Skip to contentSkip to conversation

x86/pv: Inject #GP for implicit grant unmaps

PATCH In review Andrew Cooper 2 participants 4 messages Jul 25, 2022 – Jul 26, 2022

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/pv: Inject #GP for implicit grant unmaps

This is a debug behaviour to identify buggy kernels. Crashing the domain is the most unhelpful thing to do, because it discards the relevant context.

Instead, inject #GP[0] like other permission errors in x86. In particular, this lets the kernel provide a backtrace that's actually helpful to a developer trying to figure out what's going wrong.

As a bugfix, this always injects #GP[0] to current, not l1e_owner. It is not l1e_owner's fault if dom0 using superpowers triggers an implicit unmap.

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Wei Liu <[email protected]>
CC: Juergen Gross <[email protected]>

This is a prerequisite to investigating
https://github.com/QubesOS/qubes-issues/issues/7631 which is looking like an error in Linux's gntdev driver.

---
 xen/arch/x86/mm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
xen/arch/x86/mm.c +1 -1
index 5b81d5fbdbb2..b3393385ffb6 100644
@@ -1232,7 +1232,7 @@ void put_page_from_l1e(l1_pgentry_t l1e, struct domain *l1e_owner)
         gdprintk(XENLOG_WARNING,
                  "Attempt to implicitly unmap a granted PTE %" PRIpte "\n",
                  l1e_get_intpte(l1e));
-        domain_crash(l1e_owner);
+        pv_inject_hw_exception(TRAP_gp_fault, 0);
     }
 #endif
 
Signature · 1 line
2.11.0
Acked-by on v1 1/1
#1 lore raw
Quoted 11 lines · Andrew Cooper

This is a debug behaviour to identify buggy kernels. Crashing the domain is the most unhelpful thing to do, because it discards the relevant context.

Instead, inject #GP[0] like other permission errors in x86. In particular, this lets the kernel provide a backtrace that's actually helpful to a developer trying to figure out what's going wrong.

As a bugfix, this always injects #GP[0] to current, not l1e_owner. It is not l1e_owner's fault if dom0 using superpowers triggers an implicit unmap.

Albeit preferably with ...

xen/arch/x86/mm.c void put_page_from_l1e(l1_pgentry_t l1e, struct domain *l1e_owner) · line 1237

Quoted 10 lines
xen/arch/x86/mm.c +1 -1
@@ -1232,7 +1232,7 @@ void put_page_from_l1e(l1_pgentry_t l1e, struct domain *l1e_owner)
         gdprintk(XENLOG_WARNING,
                  "Attempt to implicitly unmap a granted PTE %" PRIpte "\n",
                  l1e_get_intpte(l1e));
-        domain_crash(l1e_owner);
+        pv_inject_hw_exception(TRAP_gp_fault, 0);
     }
 #endif

... the gdprintk() adjusted to also log l1e_owner.

Jan

#2 lore raw
Quoted 16 lines · Jan Beulich
Quoted 11 lines · Andrew Cooper

This is a debug behaviour to identify buggy kernels. Crashing the domain is the most unhelpful thing to do, because it discards the relevant context.

Instead, inject #GP[0] like other permission errors in x86. In particular, this lets the kernel provide a backtrace that's actually helpful to a developer trying to figure out what's going wrong.

As a bugfix, this always injects #GP[0] to current, not l1e_owner. It is not l1e_owner's fault if dom0 using superpowers triggers an implicit unmap.

Albeit preferably with ...

Quoted 11 lines

xen/arch/x86/mm.c void put_page_from_l1e(l1_pgentry_t l1e, struct domain *l1e_owner) · line 1237

Quoted 10 lines
xen/arch/x86/mm.c +1 -1
@@ -1232,7 +1232,7 @@ void put_page_from_l1e(l1_pgentry_t l1e, struct domain *l1e_owner)
         gdprintk(XENLOG_WARNING,
                  "Attempt to implicitly unmap a granted PTE %" PRIpte "\n",
                  l1e_get_intpte(l1e));
-        domain_crash(l1e_owner);
+        pv_inject_hw_exception(TRAP_gp_fault, 0);
     }
 #endif

... the gdprintk() adjusted to also log l1e_owner.

Ok, how about this incremental?

xen/arch/x86/mm.c +0 -0
index b3393385ffb6..74054fb5f4ee 100644
@@ -1229,9 +1229,9 @@ void put_page_from_l1e(l1_pgentry_t l1e, struct

domain *l1e_owner)
     if ( (l1e_get_flags(l1e) & _PAGE_GNTTAB) && !l1e_owner->is_shutting_down && !l1e_owner->is_dying ) {
-        gdprintk(XENLOG_WARNING,
-                 "Attempt to implicitly unmap a granted PTE %" PRIpte "\n", -                 l1e_get_intpte(l1e));
+        gprintk(XENLOG_WARNING,
+                "Attempt to implicitly %pd's gntmap PTE %" PRIpte "\n", +                l1e_owner, l1e_get_intpte(l1e)); pv_inject_hw_exception(TRAP_gp_fault, 0);
     }
 #endif

The printk() needs to not be omitted in release builds which happen to have this logic compiled in.

~Andrew

#3 lore raw
Quoted 38 lines · Andrew Cooper
Quoted 16 lines · Jan Beulich
Quoted 11 lines · Andrew Cooper

This is a debug behaviour to identify buggy kernels. Crashing the domain is the most unhelpful thing to do, because it discards the relevant context.

Instead, inject #GP[0] like other permission errors in x86. In particular, this lets the kernel provide a backtrace that's actually helpful to a developer trying to figure out what's going wrong.

As a bugfix, this always injects #GP[0] to current, not l1e_owner. It is not l1e_owner's fault if dom0 using superpowers triggers an implicit unmap.

Albeit preferably with ...

Quoted 11 lines

xen/arch/x86/mm.c void put_page_from_l1e(l1_pgentry_t l1e, struct domain *l1e_owner) · line 1237

Quoted 10 lines
xen/arch/x86/mm.c +1 -1
@@ -1232,7 +1232,7 @@ void put_page_from_l1e(l1_pgentry_t l1e, struct domain *l1e_owner)
         gdprintk(XENLOG_WARNING,
                  "Attempt to implicitly unmap a granted PTE %" PRIpte "\n",
                  l1e_get_intpte(l1e));
-        domain_crash(l1e_owner);
+        pv_inject_hw_exception(TRAP_gp_fault, 0);
     }
 #endif

... the gdprintk() adjusted to also log l1e_owner.

Ok, how about this incremental?

xen/arch/x86/mm.c +0 -0
index b3393385ffb6..74054fb5f4ee 100644
@@ -1229,9 +1229,9 @@ void put_page_from_l1e(l1_pgentry_t l1e, struct
 domain *l1e_owner)

     if ( (l1e_get_flags(l1e) & _PAGE_GNTTAB) &&

Quoted 8 lines

          !l1e_owner->is_shutting_down && !l1e_owner->is_dying ) {
-        gdprintk(XENLOG_WARNING,
-                 "Attempt to implicitly unmap a granted PTE %" PRIpte "\n", -                 l1e_get_intpte(l1e));
+        gprintk(XENLOG_WARNING,
+                "Attempt to implicitly %pd's gntmap PTE %" PRIpte "\n", +                l1e_owner, l1e_get_intpte(l1e));

DYM to drop "unmap"? With it restored (or anything similar to that effect), fine with me.

Jan