Index: head/sys/amd64/vmm/amd/amdvi_hw.c =================================================================== --- head/sys/amd64/vmm/amd/amdvi_hw.c (revision 357973) +++ head/sys/amd64/vmm/amd/amdvi_hw.c (revision 357974) @@ -1,1459 +1,1460 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2016, Anish Gupta (anish@freebsd.org) * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "pcib_if.h" #include "io/iommu.h" #include "amdvi_priv.h" SYSCTL_DECL(_hw_vmm); -SYSCTL_NODE(_hw_vmm, OID_AUTO, amdvi, CTLFLAG_RW, NULL, NULL); +SYSCTL_NODE(_hw_vmm, OID_AUTO, amdvi, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, + NULL); #define MOD_INC(a, s, m) (((a) + (s)) % ((m) * (s))) #define MOD_DEC(a, s, m) (((a) - (s)) % ((m) * (s))) /* Print RID or device ID in PCI string format. */ #define RID2PCI_STR(d) PCI_RID2BUS(d), PCI_RID2SLOT(d), PCI_RID2FUNC(d) static void amdvi_dump_cmds(struct amdvi_softc *softc, int count); static void amdvi_print_dev_cap(struct amdvi_softc *softc); MALLOC_DEFINE(M_AMDVI, "amdvi", "amdvi"); extern device_t *ivhd_devs; extern int ivhd_count; SYSCTL_INT(_hw_vmm_amdvi, OID_AUTO, count, CTLFLAG_RDTUN, &ivhd_count, 0, NULL); static int amdvi_enable_user = 0; SYSCTL_INT(_hw_vmm_amdvi, OID_AUTO, enable, CTLFLAG_RDTUN, &amdvi_enable_user, 0, NULL); TUNABLE_INT("hw.vmm.amdvi_enable", &amdvi_enable_user); #ifdef AMDVI_ATS_ENABLE /* XXX: ATS is not tested. */ static int amdvi_enable_iotlb = 1; SYSCTL_INT(_hw_vmm_amdvi, OID_AUTO, iotlb_enabled, CTLFLAG_RDTUN, &amdvi_enable_iotlb, 0, NULL); TUNABLE_INT("hw.vmm.enable_iotlb", &amdvi_enable_iotlb); #endif static int amdvi_host_ptp = 1; /* Use page tables for host. */ SYSCTL_INT(_hw_vmm_amdvi, OID_AUTO, host_ptp, CTLFLAG_RDTUN, &amdvi_host_ptp, 0, NULL); TUNABLE_INT("hw.vmm.amdvi.host_ptp", &amdvi_host_ptp); /* Page table level used <= supported by h/w[v1=7]. */ int amdvi_ptp_level = 4; SYSCTL_INT(_hw_vmm_amdvi, OID_AUTO, ptp_level, CTLFLAG_RDTUN, &amdvi_ptp_level, 0, NULL); TUNABLE_INT("hw.vmm.amdvi.ptp_level", &amdvi_ptp_level); /* Disable fault event reporting. */ static int amdvi_disable_io_fault = 0; SYSCTL_INT(_hw_vmm_amdvi, OID_AUTO, disable_io_fault, CTLFLAG_RDTUN, &amdvi_disable_io_fault, 0, NULL); TUNABLE_INT("hw.vmm.amdvi.disable_io_fault", &amdvi_disable_io_fault); static uint32_t amdvi_dom_id = 0; /* 0 is reserved for host. */ SYSCTL_UINT(_hw_vmm_amdvi, OID_AUTO, domain_id, CTLFLAG_RD, &amdvi_dom_id, 0, NULL); /* * Device table entry. * Bus(256) x Dev(32) x Fun(8) x DTE(256 bits or 32 bytes). * = 256 * 2 * PAGE_SIZE. */ static struct amdvi_dte amdvi_dte[PCI_NUM_DEV_MAX] __aligned(PAGE_SIZE); CTASSERT(PCI_NUM_DEV_MAX == 0x10000); CTASSERT(sizeof(amdvi_dte) == 0x200000); static SLIST_HEAD (, amdvi_domain) dom_head; static inline uint32_t amdvi_pci_read(struct amdvi_softc *softc, int off) { return (pci_cfgregread(PCI_RID2BUS(softc->pci_rid), PCI_RID2SLOT(softc->pci_rid), PCI_RID2FUNC(softc->pci_rid), off, 4)); } #ifdef AMDVI_ATS_ENABLE /* XXX: Should be in pci.c */ /* * Check if device has ATS capability and its enabled. * If ATS is absent or disabled, return (-1), otherwise ATS * queue length. */ static int amdvi_find_ats_qlen(uint16_t devid) { device_t dev; uint32_t off, cap; int qlen = -1; dev = pci_find_bsf(PCI_RID2BUS(devid), PCI_RID2SLOT(devid), PCI_RID2FUNC(devid)); if (!dev) { return (-1); } #define PCIM_ATS_EN BIT(31) if (pci_find_extcap(dev, PCIZ_ATS, &off) == 0) { cap = pci_read_config(dev, off + 4, 4); qlen = (cap & 0x1F); qlen = qlen ? qlen : 32; printf("AMD-Vi: PCI device %d.%d.%d ATS %s qlen=%d\n", RID2PCI_STR(devid), (cap & PCIM_ATS_EN) ? "enabled" : "Disabled", qlen); qlen = (cap & PCIM_ATS_EN) ? qlen : -1; } return (qlen); } /* * Check if an endpoint device support device IOTLB or ATS. */ static inline bool amdvi_dev_support_iotlb(struct amdvi_softc *softc, uint16_t devid) { struct ivhd_dev_cfg *cfg; int qlen, i; bool pci_ats, ivhd_ats; qlen = amdvi_find_ats_qlen(devid); if (qlen < 0) return (false); KASSERT(softc, ("softc is NULL")); cfg = softc->dev_cfg; ivhd_ats = false; for (i = 0; i < softc->dev_cfg_cnt; i++) { if ((cfg->start_id <= devid) && (cfg->end_id >= devid)) { ivhd_ats = cfg->enable_ats; break; } cfg++; } pci_ats = (qlen < 0) ? false : true; if (pci_ats != ivhd_ats) device_printf(softc->dev, "BIOS bug: mismatch in ATS setting for %d.%d.%d," "ATS inv qlen = %d\n", RID2PCI_STR(devid), qlen); /* Ignore IVRS setting and respect PCI setting. */ return (pci_ats); } #endif /* Enable IOTLB support for IOMMU if its supported. */ static inline void amdvi_hw_enable_iotlb(struct amdvi_softc *softc) { #ifndef AMDVI_ATS_ENABLE softc->iotlb = false; #else bool supported; supported = (softc->ivhd_flag & IVHD_FLAG_IOTLB) ? true : false; if (softc->pci_cap & AMDVI_PCI_CAP_IOTLB) { if (!supported) device_printf(softc->dev, "IOTLB disabled by BIOS.\n"); if (supported && !amdvi_enable_iotlb) { device_printf(softc->dev, "IOTLB disabled by user.\n"); supported = false; } } else supported = false; softc->iotlb = supported; #endif } static int amdvi_init_cmd(struct amdvi_softc *softc) { struct amdvi_ctrl *ctrl = softc->ctrl; ctrl->cmd.len = 8; /* Use 256 command buffer entries. */ softc->cmd_max = 1 << ctrl->cmd.len; softc->cmd = malloc(sizeof(struct amdvi_cmd) * softc->cmd_max, M_AMDVI, M_WAITOK | M_ZERO); if ((uintptr_t)softc->cmd & PAGE_MASK) panic("AMDVi: Command buffer not aligned on page boundary."); ctrl->cmd.base = vtophys(softc->cmd) / PAGE_SIZE; /* * XXX: Reset the h/w pointers in case IOMMU is restarting, * h/w doesn't clear these pointers based on empirical data. */ ctrl->cmd_tail = 0; ctrl->cmd_head = 0; return (0); } /* * Note: Update tail pointer after we have written the command since tail * pointer update cause h/w to execute new commands, see section 3.3 * of AMD IOMMU spec ver 2.0. */ /* Get the command tail pointer w/o updating it. */ static struct amdvi_cmd * amdvi_get_cmd_tail(struct amdvi_softc *softc) { struct amdvi_ctrl *ctrl; struct amdvi_cmd *tail; KASSERT(softc, ("softc is NULL")); KASSERT(softc->cmd != NULL, ("cmd is NULL")); ctrl = softc->ctrl; KASSERT(ctrl != NULL, ("ctrl is NULL")); tail = (struct amdvi_cmd *)((uint8_t *)softc->cmd + ctrl->cmd_tail); return (tail); } /* * Update the command tail pointer which will start command execution. */ static void amdvi_update_cmd_tail(struct amdvi_softc *softc) { struct amdvi_ctrl *ctrl; int size; size = sizeof(struct amdvi_cmd); KASSERT(softc->cmd != NULL, ("cmd is NULL")); ctrl = softc->ctrl; KASSERT(ctrl != NULL, ("ctrl is NULL")); ctrl->cmd_tail = MOD_INC(ctrl->cmd_tail, size, softc->cmd_max); softc->total_cmd++; #ifdef AMDVI_DEBUG_CMD device_printf(softc->dev, "cmd_tail: %s Tail:0x%x, Head:0x%x.\n", ctrl->cmd_tail, ctrl->cmd_head); #endif } /* * Various commands supported by IOMMU. */ /* Completion wait command. */ static void amdvi_cmd_cmp(struct amdvi_softc *softc, const uint64_t data) { struct amdvi_cmd *cmd; uint64_t pa; cmd = amdvi_get_cmd_tail(softc); KASSERT(cmd != NULL, ("Cmd is NULL")); pa = vtophys(&softc->cmp_data); cmd->opcode = AMDVI_CMP_WAIT_OPCODE; cmd->word0 = (pa & 0xFFFFFFF8) | AMDVI_CMP_WAIT_STORE; cmd->word1 = (pa >> 32) & 0xFFFFF; cmd->addr = data; amdvi_update_cmd_tail(softc); } /* Invalidate device table entry. */ static void amdvi_cmd_inv_dte(struct amdvi_softc *softc, uint16_t devid) { struct amdvi_cmd *cmd; cmd = amdvi_get_cmd_tail(softc); KASSERT(cmd != NULL, ("Cmd is NULL")); cmd->opcode = AMDVI_INVD_DTE_OPCODE; cmd->word0 = devid; amdvi_update_cmd_tail(softc); #ifdef AMDVI_DEBUG_CMD device_printf(softc->dev, "Invalidated DTE:0x%x\n", devid); #endif } /* Invalidate IOMMU page, use for invalidation of domain. */ static void amdvi_cmd_inv_iommu_pages(struct amdvi_softc *softc, uint16_t domain_id, uint64_t addr, bool guest_nested, bool pde, bool page) { struct amdvi_cmd *cmd; cmd = amdvi_get_cmd_tail(softc); KASSERT(cmd != NULL, ("Cmd is NULL")); cmd->opcode = AMDVI_INVD_PAGE_OPCODE; cmd->word1 = domain_id; /* * Invalidate all addresses for this domain. */ cmd->addr = addr; cmd->addr |= pde ? AMDVI_INVD_PAGE_PDE : 0; cmd->addr |= page ? AMDVI_INVD_PAGE_S : 0; amdvi_update_cmd_tail(softc); } #ifdef AMDVI_ATS_ENABLE /* Invalidate device IOTLB. */ static void amdvi_cmd_inv_iotlb(struct amdvi_softc *softc, uint16_t devid) { struct amdvi_cmd *cmd; int qlen; if (!softc->iotlb) return; qlen = amdvi_find_ats_qlen(devid); if (qlen < 0) { panic("AMDVI: Invalid ATS qlen(%d) for device %d.%d.%d\n", qlen, RID2PCI_STR(devid)); } cmd = amdvi_get_cmd_tail(softc); KASSERT(cmd != NULL, ("Cmd is NULL")); #ifdef AMDVI_DEBUG_CMD device_printf(softc->dev, "Invalidate IOTLB devID 0x%x" " Qlen:%d\n", devid, qlen); #endif cmd->opcode = AMDVI_INVD_IOTLB_OPCODE; cmd->word0 = devid; cmd->word1 = qlen; cmd->addr = AMDVI_INVD_IOTLB_ALL_ADDR | AMDVI_INVD_IOTLB_S; amdvi_update_cmd_tail(softc); } #endif #ifdef notyet /* For Interrupt Remap. */ static void amdvi_cmd_inv_intr_map(struct amdvi_softc *softc, uint16_t devid) { struct amdvi_cmd *cmd; cmd = amdvi_get_cmd_tail(softc); KASSERT(cmd != NULL, ("Cmd is NULL")); cmd->opcode = AMDVI_INVD_INTR_OPCODE; cmd->word0 = devid; amdvi_update_cmd_tail(softc); #ifdef AMDVI_DEBUG_CMD device_printf(softc->dev, "Invalidate INTR map of devID 0x%x\n", devid); #endif } #endif /* Invalidate domain using INVALIDATE_IOMMU_PAGES command. */ static void amdvi_inv_domain(struct amdvi_softc *softc, uint16_t domain_id) { struct amdvi_cmd *cmd; cmd = amdvi_get_cmd_tail(softc); KASSERT(cmd != NULL, ("Cmd is NULL")); /* * See section 3.3.3 of IOMMU spec rev 2.0, software note * for invalidating domain. */ amdvi_cmd_inv_iommu_pages(softc, domain_id, AMDVI_INVD_PAGE_ALL_ADDR, false, true, true); #ifdef AMDVI_DEBUG_CMD device_printf(softc->dev, "Invalidate domain:0x%x\n", domain_id); #endif } static bool amdvi_cmp_wait(struct amdvi_softc *softc) { struct amdvi_ctrl *ctrl; const uint64_t VERIFY = 0xA5A5; volatile uint64_t *read; int i; bool status; ctrl = softc->ctrl; read = &softc->cmp_data; *read = 0; amdvi_cmd_cmp(softc, VERIFY); /* Wait for h/w to update completion data. */ for (i = 0; i < 100 && (*read != VERIFY); i++) { DELAY(1000); /* 1 ms */ } status = (VERIFY == softc->cmp_data) ? true : false; #ifdef AMDVI_DEBUG_CMD if (status) device_printf(softc->dev, "CMD completion DONE Tail:0x%x, " "Head:0x%x, loop:%d.\n", ctrl->cmd_tail, ctrl->cmd_head, loop); #endif return (status); } static void amdvi_wait(struct amdvi_softc *softc) { struct amdvi_ctrl *ctrl; int i; KASSERT(softc, ("softc is NULL")); ctrl = softc->ctrl; KASSERT(ctrl != NULL, ("ctrl is NULL")); /* Don't wait if h/w is not enabled. */ if ((ctrl->control & AMDVI_CTRL_EN) == 0) return; for (i = 0; i < 10; i++) { if (amdvi_cmp_wait(softc)) return; } device_printf(softc->dev, "Error: completion failed" " tail:0x%x, head:0x%x.\n", ctrl->cmd_tail, ctrl->cmd_head); /* Dump the last command. */ amdvi_dump_cmds(softc, 1); } static void amdvi_dump_cmds(struct amdvi_softc *softc, int count) { struct amdvi_ctrl *ctrl; struct amdvi_cmd *cmd; int off, i; ctrl = softc->ctrl; device_printf(softc->dev, "Dump last %d command(s):\n", count); /* * If h/w is stuck in completion, it is the previous command, * start dumping from previous command onward. */ off = MOD_DEC(ctrl->cmd_head, sizeof(struct amdvi_cmd), softc->cmd_max); for (i = 0; off != ctrl->cmd_tail && i < count; i++) { cmd = (struct amdvi_cmd *)((uint8_t *)softc->cmd + off); printf(" [CMD%d, off:0x%x] opcode= 0x%x 0x%x" " 0x%x 0x%lx\n", i, off, cmd->opcode, cmd->word0, cmd->word1, cmd->addr); off = (off + sizeof(struct amdvi_cmd)) % (softc->cmd_max * sizeof(struct amdvi_cmd)); } } static int amdvi_init_event(struct amdvi_softc *softc) { struct amdvi_ctrl *ctrl; ctrl = softc->ctrl; ctrl->event.len = 8; softc->event_max = 1 << ctrl->event.len; softc->event = malloc(sizeof(struct amdvi_event) * softc->event_max, M_AMDVI, M_WAITOK | M_ZERO); if ((uintptr_t)softc->event & PAGE_MASK) { device_printf(softc->dev, "Event buffer not aligned on page."); return (false); } ctrl->event.base = vtophys(softc->event) / PAGE_SIZE; /* Reset the pointers. */ ctrl->evt_head = 0; ctrl->evt_tail = 0; return (0); } static inline void amdvi_decode_evt_flag(uint16_t flag) { flag &= AMDVI_EVENT_FLAG_MASK; printf(" 0x%b]\n", flag, "\020" "\001GN" "\002NX" "\003US" "\004I" "\005PR" "\006RW" "\007PE" "\010RZ" "\011TR" ); } /* See section 2.5.4 of AMD IOMMU spec ver 2.62.*/ static inline void amdvi_decode_evt_flag_type(uint8_t type) { switch (AMDVI_EVENT_FLAG_TYPE(type)) { case 0: printf("RSVD\n"); break; case 1: printf("Master Abort\n"); break; case 2: printf("Target Abort\n"); break; case 3: printf("Data Err\n"); break; default: break; } } static void amdvi_decode_inv_dte_evt(uint16_t devid, uint16_t domid, uint64_t addr, uint16_t flag) { printf("\t[IO_PAGE_FAULT EVT: devId:0x%x DomId:0x%x" " Addr:0x%lx", devid, domid, addr); amdvi_decode_evt_flag(flag); } static void amdvi_decode_pf_evt(uint16_t devid, uint16_t domid, uint64_t addr, uint16_t flag) { printf("\t[IO_PAGE_FAULT EVT: devId:0x%x DomId:0x%x" " Addr:0x%lx", devid, domid, addr); amdvi_decode_evt_flag(flag); } static void amdvi_decode_dte_hwerr_evt(uint16_t devid, uint16_t domid, uint64_t addr, uint16_t flag) { printf("\t[DEV_TAB_HW_ERR EVT: devId:0x%x DomId:0x%x" " Addr:0x%lx", devid, domid, addr); amdvi_decode_evt_flag(flag); amdvi_decode_evt_flag_type(flag); } static void amdvi_decode_page_hwerr_evt(uint16_t devid, uint16_t domid, uint64_t addr, uint16_t flag) { printf("\t[PAGE_TAB_HW_ERR EVT: devId:0x%x DomId:0x%x" " Addr:0x%lx", devid, domid, addr); amdvi_decode_evt_flag(flag); amdvi_decode_evt_flag_type(AMDVI_EVENT_FLAG_TYPE(flag)); } static void amdvi_decode_evt(struct amdvi_event *evt) { struct amdvi_cmd *cmd; switch (evt->opcode) { case AMDVI_EVENT_INVALID_DTE: amdvi_decode_inv_dte_evt(evt->devid, evt->pasid_domid, evt->addr, evt->flag); break; case AMDVI_EVENT_PFAULT: amdvi_decode_pf_evt(evt->devid, evt->pasid_domid, evt->addr, evt->flag); break; case AMDVI_EVENT_DTE_HW_ERROR: amdvi_decode_dte_hwerr_evt(evt->devid, evt->pasid_domid, evt->addr, evt->flag); break; case AMDVI_EVENT_PAGE_HW_ERROR: amdvi_decode_page_hwerr_evt(evt->devid, evt->pasid_domid, evt->addr, evt->flag); break; case AMDVI_EVENT_ILLEGAL_CMD: /* FALL THROUGH */ case AMDVI_EVENT_CMD_HW_ERROR: printf("\t[%s EVT]\n", (evt->opcode == AMDVI_EVENT_ILLEGAL_CMD) ? "ILLEGAL CMD" : "CMD HW ERR"); cmd = (struct amdvi_cmd *)PHYS_TO_DMAP(evt->addr); printf("\tCMD opcode= 0x%x 0x%x 0x%x 0x%lx\n", cmd->opcode, cmd->word0, cmd->word1, cmd->addr); break; case AMDVI_EVENT_IOTLB_TIMEOUT: printf("\t[IOTLB_INV_TIMEOUT devid:0x%x addr:0x%lx]\n", evt->devid, evt->addr); break; case AMDVI_EVENT_INVALID_DTE_REQ: printf("\t[INV_DTE devid:0x%x addr:0x%lx type:0x%x tr:%d]\n", evt->devid, evt->addr, evt->flag >> 9, (evt->flag >> 8) & 1); break; case AMDVI_EVENT_INVALID_PPR_REQ: case AMDVI_EVENT_COUNTER_ZERO: printf("AMD-Vi: v2 events.\n"); break; default: printf("Unsupported AMD-Vi event:%d\n", evt->opcode); } } static void amdvi_print_events(struct amdvi_softc *softc) { struct amdvi_ctrl *ctrl; struct amdvi_event *event; int i, size; ctrl = softc->ctrl; size = sizeof(struct amdvi_event); for (i = 0; i < softc->event_max; i++) { event = &softc->event[ctrl->evt_head / size]; if (!event->opcode) break; device_printf(softc->dev, "\t[Event%d: Head:0x%x Tail:0x%x]\n", i, ctrl->evt_head, ctrl->evt_tail); amdvi_decode_evt(event); ctrl->evt_head = MOD_INC(ctrl->evt_head, size, softc->event_max); } } static int amdvi_init_dte(struct amdvi_softc *softc) { struct amdvi_ctrl *ctrl; ctrl = softc->ctrl; ctrl->dte.base = vtophys(amdvi_dte) / PAGE_SIZE; ctrl->dte.size = 0x1FF; /* 2MB device table. */ return (0); } /* * Not all capabilities of IOMMU are available in ACPI IVHD flag * or EFR entry, read directly from device. */ static int amdvi_print_pci_cap(device_t dev) { struct amdvi_softc *softc; uint32_t off, cap; softc = device_get_softc(dev); off = softc->cap_off; /* * Section 3.7.1 of IOMMU sepc rev 2.0. * Read capability from device. */ cap = amdvi_pci_read(softc, off); /* Make sure capability type[18:16] is 3. */ KASSERT((((cap >> 16) & 0x7) == 0x3), ("Not a IOMMU capability 0x%x@0x%x", cap, off)); softc->pci_cap = cap >> 24; device_printf(softc->dev, "PCI cap 0x%x@0x%x feature:%b\n", cap, off, softc->pci_cap, "\20\1IOTLB\2HT\3NPCache\4EFR\5CapExt"); return (0); } static void amdvi_event_intr(void *arg) { struct amdvi_softc *softc; struct amdvi_ctrl *ctrl; softc = (struct amdvi_softc *)arg; ctrl = softc->ctrl; device_printf(softc->dev, "EVT INTR %ld Status:0x%x" " EVT Head:0x%x Tail:0x%x]\n", softc->event_intr_cnt++, ctrl->status, ctrl->evt_head, ctrl->evt_tail); printf(" [CMD Total 0x%lx] Tail:0x%x, Head:0x%x.\n", softc->total_cmd, ctrl->cmd_tail, ctrl->cmd_head); amdvi_print_events(softc); ctrl->status &= AMDVI_STATUS_EV_OF | AMDVI_STATUS_EV_INTR; } static void amdvi_free_evt_intr_res(device_t dev) { struct amdvi_softc *softc; softc = device_get_softc(dev); if (softc->event_tag != NULL) { bus_teardown_intr(dev, softc->event_res, softc->event_tag); } if (softc->event_res != NULL) { bus_release_resource(dev, SYS_RES_IRQ, softc->event_rid, softc->event_res); } bus_delete_resource(dev, SYS_RES_IRQ, softc->event_rid); PCIB_RELEASE_MSI(device_get_parent(device_get_parent(dev)), dev, 1, &softc->event_irq); } static bool amdvi_alloc_intr_resources(struct amdvi_softc *softc) { struct amdvi_ctrl *ctrl; device_t dev, pcib; device_t mmio_dev; uint64_t msi_addr; uint32_t msi_data; int err; dev = softc->dev; pcib = device_get_parent(device_get_parent(dev)); mmio_dev = pci_find_bsf(PCI_RID2BUS(softc->pci_rid), PCI_RID2SLOT(softc->pci_rid), PCI_RID2FUNC(softc->pci_rid)); if (device_is_attached(mmio_dev)) { device_printf(dev, "warning: IOMMU device is claimed by another driver %s\n", device_get_driver(mmio_dev)->name); } softc->event_irq = -1; softc->event_rid = 0; /* * Section 3.7.1 of IOMMU rev 2.0. With MSI, there is only one * interrupt. XXX: Enable MSI/X support. */ err = PCIB_ALLOC_MSI(pcib, dev, 1, 1, &softc->event_irq); if (err) { device_printf(dev, "Couldn't find event MSI IRQ resource.\n"); return (ENOENT); } err = bus_set_resource(dev, SYS_RES_IRQ, softc->event_rid, softc->event_irq, 1); if (err) { device_printf(dev, "Couldn't set event MSI resource.\n"); return (ENXIO); } softc->event_res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &softc->event_rid, RF_ACTIVE); if (!softc->event_res) { device_printf(dev, "Unable to allocate event INTR resource.\n"); return (ENOMEM); } if (bus_setup_intr(dev, softc->event_res, INTR_TYPE_MISC | INTR_MPSAFE, NULL, amdvi_event_intr, softc, &softc->event_tag)) { device_printf(dev, "Fail to setup event intr\n"); bus_release_resource(softc->dev, SYS_RES_IRQ, softc->event_rid, softc->event_res); softc->event_res = NULL; return (ENXIO); } bus_describe_intr(dev, softc->event_res, softc->event_tag, "fault"); err = PCIB_MAP_MSI(pcib, dev, softc->event_irq, &msi_addr, &msi_data); if (err) { device_printf(dev, "Event interrupt config failed, err=%d.\n", err); amdvi_free_evt_intr_res(softc->dev); return (err); } /* Clear interrupt status bits. */ ctrl = softc->ctrl; ctrl->status &= AMDVI_STATUS_EV_OF | AMDVI_STATUS_EV_INTR; /* Now enable MSI interrupt. */ pci_enable_msi(mmio_dev, msi_addr, msi_data); return (0); } static void amdvi_print_dev_cap(struct amdvi_softc *softc) { struct ivhd_dev_cfg *cfg; int i; cfg = softc->dev_cfg; for (i = 0; i < softc->dev_cfg_cnt; i++) { device_printf(softc->dev, "device [0x%x - 0x%x]" "config:%b%s\n", cfg->start_id, cfg->end_id, cfg->data, "\020\001INIT\002ExtInt\003NMI" "\007LINT0\008LINT1", cfg->enable_ats ? "ATS enabled" : ""); cfg++; } } static int amdvi_handle_sysctl(SYSCTL_HANDLER_ARGS) { struct amdvi_softc *softc; int result, type, error = 0; softc = (struct amdvi_softc *)arg1; type = arg2; switch (type) { case 0: result = softc->ctrl->cmd_head; error = sysctl_handle_int(oidp, &result, 0, req); break; case 1: result = softc->ctrl->cmd_tail; error = sysctl_handle_int(oidp, &result, 0, req); break; case 2: result = softc->ctrl->evt_head; error = sysctl_handle_int(oidp, &result, 0, req); break; case 3: result = softc->ctrl->evt_tail; error = sysctl_handle_int(oidp, &result, 0, req); break; default: device_printf(softc->dev, "Unknown sysctl:%d\n", type); } return (error); } static void amdvi_add_sysctl(struct amdvi_softc *softc) { struct sysctl_oid_list *child; struct sysctl_ctx_list *ctx; device_t dev; dev = softc->dev; ctx = device_get_sysctl_ctx(dev); child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev)); SYSCTL_ADD_ULONG(ctx, child, OID_AUTO, "event_intr_count", CTLFLAG_RD, &softc->event_intr_cnt, "Event interrupt count"); SYSCTL_ADD_ULONG(ctx, child, OID_AUTO, "command_count", CTLFLAG_RD, &softc->total_cmd, "Command submitted count"); SYSCTL_ADD_U16(ctx, child, OID_AUTO, "pci_rid", CTLFLAG_RD, &softc->pci_rid, 0, "IOMMU RID"); SYSCTL_ADD_U16(ctx, child, OID_AUTO, "start_dev_rid", CTLFLAG_RD, &softc->start_dev_rid, 0, "Start of device under this IOMMU"); SYSCTL_ADD_U16(ctx, child, OID_AUTO, "end_dev_rid", CTLFLAG_RD, &softc->end_dev_rid, 0, "End of device under this IOMMU"); SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "command_head", - CTLTYPE_UINT | CTLFLAG_RD, softc, 0, + CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, softc, 0, amdvi_handle_sysctl, "IU", "Command head"); SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "command_tail", - CTLTYPE_UINT | CTLFLAG_RD, softc, 1, + CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, softc, 1, amdvi_handle_sysctl, "IU", "Command tail"); SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "event_head", - CTLTYPE_UINT | CTLFLAG_RD, softc, 2, + CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, softc, 2, amdvi_handle_sysctl, "IU", "Command head"); SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "event_tail", - CTLTYPE_UINT | CTLFLAG_RD, softc, 3, + CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, softc, 3, amdvi_handle_sysctl, "IU", "Command tail"); } int amdvi_setup_hw(struct amdvi_softc *softc) { device_t dev; int status; dev = softc->dev; amdvi_hw_enable_iotlb(softc); amdvi_print_dev_cap(softc); if ((status = amdvi_print_pci_cap(dev)) != 0) { device_printf(dev, "PCI capability.\n"); return (status); } if ((status = amdvi_init_cmd(softc)) != 0) { device_printf(dev, "Couldn't configure command buffer.\n"); return (status); } if ((status = amdvi_init_event(softc)) != 0) { device_printf(dev, "Couldn't configure event buffer.\n"); return (status); } if ((status = amdvi_init_dte(softc)) != 0) { device_printf(dev, "Couldn't configure device table.\n"); return (status); } if ((status = amdvi_alloc_intr_resources(softc)) != 0) { return (status); } amdvi_add_sysctl(softc); return (0); } int amdvi_teardown_hw(struct amdvi_softc *softc) { device_t dev; dev = softc->dev; /* * Called after disable, h/w is stopped by now, free all the resources. */ amdvi_free_evt_intr_res(dev); if (softc->cmd) free(softc->cmd, M_AMDVI); if (softc->event) free(softc->event, M_AMDVI); return (0); } /*********** bhyve interfaces *********************/ static int amdvi_init(void) { if (!ivhd_count) { return (EIO); } if (!amdvi_enable_user && ivhd_count) { printf("bhyve: Found %d AMD-Vi/IOMMU device(s), " "use hw.vmm.amdvi.enable=1 to enable pass-through.\n", ivhd_count); return (EINVAL); } return (0); } static void amdvi_cleanup(void) { /* Nothing. */ } static uint16_t amdvi_domainId(void) { /* * If we hit maximum domain limit, rollover leaving host * domain(0). * XXX: make sure that this domain is not used. */ if (amdvi_dom_id == AMDVI_MAX_DOMAIN) amdvi_dom_id = 1; return ((uint16_t)amdvi_dom_id++); } static void amdvi_do_inv_domain(uint16_t domain_id, bool create) { struct amdvi_softc *softc; int i; for (i = 0; i < ivhd_count; i++) { softc = device_get_softc(ivhd_devs[i]); KASSERT(softc, ("softc is NULL")); /* * If not present pages are cached, invalidate page after * creating domain. */ #if 0 if (create && ((softc->pci_cap & AMDVI_PCI_CAP_NPCACHE) == 0)) continue; #endif amdvi_inv_domain(softc, domain_id); amdvi_wait(softc); } } static void * amdvi_create_domain(vm_paddr_t maxaddr) { struct amdvi_domain *dom; dom = malloc(sizeof(struct amdvi_domain), M_AMDVI, M_ZERO | M_WAITOK); dom->id = amdvi_domainId(); //dom->maxaddr = maxaddr; #ifdef AMDVI_DEBUG_CMD printf("Created domain #%d\n", dom->id); #endif /* * Host domain(#0) don't create translation table. */ if (dom->id || amdvi_host_ptp) dom->ptp = malloc(PAGE_SIZE, M_AMDVI, M_WAITOK | M_ZERO); dom->ptp_level = amdvi_ptp_level; amdvi_do_inv_domain(dom->id, true); SLIST_INSERT_HEAD(&dom_head, dom, next); return (dom); } static void amdvi_free_ptp(uint64_t *ptp, int level) { int i; if (level < 1) return; for (i = 0; i < NPTEPG ; i++) { if ((ptp[i] & AMDVI_PT_PRESENT) == 0) continue; /* XXX: Add super-page or PTE mapping > 4KB. */ #ifdef notyet /* Super-page mapping. */ if (AMDVI_PD_SUPER(ptp[i])) continue; #endif amdvi_free_ptp((uint64_t *)PHYS_TO_DMAP(ptp[i] & AMDVI_PT_MASK), level - 1); } free(ptp, M_AMDVI); } static void amdvi_destroy_domain(void *arg) { struct amdvi_domain *domain; domain = (struct amdvi_domain *)arg; KASSERT(domain, ("domain is NULL")); #ifdef AMDVI_DEBUG_CMD printf("Destroying domain %d\n", domain->id); #endif if (domain->ptp) amdvi_free_ptp(domain->ptp, domain->ptp_level); amdvi_do_inv_domain(domain->id, false); SLIST_REMOVE(&dom_head, domain, amdvi_domain, next); free(domain, M_AMDVI); } static uint64_t amdvi_set_pt(uint64_t *pt, int level, vm_paddr_t gpa, vm_paddr_t hpa, uint64_t pg_size, bool create) { uint64_t *page, pa; int shift, index; const int PT_SHIFT = 9; const int PT_INDEX_MASK = (1 << PT_SHIFT) - 1; /* Based on PT_SHIFT */ if (!pg_size) return (0); if (hpa & (pg_size - 1)) { printf("HPA is not size aligned.\n"); return (0); } if (gpa & (pg_size - 1)) { printf("HPA is not size aligned.\n"); return (0); } shift = PML4SHIFT; while ((shift > PAGE_SHIFT) && (pg_size < (1UL << shift))) { index = (gpa >> shift) & PT_INDEX_MASK; if ((pt[index] == 0) && create) { page = malloc(PAGE_SIZE, M_AMDVI, M_WAITOK | M_ZERO); pa = vtophys(page); pt[index] = pa | AMDVI_PT_PRESENT | AMDVI_PT_RW | ((level - 1) << AMDVI_PD_LEVEL_SHIFT); } #ifdef AMDVI_DEBUG_PTE if ((gpa % 0x1000000) == 0) printf("[level%d, shift = %d]PTE:0x%lx\n", level, shift, pt[index]); #endif #define PTE2PA(x) ((uint64_t)(x) & AMDVI_PT_MASK) pa = PTE2PA(pt[index]); pt = (uint64_t *)PHYS_TO_DMAP(pa); shift -= PT_SHIFT; level--; } /* Leaf entry. */ index = (gpa >> shift) & PT_INDEX_MASK; if (create) { pt[index] = hpa | AMDVI_PT_RW | AMDVI_PT_PRESENT; } else pt[index] = 0; #ifdef AMDVI_DEBUG_PTE if ((gpa % 0x1000000) == 0) printf("[Last level%d, shift = %d]PTE:0x%lx\n", level, shift, pt[index]); #endif return (1ULL << shift); } static uint64_t amdvi_update_mapping(struct amdvi_domain *domain, vm_paddr_t gpa, vm_paddr_t hpa, uint64_t size, bool create) { uint64_t mapped, *ptp, len; int level; KASSERT(domain, ("domain is NULL")); level = domain->ptp_level; KASSERT(level, ("Page table level is 0")); ptp = domain->ptp; KASSERT(ptp, ("PTP is NULL")); mapped = 0; while (mapped < size) { len = amdvi_set_pt(ptp, level, gpa + mapped, hpa + mapped, PAGE_SIZE, create); if (!len) { printf("Error: Couldn't map HPA:0x%lx GPA:0x%lx\n", hpa, gpa); return (0); } mapped += len; } return (mapped); } static uint64_t amdvi_create_mapping(void *arg, vm_paddr_t gpa, vm_paddr_t hpa, uint64_t len) { struct amdvi_domain *domain; domain = (struct amdvi_domain *)arg; if (domain->id && !domain->ptp) { printf("ptp is NULL"); return (-1); } /* * If host domain is created w/o page table, skip IOMMU page * table set-up. */ if (domain->ptp) return (amdvi_update_mapping(domain, gpa, hpa, len, true)); else return (len); } static uint64_t amdvi_destroy_mapping(void *arg, vm_paddr_t gpa, uint64_t len) { struct amdvi_domain *domain; domain = (struct amdvi_domain *)arg; /* * If host domain is created w/o page table, skip IOMMU page * table set-up. */ if (domain->ptp) return (amdvi_update_mapping(domain, gpa, 0, len, false)); return (len); } static struct amdvi_softc * amdvi_find_iommu(uint16_t devid) { struct amdvi_softc *softc; int i; for (i = 0; i < ivhd_count; i++) { softc = device_get_softc(ivhd_devs[i]); if ((devid >= softc->start_dev_rid) && (devid <= softc->end_dev_rid)) return (softc); } /* * XXX: BIOS bug, device not in IVRS table, assume its from first IOMMU. */ printf("BIOS bug device(%d.%d.%d) doesn't have IVHD entry.\n", RID2PCI_STR(devid)); return (device_get_softc(ivhd_devs[0])); } /* * Set-up device table entry. * IOMMU spec Rev 2.0, section 3.2.2.2, some of the fields must * be set concurrently, e.g. read and write bits. */ static void amdvi_set_dte(struct amdvi_domain *domain, uint16_t devid, bool enable) { struct amdvi_softc *softc; struct amdvi_dte* temp; KASSERT(domain, ("domain is NULL for pci_rid:0x%x\n", devid)); softc = amdvi_find_iommu(devid); KASSERT(softc, ("softc is NULL for pci_rid:0x%x\n", devid)); temp = &amdvi_dte[devid]; #ifdef AMDVI_ATS_ENABLE /* If IOMMU and device support IOTLB, enable it. */ if (amdvi_dev_support_iotlb(softc, devid) && softc->iotlb) temp->iotlb_enable = 1; #endif /* Avoid duplicate I/O faults. */ temp->sup_second_io_fault = 1; temp->sup_all_io_fault = amdvi_disable_io_fault; temp->dt_valid = 1; temp->domain_id = domain->id; if (enable) { if (domain->ptp) { temp->pt_base = vtophys(domain->ptp) >> 12; temp->pt_level = amdvi_ptp_level; } /* * XXX: Page table valid[TV] bit must be set even if host domain * page tables are not enabled. */ temp->pt_valid = 1; temp->read_allow = 1; temp->write_allow = 1; } } static void amdvi_inv_device(uint16_t devid) { struct amdvi_softc *softc; softc = amdvi_find_iommu(devid); KASSERT(softc, ("softc is NULL")); amdvi_cmd_inv_dte(softc, devid); #ifdef AMDVI_ATS_ENABLE if (amdvi_dev_support_iotlb(softc, devid)) amdvi_cmd_inv_iotlb(softc, devid); #endif amdvi_wait(softc); } static void amdvi_add_device(void *arg, uint16_t devid) { struct amdvi_domain *domain; domain = (struct amdvi_domain *)arg; KASSERT(domain != NULL, ("domain is NULL")); #ifdef AMDVI_DEBUG_CMD printf("Assigning device(%d.%d.%d) to domain:%d\n", RID2PCI_STR(devid), domain->id); #endif amdvi_set_dte(domain, devid, true); amdvi_inv_device(devid); } static void amdvi_remove_device(void *arg, uint16_t devid) { struct amdvi_domain *domain; domain = (struct amdvi_domain *)arg; #ifdef AMDVI_DEBUG_CMD printf("Remove device(0x%x) from domain:%d\n", devid, domain->id); #endif amdvi_set_dte(domain, devid, false); amdvi_inv_device(devid); } static void amdvi_enable(void) { struct amdvi_ctrl *ctrl; struct amdvi_softc *softc; uint64_t val; int i; for (i = 0; i < ivhd_count; i++) { softc = device_get_softc(ivhd_devs[i]); KASSERT(softc, ("softc is NULL\n")); ctrl = softc->ctrl; KASSERT(ctrl, ("ctrl is NULL\n")); val = ( AMDVI_CTRL_EN | AMDVI_CTRL_CMD | AMDVI_CTRL_ELOG | AMDVI_CTRL_ELOGINT | AMDVI_CTRL_INV_TO_1S); if (softc->ivhd_flag & IVHD_FLAG_COH) val |= AMDVI_CTRL_COH; if (softc->ivhd_flag & IVHD_FLAG_HTT) val |= AMDVI_CTRL_HTT; if (softc->ivhd_flag & IVHD_FLAG_RPPW) val |= AMDVI_CTRL_RPPW; if (softc->ivhd_flag & IVHD_FLAG_PPW) val |= AMDVI_CTRL_PPW; if (softc->ivhd_flag & IVHD_FLAG_ISOC) val |= AMDVI_CTRL_ISOC; ctrl->control = val; } } static void amdvi_disable(void) { struct amdvi_ctrl *ctrl; struct amdvi_softc *softc; int i; for (i = 0; i < ivhd_count; i++) { softc = device_get_softc(ivhd_devs[i]); KASSERT(softc, ("softc is NULL\n")); ctrl = softc->ctrl; KASSERT(ctrl, ("ctrl is NULL\n")); ctrl->control = 0; } } static void amdvi_inv_tlb(void *arg) { struct amdvi_domain *domain; domain = (struct amdvi_domain *)arg; KASSERT(domain, ("domain is NULL")); amdvi_do_inv_domain(domain->id, false); } struct iommu_ops iommu_ops_amd = { amdvi_init, amdvi_cleanup, amdvi_enable, amdvi_disable, amdvi_create_domain, amdvi_destroy_domain, amdvi_create_mapping, amdvi_destroy_mapping, amdvi_add_device, amdvi_remove_device, amdvi_inv_tlb }; Index: head/sys/amd64/vmm/amd/npt.c =================================================================== --- head/sys/amd64/vmm/amd/npt.c (revision 357973) +++ head/sys/amd64/vmm/amd/npt.c (revision 357974) @@ -1,87 +1,88 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2013 Anish Gupta (akgupt3@gmail.com) * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include "npt.h" SYSCTL_DECL(_hw_vmm); -SYSCTL_NODE(_hw_vmm, OID_AUTO, npt, CTLFLAG_RW, NULL, NULL); +SYSCTL_NODE(_hw_vmm, OID_AUTO, npt, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, + NULL); static int npt_flags; SYSCTL_INT(_hw_vmm_npt, OID_AUTO, pmap_flags, CTLFLAG_RD, &npt_flags, 0, NULL); #define NPT_IPIMASK 0xFF /* * AMD nested page table init. */ int svm_npt_init(int ipinum) { int enable_superpage = 1; npt_flags = ipinum & NPT_IPIMASK; TUNABLE_INT_FETCH("hw.vmm.npt.enable_superpage", &enable_superpage); if (enable_superpage) npt_flags |= PMAP_PDE_SUPERPAGE; return (0); } static int npt_pinit(pmap_t pmap) { return (pmap_pinit_type(pmap, PT_RVI, npt_flags)); } struct vmspace * svm_npt_alloc(vm_offset_t min, vm_offset_t max) { return (vmspace_alloc(min, max, npt_pinit)); } void svm_npt_free(struct vmspace *vmspace) { vmspace_free(vmspace); } Index: head/sys/amd64/vmm/amd/svm.c =================================================================== --- head/sys/amd64/vmm/amd/svm.c (revision 357973) +++ head/sys/amd64/vmm/amd/svm.c (revision 357974) @@ -1,2304 +1,2305 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2013, Anish Gupta (akgupt3@gmail.com) * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "vmm_lapic.h" #include "vmm_stat.h" #include "vmm_ktr.h" #include "vmm_ioport.h" #include "vatpic.h" #include "vlapic.h" #include "vlapic_priv.h" #include "x86.h" #include "vmcb.h" #include "svm.h" #include "svm_softc.h" #include "svm_msr.h" #include "npt.h" SYSCTL_DECL(_hw_vmm); -SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW, NULL, NULL); +SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, + NULL); /* * SVM CPUID function 0x8000_000A, edx bit decoding. */ #define AMD_CPUID_SVM_NP BIT(0) /* Nested paging or RVI */ #define AMD_CPUID_SVM_LBR BIT(1) /* Last branch virtualization */ #define AMD_CPUID_SVM_SVML BIT(2) /* SVM lock */ #define AMD_CPUID_SVM_NRIP_SAVE BIT(3) /* Next RIP is saved */ #define AMD_CPUID_SVM_TSC_RATE BIT(4) /* TSC rate control. */ #define AMD_CPUID_SVM_VMCB_CLEAN BIT(5) /* VMCB state caching */ #define AMD_CPUID_SVM_FLUSH_BY_ASID BIT(6) /* Flush by ASID */ #define AMD_CPUID_SVM_DECODE_ASSIST BIT(7) /* Decode assist */ #define AMD_CPUID_SVM_PAUSE_INC BIT(10) /* Pause intercept filter. */ #define AMD_CPUID_SVM_PAUSE_FTH BIT(12) /* Pause filter threshold */ #define AMD_CPUID_SVM_AVIC BIT(13) /* AVIC present */ #define VMCB_CACHE_DEFAULT (VMCB_CACHE_ASID | \ VMCB_CACHE_IOPM | \ VMCB_CACHE_I | \ VMCB_CACHE_TPR | \ VMCB_CACHE_CR2 | \ VMCB_CACHE_CR | \ VMCB_CACHE_DR | \ VMCB_CACHE_DT | \ VMCB_CACHE_SEG | \ VMCB_CACHE_NP) static uint32_t vmcb_clean = VMCB_CACHE_DEFAULT; SYSCTL_INT(_hw_vmm_svm, OID_AUTO, vmcb_clean, CTLFLAG_RDTUN, &vmcb_clean, 0, NULL); static MALLOC_DEFINE(M_SVM, "svm", "svm"); static MALLOC_DEFINE(M_SVM_VLAPIC, "svm-vlapic", "svm-vlapic"); static uint32_t svm_feature = ~0U; /* AMD SVM features. */ SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RDTUN, &svm_feature, 0, "SVM features advertised by CPUID.8000000AH:EDX"); static int disable_npf_assist; SYSCTL_INT(_hw_vmm_svm, OID_AUTO, disable_npf_assist, CTLFLAG_RWTUN, &disable_npf_assist, 0, NULL); /* Maximum ASIDs supported by the processor */ static uint32_t nasid; SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, num_asids, CTLFLAG_RDTUN, &nasid, 0, "Number of ASIDs supported by this processor"); /* Current ASID generation for each host cpu */ static struct asid asid[MAXCPU]; /* * SVM host state saved area of size 4KB for each core. */ static uint8_t hsave[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE); static VMM_STAT_AMD(VCPU_EXITINTINFO, "VM exits during event delivery"); static VMM_STAT_AMD(VCPU_INTINFO_INJECTED, "Events pending at VM entry"); static VMM_STAT_AMD(VMEXIT_VINTR, "VM exits due to interrupt window"); static int svm_setreg(void *arg, int vcpu, int ident, uint64_t val); static __inline int flush_by_asid(void) { return (svm_feature & AMD_CPUID_SVM_FLUSH_BY_ASID); } static __inline int decode_assist(void) { return (svm_feature & AMD_CPUID_SVM_DECODE_ASSIST); } static void svm_disable(void *arg __unused) { uint64_t efer; efer = rdmsr(MSR_EFER); efer &= ~EFER_SVM; wrmsr(MSR_EFER, efer); } /* * Disable SVM on all CPUs. */ static int svm_cleanup(void) { smp_rendezvous(NULL, svm_disable, NULL, NULL); return (0); } /* * Verify that all the features required by bhyve are available. */ static int check_svm_features(void) { u_int regs[4]; /* CPUID Fn8000_000A is for SVM */ do_cpuid(0x8000000A, regs); svm_feature &= regs[3]; /* * The number of ASIDs can be configured to be less than what is * supported by the hardware but not more. */ if (nasid == 0 || nasid > regs[1]) nasid = regs[1]; KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid)); /* bhyve requires the Nested Paging feature */ if (!