Skip to contentSkip to conversation

RISC-V: Mask out the F extension on systems without D

PATCH In review Palmer Dabbelt 2 participants 3 messages Aug 27, 2018 – Aug 28, 2018 RISC-V

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 RISC-V: Mask out the F extension on systems without D

The RISC-V Linux port doesn't support systems that have the F extension but don't have the D extension -- we actually don't support systems without D either, but Alan's patch set is rectifying that soon. For now I think we can leave this in a semi-broken state and just wait for Alan's patch set to get merged for proper non-FPU support -- the patch set is starting to look good, so doing something in-between doesn't seem like it's worth the work.

I don't think it's worth fretting about support for systems with F but not D for now: our glibc ABIs are IMAC and IMAFDC so they probably won't end up being popular. We can always extend this in the future.

---
 arch/riscv/kernel/cpufeature.c | 7 +++++++
 1 file changed, 7 insertions(+)
arch/riscv/kernel/cpufeature.c +7 -0
index 17011a870044..652d102ffa06 100644
@@ -57,5 +57,12 @@ void riscv_fill_hwcap(void)
 	for (i = 0; i < strlen(isa); ++i)
 		elf_hwcap |= isa2hwcap[(unsigned char)(isa[i])];
 
+	/* We don't support systems with F but without D, so mask those out
+	 * here. */
+	if ((elf_hwcap & COMPAT_HWCAP_ISA_F) && !(elf_hwcap & COMPAT_HWCAP_ISA_D)) {
+		pr_info("This kernel does not support systems with F but not D");
+		elf_hwcap &= ~COMPAT_HWCAP_ISA_F;
+	}
+
 	pr_info("elf_hwcap is 0x%lx", elf_hwcap);
 }
Signature · 1 line
2.16.4
#1 lore raw

Hi Palmer,

Quoted 23 lines · Palmer Dabbelt

The RISC-V Linux port doesn't support systems that have the F extension but don't have the D extension -- we actually don't support systems without D either, but Alan's patch set is rectifying that soon. For now I think we can leave this in a semi-broken state and just wait for Alan's patch set to get merged for proper non-FPU support -- the patch set is starting to look good, so doing something in-between doesn't seem like it's worth the work.

I don't think it's worth fretting about support for systems with F but not D for now: our glibc ABIs are IMAC and IMAFDC so they probably won't end up being popular. We can always extend this in the future.

---
 arch/riscv/kernel/cpufeature.c | 7 +++++++
 1 file changed, 7 insertions(+)
arch/riscv/kernel/cpufeature.c +0 -0
index 17011a870044..652d102ffa06 100644

arch/riscv/kernel/cpufeature.c void riscv_fill_hwcap(void) · line 66

Quoted 10 lines
arch/riscv/kernel/cpufeature.c +7 -0
@@ -57,5 +57,12 @@ void riscv_fill_hwcap(void)
 	for (i = 0; i < strlen(isa); ++i)
 		elf_hwcap |= isa2hwcap[(unsigned char)(isa[i])];
 
+	/* We don't support systems with F but without D, so mask those out
+	 * here. */
+	if ((elf_hwcap & COMPAT_HWCAP_ISA_F) && !(elf_hwcap & COMPAT_HWCAP_ISA_D)) {
+		pr_info("This kernel does not support systems with F but not D");
+		elf_hwcap &= ~COMPAT_HWCAP_ISA_F;
+	}
+

The commit message does address the problem and this patch does provide checks and helpful information to users, but I wonder if we really need this patch, for two reasons:

* Just as you mentioned, current glibc ABI does not support such a thing as IMAFC, so probably no one has had trouble with this. To be honest, I suppose that anybody (RISC-V enthusiasts or vendors) who really need F-only support in kernel should get themself involved in the development by sending patches to improve.

* There are corner cases to let a F-only machine to pass the check in this patch. For instance, a vendor decides to name her extension ISA as doom, and supports single-precision FP only, so her ISA string would be

IMAFCXdoom.

The variable elf_hwcap is calculated at the loop in line 57,58, the 'd' from Xdoom would bypass the check, while the underlying machine does not support double-precision FP.

Quoted 5 lines
arch/riscv/kernel/cpufeature.c +0 -0
 	pr_info("elf_hwcap is 0x%lx", elf_hwcap);
 }
Signature · 1 line
2.16.4

I don't know if the reasons make sense to you, but anyway that's all I would like to say about this patch.

Alan

#2 lore raw
Quoted 51 lines · alankao at andestech.com

Hi Palmer,

Quoted 23 lines · Palmer Dabbelt

The RISC-V Linux port doesn't support systems that have the F extension but don't have the D extension -- we actually don't support systems without D either, but Alan's patch set is rectifying that soon. For now I think we can leave this in a semi-broken state and just wait for Alan's patch set to get merged for proper non-FPU support -- the patch set is starting to look good, so doing something in-between doesn't seem like it's worth the work.

I don't think it's worth fretting about support for systems with F but not D for now: our glibc ABIs are IMAC and IMAFDC so they probably won't end up being popular. We can always extend this in the future.

---
 arch/riscv/kernel/cpufeature.c | 7 +++++++
 1 file changed, 7 insertions(+)
arch/riscv/kernel/cpufeature.c +0 -0
index 17011a870044..652d102ffa06 100644

arch/riscv/kernel/cpufeature.c void riscv_fill_hwcap(void) · line 66

Quoted 10 lines
arch/riscv/kernel/cpufeature.c +7 -0
@@ -57,5 +57,12 @@ void riscv_fill_hwcap(void)
 	for (i = 0; i < strlen(isa); ++i)
 		elf_hwcap |= isa2hwcap[(unsigned char)(isa[i])];
 
+	/* We don't support systems with F but without D, so mask those out
+	 * here. */
+	if ((elf_hwcap & COMPAT_HWCAP_ISA_F) && !(elf_hwcap & COMPAT_HWCAP_ISA_D)) {
+		pr_info("This kernel does not support systems with F but not D");
+		elf_hwcap &= ~COMPAT_HWCAP_ISA_F;
+	}
+

The commit message does address the problem and this patch does provide checks and helpful information to users, but I wonder if we really need this patch, for two reasons:

* Just as you mentioned, current glibc ABI does not support such a thing as IMAFC, so probably no one has had trouble with this. To be honest, I suppose that anybody (RISC-V enthusiasts or vendors) who really need F-only support in kernel should get themself involved in the development by sending patches to improve.

* There are corner cases to let a F-only machine to pass the check in this patch. For instance, a vendor decides to name her extension ISA as doom, and supports single-precision FP only, so her ISA string would be

Quoted 5 lines

IMAFCXdoom.

The variable elf_hwcap is calculated at the loop in line 57,58, the 'd' from Xdoom would bypass the check, while the underlying machine does not support double-precision FP.

Ah, yes, that makes sense. I'd go the other way here and just be strict about parsing the ISA string: it's defined to be listed in a particular order, so we should really only be accepting legal ISA strings.

I'll submit a second patch to fix this behavior.

Quoted 11 lines
Quoted 6 lines
arch/riscv/kernel/cpufeature.c +0 -1
 	pr_info("elf_hwcap is 0x%lx", elf_hwcap);
 }
--

2.16.4

I don't know if the reasons make sense to you, but anyway that's all I would like to say about this patch.

Alan