scripts..."
     ln -sf "$srcdir/scripts"
-
-    echo "linking errata.txt..."
-    ln -sf "$srcdir/errata.txt"
 fi
 
 # link lib/asm for the architecture
Signature · 1 line
2.18.1
#3 lore raw

PATCH 3/11 x86: avoid multiply defined symbol

In reply to Thomas Huth

From: Paolo Bonzini <[email protected]>

Fedora 32 croaks about a symbol that is defined twice, fix it.

---
 lib/x86/fault_test.c | 2 +-
 lib/x86/usermode.c   | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
lib/x86/fault_test.c +1 -1
index 078dae3..e15a218 100644
@@ -1,6 +1,6 @@
 #include "fault_test.h"
 
-jmp_buf jmpbuf;
+static jmp_buf jmpbuf;
 
 static void restore_exec_to_jmpbuf(void)
 {
lib/x86/usermode.c +1 -1
index f01ad9b..f032523 100644
@@ -14,7 +14,7 @@
 #define USERMODE_STACK_SIZE	0x2000
 #define RET_TO_KERNEL_IRQ	0x20
 
-jmp_buf jmpbuf;
+static jmp_buf jmpbuf;
 
 static void restore_exec_to_jmpbuf(void)
 {
Signature · 1 line
2.18.1
#4 lore raw

PATCH 4/11 Fixes for the umip test

In reply to Thomas Huth

When compiling umip.c with -O2 instead of -O1, there are currently two problems. First, the compiler complains:

x86/umip.c: In function ‘do_ring3’:
x86/umip.c:162:37: error: array subscript 4096 is above array bounds of

    ‘unsigned char[4096]’ [-Werror=array-bounds]
       [user_stack_top]"m"(user_stack[sizeof user_stack]),
                           ~~~~~~~~~~^~~~~~~~~~~~~~~~~~~

This can be fixed by initializing the stack to point to one of the last bytes of the array instead.

The second problem is that some tests are failing - and this is due to the fact that the GP_ASM macro uses inline asm without the "volatile" keyword - so that the compiler reorders this code in certain cases where it should not. Fix it by adding "volatile" here.

Message-Id: <[email protected]>

---
 x86/umip.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
x86/umip.c +4 -2
index 7eee294..37f1ab8 100644
@@ -22,7 +22,8 @@ static void gp_handler(struct ex_regs *regs)
 
 
 #define GP_ASM(stmt, in, clobber)                  \
-     asm ("mov" W " $1f, %[expected_rip]\n\t"      \
+    asm volatile (                                 \
+          "mov" W " $1f, %[expected_rip]\n\t"      \
           "movl $2f-1f, %[skip_count]\n\t"         \
           "1: " stmt "\n\t"                        \
           "2: "                                    \
@@ -159,7 +160,8 @@ static int do_ring3(void (*fn)(const char *), const char *arg)
 		  : [ret] "=&a" (ret)
 		  : [user_ds] "i" (USER_DS),
 		    [user_cs] "i" (USER_CS),
-		    [user_stack_top]"m"(user_stack[sizeof user_stack]),
+		    [user_stack_top]"m"(user_stack[sizeof(user_stack) -
+						   sizeof(long)]),
 		    [fn]"r"(fn),
 		    [arg]"D"(arg),
 		    [kernel_ds]"i"(KERNEL_DS),
Signature · 1 line
2.18.1
#5 lore raw

PATCH 5/11 Always compile the kvm-unit-tests with -fno-common

In reply to Thomas Huth

The new GCC v10 uses -fno-common by default. To avoid that we commit code that declares global variables twice and thus fails to link with the latest version, we should also compile with -fno-common when using older versions of the compiler. However, this now also means that we can not play the trick with the common auxinfo struct anymore. Thus declare it as extern in the header now and link auxinfo.c on x86, too.

---
 Makefile            | 2 +-
 lib/auxinfo.h       | 3 +--
 x86/Makefile.common | 1 +
 3 files changed, 3 insertions(+), 3 deletions(-)
Makefile +1 -1
index 754ed65..3ff2f91 100644
@@ -49,7 +49,7 @@ include $(SRCDIR)/$(TEST_DIR)/Makefile
 cc-option = $(shell if $(CC) -Werror $(1) -S -o /dev/null -xc /dev/null \
               > /dev/null 2>&1; then echo "$(1)"; else echo "$(2)"; fi ;)
 
