| Message ID | f2186827-62e6-4b24-8a6c-0c2a9499c232@suse.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | x86: slightly simplify MB2/EFI "magic" check | expand |
On Thu, Aug 8, 2024 at 9:49 AM Jan Beulich <jbeulich@suse.com> wrote: > > A few dozen lines down from here we repeatedly use a pattern involving > just a single (conditional) branch. Do so also when checking for the > boot loader magic value. > > Signed-off-by: Jan Beulich <jbeulich@suse.com> > --- > I further question the placement of the clearing of vga_text_buffer, > just out of context: Shouldn't that be placed with the increments of > efi_platform and skip_realmode? Or else is the terminology in comments > ("on EFI platforms") wrong in one of the two places? In the end, if we > are entered at __efi64_mb2_start but the magic doesn't match, we simply > don't know what environment we're in. There may or may not be a VGA > console at the default address, so we may as well (try to) write to it > (just like we do when entered at start). > I don't think __efi64_mb2_start should be ever get called by anything else than multiboot2. Was EFI supported by multiboot version 1 maybe? As far as I can see we will print an error and halt the machine so we expect something really wrong to have happened. We could indeed be in a position where we have VGA available. Or EFI? Or not? As said something really weird and unexpected happened. Maybe the safer way would be to print on serial and try to print on VGA in this case. At the moment we mix the 2 prints (each character is duplicated), printing all message to serial and then trying to print all message to VGA and then halt could be an option. > --- a/xen/arch/x86/boot/head.S > +++ b/xen/arch/x86/boot/head.S > @@ -233,13 +233,11 @@ __efi64_mb2_start: > > /* Check for Multiboot2 bootloader. */ > cmp $MULTIBOOT2_BOOTLOADER_MAGIC,%eax > - je .Lefi_multiboot2_proto > > /* Jump to .Lnot_multiboot after switching CPU to x86_32 mode. */ > lea .Lnot_multiboot(%rip), %r15 > - jmp x86_32_switch > + jne x86_32_switch > It looks good to me. Maybe a bit less readable. > -.Lefi_multiboot2_proto: > /* Zero EFI SystemTable, EFI ImageHandle addresses and cmdline. */ > xor %esi,%esi > xor %edi,%edi > Frediano
On 08.08.2024 11:36, Frediano Ziglio wrote: > On Thu, Aug 8, 2024 at 9:49 AM Jan Beulich <jbeulich@suse.com> wrote: >> >> A few dozen lines down from here we repeatedly use a pattern involving >> just a single (conditional) branch. Do so also when checking for the >> boot loader magic value. >> >> Signed-off-by: Jan Beulich <jbeulich@suse.com> >> --- >> I further question the placement of the clearing of vga_text_buffer, >> just out of context: Shouldn't that be placed with the increments of >> efi_platform and skip_realmode? Or else is the terminology in comments >> ("on EFI platforms") wrong in one of the two places? In the end, if we >> are entered at __efi64_mb2_start but the magic doesn't match, we simply >> don't know what environment we're in. There may or may not be a VGA >> console at the default address, so we may as well (try to) write to it >> (just like we do when entered at start). > > I don't think __efi64_mb2_start should be ever get called by anything > else than multiboot2. Was EFI supported by multiboot version 1 maybe? No, there was no EFI with MB1. > As far as I can see we will print an error and halt the machine so we > expect something really wrong to have happened. > We could indeed be in a position where we have VGA available. Or EFI? > Or not? As said something really weird and unexpected happened. Maybe > the safer way would be to print on serial and try to print on VGA in > this case. At the moment we mix the 2 prints (each character is > duplicated), printing all message to serial and then trying to print > all message to VGA and then halt could be an option. Sounds like you see something wrong with the "mixing" as you call it? I'm suggesting exactly that: Try to output something to both possible channels. Whether there's serial at the default port we don't know either, after all. What we kind of know is that we're in 64-bit mode (yet even that could probably do with verifying, when we already have to assume something is very broken). Even there I'd expect the VGA range to be mapped by the identity page tables that our caller must have put in place. Jan
