Skip to contentSkip to conversation

x86, sched: Bail out of frequency invariance if turbo frequency is unknown

bug report Dan Carpenter 2 participants 5 messages May 5, 2022 – May 13, 2022 Kernel janitors

lore
Share & tools
lore
#0 lore raw

Hello Giovanni Gherdovich,

The patch 51beea8862a3: "x86, sched: Bail out of frequency invariance if turbo frequency is unknown" from May 31, 2020, leads to the following Smatch static checker warning:

	arch/x86/kernel/cpu/aperfmperf.c:274 intel_set_max_freq_ratio()
	error: uninitialized symbol 'turbo_freq'.

arch/x86/kernel/cpu/aperfmperf.c

    242 static bool __init intel_set_max_freq_ratio(void)
    243 {
    244         u64 base_freq, turbo_freq;
    245         u64 turbo_ratio;
    246 
    247         if (slv_set_max_freq_ratio(&base_freq, &turbo_freq))
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Imagine this fails.

    248                 goto out;
    249 
    250         if (x86_match_cpu(has_glm_turbo_ratio_limits) &&
    251             skx_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
    252                 goto out;
    253 
    254         if (x86_match_cpu(has_knl_turbo_ratio_limits) &&
    255             knl_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
    256                 goto out;
    257 
    258         if (x86_match_cpu(has_skx_turbo_ratio_limits) &&
    259             skx_set_max_freq_ratio(&base_freq, &turbo_freq, 4))
    260                 goto out;
    261 
    262         if (core_set_max_freq_ratio(&base_freq, &turbo_freq))
    263                 goto out;
    264 
    265         return false;
    266 
    267 out:
    268         /*
    269          * Some hypervisors advertise X86_FEATURE_APERFMPERF
    270          * but then fill all MSR's with zeroes.
    271          * Some CPUs have turbo boost but don't declare any turbo ratio
    272          * in MSR_TURBO_RATIO_LIMIT.
    273          */
--> 274         if (!base_freq || !turbo_freq) {
                    ^^^^^^^^^^^^^^^^^^^^^^^^^

Uninitialized. Although I notice that base_freq is also unintialized and that predates your patch... So I should probably send this bug report to someone else... Sorry?

    275                 pr_debug("Couldn't determine cpu base or turbo frequency, necessary for scale-invariant accounting.\n");
    276                 return false;
    277         }
    278 
    279         turbo_ratio = div_u64(turbo_freq * SCHED_CAPACITY_SCALE, base_freq);
    280         if (!turbo_ratio) {
    281                 pr_debug("Non-zero turbo and base frequencies led to a 0 ratio.\n");
    282                 return false;
    283         }
    284 
    285         arch_turbo_freq_ratio = turbo_ratio;
    286         arch_set_max_freq_ratio(turbo_disabled());
    287 
    288         return true;
    289 }

regards,
dan carpenter

#1 lore raw
Quoted 48 lines · Dan Carpenter

Hello Giovanni Gherdovich,

The patch 51beea8862a3: "x86, sched: Bail out of frequency invariance if turbo frequency is unknown" from May 31, 2020, leads to the following Smatch static checker warning:

	arch/x86/kernel/cpu/aperfmperf.c:274 intel_set_max_freq_ratio()
	error: uninitialized symbol 'turbo_freq'.

arch/x86/kernel/cpu/aperfmperf.c

    242 static bool __init intel_set_max_freq_ratio(void)
    243 {
    244         u64 base_freq, turbo_freq;
    245         u64 turbo_ratio;
    246 
    247         if (slv_set_max_freq_ratio(&base_freq, &turbo_freq))
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Imagine this fails.

    248                 goto out;
    249 
    250         if (x86_match_cpu(has_glm_turbo_ratio_limits) &&
    251             skx_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
    252                 goto out;
    253 
    254         if (x86_match_cpu(has_knl_turbo_ratio_limits) &&
    255             knl_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
    256                 goto out;
    257 
    258         if (x86_match_cpu(has_skx_turbo_ratio_limits) &&
    259             skx_set_max_freq_ratio(&base_freq, &turbo_freq, 4))
    260                 goto out;
    261 
    262         if (core_set_max_freq_ratio(&base_freq, &turbo_freq))
    263                 goto out;
    264 
    265         return false;
    266 
    267 out:
    268         /*
    269          * Some hypervisors advertise X86_FEATURE_APERFMPERF
    270          * but then fill all MSR's with zeroes.
    271          * Some CPUs have turbo boost but don't declare any turbo ratio
    272          * in MSR_TURBO_RATIO_LIMIT.
    273          */
--> 274         if (!base_freq || !turbo_freq) {
                    ^^^^^^^^^^^^^^^^^^^^^^^^^
Quoted 4 lines

Uninitialized. Although I notice that base_freq is also unintialized and that predates your patch... So I should probably send this bug report to someone else... Sorry?

