Re: [PATCH v5 1/9] x86/split_lock: Rework the initialization flow of split lock detection

From: Xiaoyao Li
Date: Tue Mar 24 2020 - 20:18:13 EST


On 3/24/2020 6:29 PM, Thomas Gleixner wrote:
Xiaoyao Li <xiaoyao.li@xxxxxxxxx> writes:
On 3/24/2020 4:24 AM, Thomas Gleixner wrote:
--- a/arch/x86/kernel/cpu/intel.c
+++ b/arch/x86/kernel/cpu/intel.c
@@ -45,6 +45,7 @@ enum split_lock_detect_state {
* split lock detect, unless there is a command line override.
*/
static enum split_lock_detect_state sld_state = sld_off;
+static DEFINE_PER_CPU(u64, msr_test_ctrl_cache);

I used percpu cache in v3, but people prefer Tony's cache for reserved
bits[1].

If you prefer percpu cache, I'll use it in next version.

I'm fine with the single variable.

static void __init split_lock_setup(void)
{
char arg[20];
int i, ret;
+ if (!split_lock_verify_msr(true) || !split_lock_verify_msr(false)) {
+ pr_info("MSR access failed: Disabled\n");
+ return;
+ }
+

I did similar thing like this in my v3, however Sean raised concern that
toggling MSR bit before parsing kernel param is bad behavior. [2]

That's trivial enough to fix.

Thanks,

tglx

8<---------------
--- a/arch/x86/kernel/cpu/intel.c
+++ b/arch/x86/kernel/cpu/intel.c
@@ -44,7 +44,8 @@ enum split_lock_detect_state {
* split_lock_setup() will switch this to sld_warn on systems that support
* split lock detect, unless there is a command line override.
*/
-static enum split_lock_detect_state sld_state = sld_off;
+static enum split_lock_detect_state sld_state __ro_after_init = sld_off;
+static u64 msr_test_ctrl_cache __ro_after_init;
/*
* Processors which have self-snooping capability can handle conflicting
@@ -984,78 +985,85 @@ static inline bool match_option(const ch
return len == arglen && !strncmp(arg, opt, len);
}
+static bool __init split_lock_verify_msr(bool on)
+{
+ u64 ctrl, tmp;
+
+ if (rdmsrl_safe(MSR_TEST_CTRL, &ctrl))
+ return false;
+ if (on)
+ ctrl |= MSR_TEST_CTRL_SPLIT_LOCK_DETECT;
+ else
+ ctrl &= ~MSR_TEST_CTRL_SPLIT_LOCK_DETECT;
+ if (wrmsrl_safe(MSR_TEST_CTRL, ctrl))
+ return false;
+ rdmsrl(MSR_TEST_CTRL, tmp);
+ return ctrl == tmp;
+}
+
static void __init split_lock_setup(void)
{
+ enum split_lock_detect_state state = sld_warn;
char arg[20];
int i, ret;
- setup_force_cpu_cap(X86_FEATURE_SPLIT_LOCK_DETECT);
- sld_state = sld_warn;
+ if (!split_lock_verify_msr(false)) {
+ pr_info("MSR access failed: Disabled\n");
+ return;
+ }
ret = cmdline_find_option(boot_command_line, "split_lock_detect",
arg, sizeof(arg));
if (ret >= 0) {
for (i = 0; i < ARRAY_SIZE(sld_options); i++) {
if (match_option(arg, ret, sld_options[i].option)) {
- sld_state = sld_options[i].state;
+ state = sld_options[i].state;
break;
}
}
}
- switch (sld_state) {
+ switch (state) {
case sld_off:
pr_info("disabled\n");
- break;
-
+ return;
Here, when sld_off, it just returns without setup_force_cpu_cap(X86_FEATURE_SPLIT_LOCK_DETECT).

So for APs, it won't clear SLD bit in split_lock_init().

