Skip to contentSkip to conversation

x86/boot: Fix zap_low_mappings() to map less of the trampoline

PATCH In review Andrew Cooper 3 participants 8 messages Jan 6, 2025 – Mar 17, 2025

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/boot: Fix zap_low_mappings() to map less of the trampoline

Regular data access into the trampoline is via the directmap.

As now discussed quite extensively in asm/trampoline.h, the trampoline is arranged so that only the AP and S3 paths need an identity mapping, and that they fit within a single page.

Right now, PFN_UP(trampoline_end - trampoline_start) is 2, causing more than expected of the trampoline to be mapped. Cut it down just the single page it ought to be.

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>

There's not an obvious candidate for a Fixes tag.

---
 xen/arch/x86/x86_64/mm.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
xen/arch/x86/x86_64/mm.c +6 -4
index 389d813ebe63..d4e6a9c0a2e0 100644
@@ -718,14 +718,16 @@ void __init zap_low_mappings(void)
 {
     BUG_ON(num_online_cpus() != 1);
 
-    /* Remove aliased mapping of first 1:1 PML4 entry. */
+    /* Stop using l?_bootmap[] mappings. */
     l4e_write(&idle_pg_table[0], l4e_empty());
     flush_local(FLUSH_TLB_GLOBAL);
 
-    /* Replace with mapping of the boot trampoline only. */
+    /*
+     * Insert an identity mapping of the AP/S3 part of the trampoline, which
+     * is arranged to fit in a single page.
+     */
     map_pages_to_xen(trampoline_phys, maddr_to_mfn(trampoline_phys),
-                     PFN_UP(trampoline_end - trampoline_start),
-                     __PAGE_HYPERVISOR_RX);
+                     1, __PAGE_HYPERVISOR_RX);
 }
 
 int setup_compat_arg_xlat(struct vcpu *v)
Signature · 1 line
2.39.5
Acked-by on v1 1/1
#1 lore raw
Quoted 11 lines · Andrew Cooper

Regular data access into the trampoline is via the directmap.

As now discussed quite extensively in asm/trampoline.h, the trampoline is arranged so that only the AP and S3 paths need an identity mapping, and that they fit within a single page.

Right now, PFN_UP(trampoline_end - trampoline_start) is 2, causing more than expected of the trampoline to be mapped. Cut it down just the single page it ought to be.

on the basis that this improves things. However, ...

Quoted 10 lines
xen/arch/x86/x86_64/mm.c +1 -1
@@ -718,14 +718,16 @@ void __init zap_low_mappings(void)
 {
     BUG_ON(num_online_cpus() != 1);
 
-    /* Remove aliased mapping of first 1:1 PML4 entry. */
+    /* Stop using l?_bootmap[] mappings. */
     l4e_write(&idle_pg_table[0], l4e_empty());
     flush_local(FLUSH_TLB_GLOBAL);

xen/arch/x86/x86_64/mm.c void __init zap_low_mappings(void) · line 730

Quoted 10 lines
xen/arch/x86/x86_64/mm.c +5 -3
 
-    /* Replace with mapping of the boot trampoline only. */
+    /*
+     * Insert an identity mapping of the AP/S3 part of the trampoline, which
+     * is arranged to fit in a single page.
+     */
     map_pages_to_xen(trampoline_phys, maddr_to_mfn(trampoline_phys),
-                     PFN_UP(trampoline_end - trampoline_start),
-                     __PAGE_HYPERVISOR_RX);
+                     1, __PAGE_HYPERVISOR_RX);

... literal numbers like this - however well they are commented - are potentially problematic to locate in case something changes significantly. The 1 here really would want connecting with the .equ establishing wakeup_stack.

Jan

#2 lore raw
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Andrew Cooper

Regular data access into the trampoline is via the directmap.

As now discussed quite extensively in asm/trampoline.h, the trampoline is arranged so that only the AP and S3 paths need an identity mapping, and that they fit within a single page.