-COMMON_CFLAGS += -g $(autodepend-flags) -fno-strict-aliasing
+COMMON_CFLAGS += -g $(autodepend-flags) -fno-strict-aliasing -fno-common
 COMMON_CFLAGS += -Wall -Wwrite-strings -Wempty-body -Wuninitialized
 COMMON_CFLAGS += -Wignored-qualifiers -Werror
 
lib/auxinfo.h +1 -2
index 08b96f8..a46a1e6 100644
@@ -13,7 +13,6 @@ struct auxinfo {
 	unsigned long flags;
 };
 
-/* No extern!  Define a common symbol.  */
-struct auxinfo auxinfo;
+extern struct auxinfo auxinfo;
 #endif
 #endif
x86/Makefile.common +1 -0
index ab67ca0..2ea9c9f 100644
@@ -5,6 +5,7 @@ all: directories test_cases
 cflatobjs += lib/pci.o
 cflatobjs += lib/pci-edu.o
 cflatobjs += lib/alloc.o
+cflatobjs += lib/auxinfo.o
 cflatobjs += lib/vmalloc.o
 cflatobjs += lib/alloc_page.o
 cflatobjs += lib/alloc_phys.o
Signature · 1 line
2.18.1
#6 lore raw

PATCH 6/11 Fix powerpc issue with the linker from Fedora 32

In reply to Thomas Huth

The linker from Fedora 32 complains:

powerpc64-linux-gnu-ld: powerpc/selftest.elf: error: PHDR segment not
covered by LOAD segment

Let's introduce some fake PHDRs to the linker script to get this working again.

---
 powerpc/flat.lds | 19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)
powerpc/flat.lds +16 -3
index 53221e8..5eed368 100644
@@ -1,7 +1,17 @@
 
+PHDRS
+{
+    text PT_LOAD FLAGS(5);
+    data PT_LOAD FLAGS(6);
+}
+
 SECTIONS
 {
-    .text : { *(.init) *(.text) *(.text.*) }
+    .text : {
+        *(.init)
+        *(.text)
+        *(.text.*)
+    } :text
     . = ALIGN(64K);
     etext = .;
     .opd : { *(.opd) }
@@ -19,9 +29,12 @@ SECTIONS
     .data : {
         *(.data)
         *(.data.rel*)
-    }
+    } :data
     . = ALIGN(16);
-    .rodata : { *(.rodata) *(.rodata.*) }
+    .rodata : {
+        *(.rodata)
+        *(.rodata.*)
+    } :data
     . = ALIGN(16);
     .bss : { *(.bss) }
     . = ALIGN(256);
Signature · 1 line
2.18.1
#7 lore raw

PATCH 7/11 Update the gitlab-ci to Fedora 32

In reply to Thomas Huth

Fedora 30 is end of life, let's use the version 32 instead.

Unfortunately, we have to disable taskswitch2 in the gitlab-ci now. It does not seem to work anymore with the latest version of gcc and/or QEMU. We still check it in the travis-ci, though, so until somebody has some spare time to debug this issue, it should be ok to disable it here.

---
 .gitlab-ci.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
.gitlab-ci.yml +2 -2
index 3093239..13e1a1f 100644
@@ -1,4 +1,4 @@
-image: fedora:30
+image: fedora:32
 
 before_script:
  - dnf update -y
@@ -77,6 +77,6 @@ build-i386:
  - ./configure --arch=i386
  - make -j2
  - ACCEL=tcg ./run_tests.sh
-     cmpxchg8b eventinj port80 setjmp sieve tsc taskswitch taskswitch2 umip
+     cmpxchg8b eventinj port80 setjmp sieve tsc taskswitch umip
      | tee results.txt
  - if grep -q FAIL results.txt ; then exit 1 ; fi
Signature · 1 line
2.18.1
#8 lore raw

PATCH 8/11 x86: use a non-negative number in shift

In reply to Thomas Huth

From: Bill Wendling <[email protected]>

Shifting a negative number is undefined. Clang complains about it:

x86/svm.c:1131:38: error: shifting a negative signed value is undefined [-Werror,-Wshift-negative-value]

    test->vmcb->control.tsc_offset = TSC_OFFSET_VALUE;

Using "~0ull" results in identical asm code:

	before: movabsq $-281474976710656, %rsi
	after:  movabsq $-281474976710656, %rsi
[thuth: Rebased to master - code is in svm_tests.c instead of svm.c now]
---
 x86/svm_tests.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
x86/svm_tests.c +1 -1
index 2f53b8f..be1bddf 100644
@@ -844,7 +844,7 @@ static bool npt_rw_l1mmio_check(struct svm_test *test)
 }
 
