Skip to contentSkip to conversation

KVM: my debug patch queue

PATCH Superseded Maxim Levitsky 5 participants 32 messages Mar 15, 2021 – Mar 18, 2021 1 message not in the archive

Subsystems (4)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

cover KVM: my debug patch queue

Hi!

I would like to publish two debug features which were needed for other stuff I work on.

One is the reworked lx-symbols script which now actually works on at least gdb 9.1 (gdb 9.2 was reported to fail to load the debug symbols from the kernel for some reason, not related to this patch) and upstream qemu.

The other feature is the ability to trap all guest exceptions (on SVM for now) and see them in kvmtrace prior to potential merge to double/triple fault.

This can be very useful and I already had to manually patch KVM a few times for this.
I will, once time permits, implement this feature on Intel as well.

Best regards,
Maxim Levitsky

Maxim Levitsky (3):
scripts/gdb: rework lx-symbols gdb script
KVM: x86: guest debug: don't inject interrupts while single stepping
KVM: SVM: allow to intercept all exceptions for debug

 arch/x86/include/asm/kvm_host.h |   2 +
 arch/x86/kvm/svm/svm.c          |  77 ++++++++++++++++++++++-
 arch/x86/kvm/svm/svm.h          |   5 +-
 arch/x86/kvm/x86.c              |  11 +++-
 kernel/module.c                 |   8 ++-
 scripts/gdb/linux/symbols.py    | 106 +++++++++++++++++++++++---------
 6 files changed, 174 insertions(+), 35 deletions(-)
Signature · 1 line
2.26.2
#1 lore raw

PATCH 1/3 scripts/gdb: rework lx-symbols gdb script

Fix several issues that are present in lx-symbols script:

* Track module unloads by placing another software breakpoint at 'free_module' (force uninline this symbol just in case), and use remove-symbol-file gdb command to unload the symobls of the module that is unloading.

That gives the gdb a chance to mark all software breakpoints from this module as pending again.
Also remove the module from the 'known' module list once it is unloaded.

* Since we now track module unload, we don't need to reload all symbols anymore when 'known' module loaded again (that can't happen anymore). This allows reloading a module in the debugged kernel to finish much faster, while lx-symbols tracks module loads and unloads.

* Disable/enable all gdb breakpoints on both module load and unload breakpoint hits, and not only in 'load_all_symbols' as was done before. (load_all_symbols is no longer called on breakpoint hit)
That allows gdb to avoid getting confused about the state of the (now two) internal breakpoints we place.

Otherwise it will leave them in the kernel code segment, when continuing which triggers a guest kernel panic as soon as it skips over the 'int3' instruction and executes the garbage tail of the optcode on which the breakpoint was placed.

---
 kernel/module.c              |   8 ++-
 scripts/gdb/linux/symbols.py | 106 +++++++++++++++++++++++++----------
 2 files changed, 83 insertions(+), 31 deletions(-)
kernel/module.c +6 -2
index 30479355ab850..ea81fc06ea1f5 100644
@@ -901,8 +901,12 @@ int module_refcount(struct module *mod)
 }
 EXPORT_SYMBOL(module_refcount);
 
-/* This exists whether we can unload or not */
-static void free_module(struct module *mod);
+/* This exists whether we can unload or not
+ * Keep it uninlined to provide a reliable breakpoint target,
+ * e.g. for the gdb helper command 'lx-symbols'.
+ */
+
+static noinline void free_module(struct module *mod);
 
 SYSCALL_DEFINE2(delete_module, const char __user *, name_user,
 		unsigned int, flags)
scripts/gdb/linux/symbols.py +77 -29

164 lines not on this page · raw message

Signature · 1 line
2.26.2
maintainer
#2 lore raw

Re: [PATCH 1/3] scripts/gdb: rework lx-symbols gdb script

Quoted 56 lines · Maxim Levitsky

Fix several issues that are present in lx-symbols script:

* Track module unloads by placing another software breakpoint at 'free_module' (force uninline this symbol just in case), and use remove-symbol-file gdb command to unload the symobls of the module that is unloading.

That gives the gdb a chance to mark all software breakpoints from this module as pending again.
Also remove the module from the 'known' module list once it is unloaded.

* Since we now track module unload, we don't need to reload all symbols anymore when 'known' module loaded again (that can't happen anymore). This allows reloading a module in the debugged kernel to finish much faster, while lx-symbols tracks module loads and unloads.

* Disable/enable all gdb breakpoints on both module load and unload breakpoint hits, and not only in 'load_all_symbols' as was done before. (load_all_symbols is no longer called on breakpoint hit)
That allows gdb to avoid getting confused about the state of the (now two) internal breakpoints we place.

Otherwise it will leave them in the kernel code segment, when continuing which triggers a guest kernel panic as soon as it skips over the 'int3' instruction and executes the garbage tail of the optcode on which the breakpoint was placed.

---
 kernel/module.c              |   8 ++-
 scripts/gdb/linux/symbols.py | 106 +++++++++++++++++++++++++----------
 2 files changed, 83 insertions(+), 31 deletions(-)
kernel/module.c +6 -2
index 30479355ab850..ea81fc06ea1f5 100644
@@ -901,8 +901,12 @@ int module_refcount(struct module *mod)
 }
 EXPORT_SYMBOL(module_refcount);
 
-/* This exists whether we can unload or not */
-static void free_module(struct module *mod);
+/* This exists whether we can unload or not
+ * Keep it uninlined to provide a reliable breakpoint target,
+ * e.g. for the gdb helper command 'lx-symbols'.
+ */
+
+static noinline void free_module(struct module *mod);
 
 SYSCALL_DEFINE2(delete_module, const char __user *, name_user,
 		unsigned int, flags)
scripts/gdb/linux/symbols.py +0 -0
index 1be9763cf8bb2..4ce879548a1ae 100644

scripts/gdb/linux/symbols.py import re · line 20

Quoted 4 lines
scripts/gdb/linux/symbols.py +1 -0
@@ -17,6 +17,24 @@ import re
 
 from linux import modules, utils
 
+def save_state():

Naming is a bit too generic. And it's not only saving the state, it's also disabling things.

scripts/gdb/linux/symbols.py import re · line 33

Quoted 13 lines
scripts/gdb/linux/symbols.py +13 -0
+        breakpoints = []
+        if hasattr(gdb, 'breakpoints') and not gdb.breakpoints() is None:
+            for bp in gdb.breakpoints():
+                breakpoints.append({'breakpoint': bp, 'enabled': bp.enabled})
+                bp.enabled = False
+
+        show_pagination = gdb.execute("show pagination", to_string=True)
+        pagination = show_pagination.endswith("on.\n")
+        gdb.execute("set pagination off")
+
+        return {"breakpoints":breakpoints, "show_pagination": show_pagination}
+
+def load_state(state):

Maybe rather something with "restore", to make naming balanced. Or is there a use case where "state" is not coming from the function above?

Quoted 8 lines
scripts/gdb/linux/symbols.py +4 -0
+    for breakpoint in state["breakpoints"]:
+        breakpoint['breakpoint'].enabled = breakpoint['enabled']
+    gdb.execute("set pagination %s" % ("on" if state["show_pagination"] else "off"))
+
 
 if hasattr(gdb, 'Breakpoint'):
     class LoadModuleBreakpoint(gdb.Breakpoint):

scripts/gdb/linux/symbols.py if hasattr(gdb, 'Breakpoint'): · line 53

Quoted 6 lines
scripts/gdb/linux/symbols.py +3 -0
@@ -30,26 +48,38 @@ if hasattr(gdb, 'Breakpoint'):
             module_name = module['name'].string()
             cmd = self.gdb_command
 
+            # module already loaded, false alarm
+            if module_name in cmd.loaded_modules:
+                return False

Possibly at all, now that we track unloading? Can our state tracking become out-of-sync?

Quoted 27 lines
scripts/gdb/linux/symbols.py +11 -9
+
             # enforce update if object file is not found
             cmd.module_files_updated = False
 
             # Disable pagination while reporting symbol (re-)loading.
             # The console input is blocked in this context so that we would
             # get stuck waiting for the user to acknowledge paged output.
-            show_pagination = gdb.execute("show pagination", to_string=True)
-            pagination = show_pagination.endswith("on.\n")
-            gdb.execute("set pagination off")
+            state = save_state()
+            cmd.load_module_symbols(module)
+            load_state(state)
+            return False
 
-            if module_name in cmd.loaded_modules:
-                gdb.write("refreshing all symbols to reload module "
-                          "'{0}'\n".format(module_name))
-                cmd.load_all_symbols()
-            else:
-                cmd.load_module_symbols(module)
+    class UnLoadModuleBreakpoint(gdb.Breakpoint):
+        def __init__(self, spec, gdb_command):
+            super(UnLoadModuleBreakpoint, self).__init__(spec, internal=True)
+            self.silent = True
+            self.gdb_command = gdb_command
+

scripts/gdb/linux/symbols.py if hasattr(gdb, 'Breakpoint'): · line 79

Quoted 10 lines
scripts/gdb/linux/symbols.py +6 -2
+        def stop(self):
+            module = gdb.parse_and_eval("mod")
+            module_name = module['name'].string()
+            cmd = self.gdb_command
 
-            # restore pagination state
-            gdb.execute("set pagination %s" % ("on" if pagination else "off"))
+            if not module_name in cmd.loaded_modules:
+                return False
 

Same question as above. For robustness, checking is not bad. But maybe it's worth reporting as well.