Hi, On Thu Aug 8, 2024 at 9:49 AM BST, Jan Beulich wrote: > A few dozen lines down from here we repeatedly use a pattern involving > just a single (conditional) branch. Do so also when checking for the > boot loader magic value. > > Signed-off-by: Jan Beulich <jbeulich@suse.com> > --- > I further question the placement of the clearing of vga_text_buffer, > just out of context: Shouldn't that be placed with the increments of > efi_platform and skip_realmode? Or else is the terminology in comments > ("on EFI platforms") wrong in one of the two places? In the end, if we > are entered at __efi64_mb2_start but the magic doesn't match, we simply > don't know what environment we're in. There may or may not be a VGA > console at the default address, so we may as well (try to) write to it > (just like we do when entered at start). It's fair to assume we're in 64bits, and in that situation it's also fair to assume the text console is long gone (pun intended). Seeing how this would be a boot protocol bug, I think the most reasonable thing to do is to leave a poison value in RAX and then hang (deadbeef..., badcafe..., take a pick). That would point the poor sod debugging this to the right part of the code. Though this is a largely theoretical issue that I can't see happening in practice. > > --- a/xen/arch/x86/boot/head.S > +++ b/xen/arch/x86/boot/head.S > @@ -233,13 +233,11 @@ __efi64_mb2_start: > > /* Check for Multiboot2 bootloader. */ > cmp $MULTIBOOT2_BOOTLOADER_MAGIC,%eax > - je .Lefi_multiboot2_proto > > /* Jump to .Lnot_multiboot after switching CPU to x86_32 mode. */ > lea .Lnot_multiboot(%rip), %r15 I don't think there's much benefit to this, but it would read more naturally if lea was before cmp. Then cmp would be next to its (new) associated jne. > - jmp x86_32_switch > + jne x86_32_switch > > -.Lefi_multiboot2_proto: > /* Zero EFI SystemTable, EFI ImageHandle addresses and cmdline. */ > xor %esi,%esi > xor %edi,%edi Cheers, Alejandro
On 12.08.2024 16:34, Alejandro Vallejo wrote: > On Thu Aug 8, 2024 at 9:49 AM BST, Jan Beulich wrote: >> A few dozen lines down from here we repeatedly use a pattern involving >> just a single (conditional) branch. Do so also when checking for the >> boot loader magic value. >> >> Signed-off-by: Jan Beulich <jbeulich@suse.com> >> --- >> I further question the placement of the clearing of vga_text_buffer, >> just out of context: Shouldn't that be placed with the increments of >> efi_platform and skip_realmode? Or else is the terminology in comments >> ("on EFI platforms") wrong in one of the two places? In the end, if we >> are entered at __efi64_mb2_start but the magic doesn't match, we simply >> don't know what environment we're in. There may or may not be a VGA >> console at the default address, so we may as well (try to) write to it >> (just like we do when entered at start). > > It's fair to assume we're in 64bits, and in that situation it's also fair to > assume the text console is long gone (pun intended). How is being in 64-bit mode correlated with there being a VGA console? (I question "fair to assume" here anyway. We're on a path where we know something's wonky.) >> --- a/xen/arch/x86/boot/head.S >> +++ b/xen/arch/x86/boot/head.S >> @@ -233,13 +233,11 @@ __efi64_mb2_start: >> >> /* Check for Multiboot2 bootloader. */ >> cmp $MULTIBOOT2_BOOTLOADER_MAGIC,%eax >> - je .Lefi_multiboot2_proto >> >> /* Jump to .Lnot_multiboot after switching CPU to x86_32 mode. */ >> lea .Lnot_multiboot(%rip), %r15 > > I don't think there's much benefit to this, but it would read more naturally if > lea was before cmp. Then cmp would be next to its (new) associated jne. You did look at the pattern though that I'm referring to in the description? Knowing that generally paring the CMP/TEST with the Jcc, I would have switched things around. Yet I wanted to make things as similar as possible, in the hope that be(com)ing consistent would make it easiest to get such a minor change in. Jan
Hi, On Mon Aug 12, 2024 at 3:43 PM BST, Jan Beulich wrote: > On 12.08.2024 16:34, Alejandro Vallejo wrote: > > On Thu Aug 8, 2024 at 9:49 AM BST, Jan Beulich wrote: > >> A few dozen lines down from here we repeatedly use a pattern involving > >> just a single (conditional) branch. Do so also when checking for the > >> boot loader magic value. > >> > >> Signed-off-by: Jan Beulich <jbeulich@suse.com> > >> --- > >> I further question the placement of the clearing of vga_text_buffer, > >> just out of context: Shouldn't that be placed with the increments of > >> efi_platform and skip_realmode? Or else is the terminology in comments > >> ("on EFI platforms") wrong in one of the two places? In the end, if we > >> are entered at __efi64_mb2_start but the magic doesn't match, we simply > >> don't know what environment we're in. There may or may not be a VGA > >> console at the default address, so we may as well (try to) write to it > >> (just like we do when entered at start). > > > > It's fair to assume we're in 64bits, and in that situation it's also fair to > > assume the text console is long gone (pun intended). > > How is being in 64-bit mode correlated with there being a VGA console? > (I question "fair to assume" here anyway. We're on a path where we know > something's wonky.) The only way in which you could plausibly have a text-mode console in 64bits is if you booted from BIOS and didn't set a VESA mode, so it boils down to what failure modes you want to consider. For "anything goes" you're right that we can't even be sure of being in 64bit (or 32bit) mode, but that's too draconian an assumption to try to uphold, imo. I think that while details in the boot protocol might be incorrect (like the magic tag), broad