Quoted 6 lines
Quoted 5 lines

Right now, PFN_UP(trampoline_end - trampoline_start) is 2, causing more than expected of the trampoline to be mapped. Cut it down just the single page it ought to be.

Thanks.  However,

Quoted 13 lines

on the basis that this improves things. However, ...

Quoted 11 lines
xen/arch/x86/x86_64/mm.c +1 -1
@@ -718,14 +718,16 @@ void __init zap_low_mappings(void)
 {
     BUG_ON(num_online_cpus() != 1);
 
-    /* Remove aliased mapping of first 1:1 PML4 entry. */
+    /* Stop using l?_bootmap[] mappings. */
     l4e_write(&idle_pg_table[0], l4e_empty());
     flush_local(FLUSH_TLB_GLOBAL);
 
Quoted 13 lines

xen/arch/x86/x86_64/mm.c void __init zap_low_mappings(void) · line 730

Quoted 9 lines
xen/arch/x86/x86_64/mm.c +5 -3
-    /* Replace with mapping of the boot trampoline only. */
+    /*
+     * Insert an identity mapping of the AP/S3 part of the trampoline, which
+     * is arranged to fit in a single page.
+     */
     map_pages_to_xen(trampoline_phys, maddr_to_mfn(trampoline_phys),
-                     PFN_UP(trampoline_end - trampoline_start),
-                     __PAGE_HYPERVISOR_RX);
+                     1, __PAGE_HYPERVISOR_RX);

... literal numbers like this - however well they are commented - are potentially problematic to locate in case something changes significantly. The 1 here really would want connecting with the .equ establishing wakeup_stack.

how do you propose doing this?

PFN_UP(wakeup_stack - trampoline_start) doesn't have the same obvious connection, and it would involve partly undoing 7d73c6f196a5 which hid the symbol recently.

While 1 isn't ideal, it is next to a comment explaining what's going on, and it's not going to go stale in a silent way...  (It's also not liable to go stale either.)

~Andrew

#3 lore raw
Quoted 46 lines · Andrew Cooper
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Andrew Cooper

Regular data access into the trampoline is via the directmap.

As now discussed quite extensively in asm/trampoline.h, the trampoline is arranged so that only the AP and S3 paths need an identity mapping, and that they fit within a single page.

Quoted 6 lines
Quoted 5 lines

Right now, PFN_UP(trampoline_end - trampoline_start) is 2, causing more than expected of the trampoline to be mapped. Cut it down just the single page it ought to be.

Thanks.  However,

Quoted 13 lines

on the basis that this improves things. However, ...

Quoted 11 lines
xen/arch/x86/x86_64/mm.c +1 -1
@@ -718,14 +718,16 @@ void __init zap_low_mappings(void)
 {
     BUG_ON(num_online_cpus() != 1);
 
-    /* Remove aliased mapping of first 1:1 PML4 entry. */
+    /* Stop using l?_bootmap[] mappings. */
     l4e_write(&idle_pg_table[0], l4e_empty());
     flush_local(FLUSH_TLB_GLOBAL);
 
Quoted 13 lines

xen/arch/x86/x86_64/mm.c void __init zap_low_mappings(void) · line 730

Quoted 9 lines
xen/arch/x86/x86_64/mm.c +5 -3
-    /* Replace with mapping of the boot trampoline only. */
+    /*
+     * Insert an identity mapping of the AP/S3 part of the trampoline, which
+     * is arranged to fit in a single page.
+     */
     map_pages_to_xen(trampoline_phys, maddr_to_mfn(trampoline_phys),
-                     PFN_UP(trampoline_end - trampoline_start),
-                     __PAGE_HYPERVISOR_RX);
+                     1, __PAGE_HYPERVISOR_RX);

... literal numbers like this - however well they are commented - are potentially problematic to locate in case something changes significantly. The 1 here really would want connecting with the .equ establishing wakeup_stack.

how do you propose doing this?

Quoted 7 lines

PFN_UP(wakeup_stack - trampoline_start) doesn't have the same obvious connection, and it would involve partly undoing 7d73c6f196a5 which hid the symbol recently.