Hello Dan,

I'm the right person for this report, as I think I wrote both the buggy patch and the buggy code that predates it. I'm taking a moment to check what's going on, as that really look like a dumb mistake.

Thanks for the report, I'll follow up.

Giovanni

#2 lore raw
In reply to Dan Carpenter
Quoted 18 lines · Dan Carpenter

Hello Giovanni Gherdovich,

The patch 51beea8862a3: "x86, sched: Bail out of frequency invariance if turbo frequency is unknown" from May 31, 2020, leads to the following Smatch static checker warning:

	arch/x86/kernel/cpu/aperfmperf.c:274 intel_set_max_freq_ratio()
	error: uninitialized symbol 'turbo_freq'.

arch/x86/kernel/cpu/aperfmperf.c

    242 static bool __init intel_set_max_freq_ratio(void)
    243 {
    244         u64 base_freq, turbo_freq;
    245         u64 turbo_ratio;
    246 
    247         if (slv_set_max_freq_ratio(&base_freq, &turbo_freq))
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Quoted 3 lines

Imagine this fails.

    248                 goto out;

I think this is a false positive of the static checker; I'll send the patch to initialize the variables anyway as it looks better.

When slv_set_max_freq_ratio() fails, it returns false, and we move on to the next "if" statement. "goto out" is executed only if slv_set_max_freq_ratio() succeeds (returns true), and writes data in its input parameters.

I give you that the code looks wrong, because apparently when people (and static analyzers) read that, they think "if (problem) goto error;" but this function is written like "if (it_works) goto the_end;".

When I received your report I thought the code was wrong, too. You already sent me this report two years ago,
https://www.spinics.net/lists/linux-kernel-janitors/msg51372.html
and I didn't reply (I should have).

Quoted 9 lines
    249 
    250         if (x86_match_cpu(has_glm_turbo_ratio_limits) &&
    251             skx_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
    252                 goto out;
    253 
    254         if (x86_match_cpu(has_knl_turbo_ratio_limits) &&
    255             knl_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
    256                 goto out;
    257 
Quoted 8 lines
    258         if (x86_match_cpu(has_skx_turbo_ratio_limits) &&
    259             skx_set_max_freq_ratio(&base_freq, &turbo_freq, 4))
    260                 goto out;
    261 
    262         if (core_set_max_freq_ratio(&base_freq, &turbo_freq))
    263                 goto out;
    264 
    265         return false;

The "return false" above is what happens if none of the "if" matched, and this is when base_freq and turbo_freq are uninitialized.

Quoted 11 lines

266
267 out:
268 /*
269 * Some hypervisors advertise X86_FEATURE_APERFMPERF
270 * but then fill all MSR's with zeroes.
271 * Some CPUs have turbo boost but don't declare any turbo ratio
272 * in MSR_TURBO_RATIO_LIMIT.
273 */

--> 274         if (!base_freq || !turbo_freq) {
                    ^^^^^^^^^^^^^^^^^^^^^^^^^

Uninitialized.

I know it looks sketchy (and makes you think I forgot the initialization), but the condition here is checking for when a *_set_max_freq_ratio() above matched, but received zeroes from reading the corresponding MSR.

Anyways, patch incoming to make this look better, but I don't think it's a "fixes: [...]".

Quoted 22 lines

Although I notice that base_freq is also unintialized and that predates your patch... So I should probably send this bug report to someone else... Sorry?

    275                 pr_debug("Couldn't determine cpu base or turbo frequency, necessary for scale-invariant accounting.\n");
    276                 return false;
    277         }
    278 
    279         turbo_ratio = div_u64(turbo_freq * SCHED_CAPACITY_SCALE, base_freq);
    280         if (!turbo_ratio) {
    281                 pr_debug("Non-zero turbo and base frequencies led to a 0 ratio.\n");
    282                 return false;
    283         }
    284 
    285         arch_turbo_freq_ratio = turbo_ratio;
    286         arch_set_max_freq_ratio(turbo_disabled());
    287 
    288         return true;
    289 }

regards,
dan carpenter

#3 lore raw
Quoted 26 lines · Giovanni Gherdovich
Quoted 18 lines · Dan Carpenter

Hello Giovanni Gherdovich,

The patch 51beea8862a3: "x86, sched: Bail out of frequency invariance if turbo frequency is unknown" from May 31, 2020, leads to the following Smatch static checker warning:

	arch/x86/kernel/cpu/aperfmperf.c:274 intel_set_max_freq_ratio()
	error: uninitialized symbol 'turbo_freq'.

arch/x86/kernel/cpu/aperfmperf.c

    242 static bool __init intel_set_max_freq_ratio(void)
    243 {
    244         u64 base_freq, turbo_freq;
    245         u64 turbo_ratio;
    246 
    247         if (slv_set_max_freq_ratio(&base_freq, &turbo_freq))
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Quoted 3 lines

Imagine this fails.