Quoted 92 lines
scripts/gdb/linux/symbols.py +32 -18
+            state = save_state()
+            cmd.unload_module_symbols(module)
+            load_state(state)
             return False
 
 
@@ -64,8 +94,9 @@ lx-symbols command."""
     module_paths = []
     module_files = []
     module_files_updated = False
-    loaded_modules = []
-    breakpoint = None
+    loaded_modules = {}
+    module_load_breakpoint = None
+    module_unload_breakpoint = None
 
     def __init__(self):
         super(LxSymbols, self).__init__("lx-symbols", gdb.COMMAND_FILES,
@@ -129,21 +160,32 @@ lx-symbols command."""
                 filename=module_file,
                 addr=module_addr,
                 sections=self._section_arguments(module))
+
             gdb.execute(cmdline, to_string=True)
-            if module_name not in self.loaded_modules:
-                self.loaded_modules.append(module_name)
+            self.loaded_modules[module_name] = {"module_file": module_file,
+                                                "module_addr": module_addr}
         else:
             gdb.write("no module object found for '{0}'\n".format(module_name))
 
+    def unload_module_symbols(self, module):
+        module_name = module['name'].string()
+
+        module_file = self.loaded_modules[module_name]["module_file"]
+        module_addr = self.loaded_modules[module_name]["module_addr"]
+
+        gdb.write("unloading @{addr}: {filename}\n".format(
+            addr=module_addr, filename=module_file))
+        cmdline = "remove-symbol-file {filename}".format(
+            filename=module_file)
+
+        gdb.execute(cmdline, to_string=True)
+        del self.loaded_modules[module_name]
+
+
     def load_all_symbols(self):
         gdb.write("loading vmlinux\n")
 
-        # Dropping symbols will disable all breakpoints. So save their states
-        # and restore them afterward.
-        saved_states = []
-        if hasattr(gdb, 'breakpoints') and not gdb.breakpoints() is None:
-            for bp in gdb.breakpoints():
-                saved_states.append({'breakpoint': bp, 'enabled': bp.enabled})
+        state = save_state()
 
         # drop all current symbols and reload vmlinux
         orig_vmlinux = 'vmlinux'
@@ -153,15 +195,14 @@ lx-symbols command."""
         gdb.execute("symbol-file", to_string=True)
         gdb.execute("symbol-file {0}".format(orig_vmlinux))
 
-        self.loaded_modules = []
+        self.loaded_modules = {}
         module_list = modules.module_list()
         if not module_list:
             gdb.write("no modules found\n")
         else:
             [self.load_module_symbols(module) for module in module_list]
 
-        for saved_state in saved_states:
-            saved_state['breakpoint'].enabled = saved_state['enabled']
+        load_state(state)
 
     def invoke(self, arg, from_tty):
         self.module_paths = [os.path.expanduser(p) for p in arg.split()]
@@ -174,11 +215,18 @@ lx-symbols command."""
         self.load_all_symbols()
 
         if hasattr(gdb, 'Breakpoint'):
-            if self.breakpoint is not None:
-                self.breakpoint.delete()
-                self.breakpoint = None
-            self.breakpoint = LoadModuleBreakpoint(
-                "kernel/module.c:do_init_module", self)
+            if self.module_load_breakpoint is not None:
+                self.module_load_breakpoint.delete()
+                self.module_load_breakpoint = None
+            self.module_load_breakpoint = \
+                LoadModuleBreakpoint("kernel/module.c:do_init_module", self)

scripts/gdb/linux/symbols.py lx-symbols command.""" · line 232

Quoted 10 lines
scripts/gdb/linux/symbols.py +7 -0
+
+            if self.module_unload_breakpoint is not None:
+                self.module_unload_breakpoint.delete()
+                self.module_unload_breakpoint = None
+            self.module_unload_breakpoint = \
+                UnLoadModuleBreakpoint("kernel/module.c:free_module", self)
+
         else:
             gdb.write("Note: symbol update on module loading not supported "
                       "with this gdb version\n")

Good improvement!

Jan

Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux
#3 lore raw

Re: [PATCH 1/3] scripts/gdb: rework lx-symbols gdb script

Quoted 65 lines · Jan Kiszka
Quoted 56 lines · Maxim Levitsky

Fix several issues that are present in lx-symbols script:

* Track module unloads by placing another software breakpoint at 'free_module' (force uninline this symbol just in case), and use remove-symbol-file gdb command to unload the symobls of the module that is unloading.

That gives the gdb a chance to mark all software breakpoints from this module as pending again.
Also remove the module from the 'known' module list once it is unloaded.

* Since we now track module unload, we don't need to reload all symbols anymore when 'known' module loaded again (that can't happen anymore). This allows reloading a module in the debugged kernel to finish much faster, while lx-symbols tracks module loads and unloads.

* Disable/enable all gdb breakpoints on both module load and unload breakpoint hits, and not only in 'load_all_symbols' as was done before. (load_all_symbols is no longer called on breakpoint hit)
That allows gdb to avoid getting confused about the state of the (now two) internal breakpoints we place.

Otherwise it will leave them in the kernel code segment, when continuing which triggers a guest kernel panic as soon as it skips over the 'int3' instruction and executes the garbage tail of the optcode on which the breakpoint was placed.

---
 kernel/module.c              |   8 ++-
 scripts/gdb/linux/symbols.py | 106 +++++++++++++++++++++++++----------
 2 files changed, 83 insertions(+), 31 deletions(-)
kernel/module.c +6 -2
index 30479355ab850..ea81fc06ea1f5 100644
@@ -901,8 +901,12 @@ int module_refcount(struct module *mod)
 }
 EXPORT_SYMBOL(module_refcount);
 
-/* This exists whether we can unload or not */
-static void free_module(struct module *mod);
+/* This exists whether we can unload or not
+ * Keep it uninlined to provide a reliable breakpoint target,
+ * e.g. for the gdb helper command 'lx-symbols'.
+ */
+
+static noinline void free_module(struct module *mod);
 
 SYSCALL_DEFINE2(delete_module, const char __user *, name_user,
 		unsigned int, flags)
scripts/gdb/linux/symbols.py +0 -0
index 1be9763cf8bb2..4ce879548a1ae 100644

scripts/gdb/linux/symbols.py import re · line 20

Quoted 4 lines
scripts/gdb/linux/symbols.py +1 -0
@@ -17,6 +17,24 @@ import re
 
 from linux import modules, utils
 
+def save_state():

Naming is a bit too generic. And it's not only saving the state, it's also disabling things.

Quoted 16 lines

scripts/gdb/linux/symbols.py import re · line 33

Quoted 13 lines
scripts/gdb/linux/symbols.py +13 -0
+        breakpoints = []
+        if hasattr(gdb, 'breakpoints') and not gdb.breakpoints() is None:
+            for bp in gdb.breakpoints():
+                breakpoints.append({'breakpoint': bp, 'enabled': bp.enabled})
+                bp.enabled = False
+
+        show_pagination = gdb.execute("show pagination", to_string=True)
+        pagination = show_pagination.endswith("on.\n")
+        gdb.execute("set pagination off")
+
+        return {"breakpoints":breakpoints, "show_pagination": show_pagination}
+
+def load_state(state):

Maybe rather something with "restore", to make naming balanced. Or is there a use case where "state" is not coming from the function above?

I didn't put much thought into naming these functions. I'll think of something better.

Quoted 12 lines
Quoted 12 lines
scripts/gdb/linux/symbols.py +4 -0
+    for breakpoint in state["breakpoints"]:
+        breakpoint['breakpoint'].enabled = breakpoint['enabled']
+    gdb.execute("set pagination %s" % ("on" if state["show_pagination"] else "off"))
+
 
 if hasattr(gdb, 'Breakpoint'):
     class LoadModuleBreakpoint(gdb.Breakpoint):
@@ -30,26 +48,38 @@ if hasattr(gdb, 'Breakpoint'):
             module_name = module['name'].string()
             cmd = self.gdb_command
 
Quoted 6 lines

scripts/gdb/linux/symbols.py if hasattr(gdb, 'Breakpoint'): · line 53

Quoted 3 lines
scripts/gdb/linux/symbols.py +3 -0
+            # module already loaded, false alarm
+            if module_name in cmd.loaded_modules:
+                return False

Possibly at all, now that we track unloading? Can our state tracking become out-of-sync?

Sadly yes and that happens a lot unless kvm is patched to avoid injecting interrupts on a single step.

What is happening is that this breakpoint is hit, then symbols are loaded which is a relatively slow process, and by the time the gdb resumes the guest, a timer interrupt is already pending (kernel local apic timer is running while vcpus are stopped), which makes the guest kernel take the interrupt (interrupts are not disabled in these two intercepted functions) and eventually return to the breakpoint, and trigger its python handler again.

This happens so often that especially with multiple vcpus, it is possible to enter live-lock like state where guest+gdb are stuck in this loop forever.

When KVM is patched, that indeed shouldn't happen but the check won't hurt here.

Plus due to the feedback I received on the patch 2, I will end up implementing the 'don't inject interrupts on single step' as a new KVM debug feature flag, thus you will need a newer qemu to make use of it.

Quoted 138 lines
Quoted 28 lines
scripts/gdb/linux/symbols.py +11 -9
+
             # enforce update if object file is not found
             cmd.module_files_updated = False
 
             # Disable pagination while reporting symbol (re-)loading.
             # The console input is blocked in this context so that we would
             # get stuck waiting for the user to acknowledge paged output.
-            show_pagination = gdb.execute("show pagination", to_string=True)
-            pagination = show_pagination.endswith("on.\n")
-            gdb.execute("set pagination off")
+            state = save_state()
+            cmd.load_module_symbols(module)
+            load_state(state)
+            return False
 
-            if module_name in cmd.loaded_modules:
-                gdb.write("refreshing all symbols to reload module "
-                          "'{0}'\n".format(module_name))
-                cmd.load_all_symbols()
-            else:
-                cmd.load_module_symbols(module)
+    class UnLoadModuleBreakpoint(gdb.Breakpoint):
+        def __init__(self, spec, gdb_command):
+            super(UnLoadModuleBreakpoint, self).__init__(spec, internal=True)
+            self.silent = True
+            self.gdb_command = gdb_command
+

scripts/gdb/linux/symbols.py if hasattr(gdb, 'Breakpoint'): · line 79

Quoted 10 lines
scripts/gdb/linux/symbols.py +6 -2
+        def stop(self):
+            module = gdb.parse_and_eval("mod")
+            module_name = module['name'].string()
+            cmd = self.gdb_command
 
-            # restore pagination state
-            gdb.execute("set pagination %s" % ("on" if pagination else "off"))
+            if not module_name in cmd.loaded_modules:
+                return False
 

Same question as above. For robustness, checking is not bad. But maybe it's worth reporting as well.

