14faea7e0STomasz Jeznach /* 24faea7e0STomasz Jeznach * RISC-V IOMMU - Hardware Performance Monitor (HPM) helpers 34faea7e0STomasz Jeznach * 44faea7e0STomasz Jeznach * Copyright (C) 2022-2023 Rivos Inc. 54faea7e0STomasz Jeznach * 64faea7e0STomasz Jeznach * This program is free software; you can redistribute it and/or modify it 74faea7e0STomasz Jeznach * under the terms and conditions of the GNU General Public License, 84faea7e0STomasz Jeznach * version 2 or later, as published by the Free Software Foundation. 94faea7e0STomasz Jeznach * 104faea7e0STomasz Jeznach * This program is distributed in the hope that it will be useful, 114faea7e0STomasz Jeznach * but WITHOUT ANY WARRANTY; without even the implied warranty of 124faea7e0STomasz Jeznach * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 134faea7e0STomasz Jeznach * GNU General Public License for more details. 144faea7e0STomasz Jeznach * 154faea7e0STomasz Jeznach * You should have received a copy of the GNU General Public License along 164faea7e0STomasz Jeznach * with this program; if not, see <http://www.gnu.org/licenses/>. 174faea7e0STomasz Jeznach */ 184faea7e0STomasz Jeznach 194faea7e0STomasz Jeznach #include "qemu/osdep.h" 204faea7e0STomasz Jeznach #include "qemu/timer.h" 214faea7e0STomasz Jeznach #include "cpu_bits.h" 224faea7e0STomasz Jeznach #include "riscv-iommu-hpm.h" 234faea7e0STomasz Jeznach #include "riscv-iommu.h" 244faea7e0STomasz Jeznach #include "riscv-iommu-bits.h" 254faea7e0STomasz Jeznach #include "trace.h" 264faea7e0STomasz Jeznach 274faea7e0STomasz Jeznach /* For now we assume IOMMU HPM frequency to be 1GHz so 1-cycle is of 1-ns. */ 284faea7e0STomasz Jeznach static inline uint64_t get_cycles(void) 294faea7e0STomasz Jeznach { 304faea7e0STomasz Jeznach return qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL); 314faea7e0STomasz Jeznach } 324faea7e0STomasz Jeznach 334faea7e0STomasz Jeznach uint64_t riscv_iommu_hpmcycle_read(RISCVIOMMUState *s) 344faea7e0STomasz Jeznach { 354faea7e0STomasz Jeznach const uint64_t cycle = riscv_iommu_reg_get64( 364faea7e0STomasz Jeznach s, RISCV_IOMMU_REG_IOHPMCYCLES); 374faea7e0STomasz Jeznach const uint32_t inhibit = riscv_iommu_reg_get32( 384faea7e0STomasz Jeznach s, RISCV_IOMMU_REG_IOCOUNTINH); 394faea7e0STomasz Jeznach const uint64_t ctr_prev = s->hpmcycle_prev; 404faea7e0STomasz Jeznach const uint64_t ctr_val = s->hpmcycle_val; 414faea7e0STomasz Jeznach 424faea7e0STomasz Jeznach if (get_field(inhibit, RISCV_IOMMU_IOCOUNTINH_CY)) { 434faea7e0STomasz Jeznach /* 444faea7e0STomasz Jeznach * Counter should not increment if inhibit bit is set. We can't really 454faea7e0STomasz Jeznach * stop the QEMU_CLOCK_VIRTUAL, so we just return the last updated 464faea7e0STomasz Jeznach * counter value to indicate that counter was not incremented. 