There's a default warning message that gets printed, however,
there are various failure conditions:
- a msr read can fail
- a msr write can fail
- a msr has an unexpected value
- all msrs have unexpected values (disable PMU)
Also, commit commit 005bd0077a79 ("perf/x86: Modify error message in
virtualized environment") completely removed printing the msr in
question but these messages could be helpful for debugging vPMUs as
well. Add them back and change them to pr_debugs, this keeps the
behavior the same for baremetal.
Lastly, use %llx to silence checkpatch
Signed-off-by: Bandan Das <[email protected]>
---
arch/x86/events/core.c | 66 ++++++++++++++++++++++++++++++++----------
1 file changed, 50 insertions(+), 16 deletions(-)
diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index e2b1447192a8..786e03893a0c 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -192,9 +192,16 @@ static void release_pmc_hardware(void) {}
static bool check_hw_exists(void)
{
- u64 val, val_fail = -1, val_new= ~0;
- int i, reg, reg_fail = -1, ret = 0;
- int bios_fail = 0;
+ u64 val = -1, val_fail = -1, val_new = ~0;
+ int i, reg = -1, reg_fail = -1, ret = 0;
+ bool virt = boot_cpu_has(X86_FEATURE_HYPERVISOR) ? true : false;
+ enum {
+ READ_FAIL = 1,
+ WRITE_FAIL = 2,
+ PMU_FAIL = 3,
+ BIOS_FAIL = 4,
+ };
+ int status = 0;
int reg_safe = -1;
/*
@@ -204,10 +211,13 @@ static bool check_hw_exists(void)
for (i = 0; i < x86_pmu.num_counters; i++) {
reg = x86_pmu_config_addr(i);
ret = rdmsrl_safe(reg, &val);
- if (ret)
+ if (ret) {
+ status = READ_FAIL;
goto msr_fail;
+ }
+
if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
- bios_fail = 1;
+ status = BIOS_FAIL;
val_fail = val;
reg_fail = reg;
} else {
@@ -218,11 +228,13 @@ static bool check_hw_exists(void)
if (x86_pmu.num_counters_fixed) {
reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
ret = rdmsrl_safe(reg, &val);
- if (ret)
+ if (ret) {
+ status = READ_FAIL;
goto msr_fail;
+ }
for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
if (val & (0x03 << i*4)) {
- bios_fail = 1;
+ status = BIOS_FAIL;
val_fail = val;
reg_fail = reg;
}
@@ -236,7 +248,7 @@ static bool check_hw_exists(void)
*/
if (reg_safe == -1) {
- reg = reg_safe;
+ status = PMU_FAIL;
goto msr_fail;
}
@@ -246,18 +258,22 @@ static bool check_hw_exists(void)
* (qemu/kvm) that don't trap on the MSR access and always return 0s.
*/
reg = x86_pmu_event_addr(reg_safe);
- if (rdmsrl_safe(reg, &val))
+ if (rdmsrl_safe(reg, &val)) {
+ status = READ_FAIL;
goto msr_fail;
+ }
val ^= 0xffffUL;
ret = wrmsrl_safe(reg, val);
ret |= rdmsrl_safe(reg, &val_new);
- if (ret || val != val_new)
+ if (ret || val != val_new) {
+ status = WRITE_FAIL;
goto msr_fail;
+ }
/*
* We still allow the PMU driver to operate:
*/
- if (bios_fail) {
+ if (status == BIOS_FAIL) {
pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
reg_fail, val_fail);
@@ -266,12 +282,30 @@ static bool check_hw_exists(void)
return true;
msr_fail:
- if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
+ if (virt)
pr_cont("PMU not available due to virtualization, using software events only.\n");
- } else {
- pr_cont("Broken PMU hardware detected, using software events only.\n");
- pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
- reg, val_new);
+ switch (status) {
+ case READ_FAIL:
+ if (virt)
+ pr_debug("Failed to read perfctr msr (MSR %x)\n", reg);
+ else
+ pr_err("Failed to read perfctr msr (MSR %x)\n", reg);
+ break;
+ case WRITE_FAIL:
+ if (virt)
+ pr_debug("Failed to write perfctr msr (MSR %x, wrote: %llx, read: %llx)\n",
+ reg, val, val_new);
+ else
+ pr_err("Failed to write perfctr msr (MSR %x, wrote: %llx, read: %llx)\n",
+ reg, val, val_new);
+ break;
+ case PMU_FAIL:
+ /* fall through for default message */
+ default:
+ if (virt)
+ pr_debug("Broken PMU hardware detected, using software events only.\n");
+ else
+ pr_cont("Broken PMU hardware detected, using software events only.\n");
}
return false;
--
2.19.2
On Fri, Apr 12, 2019 at 03:09:17PM -0400, Bandan Das wrote:
SNIP
> msr_fail:
> - if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
> + if (virt)
> pr_cont("PMU not available due to virtualization, using software events only.\n");
> - } else {
> - pr_cont("Broken PMU hardware detected, using software events only.\n");
> - pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
> - reg, val_new);
> + switch (status) {
> + case READ_FAIL:
> + if (virt)
> + pr_debug("Failed to read perfctr msr (MSR %x)\n", reg);
> + else
> + pr_err("Failed to read perfctr msr (MSR %x)\n", reg);
> + break;
> + case WRITE_FAIL:
> + if (virt)
> + pr_debug("Failed to write perfctr msr (MSR %x, wrote: %llx, read: %llx)\n",
> + reg, val, val_new);
> + else
> + pr_err("Failed to write perfctr msr (MSR %x, wrote: %llx, read: %llx)\n",
> + reg, val, val_new);
> + break;
> + case PMU_FAIL:
> + /* fall through for default message */
> + default:
> + if (virt)
> + pr_debug("Broken PMU hardware detected, using software events only.\n");
> + else
> + pr_cont("Broken PMU hardware detected, using software events only.\n");
hum, why can't we have just one pr_err for both virt and bare metal?