Quoted 96 lines
scripts/gdb/linux/symbols.py +37 -18
+            state = save_state()
+            cmd.unload_module_symbols(module)
+            load_state(state)
             return False
 
 
@@ -64,8 +94,9 @@ lx-symbols command."""
     module_paths = []
     module_files = []
     module_files_updated = False
-    loaded_modules = []
-    breakpoint = None
+    loaded_modules = {}
+    module_load_breakpoint = None
+    module_unload_breakpoint = None
 
     def __init__(self):
         super(LxSymbols, self).__init__("lx-symbols", gdb.COMMAND_FILES,
@@ -129,21 +160,32 @@ lx-symbols command."""
                 filename=module_file,
                 addr=module_addr,
                 sections=self._section_arguments(module))
+
             gdb.execute(cmdline, to_string=True)
-            if module_name not in self.loaded_modules:
-                self.loaded_modules.append(module_name)
+            self.loaded_modules[module_name] = {"module_file": module_file,
+                                                "module_addr": module_addr}
         else:
             gdb.write("no module object found for '{0}'\n".format(module_name))
 
+    def unload_module_symbols(self, module):
+        module_name = module['name'].string()
+
+        module_file = self.loaded_modules[module_name]["module_file"]
+        module_addr = self.loaded_modules[module_name]["module_addr"]
+
+        gdb.write("unloading @{addr}: {filename}\n".format(
+            addr=module_addr, filename=module_file))
+        cmdline = "remove-symbol-file {filename}".format(
+            filename=module_file)
+
+        gdb.execute(cmdline, to_string=True)
+        del self.loaded_modules[module_name]
+
+
     def load_all_symbols(self):
         gdb.write("loading vmlinux\n")
 
-        # Dropping symbols will disable all breakpoints. So save their states
-        # and restore them afterward.
-        saved_states = []
-        if hasattr(gdb, 'breakpoints') and not gdb.breakpoints() is None:
-            for bp in gdb.breakpoints():
-                saved_states.append({'breakpoint': bp, 'enabled': bp.enabled})
+        state = save_state()
 
         # drop all current symbols and reload vmlinux
         orig_vmlinux = 'vmlinux'
@@ -153,15 +195,14 @@ lx-symbols command."""
         gdb.execute("symbol-file", to_string=True)
         gdb.execute("symbol-file {0}".format(orig_vmlinux))
 
-        self.loaded_modules = []
+        self.loaded_modules = {}
         module_list = modules.module_list()
         if not module_list:
             gdb.write("no modules found\n")
         else:
             [self.load_module_symbols(module) for module in module_list]
 
-        for saved_state in saved_states:
-            saved_state['breakpoint'].enabled = saved_state['enabled']
+        load_state(state)
 
     def invoke(self, arg, from_tty):
         self.module_paths = [os.path.expanduser(p) for p in arg.split()]
@@ -174,11 +215,18 @@ lx-symbols command."""
         self.load_all_symbols()
 
         if hasattr(gdb, 'Breakpoint'):
-            if self.breakpoint is not None:
-                self.breakpoint.delete()
-                self.breakpoint = None
-            self.breakpoint = LoadModuleBreakpoint(
-                "kernel/module.c:do_init_module", self)
+            if self.module_load_breakpoint is not None:
+                self.module_load_breakpoint.delete()
+                self.module_load_breakpoint = None
+            self.module_load_breakpoint = \
+                LoadModuleBreakpoint("kernel/module.c:do_init_module", self)
+
+            if self.module_unload_breakpoint is not None:
+                self.module_unload_breakpoint.delete()
+                self.module_unload_breakpoint = None
+            self.module_unload_breakpoint = \
Quoted 8 lines

scripts/gdb/linux/symbols.py lx-symbols command.""" · line 232

Quoted 6 lines
scripts/gdb/linux/symbols.py +2 -0
+                UnLoadModuleBreakpoint("kernel/module.c:free_module", self)
+
         else:
             gdb.write("Note: symbol update on module loading not supported "
                       "with this gdb version\n")

Good improvement!

Thanks a lot!

Best regards,
Maxim Levitsky

Quoted 3 lines

Jan

#4 lore raw

PATCH 2/3 KVM: x86: guest debug: don't inject interrupts while single stepping

In reply to Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +6 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;
+
 	/*
 	 * Finally, inject interrupt events.  If an event cannot be injected
 	 * due to architectural conditions (e.g. IF=0) a window-open exit
Signature · 1 line
2.26.2
#5 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 39 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 8 lines
arch/x86/kvm/x86.c +5 -0
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Quoted 7 lines
arch/x86/kvm/x86.c +1 -0
+
 	/*
 	 * Finally, inject interrupt events.  If an event cannot be injected
 	 * due to architectural conditions (e.g. IF=0) a window-open exit
Signature · 1 line
2.26.2
maintainer
#6 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 10 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Does this happen to address this suspicious workaround in the kernel?

        /*
         * The kernel doesn't use TF single-step outside of:
         *
         *  - Kprobes, consumed through kprobe_debug_handler()
         *  - KGDB, consumed through notify_debug()
         *
         * So if we get here with DR_STEP set, something is wonky.
         *
         * A known way to trigger this is through QEMU's GDB stub,
         * which leaks #DB into the guest and causes IST recursion.
         */
        if (WARN_ON_ONCE(dr6 & DR_STEP))
                regs->flags &= ~X86_EFLAGS_TF;

(arch/x86/kernel/traps.c, exc_debug_kernel)

I wonder why this got merged while no one fixed QEMU/KVM, for years? Oh, yeah, question to myself as well, dancing around broken guest debugging for a long time while trying to fix other issues...

Jan

Quoted 7 lines
Quoted 7 lines
arch/x86/kvm/x86.c +1 -0
+
 	/*
 	 * Finally, inject interrupt events.  If an event cannot be injected
 	 * due to architectural conditions (e.g. IF=0) a window-open exit
Signature · 1 line
2.26.2
Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux
#7 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 71 lines · Jan Kiszka
Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 10 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Does this happen to address this suspicious workaround in the kernel?

        /*
         * The kernel doesn't use TF single-step outside of:
         *
         *  - Kprobes, consumed through kprobe_debug_handler()
         *  - KGDB, consumed through notify_debug()
         *
         * So if we get here with DR_STEP set, something is wonky.
         *
         * A known way to trigger this is through QEMU's GDB stub,
         * which leaks #DB into the guest and causes IST recursion.
         */
        if (WARN_ON_ONCE(dr6 & DR_STEP))
                regs->flags &= ~X86_EFLAGS_TF;
Quoted 5 lines

(arch/x86/kernel/traps.c, exc_debug_kernel)

I wonder why this got merged while no one fixed QEMU/KVM, for years? Oh, yeah, question to myself as well, dancing around broken guest debugging for a long time while trying to fix other issues...

To be honest I didn't see that warning even once, but I can imagine KVM leaking #DB due to bugs in that code. That area historically didn't receive much attention since it can only be triggered by KVM_GET/SET_GUEST_DEBUG which isn't used in production.

The only issue that I on the other hand did
see which is mostly gdb fault is that it fails to remove a software breakpoint when resuming over it, if that breakpoint's python handler messes up with gdb's symbols, which is what lx-symbols does.

And that despite the fact that lx-symbol doesn't mess with the object (that is the kernel) where the breakpoint is defined.

Just adding/removing one symbol file is enough to trigger this issue.

Since lx-symbols already works this around when it reloads all symbols, I extended that workaround to happen also when loading/unloading only a single symbol file.

Best regards,
Maxim Levitsky

Quoted 10 lines

Jan

Quoted 7 lines
Quoted 7 lines
arch/x86/kvm/x86.c +1 -0
+
 	/*
 	 * Finally, inject interrupt events.  If an event cannot be injected
 	 * due to architectural conditions (e.g. IF=0) a window-open exit
Signature · 1 line
2.26.2
maintainer
#8 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 74 lines · Maxim Levitsky
Quoted 73 lines · Jan Kiszka
Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 10 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Does this happen to address this suspicious workaround in the kernel?

        /*
         * The kernel doesn't use TF single-step outside of:
         *
         *  - Kprobes, consumed through kprobe_debug_handler()
         *  - KGDB, consumed through notify_debug()
         *
         * So if we get here with DR_STEP set, something is wonky.
         *
         * A known way to trigger this is through QEMU's GDB stub,
         * which leaks #DB into the guest and causes IST recursion.
         */
        if (WARN_ON_ONCE(dr6 & DR_STEP))
                regs->flags &= ~X86_EFLAGS_TF;

(arch/x86/kernel/traps.c, exc_debug_kernel)

Quoted 8 lines
Quoted 3 lines

I wonder why this got merged while no one fixed QEMU/KVM, for years? Oh, yeah, question to myself as well, dancing around broken guest debugging for a long time while trying to fix other issues...

To be honest I didn't see that warning even once, but I can imagine KVM leaking #DB due to bugs in that code. That area historically didn't receive much attention since it can only be triggered by KVM_GET/SET_GUEST_DEBUG which isn't used in production.

I've triggered it recently while debugging a guest, that's why I got aware of the code path. Long ago, all this used to work (soft BPs, single-stepping etc.)

Quoted 6 lines

The only issue that I on the other hand did
see which is mostly gdb fault is that it fails to remove a software breakpoint when resuming over it, if that breakpoint's python handler messes up with gdb's symbols, which is what lx-symbols does.

Quoted 8 lines

And that despite the fact that lx-symbol doesn't mess with the object (that is the kernel) where the breakpoint is defined.

Just adding/removing one symbol file is enough to trigger this issue.

Since lx-symbols already works this around when it reloads all symbols, I extended that workaround to happen also when loading/unloading only a single symbol file.

You have no issue with interactive debugging when NOT using gdb scripts / lx-symbol?

Jan

Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux
#9 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 96 lines · Jan Kiszka
Quoted 74 lines · Maxim Levitsky
Quoted 73 lines · Jan Kiszka
Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 10 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Does this happen to address this suspicious workaround in the kernel?

        /*
         * The kernel doesn't use TF single-step outside of:
         *
         *  - Kprobes, consumed through kprobe_debug_handler()
         *  - KGDB, consumed through notify_debug()
         *
         * So if we get here with DR_STEP set, something is wonky.
         *
         * A known way to trigger this is through QEMU's GDB stub,
         * which leaks #DB into the guest and causes IST recursion.
         */
        if (WARN_ON_ONCE(dr6 & DR_STEP))
                regs->flags &= ~X86_EFLAGS_TF;

(arch/x86/kernel/traps.c, exc_debug_kernel)

Quoted 8 lines
Quoted 3 lines

I wonder why this got merged while no one fixed QEMU/KVM, for years? Oh, yeah, question to myself as well, dancing around broken guest debugging for a long time while trying to fix other issues...

To be honest I didn't see that warning even once, but I can imagine KVM leaking #DB due to bugs in that code. That area historically didn't receive much attention since it can only be triggered by KVM_GET/SET_GUEST_DEBUG which isn't used in production.

I've triggered it recently while debugging a guest, that's why I got aware of the code path. Long ago, all this used to work (soft BPs, single-stepping etc.)

Quoted 8 lines

The only issue that I on the other hand did
see which is mostly gdb fault is that it fails to remove a software breakpoint when resuming over it, if that breakpoint's python handler messes up with gdb's symbols, which is what lx-symbols does.

And that despite the fact that lx-symbol doesn't mess with the object (that is the kernel) where the breakpoint is defined.

Quoted 8 lines
Quoted 5 lines

Just adding/removing one symbol file is enough to trigger this issue.

Since lx-symbols already works this around when it reloads all symbols, I extended that workaround to happen also when loading/unloading only a single symbol file.

You have no issue with interactive debugging when NOT using gdb scripts / lx-symbol?

To be honest I don't use guest debugging that much, so I probably missed some issues.

Now that I fixed lx-symbols though I'll probably use guest debugging much more.
I will keep an eye on any issues that I find.

The main push to fix lx-symbols actually came from me wanting to understand if there is something broken with KVM's guest debugging knowing that lx-symbols crashes the guest when module is loaded after lx-symbols was executed.

That lx-symbols related guest crash I traced to issue with gdb as I explained, and the lack of blocking of the interrupts on single step is not a bug but more a missing feature that should be implemented to make single step easier to use.