(svm_feature & AMD_CPUID_SVM_NP)) { printf("SVM: Nested Paging feature not available.\n"); return (ENXIO); } /* bhyve requires the NRIP Save feature */ if (!(svm_feature & AMD_CPUID_SVM_NRIP_SAVE)) { printf("SVM: NRIP Save feature not available.\n"); return (ENXIO); } return (0); } static void svm_enable(void *arg __unused) { uint64_t efer; efer = rdmsr(MSR_EFER); efer |= EFER_SVM; wrmsr(MSR_EFER, efer); wrmsr(MSR_VM_HSAVE_PA, vtophys(hsave[curcpu])); } /* * Return 1 if SVM is enabled on this processor and 0 otherwise. */ static int svm_available(void) { uint64_t msr; /* Section 15.4 Enabling SVM from APM2. */ if ((amd_feature2 & AMDID2_SVM) == 0) { printf("SVM: not available.\n"); return (0); } msr = rdmsr(MSR_VM_CR); if ((msr & VM_CR_SVMDIS) != 0) { printf("SVM: disabled by BIOS.\n"); return (0); } return (1); } static int svm_init(int ipinum) { int error, cpu; if (!svm_available()) return (ENXIO); error = check_svm_features(); if (error) return (error); vmcb_clean &= VMCB_CACHE_DEFAULT; for (cpu = 0; cpu < MAXCPU; cpu++) { /* * Initialize the host ASIDs to their "highest" valid values. * * The next ASID allocation will rollover both 'gen' and 'num' * and start off the sequence at {1,1}. */ asid[cpu].gen = ~0UL; asid[cpu].num = nasid - 1; } svm_msr_init(); svm_npt_init(ipinum); /* Enable SVM on all CPUs */ smp_rendezvous(NULL, svm_enable, NULL, NULL); return (0); } static void svm_restore(void) { svm_enable(NULL); } /* Pentium compatible MSRs */ #define MSR_PENTIUM_START 0 #define MSR_PENTIUM_END 0x1FFF /* AMD 6th generation and Intel compatible MSRs */ #define MSR_AMD6TH_START 0xC0000000UL #define MSR_AMD6TH_END 0xC0001FFFUL /* AMD 7th and 8th generation compatible MSRs */ #define MSR_AMD7TH_START 0xC0010000UL #define MSR_AMD7TH_END 0xC0011FFFUL /* * Get the index and bit position for a MSR in permission bitmap. * Two bits are used for each MSR: lower bit for read and higher bit for write. */ static int svm_msr_index(uint64_t msr, int *index, int *bit) { uint32_t base, off; *index = -1; *bit = (msr % 4) * 2; base = 0; if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) { *index = msr / 4; return (0); } base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1); if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) { off = (msr - MSR_AMD6TH_START); *index = (off + base) / 4; return (0); } base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1); if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) { off = (msr - MSR_AMD7TH_START); *index = (off + base) / 4; return (0); } return (EINVAL); } /* * Allow vcpu to read or write the 'msr' without trapping into the hypervisor. */ static void svm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write) { int index, bit, error; error = svm_msr_index(msr, &index, &bit); KASSERT(error == 0, ("%s: invalid msr %#lx", __func__, msr)); KASSERT(index >= 0 && index < SVM_MSR_BITMAP_SIZE, ("%s: invalid index %d for msr %#lx", __func__, index, msr)); KASSERT(bit >= 0 && bit <= 6, ("%s: invalid bit position %d " "msr %#lx", __func__, bit, msr)); if (read) perm_bitmap[index] &= ~(1UL << bit); if (write) perm_bitmap[index] &= ~(2UL << bit); } static void svm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr) { svm_msr_perm(perm_bitmap, msr, true, true); } static void svm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr) { svm_msr_perm(perm_bitmap, msr, true, false); } static __inline int svm_get_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask) { struct vmcb_ctrl *ctrl; KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx)); ctrl = svm_get_vmcb_ctrl(sc, vcpu); return (ctrl->intercept[idx] & bitmask ? 1 : 0); } static __inline void svm_set_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask, int enabled) { struct vmcb_ctrl *ctrl; uint32_t oldval; KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx)); ctrl = svm_get_vmcb_ctrl(sc, vcpu); oldval = ctrl->intercept[idx]; if (enabled) ctrl->intercept[idx] |= bitmask; else ctrl->intercept[idx] &= ~bitmask; if (ctrl->intercept[idx] != oldval) { svm_set_dirty(sc, vcpu, VMCB_CACHE_I); VCPU_CTR3(sc->vm, vcpu, "intercept[%d] modified " "from %#x to %#x", idx, oldval, ctrl->intercept[idx]); } } static __inline void svm_disable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask) { svm_set_intercept(sc, vcpu, off, bitmask, 0); } static __inline void svm_enable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask) { svm_set_intercept(sc, vcpu, off, bitmask, 1); } static void vmcb_init(struct svm_softc *sc, int vcpu, uint64_t iopm_base_pa, uint64_t msrpm_base_pa, uint64_t np_pml4) { struct vmcb_ctrl *ctrl; struct vmcb_state *state; uint32_t mask; int n; ctrl = svm_get_vmcb_ctrl(sc, vcpu); state = svm_get_vmcb_state(sc, vcpu); ctrl->iopm_base_pa = iopm_base_pa; ctrl->msrpm_base_pa = msrpm_base_pa; /* Enable nested paging */ ctrl->np_enable = 1; ctrl->n_cr3 = np_pml4; /* * Intercept accesses to the control registers that are not shadowed * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8. */ for (n = 0; n < 16; n++) { mask = (BIT(n) << 16) | BIT(n); if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8) svm_disable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask); else svm_enable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask); } /* * Intercept everything when tracing guest exceptions otherwise * just intercept machine check exception. */ if (vcpu_trace_exceptions(sc->vm, vcpu)) { for (n = 0; n < 32; n++) { /* * Skip unimplemented vectors in the exception bitmap. */ if (n == 2 || n == 9) { continue; } svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(n)); } } else { svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC)); } /* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */ svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_FERR_FREEZE); svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MONITOR); svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MWAIT); /* * From section "Canonicalization and Consistency Checks" in APMv2 * the VMRUN intercept bit must be set to pass the consistency check. */ svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN); /* * The ASID will be set to a non-zero value just before VMRUN. */ ctrl->asid = 0; /* * Section 15.21.1, Interrupt Masking in EFLAGS * Section 15.21.2, Virtualizing APIC.TPR * * This must be set for %rflag and %cr8 isolation of guest and host. */ ctrl->v_intr_masking = 1; /* Enable Last Branch Record aka LBR for debugging */ ctrl->lbr_virt_en = 1; state->dbgctl = BIT(0); /* EFER_SVM must always be set when the guest is executing */ state->efer = EFER_SVM; /* Set up the PAT to power-on state */ state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK) | PAT_VALUE(1, PAT_WRITE_THROUGH) | PAT_VALUE(2, PAT_UNCACHED) | PAT_VALUE(3, PAT_UNCACHEABLE) | PAT_VALUE(4, PAT_WRITE_BACK) | PAT_VALUE(5, PAT_WRITE_THROUGH) | PAT_VALUE(6, PAT_UNCACHED) | PAT_VALUE(7, PAT_UNCACHEABLE); /* Set up DR6/7 to power-on state */ state->dr6 = DBREG_DR6_RESERVED1; state->dr7 = DBREG_DR7_RESERVED1; } /* * Initialize a virtual machine. */ static void * svm_vminit(struct vm *vm, pmap_t pmap) { struct svm_softc *svm_sc; struct svm_vcpu *vcpu; vm_paddr_t msrpm_pa, iopm_pa, pml4_pa; int i; uint16_t maxcpus; svm_sc = malloc(sizeof (*svm_sc), M_SVM, M_WAITOK | M_ZERO); if (((uintptr_t)svm_sc & PAGE_MASK) != 0) panic("malloc of svm_softc not aligned on page boundary"); svm_sc->msr_bitmap = contigmalloc(SVM_MSR_BITMAP_SIZE, M_SVM, M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0); if (svm_sc->msr_bitmap == NULL) panic("contigmalloc of SVM MSR bitmap failed"); svm_sc->iopm_bitmap = contigmalloc(SVM_IO_BITMAP_SIZE, M_SVM, M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0); if (svm_sc->iopm_bitmap == NULL) panic("contigmalloc of SVM IO bitmap failed"); svm_sc->vm = vm; svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pml4); /* * Intercept read and write accesses to all MSRs. */ memset(svm_sc->msr_bitmap, 0xFF, SVM_MSR_BITMAP_SIZE); /* * Access to the following MSRs is redirected to the VMCB when the * guest is executing. Therefore it is safe to allow the guest to * read/write these MSRs directly without hypervisor involvement. */ svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR); svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT); svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC); /* * Intercept writes to make sure that the EFER_SVM bit is not cleared. */ svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER); /* Intercept access to all I/O ports. */ memset(svm_sc->iopm_bitmap, 0xFF, SVM_IO_BITMAP_SIZE); iopm_pa = vtophys(svm_sc->iopm_bitmap); msrpm_pa = vtophys(svm_sc->msr_bitmap); pml4_pa = svm_sc->nptp; maxcpus = vm_get_maxcpus(svm_sc->vm); for (i = 0; i < maxcpus; i++) { vcpu = svm_get_vcpu(svm_sc, i); vcpu->nextrip = ~0; vcpu->lastcpu = NOCPU; vcpu->vmcb_pa = vtophys(&vcpu->vmcb); vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa); svm_msr_guest_init(svm_sc, i); } return (svm_sc); } /* * Collateral for a generic SVM VM-exit. */ static void vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2) { vme->exitcode = VM_EXITCODE_SVM; vme->u.svm.exitcode = code; vme->u.svm.exitinfo1 = info1; vme->u.svm.exitinfo2 = info2; } static int svm_cpl(struct vmcb_state *state) { /* * From APMv2: * "Retrieve the CPL from the CPL field in the VMCB, not * from any segment DPL" */ return (state->cpl); } static enum vm_cpu_mode svm_vcpu_mode(struct vmcb *vmcb) { struct vmcb_segment seg; struct vmcb_state *state; int error; state = &vmcb->state; if (state->efer & EFER_LMA) { error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg); KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__, error)); /* * Section 4.8.1 for APM2, check if Code Segment has * Long attribute set in descriptor. */ if (seg.attrib & VMCB_CS_ATTRIB_L) return (CPU_MODE_64BIT); else return (CPU_MODE_COMPATIBILITY); } else if (state->cr0 & CR0_PE) { return (CPU_MODE_PROTECTED); } else { return (CPU_MODE_REAL); } } static enum vm_paging_mode svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer) { if ((cr0 & CR0_PG) == 0) return (PAGING_MODE_FLAT); if ((cr4 & CR4_PAE) == 0) return (PAGING_MODE_32); if (efer & EFER_LME) return (PAGING_MODE_64); else return (PAGING_MODE_PAE); } /* * ins/outs utility routines */ static uint64_t svm_inout_str_index(struct svm_regctx *regs, int in) { uint64_t val; val = in ? regs->sctx_rdi : regs->sctx_rsi; return (val); } static uint64_t svm_inout_str_count(struct svm_regctx *regs, int rep) { uint64_t val; val = rep ? regs->sctx_rcx : 1; return (val); } static void svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1, int in, struct vm_inout_str *vis) { int error, s; if (in) { vis->seg_name = VM_REG_GUEST_ES; } else { /* The segment field has standard encoding */ s = (info1 >> 10) & 0x7; vis->seg_name = vm_segment_name(s); } error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc); KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error)); } static int svm_inout_str_addrsize(uint64_t info1) { uint32_t size; size = (info1 >> 7) & 0x7; switch (size) { case 1: return (2); /* 16 bit */ case 2: return (4); /* 32 bit */ case 4: return (8); /* 64 bit */ default: panic("%s: invalid size encoding %d", __func__, size); } } static void svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging) { struct vmcb_state *state; state = &vmcb->state; paging->cr3 = state->cr3; paging->cpl = svm_cpl(state); paging->cpu_mode = svm_vcpu_mode(vmcb); paging->paging_mode = svm_paging_mode(state->cr0, state->cr4, state->efer); } #define UNHANDLED 0 /* * Handle guest I/O intercept. */ static int svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit) { struct vmcb_ctrl *ctrl; struct vmcb_state *state; struct svm_regctx *regs; struct vm_inout_str *vis; uint64_t info1; int inout_string; state = svm_get_vmcb_state(svm_sc, vcpu); ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu); regs = svm_get_guest_regctx(svm_sc, vcpu); info1 = ctrl->exitinfo1; inout_string = info1 & BIT(2) ? 1 : 0; /* * The effective segment number in EXITINFO1[12:10] is populated * only if the processor has the DecodeAssist capability. * * XXX this is not specified explicitly in APMv2 but can be verified * empirically. */ if (inout_string && !decode_assist()) return (UNHANDLED); vmexit->exitcode = VM_EXITCODE_INOUT; vmexit->u.inout.in = (info1 & BIT(0)) ? 1 : 0; vmexit->u.inout.string = inout_string; vmexit->u.inout.rep = (info1 & BIT(3)) ? 1 : 0; vmexit->u.inout.bytes = (info1 >> 4) & 0x7; vmexit->u.inout.port = (uint16_t)(info1 >> 16); vmexit->u.inout.eax = (uint32_t)(state->rax); if (inout_string) { vmexit->exitcode = VM_EXITCODE_INOUT_STR; vis = &vmexit->u.inout_str; svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging); vis->rflags = state->rflags; vis->cr0 = state->cr0; vis->index = svm_inout_str_index(regs, vmexit->u.inout.in); vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep); vis->addrsize = svm_inout_str_addrsize(info1); svm_inout_str_seginfo(svm_sc, vcpu, info1, vmexit->u.inout.in, vis); } return (UNHANDLED); } static int npf_fault_type(uint64_t exitinfo1) { if (exitinfo1 & VMCB_NPF_INFO1_W) return (VM_PROT_WRITE); else if (exitinfo1 & VMCB_NPF_INFO1_ID) return (VM_PROT_EXECUTE); else return (VM_PROT_READ); } static bool svm_npf_emul_fault(uint64_t exitinfo1) { if (exitinfo1 & VMCB_NPF_INFO1_ID) { return (false); } if (exitinfo1 & VMCB_NPF_INFO1_GPT) { return (false); } if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) { return (false); } return (true); } static void svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit) { struct vm_guest_paging *paging; struct vmcb_segment seg; struct vmcb_ctrl *ctrl; char *inst_bytes; int error, inst_len; ctrl = &vmcb->ctrl; paging = &vmexit->u.inst_emul.paging; vmexit->exitcode = VM_EXITCODE_INST_EMUL; vmexit->u.inst_emul.gpa = gpa; vmexit->u.inst_emul.gla = VIE_INVALID_GLA; svm_paging_info(vmcb, paging); error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg); KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error)); switch(paging->cpu_mode) { case CPU_MODE_REAL: vmexit->u.inst_emul.cs_base = seg.base; vmexit->u.inst_emul.cs_d = 0; break; case CPU_MODE_PROTECTED: case CPU_MODE_COMPATIBILITY: vmexit->u.inst_emul.cs_base = seg.base; /* * Section 4.8.1 of APM2, Default Operand Size or D bit. */ vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ? 1 : 0; break; default: vmexit->u.inst_emul.cs_base = 0; vmexit->u.inst_emul.cs_d = 0; break; } /* * Copy the instruction bytes into 'vie' if available. */ if (decode_assist() && !disable_npf_assist) { inst_len = ctrl->inst_len; inst_bytes = ctrl->inst_bytes; } else { inst_len = 0; inst_bytes = NULL; } vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len); } #ifdef KTR static const char * intrtype_to_str(int intr_type) { switch (intr_type) { case VMCB_EVENTINJ_TYPE_INTR: return ("hwintr"); case VMCB_EVENTINJ_TYPE_NMI: return ("nmi"); case VMCB_EVENTINJ_TYPE_INTn: return ("swintr"); case VMCB_EVENTINJ_TYPE_EXCEPTION: return ("exception"); default: panic("%s: unknown intr_type %d", __func__, intr_type); } } #endif /* * Inject an event to vcpu as described in section 15.20, "Event injection". */ static void svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector, uint32_t error, bool ec_valid) { struct vmcb_ctrl *ctrl; ctrl = svm_get_vmcb_ctrl(sc, vcpu); KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event already pending %#lx", __func__, ctrl->eventinj)); KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d", __func__, vector)); switch (intr_type) { case VMCB_EVENTINJ_TYPE_INTR: case VMCB_EVENTINJ_TYPE_NMI: case VMCB_EVENTINJ_TYPE_INTn: break; case VMCB_EVENTINJ_TYPE_EXCEPTION: if (vector >= 0 && vector <= 31 && vector != 2) break; /* FALLTHROUGH */ default: panic("%s: invalid intr_type/vector: %d/%d", __func__, intr_type, vector); } ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID; if (ec_valid) { ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID; ctrl->eventinj |= (uint64_t)error << 32; VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x", intrtype_to_str(intr_type), vector, error); } else { VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d", intrtype_to_str(intr_type), vector); } } static void svm_update_virqinfo(struct svm_softc *sc, int vcpu) { struct vm *vm; struct vlapic *vlapic; struct vmcb_ctrl *ctrl; vm = sc->vm; vlapic = vm_lapic(vm, vcpu); ctrl = svm_get_vmcb_ctrl(sc, vcpu); /* Update %cr8 in the emulated vlapic */ vlapic_set_cr8(vlapic, ctrl->v_tpr); /* Virtual interrupt injection is not used. */ KASSERT(ctrl->v_intr_vector == 0, ("%s: invalid " "v_intr_vector %d", __func__, ctrl->v_intr_vector)); } static void svm_save_intinfo(struct svm_softc *svm_sc, int vcpu) { struct vmcb_ctrl *ctrl; uint64_t intinfo; ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu); intinfo = ctrl->exitintinfo; if (!VMCB_EXITINTINFO_VALID(intinfo)) return; /* * From APMv2, Section "Intercepts during IDT interrupt delivery" * * If a #VMEXIT happened during event delivery then record the event * that was being delivered. */ VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n", intinfo, VMCB_EXITINTINFO_VECTOR(intinfo)); vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1); vm_exit_intinfo(svm_sc->vm, vcpu, intinfo); } #ifdef INVARIANTS static __inline int vintr_intercept_enabled(struct svm_softc *sc, int vcpu) { return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR)); } #endif static __inline void enable_intr_window_exiting(struct svm_softc *sc, int vcpu) { struct vmcb_ctrl *ctrl; ctrl = svm_get_vmcb_ctrl(sc, vcpu); if (ctrl->v_irq && ctrl->v_intr_vector == 0) { KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__)); KASSERT(vintr_intercept_enabled(sc, vcpu), ("%s: vintr intercept should be enabled", __func__)); return; } VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting"); ctrl->v_irq = 1; ctrl->v_ign_tpr = 1; ctrl->v_intr_vector = 0; svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR); } static __inline void disable_intr_window_exiting(struct svm_softc *sc, int vcpu) { struct vmcb_ctrl *ctrl; ctrl = svm_get_vmcb_ctrl(sc, vcpu); if (!ctrl->v_irq && ctrl->v_intr_vector == 0) { KASSERT(!vintr_intercept_enabled(sc, vcpu), ("%s: vintr intercept should be disabled", __func__)); return; } VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting"); ctrl->v_irq = 0; ctrl->v_intr_vector = 0; svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR); } static int svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val) { struct vmcb_ctrl *ctrl; int oldval, newval; ctrl = svm_get_vmcb_ctrl(sc, vcpu); oldval = ctrl->intr_shadow; newval = val ? 1 : 0; if (newval != oldval) { ctrl->intr_shadow = newval; VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval); } return (0); } static int svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val) { struct vmcb_ctrl *ctrl; ctrl = svm_get_vmcb_ctrl(sc, vcpu); *val = ctrl->intr_shadow; return (0); } /* * Once an NMI is injected it blocks delivery of further NMIs until the handler * executes an IRET. The IRET intercept is enabled when an NMI is injected to * to track when the vcpu is done handling the NMI. */ static int nmi_blocked(struct svm_softc *sc, int vcpu) { int blocked; blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET); return (blocked); } static void enable_nmi_blocking(struct svm_softc *sc, int vcpu) { KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked")); VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled"); svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET); } static void clear_nmi_blocking(struct svm_softc *sc, int vcpu) { int error; KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked")); VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared"); /* * When the IRET intercept is cleared the vcpu will attempt to execute * the "iret" when it runs next. However, it is possible to inject * another NMI into the vcpu before the "iret" has actually executed. * * For e.g. if the "iret" encounters a #NPF when accessing the stack * it will trap back into the hypervisor. If an NMI is pending for * the vcpu it will be injected into the guest. * * XXX this needs to be fixed */ svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET); /* * Set 'intr_shadow' to prevent an NMI from being injected on the * immediate VMRUN. */ error = svm_modify_intr_shadow(sc, vcpu, 1); KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error)); } #define EFER_MBZ_BITS 0xFFFFFFFFFFFF0200UL static int svm_write_efer(struct svm_softc *sc, int vcpu, uint64_t newval, bool *retu) { struct vm_exit *vme; struct vmcb_state *state; uint64_t changed, lma, oldval; int error; state = svm_get_vmcb_state(sc, vcpu); oldval = state->efer; VCPU_CTR2(sc->vm, vcpu, "wrmsr(efer) %#lx/%#lx", oldval, newval); newval &= ~0xFE; /* clear the Read-As-Zero (RAZ) bits */ changed = oldval ^ newval; if (newval & EFER_MBZ_BITS) goto gpf; /* APMv2 Table 14-5 "Long-Mode Consistency Checks" */ if (changed & EFER_LME) { if (state->cr0 & CR0_PG) goto gpf; } /* EFER.LMA = EFER.LME & CR0.PG */ if ((newval & EFER_LME) != 0 && (state->cr0 & CR0_PG) != 0) lma = EFER_LMA; else lma = 0; if ((newval & EFER_LMA) != lma) goto gpf; if (newval & EFER_NXE) { if (!vm_cpuid_capability(sc->vm, vcpu, VCC_NO_EXECUTE)) goto gpf; } /* * XXX bhyve does not enforce segment limits in 64-bit mode. Until * this is fixed flag guest attempt to set EFER_LMSLE as an error. */ if (newval & EFER_LMSLE) { vme = vm_exitinfo(sc->vm, vcpu); vm_exit_svm(vme, VMCB_EXIT_MSR, 1, 0); *retu = true; return (0); } if (newval & EFER_FFXSR) { if (!vm_cpuid_capability(sc->vm, vcpu, VCC_FFXSR)) goto gpf; } if (newval & EFER_TCE) { if (!vm_cpuid_capability(sc->vm, vcpu, VCC_TCE)) goto gpf; } error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, newval); KASSERT(error == 0, ("%s: error %d updating efer", __func__, error)); return (0); gpf: vm_inject_gp(sc->vm, vcpu); return (0); } static int emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val, bool *retu) { int error; if (lapic_msr(num)) error = lapic_wrmsr(sc->vm, vcpu, num, val, retu); else if (num == MSR_EFER) error = svm_write_efer(sc, vcpu, val, retu); else error = svm_wrmsr(sc, vcpu, num, val, retu); return (error); } static int emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu) { struct vmcb_state *state; struct svm_regctx *ctx; uint64_t result; int error; if (lapic_msr(num)) error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu); else error = svm_rdmsr(sc, vcpu, num, &result, retu); if (error == 0) { state = svm_get_vmcb_state(sc, vcpu); ctx = svm_get_guest_regctx(sc, vcpu); state->rax = result & 0xffffffff; ctx->sctx_rdx = result >> 32; } return (error); } #ifdef KTR static const char * exit_reason_to_str(uint64_t reason) { static char reasonbuf[32]; switch (reason) { case VMCB_EXIT_INVALID: return ("invalvmcb"); case VMCB_EXIT_SHUTDOWN: return ("shutdown"); case VMCB_EXIT_NPF: return ("nptfault"); case VMCB_EXIT_PAUSE: return ("pause"); case VMCB_EXIT_HLT: return ("hlt"); case VMCB_EXIT_CPUID: return ("cpuid"); case VMCB_EXIT_IO: return ("inout"); case VMCB_EXIT_MC: return ("mchk"); case VMCB_EXIT_INTR: return ("extintr"); case VMCB_EXIT_NMI: return ("nmi"); case VMCB_EXIT_VINTR: return ("vintr"); case VMCB_EXIT_MSR: return ("msr"); case VMCB_EXIT_IRET: return ("iret"); case VMCB_EXIT_MONITOR: return ("monitor"); case VMCB_EXIT_MWAIT: return ("mwait"); default: snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason); return (reasonbuf); } } #endif /* KTR */ /* * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs * that are due to instruction intercepts as well as MSR and IOIO intercepts * and exceptions caused by INT3, INTO and BOUND instructions. * * Return 1 if the nRIP is valid and 0 otherwise. */ static int nrip_valid(uint64_t exitcode) { switch (exitcode) { case 0x00 ... 0x0F: /* read of CR0 through CR15 */ case 0x10 ... 0x1F: /* write of CR0 through CR15 */ case 0x20 ... 0x2F: /* read of DR0 through DR15 */ case 0x30 ... 0x3F: /* write of DR0 through DR15 */ case 0x43: /* INT3 */ case 0x44: /* INTO */ case 0x45: /* BOUND */ case 0x65 ... 0x7C: /* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */ case 0x80 ... 0x8D: /* VMEXIT_VMRUN ... VMEXIT_XSETBV */ return (1); default: return (0); } } static int svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit) { struct vmcb *vmcb; struct vmcb_state *state; struct vmcb_ctrl *ctrl; struct svm_regctx *ctx; uint64_t code, info1, info2, val; uint32_t eax, ecx, edx; int error, errcode_valid, handled, idtvec, reflect; bool retu; ctx = svm_get_guest_regctx(svm_sc, vcpu); vmcb = svm_get_vmcb(svm_sc, vcpu); state = &vmcb->state; ctrl = &vmcb->ctrl; handled = 0; code = ctrl->exitcode; info1 = ctrl->exitinfo1; info2 = ctrl->exitinfo2; vmexit->exitcode = VM_EXITCODE_BOGUS; vmexit->rip = state->rip; vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0; vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1); /* * #VMEXIT(INVALID) needs to be handled early because the VMCB is * in an inconsistent state and can trigger assertions that would * never happen otherwise. */ if (code == VMCB_EXIT_INVALID) { vm_exit_svm(vmexit, code, info1, info2); return (0); } KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event " "injection valid bit is set %#lx", __func__, ctrl->eventinj)); KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15, ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)", vmexit->inst_length, code, info1, info2)); svm_update_virqinfo(svm_sc, vcpu); svm_save_intinfo(svm_sc, vcpu); switch (code) { case VMCB_EXIT_IRET: /* * Restart execution at "iret" but with the intercept cleared. */ vmexit->inst_length = 0; clear_nmi_blocking(svm_sc, vcpu); handled = 1; break; case VMCB_EXIT_VINTR: /* interrupt window exiting */ vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1); handled = 1; break; case VMCB_EXIT_INTR: /* external interrupt */ vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1); handled = 1; break; case VMCB_EXIT_NMI: /* external NMI */ handled = 1; break; case 0x40 ... 0x5F: vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1); reflect = 1; idtvec = code - 0x40; switch (idtvec) { case IDT_MC: /* * Call the machine check handler by hand. Also don't * reflect the machine check back into the guest. */ reflect = 0; VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler"); __asm __volatile("int $18"); break; case IDT_PF: error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2, info2); KASSERT(error == 0, ("%s: error %d updating cr2", __func__, error)); /* fallthru */ case IDT_NP: case IDT_SS: case IDT_GP: case IDT_AC: case IDT_TS: errcode_valid = 1; break; case IDT_DF: errcode_valid = 1; info1 = 0; break; case IDT_BP: case IDT_OF: case IDT_BR: /* * The 'nrip' field is populated for INT3, INTO and * BOUND exceptions and this also implies that * 'inst_length' is non-zero. * * Reset 'inst_length' to zero so the guest %rip at * event injection is identical to what it was when * the exception originally happened. */ VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d " "to zero before injecting exception %d", vmexit->inst_length, idtvec); vmexit->inst_length = 0; /* fallthru */ default: errcode_valid = 0; info1 = 0; break; } KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) " "when reflecting exception %d into guest", vmexit->inst_length, idtvec)); if (reflect) { /* Reflect the exception back into the guest */ VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception " "%d/%#x into the guest", idtvec, (int)info1); error = vm_inject_exception(svm_sc->vm, vcpu, idtvec, errcode_valid, info1, 0); KASSERT(error == 0, ("%s: vm_inject_exception error %d", __func__, error)); } handled = 1; break; case VMCB_EXIT_MSR: /* MSR access. */ eax = state->rax; ecx = ctx->sctx_rcx; edx = ctx->sctx_rdx; retu = false; if (info1) { vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1); val = (uint64_t)edx << 32 | eax; VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx", ecx, val); if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) { vmexit->exitcode = VM_EXITCODE_WRMSR; vmexit->u.msr.code = ecx; vmexit->u.msr.wval = val; } else if (!retu) { handled = 1; } else { KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, ("emulate_wrmsr retu with bogus exitcode")); } } else { VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx); vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1); if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) { vmexit->exitcode = VM_EXITCODE_RDMSR; vmexit->u.msr.code = ecx; } else if (!retu) { handled = 1; } else { KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, ("emulate_rdmsr retu with bogus exitcode")); } } break; case VMCB_EXIT_IO: handled = svm_handle_io(svm_sc, vcpu, vmexit); vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1); break; case VMCB_EXIT_CPUID: vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1); handled = x86_emulate_cpuid(svm_sc->vm, vcpu, (uint32_t *)&state->rax, (uint32_t *)&ctx->sctx_rbx, (uint32_t *)&ctx->sctx_rcx, (uint32_t *)&ctx->sctx_rdx); break; case VMCB_EXIT_HLT: vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1); vmexit->exitcode = VM_EXITCODE_HLT; vmexit->u.hlt.rflags = state->rflags; break; case VMCB_EXIT_PAUSE: vmexit->exitcode = VM_EXITCODE_PAUSE; vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1); break; case VMCB_EXIT_NPF: /* EXITINFO2 contains the faulting guest physical address */ if (info1 & VMCB_NPF_INFO1_RSV) { VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with " "reserved bits set: info1(%#lx) info2(%#lx)", info1, info2); } else if (vm_mem_allocated(svm_sc->vm, vcpu, info2)) { vmexit->exitcode = VM_EXITCODE_PAGING; vmexit->u.paging.gpa = info2; vmexit->u.paging.fault_type = npf_fault_type(info1); vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1); VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault " "on gpa %#lx/%#lx at rip %#lx", info2, info1, state->rip); } else if (svm_npf_emul_fault(info1)) { svm_handle_inst_emul(vmcb, info2, vmexit); vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1); VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault " "for gpa %#lx/%#lx at rip %#lx", info2, info1, state->rip); } break; case VMCB_EXIT_MONITOR: vmexit->exitcode = VM_EXITCODE_MONITOR; break; case VMCB_EXIT_MWAIT: vmexit->exitcode = VM_EXITCODE_MWAIT; break; default: vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1); break; } VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d", handled ? "handled" : "unhandled", exit_reason_to_str(code), vmexit->rip, vmexit->inst_length); if (handled) { vmexit->rip += vmexit->inst_length; vmexit->inst_length = 0; state->rip = vmexit->rip; } else { if (vmexit->exitcode == VM_EXITCODE_BOGUS) { /* * If this VM exit was not claimed by anybody then * treat it as a generic SVM exit. */ vm_exit_svm(vmexit, code, info1, info2); } else { /* * The exitcode and collateral have been populated. * The VM exit will be processed further in userland. */ } } return (handled); } static void svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu) { uint64_t intinfo; if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo)) return; KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not " "valid: %#lx", __func__, intinfo)); svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo), VMCB_EXITINTINFO_VECTOR(intinfo), VMCB_EXITINTINFO_EC(intinfo), VMCB_EXITINTINFO_EC_VALID(intinfo)); vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1); VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo); } /* * Inject event to virtual cpu. */ static void svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic) { struct vmcb_ctrl *ctrl; struct vmcb_state *state; struct svm_vcpu *vcpustate; uint8_t v_tpr; int vector, need_intr_window; int extint_pending; state = svm_get_vmcb_state(sc, vcpu); ctrl = svm_get_vmcb_ctrl(sc, vcpu); vcpustate = svm_get_vcpu(sc, vcpu); need_intr_window = 0; if (vcpustate->nextrip != state->rip) { ctrl->intr_shadow = 0; VCPU_CTR2(sc->vm, vcpu, "Guest interrupt blocking " "cleared due to rip change: %#lx/%#lx", vcpustate->nextrip, state->rip); } /* * Inject pending events or exceptions for this vcpu. * * An event might be pending because the previous #VMEXIT happened * during event delivery (i.e. ctrl->exitintinfo). * * An event might also be pending because an exception was injected * by the hypervisor (e.g. #PF during instruction emulation). */ svm_inj_intinfo(sc, vcpu); /* NMI event has priority over interrupts. */ if (vm_nmi_pending(sc->vm, vcpu)) { if (nmi_blocked(sc, vcpu)) { /* * Can't inject another NMI if the guest has not * yet executed an "iret" after the last NMI. */ VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due " "to NMI-blocking"); } else if (ctrl->intr_shadow) { /* * Can't inject an NMI if the vcpu is in an intr_shadow. */ VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to " "interrupt shadow"); need_intr_window = 1; goto done; } else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) { /* * If there is already an exception/interrupt pending * then defer the NMI until after that. */ VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to " "eventinj %#lx", ctrl->eventinj); /* * Use self-IPI to trigger a VM-exit as soon as * possible after the event injection is completed. * * This works only if the external interrupt exiting * is at a lower priority than the event injection. * * Although not explicitly specified in APMv2 the * relative priorities were verified empirically. */ ipi_cpu(curcpu, IPI_AST); /* XXX vmm_ipinum? */ } else { vm_nmi_clear(sc->vm, vcpu); /* Inject NMI, vector number is not used */ svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI, IDT_NMI, 0, false); /* virtual NMI blocking is now in effect */ enable_nmi_blocking(sc, vcpu); VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI"); } } extint_pending = vm_extint_pending(sc->vm, vcpu); if (!extint_pending) { if (!vlapic_pending_intr(vlapic, &vector)) goto done; KASSERT(vector >= 16 && vector <= 255, ("invalid vector %d from local APIC", vector)); } else { /* Ask the legacy pic for a vector to inject */ vatpic_pending_intr(sc->vm, &vector); KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR", vector)); } /* * If the guest has disabled interrupts or is in an interrupt shadow * then we cannot inject the pending interrupt. */ if ((state->rflags & PSL_I) == 0) { VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to " "rflags %#lx", vector, state->rflags); need_intr_window = 1; goto done; } if (ctrl->intr_shadow) { VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to " "interrupt shadow", vector); need_intr_window = 1; goto done; } if (ctrl->eventinj & VMCB_EVENTINJ_VALID) { VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to " "eventinj %#lx", vector, ctrl->eventinj); need_intr_window = 1; goto done; } svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false); if (!extint_pending) { vlapic_intr_accepted(vlapic, vector); } else { vm_extint_clear(sc->vm, vcpu); vatpic_intr_accepted(sc->vm, vector); } /* * Force a VM-exit as soon as the vcpu is ready to accept another * interrupt. This is done because the PIC might have another vector * that it wants to inject. Also, if the APIC has a pending interrupt * that was preempted by the ExtInt then it allows us to inject the * APIC vector as soon as possible. */ need_intr_window = 1; done: /* * The guest can modify the TPR by writing to %CR8. In guest mode * the processor reflects this write to V_TPR without hypervisor * intervention. * * The guest can also modify the TPR by writing to it via the memory * mapped APIC page. In this case, the write will be emulated by the * hypervisor. For this reason V_TPR must be updated before every * VMRUN. */ v_tpr = vlapic_get_cr8(vlapic); KASSERT(v_tpr <= 15, ("invalid v_tpr %#x", v_tpr)); if (ctrl->v_tpr != v_tpr) { VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x", ctrl->v_tpr, v_tpr); ctrl->v_tpr = v_tpr; svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); } if (need_intr_window) { /* * We use V_IRQ in conjunction with the VINTR intercept to * trap into the hypervisor as soon as a virtual interrupt * can be delivered. * * Since injected events are not subject to intercept checks * we need to ensure that the V_IRQ is not actually going to * be delivered on VM entry. The KASSERT below enforces this. */ KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 || (state->rflags & PSL_I) == 0 || ctrl->intr_shadow, ("Bogus intr_window_exiting: eventinj (%#lx), " "intr_shadow (%u), rflags (%#lx)", ctrl->eventinj, ctrl->intr_shadow, state->rflags)); enable_intr_window_exiting(sc, vcpu); } else { disable_intr_window_exiting(sc, vcpu); } } static __inline void restore_host_tss(void) { struct system_segment_descriptor *tss_sd; /* * The TSS descriptor was in use prior to launching the guest so it * has been marked busy. * * 'ltr' requires the descriptor to be marked available so change the * type