Skip to contentSkip to conversation

linux-user/alpha: populate AT_HWCAP from env->amask

PATCH In review Matt Turner 3 participants 4 messages Jul 13 – Jul 16

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 linux-user/alpha: populate AT_HWCAP from env->amask

Alpha has never set AT_HWCAP in linux-user emulation, so getauxval(AT_HWCAP) always returned 0 regardless of the emulated CPU model.

The Linux kernel computes ELF_HWCAP as ~amask(-1), i.e. the set of ISA extension bits that the amask instruction reports as supported (cleared in its output). env->amask stores exactly those bits with the same layout (BWX=0x1, FIX=0x2, CIX=0x4, MVI=0x100, TRAP=0x200, PREFETCH=0x1000), so returning it directly from get_elf_hwcap matches the kernel convention.

Add HAVE_ELF_HWCAP to target_elf.h and implement get_elf_hwcap() in elfload.c to expose the emulated CPU's capability mask to user-space programs via the auxiliary vector.

Without this fix, programs using getauxval(AT_HWCAP) to detect BWX/FIX/CIX (such as glibc's memcpy or JIT compilers targeting Alpha) incorrectly concluded that no extensions were available even when emulating ev56+.

---
 linux-user/alpha/elfload.c    | 11 +++++++++++
 linux-user/alpha/target_elf.h |  1 +
 2 files changed, 12 insertions(+)
linux-user/alpha/elfload.c +11 -0
index 1969f620a5..7be9e466b6 100644
@@ -6,6 +6,17 @@
 #include "target_elf.h"
 
 
+abi_ulong get_elf_hwcap(CPUState *cs)
+{
+    /*
+     * The Linux kernel computes ELF_HWCAP as ~amask(-1), which clears a bit
+     * for each supported ISA extension.  env->amask stores exactly those bits
+     * set for the extensions supported by the emulated CPU model, matching
+     * the kernel's convention: bit set in AT_HWCAP ↔ extension present.
+     */
+    return cpu_env(cs)->amask;
+}
+
 void elf_core_copy_regs(target_elf_gregset_t *r, const CPUAlphaState *env)
 {
     int i;
linux-user/alpha/target_elf.h +1 -0
index 4987ae3944..dd90c6f783 100644
@@ -12,6 +12,7 @@
 #define ELF_MACHINE             EM_ALPHA
 
 #define HAVE_ELF_CORE_DUMP      1
+#define HAVE_ELF_HWCAP          1
 
 /*
  * Matches the kernel's elf_gregset_t (ELF_NGREG = 33):
Signature · 1 line
2.54.0
Reviewed-by on v1 1/1
#1 lore raw
Quoted 13 lines · Matt Turner

Alpha has never set AT_HWCAP in linux-user emulation, so getauxval(AT_HWCAP) always returned 0 regardless of the emulated CPU model.

The Linux kernel computes ELF_HWCAP as ~amask(-1), i.e. the set of ISA extension bits that the amask instruction reports as supported (cleared in its output). env->amask stores exactly those bits with the same layout (BWX=0x1, FIX=0x2, CIX=0x4, MVI=0x100, TRAP=0x200, PREFETCH=0x1000), so returning it directly from get_elf_hwcap matches the kernel convention.

Add HAVE_ELF_HWCAP to target_elf.h and implement get_elf_hwcap() in elfload.c to expose the emulated CPU's capability mask to user-space programs via the auxiliary vector.

Quoted 7 lines

Without this fix, programs using getauxval(AT_HWCAP) to detect BWX/FIX/CIX (such as glibc's memcpy or JIT compilers targeting Alpha) incorrectly concluded that no extensions were available even when emulating ev56+.

---
  linux-user/alpha/elfload.c    | 11 +++++++++++
  linux-user/alpha/target_elf.h |  1 +
  2 files changed, 12 insertions(+)

I'll include this patch in the next linux-user pull request. Helge

#2 lore raw
In reply to Matt Turner
Quoted 9 lines · Matt Turner

Alpha has never set AT_HWCAP in linux-user emulation, so getauxval(AT_HWCAP) always returned 0 regardless of the emulated CPU model.