Another issue which isn't a bug is that you can't place a software breakpoint if kernel is not loaded (since there is no code in memory) or if the kernel haven't done basic paging initialization (since there is no paging yet to know where to place the breakpoint). Hardware breakpoints work for this fine though.

So in summary I haven't found any major issues with KVM's guest debug yet.

If I do notice issues with guest debug, I will try to isolate and debug them.
For the issue that you mentioned, do you have a way to reproduce it?

Best regards,
Maxim Levitsky

Quoted 3 lines

Jan

maintainer
#10 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 119 lines · Maxim Levitsky
Quoted 96 lines · Jan Kiszka
Quoted 74 lines · Maxim Levitsky
Quoted 73 lines · Jan Kiszka
Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 10 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Does this happen to address this suspicious workaround in the kernel?

        /*
         * The kernel doesn't use TF single-step outside of:
         *
         *  - Kprobes, consumed through kprobe_debug_handler()
         *  - KGDB, consumed through notify_debug()
         *
         * So if we get here with DR_STEP set, something is wonky.
         *
         * A known way to trigger this is through QEMU's GDB stub,
         * which leaks #DB into the guest and causes IST recursion.
         */
        if (WARN_ON_ONCE(dr6 & DR_STEP))
                regs->flags &= ~X86_EFLAGS_TF;

(arch/x86/kernel/traps.c, exc_debug_kernel)

Quoted 8 lines
Quoted 3 lines

I wonder why this got merged while no one fixed QEMU/KVM, for years? Oh, yeah, question to myself as well, dancing around broken guest debugging for a long time while trying to fix other issues...

To be honest I didn't see that warning even once, but I can imagine KVM leaking #DB due to bugs in that code. That area historically didn't receive much attention since it can only be triggered by KVM_GET/SET_GUEST_DEBUG which isn't used in production.

I've triggered it recently while debugging a guest, that's why I got aware of the code path. Long ago, all this used to work (soft BPs, single-stepping etc.)

Quoted 8 lines

The only issue that I on the other hand did
see which is mostly gdb fault is that it fails to remove a software breakpoint when resuming over it, if that breakpoint's python handler messes up with gdb's symbols, which is what lx-symbols does.

And that despite the fact that lx-symbol doesn't mess with the object (that is the kernel) where the breakpoint is defined.

Quoted 8 lines
Quoted 5 lines

Just adding/removing one symbol file is enough to trigger this issue.

Since lx-symbols already works this around when it reloads all symbols, I extended that workaround to happen also when loading/unloading only a single symbol file.

You have no issue with interactive debugging when NOT using gdb scripts / lx-symbol?

To be honest I don't use guest debugging that much, so I probably missed some issues.

Now that I fixed lx-symbols though I'll probably use guest debugging much more.
I will keep an eye on any issues that I find.

The main push to fix lx-symbols actually came from me wanting to understand if there is something broken with KVM's guest debugging knowing that lx-symbols crashes the guest when module is loaded after lx-symbols was executed.

Quoted 4 lines

That lx-symbols related guest crash I traced to issue with gdb as I explained, and the lack of blocking of the interrupts on single step is not a bug but more a missing feature that should be implemented to make single step easier to use.

Again, this used to work fine. But maybe this patch can change the picture by avoid that the unavoidable short TF leakage into the guest escalates beyond the single instruction to step over.

Quoted 7 lines

Another issue which isn't a bug is that you can't place a software breakpoint if kernel is not loaded (since there is no code in memory) or if the kernel haven't done basic paging initialization (since there is no paging yet to know where to place the breakpoint). Hardware breakpoints work for this fine though.

Quoted 6 lines

So in summary I haven't found any major issues with KVM's guest debug yet.

If I do notice issues with guest debug, I will try to isolate and debug them.
For the issue that you mentioned, do you have a way to reproduce it?

I need to spend some time in factoring out a clean test setup, will come back to you. I'm always pushing this to the back - and then grumble when hitting it while debugging something urgent. Your patch is a nice reason to do this systematically now.

Jan

Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux
#11 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 120 lines · Jan Kiszka
Quoted 119 lines · Maxim Levitsky
Quoted 96 lines · Jan Kiszka
Quoted 74 lines · Maxim Levitsky
Quoted 73 lines · Jan Kiszka
Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 10 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Does this happen to address this suspicious workaround in the kernel?

        /*
         * The kernel doesn't use TF single-step outside of:
         *
         *  - Kprobes, consumed through kprobe_debug_handler()
         *  - KGDB, consumed through notify_debug()
         *
         * So if we get here with DR_STEP set, something is wonky.
         *
         * A known way to trigger this is through QEMU's GDB stub,
         * which leaks #DB into the guest and causes IST recursion.
         */
        if (WARN_ON_ONCE(dr6 & DR_STEP))
                regs->flags &= ~X86_EFLAGS_TF;

(arch/x86/kernel/traps.c, exc_debug_kernel)

Quoted 8 lines
Quoted 3 lines

I wonder why this got merged while no one fixed QEMU/KVM, for years? Oh, yeah, question to myself as well, dancing around broken guest debugging for a long time while trying to fix other issues...

To be honest I didn't see that warning even once, but I can imagine KVM leaking #DB due to bugs in that code. That area historically didn't receive much attention since it can only be triggered by KVM_GET/SET_GUEST_DEBUG which isn't used in production.

I've triggered it recently while debugging a guest, that's why I got aware of the code path. Long ago, all this used to work (soft BPs, single-stepping etc.)

Quoted 8 lines

The only issue that I on the other hand did
see which is mostly gdb fault is that it fails to remove a software breakpoint when resuming over it, if that breakpoint's python handler messes up with gdb's symbols, which is what lx-symbols does.

And that despite the fact that lx-symbol doesn't mess with the object (that is the kernel) where the breakpoint is defined.

Quoted 8 lines
Quoted 5 lines

Just adding/removing one symbol file is enough to trigger this issue.

Since lx-symbols already works this around when it reloads all symbols, I extended that workaround to happen also when loading/unloading only a single symbol file.

You have no issue with interactive debugging when NOT using gdb scripts / lx-symbol?

To be honest I don't use guest debugging that much, so I probably missed some issues.

Now that I fixed lx-symbols though I'll probably use guest debugging much more.
I will keep an eye on any issues that I find.

The main push to fix lx-symbols actually came from me wanting to understand if there is something broken with KVM's guest debugging knowing that lx-symbols crashes the guest when module is loaded after lx-symbols was executed.

Quoted 8 lines
Quoted 4 lines

That lx-symbols related guest crash I traced to issue with gdb as I explained, and the lack of blocking of the interrupts on single step is not a bug but more a missing feature that should be implemented to make single step easier to use.

Again, this used to work fine. But maybe this patch can change the picture by avoid that the unavoidable short TF leakage into the guest escalates beyond the single instruction to step over.

Actually now I think I understand what is going on.

The TF flag isn't auto cleared as RF flag is, and if the instruction which is single stepped gets an interrupt it is pushed onto the interrupt stack. (then it is cleared for the duration of the interrupt handler) Since we use the TF flag for single stepping the guest, this indeed can cause it to be leaked.

So this patch actually should mitigate this almost completely.

Also now I understand why Intel has the 'monitor trap' feature, I think it is exactly for the cases when hypervisor wants to single step the guest without the fear of changing of the guest visible cpu state.

KVM on VMX should probably switch to using monitor trap for single stepping.

Best regards,
Maxim Levitsky

Quoted 10 lines
Quoted 10 lines

Another issue which isn't a bug is that you can't place a software breakpoint if kernel is not loaded (since there is no code in memory) or if the kernel haven't done basic paging initialization (since there is no paging yet to know where to place the breakpoint). Hardware breakpoints work for this fine though.

So in summary I haven't found any major issues with KVM's guest debug yet.

Quoted 8 lines
Quoted 3 lines

If I do notice issues with guest debug, I will try to isolate and debug them.
For the issue that you mentioned, do you have a way to reproduce it?

I need to spend some time in factoring out a clean test setup, will come back to you. I'm always pushing this to the back - and then grumble when hitting it while debugging something urgent. Your patch is a nice reason to do this systematically now.

Thanks for the review,
Best regards,
Maxim Levitsky

Quoted 3 lines

Jan

maintainer
#12 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 139 lines · Maxim Levitsky
Quoted 120 lines · Jan Kiszka
Quoted 119 lines · Maxim Levitsky
Quoted 96 lines · Jan Kiszka
Quoted 74 lines · Maxim Levitsky
Quoted 73 lines · Jan Kiszka
Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 10 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Does this happen to address this suspicious workaround in the kernel?

        /*
         * The kernel doesn't use TF single-step outside of:
         *
         *  - Kprobes, consumed through kprobe_debug_handler()
         *  - KGDB, consumed through notify_debug()
         *
         * So if we get here with DR_STEP set, something is wonky.
         *
         * A known way to trigger this is through QEMU's GDB stub,
         * which leaks #DB into the guest and causes IST recursion.
         */
        if (WARN_ON_ONCE(dr6 & DR_STEP))
                regs->flags &= ~X86_EFLAGS_TF;

(arch/x86/kernel/traps.c, exc_debug_kernel)

Quoted 8 lines
Quoted 3 lines

I wonder why this got merged while no one fixed QEMU/KVM, for years? Oh, yeah, question to myself as well, dancing around broken guest debugging for a long time while trying to fix other issues...

To be honest I didn't see that warning even once, but I can imagine KVM leaking #DB due to bugs in that code. That area historically didn't receive much attention since it can only be triggered by KVM_GET/SET_GUEST_DEBUG which isn't used in production.

I've triggered it recently while debugging a guest, that's why I got aware of the code path. Long ago, all this used to work (soft BPs, single-stepping etc.)

Quoted 8 lines

The only issue that I on the other hand did
see which is mostly gdb fault is that it fails to remove a software breakpoint when resuming over it, if that breakpoint's python handler messes up with gdb's symbols, which is what lx-symbols does.

And that despite the fact that lx-symbol doesn't mess with the object (that is the kernel) where the breakpoint is defined.

Quoted 8 lines
Quoted 5 lines

Just adding/removing one symbol file is enough to trigger this issue.

Since lx-symbols already works this around when it reloads all symbols, I extended that workaround to happen also when loading/unloading only a single symbol file.

You have no issue with interactive debugging when NOT using gdb scripts / lx-symbol?

To be honest I don't use guest debugging that much, so I probably missed some issues.

Now that I fixed lx-symbols though I'll probably use guest debugging much more.
I will keep an eye on any issues that I find.

The main push to fix lx-symbols actually came from me wanting to understand if there is something broken with KVM's guest debugging knowing that lx-symbols crashes the guest when module is loaded after lx-symbols was executed.