474faea7e0STomasz Jeznach */ 484faea7e0STomasz Jeznach return (ctr_val & RISCV_IOMMU_IOHPMCYCLES_COUNTER) | 494faea7e0STomasz Jeznach (cycle & RISCV_IOMMU_IOHPMCYCLES_OVF); 504faea7e0STomasz Jeznach } 514faea7e0STomasz Jeznach 524faea7e0STomasz Jeznach return (ctr_val + get_cycles() - ctr_prev) | 534faea7e0STomasz Jeznach (cycle & RISCV_IOMMU_IOHPMCYCLES_OVF); 544faea7e0STomasz Jeznach } 5511ecf24cSTomasz Jeznach 5611ecf24cSTomasz Jeznach static void hpm_incr_ctr(RISCVIOMMUState *s, uint32_t ctr_idx) 5711ecf24cSTomasz Jeznach { 5811ecf24cSTomasz Jeznach const uint32_t off = ctr_idx << 3; 5911ecf24cSTomasz Jeznach uint64_t cntr_val; 6011ecf24cSTomasz Jeznach 6111ecf24cSTomasz Jeznach cntr_val = ldq_le_p(&s->regs_rw[RISCV_IOMMU_REG_IOHPMCTR_BASE + off]); 6211ecf24cSTomasz Jeznach stq_le_p(&s->regs_rw[RISCV_IOMMU_REG_IOHPMCTR_BASE + off], cntr_val + 1); 6311ecf24cSTomasz Jeznach 6411ecf24cSTomasz Jeznach /* Handle the overflow scenario. */ 6511ecf24cSTomasz Jeznach if (cntr_val == UINT64_MAX) { 6611ecf24cSTomasz Jeznach /* 6711ecf24cSTomasz Jeznach * Generate interrupt only if OF bit is clear. +1 to offset the cycle 6811ecf24cSTomasz Jeznach * register OF bit. 6911ecf24cSTomasz Jeznach */ 7011ecf24cSTomasz Jeznach const uint32_t ovf = 7111ecf24cSTomasz Jeznach riscv_iommu_reg_mod32(s, RISCV_IOMMU_REG_IOCOUNTOVF, 7211ecf24cSTomasz Jeznach BIT(ctr_idx + 1), 0); 7311ecf24cSTomasz Jeznach if (!get_field(ovf, BIT(ctr_idx + 1))) { 7411ecf24cSTomasz Jeznach riscv_iommu_reg_mod64(s, 7511ecf24cSTomasz Jeznach RISCV_IOMMU_REG_IOHPMEVT_BASE + off, 7611ecf24cSTomasz Jeznach RISCV_IOMMU_IOHPMEVT_OF, 7711ecf24cSTomasz Jeznach 0); 7811ecf24cSTomasz Jeznach riscv_iommu_notify(s, RISCV_IOMMU_INTR_PM); 7911ecf24cSTomasz Jeznach } 8011ecf24cSTomasz Jeznach } 8111ecf24cSTomasz Jeznach } 8211ecf24cSTomasz Jeznach 8311ecf24cSTomasz Jeznach void riscv_iommu_hpm_incr_ctr(RISCVIOMMUState *s, RISCVIOMMUContext *ctx, 8411ecf24cSTomasz Jeznach unsigned event_id) 8511ecf24cSTomasz Jeznach { 8611ecf24cSTomasz Jeznach const uint32_t inhibit = riscv_iommu_reg_get32( 8711ecf24cSTomasz Jeznach s, RISCV_IOMMU_REG_IOCOUNTINH); 8811ecf24cSTomasz Jeznach uint32_t did_gscid; 8911ecf24cSTomasz Jeznach uint32_t pid_pscid; 9011ecf24cSTomasz Jeznach uint32_t ctr_idx; 9111ecf24cSTomasz Jeznach gpointer value; 9211ecf24cSTomasz Jeznach uint32_t ctrs; 9311ecf24cSTomasz Jeznach uint64_t evt; 9411ecf24cSTomasz Jeznach 9511ecf24cSTomasz Jeznach if (!