to "64-bit available TSS". */ tss_sd = PCPU_GET(tss); tss_sd->sd_type = SDT_SYSTSS; ltr(GSEL(GPROC0_SEL, SEL_KPL)); } static void check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu) { struct svm_vcpu *vcpustate; struct vmcb_ctrl *ctrl; long eptgen; bool alloc_asid; KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not " "active on cpu %u", __func__, thiscpu)); vcpustate = svm_get_vcpu(sc, vcpuid); ctrl = svm_get_vmcb_ctrl(sc, vcpuid); /* * The TLB entries associated with the vcpu's ASID are not valid * if either of the following conditions is true: * * 1. The vcpu's ASID generation is different than the host cpu's * ASID generation. This happens when the vcpu migrates to a new * host cpu. It can also happen when the number of vcpus executing * on a host cpu is greater than the number of ASIDs available. * * 2. The pmap generation number is different than the value cached in * the 'vcpustate'. This happens when the host invalidates pages * belonging to the guest. * * asidgen eptgen Action * mismatch mismatch * 0 0 (a) * 0 1 (b1) or (b2) * 1 0 (c) * 1 1 (d) * * (a) There is no mismatch in eptgen or ASID generation and therefore * no further action is needed. * * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is * retained and the TLB entries associated with this ASID * are flushed by VMRUN. * * (b2) If the cpu does not support FlushByAsid then a new ASID is * allocated. * * (c) A new ASID is allocated. * * (d) A new ASID is allocated. */ alloc_asid = false; eptgen = pmap->pm_eptgen; ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING; if (vcpustate->asid.gen != asid[thiscpu].gen) { alloc_asid = true; /* (c) and (d) */ } else if (vcpustate->eptgen != eptgen) { if (flush_by_asid()) ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST; /* (b1) */ else alloc_asid = true; /* (b2) */ } else { /* * This is the common case (a). */ KASSERT(!alloc_asid, ("ASID allocation not necessary")); KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING, ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl)); } if (alloc_asid) { if (++asid[thiscpu].num >= nasid) { asid[thiscpu].num = 1; if (++asid[thiscpu].gen == 0) asid[thiscpu].gen = 1; /* * If this cpu does not support "flush-by-asid" * then flush the entire TLB on a generation * bump. Subsequent ASID allocation in this * generation can be done without a TLB flush. */ if (!flush_by_asid()) ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL; } vcpustate->asid.gen = asid[thiscpu].gen; vcpustate->asid.num = asid[thiscpu].num; ctrl->asid = vcpustate->asid.num; svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID); /* * If this cpu supports "flush-by-asid" then the TLB * was not flushed after the generation bump. The TLB * is flushed selectively after every new ASID allocation. */ if (flush_by_asid()) ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST; } vcpustate->eptgen = eptgen; KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero")); KASSERT(ctrl->asid == vcpustate->asid.num, ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num)); } static __inline void disable_gintr(void) { __asm __volatile("clgi"); } static __inline void enable_gintr(void) { __asm __volatile("stgi"); } static __inline void svm_dr_enter_guest(struct svm_regctx *gctx) { /* Save host control debug registers. */ gctx->host_dr7 = rdr7(); gctx->host_debugctl = rdmsr(MSR_DEBUGCTLMSR); /* * Disable debugging in DR7 and DEBUGCTL to avoid triggering * exceptions in the host based on the guest DRx values. The * guest DR6, DR7, and DEBUGCTL are saved/restored in the * VMCB. */ load_dr7(0); wrmsr(MSR_DEBUGCTLMSR, 0); /* Save host debug registers. */ gctx->host_dr0 = rdr0(); gctx->host_dr1 = rdr1(); gctx->host_dr2 = rdr2(); gctx->host_dr3 = rdr3(); gctx->host_dr6 = rdr6(); /* Restore guest debug registers. */ load_dr0(gctx->sctx_dr0); load_dr1(gctx->sctx_dr1); load_dr2(gctx->sctx_dr2); load_dr3(gctx->sctx_dr3); } static __inline void svm_dr_leave_guest(struct svm_regctx *gctx) { /* Save guest debug registers. */ gctx->sctx_dr0 = rdr0(); gctx->sctx_dr1 = rdr1(); gctx->sctx_dr2 = rdr2(); gctx->sctx_dr3 = rdr3(); /* * Restore host debug registers. Restore DR7 and DEBUGCTL * last. */ load_dr0(gctx->host_dr0); load_dr1(gctx->host_dr1); load_dr2(gctx->host_dr2); load_dr3(gctx->host_dr3); load_dr6(gctx->host_dr6); wrmsr(MSR_DEBUGCTLMSR, gctx->host_debugctl); load_dr7(gctx->host_dr7); } /* * Start vcpu with specified RIP. */ static int svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap, struct vm_eventinfo *evinfo) { struct svm_regctx *gctx; struct svm_softc *svm_sc; struct svm_vcpu *vcpustate; struct vmcb_state *state; struct vmcb_ctrl *ctrl; struct vm_exit *vmexit; struct vlapic *vlapic; struct vm *vm; uint64_t vmcb_pa; int handled; uint16_t ldt_sel; svm_sc = arg; vm = svm_sc->vm; vcpustate = svm_get_vcpu(svm_sc, vcpu); state = svm_get_vmcb_state(svm_sc, vcpu); ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu); vmexit = vm_exitinfo(vm, vcpu); vlapic = vm_lapic(vm, vcpu); gctx = svm_get_guest_regctx(svm_sc, vcpu); vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa; if (vcpustate->lastcpu != curcpu) { /* * Force new ASID allocation by invalidating the generation. */ vcpustate->asid.gen = 0; /* * Invalidate the VMCB state cache by marking all fields dirty. */ svm_set_dirty(svm_sc, vcpu, 0xffffffff); /* * XXX * Setting 'vcpustate->lastcpu' here is bit premature because * we may return from this function without actually executing * the VMRUN instruction. This could happen if a rendezvous * or an AST is pending on the first time through the loop. * * This works for now but any new side-effects of vcpu * migration should take this case into account. */ vcpustate->lastcpu = curcpu; vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1); } svm_msr_guest_enter(svm_sc, vcpu); /* Update Guest RIP */ state->rip = rip; do { /* * Disable global interrupts to guarantee atomicity during * loading of guest state. This includes not only the state * loaded by the "vmrun" instruction but also software state * maintained by the hypervisor: suspended and rendezvous * state, NPT generation number, vlapic interrupts etc. */ disable_gintr(); if (vcpu_suspended(evinfo)) { enable_gintr(); vm_exit_suspended(vm, vcpu, state->rip); break; } if (vcpu_rendezvous_pending(evinfo)) { enable_gintr(); vm_exit_rendezvous(vm, vcpu, state->rip); break; } if (vcpu_reqidle(evinfo)) { enable_gintr(); vm_exit_reqidle(vm, vcpu, state->rip); break; } /* We are asked to give the cpu by scheduler. */ if (vcpu_should_yield(vm, vcpu)) { enable_gintr(); vm_exit_astpending(vm, vcpu, state->rip); break; } if (vcpu_debugged(vm, vcpu)) { enable_gintr(); vm_exit_debug(vm, vcpu, state->rip); break; } /* * #VMEXIT resumes the host with the guest LDTR, so * save the current LDT selector so it can be restored * after an exit. The userspace hypervisor probably * doesn't use a LDT, but save and restore it to be * safe. */ ldt_sel = sldt(); svm_inj_interrupts(svm_sc, vcpu, vlapic); /* Activate the nested pmap on 'curcpu' */ CPU_SET_ATOMIC_ACQ(curcpu, &pmap->pm_active); /* * Check the pmap generation and the ASID generation to * ensure that the vcpu does not use stale TLB mappings. */ check_asid(svm_sc, vcpu, pmap, curcpu); ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty; vcpustate->dirty = 0; VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean); /* Launch Virtual Machine. */ VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip); svm_dr_enter_guest(gctx); svm_launch(vmcb_pa, gctx, get_pcpu()); svm_dr_leave_guest(gctx); CPU_CLR_ATOMIC(curcpu, &pmap->pm_active); /* * The host GDTR and IDTR is saved by VMRUN and restored * automatically on #VMEXIT. However, the host TSS needs * to be restored explicitly. */ restore_host_tss(); /* Restore host LDTR. */ lldt(ldt_sel); /* #VMEXIT disables interrupts so re-enable them here. */ enable_gintr(); /* Update 'nextrip' */ vcpustate->nextrip = state->rip; /* Handle #VMEXIT and if required return to user space. */ handled = svm_vmexit(svm_sc, vcpu, vmexit); } while (handled); svm_msr_guest_exit(svm_sc, vcpu); return (0); } static void svm_vmcleanup(void *arg) { struct svm_softc *sc = arg; contigfree(sc->iopm_bitmap, SVM_IO_BITMAP_SIZE, M_SVM); contigfree(sc->msr_bitmap, SVM_MSR_BITMAP_SIZE, M_SVM); free(sc, M_SVM); } static register_t * swctx_regptr(struct svm_regctx *regctx, int reg) { switch (reg) { case VM_REG_GUEST_RBX: return (®ctx->sctx_rbx); case VM_REG_GUEST_RCX: return (®ctx->sctx_rcx); case VM_REG_GUEST_RDX: return (®ctx->sctx_rdx); case VM_REG_GUEST_RDI: return (®ctx->sctx_rdi); case VM_REG_GUEST_RSI: return (®ctx->sctx_rsi); case VM_REG_GUEST_RBP: return (®ctx->sctx_rbp); case VM_REG_GUEST_R8: return (®ctx->sctx_r8); case VM_REG_GUEST_R9: return (®ctx->sctx_r9); case VM_REG_GUEST_R10: return (®ctx->sctx_r10); case VM_REG_GUEST_R11: return (®ctx->sctx_r11); case VM_REG_GUEST_R12: return (®ctx->sctx_r12); case VM_REG_GUEST_R13: return (®ctx->sctx_r13); case VM_REG_GUEST_R14: return (®ctx->sctx_r14); case VM_REG_GUEST_R15: return (®ctx->sctx_r15); case VM_REG_GUEST_DR0: return (®ctx->sctx_dr0); case VM_REG_GUEST_DR1: return (®ctx->sctx_dr1); case VM_REG_GUEST_DR2: return (®ctx->sctx_dr2); case VM_REG_GUEST_DR3: return (®ctx->sctx_dr3); default: return (NULL); } } static int svm_getreg(void *arg, int vcpu, int ident, uint64_t *val) { struct svm_softc *svm_sc; register_t *reg; svm_sc = arg; if (ident == VM_REG_GUEST_INTR_SHADOW) { return (svm_get_intr_shadow(svm_sc, vcpu, val)); } if (vmcb_read(svm_sc, vcpu, ident, val) == 0) { return (0); } reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident); if (reg != NULL) { *val = *reg; return (0); } VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident); return (EINVAL); } static int svm_setreg(void *arg, int vcpu, int ident, uint64_t val) { struct svm_softc *svm_sc; register_t *reg; svm_sc = arg; if (ident == VM_REG_GUEST_INTR_SHADOW) { return (svm_modify_intr_shadow(svm_sc, vcpu, val)); } if (vmcb_write(svm_sc, vcpu, ident, val) == 0) { return (0); } reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident); if (reg != NULL) { *reg = val; return (0); } if (ident == VM_REG_GUEST_ENTRY_INST_LENGTH) { /* Ignore. */ return (0); } /* * XXX deal with CR3 and invalidate TLB entries tagged with the * vcpu's ASID. This needs to be treated differently depending on * whether 'running' is true/false. */ VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident); return (EINVAL); } static int svm_setcap(void *arg, int vcpu, int type, int val) { struct svm_softc *sc; int error; sc = arg; error = 0; switch (type) { case VM_CAP_HALT_EXIT: svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_HLT, val); break; case VM_CAP_PAUSE_EXIT: svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_PAUSE, val); break; case VM_CAP_UNRESTRICTED_GUEST: /* Unrestricted guest execution cannot be disabled in SVM */ if (val == 0) error = EINVAL; break; default: error = ENOENT; break; } return (error); } static int svm_getcap(void *arg, int vcpu, int type, int *retval) { struct svm_softc *sc; int error; sc = arg; error = 0; switch (type) { case VM_CAP_HALT_EXIT: *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_HLT); break; case VM_CAP_PAUSE_EXIT: *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_PAUSE); break; case VM_CAP_UNRESTRICTED_GUEST: *retval = 1; /* unrestricted guest is always enabled */ break; default: error = ENOENT; break; } return (error); } static struct vlapic * svm_vlapic_init(void *arg, int vcpuid) { struct svm_softc *svm_sc; struct vlapic *vlapic; svm_sc = arg; vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO); vlapic->vm = svm_sc->vm; vlapic->vcpuid = vcpuid; vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid]; vlapic_init(vlapic); return (vlapic); } static void svm_vlapic_cleanup(void *arg, struct vlapic *vlapic) { vlapic_cleanup(vlapic); free(vlapic, M_SVM_VLAPIC); } struct vmm_ops vmm_ops_amd = { .init = svm_init, .cleanup = svm_cleanup, .resume = svm_restore, .vminit = svm_vminit, .vmrun = svm_vmrun, .vmcleanup = svm_vmcleanup, .vmgetreg = svm_getreg, .vmsetreg = svm_setreg, .vmgetdesc = vmcb_getdesc, .vmsetdesc = vmcb_setdesc, .vmgetcap = svm_getcap, .vmsetcap = svm_setcap, .vmspace_alloc = svm_npt_alloc, .vmspace_free = svm_npt_free, .vlapic_init = svm_vlapic_init, .vlapic_cleanup = svm_vlapic_cleanup, }; Index: head/sys/amd64/vmm/intel/ept.c =================================================================== --- head/sys/amd64/vmm/intel/ept.c (revision 357973) +++ head/sys/amd64/vmm/intel/ept.c (revision 357974) @@ -1,207 +1,208 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include "vmx_cpufunc.h" #include "ept.h" #define EPT_SUPPORTS_EXEC_ONLY(cap) ((cap) & (1UL << 0)) #define EPT_PWL4(cap) ((cap) & (1UL << 6)) #define EPT_MEMORY_TYPE_WB(cap) ((cap) & (1UL << 14)) #define EPT_PDE_SUPERPAGE(cap) ((cap) & (1UL << 16)) /* 2MB pages */ #define EPT_PDPTE_SUPERPAGE(cap) ((cap) & (1UL << 17)) /* 1GB pages */ #define INVEPT_SUPPORTED(cap) ((cap) & (1UL << 20)) #define AD_BITS_SUPPORTED(cap) ((cap) & (1UL << 21)) #define INVVPID_SUPPORTED(cap) ((cap) & (1UL << 32)) #define INVVPID_ALL_TYPES_MASK 0xF0000000000UL #define INVVPID_ALL_TYPES_SUPPORTED(cap) \ (((cap) & INVVPID_ALL_TYPES_MASK) == INVVPID_ALL_TYPES_MASK) #define INVEPT_ALL_TYPES_MASK 0x6000000UL #define INVEPT_ALL_TYPES_SUPPORTED(cap) \ (((cap) & INVEPT_ALL_TYPES_MASK) == INVEPT_ALL_TYPES_MASK) #define EPT_PWLEVELS 4 /* page walk levels */ #define EPT_ENABLE_AD_BITS (1 << 6) SYSCTL_DECL(_hw_vmm); -SYSCTL_NODE(_hw_vmm, OID_AUTO, ept, CTLFLAG_RW, NULL, NULL); +SYSCTL_NODE(_hw_vmm, OID_AUTO, ept, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, + NULL); static int ept_enable_ad_bits; static int ept_pmap_flags; SYSCTL_INT(_hw_vmm_ept, OID_AUTO, pmap_flags, CTLFLAG_RD, &ept_pmap_flags, 0, NULL); int ept_init(int ipinum) { int use_hw_ad_bits, use_superpages, use_exec_only; uint64_t cap; cap = rdmsr(MSR_VMX_EPT_VPID_CAP); /* * Verify that: * - page walk length is 4 steps * - extended page tables can be laid out in write-back memory * - invvpid instruction with all possible types is supported * - invept instruction with all possible types is supported */ if (!EPT_PWL4(cap) || !EPT_MEMORY_TYPE_WB(cap) || !INVVPID_SUPPORTED(cap) || !INVVPID_ALL_TYPES_SUPPORTED(cap) || !INVEPT_SUPPORTED(cap) || !INVEPT_ALL_TYPES_SUPPORTED(cap)) return (EINVAL); ept_pmap_flags = ipinum & PMAP_NESTED_IPIMASK; use_superpages = 1; TUNABLE_INT_FETCH("hw.vmm.ept.use_superpages", &use_superpages); if (use_superpages && EPT_PDE_SUPERPAGE(cap)) ept_pmap_flags |= PMAP_PDE_SUPERPAGE; /* 2MB superpage */ use_hw_ad_bits = 1; TUNABLE_INT_FETCH("hw.vmm.ept.use_hw_ad_bits", &use_hw_ad_bits); if (use_hw_ad_bits && AD_BITS_SUPPORTED(cap)) ept_enable_ad_bits = 1; else ept_pmap_flags |= PMAP_EMULATE_AD_BITS; use_exec_only = 1; TUNABLE_INT_FETCH("hw.vmm.ept.use_exec_only", &use_exec_only); if (use_exec_only && EPT_SUPPORTS_EXEC_ONLY(cap)) ept_pmap_flags |= PMAP_SUPPORTS_EXEC_ONLY; return (0); } #if 0 static void ept_dump(uint64_t *ptp, int nlevels) { int i, t, tabs; uint64_t *ptpnext, ptpval; if (--nlevels < 0) return; tabs = 3 - nlevels; for (t = 0; t < tabs; t++) printf("\t"); printf("PTP = %p\n", ptp); for (i = 0; i < 512; i++) { ptpval = ptp[i]; if (ptpval == 0) continue; for (t = 0; t < tabs; t++) printf("\t"); printf("%3d 0x%016lx\n", i, ptpval); if (nlevels != 0 && (ptpval & EPT_PG_SUPERPAGE) == 0) { ptpnext = (uint64_t *) PHYS_TO_DMAP(ptpval & EPT_ADDR_MASK); ept_dump(ptpnext, nlevels); } } } #endif static void invept_single_context(void *arg) { struct invept_desc desc = *(struct invept_desc *)arg; invept(INVEPT_TYPE_SINGLE_CONTEXT, desc); } void ept_invalidate_mappings(u_long eptp) { struct invept_desc invept_desc = { 0 }; invept_desc.eptp = eptp; smp_rendezvous(NULL, invept_single_context, NULL, &invept_desc); } static int ept_pinit(pmap_t pmap) { return (pmap_pinit_type(pmap, PT_EPT, ept_pmap_flags)); } struct vmspace * ept_vmspace_alloc(vm_offset_t min, vm_offset_t max) { return (vmspace_alloc(min, max, ept_pinit)); } void ept_vmspace_free(struct vmspace *vmspace) { vmspace_free(vmspace); } uint64_t eptp(uint64_t pml4) { uint64_t eptp_val; eptp_val = pml4 | (EPT_PWLEVELS - 1) << 3 | PAT_WRITE_BACK; if (ept_enable_ad_bits) eptp_val |= EPT_ENABLE_AD_BITS; return (eptp_val); } Index: head/sys/amd64/vmm/intel/vmx.c =================================================================== --- head/sys/amd64/vmm/intel/vmx.c (revision 357973) +++ head/sys/amd64/vmm/intel/vmx.c (revision 357974) @@ -1,3836 +1,3839 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * Copyright (c) 2018 Joyent, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "vmm_lapic.h" #include "vmm_host.h" #include "vmm_ioport.h" #include "vmm_ktr.h" #include "vmm_stat.h" #include "vatpic.h" #include "vlapic.h" #include "vlapic_priv.h" #include "ept.h" #include "vmx_cpufunc.h" #include "vmx.h" #include "vmx_msr.h" #include "x86.h" #include "vmx_controls.h" #define PINBASED_CTLS_ONE_SETTING \ (PINBASED_EXTINT_EXITING | \ PINBASED_NMI_EXITING | \ PINBASED_VIRTUAL_NMI) #define PINBASED_CTLS_ZERO_SETTING 0 #define PROCBASED_CTLS_WINDOW_SETTING \ (PROCBASED_INT_WINDOW_EXITING | \ PROCBASED_NMI_WINDOW_EXITING) #define PROCBASED_CTLS_ONE_SETTING \ (PROCBASED_SECONDARY_CONTROLS | \ PROCBASED_MWAIT_EXITING | \ PROCBASED_MONITOR_EXITING | \ PROCBASED_IO_EXITING | \ PROCBASED_MSR_BITMAPS | \ PROCBASED_CTLS_WINDOW_SETTING | \ PROCBASED_CR8_LOAD_EXITING | \ PROCBASED_CR8_STORE_EXITING) #define PROCBASED_CTLS_ZERO_SETTING \ (PROCBASED_CR3_LOAD_EXITING | \ PROCBASED_CR3_STORE_EXITING | \ PROCBASED_IO_BITMAPS) #define PROCBASED_CTLS2_ONE_SETTING PROCBASED2_ENABLE_EPT #define PROCBASED_CTLS2_ZERO_SETTING 0 #define VM_EXIT_CTLS_ONE_SETTING \ (VM_EXIT_SAVE_DEBUG_CONTROLS | \ VM_EXIT_HOST_LMA | \ VM_EXIT_SAVE_EFER | \ VM_EXIT_LOAD_EFER | \ VM_EXIT_ACKNOWLEDGE_INTERRUPT) #define VM_EXIT_CTLS_ZERO_SETTING 0 #define VM_ENTRY_CTLS_ONE_SETTING \ (VM_ENTRY_LOAD_DEBUG_CONTROLS | \ VM_ENTRY_LOAD_EFER) #define VM_ENTRY_CTLS_ZERO_SETTING \ (VM_ENTRY_INTO_SMM | \ VM_ENTRY_DEACTIVATE_DUAL_MONITOR) #define HANDLED 1 #define UNHANDLED 0 static MALLOC_DEFINE(M_VMX, "vmx", "vmx"); static MALLOC_DEFINE(M_VLAPIC, "vlapic", "vlapic"); SYSCTL_DECL(_hw_vmm); -SYSCTL_NODE(_hw_vmm, OID_AUTO, vmx, CTLFLAG_RW, NULL, NULL); +SYSCTL_NODE(_hw_vmm, OID_AUTO, vmx, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, + NULL); int vmxon_enabled[MAXCPU]; static char vmxon_region[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE); static uint32_t pinbased_ctls, procbased_ctls, procbased_ctls2; static uint32_t exit_ctls, entry_ctls; static uint64_t cr0_ones_mask, cr0_zeros_mask; SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr0_ones_mask, CTLFLAG_RD, &cr0_ones_mask, 0, NULL); SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr0_zeros_mask, CTLFLAG_RD, &cr0_zeros_mask, 0, NULL); static uint64_t cr4_ones_mask, cr4_zeros_mask; SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr4_ones_mask, CTLFLAG_RD, &cr4_ones_mask, 0, NULL); SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr4_zeros_mask, CTLFLAG_RD, &cr4_zeros_mask, 0, NULL); static int vmx_initialized; SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, initialized, CTLFLAG_RD, &vmx_initialized, 0, "Intel VMX initialized"); /* * Optional capabilities */ -static SYSCTL_NODE(_hw_vmm_vmx, OID_AUTO, cap, CTLFLAG_RW, NULL, NULL); +static SYSCTL_NODE(_hw_vmm_vmx, OID_AUTO, cap, + CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, + NULL); static int cap_halt_exit; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, halt_exit, CTLFLAG_RD, &cap_halt_exit, 0, "HLT triggers a VM-exit"); static int cap_pause_exit; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, pause_exit, CTLFLAG_RD, &cap_pause_exit, 0, "PAUSE triggers a VM-exit"); static int cap_unrestricted_guest; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, unrestricted_guest, CTLFLAG_RD, &cap_unrestricted_guest, 0, "Unrestricted guests"); static int cap_monitor_trap; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, monitor_trap, CTLFLAG_RD, &cap_monitor_trap, 0, "Monitor trap flag"); static int cap_invpcid; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, invpcid, CTLFLAG_RD, &cap_invpcid, 0, "Guests are allowed to use INVPCID"); static int virtual_interrupt_delivery; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, virtual_interrupt_delivery, CTLFLAG_RD, &virtual_interrupt_delivery, 0, "APICv virtual interrupt delivery support"); static int posted_interrupts; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, posted_interrupts, CTLFLAG_RD, &posted_interrupts, 0, "APICv posted interrupt support"); static int pirvec = -1; SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, posted_interrupt_vector, CTLFLAG_RD, &pirvec, 0, "APICv posted interrupt vector"); static struct unrhdr *vpid_unr; static u_int vpid_alloc_failed; SYSCTL_UINT(_hw_vmm_vmx, OID_AUTO, vpid_alloc_failed, CTLFLAG_RD, &vpid_alloc_failed, 0, NULL); int guest_l1d_flush; SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, l1d_flush, CTLFLAG_RD, &guest_l1d_flush, 0, NULL); int guest_l1d_flush_sw; SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, l1d_flush_sw, CTLFLAG_RD, &guest_l1d_flush_sw, 0, NULL); static struct msr_entry msr_load_list[1] __aligned(16); /* * The definitions of SDT probes for VMX. */ SDT_PROBE_DEFINE3(vmm, vmx, exit, entry, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, taskswitch, "struct vmx *", "int", "struct vm_exit *", "struct vm_task_switch *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, craccess, "struct vmx *", "int", "struct vm_exit *", "uint64_t"); SDT_PROBE_DEFINE4(vmm, vmx, exit, rdmsr, "struct vmx *", "int", "struct vm_exit *", "uint32_t"); SDT_PROBE_DEFINE5(vmm, vmx, exit, wrmsr, "struct vmx *", "int", "struct vm_exit *", "uint32_t", "uint64_t"); SDT_PROBE_DEFINE3(vmm, vmx, exit, halt, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, mtrap, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, pause, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, intrwindow, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, interrupt, "struct vmx *", "int", "struct vm_exit *", "uint32_t"); SDT_PROBE_DEFINE3(vmm, vmx, exit, nmiwindow, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, inout, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, cpuid, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE5(vmm, vmx, exit, exception, "struct vmx *", "int", "struct vm_exit *", "uint32_t", "int"); SDT_PROBE_DEFINE5(vmm, vmx, exit, nestedfault, "struct vmx *", "int", "struct vm_exit *", "uint64_t", "uint64_t"); SDT_PROBE_DEFINE4(vmm, vmx, exit, mmiofault, "struct vmx *", "int", "struct vm_exit *", "uint64_t"); SDT_PROBE_DEFINE3(vmm, vmx, exit, eoi, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, apicaccess, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, apicwrite, "struct vmx *", "int", "struct vm_exit *", "struct vlapic *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, xsetbv, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, monitor, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, mwait, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, vminsn, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, unknown, "struct vmx *", "int", "struct vm_exit *", "uint32_t"); SDT_PROBE_DEFINE4(vmm, vmx, exit, return, "struct vmx *", "int", "struct vm_exit *", "int"); /* * Use the last page below 4GB as the APIC access address. This address is * occupied by the boot firmware so it is guaranteed that it will not conflict * with a page in system memory. */ #define APIC_ACCESS_ADDRESS 0xFFFFF000 static int vmx_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc); static int vmx_getreg(void *arg, int vcpu, int reg, uint64_t *retval); static int vmxctx_setreg(struct vmxctx *vmxctx, int reg, uint64_t val); static void vmx_inject_pir(struct vlapic *vlapic); #ifdef KTR static const char * exit_reason_to_str(int reason) { static char reasonbuf[32]; switch (reason) { case EXIT_REASON_EXCEPTION: return "exception"; case EXIT_REASON_EXT_INTR: return "extint"; case EXIT_REASON_TRIPLE_FAULT: return "triplefault"; case EXIT_REASON_INIT: return "init"; case EXIT_REASON_SIPI: return "sipi"; case EXIT_REASON_IO_SMI: return "iosmi"; case EXIT_REASON_SMI: return "smi"; case EXIT_REASON_INTR_WINDOW: return "intrwindow"; case EXIT_REASON_NMI_WINDOW: return "nmiwindow"; case EXIT_REASON_TASK_SWITCH: return "taskswitch"; case EXIT_REASON_CPUID: return "cpuid"; case EXIT_REASON_GETSEC: return "getsec"; case EXIT_REASON_HLT: return "hlt"; case EXIT_REASON_INVD: return "invd"; case EXIT_REASON_INVLPG: return "invlpg"; case EXIT_REASON_RDPMC: return "rdpmc"; case EXIT_REASON_RDTSC: return "rdtsc"; case EXIT_REASON_RSM: return "rsm"; case EXIT_REASON_VMCALL: return "vmcall"; case EXIT_REASON_VMCLEAR: return "vmclear"; case EXIT_REASON_VMLAUNCH: return "vmlaunch"; case EXIT_REASON_VMPTRLD: return "vmptrld"; case EXIT_REASON_VMPTRST: return "vmptrst"; case EXIT_REASON_VMREAD: return "vmread"; case EXIT_REASON_VMRESUME: return "vmresume"; case EXIT_REASON_VMWRITE: return "vmwrite"; case EXIT_REASON_VMXOFF: return "vmxoff"; case EXIT_REASON_VMXON: return "vmxon"; case EXIT_REASON_CR_ACCESS: return "craccess"; case EXIT_REASON_DR_ACCESS: return "draccess"; case EXIT_REASON_INOUT: return "inout"; case EXIT_REASON_RDMSR: return "rdmsr"; case EXIT_REASON_WRMSR: return "wrmsr"; case EXIT_REASON_INVAL_VMCS: return "invalvmcs"; case EXIT_REASON_INVAL_MSR: return "invalmsr"; case EXIT_REASON_MWAIT: return "mwait"; case EXIT_REASON_MTF: return "mtf"; case EXIT_REASON_MONITOR: return "monitor"; case EXIT_REASON_PAUSE: return "pause"; case EXIT_REASON_MCE_DURING_ENTRY: return "mce-during-entry"; case EXIT_REASON_TPR: return "tpr"; case EXIT_REASON_APIC_ACCESS: return "apic-access"; case EXIT_REASON_GDTR_IDTR: return "gdtridtr"; case EXIT_REASON_LDTR_TR: return "ldtrtr"; case EXIT_REASON_EPT_FAULT: return "eptfault"; case EXIT_REASON_EPT_MISCONFIG: return "eptmisconfig"; case EXIT_REASON_INVEPT: return "invept"; case EXIT_REASON_RDTSCP: return "rdtscp"; case EXIT_REASON_VMX_PREEMPT: return "vmxpreempt"; case EXIT_REASON_INVVPID: return "invvpid"; case EXIT_REASON_WBINVD: return "wbinvd"; case EXIT_REASON_XSETBV: return "xsetbv"; case EXIT_REASON_APIC_WRITE: return "apic-write"; default: snprintf(reasonbuf, sizeof(reasonbuf), "%d", reason); return (reasonbuf); } } #endif /* KTR */ static int vmx_allow_x2apic_msrs(struct vmx *vmx) { int i, error; error = 0; /* * Allow readonly access to the following x2APIC MSRs from the guest. */ error += guest_msr_ro(vmx, MSR_APIC_ID); error += guest_msr_ro(vmx, MSR_APIC_VERSION); error += guest_msr_ro(vmx, MSR_APIC_LDR); error += guest_msr_ro(vmx, MSR_APIC_SVR); for (i = 0; i < 8; i++) error += guest_msr_ro(vmx, MSR_APIC_ISR0 + i); for (i = 0; i < 8; i++) error += guest_msr_ro(vmx, MSR_APIC_TMR0 + i); for (i = 0; i < 8; i++) error += guest_msr_ro(vmx, MSR_APIC_IRR0 + i); error += guest_msr_ro(vmx, MSR_APIC_ESR); error += guest_msr_ro(vmx, MSR_APIC_LVT_TIMER); error += guest_msr_ro(vmx, MSR_APIC_LVT_THERMAL); error += guest_msr_ro(vmx, MSR_APIC_LVT_PCINT); error += guest_msr_ro(vmx, MSR_APIC_LVT_LINT0); error += guest_msr_ro(vmx, MSR_APIC_LVT_LINT1); error += guest_msr_ro(vmx, MSR_APIC_LVT_ERROR); error += guest_msr_ro(vmx, MSR_APIC_ICR_TIMER); error += guest_msr_ro(vmx, MSR_APIC_DCR_TIMER); error += guest_msr_ro(vmx, MSR_APIC_ICR); /* * Allow TPR, EOI and SELF_IPI MSRs to be read and written by the guest. * * These registers get special treatment described in the section * "Virtualizing MSR-Based APIC Accesses". */ error += guest_msr_rw(vmx, MSR_APIC_TPR); error += guest_msr_rw(vmx, MSR_APIC_EOI); error += guest_msr_rw(vmx, MSR_APIC_SELF_IPI); return (error); } u_long vmx_fix_cr0(u_long cr0) { return ((cr0 | cr0_ones_mask) & ~cr0_zeros_mask); } u_long vmx_fix_cr4(u_long cr4) { return ((cr4 | cr4_ones_mask) & ~cr4_zeros_mask); } static void vpid_free(int vpid) { if (vpid < 0 || vpid > 0xffff) panic("vpid_free: invalid vpid %d", vpid); /* * VPIDs [0,VM_MAXCPU] are special and are not allocated from * the unit number allocator. */ if (vpid > VM_MAXCPU) free_unr(vpid_unr, vpid); } static void vpid_alloc(uint16_t *vpid, int num) { int i, x; if (num <= 0 || num > VM_MAXCPU) panic("invalid number of vpids requested: %d", num); /* * If the "enable vpid" execution control is not enabled then the * VPID is required to be 0 for all vcpus. */ if ((procbased_ctls2 & PROCBASED2_ENABLE_VPID) == 0) { for (i = 0; i < num; i++) vpid[i] = 0; return; } /* * Allocate a unique VPID for each vcpu from the unit number allocator. */ for (i = 0; i < num; i++) { x = alloc_unr(vpid_unr); if (x == -1) break; else vpid[i] = x; } if (i < num) { atomic_add_int(&vpid_alloc_failed, 1); /* * If the unit number allocator does not have enough unique * VPIDs then we need to allocate from the [1,VM_MAXCPU] range. * * These VPIDs are not be unique across VMs but this does not * affect correctness because the combined mappings are also * tagged with the EP4TA which is unique for each VM. * * It is still sub-optimal because the invvpid will invalidate * combined mappings for a particular VPID across all EP4TAs. */ while (i-- > 0) vpid_free(vpid[i]); for (i = 0; i < num; i++) vpid[i] = i + 1; } } static void vpid_init(void) { /* * VPID 0 is required when the "enable VPID" execution control is * disabled. * * VPIDs [1,VM_MAXCPU] are used as the "overflow namespace" when the * unit number allocator does not have sufficient unique VPIDs to * satisfy the allocation. * * The remaining VPIDs are managed by the unit number allocator. */ vpid_unr = new_unrhdr(VM_MAXCPU + 1, 0xffff, NULL); } static void vmx_disable(void *arg __unused) { struct invvpid_desc invvpid_desc = { 0 }; struct invept_desc invept_desc = { 0 }; if (vmxon_enabled[curcpu]) { /* * See sections 25.3.3.3 and 25.3.3.4 in Intel Vol 3b. * * VMXON or VMXOFF are not required to invalidate any TLB * caching structures. This prevents potential retention of * cached information in the TLB between distinct VMX episodes. */ invvpid(INVVPID_TYPE_ALL_CONTEXTS, invvpid_desc); invept(INVEPT_TYPE_ALL_CONTEXTS, invept_desc); vmxoff(); } load_cr4(rcr4() & ~CR4_VMXE); } static int vmx_cleanup(void) { if (pirvec >= 0) lapic_ipi_free(pirvec); if (vpid_unr != NULL) { delete_unrhdr(vpid_unr); vpid_unr = NULL; } if (nmi_flush_l1d_sw == 1) nmi_flush_l1d_sw = 0; smp_rendezvous(NULL, vmx_disable, NULL, NULL); return (0); } static void vmx_enable(void *arg __unused) { int error; uint64_t feature_control; feature_control = rdmsr(MSR_IA32_FEATURE_CONTROL); if ((feature_control & IA32_FEATURE_CONTROL_LOCK) == 0 || (feature_control & IA32_FEATURE_CONTROL_VMX_EN) == 0) { wrmsr(MSR_IA32_FEATURE_CONTROL, feature_control | IA32_FEATURE_CONTROL_VMX_EN | IA32_FEATURE_CONTROL_LOCK); } load_cr4(rcr4() | CR4_VMXE); *(uint32_t *)vmxon_region[curcpu] = vmx_revision(); error = vmxon(vmxon_region[curcpu]); if (error == 0) vmxon_enabled[curcpu] = 1; } static void vmx_restore(void) { if (vmxon_enabled[curcpu]) vmxon(vmxon_region[curcpu]); } static int vmx_init(int ipinum) { int error, use_tpr_shadow; uint64_t basic, fixed0, fixed1, feature_control; uint32_t tmp, procbased2_vid_bits; /* CPUID.1:ECX[bit 5] must be 1 for processor to support VMX */ if (!(cpu_feature2 & CPUID2_VMX)) { printf("vmx_init: processor does not support VMX operation\n"); return (ENXIO); } /* * Verify that MSR_IA32_FEATURE_CONTROL lock and VMXON enable bits * are set (bits 0 and 2 respectively). */ feature_control = rdmsr(MSR_IA32_FEATURE_CONTROL); if ((feature_control & IA32_FEATURE_CONTROL_LOCK) == 1 && (feature_control & IA32_FEATURE_CONTROL_VMX_EN) == 0) { printf("vmx_init: VMX operation disabled by BIOS\n"); return (ENXIO); } /* * Verify capabilities MSR_VMX_BASIC: * - bit 54 indicates support for INS/OUTS decoding */ basic = rdmsr(MSR_VMX_BASIC); if ((basic & (1UL << 54)) == 0) { printf("vmx_init: processor does not support desired basic " "capabilities\n"); return (EINVAL); } /* Check support for primary processor-based VM-execution controls */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_CTLS_ONE_SETTING, PROCBASED_CTLS_ZERO_SETTING, &procbased_ctls); if (error) { printf("vmx_init: processor does not support desired primary " "processor-based controls\n"); return (error); } /* Clear the processor-based ctl bits that are set on demand */ procbased_ctls &= ~PROCBASED_CTLS_WINDOW_SETTING; /* Check support for secondary processor-based VM-execution controls */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED_CTLS2_ONE_SETTING, PROCBASED_CTLS2_ZERO_SETTING, &procbased_ctls2); if (error) { printf("vmx_init: processor does not support desired secondary " "processor-based controls\n"); return (error); } /* Check support for VPID */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED2_ENABLE_VPID, 0, &tmp); if (error == 0) procbased_ctls2 |= PROCBASED2_ENABLE_VPID; /* Check support for pin-based VM-execution controls */ error = vmx_set_ctlreg(MSR_VMX_PINBASED_CTLS, MSR_VMX_TRUE_PINBASED_CTLS, PINBASED_CTLS_ONE_SETTING, PINBASED_CTLS_ZERO_SETTING, &pinbased_ctls); if (error) { printf("vmx_init: processor does not support desired " "pin-based controls\n"); return (error); } /* Check support for VM-exit controls */ error = vmx_set_ctlreg(MSR_VMX_EXIT_CTLS, MSR_VMX_TRUE_EXIT_CTLS, VM_EXIT_CTLS_ONE_SETTING, VM_EXIT_CTLS_ZERO_SETTING, &exit_ctls); if (error) { printf("vmx_init: processor does not support desired " "exit controls\n"); return (error); } /* Check support for VM-entry controls */ error = vmx_set_ctlreg(MSR_VMX_ENTRY_CTLS, MSR_VMX_TRUE_ENTRY_CTLS, VM_ENTRY_CTLS_ONE_SETTING, VM_ENTRY_CTLS_ZERO_SETTING, &entry_ctls); if (error) { printf("vmx_init: processor does not support desired " "entry controls\n"); return (error); } /* * Check support for optional features by testing them * as individual bits */ cap_halt_exit = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_HLT_EXITING, 0, &tmp) == 0); cap_monitor_trap = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_PROCBASED_CTLS, PROCBASED_MTF, 0, &tmp) == 0); cap_pause_exit = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_PAUSE_EXITING, 0, &tmp) == 0); cap_unrestricted_guest = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED2_UNRESTRICTED_GUEST, 0, &tmp) == 0); cap_invpcid = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED2_ENABLE_INVPCID, 0, &tmp) == 0); /* * Check support for virtual interrupt delivery. */ procbased2_vid_bits = (PROCBASED2_VIRTUALIZE_APIC_ACCESSES | PROCBASED2_VIRTUALIZE_X2APIC_MODE | PROCBASED2_APIC_REGISTER_VIRTUALIZATION | PROCBASED2_VIRTUAL_INTERRUPT_DELIVERY); use_tpr_shadow = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_USE_TPR_SHADOW, 0, &tmp) == 0); error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, procbased2_vid_bits, 0, &tmp); if (error == 0 && use_tpr_shadow) { virtual_interrupt_delivery = 1; TUNABLE_INT_FETCH("hw.vmm.vmx.use_apic_vid", &virtual_interrupt_delivery); } if (virtual_interrupt_delivery) { procbased_ctls |= PROCBASED_USE_TPR_SHADOW; procbased_ctls2 |= procbased2_vid_bits; procbased_ctls2 &= ~PROCBASED2_VIRTUALIZE_X2APIC_MODE; /* * No need to emulate accesses to %CR8 if virtual * interrupt delivery is enabled. */ procbased_ctls &= ~PROCBASED_CR8_LOAD_EXITING; procbased_ctls &= ~PROCBASED_CR8_STORE_EXITING; /* * Check for Posted Interrupts only if Virtual Interrupt * Delivery is enabled. */ error = vmx_set_ctlreg(MSR_VMX_PINBASED_CTLS, MSR_VMX_TRUE_PINBASED_CTLS, PINBASED_POSTED_INTERRUPT, 0, &tmp); if (error == 0) { pirvec = lapic_ipi_alloc(pti ? &IDTVEC(justreturn1_pti) : &IDTVEC(justreturn)); if (pirvec < 0) { if (bootverbose) { printf("vmx_init: unable to allocate " "posted interrupt vector\n"); } } else { posted_interrupts = 1; TUNABLE_INT_FETCH("hw.vmm.vmx.use_apic_pir", &posted_interrupts); } } } if (posted_interrupts) pinbased_ctls |= PINBASED_POSTED_INTERRUPT; /* Initialize EPT */ error = ept_init(ipinum); if (error) { printf("vmx_init: ept initialization failed (%d)\n", error); return (error); } guest_l1d_flush = (cpu_ia32_arch_caps & IA32_ARCH_CAP_SKIP_L1DFL_VMENTRY) == 0; TUNABLE_INT_FETCH("hw.vmm.l1d_flush", &guest_l1d_flush); /* * L1D cache flush is enabled. Use IA32_FLUSH_CMD MSR when * available. Otherwise fall back to the software flush * method which loads enough data from the kernel text to * flush existing L1D content, both on VMX entry and on NMI * return. */ if (guest_l1d_flush) { if ((cpu_stdext_feature3 & CPUID_STDEXT3_L1D_FLUSH) == 0) { guest_l1d_flush_sw = 1; TUNABLE_INT_FETCH("hw.vmm.l1d_flush_sw", &guest_l1d_flush_sw); } if (guest_l1d_flush_sw) { if (nmi_flush_l1d_sw <= 1) nmi_flush_l1d_sw = 1; } else { msr_load_list[0].index = MSR_IA32_FLUSH_CMD; msr_load_list[0].val = IA32_FLUSH_CMD_L1D; } } /* * Stash the cr0 and cr4 bits that must be fixed to 0 or 1 */ fixed0 = rdmsr(MSR_VMX_CR0_FIXED0); fixed1 = rdmsr(MSR_VMX_CR0_FIXED1); cr0_ones_mask = fixed0 & fixed1; cr0_zeros_mask = ~fixed0 & ~fixed1; /* * CR0_PE and CR0_PG can be set to zero in VMX non-root operation * if unrestricted guest execution is allowed. */ if (cap_unrestricted_guest) cr0_ones_mask &= ~(CR0_PG | CR0_PE); /* * Do not allow the guest to set CR0_NW or CR0_CD. */ cr0_zeros_mask |= (CR0_NW | CR0_CD); fixed0 = rdmsr(MSR_VMX_CR4_FIXED0); fixed1 = rdmsr(MSR_VMX_CR4_FIXED1); cr4_ones_mask = fixed0 & fixed1; cr4_zeros_mask = ~fixed0 & ~fixed1; vpid_init(); vmx_msr_init(); /* enable VMX operation */ smp_rendezvous(NULL, vmx_enable, NULL, NULL); vmx_initialized = 1; return (0); } static void vmx_trigger_hostintr(int vector) { uintptr_t func; struct gate_descriptor *gd; gd = &idt[vector]; KASSERT(vector >= 32 && vector <= 255, ("vmx_trigger_hostintr: " "invalid vector %d", vector)); KASSERT(gd->gd_p == 1, ("gate descriptor for vector %d not present", vector)); KASSERT(gd->gd_type == SDT_SYSIGT, ("gate descriptor for vector %d " "has invalid type %d", vector, gd->gd_type)); KASSERT(gd->gd_dpl == SEL_KPL, ("gate descriptor for vector %d " "has invalid dpl %d", vector, gd->gd_dpl)); KASSERT(gd->gd_selector == GSEL(GCODE_SEL, SEL_KPL), ("gate descriptor " "for vector %d has invalid selector %d", vector, gd->gd_selector)); KASSERT(gd->gd_ist == 0, ("gate descriptor for vector %d has invalid " "IST %d", vector, gd->gd_ist)); func = ((long)gd->gd_hioffset << 16 | gd->gd_looffset); vmx_call_isr(func); } static int vmx_setup_cr_shadow(int which, struct vmcs *vmcs, uint32_t initial) { int error, mask_ident, shadow_ident; uint64_t mask_value; if (which != 0 && which != 4) panic("vmx_setup_cr_shadow: unknown cr%d", which); if (which == 0) { mask_ident = VMCS_CR0_MASK; mask_value = cr0_ones_mask | cr0_zeros_mask; shadow_ident = VMCS_CR0_SHADOW; } else { mask_ident = VMCS_CR4_MASK; mask_value = cr4_ones_mask | cr4_zeros_mask; shadow_ident = VMCS_CR4_SHADOW; } error = vmcs_setreg(vmcs, 0, VMCS_IDENT(mask_ident), mask_value); if (error) return (error); error = vmcs_setreg(vmcs, 0, VMCS_IDENT(shadow_ident), initial); if (error) return (error); return (0); } #define vmx_setup_cr0_shadow(vmcs,init) vmx_setup_cr_shadow(0, (vmcs), (init)) #define vmx_setup_cr4_shadow(vmcs,init) vmx_setup_cr_shadow(4, (vmcs), (init)) static void * vmx_vminit(struct vm *vm, pmap_t pmap) { uint16_t vpid[VM_MAXCPU]; int i, error; struct vmx *vmx; struct vmcs *vmcs; uint32_t exc_bitmap; uint16_t maxcpus; vmx = malloc(sizeof(struct vmx), M_VMX, M_WAITOK | M_ZERO); if ((uintptr_t)vmx & PAGE_MASK) { panic("malloc of struct vmx not aligned on %d byte boundary", PAGE_SIZE); } vmx->vm = vm; vmx->eptp = eptp(vtophys((vm_offset_t)pmap->pm_pml4)); /* * Clean up EPTP-tagged guest physical and combined mappings * * VMX transitions are not required to invalidate any guest physical * mappings. So, it may be possible for stale guest physical mappings * to be present in the processor TLBs. * * Combined mappings for this EP4TA are also invalidated for all VPIDs. */ ept_invalidate_mappings(vmx->eptp); msr_bitmap_initialize(vmx->msr_bitmap); /* * It is safe to allow direct access to MSR_GSBASE and MSR_FSBASE. * The guest FSBASE and GSBASE are saved and restored during * vm-exit and vm-entry respectively. The host FSBASE and GSBASE are * always restored from the vmcs host state area on vm-exit. * * The SYSENTER_CS/ESP/EIP MSRs are identical to FS/GSBASE in * how they are saved/restored so can be directly accessed by the * guest. * * MSR_EFER is saved and restored in the guest VMCS area on a * VM exit and entry respectively. It is also restored from the * host VMCS area on a VM exit. * * The TSC MSR is exposed read-only. Writes are disallowed as * that will impact the host TSC. If the guest does a write * the "use TSC offsetting" execution control is enabled and the * difference between the host TSC and the guest TSC is written * into the TSC offset in the VMCS. */ if (guest_msr_rw(vmx, MSR_GSBASE) || guest_msr_rw(vmx, MSR_FSBASE) || guest_msr_rw(vmx, MSR_SYSENTER_CS_MSR) || guest_msr_rw(vmx, MSR_SYSENTER_ESP_MSR) || guest_msr_rw(vmx, MSR_SYSENTER_EIP_MSR) || guest_msr_rw(vmx, MSR_EFER) || guest_msr_ro(vmx, MSR_TSC)) panic("vmx_vminit: error setting guest msr access"); vpid_alloc(vpid, VM_MAXCPU); if (virtual_interrupt_delivery) { error = vm_map_mmio(vm, DEFAULT_APIC_BASE, PAGE_SIZE, APIC_ACCESS_ADDRESS); /* XXX this should really return an error to the caller */ KASSERT(error == 0, ("vm_map_mmio(apicbase) error %d", error)); } maxcpus = vm_get_maxcpus(vm); for (i = 0; i < maxcpus; i++) { vmcs = &vmx->vmcs[i]; vmcs->identifier = vmx_revision(); error = vmclear(vmcs); if (error != 0) { panic("vmx_vminit: vmclear error %d on vcpu %d\n", error, i); } vmx_msr_guest_init(vmx, i); error = vmcs_init(vmcs); KASSERT(error == 0, ("vmcs_init error %d", error)); VMPTRLD(vmcs); error = 0; error += vmwrite(VMCS_HOST_RSP, (u_long)&vmx->ctx[i]); error += vmwrite(VMCS_EPTP, vmx->eptp); error += vmwrite(VMCS_PIN_BASED_CTLS, pinbased_ctls); error += vmwrite(VMCS_PRI_PROC_BASED_CTLS, procbased_ctls); error += vmwrite(VMCS_SEC_PROC_BASED_CTLS, procbased_ctls2); error += vmwrite(VMCS_EXIT_CTLS, exit_ctls); error += vmwrite(VMCS_ENTRY_CTLS, entry_ctls); error += vmwrite(VMCS_MSR_BITMAP, vtophys(vmx->msr_bitmap)); error += vmwrite(VMCS_VPID, vpid[i]); if (guest_l1d_flush && !guest_l1d_flush_sw) { vmcs_write(VMCS_ENTRY_MSR_LOAD, pmap_kextract( (vm_offset_t)&msr_load_list[0])); vmcs_write(VMCS_ENTRY_MSR_LOAD_COUNT, nitems(msr_load_list)); vmcs_write(VMCS_EXIT_MSR_STORE, 0); vmcs_write(VMCS_EXIT_MSR_STORE_COUNT, 0); } /* exception bitmap */ if (vcpu_trace_exceptions(vm, i)) exc_bitmap = 0xffffffff; else exc_bitmap = 1 << IDT_MC; error += vmwrite(VMCS_EXCEPTION_BITMAP, exc_bitmap); vmx->ctx[i].guest_dr6 = DBREG_DR6_RESERVED1; error += vmwrite(VMCS_GUEST_DR7, DBREG_DR7_RESERVED1); if (virtual_interrupt_delivery) { error += vmwrite(VMCS_APIC_ACCESS, APIC_ACCESS_ADDRESS); error += vmwrite(VMCS_VIRTUAL_APIC, vtophys(&vmx->apic_page[i])); error += vmwrite(VMCS_EOI_EXIT0, 0); error += vmwrite(VMCS_EOI_EXIT1, 0); error += vmwrite(VMCS_EOI_EXIT2, 0); error += vmwrite(VMCS_EOI_EXIT3, 0); } if (posted_interrupts) { error += vmwrite(VMCS_PIR_VECTOR, pirvec); error += vmwrite(VMCS_PIR_DESC, vtophys(&vmx->pir_desc[i])); } VMCLEAR(vmcs); KASSERT(error == 0, ("vmx_vminit: error customizing the vmcs")); vmx->cap[i].set = 0; vmx->cap[i].proc_ctls = procbased_ctls; vmx->cap[i].proc_ctls2 = procbased_ctls2; vmx->cap[i].exc_bitmap = exc_bitmap; vmx->state[i].nextrip = ~0; vmx->state[i].lastcpu = NOCPU; vmx->state[i].vpid = vpid[i]; /* * Set up the CR0/4 shadows, and init the read shadow * to the power-on register value from the Intel Sys Arch. * CR0 - 0x60000010 * CR4 - 0 */ error = vmx_setup_cr0_shadow(vmcs, 0x60000010); if (error != 0) panic("vmx_setup_cr0_shadow %d", error); error = vmx_setup_cr4_shadow(vmcs, 0); if (error != 0) panic("vmx_setup_cr4_shadow %d", error); vmx->ctx[i].pmap = pmap; } return (vmx); } static int vmx_handle_cpuid(struct vm *vm, int vcpu, struct vmxctx *vmxctx) { int handled, func; func = vmxctx->guest_rax; handled = x86_emulate_cpuid(vm, vcpu, (uint32_t*)(&vmxctx->guest_rax), (uint32_t*)(&vmxctx->guest_rbx), (uint32_t*)(&vmxctx->guest_rcx), (uint32_t*)(&vmxctx->guest_rdx)); return (handled); } static __inline void vmx_run_trace(struct vmx *vmx, int vcpu) { #ifdef KTR VCPU_CTR1(vmx->vm, vcpu, "Resume execution at %#lx", vmcs_guest_rip()); #endif } static __inline void vmx_exit_trace(struct vmx *vmx, int vcpu, uint64_t rip, uint32_t exit_reason, int handled) { #ifdef KTR VCPU_CTR3(vmx->vm, vcpu, "%s %s vmexit at 0x%0lx", handled ? "handled" : "unhandled", exit_reason_to_str(exit_reason), rip); #endif } static __inline void vmx_astpending_trace(struct vmx *vmx, int vcpu, uint64_t rip) { #ifdef KTR VCPU_CTR1(vmx->vm, vcpu, "astpending vmexit at 0x%0lx", rip); #endif } static VMM_STAT_INTEL(VCPU_INVVPID_SAVED, "Number of vpid invalidations saved"); static VMM_STAT_INTEL(VCPU_INVVPID_DONE, "Number of vpid invalidations done"); /* * Invalidate guest mappings identified by its vpid from the TLB. */ static __inline void vmx_invvpid(struct vmx *vmx, int vcpu, pmap_t pmap, int running) { struct vmxstate *vmxstate; struct invvpid_desc invvpid_desc; vmxstate = &vmx->state[vcpu]; if (vmxstate->vpid == 0) return; if (!running) { /* * Set the 'lastcpu' to an invalid host cpu. * * This will invalidate TLB entries tagged with the vcpu's * vpid the next time it runs via vmx_set_pcpu_defaults(). */ vmxstate->lastcpu = NOCPU; return; } KASSERT(curthread->td_critnest > 0, ("%s: vcpu %d running outside " "critical section", __func__, vcpu)); /* * Invalidate all mappings tagged with 'vpid' * * We do this because this vcpu was executing on a different host * cpu when it last ran. We do not track whether it invalidated * mappings associated with its 'vpid' during that run. So we must * assume that the mappings associated with 'vpid' on 'curcpu' are * stale and invalidate them. * * Note that we incur this penalty only when the scheduler chooses to * move the thread associated with this vcpu between host cpus. * * Note also that this will invalidate mappings tagged with 'vpid' * for "all" EP4TAs. */ if (pmap->pm_eptgen == vmx->eptgen[curcpu]) { invvpid_desc._res1 = 0; invvpid_desc._res2 = 0; invvpid_desc.vpid = vmxstate->vpid; invvpid_desc.linear_addr = 0; invvpid(INVVPID_TYPE_SINGLE_CONTEXT, invvpid_desc); vmm_stat_incr(vmx->vm, vcpu, VCPU_INVVPID_DONE, 1); } else { /* * The invvpid can be skipped if an invept is going to * be performed before entering the guest. The invept * will invalidate combined mappings tagged with * 'vmx->eptp' for all vpids. */ vmm_stat_incr(vmx->vm, vcpu, VCPU_INVVPID_SAVED, 1); } } static void vmx_set_pcpu_defaults(struct vmx *vmx, int vcpu, pmap_t pmap) { struct vmxstate *vmxstate; vmxstate = &vmx->state[vcpu]; if (vmxstate->lastcpu == curcpu) return; vmxstate->lastcpu = curcpu; vmm_stat_incr(vmx->vm, vcpu, VCPU_MIGRATIONS, 1); vmcs_write(VMCS_HOST_TR_BASE, vmm_get_host_trbase()); vmcs_write(VMCS_HOST_GDTR_BASE, vmm_get_host_gdtrbase()); vmcs_write(VMCS_HOST_GS_BASE, vmm_get_host_gsbase()); vmx_invvpid(vmx, vcpu, pmap, 1); } /* * We depend on 'procbased_ctls' to have the Interrupt Window Exiting bit set. */ CTASSERT((PROCBASED_CTLS_ONE_SETTING & PROCBASED_INT_WINDOW_EXITING) != 0); static void __inline vmx_set_int_window_exiting(struct vmx *vmx, int vcpu) { if ((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) == 0) { vmx->cap[vcpu].proc_ctls |= PROCBASED_INT_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Enabling interrupt window exiting"); } } static void __inline vmx_clear_int_window_exiting(struct vmx *vmx, int vcpu) { KASSERT((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0, ("intr_window_exiting not set: %#x", vmx->cap[vcpu].proc_ctls)); vmx->cap[vcpu].proc_ctls &= ~PROCBASED_INT_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Disabling interrupt window exiting"); } static void __inline vmx_set_nmi_window_exiting(struct vmx *vmx, int vcpu) { if ((vmx->cap[vcpu].proc_ctls & PROCBASED_NMI_WINDOW_EXITING) == 0) { vmx->cap[vcpu].proc_ctls |= PROCBASED_NMI_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Enabling NMI window exiting"); } } static void __inline vmx_clear_nmi_window_exiting(struct vmx *vmx, int vcpu) { KASSERT((vmx->cap[vcpu].proc_ctls & PROCBASED_NMI_WINDOW_EXITING) != 0, ("nmi_window_exiting not set %#x", vmx->cap[vcpu].proc_ctls)); vmx->cap[vcpu].proc_ctls &= ~PROCBASED_NMI_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Disabling NMI window exiting"); } int vmx_set_tsc_offset(struct vmx *vmx, int vcpu, uint64_t offset) { int error; if ((vmx->cap[vcpu].proc_ctls & PROCBASED_TSC_OFFSET) == 0) { vmx->cap[vcpu].proc_ctls |= PROCBASED_TSC_OFFSET; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Enabling TSC offsetting"); } error = vmwrite(VMCS_TSC_OFFSET, offset); return (error); } #define NMI_BLOCKING (VMCS_INTERRUPTIBILITY_NMI_BLOCKING | \ VMCS_INTERRUPTIBILITY_MOVSS_BLOCKING) #define HWINTR_BLOCKING (VMCS_INTERRUPTIBILITY_STI_BLOCKING | \ VMCS_INTERRUPTIBILITY_MOVSS_BLOCKING) static void vmx_inject_nmi(struct vmx *vmx, int vcpu) { uint32_t gi, info; gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); KASSERT((gi & NMI_BLOCKING) == 0, ("vmx_inject_nmi: invalid guest " "interruptibility-state %#x", gi)); info = vmcs_read(VMCS_ENTRY_INTR_INFO); KASSERT((info & VMCS_INTR_VALID) == 0, ("vmx_inject_nmi: invalid " "VM-entry interruption information %#x", info)); /* * Inject the virtual NMI. The vector must be the NMI IDT entry * or the VMCS entry check will fail. */ info = IDT_NMI | VMCS_INTR_T_NMI | VMCS_INTR_VALID; vmcs_write(VMCS_ENTRY_INTR_INFO, info); VCPU_CTR0(vmx->vm, vcpu, "Injecting vNMI"); /* Clear the request */ vm_nmi_clear(vmx->vm, vcpu); } static void vmx_inject_interrupts(struct vmx *vmx, int vcpu, struct vlapic *vlapic, uint64_t guestrip) { int vector, need_nmi_exiting, extint_pending; uint64_t rflags, entryinfo; uint32_t gi, info; if (vmx->state[vcpu].nextrip != guestrip) { gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if (gi & HWINTR_BLOCKING) { VCPU_CTR2(vmx->vm, vcpu, "Guest interrupt blocking " "cleared due to rip change: %#lx/%#lx", vmx->state[vcpu].nextrip, guestrip); gi &= ~HWINTR_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } } if (vm_entry_intinfo(vmx->vm, vcpu, &entryinfo)) { KASSERT((entryinfo & VMCS_INTR_VALID) != 0, ("%s: entry " "intinfo is not valid: %#lx", __func__, entryinfo)); info = vmcs_read(VMCS_ENTRY_INTR_INFO); KASSERT((info & VMCS_INTR_VALID) == 0, ("%s: cannot inject " "pending exception: %#lx/%#x", __func__, entryinfo, info)); info = entryinfo; vector = info & 0xff; if (vector == IDT_BP || vector == IDT_OF) { /* * VT-x requires #BP and #OF to be injected as software * exceptions. */ info &= ~VMCS_INTR_T_MASK; info |= VMCS_INTR_T_SWEXCEPTION; } if (info & VMCS_INTR_DEL_ERRCODE) vmcs_write(VMCS_ENTRY_EXCEPTION_ERROR, entryinfo >> 32); vmcs_write(VMCS_ENTRY_INTR_INFO, info); } if (vm_nmi_pending(vmx->vm, vcpu)) { /* * If there are no conditions blocking NMI injection then * inject it directly here otherwise enable "NMI window * exiting" to inject it as soon as we can. * * We also check for STI_BLOCKING because some implementations * don't allow NMI injection in this case. If we are running * on a processor that doesn't have this restriction it will * immediately exit and the NMI will be injected in the * "NMI window exiting" handler. */ need_nmi_exiting = 1; gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if ((gi & (HWINTR_BLOCKING | NMI_BLOCKING)) == 0) { info = vmcs_read(VMCS_ENTRY_INTR_INFO); if ((info & VMCS_INTR_VALID) == 0) { vmx_inject_nmi(vmx, vcpu); need_nmi_exiting = 0; } else { VCPU_CTR1(vmx->vm, vcpu, "Cannot inject NMI " "due to VM-entry intr info %#x", info); } } else { VCPU_CTR1(vmx->vm, vcpu, "Cannot inject NMI due to " "Guest Interruptibility-state %#x", gi); } if (need_nmi_exiting) vmx_set_nmi_window_exiting(vmx, vcpu); } extint_pending = vm_extint_pending(vmx->vm, vcpu); if (!extint_pending && virtual_interrupt_delivery) { vmx_inject_pir(vlapic); return; } /* * If interrupt-window exiting is already in effect then don't bother * checking for pending interrupts. This is just an optimization and * not needed for correctness. */ if ((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0) { VCPU_CTR0(vmx->vm, vcpu, "Skip interrupt injection due to " "pending int_window_exiting"); return; } if (!extint_pending) { /* Ask the local apic for a vector to inject */ if (!vlapic_pending_intr(vlapic, &vector)) return; /* * From the Intel SDM, Volume 3, Section "Maskable * Hardware Interrupts": * - maskable interrupt vectors [16,255] can be delivered * through the local APIC. */ KASSERT(vector >= 16 && vector <= 255, ("invalid vector %d from local APIC", vector)); } else { /* Ask the legacy pic for a vector to inject */ vatpic_pending_intr(vmx->vm, &vector); /* * From the Intel SDM, Volume 3, Section "Maskable * Hardware Interrupts": * - maskable interrupt vectors [0,255] can be delivered * through the INTR pin. */ KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR", vector)); } /* Check RFLAGS.IF and the interruptibility state of the guest */ rflags = vmcs_read(VMCS_GUEST_RFLAGS); if ((rflags & PSL_I) == 0) { VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to " "rflags %#lx", vector, rflags); goto cantinject; } gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if (gi & HWINTR_BLOCKING) { VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to " "Guest Interruptibility-state %#x", vector, gi); goto cantinject; } info = vmcs_read(VMCS_ENTRY_INTR_INFO); if (info & VMCS_INTR_VALID) { /* * This is expected and could happen for multiple reasons: * - A vectoring VM-entry was aborted due to astpending * - A VM-exit happened during event injection. * - An exception was injected above. * - An NMI was injected above or after "NMI window exiting" */ VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to " "VM-entry intr info %#x", vector, info); goto cantinject; } /* Inject the interrupt */ info = VMCS_INTR_T_HWINTR | VMCS_INTR_VALID; info |= vector; vmcs_write(VMCS_ENTRY_INTR_INFO, info); if (!extint_pending) { /* Update the Local APIC ISR */ vlapic_intr_accepted(vlapic, vector); } else { vm_extint_clear(vmx->vm, vcpu); vatpic_intr_accepted(vmx->vm, vector); /* * After we accepted the current ExtINT the PIC may * have posted another one. If that is the case, set * the Interrupt Window Exiting execution control so * we can inject that one too. * * Also, interrupt window exiting allows us to inject any * pending APIC vector that was preempted by the ExtINT * as soon as possible. This applies both for the software * emulated vlapic and the hardware assisted virtual APIC. */ vmx_set_int_window_exiting(vmx, vcpu); } VCPU_CTR1(vmx->vm, vcpu, "Injecting hwintr at vector %d", vector); return; cantinject: /* * Set the Interrupt Window Exiting execution control so we can inject * the interrupt as soon as blocking condition goes away. */ vmx_set_int_window_exiting(vmx, vcpu); } /* * If the Virtual NMIs execution control is '1' then the logical processor * tracks virtual-NMI blocking in the Guest Interruptibility-state field of * the VMCS. An IRET instruction in VMX non-root operation will remove any * virtual-NMI blocking. * * This unblocking occurs even if the IRET causes a fault. In this case the * hypervisor needs to restore virtual-NMI blocking before resuming the guest. */ static void vmx_restore_nmi_blocking(struct vmx *vmx, int vcpuid) { uint32_t gi; VCPU_CTR0(vmx->vm, vcpuid, "Restore Virtual-NMI blocking"); gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); gi |= VMCS_INTERRUPTIBILITY_NMI_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } static void vmx_clear_nmi_blocking(struct vmx *vmx, int vcpuid) { uint32_t gi; VCPU_CTR0(vmx->vm, vcpuid, "Clear Virtual-NMI blocking"); gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); gi &= ~VMCS_INTERRUPTIBILITY_NMI_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } static void vmx_assert_nmi_blocking(struct vmx *vmx, int vcpuid) { uint32_t gi; gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); KASSERT(gi & VMCS_INTERRUPTIBILITY_NMI_BLOCKING, ("NMI blocking is not in effect %#x", gi)); } static int vmx_emulate_xsetbv(struct vmx *vmx, int vcpu, struct vm_exit *vmexit) { struct vmxctx *vmxctx; uint64_t xcrval; const struct xsave_limits *limits; vmxctx = &vmx->ctx[vcpu]; limits = vmm_get_xsave_limits(); /* * Note that the processor raises a GP# fault on its own if * xsetbv is executed for CPL != 0, so we do not have to * emulate that fault here. */ /* Only xcr0 is supported. */ if (vmxctx->guest_rcx != 0) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* We only handle xcr0 if both the host and guest have XSAVE enabled. */ if (!limits->xsave_enabled || !(vmcs_read(VMCS_GUEST_CR4) & CR4_XSAVE)) { vm_inject_ud(vmx->vm, vcpu); return (HANDLED); } xcrval = vmxctx->guest_rdx << 32 | (vmxctx->guest_rax & 0xffffffff); if ((xcrval & ~limits->xcr0_allowed) != 0) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } if (!(xcrval & XFEATURE_ENABLED_X87)) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* AVX (YMM_Hi128) requires SSE. */ if (xcrval & XFEATURE_ENABLED_AVX && (xcrval & XFEATURE_AVX) != XFEATURE_AVX) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* * AVX512 requires base AVX (YMM_Hi128) as well as OpMask, * ZMM_Hi256, and Hi16_ZMM. */ if (xcrval & XFEATURE_AVX512 && (xcrval & (XFEATURE_AVX512 | XFEATURE_AVX)) != (XFEATURE_AVX512 | XFEATURE_AVX)) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* * Intel MPX requires both bound register state flags to be * set. */ if (((xcrval & XFEATURE_ENABLED_BNDREGS) != 0) != ((xcrval & XFEATURE_ENABLED_BNDCSR) != 0)) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* * This runs "inside" vmrun() with the guest's FPU state, so * modifying xcr0 directly modifies the guest's xcr0, not the * host's. */ load_xcr(0, xcrval); return (HANDLED); } static uint64_t vmx_get_guest_reg(struct vmx *vmx, int vcpu, int ident) { const struct vmxctx *vmxctx; vmxctx = &vmx->ctx[vcpu]; switch (ident) { case 0: return (vmxctx->guest_rax); case 1: return (vmxctx->guest_rcx); case 2: return (vmxctx->guest_rdx); case 3: return (vmxctx->guest_rbx); case 4: return (vmcs_read(VMCS_GUEST_RSP)); case 5: return (vmxctx->guest_rbp); case 6: return (vmxctx->guest_rsi); case 7: return (vmxctx->guest_rdi); case 8: return (vmxctx->guest_r8); case 9: return (vmxctx->guest_r9); case 10: return (vmxctx->guest_r10); case 11: return (vmxctx->guest_r11); case 12: return (vmxctx->guest_r12); case 13: return (vmxctx->guest_r13); case 14: return (vmxctx->guest_r14); case 15: return (vmxctx->guest_r15); default: panic("invalid vmx register %d", ident); } } static void vmx_set_guest_reg(struct vmx *vmx, int vcpu, int ident, uint64_t regval) { struct vmxctx *vmxctx; vmxctx = &vmx->ctx[vcpu]; switch (ident) { case 0: vmxctx->guest_rax = regval; break; case 1: vmxctx->guest_rcx = regval; break; case 2: vmxctx->guest_rdx = regval; break; case 3: vmxctx->guest_rbx = regval; break; case 4: vmcs_write(VMCS_GUEST_RSP, regval); break; case 5: vmxctx->guest_rbp = regval; break; case 6: vmxctx->guest_rsi = regval; break; case 7: vmxctx->guest_rdi = regval; break; case 8: vmxctx->guest_r8 = regval; break; case 9: vmxctx->guest_r9 = regval; break; case 10: vmxctx->guest_r10 = regval; break; case 11: vmxctx->guest_r11 = regval; break; case 12: vmxctx->guest_r12 = regval; break; case 13: vmxctx->guest_r13 = regval; break; case 14: vmxctx->guest_r14 = regval; break; case 15: vmxctx->guest_r15 = regval; break; default: panic("invalid vmx register %d", ident); } } static int vmx_emulate_cr0_access(struct vmx *vmx, int vcpu, uint64_t exitqual) { uint64_t crval, regval; /* We only handle mov to %cr0 at this time */ if ((exitqual & 0xf0) != 0x00) return (UNHANDLED); regval = vmx_get_guest_reg(vmx, vcpu, (exitqual >> 8) & 0xf); vmcs_write(VMCS_CR0_SHADOW, regval); crval = regval | cr0_ones_mask; crval &= ~cr0_zeros_mask; vmcs_write(VMCS_GUEST_CR0, crval); if (regval & CR0_PG) { uint64_t efer, entry_ctls; /* * If CR0.PG is 1 and EFER.LME is 1 then EFER.LMA and * the "IA-32e mode guest" bit in VM-entry control must be * equal. */ efer = vmcs_read(VMCS_GUEST_IA32_EFER); if (efer & EFER_LME) { efer |= EFER_LMA; vmcs_write(VMCS_GUEST_IA32_EFER, efer); entry_ctls = vmcs_read(VMCS_ENTRY_CTLS); entry_ctls |= VM_ENTRY_GUEST_LMA; vmcs_write(VMCS_ENTRY_CTLS, entry_ctls); } } return (HANDLED); } static int vmx_emulate_cr4_access(struct vmx *vmx, int vcpu, uint64_t exitqual) { uint64_t crval, regval; /* We only handle mov to %cr4 at this time */ if ((exitqual & 0xf0) != 0x00) return (UNHANDLED); regval = vmx_get_guest_reg(vmx, vcpu, (exitqual >> 8) & 0xf); vmcs_write(VMCS_CR4_SHADOW, regval); crval = regval | cr4_ones_mask; crval &= ~cr4_zeros_mask; vmcs_write(VMCS_GUEST_CR4, crval); return (HANDLED); } static int vmx_emulate_cr8_access(struct vmx *vmx, int vcpu, uint64_t exitqual) { struct vlapic *vlapic; uint64_t cr8; int regnum; /* We only handle mov %cr8 to/from a register at this time. */ if ((exitqual & 0xe0) != 0x00) { return (UNHANDLED); } vlapic = vm_lapic(vmx->vm, vcpu); regnum = (exitqual >> 8) & 0xf; if (exitqual & 0x10) { cr8 = vlapic_get_cr8(vlapic); vmx_set_guest_reg(vmx, vcpu, regnum, cr8); } else { cr8 = vmx_get_guest_reg(vmx, vcpu, regnum); vlapic_set_cr8(vlapic, cr8); } return (HANDLED); } /* * From section "Guest Register State" in the Intel SDM: CPL = SS.DPL */ static int vmx_cpl(void) { uint32_t ssar; ssar = vmcs_read(VMCS_GUEST_SS_ACCESS_RIGHTS); return ((ssar >> 5) & 0x3); } static enum vm_cpu_mode vmx_cpu_mode(void) { uint32_t csar; if (vmcs_read(VMCS_GUEST_IA32_EFER) & EFER_LMA) { csar = vmcs_read(VMCS_GUEST_CS_ACCESS_RIGHTS); if (csar & 0x2000) return (CPU_MODE_64BIT); /* CS.L = 1 */ else return (CPU_MODE_COMPATIBILITY); } else if (vmcs_read(VMCS_GUEST_CR0) & CR0_PE) { return (CPU_MODE_PROTECTED); } else { return (CPU_MODE_REAL); } } static enum vm_paging_mode vmx_paging_mode(void) { if (!(vmcs_read(VMCS_GUEST_CR0) & CR0_PG)) return (PAGING_MODE_FLAT); if (!(vmcs_read(VMCS_GUEST_CR4) & CR4_PAE)) return (PAGING_MODE_32); if (vmcs_read(VMCS_GUEST_IA32_EFER) & EFER_LME) return (PAGING_MODE_64); else return (PAGING_MODE_PAE); } static uint64_t inout_str_index(struct vmx *vmx, int vcpuid, int in) { uint64_t val; int error; enum vm_reg_name reg; reg = in ? VM_REG_GUEST_RDI : VM_REG_GUEST_RSI; error = vmx_getreg(vmx, vcpuid, reg, &val); KASSERT(error == 0, ("%s: vmx_getreg error %d", __func__, error)); return (val); } static uint64_t inout_str_count(struct vmx *vmx, int vcpuid, int rep) { uint64_t val; int error; if (rep) { error = vmx_getreg(vmx, vcpuid, VM_REG_GUEST_RCX, &val); KASSERT(!error, ("%s: vmx_getreg error %d", __func__, error)); } else { val = 1; } return (val); } static int inout_str_addrsize(uint32_t inst_info) { uint32_t size; size = (inst_info >> 7) & 0x7; switch (size) { case 0: return (2); /* 16 bit */ case 1: return (4); /* 32 bit */ case 2: return (8); /* 64 bit */ default: panic("%s: invalid size encoding %d", __func__, size); } } static void inout_str_seginfo(struct vmx *vmx, int vcpuid, uint32_t inst_info, int in, struct vm_inout_str *vis) { int error, s; if (in) { vis->seg_name = VM_REG_GUEST_ES; } else { s = (inst_info >> 15) & 0x7; vis->seg_name = vm_segment_name(s); } error = vmx_getdesc(vmx, vcpuid, vis->seg_name, &vis->seg_desc); KASSERT(error == 0, ("%s: vmx_getdesc error %d", __func__, error)); } static void vmx_paging_info(struct vm_guest_paging *paging) { paging->cr3 = vmcs_guest_cr3(); paging->cpl = vmx_cpl(); paging->cpu_mode = vmx_cpu_mode(); paging->paging_mode = vmx_paging_mode(); } static void vmexit_inst_emul(struct vm_exit *vmexit, uint64_t gpa, uint64_t gla) { struct vm_guest_paging *paging; uint32_t csar; paging = &vmexit->u.inst_emul.paging; vmexit->exitcode = VM_EXITCODE_INST_EMUL; vmexit->inst_length = 0; vmexit->u.inst_emul.gpa = gpa; vmexit->u.inst_emul.gla = gla; vmx_paging_info(paging); switch (paging->cpu_mode) { case CPU_MODE_REAL: vmexit->u.inst_emul.cs_base = vmcs_read(VMCS_GUEST_CS_BASE); vmexit->u.inst_emul.cs_d = 0; break; case CPU_MODE_PROTECTED: case CPU_MODE_COMPATIBILITY: vmexit->u.inst_emul.cs_base = vmcs_read(VMCS_GUEST_CS_BASE); csar = vmcs_read(VMCS_GUEST_CS_ACCESS_RIGHTS); vmexit->u.inst_emul.cs_d = SEG_DESC_DEF32(csar); break; default: vmexit->u.inst_emul.cs_base = 0; vmexit->u.inst_emul.cs_d = 0; break; } vie_init(&vmexit->u.inst_emul.vie, NULL, 0); } static int ept_fault_type(uint64_t ept_qual) { int fault_type; if (ept_qual & EPT_VIOLATION_DATA_WRITE) fault_type = VM_PROT_WRITE; else if (ept_qual & EPT_VIOLATION_INST_FETCH) fault_type = VM_PROT_EXECUTE; else fault_type= VM_PROT_READ; return (fault_type); } static bool ept_emulation_fault(uint64_t ept_qual) { int read, write; /* EPT fault on an instruction fetch doesn't make sense here */ if (ept_qual & EPT_VIOLATION_INST_FETCH) return (false); /* EPT fault must be a read fault or a write fault */ read = ept_qual & EPT_VIOLATION_DATA_READ ? 1 : 0; write = ept_qual & EPT_VIOLATION_DATA_WRITE ? 1 : 0; if ((read | write) == 0) return (false); /* * The EPT violation must have been caused by accessing a * guest-physical address that is a translation of a guest-linear * address. */ if ((ept_qual & EPT_VIOLATION_GLA_VALID) == 0 || (ept_qual & EPT_VIOLATION_XLAT_VALID) == 0) { return (false); } return (true); } static __inline int apic_access_virtualization(struct vmx *vmx, int vcpuid) { uint32_t proc_ctls2; proc_ctls2 = vmx->cap[vcpuid].proc_ctls2; return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_APIC_ACCESSES) ? 1 : 0); } static __inline int x2apic_virtualization(struct vmx *vmx, int vcpuid) { uint32_t proc_ctls2; proc_ctls2 = vmx->cap[vcpuid].proc_ctls2; return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_X2APIC_MODE) ? 1 : 0); } static int vmx_handle_apic_write(struct vmx *vmx, int vcpuid, struct vlapic *vlapic, uint64_t qual) { int error, handled, offset; uint32_t *apic_regs, vector; bool retu; handled = HANDLED; offset = APIC_WRITE_OFFSET(qual); if (!apic_access_virtualization(vmx, vcpuid)) { /* * In general there should not be any APIC write VM-exits * unless APIC-access virtualization is enabled. * * However self-IPI virtualization can legitimately trigger * an APIC-write VM-exit so treat it specially. */ if (x2apic_virtualization(vmx, vcpuid) && offset == APIC_OFFSET_SELF_IPI) { apic_regs = (uint32_t *)(vlapic->apic_page); vector = apic_regs[APIC_OFFSET_SELF_IPI / 4]; vlapic_self_ipi_handler(vlapic, vector); return (HANDLED); } else return (UNHANDLED); } switch (offset) { case APIC_OFFSET_ID: vlapic_id_write_handler(vlapic); break; case APIC_OFFSET_LDR: vlapic_ldr_write_handler(vlapic); break; case APIC_OFFSET_DFR: vlapic_dfr_write_handler(vlapic); break; case APIC_OFFSET_SVR: vlapic_svr_write_handler(vlapic); break; case APIC_OFFSET_ESR: vlapic_esr_write_handler(vlapic); break; case APIC_OFFSET_ICR_LOW: retu = false; error = vlapic_icrlo_write_handler(vlapic, &retu); if (error != 0 || retu) handled = UNHANDLED; break; case APIC_OFFSET_CMCI_LVT: case APIC_OFFSET_TIMER_LVT ... APIC_OFFSET_ERROR_LVT: vlapic_lvt_write_handler(vlapic, offset); break; case APIC_OFFSET_TIMER_ICR: vlapic_icrtmr_write_handler(vlapic); break; case APIC_OFFSET_TIMER_DCR: vlapic_dcr_write_handler(vlapic); break; default: handled = UNHANDLED; break; } return (handled); } static bool apic_access_fault(struct vmx *vmx, int vcpuid, uint64_t gpa) { if (apic_access_virtualization(vmx, vcpuid) && (gpa >= DEFAULT_APIC_BASE && gpa < DEFAULT_APIC_BASE + PAGE_SIZE)) return (true); else return (false); } static int vmx_handle_apic_access(struct vmx *vmx, int vcpuid, struct vm_exit *vmexit) { uint64_t qual; int access_type, offset, allowed; if (!apic_access_virtualization(vmx, vcpuid)) return (UNHANDLED); qual = vmexit->u.vmx.exit_qualification; access_type = APIC_ACCESS_TYPE(qual); offset = APIC_ACCESS_OFFSET(qual); allowed = 0; if (access_type == 0) { /* * Read data access to the following registers is expected. */ switch (offset) { case APIC_OFFSET_APR: case APIC_OFFSET_PPR: case APIC_OFFSET_RRR: case APIC_OFFSET_CMCI_LVT: case APIC_OFFSET_TIMER_CCR: allowed = 1; break; default: break; } } else if (access_type == 1) { /* * Write data access to the following registers is expected. */ switch (offset) { case APIC_OFFSET_VER: case APIC_OFFSET_APR: case APIC_OFFSET_PPR: case APIC_OFFSET_RRR: case APIC_OFFSET_ISR0 ... APIC_OFFSET_ISR7: case APIC_OFFSET_TMR0 ... APIC_OFFSET_TMR7: case APIC_OFFSET_IRR0 ... APIC_OFFSET_IRR7: case APIC_OFFSET_CMCI_LVT: case APIC_OFFSET_TIMER_CCR: allowed = 1; break; default: break; } } if (allowed) { vmexit_inst_emul(vmexit, DEFAULT_APIC_BASE + offset, VIE_INVALID_GLA); } /* * Regardless of whether the APIC-access is allowed this handler * always returns UNHANDLED: * - if the access is allowed then it is handled by emulating the * instruction that caused the VM-exit (outside the critical section) * - if the access is not allowed then it will be converted to an * exitcode of VM_EXITCODE_VMX and will be dealt with in userland. */ return (UNHANDLED); } static enum task_switch_reason vmx_task_switch_reason(uint64_t qual) { int reason; reason = (qual >> 30) & 0x3; switch (reason) { case 0: return (TSR_CALL); case 1: return (TSR_IRET); case 2: return (TSR_JMP); case 3: return (TSR_IDT_GATE); default: panic("%s: invalid reason %d", __func__, reason); } } static int emulate_wrmsr(struct vmx *vmx, int vcpuid, u_int num, uint64_t val, bool *retu) { int error; if (lapic_msr(num)) error = lapic_wrmsr(vmx->vm, vcpuid, num, val, retu); else error = vmx_wrmsr(vmx, vcpuid, num, val, retu); return (error); } static int emulate_rdmsr(struct vmx *vmx, int vcpuid, u_int num, bool *retu) { struct vmxctx *vmxctx; uint64_t result; uint32_t eax, edx; int error; if (lapic_msr(num)) error = lapic_rdmsr(vmx->vm, vcpuid, num, &result, retu); else error = vmx_rdmsr(vmx, vcpuid, num, &result, retu); if (error == 0) { eax = result; vmxctx = &vmx->ctx[vcpuid]; error = vmxctx_setreg(vmxctx, VM_REG_GUEST_RAX, eax); KASSERT(error == 0, ("vmxctx_setreg(rax) error %d", error)); edx = result >> 32; error = vmxctx_setreg(vmxctx, VM_REG_GUEST_RDX, edx); KASSERT(error == 0, ("vmxctx_setreg(rdx) error %d", error)); } return (error); } static int vmx_exit_process(struct vmx *vmx, int vcpu, struct vm_exit *vmexit) { int error, errcode, errcode_valid, handled, in; struct vmxctx *vmxctx; struct vlapic *vlapic; struct vm_inout_str *vis; struct vm_task_switch *ts; uint32_t eax, ecx, edx, idtvec_info, idtvec_err, intr_info, inst_info; uint32_t intr_type, intr_vec, reason; uint64_t exitintinfo, qual, gpa; bool retu; CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_VIRTUAL_NMI) != 0); CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_NMI_EXITING) != 0); handled = UNHANDLED; vmxctx = &vmx->ctx[vcpu]; qual = vmexit->u.vmx.exit_qualification; reason = vmexit->u.vmx.exit_reason; vmexit->exitcode = VM_EXITCODE_BOGUS; vmm_stat_incr(vmx->vm, vcpu, VMEXIT_COUNT, 1); SDT_PROBE3(vmm, vmx, exit, entry, vmx, vcpu, vmexit); /* * VM-entry failures during or after loading guest state. * * These VM-exits are uncommon but must be handled specially * as most VM-exit fields are not populated as usual. */ if (__predict_false(reason == EXIT_REASON_MCE_DURING_ENTRY)) { VCPU_CTR0(vmx->vm, vcpu, "Handling MCE during VM-entry"); __asm __volatile("int $18"); return (1); } /* * VM exits that can be triggered during event delivery need to * be handled specially by re-injecting the event if the IDT * vectoring information field's valid bit is set. * * See "Information for VM Exits During Event Delivery" in Intel SDM * for details. */ idtvec_info = vmcs_idt_vectoring_info(); if (idtvec_info & VMCS_IDT_VEC_VALID) { idtvec_info &= ~(1 << 12); /* clear undefined bit */ exitintinfo = idtvec_info; if (idtvec_info & VMCS_IDT_VEC_ERRCODE_VALID) { idtvec_err = vmcs_idt_vectoring_err(); exitintinfo |= (uint64_t)idtvec_err << 32; } error = vm_exit_intinfo(vmx->vm, vcpu, exitintinfo); KASSERT(error == 0, ("%s: vm_set_intinfo error %d", __func__, error)); /* * If 'virtual NMIs' are being used and the VM-exit * happened while injecting an NMI during the previous * VM-entry, then clear "blocking by NMI" in the * Guest Interruptibility-State so the NMI can be * reinjected on the subsequent VM-entry. * * However, if the NMI was being delivered through a task * gate, then the new task must start execution with NMIs * blocked so don't clear NMI blocking in this case. */ intr_type = idtvec_info & VMCS_INTR_T_MASK; if (intr_type == VMCS_INTR_T_NMI) { if (reason != EXIT_REASON_TASK_SWITCH) vmx_clear_nmi_blocking(vmx, vcpu); else vmx_assert_nmi_blocking(vmx, vcpu); } /* * Update VM-entry instruction length if the event being * delivered was a software interrupt or software exception. */ if (intr_type == VMCS_INTR_T_SWINTR || intr_type == VMCS_INTR_T_PRIV_SWEXCEPTION || intr_type == VMCS_INTR_T_SWEXCEPTION) { vmcs_write(VMCS_ENTRY_INST_LENGTH, vmexit->inst_length); } } switch (reason) { case EXIT_REASON_TASK_SWITCH: ts = &vmexit->u.task_switch; ts->tsssel = qual & 0xffff; ts->reason = vmx_task_switch_reason(qual); ts->ext = 0; ts->errcode_valid = 0; vmx_paging_info(&ts->paging); /* * If the task switch was due to a CALL, JMP, IRET, software * interrupt (INT n) or software exception (INT3, INTO), * then the saved %rip references the instruction that caused * the task switch. The instruction length field in the VMCS * is valid in this case. * * In all other cases (e.g., NMI, hardware exception) the * saved %rip is one that would have been saved in the old TSS * had the task switch completed normally so the instruction * length field is not needed in this case and is explicitly * set to 0. */ if (ts->reason == TSR_IDT_GATE) { KASSERT(idtvec_info & VMCS_IDT_VEC_VALID, ("invalid idtvec_info %#x for IDT task switch", idtvec_info)); intr_type = idtvec_info & VMCS_INTR_T_MASK; if (intr_type != VMCS_INTR_T_SWINTR && intr_type != VMCS_INTR_T_SWEXCEPTION && intr_type != VMCS_INTR_T_PRIV_SWEXCEPTION) { /* Task switch triggered by external event */ ts->ext = 1; vmexit->inst_length = 0; if (idtvec_info & VMCS_IDT_VEC_ERRCODE_VALID) { ts->errcode_valid = 1; ts->errcode = vmcs_idt_vectoring_err(); } } } vmexit->exitcode = VM_EXITCODE_TASK_SWITCH; SDT_PROBE4(vmm, vmx, exit, taskswitch, vmx, vcpu, vmexit, ts); VCPU_CTR4(vmx->vm, vcpu, "task switch reason %d, tss 0x%04x, " "%s errcode 0x%016lx", ts->reason, ts->tsssel, ts->ext ? "external" : "internal", ((uint64_t)ts->errcode << 32) | ts->errcode_valid); break; case EXIT_REASON_CR_ACCESS: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_CR_ACCESS, 1); SDT_PROBE4(vmm, vmx, exit, craccess, vmx, vcpu, vmexit, qual); switch (qual & 0xf) { case 0: handled = vmx_emulate_cr0_access(vmx, vcpu, qual); break; case 4: handled = vmx_emulate_cr4_access(vmx, vcpu, qual); break; case 8: handled = vmx_emulate_cr8_access(vmx, vcpu, qual); break; } break; case EXIT_REASON_RDMSR: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_RDMSR, 1); retu = false; ecx = vmxctx->guest_rcx; VCPU_CTR1(vmx->vm, vcpu, "rdmsr 0x%08x", ecx); SDT_PROBE4(vmm, vmx, exit, rdmsr, vmx, vcpu, vmexit, ecx); error = emulate_rdmsr(vmx, vcpu, ecx, &retu); if (error) { vmexit->exitcode = VM_EXITCODE_RDMSR; vmexit->u.msr.code = ecx; } else if (!retu) { handled = HANDLED; } else { /* Return to userspace with a valid exitcode */ KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, ("emulate_rdmsr retu with bogus exitcode")); } break; case EXIT_REASON_WRMSR: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_WRMSR, 1); retu = false; eax = vmxctx->guest_rax; ecx = vmxctx->guest_rcx; edx = vmxctx->guest_rdx; VCPU_CTR2(vmx->vm, vcpu, "wrmsr 0x%08x value 0x%016lx", ecx, (uint64_t)edx << 32 | eax); SDT_PROBE5(vmm, vmx, exit, wrmsr, vmx, vmexit, vcpu, ecx, (uint64_t)edx << 32 | eax); error = emulate_wrmsr(vmx, vcpu, ecx, (uint64_t)edx << 32 | eax, &retu); if (error) { vmexit->exitcode = VM_EXITCODE_WRMSR; vmexit->u.msr.code = ecx; vmexit->u.msr.wval = (uint64_t)edx << 32 | eax; } else if (!retu) { handled = HANDLED; } else { /* Return to userspace with a valid exitcode */ KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, ("emulate_wrmsr retu with bogus exitcode")); } break; case EXIT_REASON_HLT: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_HLT, 1); SDT_PROBE3(vmm, vmx, exit, halt, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_HLT; vmexit->u.hlt.rflags = vmcs_read(VMCS_GUEST_RFLAGS); if (virtual_interrupt_delivery) vmexit->u.hlt.intr_status = vmcs_read(VMCS_GUEST_INTR_STATUS); else vmexit->u.hlt.intr_status = 0; break; case EXIT_REASON_MTF: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_MTRAP, 1); SDT_PROBE3(vmm, vmx, exit, mtrap, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_MTRAP; vmexit->inst_length = 0; break; case EXIT_REASON_PAUSE: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_PAUSE, 1); SDT_PROBE3(vmm, vmx, exit, pause, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_PAUSE; break; case EXIT_REASON_INTR_WINDOW: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INTR_WINDOW, 1); SDT_PROBE3(vmm, vmx, exit, intrwindow, vmx, vcpu, vmexit); vmx_clear_int_window_exiting(vmx, vcpu); return (1); case EXIT_REASON_EXT_INTR: /* * External interrupts serve only to cause VM exits and allow * the host interrupt handler to run. * * If this external interrupt triggers a virtual interrupt * to a VM, then that state will be recorded by the * host interrupt handler in the VM's softc. We will inject * this virtual interrupt during the subsequent VM enter. */ intr_info = vmcs_read(VMCS_EXIT_INTR_INFO); SDT_PROBE4(vmm, vmx, exit, interrupt, vmx, vcpu, vmexit, intr_info); /* * XXX: Ignore this exit if VMCS_INTR_VALID is not set. * This appears to be a bug in VMware Fusion? */ if (!