Quoted 8 lines
Quoted 4 lines

That lx-symbols related guest crash I traced to issue with gdb as I explained, and the lack of blocking of the interrupts on single step is not a bug but more a missing feature that should be implemented to make single step easier to use.

Again, this used to work fine. But maybe this patch can change the picture by avoid that the unavoidable short TF leakage into the guest escalates beyond the single instruction to step over.

Actually now I think I understand what is going on.

The TF flag isn't auto cleared as RF flag is, and if the instruction which is single stepped gets an interrupt it is pushed onto the interrupt stack. (then it is cleared for the duration of the interrupt handler) Since we use the TF flag for single stepping the guest, this indeed can cause it to be leaked.

Quoted 5 lines

So this patch actually should mitigate this almost completely.

Also now I understand why Intel has the 'monitor trap' feature, I think it is exactly for the cases when hypervisor wants to single step the guest without the fear of changing of the guest visible cpu state.

Exactly.

Quoted 2 lines

KVM on VMX should probably switch to using monitor trap for single stepping.

Back then, when I was hacking on the gdb-stub and KVM support, the monitor trap flag was not yet broadly available, but the idea to once use it was already there. Now it can be considered broadly available, but it would still require some changes to get it in.

Unfortunately, we don't have such thing with SVM, even recent versions, right? So, a proper way of avoiding diverting event injections while we are having the guest in an "incorrect" state should definitely be the goal.

Given that KVM knows whether TF originates solely from guest debugging or was (also) injected by the guest, we should be able to identify the cases where your approach is best to apply. And that without any extra control knob that everyone will only forget to set.

Jan

Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux

A message that is not in the archive (sent off-list or not yet received): [email protected]

maintainer
#14 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 163 lines · Maxim Levitsky
Quoted 151 lines · Jan Kiszka
Quoted 139 lines · Maxim Levitsky
Quoted 120 lines · Jan Kiszka
Quoted 119 lines · Maxim Levitsky
Quoted 96 lines · Jan Kiszka
Quoted 74 lines · Maxim Levitsky
Quoted 73 lines · Jan Kiszka
Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 10 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Does this happen to address this suspicious workaround in the kernel?

        /*
         * The kernel doesn't use TF single-step outside of:
         *
         *  - Kprobes, consumed through kprobe_debug_handler()
         *  - KGDB, consumed through notify_debug()
         *
         * So if we get here with DR_STEP set, something is wonky.
         *
         * A known way to trigger this is through QEMU's GDB stub,
         * which leaks #DB into the guest and causes IST recursion.
         */
        if (WARN_ON_ONCE(dr6 & DR_STEP))
                regs->flags &= ~X86_EFLAGS_TF;

(arch/x86/kernel/traps.c, exc_debug_kernel)

Quoted 8 lines
Quoted 3 lines

I wonder why this got merged while no one fixed QEMU/KVM, for years? Oh, yeah, question to myself as well, dancing around broken guest debugging for a long time while trying to fix other issues...

To be honest I didn't see that warning even once, but I can imagine KVM leaking #DB due to bugs in that code. That area historically didn't receive much attention since it can only be triggered by KVM_GET/SET_GUEST_DEBUG which isn't used in production.

I've triggered it recently while debugging a guest, that's why I got aware of the code path. Long ago, all this used to work (soft BPs, single-stepping etc.)

Quoted 8 lines

The only issue that I on the other hand did
see which is mostly gdb fault is that it fails to remove a software breakpoint when resuming over it, if that breakpoint's python handler messes up with gdb's symbols, which is what lx-symbols does.

And that despite the fact that lx-symbol doesn't mess with the object (that is the kernel) where the breakpoint is defined.

Quoted 8 lines
Quoted 5 lines

Just adding/removing one symbol file is enough to trigger this issue.

Since lx-symbols already works this around when it reloads all symbols, I extended that workaround to happen also when loading/unloading only a single symbol file.

You have no issue with interactive debugging when NOT using gdb scripts / lx-symbol?

To be honest I don't use guest debugging that much, so I probably missed some issues.

Now that I fixed lx-symbols though I'll probably use guest debugging much more.
I will keep an eye on any issues that I find.

The main push to fix lx-symbols actually came from me wanting to understand if there is something broken with KVM's guest debugging knowing that lx-symbols crashes the guest when module is loaded after lx-symbols was executed.

Quoted 8 lines
Quoted 4 lines

That lx-symbols related guest crash I traced to issue with gdb as I explained, and the lack of blocking of the interrupts on single step is not a bug but more a missing feature that should be implemented to make single step easier to use.

Again, this used to work fine. But maybe this patch can change the picture by avoid that the unavoidable short TF leakage into the guest escalates beyond the single instruction to step over.

Actually now I think I understand what is going on.

The TF flag isn't auto cleared as RF flag is, and if the instruction which is single stepped gets an interrupt it is pushed onto the interrupt stack. (then it is cleared for the duration of the interrupt handler) Since we use the TF flag for single stepping the guest, this indeed can cause it to be leaked.

Quoted 5 lines

So this patch actually should mitigate this almost completely.

Also now I understand why Intel has the 'monitor trap' feature, I think it is exactly for the cases when hypervisor wants to single step the guest without the fear of changing of the guest visible cpu state.

Exactly.

Quoted 2 lines

KVM on VMX should probably switch to using monitor trap for single stepping.

Quoted 8 lines

Back then, when I was hacking on the gdb-stub and KVM support, the monitor trap flag was not yet broadly available, but the idea to once use it was already there. Now it can be considered broadly available, but it would still require some changes to get it in.

Unfortunately, we don't have such thing with SVM, even recent versions, right? So, a proper way of avoiding diverting event injections while we are having the guest in an "incorrect" state should definitely be the goal.

Yes, I am not aware of anything like monitor trap on SVM.

Quoted 1 line
Quoted 8 lines
Quoted 4 lines

Given that KVM knows whether TF originates solely from guest debugging or was (also) injected by the guest, we should be able to identify the cases where your approach is best to apply. And that without any extra control knob that everyone will only forget to set.

Well I think that the downside of this patch is that the user might actually want to single step into an interrupt handler,
and this patch makes it a bit more complicated, and changes the default behavior.

If the default makes debugging practically impossible and breaks the guest by leaking host state, that is also not OK.

We must not leak, that is priority one. So, even if we consider stepping into the interrupt a use case (surely not the default one, so having this opt-in only), we must ensure that TF will never end up on the guest stack saved for the interrupt handler. That is KVM's default responsibility.

Jan

Quoted 14 lines

I have no objections though to use this patch as is, or at least make this the new default with a new flag to override this.

Sean Christopherson, what do you think?

Best regards,
Maxim Levitsky

Quoted 3 lines

Jan

Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux
#15 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

In reply to a message not in the archive
Quoted 12 lines · Maxim Levitsky
Quoted 8 lines · Jan Kiszka

Back then, when I was hacking on the gdb-stub and KVM support, the monitor trap flag was not yet broadly available, but the idea to once use it was already there. Now it can be considered broadly available, but it would still require some changes to get it in.

Unfortunately, we don't have such thing with SVM, even recent versions, right? So, a proper way of avoiding diverting event injections while we are having the guest in an "incorrect" state should definitely be the goal.

Yes, I am not aware of anything like monitor trap on SVM.

Quoted 1 line
Quoted 7 lines
Quoted 4 lines

Given that KVM knows whether TF originates solely from guest debugging or was (also) injected by the guest, we should be able to identify the cases where your approach is best to apply. And that without any extra control knob that everyone will only forget to set.

Well I think that the downside of this patch is that the user might actually want to single step into an interrupt handler, and this patch makes it a bit more complicated, and changes the default behavior.

Yes. And, as is, this also blocks NMIs and SMIs. I suspect it also doesn't prevent weirdness if the guest is running in L2, since IRQs for L1 will cause exits from L2 during nested_ops->check_events().

Quoted 2 lines

I have no objections though to use this patch as is, or at least make this the new default with a new flag to override this.

That's less bad, but IMO still violates the principle of least surprise, e.g. someone that is single-stepping a guest and is expecting an IRQ to fire will be all kinds of confused if they see all the proper IRR, ISR, EFLAGS.IF, etc... settings, but no interrupt.

Quoted 1 line

Sean Christopherson, what do you think?

Rather than block all events in KVM, what about having QEMU "pause" the timer? E.g. save MSR_TSC_DEADLINE and APIC_TMICT (or inspect the guest to find out which flavor it's using), clear them to zero, then restore both when single-stepping is disabled. I think that will work?

maintainer
#16 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 25 lines · Sean Christopherson
Quoted 12 lines · Maxim Levitsky
Quoted 8 lines · Jan Kiszka

Back then, when I was hacking on the gdb-stub and KVM support, the monitor trap flag was not yet broadly available, but the idea to once use it was already there. Now it can be considered broadly available, but it would still require some changes to get it in.

Unfortunately, we don't have such thing with SVM, even recent versions, right? So, a proper way of avoiding diverting event injections while we are having the guest in an "incorrect" state should definitely be the goal.

Yes, I am not aware of anything like monitor trap on SVM.

Quoted 1 line
Quoted 7 lines
Quoted 4 lines

Given that KVM knows whether TF originates solely from guest debugging or was (also) injected by the guest, we should be able to identify the cases where your approach is best to apply. And that without any extra control knob that everyone will only forget to set.

Well I think that the downside of this patch is that the user might actually want to single step into an interrupt handler, and this patch makes it a bit more complicated, and changes the default behavior.

Yes. And, as is, this also blocks NMIs and SMIs. I suspect it also doesn't prevent weirdness if the guest is running in L2, since IRQs for L1 will cause exits from L2 during nested_ops->check_events().

Quoted 7 lines
Quoted 2 lines

I have no objections though to use this patch as is, or at least make this the new default with a new flag to override this.

That's less bad, but IMO still violates the principle of least surprise, e.g. someone that is single-stepping a guest and is expecting an IRQ to fire will be all kinds of confused if they see all the proper IRR, ISR, EFLAGS.IF, etc... settings, but no interrupt.

From my practical experience with debugging guests via single step, seeing an interrupt in that case is everything but handy and generally also not expected (though logical, I agree). IOW: When there is a knob for it, it will remain off in 99% of the time.

But I see the point of having some control, in an ideal world also an indication that there are pending events, permitting the user to decide what to do. But I suspect the gdb frontend and protocol does not easily permit that.

Quoted 8 lines
Quoted 2 lines

Sean Christopherson, what do you think?

Rather than block all events in KVM, what about having QEMU "pause" the timer? E.g. save MSR_TSC_DEADLINE and APIC_TMICT (or inspect the guest to find out which flavor it's using), clear them to zero, then restore both when single-stepping is disabled. I think that will work?

No one can stop the clock, and timers are only one source of interrupts. Plus they do not all come from QEMU, some also from KVM or in-kernel sources directly. Would quickly become a mess.