 #define TSC_ADJUST_VALUE    (1ll << 32)
-#define TSC_OFFSET_VALUE    (-1ll << 48)
+#define TSC_OFFSET_VALUE    (~0ull << 48)
 static bool ok;
 
 static void tsc_adjust_prepare(struct svm_test *test)
Signature · 1 line
2.18.1
#9 lore raw

PATCH 9/11 x86: use inline asm to retrieve stack pointer

In reply to Thomas Huth

From: Bill Wendling <[email protected]>

According to GCC's documentation, the only supported use for specifying registers for local variables is "to specify registers for input and output operands when calling Extended asm." Using it as a shortcut to get the value in a register isn't guaranteed to work, and clang complains that the variable is uninitialized.

Message-Id: <[email protected]>

---
 x86/vmx_tests.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
x86/vmx_tests.c +6 -2
index 0909adb..dbf5375 100644
@@ -2135,7 +2135,9 @@ static void into_guest_main(void)
 		.offset = (uintptr_t)&&into,
 		.selector = KERNEL_CS32,
 	};
-	register uintptr_t rsp asm("rsp");
+	uintptr_t rsp;
+
+	asm volatile ("mov %%rsp, %0" : "=r"(rsp));
 
 	if (fp.offset != (uintptr_t)&&into) {
 		printf("Code address too high.\n");
@@ -3231,7 +3233,9 @@ static void try_compat_invvpid(void *unused)
 		.offset = (uintptr_t)&&invvpid,
 		.selector = KERNEL_CS32,
 	};
-	register uintptr_t rsp asm("rsp");
+	uintptr_t rsp;
+
+	asm volatile ("mov %%rsp, %0" : "=r"(rsp));
 
 	TEST_ASSERT_MSG(fp.offset == (uintptr_t)&&invvpid,
 			"Code address too high.");
Signature · 1 line
2.18.1
#10 lore raw

PATCH 10/11 vmx_tests: Silence warning from Clang

In reply to Thomas Huth

Clang complains:

x86/vmx_tests.c:8429:40: error: converting the result of '<<' to a boolean
always evaluates to true [-Werror,-Wtautological-constant-compare]

         vmx_preemption_timer_zero_inject_db(1 << DB_VECTOR);
                                               ^

Looking at the code, the "1 << DB_VECTOR" is done within the function vmx_preemption_timer_zero_inject_db() indeed:

	vmcs_write(EXC_BITMAP, intercept_db ? 1 << DB_VECTOR : 0);

... so using "true" as parameter for the function should be appropriate here.

---
 x86/vmx_tests.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
x86/vmx_tests.c +1 -1
index dbf5375..c38115c 100644
@@ -8426,7 +8426,7 @@ static void vmx_preemption_timer_zero_test(void)
 	 * an event that you injected.
 	 */
 	vmx_set_test_stage(1);
-	vmx_preemption_timer_zero_inject_db(1 << DB_VECTOR);
+	vmx_preemption_timer_zero_inject_db(true);
 	vmx_preemption_timer_zero_expect_preempt_at_rip(db_fault_address);
 	vmx_preemption_timer_zero_advance_past_vmcall();
 
Signature · 1 line
2.18.1
#11 lore raw

PATCH 11/11 Compile the kvm-unit-tests also with Clang

In reply to Thomas Huth

To get some more test coverage, let's check compilation with Clang, too.

---
 .gitlab-ci.yml | 13 +++++++++++++
 1 file changed, 13 insertions(+)
.gitlab-ci.yml +13 -0
index 13e1a1f..3af53f0 100644
@@ -80,3 +80,16 @@ build-i386:
      cmpxchg8b eventinj port80 setjmp sieve tsc taskswitch umip
      | tee results.txt
  - if grep -q FAIL results.txt ; then exit 1 ; fi
+
+build-clang:
+ script:
+ - dnf install -y qemu-system-x86 clang
+ - ./configure --arch=x86_64 --cc=clang
+ - make -j2
+ - ACCEL=tcg ./run_tests.sh
+     smptest smptest3 vmexit_cpuid vmexit_mov_from_cr8
+     vmexit_mov_to_cr8 vmexit_inl_pmtimer  vmexit_ipi vmexit_ipi_halt
+     vmexit_ple_round_robin vmexit_tscdeadline vmexit_tscdeadline_immed
+     eventinj msr port80 setjmp syscall tsc rmap_chain umip intel_iommu
+     | tee results.txt
+ - grep -q PASS results.txt && ! grep -q FAIL results.txt
Signature · 1 line
2.18.1