(intr_info & VMCS_INTR_VALID)) return (1); KASSERT((intr_info & VMCS_INTR_VALID) != 0 && (intr_info & VMCS_INTR_T_MASK) == VMCS_INTR_T_HWINTR, ("VM exit interruption info invalid: %#x", intr_info)); vmx_trigger_hostintr(intr_info & 0xff); /* * This is special. We want to treat this as an 'handled' * VM-exit but not increment the instruction pointer. */ vmm_stat_incr(vmx->vm, vcpu, VMEXIT_EXTINT, 1); return (1); case EXIT_REASON_NMI_WINDOW: SDT_PROBE3(vmm, vmx, exit, nmiwindow, vmx, vcpu, vmexit); /* Exit to allow the pending virtual NMI to be injected */ if (vm_nmi_pending(vmx->vm, vcpu)) vmx_inject_nmi(vmx, vcpu); vmx_clear_nmi_window_exiting(vmx, vcpu); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_NMI_WINDOW, 1); return (1); case EXIT_REASON_INOUT: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INOUT, 1); vmexit->exitcode = VM_EXITCODE_INOUT; vmexit->u.inout.bytes = (qual & 0x7) + 1; vmexit->u.inout.in = in = (qual & 0x8) ? 1 : 0; vmexit->u.inout.string = (qual & 0x10) ? 1 : 0; vmexit->u.inout.rep = (qual & 0x20) ? 1 : 0; vmexit->u.inout.port = (uint16_t)(qual >> 16); vmexit->u.inout.eax = (uint32_t)(vmxctx->guest_rax); if (vmexit->u.inout.string) { inst_info = vmcs_read(VMCS_EXIT_INSTRUCTION_INFO); vmexit->exitcode = VM_EXITCODE_INOUT_STR; vis = &vmexit->u.inout_str; vmx_paging_info(&vis->paging); vis->rflags = vmcs_read(VMCS_GUEST_RFLAGS); vis->cr0 = vmcs_read(VMCS_GUEST_CR0); vis->index = inout_str_index(vmx, vcpu, in); vis->count = inout_str_count(vmx, vcpu, vis->inout.rep); vis->addrsize = inout_str_addrsize(inst_info); inout_str_seginfo(vmx, vcpu, inst_info, in, vis); } SDT_PROBE3(vmm, vmx, exit, inout, vmx, vcpu, vmexit); break; case EXIT_REASON_CPUID: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_CPUID, 1); SDT_PROBE3(vmm, vmx, exit, cpuid, vmx, vcpu, vmexit); handled = vmx_handle_cpuid(vmx->vm, vcpu, vmxctx); break; case EXIT_REASON_EXCEPTION: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_EXCEPTION, 1); intr_info = vmcs_read(VMCS_EXIT_INTR_INFO); KASSERT((intr_info & VMCS_INTR_VALID) != 0, ("VM exit interruption info invalid: %#x", intr_info)); intr_vec = intr_info & 0xff; intr_type = intr_info & VMCS_INTR_T_MASK; /* * If Virtual NMIs control is 1 and the VM-exit is due to a * fault encountered during the execution of IRET then we must * restore the state of "virtual-NMI blocking" before resuming * the guest. * * See "Resuming Guest Software after Handling an Exception". * See "Information for VM Exits Due to Vectored Events". */ if ((idtvec_info & VMCS_IDT_VEC_VALID) == 0 && (intr_vec != IDT_DF) && (intr_info & EXIT_QUAL_NMIUDTI) != 0) vmx_restore_nmi_blocking(vmx, vcpu); /* * The NMI has already been handled in vmx_exit_handle_nmi(). */ if (intr_type == VMCS_INTR_T_NMI) return (1); /* * Call the machine check handler by hand. Also don't reflect * the machine check back into the guest. */ if (intr_vec == IDT_MC) { VCPU_CTR0(vmx->vm, vcpu, "Vectoring to MCE handler"); __asm __volatile("int $18"); return (1); } /* * If the hypervisor has requested user exits for * debug exceptions, bounce them out to userland. */ if (intr_type == VMCS_INTR_T_SWEXCEPTION && intr_vec == IDT_BP && (vmx->cap[vcpu].set & (1 << VM_CAP_BPT_EXIT))) { vmexit->exitcode = VM_EXITCODE_BPT; vmexit->u.bpt.inst_length = vmexit->inst_length; vmexit->inst_length = 0; break; } if (intr_vec == IDT_PF) { error = vmxctx_setreg(vmxctx, VM_REG_GUEST_CR2, qual); KASSERT(error == 0, ("%s: vmxctx_setreg(cr2) error %d", __func__, error)); } /* * Software exceptions exhibit trap-like behavior. This in * turn requires populating the VM-entry instruction length * so that the %rip in the trap frame is past the INT3/INTO * instruction. */ if (intr_type == VMCS_INTR_T_SWEXCEPTION) vmcs_write(VMCS_ENTRY_INST_LENGTH, vmexit->inst_length); /* Reflect all other exceptions back into the guest */ errcode_valid = errcode = 0; if (intr_info & VMCS_INTR_DEL_ERRCODE) { errcode_valid = 1; errcode = vmcs_read(VMCS_EXIT_INTR_ERRCODE); } VCPU_CTR2(vmx->vm, vcpu, "Reflecting exception %d/%#x into " "the guest", intr_vec, errcode); SDT_PROBE5(vmm, vmx, exit, exception, vmx, vcpu, vmexit, intr_vec, errcode); error = vm_inject_exception(vmx->vm, vcpu, intr_vec, errcode_valid, errcode, 0); KASSERT(error == 0, ("%s: vm_inject_exception error %d", __func__, error)); return (1); case EXIT_REASON_EPT_FAULT: /* * If 'gpa' lies within the address space allocated to * memory then this must be a nested page fault otherwise * this must be an instruction that accesses MMIO space. */ gpa = vmcs_gpa(); if (vm_mem_allocated(vmx->vm, vcpu, gpa) || apic_access_fault(vmx, vcpu, gpa)) { vmexit->exitcode = VM_EXITCODE_PAGING; vmexit->inst_length = 0; vmexit->u.paging.gpa = gpa; vmexit->u.paging.fault_type = ept_fault_type(qual); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_NESTED_FAULT, 1); SDT_PROBE5(vmm, vmx, exit, nestedfault, vmx, vcpu, vmexit, gpa, qual); } else if (ept_emulation_fault(qual)) { vmexit_inst_emul(vmexit, gpa, vmcs_gla()); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INST_EMUL, 1); SDT_PROBE4(vmm, vmx, exit, mmiofault, vmx, vcpu, vmexit, gpa); } /* * If Virtual NMIs control is 1 and the VM-exit is due to an * EPT fault during the execution of IRET then we must restore * the state of "virtual-NMI blocking" before resuming. * * See description of "NMI unblocking due to IRET" in * "Exit Qualification for EPT Violations". */ if ((idtvec_info & VMCS_IDT_VEC_VALID) == 0 && (qual & EXIT_QUAL_NMIUDTI) != 0) vmx_restore_nmi_blocking(vmx, vcpu); break; case EXIT_REASON_VIRTUALIZED_EOI: vmexit->exitcode = VM_EXITCODE_IOAPIC_EOI; vmexit->u.ioapic_eoi.vector = qual & 0xFF; SDT_PROBE3(vmm, vmx, exit, eoi, vmx, vcpu, vmexit); vmexit->inst_length = 0; /* trap-like */ break; case EXIT_REASON_APIC_ACCESS: SDT_PROBE3(vmm, vmx, exit, apicaccess, vmx, vcpu, vmexit); handled = vmx_handle_apic_access(vmx, vcpu, vmexit); break; case EXIT_REASON_APIC_WRITE: /* * APIC-write VM exit is trap-like so the %rip is already * pointing to the next instruction. */ vmexit->inst_length = 0; vlapic = vm_lapic(vmx->vm, vcpu); SDT_PROBE4(vmm, vmx, exit, apicwrite, vmx, vcpu, vmexit, vlapic); handled = vmx_handle_apic_write(vmx, vcpu, vlapic, qual); break; case EXIT_REASON_XSETBV: SDT_PROBE3(vmm, vmx, exit, xsetbv, vmx, vcpu, vmexit); handled = vmx_emulate_xsetbv(vmx, vcpu, vmexit); break; case EXIT_REASON_MONITOR: SDT_PROBE3(vmm, vmx, exit, monitor, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_MONITOR; break; case EXIT_REASON_MWAIT: SDT_PROBE3(vmm, vmx, exit, mwait, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_MWAIT; break; case EXIT_REASON_VMCALL: case EXIT_REASON_VMCLEAR: case EXIT_REASON_VMLAUNCH: case EXIT_REASON_VMPTRLD: case EXIT_REASON_VMPTRST: case EXIT_REASON_VMREAD: case EXIT_REASON_VMRESUME: case EXIT_REASON_VMWRITE: case EXIT_REASON_VMXOFF: case EXIT_REASON_VMXON: SDT_PROBE3(vmm, vmx, exit, vminsn, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_VMINSN; break; default: SDT_PROBE4(vmm, vmx, exit, unknown, vmx, vcpu, vmexit, reason); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_UNKNOWN, 1); break; } if (handled) { /* * It is possible that control is returned to userland * even though we were able to handle the VM exit in the * kernel. * * In such a case we want to make sure that the userland * restarts guest execution at the instruction *after* * the one we just processed. Therefore we update the * guest rip in the VMCS and in 'vmexit'. */ vmexit->rip += vmexit->inst_length; vmexit->inst_length = 0; vmcs_write(VMCS_GUEST_RIP, vmexit->rip); } else { if (vmexit->exitcode == VM_EXITCODE_BOGUS) { /* * If this VM exit was not claimed by anybody then * treat it as a generic VMX exit. */ vmexit->exitcode = VM_EXITCODE_VMX; vmexit->u.vmx.status = VM_SUCCESS; vmexit->u.vmx.inst_type = 0; vmexit->u.vmx.inst_error = 0; } else { /* * The exitcode and collateral have been populated. * The VM exit will be processed further in userland. */ } } SDT_PROBE4(vmm, vmx, exit, return, vmx, vcpu, vmexit, handled); return (handled); } static __inline void vmx_exit_inst_error(struct vmxctx *vmxctx, int rc, struct vm_exit *vmexit) { KASSERT(vmxctx->inst_fail_status != VM_SUCCESS, ("vmx_exit_inst_error: invalid inst_fail_status %d", vmxctx->inst_fail_status)); vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_VMX; vmexit->u.vmx.status = vmxctx->inst_fail_status; vmexit->u.vmx.inst_error = vmcs_instruction_error(); vmexit->u.vmx.exit_reason = ~0; vmexit->u.vmx.exit_qualification = ~0; switch (rc) { case VMX_VMRESUME_ERROR: case VMX_VMLAUNCH_ERROR: case VMX_INVEPT_ERROR: vmexit->u.vmx.inst_type = rc; break; default: panic("vm_exit_inst_error: vmx_enter_guest returned %d", rc); } } /* * If the NMI-exiting VM execution control is set to '1' then an NMI in * non-root operation causes a VM-exit. NMI blocking is in effect so it is * sufficient to simply vector to the NMI handler via a software interrupt. * However, this must be done before maskable interrupts are enabled * otherwise the "iret" issued by an interrupt handler will incorrectly * clear NMI blocking. */ static __inline void vmx_exit_handle_nmi(struct vmx *vmx, int vcpuid, struct vm_exit *vmexit) { uint32_t intr_info; KASSERT((read_rflags() & PSL_I) == 0, ("interrupts enabled")); if (vmexit->u.vmx.exit_reason != EXIT_REASON_EXCEPTION) return; intr_info = vmcs_read(VMCS_EXIT_INTR_INFO); KASSERT((intr_info & VMCS_INTR_VALID) != 0, ("VM exit interruption info invalid: %#x", intr_info)); if ((intr_info & VMCS_INTR_T_MASK) == VMCS_INTR_T_NMI) { KASSERT((intr_info & 0xff) == IDT_NMI, ("VM exit due " "to NMI has invalid vector: %#x", intr_info)); VCPU_CTR0(vmx->vm, vcpuid, "Vectoring to NMI handler"); __asm __volatile("int $2"); } } static __inline void vmx_dr_enter_guest(struct vmxctx *vmxctx) { register_t rflags; /* Save host control debug registers. */ vmxctx->host_dr7 = rdr7(); vmxctx->host_debugctl = rdmsr(MSR_DEBUGCTLMSR); /* * Disable debugging in DR7 and DEBUGCTL to avoid triggering * exceptions in the host based on the guest DRx values. The * guest DR7 and DEBUGCTL are saved/restored in the VMCS. */ load_dr7(0); wrmsr(MSR_DEBUGCTLMSR, 0); /* * Disable single stepping the kernel to avoid corrupting the * guest DR6. A debugger might still be able to corrupt the * guest DR6 by setting a breakpoint after this point and then * single stepping. */ rflags = read_rflags(); vmxctx->host_tf = rflags & PSL_T; write_rflags(rflags & ~PSL_T); /* Save host debug registers. */ vmxctx->host_dr0 = rdr0(); vmxctx->host_dr1 = rdr1(); vmxctx->host_dr2 = rdr2(); vmxctx->host_dr3 = rdr3(); vmxctx->host_dr6 = rdr6(); /* Restore guest debug registers. */ load_dr0(vmxctx->guest_dr0); load_dr1(vmxctx->guest_dr1); load_dr2(vmxctx->guest_dr2); load_dr3(vmxctx->guest_dr3); load_dr6(vmxctx->guest_dr6); } static __inline void vmx_dr_leave_guest(struct vmxctx *vmxctx) { /* Save guest debug registers. */ vmxctx->guest_dr0 = rdr0(); vmxctx->guest_dr1 = rdr1(); vmxctx->guest_dr2 = rdr2(); vmxctx->guest_dr3 = rdr3(); vmxctx->guest_dr6 = rdr6(); /* * Restore host debug registers. Restore DR7, DEBUGCTL, and * PSL_T last. */ load_dr0(vmxctx->host_dr0); load_dr1(vmxctx->host_dr1); load_dr2(vmxctx->host_dr2); load_dr3(vmxctx->host_dr3); load_dr6(vmxctx->host_dr6); wrmsr(MSR_DEBUGCTLMSR, vmxctx->host_debugctl); load_dr7(vmxctx->host_dr7); write_rflags(read_rflags() | vmxctx->host_tf); } static int vmx_run(void *arg, int vcpu, register_t rip, pmap_t pmap, struct vm_eventinfo *evinfo) { int rc, handled, launched; struct vmx *vmx; struct vm *vm; struct vmxctx *vmxctx; struct vmcs *vmcs; struct vm_exit *vmexit; struct vlapic *vlapic; uint32_t exit_reason; struct region_descriptor gdtr, idtr; uint16_t ldt_sel; vmx = arg; vm = vmx->vm; vmcs = &vmx->vmcs[vcpu]; vmxctx = &vmx->ctx[vcpu]; vlapic = vm_lapic(vm, vcpu); vmexit = vm_exitinfo(vm, vcpu); launched = 0; KASSERT(vmxctx->pmap == pmap, ("pmap %p different than ctx pmap %p", pmap, vmxctx->pmap)); vmx_msr_guest_enter(vmx, vcpu); VMPTRLD(vmcs); /* * XXX * We do this every time because we may setup the virtual machine * from a different process than the one that actually runs it. * * If the life of a virtual machine was spent entirely in the context * of a single process we could do this once in vmx_vminit(). */ vmcs_write(VMCS_HOST_CR3, rcr3()); vmcs_write(VMCS_GUEST_RIP, rip); vmx_set_pcpu_defaults(vmx, vcpu, pmap); do { KASSERT(vmcs_guest_rip() == rip, ("%s: vmcs guest rip mismatch " "%#lx/%#lx", __func__, vmcs_guest_rip(), rip)); handled = UNHANDLED; /* * Interrupts are disabled from this point on until the * guest starts executing. This is done for the following * reasons: * * If an AST is asserted on this thread after the check below, * then the IPI_AST notification will not be lost, because it * will cause a VM exit due to external interrupt as soon as * the guest state is loaded. * * A posted interrupt after 'vmx_inject_interrupts()' will * not be "lost" because it will be held pending in the host * APIC because interrupts are disabled. The pending interrupt * will be recognized as soon as the guest state is loaded. * * The same reasoning applies to the IPI generated by * pmap_invalidate_ept(). */ disable_intr(); vmx_inject_interrupts(vmx, vcpu, vlapic, rip); /* * Check for vcpu suspension after injecting events because * vmx_inject_interrupts() can suspend the vcpu due to a * triple fault. */ if (vcpu_suspended(evinfo)) { enable_intr(); vm_exit_suspended(vmx->vm, vcpu, rip); break; } if (vcpu_rendezvous_pending(evinfo)) { enable_intr(); vm_exit_rendezvous(vmx->vm, vcpu, rip); break; } if (vcpu_reqidle(evinfo)) { enable_intr(); vm_exit_reqidle(vmx->vm, vcpu, rip); break; } if (vcpu_should_yield(vm, vcpu)) { enable_intr(); vm_exit_astpending(vmx->vm, vcpu, rip); vmx_astpending_trace(vmx, vcpu, rip); handled = HANDLED; break; } if (vcpu_debugged(vm, vcpu)) { enable_intr(); vm_exit_debug(vmx->vm, vcpu, rip); break; } /* * VM exits restore the base address but not the * limits of GDTR and IDTR. The VMCS only stores the * base address, so VM exits set the limits to 0xffff. * Save and restore the full GDTR and IDTR to restore * the limits. * * The VMCS does not save the LDTR at all, and VM * exits clear LDTR as if a NULL selector were loaded. * The userspace hypervisor probably doesn't use a * LDT, but save and restore it to be safe. */ sgdt(&gdtr); sidt(&idtr); ldt_sel = sldt(); vmx_run_trace(vmx, vcpu); vmx_dr_enter_guest(vmxctx); rc = vmx_enter_guest(vmxctx, vmx, launched); vmx_dr_leave_guest(vmxctx); bare_lgdt(&gdtr); lidt(&idtr); lldt(ldt_sel); /* Collect some information for VM exit processing */ vmexit->rip = rip = vmcs_guest_rip(); vmexit->inst_length = vmexit_instruction_length(); vmexit->u.vmx.exit_reason = exit_reason = vmcs_exit_reason(); vmexit->u.vmx.exit_qualification = vmcs_exit_qualification(); /* Update 'nextrip' */ vmx->state[vcpu].nextrip = rip; if (rc == VMX_GUEST_VMEXIT) { vmx_exit_handle_nmi(vmx, vcpu, vmexit); enable_intr(); handled = vmx_exit_process(vmx, vcpu, vmexit); } else { enable_intr(); vmx_exit_inst_error(vmxctx, rc, vmexit); } launched = 1; vmx_exit_trace(vmx, vcpu, rip, exit_reason, handled); rip = vmexit->rip; } while (handled); /* * If a VM exit has been handled then the exitcode must be BOGUS * If a VM exit is not handled then the exitcode must not be BOGUS */ if ((handled && vmexit->exitcode != VM_EXITCODE_BOGUS) || (!handled && vmexit->exitcode == VM_EXITCODE_BOGUS)) { panic("Mismatch between handled (%d) and exitcode (%d)", handled, vmexit->exitcode); } if (!handled) vmm_stat_incr(vm, vcpu, VMEXIT_USERSPACE, 1); VCPU_CTR1(vm, vcpu, "returning from vmx_run: exitcode %d", vmexit->exitcode); VMCLEAR(vmcs); vmx_msr_guest_exit(vmx, vcpu); return (0); } static void vmx_vmcleanup(void *arg) { int i; struct vmx *vmx = arg; uint16_t maxcpus; if (apic_access_virtualization(vmx, 0)) vm_unmap_mmio(vmx->vm, DEFAULT_APIC_BASE, PAGE_SIZE); maxcpus = vm_get_maxcpus(vmx->vm); for (i = 0; i < maxcpus; i++) vpid_free(vmx->state[i].vpid); free(vmx, M_VMX); return; } static register_t * vmxctx_regptr(struct vmxctx *vmxctx, int reg) { switch (reg) { case VM_REG_GUEST_RAX: return (&vmxctx->guest_rax); case VM_REG_GUEST_RBX: return (&vmxctx->guest_rbx); case VM_REG_GUEST_RCX: return (&vmxctx->guest_rcx); case VM_REG_GUEST_RDX: return (&vmxctx->guest_rdx); case VM_REG_GUEST_RSI: return (&vmxctx->guest_rsi); case VM_REG_GUEST_RDI: return (&vmxctx->guest_rdi); case VM_REG_GUEST_RBP: return (&vmxctx->guest_rbp); case VM_REG_GUEST_R8: return (&vmxctx->guest_r8); case VM_REG_GUEST_R9: return (&vmxctx->guest_r9); case VM_REG_GUEST_R10: return (&vmxctx->guest_r10); case VM_REG_GUEST_R11: return (&vmxctx->guest_r11); case VM_REG_GUEST_R12: return (&vmxctx->guest_r12); case VM_REG_GUEST_R13: return (&vmxctx->guest_r13); case VM_REG_GUEST_R14: return (&vmxctx->guest_r14); case VM_REG_GUEST_R15: return (&vmxctx->guest_r15); case VM_REG_GUEST_CR2: return (&vmxctx->guest_cr2); case VM_REG_GUEST_DR0: return (&vmxctx->guest_dr0); case VM_REG_GUEST_DR1: return (&vmxctx->guest_dr1); case VM_REG_GUEST_DR2: return (&vmxctx->guest_dr2); case VM_REG_GUEST_DR3: return (&vmxctx->guest_dr3); case VM_REG_GUEST_DR6: return (&vmxctx->guest_dr6); default: break; } return (NULL); } static int vmxctx_getreg(struct vmxctx *vmxctx, int reg, uint64_t *retval) { register_t *regp; if ((regp = vmxctx_regptr(vmxctx, reg)) != NULL) { *retval = *regp; return (0); } else return (EINVAL); } static int vmxctx_setreg(struct vmxctx *vmxctx, int reg, uint64_t val) { register_t *regp; if ((regp = vmxctx_regptr(vmxctx, reg)) != NULL) { *regp = val; return (0); } else return (EINVAL); } static int vmx_get_intr_shadow(struct vmx *vmx, int vcpu, int running, uint64_t *retval) { uint64_t gi; int error; error = vmcs_getreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(VMCS_GUEST_INTERRUPTIBILITY), &gi); *retval = (gi & HWINTR_BLOCKING) ? 1 : 0; return (error); } static int vmx_modify_intr_shadow(struct vmx *vmx, int vcpu, int running, uint64_t val) { struct vmcs *vmcs; uint64_t gi; int error, ident; /* * Forcing the vcpu into an interrupt shadow is not supported. */ if (val) { error = EINVAL; goto done; } vmcs = &vmx->vmcs[vcpu]; ident = VMCS_IDENT(VMCS_GUEST_INTERRUPTIBILITY); error = vmcs_getreg(vmcs, running, ident, &gi); if (error == 0) { gi &= ~HWINTR_BLOCKING; error = vmcs_setreg(vmcs, running, ident, gi); } done: VCPU_CTR2(vmx->vm, vcpu, "Setting intr_shadow to %#lx %s", val, error ? "failed" : "succeeded"); return (error); } static int vmx_shadow_reg(int reg) { int shreg; shreg = -1; switch (reg) { case VM_REG_GUEST_CR0: shreg = VMCS_CR0_SHADOW; break; case VM_REG_GUEST_CR4: shreg = VMCS_CR4_SHADOW; break; default: break; } return (shreg); } static int vmx_getreg(void *arg, int vcpu, int reg, uint64_t *retval) { int running, hostcpu; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_getreg: %s%d is running", vm_name(vmx->vm), vcpu); if (reg == VM_REG_GUEST_INTR_SHADOW) return (vmx_get_intr_shadow(vmx, vcpu, running, retval)); if (vmxctx_getreg(&vmx->ctx[vcpu], reg, retval) == 0) return (0); return (vmcs_getreg(&vmx->vmcs[vcpu], running, reg, retval)); } static int vmx_setreg(void *arg, int vcpu, int reg, uint64_t val) { int error, hostcpu, running, shadow; uint64_t ctls; pmap_t pmap; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_setreg: %s%d is running", vm_name(vmx->vm), vcpu); if (reg == VM_REG_GUEST_INTR_SHADOW) return (vmx_modify_intr_shadow(vmx, vcpu, running, val)); if (vmxctx_setreg(&vmx->ctx[vcpu], reg, val) == 0) return (0); error = vmcs_setreg(&vmx->vmcs[vcpu], running, reg, val); if (error == 0) { /* * If the "load EFER" VM-entry control is 1 then the * value of EFER.LMA must be identical to "IA-32e mode guest" * bit in the VM-entry control. */ if ((entry_ctls & VM_ENTRY_LOAD_EFER) != 0 && (reg == VM_REG_GUEST_EFER)) { vmcs_getreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(VMCS_ENTRY_CTLS), &ctls); if (val & EFER_LMA) ctls |= VM_ENTRY_GUEST_LMA; else ctls &= ~VM_ENTRY_GUEST_LMA; vmcs_setreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(VMCS_ENTRY_CTLS), ctls); } shadow = vmx_shadow_reg(reg); if (shadow > 0) { /* * Store the unmodified value in the shadow */ error = vmcs_setreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(shadow), val); } if (reg == VM_REG_GUEST_CR3) { /* * Invalidate the guest vcpu's TLB mappings to emulate * the behavior of updating %cr3. * * XXX the processor retains global mappings when %cr3 * is updated but vmx_invvpid() does not. */ pmap = vmx->ctx[vcpu].pmap; vmx_invvpid(vmx, vcpu, pmap, running); } } return (error); } static int vmx_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc) { int hostcpu, running; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_getdesc: %s%d is running", vm_name(vmx->vm), vcpu); return (vmcs_getdesc(&vmx->vmcs[vcpu], running, reg, desc)); } static int vmx_setdesc(void *arg, int vcpu, int reg, struct seg_desc *desc) { int hostcpu, running; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_setdesc: %s%d is running", vm_name(vmx->vm), vcpu); return (vmcs_setdesc(&vmx->vmcs[vcpu], running, reg, desc)); } static int vmx_getcap(void *arg, int vcpu, int type, int *retval) { struct vmx *vmx = arg; int vcap; int ret; ret = ENOENT; vcap = vmx->cap[vcpu].set; switch (type) { case VM_CAP_HALT_EXIT: if (cap_halt_exit) ret = 0; break; case VM_CAP_PAUSE_EXIT: if (cap_pause_exit) ret = 0; break; case VM_CAP_MTRAP_EXIT: if (cap_monitor_trap) ret = 0; break; case VM_CAP_UNRESTRICTED_GUEST: if (cap_unrestricted_guest) ret = 0; break; case VM_CAP_ENABLE_INVPCID: if (cap_invpcid) ret = 0; break; case VM_CAP_BPT_EXIT: ret = 0; break; default: break; } if (ret == 0) *retval = (vcap & (1 << type)) ? 1 : 0; return (ret); } static int vmx_setcap(void *arg, int vcpu, int type, int val) { struct vmx *vmx = arg; struct vmcs *vmcs = &vmx->vmcs[vcpu]; uint32_t baseval; uint32_t *pptr; int error; int flag; int reg; int retval; retval = ENOENT; pptr = NULL; switch (type) { case VM_CAP_HALT_EXIT: if (cap_halt_exit) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls; baseval = *pptr; flag = PROCBASED_HLT_EXITING; reg = VMCS_PRI_PROC_BASED_CTLS; } break; case VM_CAP_MTRAP_EXIT: if (cap_monitor_trap) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls; baseval = *pptr; flag = PROCBASED_MTF; reg = VMCS_PRI_PROC_BASED_CTLS; } break; case VM_CAP_PAUSE_EXIT: if (cap_pause_exit) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls; baseval = *pptr; flag = PROCBASED_PAUSE_EXITING; reg = VMCS_PRI_PROC_BASED_CTLS; } break; case VM_CAP_UNRESTRICTED_GUEST: if (cap_unrestricted_guest) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls2; baseval = *pptr; flag = PROCBASED2_UNRESTRICTED_GUEST; reg = VMCS_SEC_PROC_BASED_CTLS; } break; case VM_CAP_ENABLE_INVPCID: if (cap_invpcid) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls2; baseval = *pptr; flag = PROCBASED2_ENABLE_INVPCID; reg = VMCS_SEC_PROC_BASED_CTLS; } break; case VM_CAP_BPT_EXIT: retval = 0; /* Don't change the bitmap if we are tracing all exceptions. */ if (vmx->cap[vcpu].exc_bitmap != 0xffffffff) { pptr = &vmx->cap[vcpu].exc_bitmap; baseval = *pptr; flag = (1 << IDT_BP); reg = VMCS_EXCEPTION_BITMAP; } break; default: break; } if (retval) return (retval); if (pptr != NULL) { if (val) { baseval |= flag; } else { baseval &= ~flag; } VMPTRLD(vmcs); error = vmwrite(reg, baseval); VMCLEAR(vmcs); if (error) return (error); /* * Update optional stored flags, and record * setting */ *pptr = baseval; } if (val) { vmx->cap[vcpu].set |= (1 << type); } else { vmx->cap[vcpu].set &= ~(1 << type); } return (0); } struct vlapic_vtx { struct vlapic vlapic; struct pir_desc *pir_desc; struct vmx *vmx; u_int pending_prio; }; #define VPR_PRIO_BIT(vpr) (1 << ((vpr) >> 4)) #define VMX_CTR_PIR(vm, vcpuid, pir_desc, notify, vector, level, msg) \ do { \ VCPU_CTR2(vm, vcpuid, msg " assert %s-triggered vector %d", \ level ? "level" : "edge", vector); \ VCPU_CTR1(vm, vcpuid, msg " pir0 0x%016lx", pir_desc->pir[0]); \ VCPU_CTR1(vm, vcpuid, msg " pir1 0x%016lx", pir_desc->pir[1]); \ VCPU_CTR1(vm, vcpuid, msg " pir2 0x%016lx", pir_desc->pir[2]); \ VCPU_CTR1(vm, vcpuid, msg " pir3 0x%016lx", pir_desc->pir[3]); \ VCPU_CTR1(vm, vcpuid, msg " notify: %s", notify ? "yes" : "no");\ } while (0) /* * vlapic->ops handlers that utilize the APICv hardware assist described in * Chapter 29 of the Intel SDM. */ static int vmx_set_intr_ready(struct vlapic *vlapic, int vector, bool level) { struct vlapic_vtx *vlapic_vtx; struct pir_desc *pir_desc; uint64_t mask; int idx, notify = 0; vlapic_vtx = (struct vlapic_vtx *)vlapic; pir_desc = vlapic_vtx->pir_desc; /* * Keep track of interrupt requests in the PIR descriptor. This is * because the virtual APIC page pointed to by the VMCS cannot be * modified if the vcpu is running. */ idx = vector / 64; mask = 1UL << (vector % 64); atomic_set_long(&pir_desc->pir[idx], mask); /* * A notification is required whenever the 'pending' bit makes a * transition from 0->1. * * Even if the 'pending' bit is already asserted, notification about * the incoming interrupt may still be necessary. For example, if a * vCPU is HLTed with a high PPR, a low priority interrupt would cause * the 0->1 'pending' transition with a notification, but the vCPU * would ignore the interrupt for the time being. The same vCPU would * need to then be notified if a high-priority interrupt arrived which * satisfied the PPR. * * The priorities of interrupts injected while 'pending' is asserted * are tracked in a custom bitfield 'pending_prio'. Should the * to-be-injected interrupt exceed the priorities already present, the * notification is sent. The priorities recorded in 'pending_prio' are * cleared whenever the 'pending' bit makes another 0->1 transition. */ if (atomic_cmpset_long(&pir_desc->pending, 0, 1) != 0) { notify = 1; vlapic_vtx->pending_prio = 0; } else { const u_int old_prio = vlapic_vtx->pending_prio; const u_int prio_bit = VPR_PRIO_BIT(vector & APIC_TPR_INT); if ((old_prio & prio_bit) == 0 && prio_bit > old_prio) { atomic_set_int(&vlapic_vtx->pending_prio, prio_bit); notify = 1; } } VMX_CTR_PIR(vlapic->vm, vlapic->vcpuid, pir_desc, notify, vector, level, "vmx_set_intr_ready"); return (notify); } static int vmx_pending_intr(struct vlapic *vlapic, int *vecptr) { struct vlapic_vtx *vlapic_vtx; struct pir_desc *pir_desc; struct LAPIC *lapic; uint64_t pending, pirval; uint32_t ppr, vpr; int i; /* * This function is only expected to be called from the 'HLT' exit * handler which does not care about the vector that is pending. */ KASSERT(vecptr == NULL, ("vmx_pending_intr: vecptr must be NULL")); vlapic_vtx = (struct vlapic_vtx *)vlapic; pir_desc = vlapic_vtx->pir_desc; pending = atomic_load_acq_long(&pir_desc->pending); if (!pending) { /* * While a virtual interrupt may have already been * processed the actual delivery maybe pending the * interruptibility of the guest. Recognize a pending * interrupt by reevaluating virtual interrupts * following Section 29.2.1 in the Intel SDM Volume 3. */ struct vm_exit *vmexit; uint8_t rvi, ppr; vmexit = vm_exitinfo(vlapic->vm, vlapic->vcpuid); KASSERT(vmexit->exitcode == VM_EXITCODE_HLT, ("vmx_pending_intr: exitcode not 'HLT'")); rvi = vmexit->u.hlt.intr_status & APIC_TPR_INT; lapic = vlapic->apic_page; ppr = lapic->ppr & APIC_TPR_INT; if (rvi > ppr) { return (1); } return (0); } /* * If there is an interrupt pending then it will be recognized only * if its priority is greater than the processor priority. * * Special case: if the processor priority is zero then any pending * interrupt will be recognized. */ lapic = vlapic->apic_page; ppr = lapic->ppr & APIC_TPR_INT; if (ppr == 0) return (1); VCPU_CTR1(vlapic->vm, vlapic->vcpuid, "HLT with non-zero PPR %d", lapic->ppr); vpr = 0; for (i = 3; i >= 0; i--) { pirval = pir_desc->pir[i]; if (pirval != 0) { vpr = (i * 64 + flsl(pirval) - 1) & APIC_TPR_INT; break; } } /* * If the highest-priority pending interrupt falls short of the * processor priority of this vCPU, ensure that 'pending_prio' does not * have any stale bits which would preclude a higher-priority interrupt * from incurring a notification later. */ if (vpr <= ppr) { const u_int prio_bit = VPR_PRIO_BIT(vpr); const u_int old = vlapic_vtx->pending_prio; if (old > prio_bit && (old & prio_bit) == 0) { vlapic_vtx->pending_prio = prio_bit; } return (0); } return (1); } static void vmx_intr_accepted(struct vlapic *vlapic, int vector) { panic("vmx_intr_accepted: not expected to be called"); } static void vmx_set_tmr(struct vlapic *vlapic, int vector, bool level) { struct vlapic_vtx *vlapic_vtx; struct vmx *vmx; struct vmcs *vmcs; uint64_t mask, val; KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d", vector)); KASSERT(!vcpu_is_running(vlapic->vm, vlapic->vcpuid, NULL), ("vmx_set_tmr: vcpu cannot be running")); vlapic_vtx = (struct vlapic_vtx *)vlapic; vmx = vlapic_vtx->vmx; vmcs = &vmx->vmcs[vlapic->vcpuid]; mask = 1UL << (vector % 64); VMPTRLD(vmcs); val = vmcs_read(VMCS_EOI_EXIT(vector)); if (level) val |= mask; else val &= ~mask; vmcs_write(VMCS_EOI_EXIT(vector), val); VMCLEAR(vmcs); } static void vmx_enable_x2apic_mode(struct vlapic *vlapic) { struct vmx *vmx; struct vmcs *vmcs; uint32_t proc_ctls2; int vcpuid, error; vcpuid = vlapic->vcpuid; vmx = ((struct vlapic_vtx *)vlapic)->vmx; vmcs = &vmx->vmcs[vcpuid]; proc_ctls2 = vmx->cap[vcpuid].proc_ctls2; KASSERT((proc_ctls2 & PROCBASED2_VIRTUALIZE_APIC_ACCESSES) != 0, ("%s: invalid proc_ctls2 %#x", __func__, proc_ctls2)); proc_ctls2 &= ~PROCBASED2_VIRTUALIZE_APIC_ACCESSES; proc_ctls2 |= PROCBASED2_VIRTUALIZE_X2APIC_MODE; vmx->cap[vcpuid].proc_ctls2 = proc_ctls2; VMPTRLD(vmcs); vmcs_write(VMCS_SEC_PROC_BASED_CTLS, proc_ctls2); VMCLEAR(vmcs); if (vlapic->vcpuid == 0) { /* * The nested page table mappings are shared by all vcpus * so unmap the APIC access page just once. */ error = vm_unmap_mmio(vmx->vm, DEFAULT_APIC_BASE, PAGE_SIZE); KASSERT(error == 0, ("%s: vm_unmap_mmio error %d", __func__, error)); /* * The MSR bitmap is shared by all vcpus so modify it only * once in the context of vcpu 0. */ error = vmx_allow_x2apic_msrs(vmx); KASSERT(error == 0, ("%s: vmx_allow_x2apic_msrs error %d", __func__, error)); } } static void vmx_post_intr(struct vlapic *vlapic, int hostcpu) { ipi_cpu(hostcpu, pirvec); } /* * Transfer the pending interrupts in the PIR descriptor to the IRR * in the virtual APIC page. */ static void vmx_inject_pir(struct vlapic *vlapic) { struct vlapic_vtx *vlapic_vtx; struct pir_desc *pir_desc; struct LAPIC *lapic; uint64_t val, pirval; int rvi, pirbase = -1; uint16_t intr_status_old, intr_status_new; vlapic_vtx = (struct vlapic_vtx *)vlapic; pir_desc = vlapic_vtx->pir_desc; if (atomic_cmpset_long(&pir_desc->pending, 1, 0) == 0) { VCPU_CTR0(vlapic->vm, vlapic->vcpuid, "vmx_inject_pir: " "no posted interrupt pending"); return; } pirval = 0; pirbase = -1; lapic = vlapic->apic_page; val = atomic_readandclear_long(&pir_desc->pir[0]); if (val != 0) { lapic->irr0 |= val; lapic->irr1 |= val >> 32; pirbase = 0; pirval = val; } val = atomic_readandclear_long(&pir_desc->pir[1]); if (val != 0) { lapic->irr2 |= val; lapic->irr3 |= val >> 32; pirbase = 64; pirval = val; } val = atomic_readandclear_long(&pir_desc->pir[2]); if (val != 0) { lapic->irr4 |= val; lapic->irr5 |= val >> 32; pirbase = 128; pirval = val; } val = atomic_readandclear_long(&pir_desc->pir[3]); if (val != 0) { lapic->irr6 |= val; lapic->irr7 |= val >> 32; pirbase = 192; pirval = val; } VLAPIC_CTR_IRR(vlapic, "vmx_inject_pir"); /* * Update RVI so the processor can evaluate pending virtual * interrupts on VM-entry. * * It is possible for pirval to be 0 here, even though the * pending bit has been set. The scenario is: * CPU-Y is sending a posted interrupt to CPU-X, which * is running a guest and processing posted interrupts in h/w. * CPU-X will eventually exit and the state seen in s/w is * the pending bit set, but no PIR bits set. * * CPU-X CPU-Y * (vm running) (host running) * rx posted interrupt * CLEAR pending bit * SET PIR bit * READ/CLEAR PIR bits * SET pending bit * (vm exit) * pending bit set, PIR 0 */ if (pirval != 0) { rvi = pirbase + flsl(pirval) - 1; intr_status_old = vmcs_read(VMCS_GUEST_INTR_STATUS); intr_status_new = (intr_status_old & 0xFF00) | rvi; if (intr_status_new > intr_status_old) { vmcs_write(VMCS_GUEST_INTR_STATUS, intr_status_new); VCPU_CTR2(vlapic->vm, vlapic->vcpuid, "vmx_inject_pir: " "guest_intr_status changed from 0x%04x to 0x%04x", intr_status_old, intr_status_new); } } } static struct vlapic * vmx_vlapic_init(void *arg, int vcpuid) { struct vmx *vmx; struct vlapic *vlapic; struct vlapic_vtx *vlapic_vtx; vmx = arg; vlapic = malloc(sizeof(struct vlapic_vtx), M_VLAPIC, M_WAITOK | M_ZERO); vlapic->vm = vmx->vm; vlapic->vcpuid = vcpuid; vlapic->apic_page = (struct LAPIC *)&vmx->apic_page[vcpuid]; vlapic_vtx = (struct vlapic_vtx *)vlapic; vlapic_vtx->pir_desc = &vmx->pir_desc[vcpuid]; vlapic_vtx->vmx = vmx; if (virtual_interrupt_delivery) { vlapic->ops.set_intr_ready = vmx_set_intr_ready; vlapic->ops.pending_intr = vmx_pending_intr; vlapic->ops.intr_accepted = vmx_intr_accepted; vlapic->ops.set_tmr = vmx_set_tmr; vlapic->ops.enable_x2apic_mode = vmx_enable_x2apic_mode; } if (posted_interrupts) vlapic->ops.post_intr = vmx_post_intr; vlapic_init(vlapic); return (vlapic); } static void vmx_vlapic_cleanup(void *arg, struct vlapic *vlapic) { vlapic_cleanup(vlapic); free(vlapic, M_VLAPIC); } struct vmm_ops vmm_ops_intel = { .init = vmx_init, .cleanup = vmx_cleanup, .resume = vmx_restore, .vminit = vmx_vminit, .vmrun = vmx_run, .vmcleanup = vmx_vmcleanup, .vmgetreg = vmx_getreg, .vmsetreg = vmx_setreg, .vmgetdesc = vmx_getdesc, .vmsetdesc = vmx_setdesc, .vmgetcap = vmx_getcap, .vmsetcap = vmx_setcap, .vmspace_alloc = ept_vmspace_alloc, .vmspace_free = ept_vmspace_free, .vlapic_init = vmx_vlapic_init, .vlapic_cleanup = vmx_vlapic_cleanup, }; Index: head/sys/amd64/vmm/io/iommu.c =================================================================== --- head/sys/amd64/vmm/io/iommu.c (revision 357973) +++ head/sys/amd64/vmm/io/iommu.c (revision 357974) @@ -1,341 +1,342 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include "vmm_util.h" #include "vmm_mem.h" #include "iommu.h" SYSCTL_DECL(_hw_vmm); -SYSCTL_NODE(_hw_vmm, OID_AUTO, iommu, CTLFLAG_RW, 0, "bhyve iommu parameters"); +SYSCTL_NODE(_hw_vmm, OID_AUTO, iommu, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, + "bhyve iommu parameters"); static int iommu_avail; SYSCTL_INT(_hw_vmm_iommu, OID_AUTO, initialized, CTLFLAG_RD, &iommu_avail, 0, "bhyve iommu initialized?"); static int iommu_enable = 1; SYSCTL_INT(_hw_vmm_iommu, OID_AUTO, enable, CTLFLAG_RDTUN, &iommu_enable, 0, "Enable use of I/O MMU (required for PCI passthrough)."); static struct iommu_ops *ops; static void *host_domain; static eventhandler_tag add_tag, delete_tag; static __inline int IOMMU_INIT(void) { if (ops != NULL) return ((*ops->init)()); else return (ENXIO); } static __inline void IOMMU_CLEANUP(void) { if (ops != NULL && iommu_avail) (*ops->cleanup)(); } static __inline void * IOMMU_CREATE_DOMAIN(vm_paddr_t maxaddr) { if (ops != NULL && iommu_avail) return ((*ops->create_domain)(maxaddr)); else return (NULL); } static __inline void IOMMU_DESTROY_DOMAIN(void *dom) { if (ops != NULL && iommu_avail) (*ops->destroy_domain)(dom); } static __inline uint64_t IOMMU_CREATE_MAPPING(void *domain, vm_paddr_t gpa, vm_paddr_t hpa, uint64_t len) { if (ops != NULL && iommu_avail) return ((*ops->create_mapping)(domain, gpa, hpa, len)); else return (len); /* XXX */ } static __inline uint64_t IOMMU_REMOVE_MAPPING(void *domain, vm_paddr_t gpa, uint64_t len) { if (ops != NULL && iommu_avail) return ((*ops->remove_mapping)(domain, gpa, len)); else return (len); /* XXX */ } static __inline void IOMMU_ADD_DEVICE(void *domain, uint16_t rid) { if (ops != NULL && iommu_avail) (*ops->add_device)(domain, rid); } static __inline void IOMMU_REMOVE_DEVICE(void *domain, uint16_t rid) { if (ops != NULL && iommu_avail) (*ops->remove_device)(domain, rid); } static __inline void IOMMU_INVALIDATE_TLB(void *domain) { if (ops != NULL && iommu_avail) (*ops->invalidate_tlb)(domain); } static __inline void IOMMU_ENABLE(void) { if (ops != NULL && iommu_avail) (*ops->enable)(); } static __inline void IOMMU_DISABLE(void) { if (ops != NULL && iommu_avail) (*ops->disable)(); } static void iommu_pci_add(void *arg, device_t dev) { /* Add new devices to the host domain. */ iommu_add_device(host_domain, pci_get_rid(dev)); } static void iommu_pci_delete(void *arg, device_t dev) { iommu_remove_device(host_domain, pci_get_rid(dev)); } static void iommu_init(void) { int error, bus, slot, func; vm_paddr_t maxaddr; devclass_t dc; device_t dev; if (!iommu_enable) return; if (vmm_is_intel()) ops = &iommu_ops_intel; else if (vmm_is_svm()) ops = &iommu_ops_amd; else ops = NULL; error = IOMMU_INIT(); if (error) return; iommu_avail = 1; /* * Create a domain for the devices owned by the host */ maxaddr = vmm_mem_maxaddr(); host_domain = IOMMU_CREATE_DOMAIN(maxaddr); if (host_domain == NULL) { printf("iommu_init: unable to create a host domain"); IOMMU_CLEANUP(); ops = NULL; iommu_avail = 0; return; } /* * Create 1:1 mappings from '0' to 'maxaddr' for devices assigned to * the host */ iommu_create_mapping(host_domain, 0, 0, maxaddr); add_tag = EVENTHANDLER_REGISTER(pci_add_device, iommu_pci_add, NULL, 0); delete_tag = EVENTHANDLER_REGISTER(pci_delete_device, iommu_pci_delete, NULL, 0); dc = devclass_find("ppt"); for (bus = 0; bus <= PCI_BUSMAX; bus++) { for (slot = 0; slot <= PCI_SLOTMAX; slot++) { for (func = 0; func <= PCI_FUNCMAX; func++) { dev = pci_find_dbsf(0, bus, slot, func); if (dev == NULL) continue; /* Skip passthrough devices. */ if (dc != NULL && device_get_devclass(dev) == dc) continue; /* * Everything else belongs to the host * domain. */ iommu_add_device(host_domain, pci_get_rid(dev)); } } } IOMMU_ENABLE(); } void iommu_cleanup(void) { if (add_tag != NULL) { EVENTHANDLER_DEREGISTER(pci_add_device, add_tag); add_tag = NULL; } if (delete_tag != NULL) { EVENTHANDLER_DEREGISTER(pci_delete_device, delete_tag); delete_tag = NULL; } IOMMU_DISABLE(); IOMMU_DESTROY_DOMAIN(host_domain); IOMMU_CLEANUP(); } void * iommu_create_domain(vm_paddr_t maxaddr) { static volatile int iommu_initted; if (iommu_initted < 2) { if (atomic_cmpset_int(&iommu_initted, 0, 1)) { iommu_init(); atomic_store_rel_int(&iommu_initted, 2); } else while (iommu_initted == 1) cpu_spinwait(); } return (IOMMU_CREATE_DOMAIN(maxaddr)); } void iommu_destroy_domain(void *dom) { IOMMU_DESTROY_DOMAIN(dom); } void iommu_create_mapping(void *dom, vm_paddr_t gpa, vm_paddr_t hpa, size_t len) { uint64_t mapped, remaining; remaining = len; while (remaining > 0) { mapped = IOMMU_CREATE_MAPPING(dom, gpa, hpa, remaining); gpa += mapped; hpa += mapped; remaining -= mapped; } } void iommu_remove_mapping(void *dom, vm_paddr_t gpa, size_t len) { uint64_t unmapped, remaining; remaining = len; while (remaining > 0) { unmapped = IOMMU_REMOVE_MAPPING(dom, gpa, remaining); gpa += unmapped; remaining -= unmapped; } } void * iommu_host_domain(void) { return (host_domain); } void iommu_add_device(void *dom, uint16_t rid) { IOMMU_ADD_DEVICE(dom, rid); } void iommu_remove_device(void *dom, uint16_t rid) { IOMMU_REMOVE_DEVICE(dom, rid); } void iommu_invalidate_tlb(void *domain) { IOMMU_INVALIDATE_TLB(domain); } Index: head/sys/amd64/vmm/io/ppt.c =================================================================== --- head/sys/amd64/vmm/io/ppt.c (revision 357973) +++ head/sys/amd64/vmm/io/ppt.c (revision 357974) @@ -1,701 +1,702 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "vmm_lapic.h" #include "vmm_ktr.h" #include "iommu.h" #include "ppt.h" /* XXX locking */ #define MAX_MSIMSGS 32 /* * If the MSI-X table is located in the middle of a BAR then that MMIO * region gets split into two segments - one segment above the MSI-X table * and the other segment below the MSI-X table - with a hole in place of * the MSI-X table so accesses to it can be trapped and emulated. * * So, allocate a MMIO segment for each BAR register + 1 additional segment. */ #define MAX_MMIOSEGS ((PCIR_MAX_BAR_0 + 1) + 1) MALLOC_DEFINE(M_PPTMSIX, "pptmsix", "Passthru MSI-X resources"); struct pptintr_arg { /* pptintr(pptintr_arg) */ struct pptdev *pptdev; uint64_t addr; uint64_t msg_data; }; struct pptseg { vm_paddr_t gpa; size_t len; int wired; }; struct pptdev { device_t dev; struct vm *vm; /* owner of this device */ TAILQ_ENTRY(pptdev) next; struct pptseg mmio[MAX_MMIOSEGS]; struct { int num_msgs; /* guest state */ int startrid; /* host state */ struct resource *res[MAX_MSIMSGS]; void *cookie[MAX_MSIMSGS]; struct pptintr_arg arg[MAX_MSIMSGS]; } msi; struct { int num_msgs; int startrid; int msix_table_rid; int msix_pba_rid; struct resource *msix_table_res; struct resource *msix_pba_res; struct resource **res; void **cookie; struct pptintr_arg *arg; } msix; }; SYSCTL_DECL(_hw_vmm); -SYSCTL_NODE(_hw_vmm, OID_AUTO, ppt, CTLFLAG_RW, 0, "bhyve passthru devices"); +SYSCTL_NODE(_hw_vmm, OID_AUTO, ppt, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, + "bhyve passthru devices"); static int num_pptdevs; SYSCTL_INT(_hw_vmm_ppt, OID_AUTO, devices, CTLFLAG_RD, &num_pptdevs, 0, "number of pci passthru devices"); static TAILQ_HEAD(, pptdev) pptdev_list = TAILQ_HEAD_INITIALIZER(pptdev_list); static int ppt_probe(device_t dev) { int bus, slot, func; struct pci_devinfo *dinfo; dinfo = (struct pci_devinfo *)device_get_ivars(dev); bus = pci_get_bus(dev); slot = pci_get_slot(dev); func = pci_get_function(dev); /* * To qualify as a pci passthrough device a device must: * - be allowed by administrator to be used in this role * - be an endpoint device */ if ((dinfo->cfg.hdrtype & PCIM_HDRTYPE) != PCIM_HDRTYPE_NORMAL) return (ENXIO); else if (vmm_is_pptdev(bus, slot, func)) return (0); else /* * Returning BUS_PROBE_NOWILDCARD here matches devices that the * SR-IOV infrastructure specified as "ppt" passthrough devices. * All normal devices that did not have "ppt" specified as their * driver will not be matched by this. */ return (BUS_PROBE_NOWILDCARD); } static int ppt_attach(device_t dev) { struct pptdev *ppt; ppt = device_get_softc(dev); iommu_remove_device(iommu_host_domain(), pci_get_rid(dev)); num_pptdevs++; TAILQ_INSERT_TAIL(&pptdev_list, ppt, next); ppt->dev = dev; if (bootverbose) device_printf(dev, "attached\n"); return (0); } static int ppt_detach(device_t dev) { struct pptdev *ppt; ppt = device_get_softc(dev); if (ppt->vm != NULL) return (EBUSY); num_pptdevs--; TAILQ_REMOVE(&pptdev_list, ppt, next); pci_disable_busmaster(dev); iommu_add_device(iommu_host_domain(), pci_get_rid(dev)); return (0); } static device_method_t ppt_methods[] = { /* Device interface */ DEVMETHOD(device_probe, ppt_probe), DEVMETHOD(device_attach, ppt_attach), DEVMETHOD(device_detach, ppt_detach), {0, 0} }; static devclass_t ppt_devclass; DEFINE_CLASS_0(ppt, ppt_driver, ppt_methods, sizeof(struct pptdev)); DRIVER_MODULE(ppt, pci, ppt_driver, ppt_devclass, NULL, NULL); static struct pptdev * ppt_find(int bus, int slot, int func) { device_t dev; struct pptdev *ppt; int b, s, f; TAILQ_FOREACH(ppt, &pptdev_list, next) { dev = ppt->dev; b = pci_get_bus(dev); s = pci_get_slot(dev); f = pci_get_function(dev); if (bus == b && slot == s && func == f) return (ppt); } return (NULL); } static void ppt_unmap_mmio(struct vm *vm, struct pptdev *ppt) { int i; struct pptseg *seg; for (i = 0; i < MAX_MMIOSEGS; i++) { seg = &ppt->mmio[i]; if (seg->len == 0) continue; (void)vm_unmap_mmio(vm, seg->gpa, seg->len); bzero(seg, sizeof(struct pptseg)); } } static void ppt_teardown_msi(struct pptdev *ppt) { int i, rid; void *cookie; struct resource *res; if (ppt->msi.num_msgs == 0) return; for (i = 0; i < ppt->msi.num_msgs; i++) { rid = ppt->msi.startrid + i; res = ppt->msi.res[i]; cookie = ppt->msi.cookie[i]; if (cookie != NULL) bus_teardown_intr(ppt->dev, res, cookie); if (res != NULL) bus_release_resource(ppt->dev, SYS_RES_IRQ, rid, res); ppt->msi.res[i] = NULL; ppt->msi.cookie[i] = NULL; } if (ppt->msi.startrid == 1) pci_release_msi(ppt->dev); ppt->msi.num_msgs = 0; } static void ppt_teardown_msix_intr(struct pptdev *ppt, int idx) { int rid; struct resource *res; void *cookie; rid = ppt->msix.startrid + idx; res = ppt->msix.res[idx]; cookie = ppt->msix.cookie[idx]; if (cookie != NULL) bus_teardown_intr(ppt->dev, res, cookie); if (res != NULL) bus_release_resource(ppt->dev, SYS_RES_IRQ, rid, res); ppt->msix.res[idx] = NULL; ppt->msix.cookie[idx] = NULL; } static void ppt_teardown_msix(struct pptdev *ppt) { int i; if (ppt->msix.num_msgs == 0) return; for (i = 0; i < ppt->msix.num_msgs; i++) ppt_teardown_msix_intr(ppt, i); free(ppt->msix.res, M_PPTMSIX); free(ppt->msix.cookie, M_PPTMSIX); free(ppt->msix.arg, M_PPTMSIX); pci_release_msi(ppt->dev); if (ppt->msix.msix_table_res) { bus_release_resource(ppt->dev, SYS_RES_MEMORY, ppt->msix.msix_table_rid, ppt->msix.msix_table_res); ppt->msix.msix_table_res = NULL; ppt->msix.msix_table_rid = 0; } if (ppt->msix.msix_pba_res) { bus_release_resource(ppt->dev, SYS_RES_MEMORY, ppt->msix.msix_pba_rid, ppt->msix.msix_pba_res); ppt->msix.msix_pba_res = NULL; ppt->msix.msix_pba_rid = 0; } ppt->msix.num_msgs = 0; } int ppt_avail_devices(void) { return (num_pptdevs); } int ppt_assigned_devices(struct vm *vm) { struct pptdev *ppt; int num; num = 0; TAILQ_FOREACH(ppt, &pptdev_list, next) { if (ppt->vm == vm) num++; } return (num); } bool ppt_is_mmio(struct vm *vm, vm_paddr_t gpa) { int i; struct pptdev *ppt; struct pptseg *seg; TAILQ_FOREACH(ppt, &pptdev_list, next) { if (ppt->vm != vm) continue; for (i = 0; i < MAX_MMIOSEGS; i++) { seg = &ppt->mmio[i]; if (seg->len == 0) continue; if (gpa >= seg->gpa && gpa < seg->gpa + seg->len) return (true); } } return (false); } static void ppt_pci_reset(device_t dev) { if (pcie_flr(dev, max(pcie_get_max_completion_timeout(dev) / 1000, 10), true)) return; pci_power_reset(dev); } int ppt_assign_device(struct vm *vm, int bus, int slot, int func) { struct pptdev *ppt; ppt = ppt_find(bus, slot, func); if (ppt != NULL) { /* * If this device is owned by a different VM then we * cannot change its owner. */ if (ppt->vm != NULL && ppt->vm != vm) return (EBUSY); pci_save_state(ppt->dev); ppt_pci_reset(ppt->dev); pci_restore_state(ppt->dev); ppt->vm = vm; iommu_add_device(vm_iommu_domain(vm), pci_get_rid(ppt->dev)); return (0); } return (ENOENT); } int ppt_unassign_device(struct vm *vm, int bus, int slot, int func) { struct pptdev *ppt; ppt = ppt_find(bus, slot, func); if (ppt != NULL) { /* * If this device is not owned by this 'vm' then bail out. */ if (ppt->vm != vm) return (EBUSY); pci_save_state(ppt->dev); ppt_pci_reset(ppt->dev); pci_restore_state(ppt->dev); ppt_unmap_mmio(vm, ppt); ppt_teardown_msi(ppt); ppt_teardown_msix(ppt); iommu_remove_device(vm_iommu_domain(vm), pci_get_rid(ppt->dev)); ppt->vm = NULL; return (0); } return (ENOENT); } int ppt_unassign_all(struct vm *vm) { struct pptdev *ppt; int bus, slot, func; device_t dev; TAILQ_FOREACH(ppt, &pptdev_list, next) { if (ppt->vm == vm) { dev = ppt->dev; bus = pci_get_bus(dev); slot = pci_get_slot(dev); func = pci_get_function(dev); vm_unassign_pptdev(vm, bus, slot, func); } } return (0); } int ppt_map_mmio(struct vm *vm, int bus, int slot, int func, vm_paddr_t gpa, size_t len, vm_paddr_t hpa) { int i, error; struct pptseg *seg; struct pptdev *ppt; ppt = ppt_find(bus, slot, func); if (ppt != NULL) { if (ppt->vm != vm) return (EBUSY); for (i = 0; i < MAX_MMIOSEGS; i++) { seg = &ppt->mmio[i]; if (seg->len == 0) { error = vm_map_mmio(vm, gpa, len, hpa); if (error == 0) { seg->gpa = gpa; seg->len = len; } return (error); } } return (ENOSPC); } return (ENOENT); } static int pptintr(void *arg) { struct pptdev *ppt; struct pptintr_arg *pptarg; pptarg = arg; ppt = pptarg->pptdev; if (ppt->vm != NULL) lapic_intr_msi(ppt->vm, pptarg->addr, pptarg->msg_data); else { /* * XXX * This is not expected to happen - panic? */ } /* * For legacy interrupts give other filters a chance in case * the interrupt was not generated by the passthrough device. */ if (ppt->msi.startrid == 0) return (FILTER_STRAY); else return (FILTER_HANDLED); } int ppt_setup_msi(struct vm *vm, int vcpu, int bus, int slot, int func, uint64_t addr, uint64_t msg, int numvec) { int i, rid, flags; int msi_count, startrid, error, tmp; struct pptdev *ppt; if (numvec < 0 || numvec > MAX_MSIMSGS) return (EINVAL); ppt = ppt_find(bus, slot, func); if (ppt == NULL) return (ENOENT); if (ppt->vm != vm) /* Make sure we own this device */ return (EBUSY); /* Free any allocated resources */ ppt_teardown_msi(ppt); if (numvec == 0) /* nothing more to do */ return (0); flags = RF_ACTIVE; msi_count = pci_msi_count(ppt->dev); if (msi_count == 0) { startrid = 0; /* legacy interrupt */ msi_count = 1; flags |= RF_SHAREABLE; } else startrid = 1; /* MSI */ /* * The device must be capable of supporting the number of vectors * the guest wants to allocate. */ if (numvec > msi_count) return (EINVAL); /* * Make sure that we can allocate all the MSI vectors that are needed * by the guest. */ if (startrid == 1) { tmp = numvec; error = pci_alloc_msi(ppt->dev, &tmp); if (error) return (error); else if (tmp != numvec) { pci_release_msi(ppt->dev); return (ENOSPC); } else { /* success */ } } ppt->msi.startrid = startrid; /* * Allocate the irq resource and attach it to the interrupt handler. */ for (i = 0; i < numvec; i++) { ppt->msi.num_msgs = i + 1; ppt->msi.cookie[i] = NULL; rid = startrid + i; ppt->msi.res[i] = bus_alloc_resource_any(ppt->dev, SYS_RES_IRQ, &rid, flags); if (ppt->msi.res[i] == NULL) break; ppt->msi.arg[i].pptdev = ppt; ppt->msi.arg[i].addr = addr; ppt->msi.arg[i].msg_data = msg + i; error = bus_setup_intr(ppt->dev, ppt->msi.res[i], INTR_TYPE_NET | INTR_MPSAFE, pptintr, NULL, &ppt->msi.arg[i], &ppt->msi.cookie[i]); if (error != 0) break; } if (i < numvec) { ppt_teardown_msi(ppt); return (ENXIO); } return (0); } int ppt_setup_msix(struct vm *vm, int vcpu, int bus, int slot, int func, int idx, uint64_t addr, uint64_t msg, uint32_t vector_control) { struct pptdev *ppt; struct pci_devinfo *dinfo; int numvec, alloced, rid, error; size_t res_size, cookie_size, arg_size; ppt = ppt_find(bus, slot, func); if (ppt == NULL) return (ENOENT); if (ppt->vm != vm) /* Make sure we own this device */ return (EBUSY); dinfo = device_get_ivars(ppt->dev); if (!dinfo) return (ENXIO); /* * First-time configuration: * Allocate the MSI-X table * Allocate the IRQ resources * Set up some variables in ppt->msix */ if (ppt->msix.num_msgs == 0) { numvec = pci_msix_count(ppt->dev); if (numvec <= 0) return (EINVAL); ppt->msix.startrid = 1; ppt->msix.num_msgs = numvec; res_size = numvec * sizeof(ppt->msix.res[0]); cookie_size = numvec * sizeof(ppt->msix.cookie[0]); arg_size = numvec * sizeof(ppt->msix.arg[0]); ppt->msix.res = malloc(res_size, M_PPTMSIX, M_WAITOK | M_ZERO); ppt->msix.cookie = malloc(cookie_size, M_PPTMSIX, M_WAITOK | M_ZERO); ppt->msix.arg = malloc(arg_size, M_PPTMSIX, M_WAITOK | M_ZERO); rid = dinfo->cfg.msix.msix_table_bar; ppt->msix.msix_table_res = bus_alloc_resource_any(ppt->dev, SYS_RES_MEMORY, &rid, RF_ACTIVE); if (ppt->msix.msix_table_res == NULL) { ppt_teardown_msix(ppt); return (ENOSPC); } ppt->msix.msix_table_rid = rid; if (dinfo->cfg.msix.msix_table_bar != dinfo->cfg.msix.msix_pba_bar) { rid = dinfo->cfg.msix.msix_pba_bar; ppt->msix.msix_pba_res = bus_alloc_resource_any( ppt->dev, SYS_RES_MEMORY, &rid, RF_ACTIVE); if (ppt->msix.msix_pba_res == NULL) { ppt_teardown_msix(ppt); return (ENOSPC); } ppt->msix.msix_pba_rid = rid; } alloced = numvec; error = pci_alloc_msix(ppt->dev, &alloced); if (error || alloced != numvec) { ppt_teardown_msix(ppt); return (error == 0 ? ENOSPC: error); } } if ((vector_control & PCIM_MSIX_VCTRL_MASK) == 0) { /* Tear down the IRQ if it's already set up */ ppt_teardown_msix_intr(ppt, idx); /* Allocate the IRQ resource */ ppt->msix.cookie[idx] = NULL; rid = ppt->msix.startrid + idx; ppt->msix.res[idx] = bus_alloc_resource_any(ppt->dev, SYS_RES_IRQ, &rid, RF_ACTIVE); if (ppt->msix.res[idx] == NULL) return (ENXIO); ppt->msix.arg[idx].pptdev = ppt; ppt->msix.arg[idx].addr = addr; ppt->msix.arg[idx].msg_data = msg; /* Setup the MSI-X interrupt */ error = bus_setup_intr(ppt->dev, ppt->msix.res[idx], INTR_TYPE_NET | INTR_MPSAFE, pptintr, NULL, &ppt->msix.arg[idx], &ppt->msix.cookie[idx]); if (error != 0) { bus_release_resource(ppt->dev, SYS_RES_IRQ, rid, ppt->msix.res[idx]); ppt->msix.cookie[idx] = NULL; ppt->msix.res[idx] = NULL; return (ENXIO); } } else { /* Masked, tear it down if it's already been set up */ ppt_teardown_msix_intr(ppt, idx); } return (0); } Index: head/sys/amd64/vmm/io/vrtc.c =================================================================== --- head/sys/amd64/vmm/io/vrtc.c (revision 357973) +++ head/sys/amd64/vmm/io/vrtc.c (revision 357974) @@ -1,1021 +1,1022 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2014, Neel Natu (neel@freebsd.org) * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include "vmm_ktr.h" #include "vatpic.h" #include "vioapic.h" #include "vrtc.h" /* Register layout of the RTC */ struct rtcdev { uint8_t sec; uint8_t alarm_sec; uint8_t min; uint8_t alarm_min; uint8_t hour; uint8_t alarm_hour; uint8_t day_of_week; uint8_t day_of_month; uint8_t month; uint8_t year; uint8_t reg_a; uint8_t reg_b; uint8_t reg_c; uint8_t reg_d; uint8_t nvram[36]; uint8_t century; uint8_t nvram2[128 - 51]; } __packed; CTASSERT(sizeof(struct rtcdev) == 128); CTASSERT(offsetof(struct rtcdev, century) == RTC_CENTURY); struct vrtc { struct vm *vm; struct mtx mtx; struct callout callout; u_int addr; /* RTC register to read or write */ sbintime_t base_uptime; time_t base_rtctime; struct rtcdev rtcdev; }; #define VRTC_LOCK(vrtc) mtx_lock(&((vrtc)->mtx)) #define VRTC_UNLOCK(vrtc) mtx_unlock(&((vrtc)->mtx)) #define VRTC_LOCKED(vrtc) mtx_owned(&((vrtc)->mtx)) /* * RTC time is considered "broken" if: * - RTC updates are halted by the guest * - RTC date/time fields have invalid values */ #define VRTC_BROKEN_TIME ((time_t)-1) #define RTC_IRQ 8 #define RTCSB_BIN 0x04 #define RTCSB_ALL_INTRS (RTCSB_UINTR | RTCSB_AINTR | RTCSB_PINTR) #define rtc_halted(vrtc) ((vrtc->rtcdev.reg_b & RTCSB_HALT) != 0) #define aintr_enabled(vrtc) (((vrtc)->rtcdev.reg_b & RTCSB_AINTR) != 0) #define pintr_enabled(vrtc) (((vrtc)->rtcdev.reg_b & RTCSB_PINTR) != 0) #define uintr_enabled(vrtc) (((vrtc)->rtcdev.reg_b & RTCSB_UINTR) != 0) static void vrtc_callout_handler(void *arg); static void vrtc_set_reg_c(struct vrtc *vrtc, uint8_t newval); static MALLOC_DEFINE(M_VRTC, "vrtc", "bhyve virtual rtc"); SYSCTL_DECL(_hw_vmm); -SYSCTL_NODE(_hw_vmm, OID_AUTO, vrtc, CTLFLAG_RW, NULL, NULL); +SYSCTL_NODE(_hw_vmm, OID_AUTO, vrtc, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, + NULL); static int rtc_flag_broken_time = 1; SYSCTL_INT(_hw_vmm_vrtc, OID_AUTO, flag_broken_time, CTLFLAG_RDTUN, &rtc_flag_broken_time, 0, "Stop guest when invalid RTC time is detected"); static __inline bool divider_enabled(int reg_a) { /* * The RTC is counting only when dividers are not held in reset. */ return ((reg_a & 0x70) == 0x20); } static __inline bool update_enabled(struct vrtc *vrtc) { /* * RTC date/time can be updated only if: * - divider is not held in reset * - guest has not disabled updates * - the date/time fields have valid contents */ if (!divider_enabled(vrtc->rtcdev.reg_a)) return (false); if (rtc_halted(vrtc)) return (false); if (vrtc->base_rtctime == VRTC_BROKEN_TIME) return (false); return (true); } static time_t vrtc_curtime(struct vrtc *vrtc, sbintime_t *basetime) { sbintime_t now, delta; time_t t, secs; KASSERT(VRTC_LOCKED(vrtc), ("%s: vrtc not locked", __func__)); t = vrtc->base_rtctime; *basetime = vrtc->base_uptime; if (update_enabled(vrtc)) { now = sbinuptime(); delta = now - vrtc->base_uptime; KASSERT(delta >= 0, ("vrtc_curtime: uptime went backwards: " "%#lx to %#lx", vrtc->base_uptime, now)); secs = delta / SBT_1S; t += secs; *basetime += secs * SBT_1S; } return (t); } static __inline uint8_t rtcset(struct rtcdev *rtc, int val) { KASSERT(val >= 0 && val < 100, ("%s: invalid bin2bcd index %d", __func__, val)); return ((rtc->reg_b & RTCSB_BIN) ? val : bin2bcd_data[val]); } static void secs_to_rtc(time_t rtctime, struct vrtc *vrtc, int force_update) { struct clocktime ct; struct timespec ts; struct rtcdev *rtc; int hour; KASSERT(VRTC_LOCKED(vrtc), ("%s: vrtc not locked", __func__)); if (rtctime < 0) { KASSERT(rtctime == VRTC_BROKEN_TIME, ("%s: invalid vrtc time %#lx", __func__, rtctime)); return; } /* * If the RTC is halted then the guest has "ownership" of the * date/time fields. Don't update the RTC date/time fields in * this case (unless forced). */ if (rtc_halted(vrtc) && !force_update) return; ts.tv_sec = rtctime; ts.tv_nsec = 0; clock_ts_to_ct(&ts, &ct); KASSERT(ct.sec >= 0 && ct.sec <= 59, ("invalid clocktime sec %d", ct.sec)); KASSERT(ct.min >= 0 && ct.min <= 59, ("invalid clocktime min %d", ct.min)); KASSERT(ct.hour >= 0 && ct.hour <= 23, ("invalid clocktime hour %d", ct.hour)); KASSERT(ct.dow >= 0 && ct.dow <= 6, ("invalid clocktime wday %d", ct.dow)); KASSERT(ct.day >= 1 && ct.day <= 31, ("invalid clocktime mday %d", ct.day)); KASSERT(ct.mon >= 1 && ct.mon <= 12, ("invalid clocktime month %d", ct.mon)); KASSERT(ct.year >= POSIX_BASE_YEAR, ("invalid clocktime year %d", ct.year)); rtc = &vrtc->rtcdev; rtc->sec = rtcset(rtc, ct.sec); rtc->min = rtcset(rtc, ct.min); if (rtc->reg_b & RTCSB_24HR) { hour = ct.hour; } else { /* * Convert to the 12-hour format. */ switch (ct.hour) { case 0: /* 12 AM */ case 12: /* 12 PM */ hour = 12; break; default: /* * The remaining 'ct.hour' values are interpreted as: * [1 - 11] -> 1 - 11 AM * [13 - 23] -> 1 - 11 PM */ hour = ct.hour % 12; break; } } rtc->hour = rtcset(rtc, hour); if ((rtc->reg_b & RTCSB_24HR) == 0 && ct.hour >= 12) rtc->hour |= 0x80; /* set MSB to indicate PM */ rtc->day_of_week = rtcset(rtc, ct.dow + 1); rtc->day_of_month = rtcset(rtc, ct.day); rtc->month = rtcset(rtc, ct.mon); rtc->year = rtcset(rtc, ct.year % 100); rtc->century = rtcset(rtc, ct.year / 100); } static int rtcget(struct rtcdev *rtc, int val, int *retval) { uint8_t upper, lower; if (rtc->reg_b & RTCSB_BIN) { *retval = val; return (0); } lower = val & 0xf; upper = (val >> 4) & 0xf; if (lower > 9 || upper > 9) return (-1); *retval = upper * 10 + lower; return (0); } static time_t rtc_to_secs(struct vrtc *vrtc) { struct clocktime ct; struct timespec ts; struct rtcdev *rtc; struct vm *vm; int century, error, hour, pm, year; KASSERT(VRTC_LOCKED(vrtc), ("%s: vrtc not locked", __func__)); vm = vrtc->vm; rtc = &vrtc->rtcdev; bzero(&ct, sizeof(struct clocktime)); error = rtcget(rtc, rtc->sec, &ct.sec); if (error || ct.sec < 0 || ct.sec > 59) { VM_CTR2(vm, "Invalid RTC sec %#x/%d", rtc->sec, ct.sec); goto fail; } error = rtcget(rtc, rtc->min, &ct.min); if (error || ct.min < 0 || ct.min > 59) { VM_CTR2(vm, "Invalid RTC min %#x/%d", rtc->min, ct.min); goto fail; } pm = 0; hour = rtc->hour; if ((rtc->reg_b & RTCSB_24HR) == 0) { if (hour & 0x80) { hour &= ~0x80; pm = 1; } } error = rtcget(rtc, hour, &ct.hour); if ((rtc->reg_b & RTCSB_24HR) == 0) { if (ct.hour >= 1 && ct.hour <= 12) { /* * Convert from 12-hour format to internal 24-hour * representation as follows: * * 12-hour format ct.hour * 12 AM 0 * 1 - 11 AM 1 - 11 * 12 PM 12 * 1 - 11 PM 13 - 23 */ if (ct.hour == 12) ct.hour = 0; if (pm) ct.hour += 12; } else { VM_CTR2(vm, "Invalid RTC 12-hour format %#x/%d", rtc->hour, ct.hour); goto fail; } } if (error || ct.hour < 0 || ct.hour > 23) { VM_CTR2(vm, "Invalid RTC hour %#x/%d", rtc->hour, ct.hour); goto fail; } /* * Ignore 'rtc->dow' because some guests like Linux don't bother * setting it at all while others like OpenBSD/i386 set it incorrectly. * * clock_ct_to_ts() does not depend on 'ct.dow' anyways so ignore it. */ ct.dow = -1; error = rtcget(rtc, rtc->day_of_month, &ct.day); if (error || ct.day < 1 || ct.day > 31) { VM_CTR2(vm, "Invalid RTC mday %#x/%d", rtc->day_of_month, ct.day); goto fail; } error = rtcget(rtc, rtc->month, &ct.mon); if (error || ct.mon < 1 || ct.mon > 12) { VM_CTR2(vm, "Invalid RTC month %#x/%d", rtc->month, ct.mon); goto fail; } error = rtcget(rtc, rtc->year, &year); if (error || year < 0 || year > 99) { VM_CTR2(vm, "Invalid RTC year %#x/%d", rtc->year, year); goto fail; } error = rtcget(rtc, rtc->century, ¢ury); ct.year = century * 100 + year; if (error || ct.year < POSIX_BASE_YEAR) { VM_CTR2(vm, "Invalid RTC century %#x/%d", rtc->century, ct.year); goto fail; } error = clock_ct_to_ts(&ct, &ts); if (error || ts.tv_sec < 0) { VM_CTR3(vm, "Invalid RTC clocktime.date %04d-%02d-%02d", ct.year, ct.mon, ct.day); VM_CTR3(vm, "Invalid RTC clocktime.time %02d:%02d:%02d", ct.hour, ct.min, ct.sec); goto fail; } return (ts.tv_sec); /* success */ fail: /* * Stop updating the RTC if the date/time fields programmed by * the guest are invalid. */ VM_CTR0(vrtc->vm, "Invalid RTC date/time programming detected"); return (VRTC_BROKEN_TIME); } static int vrtc_time_update(struct vrtc *vrtc, time_t newtime, sbintime_t newbase) { struct rtcdev *rtc; sbintime_t oldbase; time_t oldtime; uint8_t alarm_sec, alarm_min, alarm_hour; KASSERT(VRTC_LOCKED(vrtc), ("%s: vrtc not locked", __func__)); rtc = &vrtc->rtcdev; alarm_sec = rtc->alarm_sec; alarm_min = rtc->alarm_min; alarm_hour = rtc->alarm_hour; oldtime = vrtc->base_rtctime; VM_CTR2(vrtc->vm, "Updating RTC secs from %#lx to %#lx", oldtime, newtime); oldbase = vrtc->base_uptime; VM_CTR2(vrtc->vm, "Updating RTC base uptime from %#lx to %#lx", oldbase, newbase); vrtc->base_uptime = newbase; if (newtime == oldtime) return (0); /* * If 'newtime' indicates that RTC updates are disabled then just * record that and return. There is no need to do alarm interrupt * processing in this case. */ if (newtime == VRTC_BROKEN_TIME) { vrtc->base_rtctime = VRTC_BROKEN_TIME; return (0); } /* * Return an error if RTC updates are halted by the guest. */ if (rtc_halted(vrtc)) { VM_CTR0(vrtc->vm, "RTC update halted by guest"); return (EBUSY); } do { /* * If the alarm interrupt is enabled and 'oldtime' is valid * then visit all the seconds between 'oldtime' and 'newtime' * to check for the alarm condition. * * Otherwise move the RTC time forward directly to 'newtime'. */ if (aintr_enabled(vrtc) && oldtime != VRTC_BROKEN_TIME) vrtc->base_rtctime++; else vrtc->base_rtctime = newtime; if (aintr_enabled(vrtc)) { /* * Update the RTC date/time fields before checking * if the alarm conditions are satisfied. */ secs_to_rtc(vrtc->base_rtctime, vrtc, 0); if ((alarm_sec >= 0xC0 || alarm_sec == rtc->sec) && (alarm_min >= 0xC0 || alarm_min == rtc->min) && (alarm_hour >= 0xC0 || alarm_hour == rtc->hour)) { vrtc_set_reg_c(vrtc, rtc->reg_c | RTCIR_ALARM); } } } while (vrtc->base_rtctime != newtime); if (uintr_enabled(vrtc)) vrtc_set_reg_c(vrtc, rtc->reg_c | RTCIR_UPDATE); return (0); } static sbintime_t vrtc_freq(struct vrtc *vrtc) { int ratesel; static sbintime_t pf[16] = { 0, SBT_1S / 256, SBT_1S / 128, SBT_1S / 8192, SBT_1S / 4096, SBT_1S / 2048, SBT_1S / 1024, SBT_1S / 512, SBT_1S / 256, SBT_1S / 128, SBT_1S / 64, SBT_1S / 32, SBT_1S / 16, SBT_1S / 8, SBT_1S / 4, SBT_1S / 2, }; KASSERT(VRTC_LOCKED(vrtc), ("%s: vrtc not locked", __func__)); /* * If both periodic and alarm interrupts are enabled then use the * periodic frequency to drive the callout. The minimum periodic * frequency (2 Hz) is higher than the alarm frequency (1 Hz) so * piggyback the alarm on top of it. The same argument applies to * the update interrupt. */ if (pintr_enabled(vrtc) && divider_enabled(vrtc->rtcdev.reg_a)) { ratesel = vrtc->rtcdev.reg_a & 0xf; return (pf[ratesel]); } else if (aintr_enabled(vrtc) && update_enabled(vrtc)) { return (SBT_1S); } else if (uintr_enabled(vrtc) && update_enabled(vrtc)) { return (SBT_1S); } else { return (0); } } static void vrtc_callout_reset(struct vrtc *vrtc, sbintime_t freqsbt) { KASSERT(VRTC_LOCKED(vrtc), ("%s: vrtc not locked", __func__)); if (freqsbt == 0) { if (callout_active(&vrtc->callout)) { VM_CTR0(vrtc->vm, "RTC callout stopped"); callout_stop(&vrtc->callout); } return; } VM_CTR1(vrtc->vm, "RTC callout frequency %d hz", SBT_1S / freqsbt); callout_reset_sbt(&vrtc->callout, freqsbt, 0, vrtc_callout_handler, vrtc, 0); } static void vrtc_callout_handler(void *arg) { struct vrtc *vrtc = arg; sbintime_t freqsbt, basetime; time_t rtctime; int error; VM_CTR0(vrtc->vm, "vrtc callout fired"); VRTC_LOCK(vrtc); if (callout_pending(&vrtc->callout)) /* callout was reset */ goto done; if (!callout_active(&vrtc->callout)) /* callout was stopped */ goto done; callout_deactivate(&vrtc->callout); KASSERT((vrtc->rtcdev.reg_b & RTCSB_ALL_INTRS) != 0, ("gratuitous vrtc callout")); if (pintr_enabled(vrtc)) vrtc_set_reg_c(vrtc, vrtc->rtcdev.reg_c | RTCIR_PERIOD); if (aintr_enabled(vrtc) || uintr_enabled(vrtc)) { rtctime = vrtc_curtime(vrtc, &basetime); error = vrtc_time_update(vrtc, rtctime, basetime); KASSERT(error == 0, ("%s: vrtc_time_update error %d", __func__, error)); } freqsbt = vrtc_freq(vrtc); KASSERT(freqsbt != 0, ("%s: vrtc frequency cannot be zero", __func__)); vrtc_callout_reset(vrtc, freqsbt); done: VRTC_UNLOCK(vrtc); } static __inline void vrtc_callout_check(struct vrtc *vrtc, sbintime_t freq) { int active; active = callout_active(&vrtc->callout) ? 