(s->cap & RISCV_IOMMU_CAP_HPM)) { 9611ecf24cSTomasz Jeznach return; 9711ecf24cSTomasz Jeznach } 9811ecf24cSTomasz Jeznach 9911ecf24cSTomasz Jeznach value = g_hash_table_lookup(s->hpm_event_ctr_map, 10011ecf24cSTomasz Jeznach GUINT_TO_POINTER(event_id)); 10111ecf24cSTomasz Jeznach if (value == NULL) { 10211ecf24cSTomasz Jeznach return; 10311ecf24cSTomasz Jeznach } 10411ecf24cSTomasz Jeznach 10511ecf24cSTomasz Jeznach for (ctrs = GPOINTER_TO_UINT(value); ctrs != 0; ctrs &= ctrs - 1) { 10611ecf24cSTomasz Jeznach ctr_idx = ctz32(ctrs); 10711ecf24cSTomasz Jeznach if (get_field(inhibit, BIT(ctr_idx + 1))) { 10811ecf24cSTomasz Jeznach continue; 10911ecf24cSTomasz Jeznach } 11011ecf24cSTomasz Jeznach 11111ecf24cSTomasz Jeznach evt = riscv_iommu_reg_get64(s, 11211ecf24cSTomasz Jeznach RISCV_IOMMU_REG_IOHPMEVT_BASE + (ctr_idx << 3)); 11311ecf24cSTomasz Jeznach 11411ecf24cSTomasz Jeznach /* 11511ecf24cSTomasz Jeznach * It's quite possible that event ID has been changed in counter 11611ecf24cSTomasz Jeznach * but hashtable hasn't been updated yet. We don't want to increment 11711ecf24cSTomasz Jeznach * counter for the old event ID. 11811ecf24cSTomasz Jeznach */ 11911ecf24cSTomasz Jeznach if (event_id != get_field(evt, RISCV_IOMMU_IOHPMEVT_EVENT_ID)) { 12011ecf24cSTomasz Jeznach continue; 12111ecf24cSTomasz Jeznach } 12211ecf24cSTomasz Jeznach 12311ecf24cSTomasz Jeznach if (get_field(evt, RISCV_IOMMU_IOHPMEVT_IDT)) { 12411ecf24cSTomasz Jeznach did_gscid = get_field(ctx->gatp, RISCV_IOMMU_DC_IOHGATP_GSCID); 12511ecf24cSTomasz Jeznach pid_pscid = get_field(ctx->ta, RISCV_IOMMU_DC_TA_PSCID); 12611ecf24cSTomasz Jeznach } else { 12711ecf24cSTomasz Jeznach did_gscid = ctx->devid; 12811ecf24cSTomasz Jeznach pid_pscid = ctx->process_id; 12911ecf24cSTomasz Jeznach } 13011ecf24cSTomasz Jeznach 13111ecf24cSTomasz Jeznach if (get_field(evt, RISCV_IOMMU_IOHPMEVT_PV_PSCV)) { 13211ecf24cSTomasz Jeznach /* 13311ecf24cSTomasz Jeznach * If the transaction does not have a valid process_id, counter 13411ecf24cSTomasz Jeznach * increments if device_id matches DID_GSCID. If the transaction 13511ecf24cSTomasz Jeznach * has a valid process_id, counter increments if device_id 13611ecf24cSTomasz Jeznach * matches DID_GSCID and process_id matches PID_PSCID. See 13711ecf24cSTomasz Jeznach * IOMMU Specification, Chapter 5.23. Performance-monitoring 13811ecf24cSTomasz Jeznach * event selector. 13911ecf24cSTomasz Jeznach */ 14011ecf24cSTomasz Jeznach if (ctx->process_id && 14111ecf24cSTomasz Jeznach get_field(evt, RISCV_IOMMU_IOHPMEVT_PID_PSCID) != pid_pscid) { 14211ecf24cSTomasz Jeznach continue; 14311ecf24cSTomasz Jeznach } 14411ecf24cSTomasz Jeznach } 14511ecf24cSTomasz Jeznach 14611ecf24cSTomasz Jeznach if (get_field(evt, RISCV_IOMMU_IOHPMEVT_DV_GSCV)) { 14711ecf24cSTomasz Jeznach uint32_t mask = ~0; 14811ecf24cSTomasz Jeznach 14911ecf24cSTomasz Jeznach if (get_field(evt, RISCV_IOMMU_IOHPMEVT_DMASK)) { 15011ecf24cSTomasz Jeznach /* 15111ecf24cSTomasz Jeznach * 1001 1011 mask = GSCID 15211ecf24cSTomasz Jeznach * 0000 0111 mask = mask ^ (mask + 1) 15311ecf24cSTomasz