From fac01d11722c92a186b27ee26cd429a8066adfb5 Mon Sep 17 00:00:00 2001 From: Hans de Goede Date: Sun, 23 Feb 2020 15:06:10 +0100 Subject: x86/tsc_msr: Make MSR derived TSC frequency more accurate MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The "Intel 64 and IA-32 Architectures Software Developer’s Manual Volume 4: Model-Specific Registers" has the following table for the values from freq_desc_byt: 000B: 083.3 MHz 001B: 100.0 MHz 010B: 133.3 MHz 011B: 116.7 MHz 100B: 080.0 MHz Notice how for e.g the 83.3 MHz value there are 3 significant digits, which translates to an accuracy of a 1000 ppm, where as a typical crystal oscillator is 20 - 100 ppm, so the accuracy of the frequency format used in the Software Developer’s Manual is not really helpful. As far as we know Bay Trail SoCs use a 25 MHz crystal and Cherry Trail uses a 19.2 MHz crystal, the crystal is the source clock for a root PLL which outputs 1600 and 100 MHz. It is unclear if the root PLL outputs are used directly by the CPU clock PLL or if there is another PLL in between. This does not matter though, we can model the chain of PLLs as a single PLL with a quotient equal to the quotients of all PLLs in the chain multiplied. So we can create a simplified model of the CPU clock setup using a reference clock of 100 MHz plus a quotient which gets us as close to the frequency from the SDM as possible. For the 83.3 MHz example from above this would give 100 MHz * 5 / 6 = 83 and 1/3 MHz, which matches exactly what has been measured on actual hardware. Use a simplified PLL model with a reference clock of 100 MHz for all Bay and Cherry Trail models. This has been tested on the following models: CPU freq before: CPU freq after: Intel N2840 2165.800 MHz 2166.667 MHz Intel Z3736 1332.800 MHz 1333.333 MHz Intel Z3775 1466.300 MHz 1466.667 MHz Intel Z8350 1440.000 MHz 1440.000 MHz Intel Z8750 1600.000 MHz 1600.000 MHz This fixes the time drifting by about 1 second per hour (20 - 30 seconds per day) on (some) devices which rely on the tsc_msr.c code to determine the TSC frequency. Reported-by: Vipul Kumar Suggested-by: Thomas Gleixner Signed-off-by: Hans de Goede Signed-off-by: Thomas Gleixner Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/20200223140610.59612-3-hdegoede@redhat.com --- arch/x86/kernel/tsc_msr.c | 97 +++++++++++++++++++++++++++++++++++++++++------ 1 file changed, 86 insertions(+), 11 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/tsc_msr.c b/arch/x86/kernel/tsc_msr.c index 95030895fffa..c65adaf81384 100644 --- a/arch/x86/kernel/tsc_msr.c +++ b/arch/x86/kernel/tsc_msr.c @@ -17,6 +17,28 @@ #define MAX_NUM_FREQS 16 /* 4 bits to select the frequency */ +/* + * The frequency numbers in the SDM are e.g. 83.3 MHz, which does not contain a + * lot of accuracy which leads to clock drift. As far as we know Bay Trail SoCs + * use a 25 MHz crystal and Cherry Trail uses a 19.2 MHz crystal, the crystal + * is the source clk for a root PLL which outputs 1600 and 100 MHz. It is + * unclear if the root PLL outputs are used directly by the CPU clock PLL or + * if there is another PLL in between. + * This does not matter though, we can model the chain of PLLs as a single PLL + * with a quotient equal to the quotients of all PLLs in the chain multiplied. + * So we can create a simplified model of the CPU clock setup using a reference + * clock of 100 MHz plus a quotient which gets us as close to the frequency + * from the SDM as possible. + * For the 83.3 MHz example from above this would give us 100 MHz * 5 / 6 = + * 83 and 1/3 MHz, which matches exactly what has been measured on actual hw. + */ +#define TSC_REFERENCE_KHZ 100000 + +struct muldiv { + u32 multiplier; + u32 divider; +}; + /* * If MSR_PERF_STAT[31] is set, the maximum resolved bus ratio can be * read in