Jan

Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux
#17 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 10 lines · Jan Kiszka
Quoted 5 lines · Sean Christopherson

Rather than block all events in KVM, what about having QEMU "pause" the timer? E.g. save MSR_TSC_DEADLINE and APIC_TMICT (or inspect the guest to find out which flavor it's using), clear them to zero, then restore both when single-stepping is disabled. I think that will work?

No one can stop the clock, and timers are only one source of interrupts. Plus they do not all come from QEMU, some also from KVM or in-kernel sources directly.

But are any other sources of interrupts a chronic problem? I 100% agree that this would not be a robust solution, but neither is blocking events in KVM. At least with this approach, the blast radius is somewhat contained.

Quoted 1 line

Would quickly become a mess.

Maybe, but it'd be Qemu's mess ;-)

maintainer
#18 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

Quoted 8 lines · Sean Christopherson
Quoted 7 lines · Jan Kiszka
Quoted 5 lines · Sean Christopherson

Rather than block all events in KVM, what about having QEMU "pause" the timer? E.g. save MSR_TSC_DEADLINE and APIC_TMICT (or inspect the guest to find out which flavor it's using), clear them to zero, then restore both when single-stepping is disabled. I think that will work?

Quoted 5 lines
Quoted 3 lines

No one can stop the clock, and timers are only one source of interrupts. Plus they do not all come from QEMU, some also from KVM or in-kernel sources directly.

But are any other sources of interrupts a chronic problem? I 100% agree that

If you are debugging a problem, you are not interested in seening problems of the debugger, only real ones of your target. IOW: Yes, they are, even if less likely - for idle VMs.

Quoted 7 lines

this would not be a robust solution, but neither is blocking events in KVM. At least with this approach, the blast radius is somewhat contained.

Quoted 1 line

Would quickly become a mess.

Maybe, but it'd be Qemu's mess ;-)

Nope, it would spread to KVM as well, as indicated above.

Jan

Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux
#19 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

In reply to Jan Kiszka
Quoted 38 lines · Jan Kiszka
Quoted 24 lines · Sean Christopherson
Quoted 11 lines · Maxim Levitsky
Quoted 8 lines · Jan Kiszka

Back then, when I was hacking on the gdb-stub and KVM support, the monitor trap flag was not yet broadly available, but the idea to once use it was already there. Now it can be considered broadly available, but it would still require some changes to get it in.

Unfortunately, we don't have such thing with SVM, even recent versions, right? So, a proper way of avoiding diverting event injections while we are having the guest in an "incorrect" state should definitely be the goal.

Yes, I am not aware of anything like monitor trap on SVM.

Quoted 7 lines
Quoted 4 lines

Given that KVM knows whether TF originates solely from guest debugging or was (also) injected by the guest, we should be able to identify the cases where your approach is best to apply. And that without any extra control knob that everyone will only forget to set.

Well I think that the downside of this patch is that the user might actually want to single step into an interrupt handler, and this patch makes it a bit more complicated, and changes the default behavior.

Yes. And, as is, this also blocks NMIs and SMIs. I suspect it also doesn't prevent weirdness if the guest is running in L2, since IRQs for L1 will cause exits from L2 during nested_ops->check_events().

Quoted 7 lines
Quoted 2 lines

I have no objections though to use this patch as is, or at least make this the new default with a new flag to override this.

That's less bad, but IMO still violates the principle of least surprise, e.g. someone that is single-stepping a guest and is expecting an IRQ to fire will be all kinds of confused if they see all the proper IRR, ISR, EFLAGS.IF, etc... settings, but no interrupt.

From my practical experience with debugging guests via single step, seeing an interrupt in that case is everything but handy and generally also not expected (though logical, I agree). IOW: When there is a knob for it, it will remain off in 99% of the time.

Quoted 4 lines

But I see the point of having some control, in an ideal world also an indication that there are pending events, permitting the user to decide what to do. But I suspect the gdb frontend and protocol does not easily permit that.

Qemu gdbstub actually does have control over suppression of the interrupts over a single step and it is even enabled by default:

https://qemu.readthedocs.io/en/latest/system/gdb.html
(advanced debug options)

However it is currently only implemented in TCG (software emulator) mode and not in KVM mode (I can argue that this is a qemu bug).

So my plan was to add a new kvm guest debug flag KVM_GUESTDBG_BLOCKEVENTS, and let qemu enable it when its 'NOIRQ' mode is enabled (it is by default).

However due to the discussion in this thread about the leakage of the RFLAGS.TF, I wonder if kvm should by default suppress events and have something like KVM_GUESTDBG_SSTEP_ALLOW_EVENTS to override this and wire
that to qemu's NOIRQ=false case.

This will allow older qemu to work correctly and new qemu will be able to choose the old less ideal behavior.

Quoted 9 lines
Quoted 8 lines
Quoted 2 lines

Sean Christopherson, what do you think?

Rather than block all events in KVM, what about having QEMU "pause" the timer? E.g. save MSR_TSC_DEADLINE and APIC_TMICT (or inspect the guest to find out which flavor it's using), clear them to zero, then restore both when single-stepping is disabled. I think that will work?

Quoted 3 lines

No one can stop the clock, and timers are only one source of interrupts. Plus they do not all come from QEMU, some also from KVM or in-kernel sources directly. Would quickly become a mess.

This, plus as we see, even changing with RFLAGS.TF leaks it. Changing things like MSR_TSC_DEADLINE will also make it visible to the guest, sooner or later and is a mess that I rather not get into.

It is _possible_ to disable timer interrupts 'out of band', but that is messy too if done from userspace. For example, what if the timer interrupt is already pending in local apic, when qemu decides to single step?

Also with gdbstub the user doesn't have to stop all vcpus (there is a non-stop mode), in which only some vcpus are stopped which is actually a very cool feature, and of course running vcpus can raise events.

Also interrupts can indeed come from things like vhost.

Best regards,
Maxim Levitsky

Quoted 2 lines

Jan

maintainer
#20 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

[only saw this now, or delivery to me was delayed - anyway]

Quoted 44 lines · Maxim Levitsky
Quoted 38 lines · Jan Kiszka
Quoted 24 lines · Sean Christopherson
Quoted 11 lines · Maxim Levitsky
Quoted 8 lines · Jan Kiszka

Back then, when I was hacking on the gdb-stub and KVM support, the monitor trap flag was not yet broadly available, but the idea to once use it was already there. Now it can be considered broadly available, but it would still require some changes to get it in.

Unfortunately, we don't have such thing with SVM, even recent versions, right? So, a proper way of avoiding diverting event injections while we are having the guest in an "incorrect" state should definitely be the goal.

Yes, I am not aware of anything like monitor trap on SVM.

Quoted 7 lines
Quoted 4 lines

Given that KVM knows whether TF originates solely from guest debugging or was (also) injected by the guest, we should be able to identify the cases where your approach is best to apply. And that without any extra control knob that everyone will only forget to set.

Well I think that the downside of this patch is that the user might actually want to single step into an interrupt handler, and this patch makes it a bit more complicated, and changes the default behavior.

Yes. And, as is, this also blocks NMIs and SMIs. I suspect it also doesn't prevent weirdness if the guest is running in L2, since IRQs for L1 will cause exits from L2 during nested_ops->check_events().

Quoted 7 lines
Quoted 2 lines

I have no objections though to use this patch as is, or at least make this the new default with a new flag to override this.

That's less bad, but IMO still violates the principle of least surprise, e.g. someone that is single-stepping a guest and is expecting an IRQ to fire will be all kinds of confused if they see all the proper IRR, ISR, EFLAGS.IF, etc... settings, but no interrupt.

From my practical experience with debugging guests via single step, seeing an interrupt in that case is everything but handy and generally also not expected (though logical, I agree). IOW: When there is a knob for it, it will remain off in 99% of the time.

Quoted 4 lines

But I see the point of having some control, in an ideal world also an indication that there are pending events, permitting the user to decide what to do. But I suspect the gdb frontend and protocol does not easily permit that.

Quoted 6 lines

Qemu gdbstub actually does have control over suppression of the interrupts over a single step and it is even enabled by default:

https://qemu.readthedocs.io/en/latest/system/gdb.html
(advanced debug options)

Ah, cool! Absolutely in line with what we need.

Quoted 2 lines

However it is currently only implemented in TCG (software emulator) mode and not in KVM mode (I can argue that this is a qemu bug).

Maybe the behavior of old KVM was not exposing the issue, thus no one cared. As I wrote in the other mail today, even some recent kernel do not seem to break single-stepping, for yet unknown reasons.

Quoted 11 lines

So my plan was to add a new kvm guest debug flag KVM_GUESTDBG_BLOCKEVENTS, and let qemu enable it when its 'NOIRQ' mode is enabled (it is by default).

However due to the discussion in this thread about the leakage of the RFLAGS.TF, I wonder if kvm should by default suppress events and have something like KVM_GUESTDBG_SSTEP_ALLOW_EVENTS to override this and wire
that to qemu's NOIRQ=false case.

This will allow older qemu to work correctly and new qemu will be able to choose the old less ideal behavior.

Sounds very reasonable to me.

Quoted 10 lines
Quoted 10 lines
Quoted 9 lines
Quoted 3 lines

Sean Christopherson, what do you think?

Rather than block all events in KVM, what about having QEMU "pause" the timer? E.g. save MSR_TSC_DEADLINE and APIC_TMICT (or inspect the guest to find out which flavor it's using), clear them to zero, then restore both when single-stepping is disabled. I think that will work?

Quoted 5 lines
Quoted 3 lines

No one can stop the clock, and timers are only one source of interrupts. Plus they do not all come from QEMU, some also from KVM or in-kernel sources directly. Would quickly become a mess.

This, plus as we see, even changing with RFLAGS.TF leaks it.

As I wrote: When we take events, the leakage must be stopped for that case. But that might be a bit more tricky because we need to stop on the first instruction in the interrupt handler, thus need some TF, but we must also remove it again from the flags saved for the interrupt context on the guest's interrupt/exception handler stack.

Quoted 7 lines

Changing things like MSR_TSC_DEADLINE will also make it visible to the guest, sooner or later and is a mess that I rather not get into.

It is _possible_ to disable timer interrupts 'out of band', but that is messy too if done from userspace. For example, what if the timer interrupt is already pending in local apic, when qemu decides to single step?

Quoted 6 lines

Also with gdbstub the user doesn't have to stop all vcpus (there is a non-stop mode), in which only some vcpus are stopped which is actually a very cool feature, and of course running vcpus can raise events.

Also interrupts can indeed come from things like vhost.

Exactly.

Jan

Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux
maintainer
#21 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

In reply to Maxim Levitsky
Quoted 130 lines · Maxim Levitsky
Quoted 96 lines · Jan Kiszka
Quoted 74 lines · Maxim Levitsky
Quoted 73 lines · Jan Kiszka
Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 10 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

Does this happen to address this suspicious workaround in the kernel?

