Merge branch 'dev' of https://github.com/GuillaumeDIDIER/dendrobates-t-azureus

cache_slice/src/arch.rs

@@ -135,6 +135,7 @@ pub struct XeonPerfCounters {
 }
 
 pub struct CorePerfCounters {
+    pub msr_unc_cbo_config: u64,
     pub max_slice: u16,
     pub msr_unc_perf_global_ctr: u64,
     pub val_enable_ctrs: u64,
@@ -221,6 +222,7 @@ const BROADWELL_XEON: XeonPerfCounters = XeonPerfCounters {
 
 // TODO find appropriate values
 const ALDER_LAKE_TO_RAPTOR_LAKE_CORE: CorePerfCounters = CorePerfCounters {
+    msr_unc_cbo_config: 0x396,
     max_slice: 10,
     msr_unc_perf_global_ctr: 0x2ff0,
     val_enable_ctrs: 0x20000000, // To validate
@@ -233,6 +235,7 @@ const ALDER_LAKE_TO_RAPTOR_LAKE_CORE: CorePerfCounters = CorePerfCounters {
 
 // TODO verify his on ICELAKE, and appropriate values. Also deal with backport Cypress Cove ?
 const CANNON_LAKE_TO_TIGER_LAKE_CORE: CorePerfCounters = CorePerfCounters {
+    msr_unc_cbo_config: 0x396,
     max_slice: 8, // To validate
     msr_unc_perf_global_ctr: 0xe01,
     val_enable_ctrs: 0x20000000, // To validate
@@ -244,6 +247,7 @@ const CANNON_LAKE_TO_TIGER_LAKE_CORE: CorePerfCounters = CorePerfCounters {
 };
 
 const SKYLAKE_KABYLAKE_CORE: CorePerfCounters = CorePerfCounters {
+    msr_unc_cbo_config: 0x396,
     max_slice: 7,
     msr_unc_perf_global_ctr: 0xe01,
     val_enable_ctrs: 0x20000000,
@@ -257,7 +261,8 @@ const SKYLAKE_KABYLAKE_CORE: CorePerfCounters = CorePerfCounters {
 // This is documented in Intel SDM, 20.3.4.6 (in March 2024 edition)
 
 const SANDYBRIDGE_TO_BROADWELL_CORE: CorePerfCounters = CorePerfCounters {
-    max_slice: 0,
+    msr_unc_cbo_config: 0x396,
+    max_slice: 4,
     msr_unc_perf_global_ctr: 0x391,
     // Go in MSR_UNC_PERF_GLOBAL_CTR EN (bit 29) set to one, and route PMI to core 1-4 upon overflow.
     val_enable_ctrs: 0x2000000f,

cache_slice/src/lib.rs

@@ -3,7 +3,7 @@
 use std::arch::x86_64::_mm_clflush;
 use crate::arch::CpuClass::{IntelCore, IntelXeon, IntelXeonSP};
 use crate::arch::{get_performance_counters_core, get_performance_counters_xeon};
-use crate::Error::UnsupportedCPU;
+use crate::Error::{InvalidParameter, UnsupportedCPU};
 use crate::msr::{read_msr_on_cpu, write_msr_on_cpu};
 
 pub mod msr;
@@ -39,7 +39,7 @@ unsafe fn monitor_xeon(addr: *const u8, cpu: u8, max_cbox: usize) -> Result<Vec<
     };
 
     if (performance_counters.max_slice as usize) < max_cbox {
-        return Err(Error::InvalidParameter);
+        return Err(InvalidParameter);
     }
 
     // Freeze counters
@@ -89,33 +89,44 @@ unsafe fn monitor_xeon(addr: *const u8, cpu: u8, max_cbox: usize) -> Result<Vec<
     Ok(results)
 }
 