While 1 isn't ideal, it is next to a comment explaining what's going on, and it's not going to go stale in a silent way...  (It's also not liable to go stale either.)

If in

.equ wakeup_stack, trampoline_start + PAGE_SIZE

PAGE_SIZE was replaced by a new (in asm/trampoline.h) TRAMPOLINE_PERM_SIZE, you could use PFN_UP(TRAMPOLINE_PERM_SIZE) here to establish a connection.

I have to admit I also don't really see why things going stale here would
(a) be unlikely and (b) be guaranteed to not go silently. We just don't know what we may need to add to the trampoline, sooner or later.

Jan

#4 lore raw
Quoted 57 lines · Jan Beulich
Quoted 46 lines · Andrew Cooper
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Andrew Cooper

Regular data access into the trampoline is via the directmap.

As now discussed quite extensively in asm/trampoline.h, the trampoline is arranged so that only the AP and S3 paths need an identity mapping, and that they fit within a single page.

Quoted 6 lines
Quoted 5 lines

Right now, PFN_UP(trampoline_end - trampoline_start) is 2, causing more than expected of the trampoline to be mapped. Cut it down just the single page it ought to be.

Thanks.  However,

Quoted 13 lines

on the basis that this improves things. However, ...

Quoted 11 lines
xen/arch/x86/x86_64/mm.c +1 -1
@@ -718,14 +718,16 @@ void __init zap_low_mappings(void)
 {
     BUG_ON(num_online_cpus() != 1);
 
-    /* Remove aliased mapping of first 1:1 PML4 entry. */
+    /* Stop using l?_bootmap[] mappings. */
     l4e_write(&idle_pg_table[0], l4e_empty());
     flush_local(FLUSH_TLB_GLOBAL);
 
Quoted 13 lines

xen/arch/x86/x86_64/mm.c void __init zap_low_mappings(void) · line 730

Quoted 9 lines
xen/arch/x86/x86_64/mm.c +5 -3
-    /* Replace with mapping of the boot trampoline only. */
+    /*
+     * Insert an identity mapping of the AP/S3 part of the trampoline, which
+     * is arranged to fit in a single page.
+     */
     map_pages_to_xen(trampoline_phys, maddr_to_mfn(trampoline_phys),
-                     PFN_UP(trampoline_end - trampoline_start),
-                     __PAGE_HYPERVISOR_RX);
+                     1, __PAGE_HYPERVISOR_RX);

... literal numbers like this - however well they are commented - are potentially problematic to locate in case something changes significantly. The 1 here really would want connecting with the .equ establishing wakeup_stack.

how do you propose doing this?

Quoted 7 lines

PFN_UP(wakeup_stack - trampoline_start) doesn't have the same obvious connection, and it would involve partly undoing 7d73c6f196a5 which hid the symbol recently.

While 1 isn't ideal, it is next to a comment explaining what's going on, and it's not going to go stale in a silent way...  (It's also not liable to go stale either.)

If in

Quoted 8 lines

.equ wakeup_stack, trampoline_start + PAGE_SIZE

PAGE_SIZE was replaced by a new (in asm/trampoline.h) TRAMPOLINE_PERM_SIZE, you could use PFN_UP(TRAMPOLINE_PERM_SIZE) here to establish a connection.

I have to admit I also don't really see why things going stale here would
(a) be unlikely and (b) be guaranteed to not go silently. We just don't know what we may need to add to the trampoline, sooner or later.

Maybe we could introduce trampoline_{ap?,runtime?}_{start,end} that covers this more narrow section of the trampoline code?

Thanks, Roger.

#5 lore raw
Quoted 55 lines · Roger Pau Monné
Quoted 54 lines · Jan Beulich
Quoted 44 lines · Andrew Cooper
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Andrew Cooper

Regular data access into the trampoline is via the directmap.

As now discussed quite extensively in asm/trampoline.h, the trampoline is arranged so that only the AP and S3 paths need an identity mapping, and that they fit within a single page.