And I remember why I used sld_not_exist, not use
setup_force_cpu_cap(X86_FEATURE_SPLIT_LOCK_DETECT)

Yes, we can call setup_force_cpu_cap(X86_FEATURE_SPLIT_LOCK_DETECT)
for sld_off case. And in split_lock_init(), explicitly calling sld_update_msr(false) to turn off sld, and calling clear_cpu_cap(c, X86_FEATURE_SPLIT_LOCK_DETECT) to clear the cap. But due to setup_force_cpu_cap(), split_lock_detect will still occurs in /proc/cpuinfo.

case sld_warn:
pr_info("warning about user-space split_locks\n");
break;
-
case sld_fatal:
pr_info("sending SIGBUS on user-space split_locks\n");
break;
}
+
+ rdmsrl(MSR_TEST_CTRL, msr_test_ctrl_cache);
+
+ if (!split_lock_verify_msr(true)) {
+ pr_info("MSR access failed: Disabled\n");
+ return;
+ }
+
+ sld_state = state;
+ setup_force_cpu_cap(X86_FEATURE_SPLIT_LOCK_DETECT);
}
/*
- * Locking is not required at the moment because only bit 29 of this
- * MSR is implemented and locking would not prevent that the operation
- * of one thread is immediately undone by the sibling thread.
- * Use the "safe" versions of rdmsr/wrmsr here because although code
- * checks CPUID and MSR bits to make sure the TEST_CTRL MSR should
- * exist, there may be glitches in virtualization that leave a guest
- * with an incorrect view of real h/w capabilities.
+ * MSR_TEST_CTRL is per core, but we treat it like a per CPU MSR. Locking
+ * is not implemented as one thread could undo the setting of the other
+ * thread immediately after dropping the lock anyway.
*/
-static bool __sld_msr_set(bool on)
+static void sld_update_msr(bool on)
{
- u64 test_ctrl_val;
-
- if (rdmsrl_safe(MSR_TEST_CTRL, &test_ctrl_val))
- return false;
+ u64 ctrl = msr_test_ctrl_cache;
if (on)
- test_ctrl_val |= MSR_TEST_CTRL_SPLIT_LOCK_DETECT;
- else
- test_ctrl_val &= ~MSR_TEST_CTRL_SPLIT_LOCK_DETECT;
-
- return !wrmsrl_safe(MSR_TEST_CTRL, test_ctrl_val);
+ ctrl |= MSR_TEST_CTRL_SPLIT_LOCK_DETECT;
+ wrmsrl(MSR_TEST_CTRL, ctrl);
}
static void split_lock_init(void)
{
- if (sld_state == sld_off)
- return;
-
- if (__sld_msr_set(true))
- return;
-
- /*
- * If this is anything other than the boot-cpu, you've done
- * funny things and you get to keep whatever pieces.
- */
- pr_warn("MSR fail -- disabled\n");
- sld_state = sld_off;
+ if (boot_cpu_has(X86_FEATURE_SPLIT_LOCK_DETECT))
+ sld_update_msr(sld_state != sld_off);
}
bool handle_user_split_lock(struct pt_regs *regs, long error_code)
@@ -1071,7 +1079,7 @@ bool handle_user_split_lock(struct pt_re
* progress and set TIF_SLD so the detection is re-enabled via
* switch_to_sld() when the task is scheduled out.
*/
- __sld_msr_set(false);
+ sld_update_msr(false);
set_tsk_thread_flag(current, TIF_SLD);
return true;
}
@@ -1085,7 +1093,7 @@ bool handle_user_split_lock(struct pt_re
*/
void switch_to_sld(unsigned long tifn)
{
- __sld_msr_set(!(tifn & _TIF_SLD));
+ sld_update_msr(!(tifn & _TIF_SLD));
}
#define SPLIT_LOCK_CPU(model) {X86_VENDOR_INTEL, 6, model, X86_FEATURE_ANY}