-fn monitor_core(addr: *const u8, cpu: u8, mut max_cbox: usize) -> Result<Vec<u64>, Error> {
+fn monitor_core(addr: *const u8, cpu: u8) -> Result<Vec<u64>, Error> {
     // Note, we need to add the workaround for one missing perf counter here.
     let performance_counters = if let Some(p) = get_performance_counters_core() {
         p
     } else {
         return Err(UnsupportedCPU);
     };
+    #[cfg(debug_assertions)]
+    eprint!("Finding the number of CBox available... ");
+    let max_cbox = (read_msr_on_cpu(performance_counters.msr_unc_cbo_config, cpu)? & 0xF) as usize; // TODO magic number (mask for bit 3:0)
+    #[cfg(debug_assertions)]
+    eprintln!("{}", max_cbox);
 
-    let workaround = if (performance_counters.max_slice as usize) + 1 == max_cbox {
-        max_cbox = performance_counters.max_slice as usize;
-        true
-    } else if (performance_counters.max_slice as usize) >= max_cbox {
-        false
-    } else {
-        return Err(Error::InvalidParameter);
-    };
-
-    write_msr_on_cpu(performance_counters.msr_unc_perf_global_ctr, cpu, performance_counters.val_disable_ctrs)?;
-
-    for i in 0..max_cbox {
-        write_msr_on_cpu(performance_counters.msr_unc_cbo_per_ctr0[i], cpu, performance_counters.val_reset_ctrs)?;
+    if max_cbox > performance_counters.max_slice as usize {
+        return Err(InvalidParameter);
     }
 
+    #[cfg(debug_assertions)]
+    eprintln!("Disabling counters");
+    write_msr_on_cpu(performance_counters.msr_unc_perf_global_ctr, cpu, performance_counters.val_disable_ctrs)?;
+
+    #[cfg(debug_assertions)]
+    eprint!("Resetting counters...");
+    for i in 0..max_cbox {
+        eprint!(" {i}");
+        write_msr_on_cpu(performance_counters.msr_unc_cbo_per_ctr0[i], cpu, performance_counters.val_reset_ctrs)?;
+    }
+    #[cfg(debug_assertions)]
+    eprintln!(" ok");
+
+    #[cfg(debug_assertions)]
+    eprintln!("Selecting events");
     for i in 0..max_cbox {
         write_msr_on_cpu(performance_counters.msr_unc_cbo_perfevtsel0[i], cpu, performance_counters.val_select_evt_core)?;
     }
 
+    #[cfg(debug_assertions)]
+    eprintln!("enabling counters");
     write_msr_on_cpu(performance_counters.msr_unc_perf_global_ctr, cpu, performance_counters.val_enable_ctrs)?;
 
     unsafe { poke(addr) };
@@ -126,6 +137,8 @@ fn monitor_core(addr: *const u8, cpu: u8, mut max_cbox: usize) -> Result<Vec<u64
 
     */
     // Read counters
+    #[cfg(debug_assertions)]
+    eprintln!("Gathering results");
     let mut results = Vec::new();
     for i in 0..max_cbox {
         let result = read_msr_on_cpu(performance_counters.msr_unc_cbo_per_ctr0[i], cpu)?;
@@ -136,15 +149,18 @@ fn monitor_core(addr: *const u8, cpu: u8, mut max_cbox: usize) -> Result<Vec<u64
         }
     }
 
+    #[cfg(debug_assertions)]
+    eprintln!("disabling counters again");
     write_msr_on_cpu(performance_counters.msr_unc_perf_global_ctr, cpu, performance_counters.val_disable_ctrs)?;
 
     Ok(results)
 }
 
+// Note: max_cbox is not used on Intel Core.
 pub unsafe fn monitor_address(addr: *const u8, cpu: u8, max_cbox: u16) -> Result<Vec<u64>, Error> {
     match arch::determine_cpu_class() {
         Some(IntelCore) => {
-            unsafe { monitor_core(addr, cpu, max_cbox as usize) }
+            unsafe { monitor_core(addr, cpu) }
         }
         Some(IntelXeon) => {
             unsafe { monitor_xeon(addr, cpu, max_cbox as usize) }