Quoted 6 lines
Quoted 5 lines

Right now, PFN_UP(trampoline_end - trampoline_start) is 2, causing more than expected of the trampoline to be mapped. Cut it down just the single page it ought to be.

Thanks.  However,

Quoted 13 lines

on the basis that this improves things. However, ...

Quoted 11 lines
xen/arch/x86/x86_64/mm.c +1 -1
@@ -718,14 +718,16 @@ void __init zap_low_mappings(void)
 {
     BUG_ON(num_online_cpus() != 1);
 
-    /* Remove aliased mapping of first 1:1 PML4 entry. */
+    /* Stop using l?_bootmap[] mappings. */
     l4e_write(&idle_pg_table[0], l4e_empty());
     flush_local(FLUSH_TLB_GLOBAL);
 
Quoted 13 lines

xen/arch/x86/x86_64/mm.c void __init zap_low_mappings(void) · line 730

Quoted 9 lines
xen/arch/x86/x86_64/mm.c +5 -3
-    /* Replace with mapping of the boot trampoline only. */
+    /*
+     * Insert an identity mapping of the AP/S3 part of the trampoline, which
+     * is arranged to fit in a single page.
+     */
     map_pages_to_xen(trampoline_phys, maddr_to_mfn(trampoline_phys),
-                     PFN_UP(trampoline_end - trampoline_start),
-                     __PAGE_HYPERVISOR_RX);
+                     1, __PAGE_HYPERVISOR_RX);

... literal numbers like this - however well they are commented - are potentially problematic to locate in case something changes significantly. The 1 here really would want connecting with the .equ establishing wakeup_stack.

how do you propose doing this?

Quoted 7 lines

PFN_UP(wakeup_stack - trampoline_start) doesn't have the same obvious connection, and it would involve partly undoing 7d73c6f196a5 which hid the symbol recently.

While 1 isn't ideal, it is next to a comment explaining what's going on, and it's not going to go stale in a silent way...  (It's also not liable to go stale either.)

If in

Quoted 7 lines
Quoted 7 lines

.equ wakeup_stack, trampoline_start + PAGE_SIZE

PAGE_SIZE was replaced by a new (in asm/trampoline.h) TRAMPOLINE_PERM_SIZE, you could use PFN_UP(TRAMPOLINE_PERM_SIZE) here to establish a connection.

I have to admit I also don't really see why things going stale here would
(a) be unlikely and (b) be guaranteed to not go silently.

The size can't go to 0 or everything will break, and if it goes larger than 1 (which it almost certainly never will), then APs and/or S3 will break, and we've got both of these in CI.

Furthermore, the actual thing which matters is:

Quoted 5 lines

/* Map the permanent trampoline page into l1_bootmap[]. */ mov     sym_esi(trampoline_phys), %ecx
lea     __PAGE_HYPERVISOR_RX(%ecx), %edx /* %edx = PTE to write  */ shr     $PAGE_SHIFT, %ecx                /* %ecx = Slot to write */ mov     %edx, sym_offs(l1_bootmap)(%esi, %ecx, 8)

which hardcodes 1 page, because there's almost certainly no chance this will ever change.

Quoted 4 lines
Quoted 2 lines

We just don't
know what we may need to add to the trampoline, sooner or later.

Maybe we could introduce trampoline_{ap?,runtime?}_{start,end} that covers this more narrow section of the trampoline code?

We already have one of those, and a linker assertion that it stays below 1k, so wakeup_stack is at least 3k.

The complexity is that the wakeup_stack overlays some init-only logic in the placed trampoline.  It's not something that exists concretely in the Xen image.

~Andrew

#6 lore raw
Quoted 84 lines · Andrew Cooper
Quoted 55 lines · Roger Pau Monné
Quoted 54 lines · Jan Beulich
Quoted 44 lines · Andrew Cooper
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Andrew Cooper

Regular data access into the trampoline is via the directmap.

As now discussed quite extensively in asm/trampoline.h, the trampoline is arranged so that only the AP and S3 paths need an identity mapping, and that they fit within a single page.