The Linux kernel computes ELF_HWCAP as ~amask(-1), i.e. the set of ISA extension bits that the amask instruction reports as supported (cleared in its output). env->amask stores exactly those bits with the same layout (BWX=0x1, FIX=0x2, CIX=0x4, MVI=0x100, TRAP=0x200, PREFETCH=0x1000), so returning it directly from get_elf_hwcap matches the kernel convention.

Quoted 7 lines

Add HAVE_ELF_HWCAP to target_elf.h and implement get_elf_hwcap() in elfload.c to expose the emulated CPU's capability mask to user-space programs via the auxiliary vector.

Without this fix, programs using getauxval(AT_HWCAP) to detect BWX/FIX/CIX (such as glibc's memcpy or JIT compilers targeting Alpha) incorrectly concluded that no extensions were available even when emulating ev56+.

In stable 10.0.x qemu series, things are different from what we have now, - there, everything is in single linux-user/elfload.c and the context is a bit different too.

Here's my backport of this change to 10.0.x:
https://gitlab.com/mjt0k/qemu/-/commit/e4c1bf9b1fe030ef1d0d4c74086cccb62c6b0442

I appreciate a quick look at the result, whenever it looks good for 10.0.x or not.

Thanks,

/mjt

Quoted 38 lines
  linux-user/alpha/elfload.c    | 11 +++++++++++
  linux-user/alpha/target_elf.h |  1 +
  2 files changed, 12 insertions(+)
linux-user/alpha/elfload.c +11 -0
index 1969f620a5..7be9e466b6 100644
@@ -6,6 +6,17 @@
  #include "target_elf.h"
 
 
+abi_ulong get_elf_hwcap(CPUState *cs)
+{
+    /*
+     * The Linux kernel computes ELF_HWCAP as ~amask(-1), which clears a bit
+     * for each supported ISA extension.  env->amask stores exactly those bits
+     * set for the extensions supported by the emulated CPU model, matching
+     * the kernel's convention: bit set in AT_HWCAP ↔ extension present.
+     */
+    return cpu_env(cs)->amask;
+}
+
  void elf_core_copy_regs(target_elf_gregset_t *r, const CPUAlphaState *env)
  {
      int i;
linux-user/alpha/target_elf.h +1 -0
index 4987ae3944..dd90c6f783 100644
@@ -12,6 +12,7 @@
  #define ELF_MACHINE             EM_ALPHA
 
  #define HAVE_ELF_CORE_DUMP      1
+#define HAVE_ELF_HWCAP          1
 
  /*
   * Matches the kernel's elf_gregset_t (ELF_NGREG = 33):
#3 lore raw
Quoted 23 lines · Michael Tokarev
Quoted 9 lines · Matt Turner

Alpha has never set AT_HWCAP in linux-user emulation, so getauxval(AT_HWCAP) always returned 0 regardless of the emulated CPU model.

The Linux kernel computes ELF_HWCAP as ~amask(-1), i.e. the set of ISA extension bits that the amask instruction reports as supported (cleared in its output). env->amask stores exactly those bits with the same layout (BWX=0x1, FIX=0x2, CIX=0x4, MVI=0x100, TRAP=0x200, PREFETCH=0x1000), so returning it directly from get_elf_hwcap matches the kernel convention.

Quoted 7 lines

Add HAVE_ELF_HWCAP to target_elf.h and implement get_elf_hwcap() in elfload.c to expose the emulated CPU's capability mask to user-space programs via the auxiliary vector.

Without this fix, programs using getauxval(AT_HWCAP) to detect BWX/FIX/CIX (such as glibc's memcpy or JIT compilers targeting Alpha) incorrectly concluded that no extensions were available even when emulating ev56+.

In stable 10.0.x qemu series, things are different from what we have now, - there, everything is in single linux-user/elfload.c and the context is a bit different too.

Quoted 5 lines

Here's my backport of this change to 10.0.x:
https://gitlab.com/mjt0k/qemu/-/commit/e4c1bf9b1fe030ef1d0d4c74086cccb62c6b0442

I appreciate a quick look at the result, whenever it looks good for 10.0.x or not.

Looks good to me. Thank you!