        /*
         * The kernel doesn't use TF single-step outside of:
         *
         *  - Kprobes, consumed through kprobe_debug_handler()
         *  - KGDB, consumed through notify_debug()
         *
         * So if we get here with DR_STEP set, something is wonky.
         *
         * A known way to trigger this is through QEMU's GDB stub,
         * which leaks #DB into the guest and causes IST recursion.
         */
        if (WARN_ON_ONCE(dr6 & DR_STEP))
                regs->flags &= ~X86_EFLAGS_TF;

(arch/x86/kernel/traps.c, exc_debug_kernel)

Quoted 8 lines
Quoted 3 lines

I wonder why this got merged while no one fixed QEMU/KVM, for years? Oh, yeah, question to myself as well, dancing around broken guest debugging for a long time while trying to fix other issues...

To be honest I didn't see that warning even once, but I can imagine KVM leaking #DB due to bugs in that code. That area historically didn't receive much attention since it can only be triggered by KVM_GET/SET_GUEST_DEBUG which isn't used in production.

I've triggered it recently while debugging a guest, that's why I got aware of the code path. Long ago, all this used to work (soft BPs, single-stepping etc.)

Quoted 8 lines

The only issue that I on the other hand did
see which is mostly gdb fault is that it fails to remove a software breakpoint when resuming over it, if that breakpoint's python handler messes up with gdb's symbols, which is what lx-symbols does.

And that despite the fact that lx-symbol doesn't mess with the object (that is the kernel) where the breakpoint is defined.

Quoted 8 lines
Quoted 5 lines

Just adding/removing one symbol file is enough to trigger this issue.

Since lx-symbols already works this around when it reloads all symbols, I extended that workaround to happen also when loading/unloading only a single symbol file.

You have no issue with interactive debugging when NOT using gdb scripts / lx-symbol?

To be honest I don't use guest debugging that much, so I probably missed some issues.

Now that I fixed lx-symbols though I'll probably use guest debugging much more.
I will keep an eye on any issues that I find.

The main push to fix lx-symbols actually came from me wanting to understand if there is something broken with KVM's guest debugging knowing that lx-symbols crashes the guest when module is loaded after lx-symbols was executed.

That lx-symbols related guest crash I traced to issue with gdb as I explained, and the lack of blocking of the interrupts on single step is not a bug but more a missing feature that should be implemented to make single step easier to use.

Another issue which isn't a bug is that you can't place a software breakpoint if kernel is not loaded (since there is no code in memory) or if the kernel haven't done basic paging initialization (since there is no paging yet to know where to place the breakpoint). Hardware breakpoints work for this fine though.

Quoted 6 lines

So in summary I haven't found any major issues with KVM's guest debug yet.

If I do notice issues with guest debug, I will try to isolate and debug them.
For the issue that you mentioned, do you have a way to reproduce it?

To pick this up again, I did some experiments and was easily able to reproduce the main problem:

- checkout linux master (1df27313f50 yesterday)
- build a fitting host kernel with KVM
- build a target kernel with defconfig + CONFIG_KVM_GUEST + CONFIG_DEBUG_INFO, no gdb scripts for now
- boot the guest

qemu-system-x86_64 linux.img -enable-kvm -smp 4 -s
-kernel bzImage -append "console=ttyS0 root=... nokaslr"

- gdb vmlinux
- tar rem :1234
- b __x64_sys_execve
- continue whenever you hit the breakpoint, and the guest will eventually hang
- or stepi, and you may end up in the interrupt handler
- specifically doing that on the serial console or setting the bp in early boot exposes the issue

I've also seen

WARNING: CPU: 3 PID: 751 at ../arch/x86/kernel/traps.c:915
exc_debug+0x16b/0x1c0

from time to time.

When I apply this patch to the host, the problems are gone.

Interestingly, my OpenSUSE 5.3.18-lp152.66-default does not show this behavior and /seems/ stable from quick testing. Not sure if there were changes in upstream between that baseline and head or if Suse is carrying a local fix.

In any case, I think we need the following:

- default disabling of event injections when guest debugging injected TF
- possibly some control interface to allow events BUT then avoids any TF injection to prevent guest state corruptions
- exposure of that interface to the gdb frontend, maybe by making it available via the QEMU monitor (monitor ...)
- a for kvm-unit-tests to trigger the above corner cases

Jan

Signature · 2 lines
Siemens AG, T RDA IOT
Corporate Competence Center Embedded Linux
#22 lore raw

Re: [PATCH 2/3] KVM: x86: guest debug: don't inject interrupts while single stepping

In reply to Sean Christopherson
Quoted 43 lines · Sean Christopherson
Quoted 42 lines · Maxim Levitsky

This change greatly helps with two issues:

* Resuming from a breakpoint is much more reliable.

When resuming execution from a breakpoint, with interrupts enabled, more often than not, KVM would inject an interrupt and make the CPU jump immediately to the interrupt handler and eventually return to the breakpoint, to trigger it again.

From the user point of view it looks like the CPU never executed a single instruction and in some cases that can even prevent forward progress, for example, when the breakpoint is placed by an automated script (e.g lx-symbols), which does something in response to the breakpoint and then continues the guest automatically.
If the script execution takes enough time for another interrupt to arrive, the guest will be stuck on the same breakpoint RIP forever.

* Normal single stepping is much more predictable, since it won't land the debugger into an interrupt handler, so it is much more usable.

(If entry to an interrupt handler is desired, the user can still place a breakpoint at it and resume the guest, which won't activate this workaround and let the gdb still stop at the interrupt handler)

Since this change is only active when guest is debugged, it won't affect KVM running normal 'production' VMs.

---
 arch/x86/kvm/x86.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/x86.c +0 -0
index a9d95f90a0487..b75d990fcf12b 100644
@@ -8458,6 +8458,12 @@ static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit
 		can_inject = false;
 	}
 
Quoted 9 lines

arch/x86/kvm/x86.c static void inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit · line 8465

Quoted 5 lines
arch/x86/kvm/x86.c +5 -0
+	/*
+	 * Don't inject interrupts while single stepping to make guest debug easier
+	 */
+	if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
+		return;

Is this something userspace can deal with? E.g. disable IRQs and/or set NMI blocking at the start of single-stepping, unwind at the end? Deviating this far from architectural behavior will end in tears at some point.

I don't worry about NMI, but for IRQs, userspace can clear EFLAGS.IF, but that can be messy to unwind, if an instruction that clears the interrupt flag was single stepped over.

There is also notion of interrupt shadow but it also is reserved for things like delaying interrupts for one cycle after sti, and such.

IMHO KVM_GUESTDBG_SINGLESTEP is already non architectural feature (userspace basically tell the KVM to single step the guest but it doesn't set TF flag or something like that), so changing its definition shouldn't be a problem.

If you worry about some automated script breaking due to the change, (I expect that KVM_GUESTDBG_SINGLESTEP is mostly used manually, especially since single stepping is never 100% reliable due to various issues like that), I can add another flag to it which will block all the interrupts. (like say KVM_GUESTDBG_BLOCKEVENTS).

In fact qemu already has single step flags, enabled over special qemu gdb extension 'maintenance packet qqemu.sstepbits'

Those single step flags allow to disable interrupts and qemu timers during the single stepping, (and both modes are enabled by default)
However kvm code in qemu ignores these bits.

What do you think?

Best regards,
Maxim Levitsky

Quoted 8 lines
Quoted 8 lines
arch/x86/kvm/x86.c +1 -0
+
 	/*
 	 * Finally, inject interrupt events.  If an event cannot be injected
 	 * due to architectural conditions (e.g. IF=0) a window-open exit
Signature · 1 line
2.26.2
#23 lore raw

PATCH 3/3 KVM: SVM: allow to intercept all exceptions for debug

In reply to Maxim Levitsky

Add a new debug module param 'debug_intercept_exceptions' which will allow the KVM to intercept any guest exception, and forward it to the guest.

This can be very useful for guest debugging and/or KVM debugging with kvm trace. This is not intended to be used on production systems.

This is based on an idea first shown here:
https://patchwork.kernel.org/project/kvm/patch/[email protected]/

---
 arch/x86/include/asm/kvm_host.h |  2 +
 arch/x86/kvm/svm/svm.c          | 77 ++++++++++++++++++++++++++++++++-
 arch/x86/kvm/svm/svm.h          |  5 ++-
 arch/x86/kvm/x86.c              |  5 ++-
 4 files changed, 85 insertions(+), 4 deletions(-)
arch/x86/include/asm/kvm_host.h +2 -0
index a52f973bdff6d..c8f44a88b3153 100644
@@ -1564,6 +1564,8 @@ int kvm_emulate_rdpmc(struct kvm_vcpu *vcpu);
 void kvm_queue_exception(struct kvm_vcpu *vcpu, unsigned nr);
 void kvm_queue_exception_e(struct kvm_vcpu *vcpu, unsigned nr, u32 error_code);
 void kvm_queue_exception_p(struct kvm_vcpu *vcpu, unsigned nr, unsigned long payload);
+void kvm_queue_exception_e_p(struct kvm_vcpu *vcpu, unsigned nr,
+			     u32 error_code, unsigned long payload);
 void kvm_requeue_exception(struct kvm_vcpu *vcpu, unsigned nr);
 void kvm_requeue_exception_e(struct kvm_vcpu *vcpu, unsigned nr, u32 error_code);
 void kvm_inject_page_fault(struct kvm_vcpu *vcpu, struct x86_exception *fault);
arch/x86/kvm/svm/svm.c +76 -1

113 lines not on this page · raw message

arch/x86/kvm/svm/svm.h +4 -1
index 8e276c4fb33df..e0ff9ca996df8 100644
@@ -32,6 +32,7 @@ static const u32 host_save_user_msrs[] = {
 #define MSRPM_OFFSETS	16
 extern u32 msrpm_offsets[MSRPM_OFFSETS] __read_mostly;
 extern bool npt_enabled;
+extern uint debug_intercept_exceptions;
 
 enum {
 	VMCB_INTERCEPTS, /* Intercept vectors, TSC offset,
@@ -333,7 +334,9 @@ static inline void clr_exception_intercept(struct vcpu_svm *svm, u32 bit)
 	struct vmcb *vmcb = svm->vmcb01.ptr;
 
 	WARN_ON_ONCE(bit >= 32);
-	vmcb_clr_intercept(&vmcb->control, INTERCEPT_EXCEPTION_OFFSET + bit);
+
+	if (!((1 << bit) & debug_intercept_exceptions))
+		vmcb_clr_intercept(&vmcb->control, INTERCEPT_EXCEPTION_OFFSET + bit);
 
 	recalc_intercepts(svm);
 }
arch/x86/kvm/x86.c +3 -2
index b75d990fcf12b..be509944622bd 100644
@@ -627,12 +627,13 @@ void kvm_queue_exception_p(struct kvm_vcpu *vcpu, unsigned nr,
 }
 EXPORT_SYMBOL_GPL(kvm_queue_exception_p);
 