Quoted 6 lines
Quoted 5 lines

Right now, PFN_UP(trampoline_end - trampoline_start) is 2, causing more than expected of the trampoline to be mapped. Cut it down just the single page it ought to be.

Thanks.  However,

Quoted 13 lines

on the basis that this improves things. However, ...

Quoted 11 lines
xen/arch/x86/x86_64/mm.c +1 -1
@@ -718,14 +718,16 @@ void __init zap_low_mappings(void)
 {
     BUG_ON(num_online_cpus() != 1);
 
-    /* Remove aliased mapping of first 1:1 PML4 entry. */
+    /* Stop using l?_bootmap[] mappings. */
     l4e_write(&idle_pg_table[0], l4e_empty());
     flush_local(FLUSH_TLB_GLOBAL);
 
Quoted 13 lines

xen/arch/x86/x86_64/mm.c void __init zap_low_mappings(void) · line 730

Quoted 9 lines
xen/arch/x86/x86_64/mm.c +5 -3
-    /* Replace with mapping of the boot trampoline only. */
+    /*
+     * Insert an identity mapping of the AP/S3 part of the trampoline, which
+     * is arranged to fit in a single page.
+     */
     map_pages_to_xen(trampoline_phys, maddr_to_mfn(trampoline_phys),
-                     PFN_UP(trampoline_end - trampoline_start),
-                     __PAGE_HYPERVISOR_RX);
+                     1, __PAGE_HYPERVISOR_RX);

... literal numbers like this - however well they are commented - are potentially problematic to locate in case something changes significantly. The 1 here really would want connecting with the .equ establishing wakeup_stack.

how do you propose doing this?

Quoted 7 lines

PFN_UP(wakeup_stack - trampoline_start) doesn't have the same obvious connection, and it would involve partly undoing 7d73c6f196a5 which hid the symbol recently.

While 1 isn't ideal, it is next to a comment explaining what's going on, and it's not going to go stale in a silent way...  (It's also not liable to go stale either.)

If in

Quoted 7 lines
Quoted 7 lines

.equ wakeup_stack, trampoline_start + PAGE_SIZE

PAGE_SIZE was replaced by a new (in asm/trampoline.h) TRAMPOLINE_PERM_SIZE, you could use PFN_UP(TRAMPOLINE_PERM_SIZE) here to establish a connection.

I have to admit I also don't really see why things going stale here would
(a) be unlikely and (b) be guaranteed to not go silently.

The size can't go to 0 or everything will break, and if it goes larger than 1 (which it almost certainly never will), then APs and/or S3 will break, and we've got both of these in CI.

Furthermore, the actual thing which matters is:

Quoted 5 lines

/* Map the permanent trampoline page into l1_bootmap[]. */ mov     sym_esi(trampoline_phys), %ecx
lea     __PAGE_HYPERVISOR_RX(%ecx), %edx /* %edx = PTE to write  */ shr     $PAGE_SHIFT, %ecx                /* %ecx = Slot to write */ mov     %edx, sym_offs(l1_bootmap)(%esi, %ecx, 8)

which hardcodes 1 page, because there's almost certainly no chance this will ever change.

Quoted 4 lines
Quoted 2 lines

We just don't
know what we may need to add to the trampoline, sooner or later.

Maybe we could introduce trampoline_{ap?,runtime?}_{start,end} that covers this more narrow section of the trampoline code?

Quoted 6 lines

We already have one of those, and a linker assertion that it stays below 1k, so wakeup_stack is at least 3k.

The complexity is that the wakeup_stack overlays some init-only logic in the placed trampoline.  It's not something that exists concretely in the Xen image.

Well - you've got an ack; while I'd prefer if connections were properly made, I agree it's unlikely the size will grow enough for it to matter. So I think you should feel free to put this in as is.

Jan

#7 lore raw
Quoted 85 lines · Jan Beulich
Quoted 84 lines · Andrew Cooper
Quoted 55 lines · Roger Pau Monné
Quoted 54 lines · Jan Beulich
Quoted 44 lines · Andrew Cooper
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Andrew Cooper

Regular data access into the trampoline is via the directmap.