MSR_PLATFORM_ID[12:8], otherwise in MSR_PERF_STAT[44:40]. @@ -26,6 +48,11 @@ */ struct freq_desc { bool use_msr_plat; + struct muldiv muldiv[MAX_NUM_FREQS]; + /* + * Some CPU frequencies in the SDM do not map to known PLL freqs, in + * that case the muldiv array is empty and the freqs array is used. + */ u32 freqs[MAX_NUM_FREQS]; u32 mask; }; @@ -47,31 +74,66 @@ static const struct freq_desc freq_desc_clv = { .mask = 0x07, }; +/* + * Bay Trail SDM MSR_FSB_FREQ frequencies simplified PLL model: + * 000: 100 * 5 / 6 = 83.3333 MHz + * 001: 100 * 1 / 1 = 100.0000 MHz + * 010: 100 * 4 / 3 = 133.3333 MHz + * 011: 100 * 7 / 6 = 116.6667 MHz + * 100: 100 * 4 / 5 = 80.0000 MHz + */ static const struct freq_desc freq_desc_byt = { .use_msr_plat = true, - .freqs = { 83300, 100000, 133300, 116700, 80000, 0, 0, 0 }, + .muldiv = { { 5, 6 }, { 1, 1 }, { 4, 3 }, { 7, 6 }, + { 4, 5 } }, .mask = 0x07, }; +/* + * Cherry Trail SDM MSR_FSB_FREQ frequencies simplified PLL model: + * 0000: 100 * 5 / 6 = 83.3333 MHz + * 0001: 100 * 1 / 1 = 100.0000 MHz + * 0010: 100 * 4 / 3 = 133.3333 MHz + * 0011: 100 * 7 / 6 = 116.6667 MHz + * 0100: 100 * 4 / 5 = 80.0000 MHz + * 0101: 100 * 14 / 15 = 93.3333 MHz + * 0110: 100 * 9 / 10 = 90.0000 MHz + * 0111: 100 * 8 / 9 = 88.8889 MHz + * 1000: 100 * 7 / 8 = 87.5000 MHz + */ static const struct freq_desc freq_desc_cht = { .use_msr_plat = true, - .freqs = { 83300, 100000, 133300, 116700, 80000, 93300, 90000, - 88900, 87500 }, + .muldiv = { { 5, 6 }, { 1, 1 }, { 4, 3 }, { 7, 6 }, + { 4, 5 }, { 14, 15 }, { 9, 10 }, { 8, 9 }, + { 7, 8 } }, .mask = 0x0f, }; +/* + * Merriefield SDM MSR_FSB_FREQ frequencies simplified PLL model: + * 0001: 100 * 1 / 1 = 100.0000 MHz + * 0010: 100 * 4 / 3 = 133.3333 MHz + */ static const struct freq_desc freq_desc_tng = { .use_msr_plat = true, - .freqs = { 0, 100000, 133300, 0, 0, 0, 0, 0 }, + .muldiv = { { 0, 0 }, { 1, 1 }, { 4, 3 } }, .mask = 0x07, }; +/* + * Moorefield SDM MSR_FSB_FREQ frequencies simplified PLL model: + * 0000: 100 * 5 / 6 = 83.3333 MHz + * 0001: 100 * 1 / 1 = 100.0000 MHz + * 0010: 100 * 4 / 3 = 133.3333 MHz + * 0011: 100 * 1 / 1 = 100.0000 MHz + */ static const struct freq_desc freq_desc_ann = { .use_msr_plat = true, - .freqs = { 83300, 100000, 133300, 100000, 0, 0, 0, 0 }, + .muldiv = { { 5, 6 }, { 1, 1 }, { 4, 3 }, { 1, 1 } }, .mask = 0x0f, }; +/* 24 MHz crystal? : 24 * 13 / 4 = 78 MHz */ static const struct freq_desc freq_desc_lgm = { .use_msr_plat = true, .freqs = { 78000, 78000, 78000, 78000, 78000, 78000, 78000, 78000 }, @@ -97,9 +159,10 @@ static const struct x86_cpu_id tsc_msr_cpu_ids[] = { */ unsigned long cpu_khz_from_msr(void) { - u32 lo, hi, ratio, freq; + u32 lo, hi, ratio, freq, tscref; const struct freq_desc *freq_desc; const struct x86_cpu_id *id; + const struct muldiv *md; unsigned long res; int index; @@ -119,12 +182,24 @@ unsigned long cpu_khz_from_msr(void) /* Get FSB FREQ ID */ rdmsr(MSR_FSB_FREQ, lo, hi); index = lo & freq_desc->mask; + md = &freq_desc->muldiv[index]; - /* Map CPU reference clock freq ID(0-7) to CPU reference clock freq(KHz) */ - freq = freq_desc->freqs[index]; - - /* TSC frequency = maximum resolved freq * maximum resolved bus ratio */ - res = freq * ratio; + /* + * Note this also catches cases where the index points to an unpopulated + * part of muldiv, in that case the else will set freq and res to 0. + */ + if (md->divider) { + tscref = TSC_REFERENCE_KHZ * md->multiplier; + freq = DIV_ROUND_CLOSEST(tscref, md->divider); + /* + * Multiplying by ratio before the division has better + * accuracy than just calculating freq * ratio. + */ + res = DIV_ROUND_CLOSEST(tscref * ratio, md->divider); + } else { + freq = freq_desc->freqs[index]; + res = freq * ratio; + } if (freq == 0) pr_err("Error MSR_FSB_FREQ index %d is unknown\n", index); -- cgit v1.2.3