strokes (like being in long mode and having a UEFI runtime) must still hold. Trying to use the serial is fine (worst-case scenario it doesn't work), but trying to use a framebuffer you're not sure about is not unlikely to triple fault your machine prematurely and then debugging it is a pain even with an emulator. > > >> --- a/xen/arch/x86/boot/head.S > >> +++ b/xen/arch/x86/boot/head.S > >> @@ -233,13 +233,11 @@ __efi64_mb2_start: > >> > >> /* Check for Multiboot2 bootloader. */ > >> cmp $MULTIBOOT2_BOOTLOADER_MAGIC,%eax > >> - je .Lefi_multiboot2_proto > >> > >> /* Jump to .Lnot_multiboot after switching CPU to x86_32 mode. */ > >> lea .Lnot_multiboot(%rip), %r15 > > > > I don't think there's much benefit to this, but it would read more naturally if > > lea was before cmp. Then cmp would be next to its (new) associated jne. > > You did look at the pattern though that I'm referring to in the description? > Knowing that generally paring the CMP/TEST with the Jcc, I would have > switched things around. Yet I wanted to make things as similar as possible, > in the hope that be(com)ing consistent would make it easiest to get such a > minor change in. > > Jan I'm not sure about the pattern you mention. Seems like a standard set of doX+check+cond_jump finished in an unconditional jump. All of them pretty normal. Regardless, I'm not arguing against this. While I happen to find it easier to mentally parse it in its current form we do save a jump instruction after your change. It's just that it'd be easier to follow with the mentioned reversal of lea and cmp. Cheers, Alejandro
On 08/08/2024 9:49 am, Jan Beulich wrote: > --- a/xen/arch/x86/boot/head.S > +++ b/xen/arch/x86/boot/head.S > @@ -233,13 +233,11 @@ __efi64_mb2_start: > > /* Check for Multiboot2 bootloader. */ > cmp $MULTIBOOT2_BOOTLOADER_MAGIC,%eax > - je .Lefi_multiboot2_proto > > /* Jump to .Lnot_multiboot after switching CPU to x86_32 mode. */ > lea .Lnot_multiboot(%rip), %r15 > - jmp x86_32_switch > + jne x86_32_switch > > -.Lefi_multiboot2_proto: > /* Zero EFI SystemTable, EFI ImageHandle addresses and cmdline. */ > xor %esi,%esi > xor %edi,%edi You've split the logical in two, and now the comment is in the wrong position. If you're going to make this change, it wants to end up reading: /* Jump to .Lnot_multiboot after switching CPU to x86_32 mode. */ lea .Lnot_multiboot(%rip), %r15 /* Check for Multiboot2 bootloader. */ cmp $MULTIBOOT2_BOOTLOADER_MAGIC,%eax jne x86_32_switch Not that it's really relevant, but this form also macrofuses nicely. ~Andrew
On 12.08.2024 23:40, Andrew Cooper wrote: > On 08/08/2024 9:49 am, Jan Beulich wrote: >> --- a/xen/arch/x86/boot/head.S >> +++ b/xen/arch/x86/boot/head.S >> @@ -233,13 +233,11 @@ __efi64_mb2_start: >> >> /* Check for Multiboot2 bootloader. */ >> cmp $MULTIBOOT2_BOOTLOADER_MAGIC,%eax >> - je .Lefi_multiboot2_proto >> >> /* Jump to .Lnot_multiboot after switching CPU to x86_32 mode. */ >> lea .Lnot_multiboot(%rip), %r15 >> - jmp x86_32_switch >> + jne x86_32_switch >> >> -.Lefi_multiboot2_proto: >> /* Zero EFI SystemTable, EFI ImageHandle addresses and cmdline. */ >> xor %esi,%esi >> xor %edi,%edi > > > You've split the logical in two, and now the comment is in the wrong > position. > > If you're going to make this change, it wants to end up reading: > > /* Jump to .Lnot_multiboot after switching CPU to x86_32 mode. */ > lea .Lnot_multiboot(%rip), %r15 > > /* Check for Multiboot2 bootloader. */ > cmp $MULTIBOOT2_BOOTLOADER_MAGIC,%eax > jne x86_32_switch > > > Not that it's really relevant, but this form also macrofuses nicely. All true, yet I'm sure you also read my response to Alejandro: Then we ought to also change the other similar instances. The goal of this simple change, after all, is to get all similar pieces of code into consistent shape. Jan
--- a/xen/arch/x86/boot/head.S +++ b/xen/arch/x86/boot/head.S @@ -233,13 +233,11 @@ __efi64_mb2_start: /* Check for Multiboot2 bootloader. */ cmp $MULTIBOOT2_BOOTLOADER_MAGIC,%eax - je .Lefi_multiboot2_proto /* Jump to .Lnot_multiboot after switching CPU to x86_32 mode. */ lea .Lnot_multiboot(%rip), %r15 - jmp x86_32_switch + jne x86_32_switch -.Lefi_multiboot2_proto: /* Zero EFI SystemTable, EFI ImageHandle addresses and cmdline. */ xor %esi,%esi xor %edi,%edi
A few dozen lines down from here we repeatedly use a pattern involving just a single (conditional) branch. Do so also when checking for the boot loader magic value. Signed-off-by: Jan Beulich <jbeulich@suse.com> --- I further question the placement of the clearing of vga_text_buffer, just out of context: Shouldn't that be placed with the increments of efi_platform and skip_realmode? Or else is the terminology in comments ("on EFI platforms") wrong in one of the two places? In the end, if we are entered at __efi64_mb2_start but the magic doesn't match, we simply don't know what environment we're in. There may or may not be a VGA console at the default address, so we may as well (try to) write to it (just like we do when entered at start).