jirka
> }
>
> return false;
> --
> 2.19.2
>
On Fri, Apr 12, 2019 at 03:09:17PM -0400, Bandan Das wrote:
>
> There's a default warning message that gets printed, however,
> there are various failure conditions:
> - a msr read can fail
> - a msr write can fail
> - a msr has an unexpected value
> - all msrs have unexpected values (disable PMU)
>
> Also, commit commit 005bd0077a79 ("perf/x86: Modify error message in
> virtualized environment") completely removed printing the msr in
> question but these messages could be helpful for debugging vPMUs as
> well. Add them back and change them to pr_debugs, this keeps the
> behavior the same for baremetal.
>
> Lastly, use %llx to silence checkpatch
Yuck... if you're debugging a hypervisor, you can bloody well run your
own kernel with additional print slattered around.
The whole make an exception for virt bullshit was already pushing it,
this is just insane.
> @@ -266,12 +282,30 @@ static bool check_hw_exists(void)
> return true;
>
> msr_fail:
> - if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
> + if (virt)
> pr_cont("PMU not available due to virtualization, using software events only.\n");
> - } else {
> - pr_cont("Broken PMU hardware detected, using software events only.\n");
> - pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
> - reg, val_new);
> + switch (status) {
> + case READ_FAIL:
> + if (virt)
> + pr_debug("Failed to read perfctr msr (MSR %x)\n", reg);
> + else
> + pr_err("Failed to read perfctr msr (MSR %x)\n", reg);
> + break;
> + case WRITE_FAIL:
> + if (virt)
> + pr_debug("Failed to write perfctr msr (MSR %x, wrote: %llx, read: %llx)\n",
> + reg, val, val_new);
> + else
> + pr_err("Failed to write perfctr msr (MSR %x, wrote: %llx, read: %llx)\n",
> + reg, val, val_new);
> + break;
> + case PMU_FAIL:
> + /* fall through for default message */
> + default:
> + if (virt)
> + pr_debug("Broken PMU hardware detected, using software events only.\n");
> + else
> + pr_cont("Broken PMU hardware detected, using software events only.\n");
> }
Hi Peter,
Peter Zijlstra <[email protected]> writes:
> On Fri, Apr 12, 2019 at 03:09:17PM -0400, Bandan Das wrote:
>>
>> There's a default warning message that gets printed, however,
>> there are various failure conditions:
>> - a msr read can fail
>> - a msr write can fail
>> - a msr has an unexpected value
>> - all msrs have unexpected values (disable PMU)
>>
>> Also, commit commit 005bd0077a79 ("perf/x86: Modify error message in
>> virtualized environment") completely removed printing the msr in
>> question but these messages could be helpful for debugging vPMUs as
>> well. Add them back and change them to pr_debugs, this keeps the
>> behavior the same for baremetal.
>>
>> Lastly, use %llx to silence checkpatch
>
> Yuck... if you're debugging a hypervisor, you can bloody well run your
> own kernel with additional print slattered around.
>
> The whole make an exception for virt bullshit was already pushing it,
> this is just insane.
>
The only virt specific parts are the pr_debugs which I can remove and
replace with unconditional pr_err()s as suggested by Jiri. Is that ok ?
Bandan
>> @@ -266,12 +282,30 @@ static bool check_hw_exists(void)
>> return true;
>>
>> msr_fail:
>> - if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
>> + if (virt)
>> pr_cont("PMU not available due to virtualization, using software events only.\n");
>> - } else {
>> - pr_cont("Broken PMU hardware detected, using software events only.\n");
>> - pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
>> - reg, val_new);
>> + switch (status) {
>> + case READ_FAIL:
>> + if (virt)
>> + pr_debug("Failed to read perfctr msr (MSR %x)\n", reg);
>> + else
>> + pr_err("Failed to read perfctr msr (MSR %x)\n", reg);
>> + break;
>> + case WRITE_FAIL:
>> + if (virt)
>> + pr_debug("Failed to write perfctr msr (MSR %x, wrote: %llx, read: %llx)\n",
>> + reg, val, val_new);
>> + else
>> + pr_err("Failed to write perfctr msr (MSR %x, wrote: %llx, read: %llx)\n",
>> + reg, val, val_new);
>> + break;
>> + case PMU_FAIL:
>> + /* fall through for default message */
>> + default:
>> + if (virt)
>> + pr_debug("Broken PMU hardware detected, using software events only.\n");
>> + else
>> + pr_cont("Broken PMU hardware detected, using software events only.\n");
>> }