    248                 goto out;

I think this is a false positive of the static checker; I'll send the patch to initialize the variables anyway as it looks better.

Quoted 7 lines

When slv_set_max_freq_ratio() fails, it returns false, and we move on to the next "if" statement. "goto out" is executed only if slv_set_max_freq_ratio() succeeds (returns true), and writes data in its input parameters.

I give you that the code looks wrong, because apparently when people (and static analyzers) read that, they think "if (problem) goto error;" but this function is written like "if (it_works) goto the_end;".

I read this wrong. But looking at it closer, I think the static analyzer is actually correct (definitely correct-ish for sure). Smatch is complaining about knl_set_max_freq_ratio().

   149  static bool __init knl_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq,
   150                                            int num_delta_fratio)
   151  {
   152          int fratio, delta_fratio, found;
   153          int err, i;
   154          u64 msr;
   155  
   156          err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
   157          if (err)
   158                  return false;
   159  
   160          *base_freq = (*base_freq >> 8) & 0xFF;      /* max P state */
   161  
   162          err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
   163          if (err)
   164                  return false;
   165  
   166          fratio = (msr >> 8) & 0xFF;
   167          i = 16;
   168          found = 0;
   169          do {
   170                  if (found >= num_delta_fratio) {
   171                          *turbo_freq = fratio;
   172                          return true;
   173                  }
   174  
   175                  delta_fratio = (msr >> (i + 5)) & 0x7;
   176  
   177                  if (delta_fratio) {
   178                          found += 1;
   179                          fratio -= delta_fratio;
   180                  }
   181  
   182                  i += 8;
   183          } while (i < 64);
   184  
   185          return true;

If we reach this "return true" then turbo_freq is not set. This may not be reachable in real life. Should it be return false?

186 }

regards,
dan carpenter

#4 lore raw
Quoted 77 lines · Dan Carpenter
Quoted 26 lines · Giovanni Gherdovich
Quoted 18 lines · Dan Carpenter

Hello Giovanni Gherdovich,

The patch 51beea8862a3: "x86, sched: Bail out of frequency invariance if turbo frequency is unknown" from May 31, 2020, leads to the following Smatch static checker warning:

	arch/x86/kernel/cpu/aperfmperf.c:274 intel_set_max_freq_ratio()
	error: uninitialized symbol 'turbo_freq'.

arch/x86/kernel/cpu/aperfmperf.c

    242 static bool __init intel_set_max_freq_ratio(void)
    243 {
    244         u64 base_freq, turbo_freq;
    245         u64 turbo_ratio;
    246 
    247         if (slv_set_max_freq_ratio(&base_freq, &turbo_freq))
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Quoted 3 lines

Imagine this fails.

    248                 goto out;

I think this is a false positive of the static checker; I'll send the patch to initialize the variables anyway as it looks better.

Quoted 7 lines

When slv_set_max_freq_ratio() fails, it returns false, and we move on to the next "if" statement. "goto out" is executed only if slv_set_max_freq_ratio() succeeds (returns true), and writes data in its input parameters.

I give you that the code looks wrong, because apparently when people (and static analyzers) read that, they think "if (problem) goto error;" but this function is written like "if (it_works) goto the_end;".

I read this wrong. But looking at it closer, I think the static analyzer is actually correct (definitely correct-ish for sure). Smatch is complaining about knl_set_max_freq_ratio().

   149  static bool __init knl_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq,
   150                                            int num_delta_fratio)
   151  {
   152          int fratio, delta_fratio, found;
   153          int err, i;
   154          u64 msr;
   155  
   156          err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
   157          if (err)
   158                  return false;
   159  
   160          *base_freq = (*base_freq >> 8) & 0xFF;      /* max P state */
   161  
   162          err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
   163          if (err)
   164                  return false;
   165  
   166          fratio = (msr >> 8) & 0xFF;
   167          i = 16;
   168          found = 0;
   169          do {
   170                  if (found >= num_delta_fratio) {
   171                          *turbo_freq = fratio;
   172                          return true;
   173                  }
   174  
   175                  delta_fratio = (msr >> (i + 5)) & 0x7;
   176  
   177                  if (delta_fratio) {
   178                          found += 1;
   179                          fratio -= delta_fratio;
   180                  }
   181  
   182                  i += 8;
   183          } while (i < 64);
   184  
   185          return true;
Quoted 3 lines

If we reach this "return true" then turbo_freq is not set. This may not be reachable in real life. Should it be return false?

That's right, it's a bug. It should return false at the end. I'll send a fix, thanks.

Giovanni