As now discussed quite extensively in asm/trampoline.h, the trampoline is arranged so that only the AP and S3 paths need an identity mapping, and that they fit within a single page.

Quoted 6 lines
Quoted 5 lines

Right now, PFN_UP(trampoline_end - trampoline_start) is 2, causing more than expected of the trampoline to be mapped. Cut it down just the single page it ought to be.

Thanks.  However,

Quoted 13 lines

on the basis that this improves things. However, ...

Quoted 11 lines
xen/arch/x86/x86_64/mm.c +1 -1
@@ -718,14 +718,16 @@ void __init zap_low_mappings(void)
 {
     BUG_ON(num_online_cpus() != 1);
 
-    /* Remove aliased mapping of first 1:1 PML4 entry. */
+    /* Stop using l?_bootmap[] mappings. */
     l4e_write(&idle_pg_table[0], l4e_empty());
     flush_local(FLUSH_TLB_GLOBAL);
 
Quoted 13 lines

xen/arch/x86/x86_64/mm.c void __init zap_low_mappings(void) · line 730

Quoted 9 lines
xen/arch/x86/x86_64/mm.c +5 -3
-    /* Replace with mapping of the boot trampoline only. */
+    /*
+     * Insert an identity mapping of the AP/S3 part of the trampoline, which
+     * is arranged to fit in a single page.
+     */
     map_pages_to_xen(trampoline_phys, maddr_to_mfn(trampoline_phys),
-                     PFN_UP(trampoline_end - trampoline_start),
-                     __PAGE_HYPERVISOR_RX);
+                     1, __PAGE_HYPERVISOR_RX);

... literal numbers like this - however well they are commented - are potentially problematic to locate in case something changes significantly. The 1 here really would want connecting with the .equ establishing wakeup_stack.

how do you propose doing this?

Quoted 7 lines

PFN_UP(wakeup_stack - trampoline_start) doesn't have the same obvious connection, and it would involve partly undoing 7d73c6f196a5 which hid the symbol recently.

While 1 isn't ideal, it is next to a comment explaining what's going on, and it's not going to go stale in a silent way...  (It's also not liable to go stale either.)

If in

Quoted 7 lines
Quoted 7 lines

.equ wakeup_stack, trampoline_start + PAGE_SIZE

PAGE_SIZE was replaced by a new (in asm/trampoline.h) TRAMPOLINE_PERM_SIZE, you could use PFN_UP(TRAMPOLINE_PERM_SIZE) here to establish a connection.

I have to admit I also don't really see why things going stale here would
(a) be unlikely and (b) be guaranteed to not go silently.

The size can't go to 0 or everything will break, and if it goes larger than 1 (which it almost certainly never will), then APs and/or S3 will break, and we've got both of these in CI.

Furthermore, the actual thing which matters is:

Quoted 5 lines

/* Map the permanent trampoline page into l1_bootmap[]. */ mov     sym_esi(trampoline_phys), %ecx
lea     __PAGE_HYPERVISOR_RX(%ecx), %edx /* %edx = PTE to write  */ shr     $PAGE_SHIFT, %ecx                /* %ecx = Slot to write */ mov     %edx, sym_offs(l1_bootmap)(%esi, %ecx, 8)

which hardcodes 1 page, because there's almost certainly no chance this will ever change.

Quoted 4 lines
Quoted 2 lines

We just don't
know what we may need to add to the trampoline, sooner or later.

Maybe we could introduce trampoline_{ap?,runtime?}_{start,end} that covers this more narrow section of the trampoline code?

We already have one of those, and a linker assertion that it stays below 1k, so wakeup_stack is at least 3k.

Quoted 6 lines
Quoted 3 lines

The complexity is that the wakeup_stack overlays some init-only logic in the placed trampoline.  It's not something that exists concretely in the Xen image.

Well - you've got an ack; while I'd prefer if connections were properly made, I agree it's unlikely the size will grow enough for it to matter. So I think you should feel free to put this in as is.

Thankyou.

~Andrew