1 : 0; KASSERT((freq == 0 && !active) || (freq != 0 && active), ("vrtc callout %s with frequency %#lx", active ? "active" : "inactive", freq)); } static void vrtc_set_reg_c(struct vrtc *vrtc, uint8_t newval) { struct rtcdev *rtc; int oldirqf, newirqf; uint8_t oldval, changed; KASSERT(VRTC_LOCKED(vrtc), ("%s: vrtc not locked", __func__)); rtc = &vrtc->rtcdev; newval &= RTCIR_ALARM | RTCIR_PERIOD | RTCIR_UPDATE; oldirqf = rtc->reg_c & RTCIR_INT; if ((aintr_enabled(vrtc) && (newval & RTCIR_ALARM) != 0) || (pintr_enabled(vrtc) && (newval & RTCIR_PERIOD) != 0) || (uintr_enabled(vrtc) && (newval & RTCIR_UPDATE) != 0)) { newirqf = RTCIR_INT; } else { newirqf = 0; } oldval = rtc->reg_c; rtc->reg_c = newirqf | newval; changed = oldval ^ rtc->reg_c; if (changed) { VM_CTR2(vrtc->vm, "RTC reg_c changed from %#x to %#x", oldval, rtc->reg_c); } if (!oldirqf && newirqf) { VM_CTR1(vrtc->vm, "RTC irq %d asserted", RTC_IRQ); vatpic_pulse_irq(vrtc->vm, RTC_IRQ); vioapic_pulse_irq(vrtc->vm, RTC_IRQ); } else if (oldirqf && !newirqf) { VM_CTR1(vrtc->vm, "RTC irq %d deasserted", RTC_IRQ); } } static int vrtc_set_reg_b(struct vrtc *vrtc, uint8_t newval) { struct rtcdev *rtc; sbintime_t oldfreq, newfreq, basetime; time_t curtime, rtctime; int error; uint8_t oldval, changed; KASSERT(VRTC_LOCKED(vrtc), ("%s: vrtc not locked", __func__)); rtc = &vrtc->rtcdev; oldval = rtc->reg_b; oldfreq = vrtc_freq(vrtc); rtc->reg_b = newval; changed = oldval ^ newval; if (changed) { VM_CTR2(vrtc->vm, "RTC reg_b changed from %#x to %#x", oldval, newval); } if (changed & RTCSB_HALT) { if ((newval & RTCSB_HALT) == 0) { rtctime = rtc_to_secs(vrtc); basetime = sbinuptime(); if (rtctime == VRTC_BROKEN_TIME) { if (rtc_flag_broken_time) return (-1); } } else { curtime = vrtc_curtime(vrtc, &basetime); KASSERT(curtime == vrtc->base_rtctime, ("%s: mismatch " "between vrtc basetime (%#lx) and curtime (%#lx)", __func__, vrtc->base_rtctime, curtime)); /* * Force a refresh of the RTC date/time fields so * they reflect the time right before the guest set * the HALT bit. */ secs_to_rtc(curtime, vrtc, 1); /* * Updates are halted so mark 'base_rtctime' to denote * that the RTC date/time is in flux. */ rtctime = VRTC_BROKEN_TIME; rtc->reg_b &= ~RTCSB_UINTR; } error = vrtc_time_update(vrtc, rtctime, basetime); KASSERT(error == 0, ("vrtc_time_update error %d", error)); } /* * Side effect of changes to the interrupt enable bits. */ if (changed & RTCSB_ALL_INTRS) vrtc_set_reg_c(vrtc, vrtc->rtcdev.reg_c); /* * Change the callout frequency if it has changed. */ newfreq = vrtc_freq(vrtc); if (newfreq != oldfreq) vrtc_callout_reset(vrtc, newfreq); else vrtc_callout_check(vrtc, newfreq); /* * The side effect of bits that control the RTC date/time format * is handled lazily when those fields are actually read. */ return (0); } static void vrtc_set_reg_a(struct vrtc *vrtc, uint8_t newval) { sbintime_t oldfreq, newfreq; uint8_t oldval, changed; KASSERT(VRTC_LOCKED(vrtc), ("%s: vrtc not locked", __func__)); newval &= ~RTCSA_TUP; oldval = vrtc->rtcdev.reg_a; oldfreq = vrtc_freq(vrtc); if (divider_enabled(oldval) && !divider_enabled(newval)) { VM_CTR2(vrtc->vm, "RTC divider held in reset at %#lx/%#lx", vrtc->base_rtctime, vrtc->base_uptime); } else if (!divider_enabled(oldval) && divider_enabled(newval)) { /* * If the dividers are coming out of reset then update * 'base_uptime' before this happens. This is done to * maintain the illusion that the RTC date/time was frozen * while the dividers were disabled. */ vrtc->base_uptime = sbinuptime(); VM_CTR2(vrtc->vm, "RTC divider out of reset at %#lx/%#lx", vrtc->base_rtctime, vrtc->base_uptime); } else { /* NOTHING */ } vrtc->rtcdev.reg_a = newval; changed = oldval ^ newval; if (changed) { VM_CTR2(vrtc->vm, "RTC reg_a changed from %#x to %#x", oldval, newval); } /* * Side effect of changes to rate select and divider enable bits. */ newfreq = vrtc_freq(vrtc); if (newfreq != oldfreq) vrtc_callout_reset(vrtc, newfreq); else vrtc_callout_check(vrtc, newfreq); } int vrtc_set_time(struct vm *vm, time_t secs) { struct vrtc *vrtc; int error; vrtc = vm_rtc(vm); VRTC_LOCK(vrtc); error = vrtc_time_update(vrtc, secs, sbinuptime()); VRTC_UNLOCK(vrtc); if (error) { VM_CTR2(vrtc->vm, "Error %d setting RTC time to %#lx", error, secs); } else { VM_CTR1(vrtc->vm, "RTC time set to %#lx", secs); } return (error); } time_t vrtc_get_time(struct vm *vm) { struct vrtc *vrtc; sbintime_t basetime; time_t t; vrtc = vm_rtc(vm); VRTC_LOCK(vrtc); t = vrtc_curtime(vrtc, &basetime); VRTC_UNLOCK(vrtc); return (t); } int vrtc_nvram_write(struct vm *vm, int offset, uint8_t value) { struct vrtc *vrtc; uint8_t *ptr; vrtc = vm_rtc(vm); /* * Don't allow writes to RTC control registers or the date/time fields. */ if (offset < offsetof(struct rtcdev, nvram[0]) || offset == RTC_CENTURY || offset >= sizeof(struct rtcdev)) { VM_CTR1(vrtc->vm, "RTC nvram write to invalid offset %d", offset); return (EINVAL); } VRTC_LOCK(vrtc); ptr = (uint8_t *)(&vrtc->rtcdev); ptr[offset] = value; VM_CTR2(vrtc->vm, "RTC nvram write %#x to offset %#x", value, offset); VRTC_UNLOCK(vrtc); return (0); } int vrtc_nvram_read(struct vm *vm, int offset, uint8_t *retval) { struct vrtc *vrtc; sbintime_t basetime; time_t curtime; uint8_t *ptr; /* * Allow all offsets in the RTC to be read. */ if (offset < 0 || offset >= sizeof(struct rtcdev)) return (EINVAL); vrtc = vm_rtc(vm); VRTC_LOCK(vrtc); /* * Update RTC date/time fields if necessary. */ if (offset < 10 || offset == RTC_CENTURY) { curtime = vrtc_curtime(vrtc, &basetime); secs_to_rtc(curtime, vrtc, 0); } ptr = (uint8_t *)(&vrtc->rtcdev); *retval = ptr[offset]; VRTC_UNLOCK(vrtc); return (0); } int vrtc_addr_handler(struct vm *vm, int vcpuid, bool in, int port, int bytes, uint32_t *val) { struct vrtc *vrtc; vrtc = vm_rtc(vm); if (bytes != 1) return (-1); if (in) { *val = 0xff; return (0); } VRTC_LOCK(vrtc); vrtc->addr = *val & 0x7f; VRTC_UNLOCK(vrtc); return (0); } int vrtc_data_handler(struct vm *vm, int vcpuid, bool in, int port, int bytes, uint32_t *val) { struct vrtc *vrtc; struct rtcdev *rtc; sbintime_t basetime; time_t curtime; int error, offset; vrtc = vm_rtc(vm); rtc = &vrtc->rtcdev; if (bytes != 1) return (-1); VRTC_LOCK(vrtc); offset = vrtc->addr; if (offset >= sizeof(struct rtcdev)) { VRTC_UNLOCK(vrtc); return (-1); } error = 0; curtime = vrtc_curtime(vrtc, &basetime); vrtc_time_update(vrtc, curtime, basetime); /* * Update RTC date/time fields if necessary. * * This is not just for reads of the RTC. The side-effect of writing * the century byte requires other RTC date/time fields (e.g. sec) * to be updated here. */ if (offset < 10 || offset == RTC_CENTURY) secs_to_rtc(curtime, vrtc, 0); if (in) { if (offset == 12) { /* * XXX * reg_c interrupt flags are updated only if the * corresponding interrupt enable bit in reg_b is set. */ *val = vrtc->rtcdev.reg_c; vrtc_set_reg_c(vrtc, 0); } else { *val = *((uint8_t *)rtc + offset); } VCPU_CTR2(vm, vcpuid, "Read value %#x from RTC offset %#x", *val, offset); } else { switch (offset) { case 10: VCPU_CTR1(vm, vcpuid, "RTC reg_a set to %#x", *val); vrtc_set_reg_a(vrtc, *val); break; case 11: VCPU_CTR1(vm, vcpuid, "RTC reg_b set to %#x", *val); error = vrtc_set_reg_b(vrtc, *val); break; case 12: VCPU_CTR1(vm, vcpuid, "RTC reg_c set to %#x (ignored)", *val); break; case 13: VCPU_CTR1(vm, vcpuid, "RTC reg_d set to %#x (ignored)", *val); break; case 0: /* * High order bit of 'seconds' is readonly. */ *val &= 0x7f; /* FALLTHRU */ default: VCPU_CTR2(vm, vcpuid, "RTC offset %#x set to %#x", offset, *val); *((uint8_t *)rtc + offset) = *val; break; } /* * XXX some guests (e.g. OpenBSD) write the century byte * outside of RTCSB_HALT so re-calculate the RTC date/time. */ if (offset == RTC_CENTURY && !rtc_halted(vrtc)) { curtime = rtc_to_secs(vrtc); error = vrtc_time_update(vrtc, curtime, sbinuptime()); KASSERT(!error, ("vrtc_time_update error %d", error)); if (curtime == VRTC_BROKEN_TIME && rtc_flag_broken_time) error = -1; } } VRTC_UNLOCK(vrtc); return (error); } void vrtc_reset(struct vrtc *vrtc) { struct rtcdev *rtc; VRTC_LOCK(vrtc); rtc = &vrtc->rtcdev; vrtc_set_reg_b(vrtc, rtc->reg_b & ~(RTCSB_ALL_INTRS | RTCSB_SQWE)); vrtc_set_reg_c(vrtc, 0); KASSERT(!callout_active(&vrtc->callout), ("rtc callout still active")); VRTC_UNLOCK(vrtc); } struct vrtc * vrtc_init(struct vm *vm) { struct vrtc *vrtc; struct rtcdev *rtc; time_t curtime; vrtc = malloc(sizeof(struct vrtc), M_VRTC, M_WAITOK | M_ZERO); vrtc->vm = vm; mtx_init(&vrtc->mtx, "vrtc lock", NULL, MTX_DEF); callout_init(&vrtc->callout, 1); /* Allow dividers to keep time but disable everything else */ rtc = &vrtc->rtcdev; rtc->reg_a = 0x20; rtc->reg_b = RTCSB_24HR; rtc->reg_c = 0; rtc->reg_d = RTCSD_PWR; /* Reset the index register to a safe value. */ vrtc->addr = RTC_STATUSD; /* * Initialize RTC time to 00:00:00 Jan 1, 1970. */ curtime = 0; VRTC_LOCK(vrtc); vrtc->base_rtctime = VRTC_BROKEN_TIME; vrtc_time_update(vrtc, curtime, sbinuptime()); secs_to_rtc(curtime, vrtc, 0); VRTC_UNLOCK(vrtc); return (vrtc); } void vrtc_cleanup(struct vrtc *vrtc) { callout_drain(&vrtc->callout); free(vrtc, M_VRTC); } Index: head/sys/amd64/vmm/vmm.c =================================================================== --- head/sys/amd64/vmm/vmm.c (revision 357973) +++ head/sys/amd64/vmm/vmm.c (revision 357974) @@ -1,2731 +1,2732 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "vmm_ioport.h" #include "vmm_ktr.h" #include "vmm_host.h" #include "vmm_mem.h" #include "vmm_util.h" #include "vatpic.h" #include "vatpit.h" #include "vhpet.h" #include "vioapic.h" #include "vlapic.h" #include "vpmtmr.h" #include "vrtc.h" #include "vmm_stat.h" #include "vmm_lapic.h" #include "io/ppt.h" #include "io/iommu.h" struct vlapic; /* * Initialization: * (a) allocated when vcpu is created * (i) initialized when vcpu is created and when it is reinitialized * (o) initialized the first time the vcpu is created * (x) initialized before use */ struct vcpu { struct mtx mtx; /* (o) protects 'state' and 'hostcpu' */ enum vcpu_state state; /* (o) vcpu state */ int hostcpu; /* (o) vcpu's host cpu */ int reqidle; /* (i) request vcpu to idle */ struct vlapic *vlapic; /* (i) APIC device model */ enum x2apic_state x2apic_state; /* (i) APIC mode */ uint64_t exitintinfo; /* (i) events pending at VM exit */ int nmi_pending; /* (i) NMI pending */ int extint_pending; /* (i) INTR pending */ int exception_pending; /* (i) exception pending */ int exc_vector; /* (x) exception collateral */ int exc_errcode_valid; uint32_t exc_errcode; struct savefpu *guestfpu; /* (a,i) guest fpu state */ uint64_t guest_xcr0; /* (i) guest %xcr0 register */ void *stats; /* (a,i) statistics */ struct vm_exit exitinfo; /* (x) exit reason and collateral */ uint64_t nextrip; /* (x) next instruction to execute */ }; #define vcpu_lock_initialized(v) mtx_initialized(&((v)->mtx)) #define vcpu_lock_init(v) mtx_init(&((v)->mtx), "vcpu lock", 0, MTX_SPIN) #define vcpu_lock(v) mtx_lock_spin(&((v)->mtx)) #define vcpu_unlock(v) mtx_unlock_spin(&((v)->mtx)) #define vcpu_assert_locked(v) mtx_assert(&((v)->mtx), MA_OWNED) struct mem_seg { size_t len; bool sysmem; struct vm_object *object; }; #define VM_MAX_MEMSEGS 3 struct mem_map { vm_paddr_t gpa; size_t len; vm_ooffset_t segoff; int segid; int prot; int flags; }; #define VM_MAX_MEMMAPS 4 /* * Initialization: * (o) initialized the first time the VM is created * (i) initialized when VM is created and when it is reinitialized * (x) initialized before use */ struct vm { void *cookie; /* (i) cpu-specific data */ void *iommu; /* (x) iommu-specific data */ struct vhpet *vhpet; /* (i) virtual HPET */ struct vioapic *vioapic; /* (i) virtual ioapic */ struct vatpic *vatpic; /* (i) virtual atpic */ struct vatpit *vatpit; /* (i) virtual atpit */ struct vpmtmr *vpmtmr; /* (i) virtual ACPI PM timer */ struct vrtc *vrtc; /* (o) virtual RTC */ volatile cpuset_t active_cpus; /* (i) active vcpus */ volatile cpuset_t debug_cpus; /* (i) vcpus stopped for debug */ int suspend; /* (i) stop VM execution */ volatile cpuset_t suspended_cpus; /* (i) suspended vcpus */ volatile cpuset_t halted_cpus; /* (x) cpus in a hard halt */ cpuset_t rendezvous_req_cpus; /* (x) rendezvous requested */ cpuset_t rendezvous_done_cpus; /* (x) rendezvous finished */ void *rendezvous_arg; /* (x) rendezvous func/arg */ vm_rendezvous_func_t rendezvous_func; struct mtx rendezvous_mtx; /* (o) rendezvous lock */ struct mem_map mem_maps[VM_MAX_MEMMAPS]; /* (i) guest address space */ struct mem_seg mem_segs[VM_MAX_MEMSEGS]; /* (o) guest memory regions */ struct vmspace *vmspace; /* (o) guest's address space */ char name[VM_MAX_NAMELEN]; /* (o) virtual machine name */ struct vcpu vcpu[VM_MAXCPU]; /* (i) guest vcpus */ /* The following describe the vm cpu topology */ uint16_t sockets; /* (o) num of sockets */ uint16_t cores; /* (o) num of cores/socket */ uint16_t threads; /* (o) num of threads/core */ uint16_t maxcpus; /* (o) max pluggable cpus */ }; static int vmm_initialized; static struct vmm_ops *ops; #define VMM_INIT(num) (ops != NULL ? (*ops->init)(num) : 0) #define VMM_CLEANUP() (ops != NULL ? (*ops->cleanup)() : 0) #define VMM_RESUME() (ops != NULL ? (*ops->resume)() : 0) #define VMINIT(vm, pmap) (ops != NULL ? (*ops->vminit)(vm, pmap): NULL) #define VMRUN(vmi, vcpu, rip, pmap, evinfo) \ (ops != NULL ? (*ops->vmrun)(vmi, vcpu, rip, pmap, evinfo) : ENXIO) #define VMCLEANUP(vmi) (ops != NULL ? (*ops->vmcleanup)(vmi) : NULL) #define VMSPACE_ALLOC(min, max) \ (ops != NULL ? (*ops->vmspace_alloc)(min, max) : NULL) #define VMSPACE_FREE(vmspace) \ (ops != NULL ? (*ops->vmspace_free)(vmspace) : ENXIO) #define VMGETREG(vmi, vcpu, num, retval) \ (ops != NULL ? (*ops->vmgetreg)(vmi, vcpu, num, retval) : ENXIO) #define VMSETREG(vmi, vcpu, num, val) \ (ops != NULL ? (*ops->vmsetreg)(vmi, vcpu, num, val) : ENXIO) #define VMGETDESC(vmi, vcpu, num, desc) \ (ops != NULL ? (*ops->vmgetdesc)(vmi, vcpu, num, desc) : ENXIO) #define VMSETDESC(vmi, vcpu, num, desc) \ (ops != NULL ? (*ops->vmsetdesc)(vmi, vcpu, num, desc) : ENXIO) #define VMGETCAP(vmi, vcpu, num, retval) \ (ops != NULL ? (*ops->vmgetcap)(vmi, vcpu, num, retval) : ENXIO) #define VMSETCAP(vmi, vcpu, num, val) \ (ops != NULL ? (*ops->vmsetcap)(vmi, vcpu, num, val) : ENXIO) #define VLAPIC_INIT(vmi, vcpu) \ (ops != NULL ? (*ops->vlapic_init)(vmi, vcpu) : NULL) #define VLAPIC_CLEANUP(vmi, vlapic) \ (ops != NULL ? (*ops->vlapic_cleanup)(vmi, vlapic) : NULL) #define fpu_start_emulating() load_cr0(rcr0() | CR0_TS) #define fpu_stop_emulating() clts() SDT_PROVIDER_DEFINE(vmm); static MALLOC_DEFINE(M_VM, "vm", "vm"); /* statistics */ static VMM_STAT(VCPU_TOTAL_RUNTIME, "vcpu total runtime"); -SYSCTL_NODE(_hw, OID_AUTO, vmm, CTLFLAG_RW, NULL, NULL); +SYSCTL_NODE(_hw, OID_AUTO, vmm, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, + NULL); /* * Halt the guest if all vcpus are executing a HLT instruction with * interrupts disabled. */ static int halt_detection_enabled = 1; SYSCTL_INT(_hw_vmm, OID_AUTO, halt_detection, CTLFLAG_RDTUN, &halt_detection_enabled, 0, "Halt VM if all vcpus execute HLT with interrupts disabled"); static int vmm_ipinum; SYSCTL_INT(_hw_vmm, OID_AUTO, ipinum, CTLFLAG_RD, &vmm_ipinum, 0, "IPI vector used for vcpu notifications"); static int trace_guest_exceptions; SYSCTL_INT(_hw_vmm, OID_AUTO, trace_guest_exceptions, CTLFLAG_RDTUN, &trace_guest_exceptions, 0, "Trap into hypervisor on all guest exceptions and reflect them back"); static void vm_free_memmap(struct vm *vm, int ident); static bool sysmem_mapping(struct vm *vm, struct mem_map *mm); static void vcpu_notify_event_locked(struct vcpu *vcpu, bool lapic_intr); #ifdef KTR static const char * vcpu_state2str(enum vcpu_state state) { switch (state) { case VCPU_IDLE: return ("idle"); case VCPU_FROZEN: return ("frozen"); case VCPU_RUNNING: return ("running"); case VCPU_SLEEPING: return ("sleeping"); default: return ("unknown"); } } #endif static void vcpu_cleanup(struct vm *vm, int i, bool destroy) { struct vcpu *vcpu = &vm->vcpu[i]; VLAPIC_CLEANUP(vm->cookie, vcpu->vlapic); if (destroy) { vmm_stat_free(vcpu->stats); fpu_save_area_free(vcpu->guestfpu); } } static void vcpu_init(struct vm *vm, int vcpu_id, bool create) { struct vcpu *vcpu; KASSERT(vcpu_id >= 0 && vcpu_id < vm->maxcpus, ("vcpu_init: invalid vcpu %d", vcpu_id)); vcpu = &vm->vcpu[vcpu_id]; if (create) { KASSERT(!vcpu_lock_initialized(vcpu), ("vcpu %d already " "initialized", vcpu_id)); vcpu_lock_init(vcpu); vcpu->state = VCPU_IDLE; vcpu->hostcpu = NOCPU; vcpu->guestfpu = fpu_save_area_alloc(); vcpu->stats = vmm_stat_alloc(); } vcpu->vlapic = VLAPIC_INIT(vm->cookie, vcpu_id); vm_set_x2apic_state(vm, vcpu_id, X2APIC_DISABLED); vcpu->reqidle = 0; vcpu->exitintinfo = 0; vcpu->nmi_pending = 0; vcpu->extint_pending = 0; vcpu->exception_pending = 0; vcpu->guest_xcr0 = XFEATURE_ENABLED_X87; fpu_save_area_reset(vcpu->guestfpu); vmm_stat_init(vcpu->stats); } int vcpu_trace_exceptions(struct vm *vm, int vcpuid) { return (trace_guest_exceptions); } struct vm_exit * vm_exitinfo(struct vm *vm, int cpuid) { struct vcpu *vcpu; if (cpuid < 0 || cpuid >= vm->maxcpus) panic("vm_exitinfo: invalid cpuid %d", cpuid); vcpu = &vm->vcpu[cpuid]; return (&vcpu->exitinfo); } static void vmm_resume(void) { VMM_RESUME(); } static int vmm_init(void) { int error; vmm_host_state_init(); vmm_ipinum = lapic_ipi_alloc(pti ? &IDTVEC(justreturn1_pti) : &IDTVEC(justreturn)); if (vmm_ipinum < 0) vmm_ipinum = IPI_AST; error = vmm_mem_init(); if (error) return (error); if (vmm_is_intel()) ops = &vmm_ops_intel; else if (vmm_is_svm()) ops = &vmm_ops_amd; else return (ENXIO); vmm_resume_p = vmm_resume; return (VMM_INIT(vmm_ipinum)); } static int vmm_handler(module_t mod, int what, void *arg) { int error; switch (what) { case MOD_LOAD: vmmdev_init(); error = vmm_init(); if (error == 0) vmm_initialized = 1; break; case MOD_UNLOAD: error = vmmdev_cleanup(); if (error == 0) { vmm_resume_p = NULL; iommu_cleanup(); if (vmm_ipinum != IPI_AST) lapic_ipi_free(vmm_ipinum); error = VMM_CLEANUP(); /* * Something bad happened - prevent new * VMs from being created */ if (error) vmm_initialized = 0; } break; default: error = 0; break; } return (error); } static moduledata_t vmm_kmod = { "vmm", vmm_handler, NULL }; /* * vmm initialization has the following dependencies: * * - VT-x initialization requires smp_rendezvous() and therefore must happen * after SMP is fully functional (after SI_SUB_SMP). */ DECLARE_MODULE(vmm, vmm_kmod, SI_SUB_SMP + 1, SI_ORDER_ANY); MODULE_VERSION(vmm, 1); static void vm_init(struct vm *vm, bool create) { int i; vm->cookie = VMINIT(vm, vmspace_pmap(vm->vmspace)); vm->iommu = NULL; vm->vioapic = vioapic_init(vm); vm->vhpet = vhpet_init(vm); vm->vatpic = vatpic_init(vm); vm->vatpit = vatpit_init(vm); vm->vpmtmr = vpmtmr_init(vm); if (create) vm->vrtc = vrtc_init(vm); CPU_ZERO(&vm->active_cpus); CPU_ZERO(&vm->debug_cpus); vm->suspend = 0; CPU_ZERO(&vm->suspended_cpus); for (i = 0; i < vm->maxcpus; i++) vcpu_init(vm, i, create); } /* * The default CPU topology is a single thread per package. */ u_int cores_per_package = 1; u_int threads_per_core = 1; int vm_create(const char *name, struct vm **retvm) { struct vm *vm; struct vmspace *vmspace; /* * If vmm.ko could not be successfully initialized then don't attempt * to create the virtual machine. */ if (!vmm_initialized) return (ENXIO); if (name == NULL || strlen(name) >= VM_MAX_NAMELEN) return (EINVAL); vmspace = VMSPACE_ALLOC(0, VM_MAXUSER_ADDRESS); if (vmspace == NULL) return (ENOMEM); vm = malloc(sizeof(struct vm), M_VM, M_WAITOK | M_ZERO); strcpy(vm->name, name); vm->vmspace = vmspace; mtx_init(&vm->rendezvous_mtx, "vm rendezvous lock", 0, MTX_DEF); vm->sockets = 1; vm->cores = cores_per_package; /* XXX backwards compatibility */ vm->threads = threads_per_core; /* XXX backwards compatibility */ vm->maxcpus = VM_MAXCPU; /* XXX temp to keep code working */ vm_init(vm, true); *retvm = vm; return (0); } void vm_get_topology(struct vm *vm, uint16_t *sockets, uint16_t *cores, uint16_t *threads, uint16_t *maxcpus) { *sockets = vm->sockets; *cores = vm->cores; *threads = vm->threads; *maxcpus = vm->maxcpus; } uint16_t vm_get_maxcpus(struct vm *vm) { return (vm->maxcpus); } int vm_set_topology(struct vm *vm, uint16_t sockets, uint16_t cores, uint16_t threads, uint16_t maxcpus) { if (maxcpus != 0) return (EINVAL); /* XXX remove when supported */ if ((sockets * cores * threads) > vm->maxcpus) return (EINVAL); /* XXX need to check sockets * cores * threads == vCPU, how? */ vm->sockets = sockets; vm->cores = cores; vm->threads = threads; vm->maxcpus = VM_MAXCPU; /* XXX temp to keep code working */ return(0); } static void vm_cleanup(struct vm *vm, bool destroy) { struct mem_map *mm; int i; ppt_unassign_all(vm); if (vm->iommu != NULL) iommu_destroy_domain(vm->iommu); if (destroy) vrtc_cleanup(vm->vrtc); else vrtc_reset(vm->vrtc); vpmtmr_cleanup(vm->vpmtmr); vatpit_cleanup(vm->vatpit); vhpet_cleanup(vm->vhpet); vatpic_cleanup(vm->vatpic); vioapic_cleanup(vm->vioapic); for (i = 0; i < vm->maxcpus; i++) vcpu_cleanup(vm, i, destroy); VMCLEANUP(vm->cookie); /* * System memory is removed from the guest address space only when * the VM is destroyed. This is because the mapping remains the same * across VM reset. * * Device memory can be relocated by the guest (e.g. using PCI BARs) * so those mappings are removed on a VM reset. */ for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (destroy || !sysmem_mapping(vm, mm)) vm_free_memmap(vm, i); } if (destroy) { for (i = 0; i < VM_MAX_MEMSEGS; i++) vm_free_memseg(vm, i); VMSPACE_FREE(vm->vmspace); vm->vmspace = NULL; } } void vm_destroy(struct vm *vm) { vm_cleanup(vm, true); free(vm, M_VM); } int vm_reinit(struct vm *vm) { int error; /* * A virtual machine can be reset only if all vcpus are suspended. */ if (CPU_CMP(&vm->suspended_cpus, &vm->active_cpus) == 0) { vm_cleanup(vm, false); vm_init(vm, false); error = 0; } else { error = EBUSY; } return (error); } const char * vm_name(struct vm *vm) { return (vm->name); } int vm_map_mmio(struct vm *vm, vm_paddr_t gpa, size_t len, vm_paddr_t hpa) { vm_object_t obj; if ((obj = vmm_mmio_alloc(vm->vmspace, gpa, len, hpa)) == NULL) return (ENOMEM); else return (0); } int vm_unmap_mmio(struct vm *vm, vm_paddr_t gpa, size_t len) { vmm_mmio_free(vm->vmspace, gpa, len); return (0); } /* * Return 'true' if 'gpa' is allocated in the guest address space. * * This function is called in the context of a running vcpu which acts as * an implicit lock on 'vm->mem_maps[]'. */ bool vm_mem_allocated(struct vm *vm, int vcpuid, vm_paddr_t gpa) { struct mem_map *mm; int i; #ifdef INVARIANTS int hostcpu, state; state = vcpu_get_state(vm, vcpuid, &hostcpu); KASSERT(state == VCPU_RUNNING && hostcpu == curcpu, ("%s: invalid vcpu state %d/%d", __func__, state, hostcpu)); #endif for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (mm->len != 0 && gpa >= mm->gpa && gpa < mm->gpa + mm->len) return (true); /* 'gpa' is sysmem or devmem */ } if (ppt_is_mmio(vm, gpa)) return (true); /* 'gpa' is pci passthru mmio */ return (false); } int vm_alloc_memseg(struct vm *vm, int ident, size_t len, bool sysmem) { struct mem_seg *seg; vm_object_t obj; if (ident < 0 || ident >= VM_MAX_MEMSEGS) return (EINVAL); if (len == 0 || (len & PAGE_MASK)) return (EINVAL); seg = &vm->mem_segs[ident]; if (seg->object != NULL) { if (seg->len == len && seg->sysmem == sysmem) return (EEXIST); else return (EINVAL); } obj = vm_object_allocate(OBJT_DEFAULT, len >> PAGE_SHIFT); if (obj == NULL) return (ENOMEM); seg->len = len; seg->object = obj; seg->sysmem = sysmem; return (0); } int vm_get_memseg(struct vm *vm, int ident, size_t *len, bool *sysmem, vm_object_t *objptr) { struct mem_seg *seg; if (ident < 0 || ident >= VM_MAX_MEMSEGS) return (EINVAL); seg = &vm->mem_segs[ident]; if (len) *len = seg->len; if (sysmem) *sysmem = seg->sysmem; if (objptr) *objptr = seg->object; return (0); } void vm_free_memseg(struct vm *vm, int ident) { struct mem_seg *seg; KASSERT(ident >= 0 && ident < VM_MAX_MEMSEGS, ("%s: invalid memseg ident %d", __func__, ident)); seg = &vm->mem_segs[ident]; if (seg->object != NULL) { vm_object_deallocate(seg->object); bzero(seg, sizeof(struct mem_seg)); } } int vm_mmap_memseg(struct vm *vm, vm_paddr_t gpa, int segid, vm_ooffset_t first, size_t len, int prot, int flags) { struct mem_seg *seg; struct mem_map *m, *map; vm_ooffset_t last; int i, error; if (prot == 0 || (prot & ~(VM_PROT_ALL)) != 0) return (EINVAL); if (flags & ~VM_MEMMAP_F_WIRED) return (EINVAL); if (segid < 0 || segid >= VM_MAX_MEMSEGS) return (EINVAL); seg = &vm->mem_segs[segid]; if (seg->object == NULL) return (EINVAL); last = first + len; if (first < 0 || first >= last || last > seg->len) return (EINVAL); if ((gpa | first | last) & PAGE_MASK) return (EINVAL); map = NULL; for (i = 0; i < VM_MAX_MEMMAPS; i++) { m = &vm->mem_maps[i]; if (m->len == 0) { map = m; break; } } if (map == NULL) return (ENOSPC); error = vm_map_find(&vm->vmspace->vm_map, seg->object, first, &gpa, len, 0, VMFS_NO_SPACE, prot, prot, 0); if (error != KERN_SUCCESS) return (EFAULT); vm_object_reference(seg->object); if (flags & VM_MEMMAP_F_WIRED) { error = vm_map_wire(&vm->vmspace->vm_map, gpa, gpa + len, VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); if (error != KERN_SUCCESS) { vm_map_remove(&vm->vmspace->vm_map, gpa, gpa + len); return (error == KERN_RESOURCE_SHORTAGE ? ENOMEM : EFAULT); } } map->gpa = gpa; map->len = len; map->segoff = first; map->segid = segid; map->prot = prot; map->flags = flags; return (0); } int vm_mmap_getnext(struct vm *vm, vm_paddr_t *gpa, int *segid, vm_ooffset_t *segoff, size_t *len, int *prot, int *flags) { struct mem_map *mm, *mmnext; int i; mmnext = NULL; for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (mm->len == 0 || mm->gpa < *gpa) continue; if (mmnext == NULL || mm->gpa < mmnext->gpa) mmnext = mm; } if (mmnext != NULL) { *gpa = mmnext->gpa; if (segid) *segid = mmnext->segid; if (segoff) *segoff = mmnext->segoff; if (len) *len = mmnext->len; if (prot) *prot = mmnext->prot; if (flags) *flags = mmnext->flags; return (0); } else { return (ENOENT); } } static void vm_free_memmap(struct vm *vm, int ident) { struct mem_map *mm; int error; mm = &vm->mem_maps[ident]; if (mm->len) { error = vm_map_remove(&vm->vmspace->vm_map, mm->gpa, mm->gpa + mm->len); KASSERT(error == KERN_SUCCESS, ("%s: vm_map_remove error %d", __func__, error)); bzero(mm, sizeof(struct mem_map)); } } static __inline bool sysmem_mapping(struct vm *vm, struct mem_map *mm) { if (mm->len != 0 && vm->mem_segs[mm->segid].sysmem) return (true); else return (false); } vm_paddr_t vmm_sysmem_maxaddr(struct vm *vm) { struct mem_map *mm; vm_paddr_t maxaddr; int i; maxaddr = 0; for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (sysmem_mapping(vm, mm)) { if (maxaddr < mm->gpa + mm->len) maxaddr = mm->gpa + mm->len; } } return (maxaddr); } static void vm_iommu_modify(struct vm *vm, bool map) { int i, sz; vm_paddr_t gpa, hpa; struct mem_map *mm; void *vp, *cookie, *host_domain; sz = PAGE_SIZE; host_domain = iommu_host_domain(); for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (!sysmem_mapping(vm, mm)) continue; if (map) { KASSERT((mm->flags & VM_MEMMAP_F_IOMMU) == 0, ("iommu map found invalid memmap %#lx/%#lx/%#x", mm->gpa, mm->len, mm->flags)); if ((mm->flags & VM_MEMMAP_F_WIRED) == 0) continue; mm->flags |= VM_MEMMAP_F_IOMMU; } else { if ((mm->flags & VM_MEMMAP_F_IOMMU) == 0) continue; mm->flags &= ~VM_MEMMAP_F_IOMMU; KASSERT((mm->flags & VM_MEMMAP_F_WIRED) != 0, ("iommu unmap found invalid memmap %#lx/%#lx/%#x", mm->gpa, mm->len, mm->flags)); } gpa = mm->gpa; while (gpa < mm->gpa + mm->len) { vp = vm_gpa_hold(vm, -1, gpa, PAGE_SIZE, VM_PROT_WRITE, &cookie); KASSERT(vp != NULL, ("vm(%s) could not map gpa %#lx", vm_name(vm), gpa)); vm_gpa_release(cookie); hpa = DMAP_TO_PHYS((uintptr_t)vp); if (map) { iommu_create_mapping(vm->iommu, gpa, hpa, sz); iommu_remove_mapping(host_domain, hpa, sz); } else { iommu_remove_mapping(vm->iommu, gpa, sz); iommu_create_mapping(host_domain, hpa, hpa, sz); } gpa += PAGE_SIZE; } } /* * Invalidate the cached translations associated with the domain * from which pages were removed. */ if (map) iommu_invalidate_tlb(host_domain); else iommu_invalidate_tlb(vm->iommu); } #define vm_iommu_unmap(vm) vm_iommu_modify((vm), false) #define vm_iommu_map(vm) vm_iommu_modify((vm), true) int vm_unassign_pptdev(struct vm *vm, int bus, int slot, int func) { int error; error = ppt_unassign_device(vm, bus, slot, func); if (error) return (error); if (ppt_assigned_devices(vm) == 0) vm_iommu_unmap(vm); return (0); } int vm_assign_pptdev(struct vm *vm, int bus, int slot, int func) { int error; vm_paddr_t maxaddr; /* Set up the IOMMU to do the 'gpa' to 'hpa' translation */ if (ppt_assigned_devices(vm) == 0) { KASSERT(vm->iommu == NULL, ("vm_assign_pptdev: iommu must be NULL")); maxaddr = vmm_sysmem_maxaddr(vm); vm->iommu = iommu_create_domain(maxaddr); if (vm->iommu == NULL) return (ENXIO); vm_iommu_map(vm); } error = ppt_assign_device(vm, bus, slot, func); return (error); } void * vm_gpa_hold(struct vm *vm, int vcpuid, vm_paddr_t gpa, size_t len, int reqprot, void **cookie) { int i, count, pageoff; struct mem_map *mm; vm_page_t m; #ifdef INVARIANTS /* * All vcpus are frozen by ioctls that modify the memory map * (e.g. VM_MMAP_MEMSEG). Therefore 'vm->memmap[]' stability is * guaranteed if at least one vcpu is in the VCPU_FROZEN state. */ int state; KASSERT(vcpuid >= -1 && vcpuid < vm->maxcpus, ("%s: invalid vcpuid %d", __func__, vcpuid)); for (i = 0; i < vm->maxcpus; i++) { if (vcpuid != -1 && vcpuid != i) continue; state = vcpu_get_state(vm, i, NULL); KASSERT(state == VCPU_FROZEN, ("%s: invalid vcpu state %d", __func__, state)); } #endif pageoff = gpa & PAGE_MASK; if (len > PAGE_SIZE - pageoff) panic("vm_gpa_hold: invalid gpa/len: 0x%016lx/%lu", gpa, len); count = 0; for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (sysmem_mapping(vm, mm) && gpa >= mm->gpa && gpa < mm->gpa + mm->len) { count = vm_fault_quick_hold_pages(&vm->vmspace->vm_map, trunc_page(gpa), PAGE_SIZE, reqprot, &m, 1); break; } } if (count == 1) { *cookie = m; return ((void *)(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)) + pageoff)); } else { *cookie = NULL; return (NULL); } } void vm_gpa_release(void *cookie) { vm_page_t m = cookie; vm_page_unwire(m, PQ_ACTIVE); } int vm_get_register(struct vm *vm, int vcpu, int reg, uint64_t *retval) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (reg >= VM_REG_LAST) return (EINVAL); return (VMGETREG(vm->cookie, vcpu, reg, retval)); } int vm_set_register(struct vm *vm, int vcpuid, int reg, uint64_t val) { struct vcpu *vcpu; int error; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (reg >= VM_REG_LAST) return (EINVAL); error = VMSETREG(vm->cookie, vcpuid, reg, val); if (error || reg != VM_REG_GUEST_RIP) return (error); /* Set 'nextrip' to match the value of %rip */ VCPU_CTR1(vm, vcpuid, "Setting nextrip to %#lx", val); vcpu = &vm->vcpu[vcpuid]; vcpu->nextrip = val; return (0); } static bool is_descriptor_table(int reg) { switch (reg) { case VM_REG_GUEST_IDTR: case VM_REG_GUEST_GDTR: return (true); default: return (false); } } static bool is_segment_register(int reg) { switch (reg) { case VM_REG_GUEST_ES: case VM_REG_GUEST_CS: case VM_REG_GUEST_SS: case VM_REG_GUEST_DS: case VM_REG_GUEST_FS: case VM_REG_GUEST_GS: case VM_REG_GUEST_TR: case VM_REG_GUEST_LDTR: return (true); default: return (false); } } int vm_get_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *desc) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (!is_segment_register(reg) && !is_descriptor_table(reg)) return (EINVAL); return (VMGETDESC(vm->cookie, vcpu, reg, desc)); } int vm_set_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *desc) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (!is_segment_register(reg) && !is_descriptor_table(reg)) return (EINVAL); return (VMSETDESC(vm->cookie, vcpu, reg, desc)); } static void restore_guest_fpustate(struct vcpu *vcpu) { /* flush host state to the pcb */ fpuexit(curthread); /* restore guest FPU state */ fpu_stop_emulating(); fpurestore(vcpu->guestfpu); /* restore guest XCR0 if XSAVE is enabled in the host */ if (rcr4() & CR4_XSAVE) load_xcr(0, vcpu->guest_xcr0); /* * The FPU is now "dirty" with the guest's state so turn on emulation * to trap any access to the FPU by the host. */ fpu_start_emulating(); } static void save_guest_fpustate(struct vcpu *vcpu) { if ((rcr0() & CR0_TS) == 0) panic("fpu emulation not enabled in host!"); /* save guest XCR0 and restore host XCR0 */ if (rcr4() & CR4_XSAVE) { vcpu->guest_xcr0 = rxcr(0); load_xcr(0, vmm_get_host_xcr0()); } /* save guest FPU state */ fpu_stop_emulating(); fpusave(vcpu->guestfpu); fpu_start_emulating(); } static VMM_STAT(VCPU_IDLE_TICKS, "number of ticks vcpu was idle"); static int vcpu_set_state_locked(struct vm *vm, int vcpuid, enum vcpu_state newstate, bool from_idle) { struct vcpu *vcpu; int error; vcpu = &vm->vcpu[vcpuid]; vcpu_assert_locked(vcpu); /* * State transitions from the vmmdev_ioctl() must always begin from * the VCPU_IDLE state. This guarantees that there is only a single * ioctl() operating on a vcpu at any point. */ if (from_idle) { while (vcpu->state != VCPU_IDLE) { vcpu->reqidle = 1; vcpu_notify_event_locked(vcpu, false); VCPU_CTR1(vm, vcpuid, "vcpu state change from %s to " "idle requested", vcpu_state2str(vcpu->state)); msleep_spin(&vcpu->state, &vcpu->mtx, "vmstat", hz); } } else { KASSERT(vcpu->state != VCPU_IDLE, ("invalid transition from " "vcpu idle state")); } if (vcpu->state == VCPU_RUNNING) { KASSERT(vcpu->hostcpu == curcpu, ("curcpu %d and hostcpu %d " "mismatch for running vcpu", curcpu, vcpu->hostcpu)); } else { KASSERT(vcpu->hostcpu == NOCPU, ("Invalid hostcpu %d for a " "vcpu that is not running", vcpu->hostcpu)); } /* * The following state transitions are allowed: * IDLE -> FROZEN -> IDLE * FROZEN -> RUNNING -> FROZEN * FROZEN -> SLEEPING -> FROZEN */ switch (vcpu->state) { case VCPU_IDLE: case VCPU_RUNNING: case VCPU_SLEEPING: error = (newstate != VCPU_FROZEN); break; case VCPU_FROZEN: error = (newstate == VCPU_FROZEN); break; default: error = 1; break; } if (error) return (EBUSY); VCPU_CTR2(vm, vcpuid, "vcpu state changed from %s to %s", vcpu_state2str(vcpu->state), vcpu_state2str(newstate)); vcpu->state = newstate; if (newstate == VCPU_RUNNING) vcpu->hostcpu = curcpu; else vcpu->hostcpu = NOCPU; if (newstate == VCPU_IDLE) wakeup(&vcpu->state); return (0); } static void vcpu_require_state(struct vm *vm, int vcpuid, enum vcpu_state newstate) { int error; if ((error = vcpu_set_state(vm, vcpuid, newstate, false)) != 0) panic("Error %d setting state to %d\n", error, newstate); } static void vcpu_require_state_locked(struct vm *vm, int vcpuid, enum vcpu_state newstate) { int error; if ((error = vcpu_set_state_locked(vm, vcpuid, newstate, false)) != 0) panic("Error %d setting state to %d", error, newstate); } #define RENDEZVOUS_CTR0(vm, vcpuid, fmt) \ do { \ if (vcpuid >= 0) \ VCPU_CTR0(vm, vcpuid, fmt); \ else \ VM_CTR0(vm, fmt); \ } while (0) static int vm_handle_rendezvous(struct vm *vm, int vcpuid) { struct thread *td; int error; KASSERT(vcpuid == -1 || (vcpuid >= 0 && vcpuid < vm->maxcpus), ("vm_handle_rendezvous: invalid vcpuid %d", vcpuid)); error = 0; td = curthread; mtx_lock(&vm->rendezvous_mtx); while (vm->rendezvous_func != NULL) { /* 'rendezvous_req_cpus' must be a subset of 'active_cpus' */ CPU_AND(&vm->rendezvous_req_cpus, &vm->active_cpus); if (vcpuid != -1 && CPU_ISSET(vcpuid, &vm->rendezvous_req_cpus) && !CPU_ISSET(vcpuid, &vm->rendezvous_done_cpus)) { VCPU_CTR0(vm, vcpuid, "Calling rendezvous func"); (*vm->rendezvous_func)(vm, vcpuid, vm->rendezvous_arg); CPU_SET(vcpuid, &vm->rendezvous_done_cpus); } if (CPU_CMP(&vm->rendezvous_req_cpus, &vm->rendezvous_done_cpus) == 0) { VCPU_CTR0(vm, vcpuid, "Rendezvous completed"); vm->rendezvous_func = NULL; wakeup(&vm->rendezvous_func); break; } RENDEZVOUS_CTR0(vm, vcpuid, "Wait for rendezvous completion"); mtx_sleep(&vm->rendezvous_func, &vm->rendezvous_mtx, 0, "vmrndv", hz); if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) { mtx_unlock(&vm->rendezvous_mtx); error = thread_check_susp(td, true); if (error != 0) return (error); mtx_lock(&vm->rendezvous_mtx); } } mtx_unlock(&vm->rendezvous_mtx); return (0); } /* * Emulate a guest 'hlt' by sleeping until the vcpu is ready to run. */ static int vm_handle_hlt(struct vm *vm, int vcpuid, bool intr_disabled, bool *retu) { struct vcpu *vcpu; const char *wmesg; struct thread *td; int error, t, vcpu_halted, vm_halted; KASSERT(!CPU_ISSET(vcpuid, &vm->halted_cpus), ("vcpu already halted")); vcpu = &vm->vcpu[vcpuid]; vcpu_halted = 0; vm_halted = 0; error = 0; td = curthread; vcpu_lock(vcpu); while (1) { /* * Do a final check for pending NMI or interrupts before * really putting this thread to sleep. Also check for * software events that would cause this vcpu to wakeup. * * These interrupts/events could have happened after the * vcpu returned from VMRUN() and before it acquired the * vcpu lock above. */ if (vm->rendezvous_func != NULL || vm->suspend || vcpu->reqidle) break; if (vm_nmi_pending(vm, vcpuid)) break; if (!intr_disabled) { if (vm_extint_pending(vm, vcpuid) || vlapic_pending_intr(vcpu->vlapic, NULL)) { break; } } /* Don't go to sleep if the vcpu thread needs to yield */ if (vcpu_should_yield(vm, vcpuid)) break; if (vcpu_debugged(vm, vcpuid)) break; /* * Some Linux guests implement "halt" by having all vcpus * execute HLT with interrupts disabled. 