-static void kvm_queue_exception_e_p(struct kvm_vcpu *vcpu, unsigned nr,
-				    u32 error_code, unsigned long payload)
+void kvm_queue_exception_e_p(struct kvm_vcpu *vcpu, unsigned nr,
+			     u32 error_code, unsigned long payload)
 {
 	kvm_multiple_exception(vcpu, nr, true, error_code,
 			       true, payload, false);
 }
+EXPORT_SYMBOL_GPL(kvm_queue_exception_e_p);
 
 int kvm_complete_insn_gp(struct kvm_vcpu *vcpu, int err)
 {
Signature · 1 line
2.26.2
#24 lore raw

Re: [PATCH 3/3] KVM: SVM: allow to intercept all exceptions for debug

Hi Maxim,

arch/x86/kvm/svm/svm.c static int invpcid_interception(struct kvm_vcpu *vcpu) · line 3046

Quoted 2 lines · Maxim Levitsky
arch/x86/kvm/svm/svm.c +1 -1
-static int (*const svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {
+static int (*svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {

Can you keep this const and always set the necessary handlers? If exceptions are not intercepted they will not be used.

arch/x86/kvm/svm/svm.h static inline void clr_exception_intercept(struct vcpu_svm *svm, u32 bit) · line 339

Quoted 8 lines
arch/x86/kvm/svm/svm.h +3 -1
@@ -333,7 +334,9 @@ static inline void clr_exception_intercept(struct vcpu_svm *svm, u32 bit)
 	struct vmcb *vmcb = svm->vmcb01.ptr;
 
 	WARN_ON_ONCE(bit >= 32);
-	vmcb_clr_intercept(&vmcb->control, INTERCEPT_EXCEPTION_OFFSET + bit);
+
+	if (!((1 << bit) & debug_intercept_exceptions))
+		vmcb_clr_intercept(&vmcb->control, INTERCEPT_EXCEPTION_OFFSET + bit);

This will break SEV-ES guests, as those will not cause an intercept but now start to get #VC exceptions on every other exception that is raised. SEV-ES guests are not prepared for that and will not even boot, so please don't enable this feature for them.

#25 lore raw

Re: [PATCH 3/3] KVM: SVM: allow to intercept all exceptions for debug

Quoted 12 lines · Joerg Roedel

Hi Maxim,

arch/x86/kvm/svm/svm.c static int invpcid_interception(struct kvm_vcpu *vcpu) · line 3046

Quoted 2 lines · Maxim Levitsky
arch/x86/kvm/svm/svm.c +1 -1
-static int (*const svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {
+static int (*svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {

Can you keep this const and always set the necessary handlers? If exceptions are not intercepted they will not be used.

Quoted 3 lines
arch/x86/kvm/svm/svm.h +0 -0
@@ -333,7 +334,9 @@ static inline void clr_exception_intercept(struct vcpu_svm *svm, u32 bit)
 	struct vmcb *vmcb = svm->vmcb01.ptr;
 
Quoted 10 lines

arch/x86/kvm/svm/svm.h static inline void clr_exception_intercept(struct vcpu_svm *svm, u32 bit) · line 339

Quoted 5 lines
arch/x86/kvm/svm/svm.h +3 -1
 	WARN_ON_ONCE(bit >= 32);
-	vmcb_clr_intercept(&vmcb->control, INTERCEPT_EXCEPTION_OFFSET + bit);
+
+	if (!((1 << bit) & debug_intercept_exceptions))
+		vmcb_clr_intercept(&vmcb->control, INTERCEPT_EXCEPTION_OFFSET + bit);

This will break SEV-ES guests, as those will not cause an intercept but now start to get #VC exceptions on every other exception that is raised. SEV-ES guests are not prepared for that and will not even boot, so please don't enable this feature for them.

I agree but what is wrong with that?
This is a debug feature, and it only can be enabled by the root, and so someone might actually want this case to happen (e.g to see if a SEV guest can cope with extra #VC exceptions).

I have nothing against not allowing this for SEV-ES guests though. What do you think?

Best regards,
Maxim Levitsky

#26 lore raw

Re: [PATCH 3/3] KVM: SVM: allow to intercept all exceptions for debug

Quoted 4 lines · Maxim Levitsky

I agree but what is wrong with that?
This is a debug feature, and it only can be enabled by the root, and so someone might actually want this case to happen (e.g to see if a SEV guest can cope with extra #VC exceptions).

That doesn't make sense, we know that and SEV-ES guest can't cope with extra #VC exceptions, so there is no point in testing this. It is more a way to shot oneself into the foot for the user and a potential source of bug reports for SEV-ES guests.

Quoted 2 lines

I have nothing against not allowing this for SEV-ES guests though. What do you think?

I think SEV-ES guests should only have the intercept bits set which guests acutally support.

Regards,

Joerg

#27 lore raw

Re: [PATCH 3/3] KVM: SVM: allow to intercept all exceptions for debug

Quoted 10 lines · Joerg Roedel
Quoted 4 lines · Maxim Levitsky

I agree but what is wrong with that?
This is a debug feature, and it only can be enabled by the root, and so someone might actually want this case to happen (e.g to see if a SEV guest can cope with extra #VC exceptions).

That doesn't make sense, we know that and SEV-ES guest can't cope with extra #VC exceptions, so there is no point in testing this. It is more a way to shot oneself into the foot for the user and a potential source of bug reports for SEV-ES guests.

But again this is a debug feature, and it is intended to allow the user to shoot himself in the foot. Bug reports for a debug feature are autoclosed. It is no different from say poking kernel memory with its built-in gdbstub, for example.

Best regards,
Maxim Levitsky

Quoted 13 lines
Quoted 4 lines

I have nothing against not allowing this for SEV-ES guests though. What do you think?

I think SEV-ES guests should only have the intercept bits set which guests acutally support

Regards,

Joerg

#28 lore raw

Re: [PATCH 3/3] KVM: SVM: allow to intercept all exceptions for debug

Quoted 2 lines · Maxim Levitsky

But again this is a debug feature, and it is intended to allow the user to shoot himself in the foot.

And one can't debug SEV-ES guests with it, so what is the point of enabling it for them too?

Regards,

Joerg

#29 lore raw

Re: [PATCH 3/3] KVM: SVM: allow to intercept all exceptions for debug

Quoted 6 lines · Joerg Roedel
Quoted 2 lines · Maxim Levitsky

But again this is a debug feature, and it is intended to allow the user to shoot himself in the foot.

And one can't debug SEV-ES guests with it, so what is the point of enabling it for them too?

Agreed. I can see myself enabling debug features by default, it would be nice to not having to go out of my way to disable them for SEV-ES/SNP guests.

Skipping SEV-ES guests should not be difficult; KVM could probably even print a message stating that the debug hook is being ignored. One thought would be to snapshot debug_intercept_exceptions at VM creation, and simply zero it out for incompatible guests. That would also allow changing debug_intercept_exceptions without reloading KVM, which IMO would be very convenient.

#30 lore raw

Re: [PATCH 3/3] KVM: SVM: allow to intercept all exceptions for debug

Quoted 7 lines · Sean Christopherson
Quoted 6 lines · Joerg Roedel
Quoted 2 lines · Maxim Levitsky

But again this is a debug feature, and it is intended to allow the user to shoot himself in the foot.

And one can't debug SEV-ES guests with it, so what is the point of enabling it for them too?

You can create a special SEV-ES guest which does handle all exceptions via #VC, or just observe it fail which can be useful for some whatever reason.

Quoted 3 lines

Agreed. I can see myself enabling debug features by default, it would be nice to not having to go out of my way to disable them for SEV-ES/SNP guests.

This does sound like a valid reason to disable this for SEV-ES.

Quoted 7 lines

Skipping SEV-ES guests should not be difficult; KVM could probably even print a message stating that the debug hook is being ignored. One thought would be to snapshot debug_intercept_exceptions at VM creation, and simply zero it out for incompatible guests. That would also allow changing debug_intercept_exceptions without reloading KVM, which IMO would be very convenient.

So all right I'll disable this for SEV-ES.
The idea to change the debug_intercept_exceptions on the fly is also a good idea, I will implement it in next version of the patches.

Thanks for the review,
Best regards,
Maxim Levitsky

#31 lore raw

Re: [PATCH 3/3] KVM: SVM: allow to intercept all exceptions for debug

Quoted 8 lines · Maxim Levitsky
Quoted 6 lines · Sean Christopherson

Skipping SEV-ES guests should not be difficult; KVM could probably even print a message stating that the debug hook is being ignored. One thought would be to snapshot debug_intercept_exceptions at VM creation, and simply zero it out for incompatible guests. That would also allow changing debug_intercept_exceptions without reloading KVM, which IMO would be very convenient.

So all right I'll disable this for SEV-ES.

Belated thought. KVM doesn't know a guest will be an SEV-ES guest until sev_es_guest_init(), and KVM currently doesn't prevent creating vCPUs before KVM_SEV_ES_INIT. But, I'm 99% confident that's a KVM bug. For your purposes, I think you can assume kvm->arch.debug_intercept_exceptions will _not_ change after vCPU creation.

Quoted 2 lines

The idea to change the debug_intercept_exceptions on the fly is also a good idea, I will implement it in next version of the patches.

Can you also move the module param to x86? It doesn't need to be wired up for VMX right away, but it makes sense to do it at some point, and ideally folks won't have to update their scripts when that happens.

#32 lore raw

Re: [PATCH 3/3] KVM: SVM: allow to intercept all exceptions for debug

In reply to Maxim Levitsky
Quoted 12 lines · Maxim Levitsky

Add a new debug module param 'debug_intercept_exceptions' which will allow the KVM to intercept any guest exception, and forward it to the guest.

This can be very useful for guest debugging and/or KVM debugging with kvm trace. This is not intended to be used on production systems.

This is based on an idea first shown here:
https://patchwork.kernel.org/project/kvm/patch/[email protected]/

Quoted 6 lines
---
 arch/x86/include/asm/kvm_host.h |  2 +
 arch/x86/kvm/svm/svm.c          | 77 ++++++++++++++++++++++++++++++++-
 arch/x86/kvm/svm/svm.h          |  5 ++-
 arch/x86/kvm/x86.c              |  5 ++-
 4 files changed, 85 insertions(+), 4 deletions(-)

Looks interesting, I'll give it a try when I get a chance in the coming days.

Thx!

Signature · 4 lines
Regards/Gruss,
    Boris.

SUSE Software Solutions Germany GmbH, GF: Felix Imendörffer, HRB 36809, AG Nürnberg