Jeznach * 1111 1000 mask = ~mask; 15411ecf24cSTomasz Jeznach */ 15511ecf24cSTomasz Jeznach mask = get_field(evt, RISCV_IOMMU_IOHPMEVT_DID_GSCID); 15611ecf24cSTomasz Jeznach mask = mask ^ (mask + 1); 15711ecf24cSTomasz Jeznach mask = ~mask; 15811ecf24cSTomasz Jeznach } 15911ecf24cSTomasz Jeznach 16011ecf24cSTomasz Jeznach if ((get_field(evt, RISCV_IOMMU_IOHPMEVT_DID_GSCID) & mask) != 16111ecf24cSTomasz Jeznach (did_gscid & mask)) { 16211ecf24cSTomasz Jeznach continue; 16311ecf24cSTomasz Jeznach } 16411ecf24cSTomasz Jeznach } 16511ecf24cSTomasz Jeznach 16611ecf24cSTomasz Jeznach hpm_incr_ctr(s, ctr_idx); 16711ecf24cSTomasz Jeznach } 16811ecf24cSTomasz Jeznach } 169*ffb37df0STomasz Jeznach 170*ffb37df0STomasz Jeznach /* Timer callback for cycle counter overflow. */ 171*ffb37df0STomasz Jeznach void riscv_iommu_hpm_timer_cb(void *priv) 172*ffb37df0STomasz Jeznach { 173*ffb37df0STomasz Jeznach RISCVIOMMUState *s = priv; 174*ffb37df0STomasz Jeznach const uint32_t inhibit = riscv_iommu_reg_get32( 175*ffb37df0STomasz Jeznach s, RISCV_IOMMU_REG_IOCOUNTINH); 176*ffb37df0STomasz Jeznach uint32_t ovf; 177*ffb37df0STomasz Jeznach 178*ffb37df0STomasz Jeznach if (get_field(inhibit, RISCV_IOMMU_IOCOUNTINH_CY)) { 179*ffb37df0STomasz Jeznach return; 180*ffb37df0STomasz Jeznach } 181*ffb37df0STomasz Jeznach 182*ffb37df0STomasz Jeznach if (s->irq_overflow_left > 0) { 183*ffb37df0STomasz Jeznach uint64_t irq_trigger_at = 184*ffb37df0STomasz Jeznach qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + s->irq_overflow_left; 185*ffb37df0STomasz Jeznach timer_mod_anticipate_ns(s->hpm_timer, irq_trigger_at); 186*ffb37df0STomasz Jeznach s->irq_overflow_left = 0; 187*ffb37df0STomasz Jeznach return; 188*ffb37df0STomasz Jeznach } 189*ffb37df0STomasz Jeznach 190*ffb37df0STomasz Jeznach ovf = riscv_iommu_reg_get32(s, RISCV_IOMMU_REG_IOCOUNTOVF); 191*ffb37df0STomasz Jeznach if (!get_field(ovf, RISCV_IOMMU_IOCOUNTOVF_CY)) { 192*ffb37df0STomasz Jeznach /* 193*ffb37df0STomasz Jeznach * We don't need to set hpmcycle_val to zero and update hpmcycle_prev to 194*ffb37df0STomasz Jeznach * current clock value. The way we calculate iohpmcycs will overflow 195*ffb37df0STomasz Jeznach * and return the correct value. This avoids the need to synchronize 196*ffb37df0STomasz Jeznach * timer callback and write callback. 197*ffb37df0STomasz Jeznach */ 198*ffb37df0STomasz Jeznach riscv_iommu_reg_mod32(s, RISCV_IOMMU_REG_IOCOUNTOVF, 199*ffb37df0STomasz Jeznach RISCV_IOMMU_IOCOUNTOVF_CY, 0); 200*ffb37df0STomasz Jeznach riscv_iommu_reg_mod64(s, RISCV_IOMMU_REG_IOHPMCYCLES, 201*ffb37df0STomasz Jeznach RISCV_IOMMU_IOHPMCYCLES_OVF, 0); 202*ffb37df0STomasz Jeznach riscv_iommu_notify(s, RISCV_IOMMU_INTR_PM); 203*ffb37df0STomasz Jeznach } 204*ffb37df0STomasz Jeznach } 205