'halted_cpus' keeps * track of the vcpus that have entered this state. When all * vcpus enter the halted state the virtual machine is halted. */ if (intr_disabled) { wmesg = "vmhalt"; VCPU_CTR0(vm, vcpuid, "Halted"); if (!vcpu_halted && halt_detection_enabled) { vcpu_halted = 1; CPU_SET_ATOMIC(vcpuid, &vm->halted_cpus); } if (CPU_CMP(&vm->halted_cpus, &vm->active_cpus) == 0) { vm_halted = 1; break; } } else { wmesg = "vmidle"; } t = ticks; vcpu_require_state_locked(vm, vcpuid, VCPU_SLEEPING); /* * XXX msleep_spin() cannot be interrupted by signals so * wake up periodically to check pending signals. */ msleep_spin(vcpu, &vcpu->mtx, wmesg, hz); vcpu_require_state_locked(vm, vcpuid, VCPU_FROZEN); vmm_stat_incr(vm, vcpuid, VCPU_IDLE_TICKS, ticks - t); if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) { vcpu_unlock(vcpu); error = thread_check_susp(td, false); if (error != 0) return (error); vcpu_lock(vcpu); } } if (vcpu_halted) CPU_CLR_ATOMIC(vcpuid, &vm->halted_cpus); vcpu_unlock(vcpu); if (vm_halted) vm_suspend(vm, VM_SUSPEND_HALT); return (0); } static int vm_handle_paging(struct vm *vm, int vcpuid, bool *retu) { int rv, ftype; struct vm_map *map; struct vcpu *vcpu; struct vm_exit *vme; vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; KASSERT(vme->inst_length == 0, ("%s: invalid inst_length %d", __func__, vme->inst_length)); ftype = vme->u.paging.fault_type; KASSERT(ftype == VM_PROT_READ || ftype == VM_PROT_WRITE || ftype == VM_PROT_EXECUTE, ("vm_handle_paging: invalid fault_type %d", ftype)); if (ftype == VM_PROT_READ || ftype == VM_PROT_WRITE) { rv = pmap_emulate_accessed_dirty(vmspace_pmap(vm->vmspace), vme->u.paging.gpa, ftype); if (rv == 0) { VCPU_CTR2(vm, vcpuid, "%s bit emulation for gpa %#lx", ftype == VM_PROT_READ ? "accessed" : "dirty", vme->u.paging.gpa); goto done; } } map = &vm->vmspace->vm_map; rv = vm_fault(map, vme->u.paging.gpa, ftype, VM_FAULT_NORMAL, NULL); VCPU_CTR3(vm, vcpuid, "vm_handle_paging rv = %d, gpa = %#lx, " "ftype = %d", rv, vme->u.paging.gpa, ftype); if (rv != KERN_SUCCESS) return (EFAULT); done: return (0); } static int vm_handle_inst_emul(struct vm *vm, int vcpuid, bool *retu) { struct vie *vie; struct vcpu *vcpu; struct vm_exit *vme; uint64_t gla, gpa, cs_base; struct vm_guest_paging *paging; mem_region_read_t mread; mem_region_write_t mwrite; enum vm_cpu_mode cpu_mode; int cs_d, error, fault; vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; KASSERT(vme->inst_length == 0, ("%s: invalid inst_length %d", __func__, vme->inst_length)); gla = vme->u.inst_emul.gla; gpa = vme->u.inst_emul.gpa; cs_base = vme->u.inst_emul.cs_base; cs_d = vme->u.inst_emul.cs_d; vie = &vme->u.inst_emul.vie; paging = &vme->u.inst_emul.paging; cpu_mode = paging->cpu_mode; VCPU_CTR1(vm, vcpuid, "inst_emul fault accessing gpa %#lx", gpa); /* Fetch, decode and emulate the faulting instruction */ if (vie->num_valid == 0) { error = vmm_fetch_instruction(vm, vcpuid, paging, vme->rip + cs_base, VIE_INST_SIZE, vie, &fault); } else { /* * The instruction bytes have already been copied into 'vie' */ error = fault = 0; } if (error || fault) return (error); if (vmm_decode_instruction(vm, vcpuid, gla, cpu_mode, cs_d, vie) != 0) { VCPU_CTR1(vm, vcpuid, "Error decoding instruction at %#lx", vme->rip + cs_base); *retu = true; /* dump instruction bytes in userspace */ return (0); } /* * Update 'nextrip' based on the length of the emulated instruction. */ vme->inst_length = vie->num_processed; vcpu->nextrip += vie->num_processed; VCPU_CTR1(vm, vcpuid, "nextrip updated to %#lx after instruction " "decoding", vcpu->nextrip); /* return to userland unless this is an in-kernel emulated device */ if (gpa >= DEFAULT_APIC_BASE && gpa < DEFAULT_APIC_BASE + PAGE_SIZE) { mread = lapic_mmio_read; mwrite = lapic_mmio_write; } else if (gpa >= VIOAPIC_BASE && gpa < VIOAPIC_BASE + VIOAPIC_SIZE) { mread = vioapic_mmio_read; mwrite = vioapic_mmio_write; } else if (gpa >= VHPET_BASE && gpa < VHPET_BASE + VHPET_SIZE) { mread = vhpet_mmio_read; mwrite = vhpet_mmio_write; } else { *retu = true; return (0); } error = vmm_emulate_instruction(vm, vcpuid, gpa, vie, paging, mread, mwrite, retu); return (error); } static int vm_handle_suspend(struct vm *vm, int vcpuid, bool *retu) { int error, i; struct vcpu *vcpu; struct thread *td; error = 0; vcpu = &vm->vcpu[vcpuid]; td = curthread; CPU_SET_ATOMIC(vcpuid, &vm->suspended_cpus); /* * Wait until all 'active_cpus' have suspended themselves. * * Since a VM may be suspended at any time including when one or * more vcpus are doing a rendezvous we need to call the rendezvous * handler while we are waiting to prevent a deadlock. */ vcpu_lock(vcpu); while (error == 0) { if (CPU_CMP(&vm->suspended_cpus, &vm->active_cpus) == 0) { VCPU_CTR0(vm, vcpuid, "All vcpus suspended"); break; } if (vm->rendezvous_func == NULL) { VCPU_CTR0(vm, vcpuid, "Sleeping during suspend"); vcpu_require_state_locked(vm, vcpuid, VCPU_SLEEPING); msleep_spin(vcpu, &vcpu->mtx, "vmsusp", hz); vcpu_require_state_locked(vm, vcpuid, VCPU_FROZEN); if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) { vcpu_unlock(vcpu); error = thread_check_susp(td, false); vcpu_lock(vcpu); } } else { VCPU_CTR0(vm, vcpuid, "Rendezvous during suspend"); vcpu_unlock(vcpu); error = vm_handle_rendezvous(vm, vcpuid); vcpu_lock(vcpu); } } vcpu_unlock(vcpu); /* * Wakeup the other sleeping vcpus and return to userspace. */ for (i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &vm->suspended_cpus)) { vcpu_notify_event(vm, i, false); } } *retu = true; return (error); } static int vm_handle_reqidle(struct vm *vm, int vcpuid, bool *retu) { struct vcpu *vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); KASSERT(vcpu->reqidle, ("invalid vcpu reqidle %d", vcpu->reqidle)); vcpu->reqidle = 0; vcpu_unlock(vcpu); *retu = true; return (0); } int vm_suspend(struct vm *vm, enum vm_suspend_how how) { int i; if (how <= VM_SUSPEND_NONE || how >= VM_SUSPEND_LAST) return (EINVAL); if (atomic_cmpset_int(&vm->suspend, 0, how) == 0) { VM_CTR2(vm, "virtual machine already suspended %d/%d", vm->suspend, how); return (EALREADY); } VM_CTR1(vm, "virtual machine successfully suspended %d", how); /* * Notify all active vcpus that they are now suspended. */ for (i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &vm->active_cpus)) vcpu_notify_event(vm, i, false); } return (0); } void vm_exit_suspended(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; KASSERT(vm->suspend > VM_SUSPEND_NONE && vm->suspend < VM_SUSPEND_LAST, ("vm_exit_suspended: invalid suspend type %d", vm->suspend)); vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_SUSPENDED; vmexit->u.suspended.how = vm->suspend; } void vm_exit_debug(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_DEBUG; } void vm_exit_rendezvous(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; KASSERT(vm->rendezvous_func != NULL, ("rendezvous not in progress")); vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_RENDEZVOUS; vmm_stat_incr(vm, vcpuid, VMEXIT_RENDEZVOUS, 1); } void vm_exit_reqidle(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_REQIDLE; vmm_stat_incr(vm, vcpuid, VMEXIT_REQIDLE, 1); } void vm_exit_astpending(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_BOGUS; vmm_stat_incr(vm, vcpuid, VMEXIT_ASTPENDING, 1); } int vm_run(struct vm *vm, struct vm_run *vmrun) { struct vm_eventinfo evinfo; int error, vcpuid; struct vcpu *vcpu; struct pcb *pcb; uint64_t tscval; struct vm_exit *vme; bool retu, intr_disabled; pmap_t pmap; vcpuid = vmrun->cpuid; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (!CPU_ISSET(vcpuid, &vm->active_cpus)) return (EINVAL); if (CPU_ISSET(vcpuid, &vm->suspended_cpus)) return (EINVAL); pmap = vmspace_pmap(vm->vmspace); vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; evinfo.rptr = &vm->rendezvous_func; evinfo.sptr = &vm->suspend; evinfo.iptr = &vcpu->reqidle; restart: critical_enter(); KASSERT(!CPU_ISSET(curcpu, &pmap->pm_active), ("vm_run: absurd pm_active")); tscval = rdtsc(); pcb = PCPU_GET(curpcb); set_pcb_flags(pcb, PCB_FULL_IRET); restore_guest_fpustate(vcpu); vcpu_require_state(vm, vcpuid, VCPU_RUNNING); error = VMRUN(vm->cookie, vcpuid, vcpu->nextrip, pmap, &evinfo); vcpu_require_state(vm, vcpuid, VCPU_FROZEN); save_guest_fpustate(vcpu); vmm_stat_incr(vm, vcpuid, VCPU_TOTAL_RUNTIME, rdtsc() - tscval); critical_exit(); if (error == 0) { retu = false; vcpu->nextrip = vme->rip + vme->inst_length; switch (vme->exitcode) { case VM_EXITCODE_REQIDLE: error = vm_handle_reqidle(vm, vcpuid, &retu); break; case VM_EXITCODE_SUSPENDED: error = vm_handle_suspend(vm, vcpuid, &retu); break; case VM_EXITCODE_IOAPIC_EOI: vioapic_process_eoi(vm, vcpuid, vme->u.ioapic_eoi.vector); break; case VM_EXITCODE_RENDEZVOUS: error = vm_handle_rendezvous(vm, vcpuid); break; case VM_EXITCODE_HLT: intr_disabled = ((vme->u.hlt.rflags & PSL_I) == 0); error = vm_handle_hlt(vm, vcpuid, intr_disabled, &retu); break; case VM_EXITCODE_PAGING: error = vm_handle_paging(vm, vcpuid, &retu); break; case VM_EXITCODE_INST_EMUL: error = vm_handle_inst_emul(vm, vcpuid, &retu); break; case VM_EXITCODE_INOUT: case VM_EXITCODE_INOUT_STR: error = vm_handle_inout(vm, vcpuid, vme, &retu); break; case VM_EXITCODE_MONITOR: case VM_EXITCODE_MWAIT: case VM_EXITCODE_VMINSN: vm_inject_ud(vm, vcpuid); break; default: retu = true; /* handled in userland */ break; } } if (error == 0 && retu == false) goto restart; VCPU_CTR2(vm, vcpuid, "retu %d/%d", error, vme->exitcode); /* copy the exit information */ bcopy(vme, &vmrun->vm_exit, sizeof(struct vm_exit)); return (error); } int vm_restart_instruction(void *arg, int vcpuid) { struct vm *vm; struct vcpu *vcpu; enum vcpu_state state; uint64_t rip; int error; vm = arg; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; state = vcpu_get_state(vm, vcpuid, NULL); if (state == VCPU_RUNNING) { /* * When a vcpu is "running" the next instruction is determined * by adding 'rip' and 'inst_length' in the vcpu's 'exitinfo'. * Thus setting 'inst_length' to zero will cause the current * instruction to be restarted. */ vcpu->exitinfo.inst_length = 0; VCPU_CTR1(vm, vcpuid, "restarting instruction at %#lx by " "setting inst_length to zero", vcpu->exitinfo.rip); } else if (state == VCPU_FROZEN) { /* * When a vcpu is "frozen" it is outside the critical section * around VMRUN() and 'nextrip' points to the next instruction. * Thus instruction restart is achieved by setting 'nextrip' * to the vcpu's %rip. */ error = vm_get_register(vm, vcpuid, VM_REG_GUEST_RIP, &rip); KASSERT(!error, ("%s: error %d getting rip", __func__, error)); VCPU_CTR2(vm, vcpuid, "restarting instruction by updating " "nextrip from %#lx to %#lx", vcpu->nextrip, rip); vcpu->nextrip = rip; } else { panic("%s: invalid state %d", __func__, state); } return (0); } int vm_exit_intinfo(struct vm *vm, int vcpuid, uint64_t info) { struct vcpu *vcpu; int type, vector; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; if (info & VM_INTINFO_VALID) { type = info & VM_INTINFO_TYPE; vector = info & 0xff; if (type == VM_INTINFO_NMI && vector != IDT_NMI) return (EINVAL); if (type == VM_INTINFO_HWEXCEPTION && vector >= 32) return (EINVAL); if (info & VM_INTINFO_RSVD) return (EINVAL); } else { info = 0; } VCPU_CTR2(vm, vcpuid, "%s: info1(%#lx)", __func__, info); vcpu->exitintinfo = info; return (0); } enum exc_class { EXC_BENIGN, EXC_CONTRIBUTORY, EXC_PAGEFAULT }; #define IDT_VE 20 /* Virtualization Exception (Intel specific) */ static enum exc_class exception_class(uint64_t info) { int type, vector; KASSERT(info & VM_INTINFO_VALID, ("intinfo must be valid: %#lx", info)); type = info & VM_INTINFO_TYPE; vector = info & 0xff; /* Table 6-4, "Interrupt and Exception Classes", Intel SDM, Vol 3 */ switch (type) { case VM_INTINFO_HWINTR: case VM_INTINFO_SWINTR: case VM_INTINFO_NMI: return (EXC_BENIGN); default: /* * Hardware exception. * * SVM and VT-x use identical type values to represent NMI, * hardware interrupt and software interrupt. * * SVM uses type '3' for all exceptions. VT-x uses type '3' * for exceptions except #BP and #OF. #BP and #OF use a type * value of '5' or '6'. Therefore we don't check for explicit * values of 'type' to classify 'intinfo' into a hardware * exception. */ break; } switch (vector) { case IDT_PF: case IDT_VE: return (EXC_PAGEFAULT); case IDT_DE: case IDT_TS: case IDT_NP: case IDT_SS: case IDT_GP: return (EXC_CONTRIBUTORY); default: return (EXC_BENIGN); } } static int nested_fault(struct vm *vm, int vcpuid, uint64_t info1, uint64_t info2, uint64_t *retinfo) { enum exc_class exc1, exc2; int type1, vector1; KASSERT(info1 & VM_INTINFO_VALID, ("info1 %#lx is not valid", info1)); KASSERT(info2 & VM_INTINFO_VALID, ("info2 %#lx is not valid", info2)); /* * If an exception occurs while attempting to call the double-fault * handler the processor enters shutdown mode (aka triple fault). */ type1 = info1 & VM_INTINFO_TYPE; vector1 = info1 & 0xff; if (type1 == VM_INTINFO_HWEXCEPTION && vector1 == IDT_DF) { VCPU_CTR2(vm, vcpuid, "triple fault: info1(%#lx), info2(%#lx)", info1, info2); vm_suspend(vm, VM_SUSPEND_TRIPLEFAULT); *retinfo = 0; return (0); } /* * Table 6-5 "Conditions for Generating a Double Fault", Intel SDM, Vol3 */ exc1 = exception_class(info1); exc2 = exception_class(info2); if ((exc1 == EXC_CONTRIBUTORY && exc2 == EXC_CONTRIBUTORY) || (exc1 == EXC_PAGEFAULT && exc2 != EXC_BENIGN)) { /* Convert nested fault into a double fault. */ *retinfo = IDT_DF; *retinfo |= VM_INTINFO_VALID | VM_INTINFO_HWEXCEPTION; *retinfo |= VM_INTINFO_DEL_ERRCODE; } else { /* Handle exceptions serially */ *retinfo = info2; } return (1); } static uint64_t vcpu_exception_intinfo(struct vcpu *vcpu) { uint64_t info = 0; if (vcpu->exception_pending) { info = vcpu->exc_vector & 0xff; info |= VM_INTINFO_VALID | VM_INTINFO_HWEXCEPTION; if (vcpu->exc_errcode_valid) { info |= VM_INTINFO_DEL_ERRCODE; info |= (uint64_t)vcpu->exc_errcode << 32; } } return (info); } int vm_entry_intinfo(struct vm *vm, int vcpuid, uint64_t *retinfo) { struct vcpu *vcpu; uint64_t info1, info2; int valid; KASSERT(vcpuid >= 0 && vcpuid < vm->maxcpus, ("invalid vcpu %d", vcpuid)); vcpu = &vm->vcpu[vcpuid]; info1 = vcpu->exitintinfo; vcpu->exitintinfo = 0; info2 = 0; if (vcpu->exception_pending) { info2 = vcpu_exception_intinfo(vcpu); vcpu->exception_pending = 0; VCPU_CTR2(vm, vcpuid, "Exception %d delivered: %#lx", vcpu->exc_vector, info2); } if ((info1 & VM_INTINFO_VALID) && (info2 & VM_INTINFO_VALID)) { valid = nested_fault(vm, vcpuid, info1, info2, retinfo); } else if (info1 & VM_INTINFO_VALID) { *retinfo = info1; valid = 1; } else if (info2 & VM_INTINFO_VALID) { *retinfo = info2; valid = 1; } else { valid = 0; } if (valid) { VCPU_CTR4(vm, vcpuid, "%s: info1(%#lx), info2(%#lx), " "retinfo(%#lx)", __func__, info1, info2, *retinfo); } return (valid); } int vm_get_intinfo(struct vm *vm, int vcpuid, uint64_t *info1, uint64_t *info2) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; *info1 = vcpu->exitintinfo; *info2 = vcpu_exception_intinfo(vcpu); return (0); } int vm_inject_exception(struct vm *vm, int vcpuid, int vector, int errcode_valid, uint32_t errcode, int restart_instruction) { struct vcpu *vcpu; uint64_t regval; int error; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (vector < 0 || vector >= 32) return (EINVAL); /* * A double fault exception should never be injected directly into * the guest. It is a derived exception that results from specific * combinations of nested faults. */ if (vector == IDT_DF) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; if (vcpu->exception_pending) { VCPU_CTR2(vm, vcpuid, "Unable to inject exception %d due to " "pending exception %d", vector, vcpu->exc_vector); return (EBUSY); } if (errcode_valid) { /* * Exceptions don't deliver an error code in real mode. */ error = vm_get_register(vm, vcpuid, VM_REG_GUEST_CR0, ®val); KASSERT(!error, ("%s: error %d getting CR0", __func__, error)); if (!(regval & CR0_PE)) errcode_valid = 0; } /* * From section 26.6.1 "Interruptibility State" in Intel SDM: * * Event blocking by "STI" or "MOV SS" is cleared after guest executes * one instruction or incurs an exception. */ error = vm_set_register(vm, vcpuid, VM_REG_GUEST_INTR_SHADOW, 0); KASSERT(error == 0, ("%s: error %d clearing interrupt shadow", __func__, error)); if (restart_instruction) vm_restart_instruction(vm, vcpuid); vcpu->exception_pending = 1; vcpu->exc_vector = vector; vcpu->exc_errcode = errcode; vcpu->exc_errcode_valid = errcode_valid; VCPU_CTR1(vm, vcpuid, "Exception %d pending", vector); return (0); } void vm_inject_fault(void *vmarg, int vcpuid, int vector, int errcode_valid, int errcode) { struct vm *vm; int error, restart_instruction; vm = vmarg; restart_instruction = 1; error = vm_inject_exception(vm, vcpuid, vector, errcode_valid, errcode, restart_instruction); KASSERT(error == 0, ("vm_inject_exception error %d", error)); } void vm_inject_pf(void *vmarg, int vcpuid, int error_code, uint64_t cr2) { struct vm *vm; int error; vm = vmarg; VCPU_CTR2(vm, vcpuid, "Injecting page fault: error_code %#x, cr2 %#lx", error_code, cr2); error = vm_set_register(vm, vcpuid, VM_REG_GUEST_CR2, cr2); KASSERT(error == 0, ("vm_set_register(cr2) error %d", error)); vm_inject_fault(vm, vcpuid, IDT_PF, 1, error_code); } static VMM_STAT(VCPU_NMI_COUNT, "number of NMIs delivered to vcpu"); int vm_inject_nmi(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; vcpu->nmi_pending = 1; vcpu_notify_event(vm, vcpuid, false); return (0); } int vm_nmi_pending(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_nmi_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; return (vcpu->nmi_pending); } void vm_nmi_clear(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_nmi_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; if (vcpu->nmi_pending == 0) panic("vm_nmi_clear: inconsistent nmi_pending state"); vcpu->nmi_pending = 0; vmm_stat_incr(vm, vcpuid, VCPU_NMI_COUNT, 1); } static VMM_STAT(VCPU_EXTINT_COUNT, "number of ExtINTs delivered to vcpu"); int vm_inject_extint(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; vcpu->extint_pending = 1; vcpu_notify_event(vm, vcpuid, false); return (0); } int vm_extint_pending(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_extint_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; return (vcpu->extint_pending); } void vm_extint_clear(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_extint_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; if (vcpu->extint_pending == 0) panic("vm_extint_clear: inconsistent extint_pending state"); vcpu->extint_pending = 0; vmm_stat_incr(vm, vcpuid, VCPU_EXTINT_COUNT, 1); } int vm_get_capability(struct vm *vm, int vcpu, int type, int *retval) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (type < 0 || type >= VM_CAP_MAX) return (EINVAL); return (VMGETCAP(vm->cookie, vcpu, type, retval)); } int vm_set_capability(struct vm *vm, int vcpu, int type, int val) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (type < 0 || type >= VM_CAP_MAX) return (EINVAL); return (VMSETCAP(vm->cookie, vcpu, type, val)); } struct vlapic * vm_lapic(struct vm *vm, int cpu) { return (vm->vcpu[cpu].vlapic); } struct vioapic * vm_ioapic(struct vm *vm) { return (vm->vioapic); } struct vhpet * vm_hpet(struct vm *vm) { return (vm->vhpet); } bool vmm_is_pptdev(int bus, int slot, int func) { int b, f, i, n, s; char *val, *cp, *cp2; bool found; /* * XXX * The length of an environment variable is limited to 128 bytes which * puts an upper limit on the number of passthru devices that may be * specified using a single environment variable. * * Work around this by scanning multiple environment variable * names instead of a single one - yuck! */ const char *names[] = { "pptdevs", "pptdevs2", "pptdevs3", NULL }; /* set pptdevs="1/2/3 4/5/6 7/8/9 10/11/12" */ found = false; for (i = 0; names[i] != NULL && !found; i++) { cp = val = kern_getenv(names[i]); while (cp != NULL && *cp != '\0') { if ((cp2 = strchr(cp, ' ')) != NULL) *cp2 = '\0'; n = sscanf(cp, "%d/%d/%d", &b, &s, &f); if (n == 3 && bus == b && slot == s && func == f) { found = true; break; } if (cp2 != NULL) *cp2++ = ' '; cp = cp2; } freeenv(val); } return (found); } void * vm_iommu_domain(struct vm *vm) { return (vm->iommu); } int vcpu_set_state(struct vm *vm, int vcpuid, enum vcpu_state newstate, bool from_idle) { int error; struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_set_run_state: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); error = vcpu_set_state_locked(vm, vcpuid, newstate, from_idle); vcpu_unlock(vcpu); return (error); } enum vcpu_state vcpu_get_state(struct vm *vm, int vcpuid, int *hostcpu) { struct vcpu *vcpu; enum vcpu_state state; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_get_run_state: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); state = vcpu->state; if (hostcpu != NULL) *hostcpu = vcpu->hostcpu; vcpu_unlock(vcpu); return (state); } int vm_activate_cpu(struct vm *vm, int vcpuid) { if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (CPU_ISSET(vcpuid, &vm->active_cpus)) return (EBUSY); VCPU_CTR0(vm, vcpuid, "activated"); CPU_SET_ATOMIC(vcpuid, &vm->active_cpus); return (0); } int vm_suspend_cpu(struct vm *vm, int vcpuid) { int i; if (vcpuid < -1 || vcpuid >= vm->maxcpus) return (EINVAL); if (vcpuid == -1) { vm->debug_cpus = vm->active_cpus; for (i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &vm->active_cpus)) vcpu_notify_event(vm, i, false); } } else { if (!CPU_ISSET(vcpuid, &vm->active_cpus)) return (EINVAL); CPU_SET_ATOMIC(vcpuid, &vm->debug_cpus); vcpu_notify_event(vm, vcpuid, false); } return (0); } int vm_resume_cpu(struct vm *vm, int vcpuid) { if (vcpuid < -1 || vcpuid >= vm->maxcpus) return (EINVAL); if (vcpuid == -1) { CPU_ZERO(&vm->debug_cpus); } else { if (!CPU_ISSET(vcpuid, &vm->debug_cpus)) return (EINVAL); CPU_CLR_ATOMIC(vcpuid, &vm->debug_cpus); } return (0); } int vcpu_debugged(struct vm *vm, int vcpuid) { return (CPU_ISSET(vcpuid, &vm->debug_cpus)); } cpuset_t vm_active_cpus(struct vm *vm) { return (vm->active_cpus); } cpuset_t vm_debug_cpus(struct vm *vm) { return (vm->debug_cpus); } cpuset_t vm_suspended_cpus(struct vm *vm) { return (vm->suspended_cpus); } void * vcpu_stats(struct vm *vm, int vcpuid) { return (vm->vcpu[vcpuid].stats); } int vm_get_x2apic_state(struct vm *vm, int vcpuid, enum x2apic_state *state) { if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); *state = vm->vcpu[vcpuid].x2apic_state; return (0); } int vm_set_x2apic_state(struct vm *vm, int vcpuid, enum x2apic_state state) { if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (state >= X2APIC_STATE_LAST) return (EINVAL); vm->vcpu[vcpuid].x2apic_state = state; vlapic_set_x2apic_state(vm, vcpuid, state); return (0); } /* * This function is called to ensure that a vcpu "sees" a pending event * as soon as possible: * - If the vcpu thread is sleeping then it is woken up. * - If the vcpu is running on a different host_cpu then an IPI will be directed * to the host_cpu to cause the vcpu to trap into the hypervisor. */ static void vcpu_notify_event_locked(struct vcpu *vcpu, bool lapic_intr) { int hostcpu; hostcpu = vcpu->hostcpu; if (vcpu->state == VCPU_RUNNING) { KASSERT(hostcpu != NOCPU, ("vcpu running on invalid hostcpu")); if (hostcpu != curcpu) { if (lapic_intr) { vlapic_post_intr(vcpu->vlapic, hostcpu, vmm_ipinum); } else { ipi_cpu(hostcpu, vmm_ipinum); } } else { /* * If the 'vcpu' is running on 'curcpu' then it must * be sending a notification to itself (e.g. SELF_IPI). * The pending event will be picked up when the vcpu * transitions back to guest context. */ } } else { KASSERT(hostcpu == NOCPU, ("vcpu state %d not consistent " "with hostcpu %d", vcpu->state, hostcpu)); if (vcpu->state == VCPU_SLEEPING) wakeup_one(vcpu); } } void vcpu_notify_event(struct vm *vm, int vcpuid, bool lapic_intr) { struct vcpu *vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); vcpu_notify_event_locked(vcpu, lapic_intr); vcpu_unlock(vcpu); } struct vmspace * vm_get_vmspace(struct vm *vm) { return (vm->vmspace); } int vm_apicid2vcpuid(struct vm *vm, int apicid) { /* * XXX apic id is assumed to be numerically identical to vcpu id */ return (apicid); } int vm_smp_rendezvous(struct vm *vm, int vcpuid, cpuset_t dest, vm_rendezvous_func_t func, void *arg) { int error, i; /* * Enforce that this function is called without any locks */ WITNESS_WARN(WARN_PANIC, NULL, "vm_smp_rendezvous"); KASSERT(vcpuid == -1 || (vcpuid >= 0 && vcpuid < vm->maxcpus), ("vm_smp_rendezvous: invalid vcpuid %d", vcpuid)); restart: mtx_lock(&vm->rendezvous_mtx); if (vm->rendezvous_func != NULL) { /* * If a rendezvous is already in progress then we need to * call the rendezvous handler in case this 'vcpuid' is one * of the targets of the rendezvous. */ RENDEZVOUS_CTR0(vm, vcpuid, "Rendezvous already in progress"); mtx_unlock(&vm->rendezvous_mtx); error = vm_handle_rendezvous(vm, vcpuid); if (error != 0) return (error); goto restart; } KASSERT(vm->rendezvous_func == NULL, ("vm_smp_rendezvous: previous " "rendezvous is still in progress")); RENDEZVOUS_CTR0(vm, vcpuid, "Initiating rendezvous"); vm->rendezvous_req_cpus = dest; CPU_ZERO(&vm->rendezvous_done_cpus); vm->rendezvous_arg = arg; vm->rendezvous_func = func; mtx_unlock(&vm->rendezvous_mtx); /* * Wake up any sleeping vcpus and trigger a VM-exit in any running * vcpus so they handle the rendezvous as soon as possible. */ for (i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &dest)) vcpu_notify_event(vm, i, false); } return (vm_handle_rendezvous(vm, vcpuid)); } struct vatpic * vm_atpic(struct vm *vm) { return (vm->vatpic); } struct vatpit * vm_atpit(struct vm *vm) { return (vm->vatpit); } struct vpmtmr * vm_pmtmr(struct vm *vm) { return (vm->vpmtmr); } struct vrtc * vm_rtc(struct vm *vm) { return (vm->vrtc); } enum vm_reg_name vm_segment_name(int seg) { static enum vm_reg_name seg_names[] = { VM_REG_GUEST_ES, VM_REG_GUEST_CS, VM_REG_GUEST_SS, VM_REG_GUEST_DS, VM_REG_GUEST_FS, VM_REG_GUEST_GS }; KASSERT(seg >= 0 && seg < nitems(seg_names), ("%s: invalid segment encoding %d", __func__, seg)); return (seg_names[seg]); } void vm_copy_teardown(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, int num_copyinfo) { int idx; for (idx = 0; idx < num_copyinfo; idx++) { if (copyinfo[idx].cookie != NULL) vm_gpa_release(copyinfo[idx].cookie); } bzero(copyinfo, num_copyinfo * sizeof(struct vm_copyinfo)); } int vm_copy_setup(struct vm *vm, int vcpuid, struct vm_guest_paging *paging, uint64_t gla, size_t len, int prot, struct vm_copyinfo *copyinfo, int num_copyinfo, int *fault) { int error, idx, nused; size_t n, off, remaining; void *hva, *cookie; uint64_t gpa; bzero(copyinfo, sizeof(struct vm_copyinfo) * num_copyinfo); nused = 0; remaining = len; while (remaining > 0) { KASSERT(nused < num_copyinfo, ("insufficient vm_copyinfo")); error = vm_gla2gpa(vm, vcpuid, paging, gla, prot, &gpa, fault); if (error || *fault) return (error); off = gpa & PAGE_MASK; n = min(remaining, PAGE_SIZE - off); copyinfo[nused].gpa = gpa; copyinfo[nused].len = n; remaining -= n; gla += n; nused++; } for (idx = 0; idx < nused; idx++) { hva = vm_gpa_hold(vm, vcpuid, copyinfo[idx].gpa, copyinfo[idx].len, prot, &cookie); if (hva == NULL) break; copyinfo[idx].hva = hva; copyinfo[idx].cookie = cookie; } if (idx != nused) { vm_copy_teardown(vm, vcpuid, copyinfo, num_copyinfo); return (EFAULT); } else { *fault = 0; return (0); } } void vm_copyin(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, void *kaddr, size_t len) { char *dst; int idx; dst = kaddr; idx = 0; while (len > 0) { bcopy(copyinfo[idx].hva, dst, copyinfo[idx].len); len -= copyinfo[idx].len; dst += copyinfo[idx].len; idx++; } } void vm_copyout(struct vm *vm, int vcpuid, const void *kaddr, struct vm_copyinfo *copyinfo, size_t len) { const char *src; int idx; src = kaddr; idx = 0; while (len > 0) { bcopy(src, copyinfo[idx].hva, copyinfo[idx].len); len -= copyinfo[idx].len; src += copyinfo[idx].len; idx++; } } /* * Return the amount of in-use and wired memory for the VM. Since * these are global stats, only return the values with for vCPU 0 */ VMM_STAT_DECLARE(VMM_MEM_RESIDENT); VMM_STAT_DECLARE(VMM_MEM_WIRED); static void vm_get_rescnt(struct vm *vm, int vcpu, struct vmm_stat_type *stat) { if (vcpu == 0) { vmm_stat_set(vm, vcpu, VMM_MEM_RESIDENT, PAGE_SIZE * vmspace_resident_count(vm->vmspace)); } } static void vm_get_wiredcnt(struct vm *vm, int vcpu, struct vmm_stat_type *stat) { if (vcpu == 0) { vmm_stat_set(vm, vcpu, VMM_MEM_WIRED, PAGE_SIZE * pmap_wired_count(vmspace_pmap(vm->vmspace))); } } VMM_STAT_FUNC(VMM_MEM_RESIDENT, "Resident memory", vm_get_rescnt); VMM_STAT_FUNC(VMM_MEM_WIRED, "Wired memory", vm_get_wiredcnt); Index: head/sys/amd64/vmm/vmm_dev.c =================================================================== --- head/sys/amd64/vmm/vmm_dev.c (revision 357973) +++ head/sys/amd64/vmm/vmm_dev.c (revision 357974) @@ -1,1165 +1,1167 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "vmm_lapic.h" #include "vmm_stat.h" #include "vmm_mem.h" #include "io/ppt.h" #include "io/vatpic.h" #include "io/vioapic.h" #include "io/vhpet.h" #include "io/vrtc.h" struct devmem_softc { int segid; char *name; struct cdev *cdev; struct vmmdev_softc *sc; SLIST_ENTRY(devmem_softc) link; }; struct vmmdev_softc { struct vm *vm; /* vm instance cookie */ struct cdev *cdev; SLIST_ENTRY(vmmdev_softc) link; SLIST_HEAD(, devmem_softc) devmem; int flags; }; #define VSC_LINKED 0x01 static SLIST_HEAD(, vmmdev_softc) head; static unsigned pr_allow_flag; static struct mtx vmmdev_mtx; static MALLOC_DEFINE(M_VMMDEV, "vmmdev", "vmmdev"); SYSCTL_DECL(_hw_vmm); static int vmm_priv_check(struct ucred *ucred); static int devmem_create_cdev(const char *vmname, int id, char *devmem); static void devmem_destroy(void *arg); static int vmm_priv_check(struct ucred *ucred) { if (jailed(ucred) && !(ucred->cr_prison->pr_allow & pr_allow_flag)) return (EPERM); return (0); } static int vcpu_lock_one(struct vmmdev_softc *sc, int vcpu) { int error; if (vcpu < 0 || vcpu >= vm_get_maxcpus(sc->vm)) return (EINVAL); error = vcpu_set_state(sc->vm, vcpu, VCPU_FROZEN, true); return (error); } static void vcpu_unlock_one(struct vmmdev_softc *sc, int vcpu) { enum vcpu_state state; state = vcpu_get_state(sc->vm, vcpu, NULL); if (state != VCPU_FROZEN) { panic("vcpu %s(%d) has invalid state %d", vm_name(sc->vm), vcpu, state); } vcpu_set_state(sc->vm, vcpu, VCPU_IDLE, false); } static int vcpu_lock_all(struct vmmdev_softc *sc) { int error, vcpu; uint16_t maxcpus; maxcpus = vm_get_maxcpus(sc->vm); for (vcpu = 0; vcpu < maxcpus; vcpu++) { error = vcpu_lock_one(sc, vcpu); if (error) break; } if (error) { while (--vcpu >= 0) vcpu_unlock_one(sc, vcpu); } return (error); } static void vcpu_unlock_all(struct vmmdev_softc *sc) { int vcpu; uint16_t maxcpus; maxcpus = vm_get_maxcpus(sc->vm); for (vcpu = 0; vcpu < maxcpus; vcpu++) vcpu_unlock_one(sc, vcpu); } static struct vmmdev_softc * vmmdev_lookup(const char *name) { struct vmmdev_softc *sc; #ifdef notyet /* XXX kernel is not compiled with invariants */ mtx_assert(&vmmdev_mtx, MA_OWNED); #endif SLIST_FOREACH(sc, &head, link) { if (strcmp(name, vm_name(sc->vm)) == 0) break; } return (sc); } static struct vmmdev_softc * vmmdev_lookup2(struct cdev *cdev) { return (cdev->si_drv1); } static int vmmdev_rw(struct cdev *cdev, struct uio *uio, int flags) { int error, off, c, prot; vm_paddr_t gpa, maxaddr; void *hpa, *cookie; struct vmmdev_softc *sc; uint16_t lastcpu; error = vmm_priv_check(curthread->td_ucred); if (error) return (error); sc = vmmdev_lookup2(cdev); if (sc == NULL) return (ENXIO); /* * Get a read lock on the guest memory map by freezing any vcpu. */ lastcpu = vm_get_maxcpus(sc->vm) - 1; error = vcpu_lock_one(sc, lastcpu); if (error) return (error); prot = (uio->uio_rw == UIO_WRITE ? VM_PROT_WRITE : VM_PROT_READ); maxaddr = vmm_sysmem_maxaddr(sc->vm); while (uio->uio_resid > 0 && error == 0) { gpa = uio->uio_offset; off = gpa & PAGE_MASK; c = min(uio->uio_resid, PAGE_SIZE - off); /* * The VM has a hole in its physical memory map. If we want to * use 'dd' to inspect memory beyond the hole we need to * provide bogus data for memory that lies in the hole. * * Since this device does not support lseek(2), dd(1) will * read(2) blocks of data to simulate the lseek(2). */ hpa = vm_gpa_hold(sc->vm, lastcpu, gpa, c, prot, &cookie); if (hpa == NULL) { if (uio->uio_rw == UIO_READ && gpa < maxaddr) error = uiomove(__DECONST(void *, zero_region), c, uio); else error = EFAULT; } else { error = uiomove(hpa, c, uio); vm_gpa_release(cookie); } } vcpu_unlock_one(sc, lastcpu); return (error); } CTASSERT(sizeof(((struct vm_memseg *)0)->name) >= VM_MAX_SUFFIXLEN + 1); static int get_memseg(struct vmmdev_softc *sc, struct vm_memseg *mseg) { struct devmem_softc *dsc; int error; bool sysmem; error = vm_get_memseg(sc->vm, mseg->segid, &mseg->len, &sysmem, NULL); if (error || mseg->len == 0) return (error); if (!sysmem) { SLIST_FOREACH(dsc, &sc->devmem, link) { if (dsc->segid == mseg->segid) break; } KASSERT(dsc != NULL, ("%s: devmem segment %d not found", __func__, mseg->segid)); error = copystr(dsc->name, mseg->name, sizeof(mseg->name), NULL); } else { bzero(mseg->name, sizeof(mseg->name)); } return (error); } static int alloc_memseg(struct vmmdev_softc *sc, struct vm_memseg *mseg) { char *name; int error; bool sysmem; error = 0; name = NULL; sysmem = true; /* * The allocation is lengthened by 1 to hold a terminating NUL. It'll * by stripped off when devfs processes the full string. */ if (VM_MEMSEG_NAME(mseg)) { sysmem = false; name = malloc(sizeof(mseg->name), M_VMMDEV, M_WAITOK); error = copystr(mseg->name, name, sizeof(mseg->name), NULL); if (error) goto done; } error = vm_alloc_memseg(sc->vm, mseg->segid, mseg->len, sysmem); if (error) goto done; if (VM_MEMSEG_NAME(mseg)) { error = devmem_create_cdev(vm_name(sc->vm), mseg->segid, name); if (error) vm_free_memseg(sc->vm, mseg->segid); else name = NULL; /* freed when 'cdev' is destroyed */ } done: free(name, M_VMMDEV); return (error); } static int vm_get_register_set(struct vm *vm, int vcpu, unsigned int count, int *regnum, uint64_t *regval) { int error, i; error = 0; for (i = 0; i < count; i++) { error = vm_get_register(vm, vcpu, regnum[i], ®val[i]); if (error) break; } return (error); } static int vm_set_register_set(struct vm *vm, int vcpu, unsigned int count, int *regnum, uint64_t *regval) { int error, i; error = 0; for (i = 0; i < count; i++) { error = vm_set_register(vm, vcpu, regnum[i], regval[i]); if (error) break; } return (error); } static int vmmdev_ioctl(struct cdev *cdev, u_long cmd, caddr_t data, int fflag, struct thread *td) { int error, vcpu, state_changed, size; cpuset_t *cpuset; struct vmmdev_softc *sc; struct vm_register *vmreg; struct vm_seg_desc *vmsegdesc; struct vm_register_set *vmregset; struct vm_run *vmrun; struct vm_exception *vmexc; struct vm_lapic_irq *vmirq; struct vm_lapic_msi *vmmsi; struct vm_ioapic_irq *ioapic_irq; struct vm_isa_irq *isa_irq; struct vm_isa_irq_trigger *isa_irq_trigger; struct vm_capability *vmcap; struct vm_pptdev *pptdev; struct vm_pptdev_mmio *pptmmio; struct vm_pptdev_msi *pptmsi; struct vm_pptdev_msix *pptmsix; struct vm_nmi *vmnmi; struct vm_stats *vmstats; struct vm_stat_desc *statdesc; struct vm_x2apic *x2apic; struct vm_gpa_pte *gpapte; struct vm_suspend *vmsuspend; struct vm_gla2gpa *gg; struct vm_activate_cpu *vac; struct vm_cpuset *vm_cpuset; struct vm_intinfo *vmii; struct vm_rtc_time *rtctime; struct vm_rtc_data *rtcdata; struct vm_memmap *mm; struct vm_cpu_topology *topology; uint64_t *regvals; int *regnums; error = vmm_priv_check(curthread->td_ucred); if (error) return (error); sc = vmmdev_lookup2(cdev); if (sc == NULL) return (ENXIO); vcpu = -1; state_changed = 0; /* * Some VMM ioctls can operate only on vcpus that are not running. */ switch (cmd) { case VM_RUN: case VM_GET_REGISTER: case VM_SET_REGISTER: case VM_GET_SEGMENT_DESCRIPTOR: case VM_SET_SEGMENT_DESCRIPTOR: case VM_GET_REGISTER_SET: case VM_SET_REGISTER_SET: case VM_INJECT_EXCEPTION: case VM_GET_CAPABILITY: case VM_SET_CAPABILITY: case VM_PPTDEV_MSI: case VM_PPTDEV_MSIX: case VM_SET_X2APIC_STATE: case VM_GLA2GPA: case VM_GLA2GPA_NOFAULT: case VM_ACTIVATE_CPU: case VM_SET_INTINFO: case VM_GET_INTINFO: case VM_RESTART_INSTRUCTION: /* * XXX fragile, handle with care * Assumes that the first field of the ioctl data is the vcpu. */ vcpu = *(int *)data; error = vcpu_lock_one(sc, vcpu); if (error) goto done; state_changed = 1; break; case VM_MAP_PPTDEV_MMIO: case VM_BIND_PPTDEV: case VM_UNBIND_PPTDEV: case VM_ALLOC_MEMSEG: case VM_MMAP_MEMSEG: case VM_REINIT: /* * ioctls that operate on the entire virtual machine must * prevent all vcpus from running. */ error = vcpu_lock_all(sc); if (error) goto done; state_changed = 2; break; case VM_GET_MEMSEG: case VM_MMAP_GETNEXT: /* * Lock a vcpu to make sure that the memory map cannot be * modified while it is being inspected. */ vcpu = vm_get_maxcpus(sc->vm) - 1; error = vcpu_lock_one(sc, vcpu); if (error) goto done; state_changed = 1; break; default: break; } switch(cmd) { case VM_RUN: vmrun = (struct vm_run *)data; error = vm_run(sc->vm, vmrun); break; case VM_SUSPEND: vmsuspend = (struct vm_suspend *)data; error = vm_suspend(sc->vm, vmsuspend->how); break; case VM_REINIT: error = vm_reinit(sc->vm); break; case VM_STAT_DESC: { statdesc = (struct vm_stat_desc *)data; error = vmm_stat_desc_copy(statdesc->index, statdesc->desc, sizeof(statdesc->desc)); break; } case VM_STATS: { CTASSERT(MAX_VM_STATS >= MAX_VMM_STAT_ELEMS); vmstats = (struct vm_stats *)data; getmicrotime(&vmstats->tv); error = vmm_stat_copy(sc->vm, vmstats->cpuid, &vmstats->num_entries, vmstats->statbuf); break; } case VM_PPTDEV_MSI: pptmsi = (struct vm_pptdev_msi *)data; error = ppt_setup_msi(sc->vm, pptmsi->vcpu, pptmsi->bus, pptmsi->slot, pptmsi->func, pptmsi->addr, pptmsi->msg, pptmsi->numvec); break; case VM_PPTDEV_MSIX: pptmsix = (struct vm_pptdev_msix *)data; error = ppt_setup_msix(sc->vm, pptmsix->vcpu, pptmsix->bus, pptmsix->slot, pptmsix->func, pptmsix->idx, pptmsix->addr, pptmsix->msg, pptmsix->vector_control); break; case VM_MAP_PPTDEV_MMIO: pptmmio = (struct vm_pptdev_mmio *)data; error = ppt_map_mmio(sc->vm, pptmmio->bus, pptmmio->slot, pptmmio->func, pptmmio->gpa, pptmmio->len, pptmmio->hpa); break; case VM_BIND_PPTDEV: pptdev = (struct vm_pptdev *)data; error = vm_assign_pptdev(sc->vm, pptdev->bus, pptdev->slot, pptdev->func); break; case VM_UNBIND_PPTDEV: pptdev = (struct vm_pptdev *)data; error = vm_unassign_pptdev(sc->vm, pptdev->bus, pptdev->slot, pptdev->func); break; case VM_INJECT_EXCEPTION: vmexc = (struct vm_exception *)data; error = vm_inject_exception(sc->vm, vmexc->cpuid, vmexc->vector, vmexc->error_code_valid, vmexc->error_code, vmexc->restart_instruction); break; case VM_INJECT_NMI: vmnmi = (struct vm_nmi *)data; error = vm_inject_nmi(sc->vm, vmnmi->cpuid); break; case VM_LAPIC_IRQ: vmirq = (struct vm_lapic_irq *)data; error = lapic_intr_edge(sc->vm, vmirq->cpuid, vmirq->vector); break; case VM_LAPIC_LOCAL_IRQ: vmirq = (struct vm_lapic_irq *)data; error = lapic_set_local_intr(sc->vm, vmirq->cpuid, vmirq->vector); break; case VM_LAPIC_MSI: vmmsi = (struct vm_lapic_msi *)data; error = lapic_intr_msi(sc->vm, vmmsi->addr, vmmsi->msg); break; case VM_IOAPIC_ASSERT_IRQ: ioapic_irq = (struct vm_ioapic_irq *)data; error = vioapic_assert_irq(sc->vm, ioapic_irq->irq); break; case VM_IOAPIC_DEASSERT_IRQ: ioapic_irq = (struct vm_ioapic_irq *)data; error = vioapic_deassert_irq(sc->vm, ioapic_irq->irq); break; case VM_IOAPIC_PULSE_IRQ: ioapic_irq = (struct vm_ioapic_irq *)data; error = vioapic_pulse_irq(sc->vm, ioapic_irq->irq); break; case VM_IOAPIC_PINCOUNT: *(int *)data = vioapic_pincount(sc->vm); break; case VM_ISA_ASSERT_IRQ: isa_irq = (struct vm_isa_irq *)data; error = vatpic_assert_irq(sc->vm, isa_irq->atpic_irq); if (error == 0 && isa_irq->ioapic_irq != -1) error = vioapic_assert_irq(sc->vm, isa_irq->ioapic_irq); break; case VM_ISA_DEASSERT_IRQ: isa_irq = (struct vm_isa_irq *)data; error = vatpic_deassert_irq(sc->vm, isa_irq->atpic_irq); if (error == 0 && isa_irq->ioapic_irq != -1) error = vioapic_deassert_irq(sc->vm, isa_irq->ioapic_irq); break; case VM_ISA_PULSE_IRQ: isa_irq = (struct vm_isa_irq *)data; error = vatpic_pulse_irq(sc->vm, isa_irq->atpic_irq); if (error == 0 && isa_irq->ioapic_irq != -1) error = vioapic_pulse_irq(sc->vm, isa_irq->ioapic_irq); break; case VM_ISA_SET_IRQ_TRIGGER: isa_irq_trigger = (struct vm_isa_irq_trigger *)data; error = vatpic_set_irq_trigger(sc->vm, isa_irq_trigger->atpic_irq, isa_irq_trigger->trigger); break; case VM_MMAP_GETNEXT: mm = (struct vm_memmap *)data; error = vm_mmap_getnext(sc->vm, &mm->gpa, &mm->segid, &mm->segoff, &mm->len, &mm->prot, &mm->flags); break; case VM_MMAP_MEMSEG: mm = (struct vm_memmap *)data; error = vm_mmap_memseg(sc->vm, mm->gpa, mm->segid, mm->segoff, mm->len, mm->prot, mm->flags); break; case VM_ALLOC_MEMSEG: error = alloc_memseg(sc, (struct vm_memseg *)data); break; case VM_GET_MEMSEG: error = get_memseg(sc, (struct vm_memseg *)data); break; case VM_GET_REGISTER: vmreg = (struct vm_register *)data; error = vm_get_register(sc->vm, vmreg->cpuid, vmreg->regnum, &vmreg->regval); break; case VM_SET_REGISTER: vmreg = (struct vm_register *)data; error = vm_set_register(sc->vm, vmreg->cpuid, vmreg->regnum, vmreg->regval); break; case VM_SET_SEGMENT_DESCRIPTOR: vmsegdesc = (struct vm_seg_desc *)data; error = vm_set_seg_desc(sc->vm, vmsegdesc->cpuid, vmsegdesc->regnum, &vmsegdesc->desc); break; case VM_GET_SEGMENT_DESCRIPTOR: vmsegdesc = (struct vm_seg_desc *)data; error = vm_get_seg_desc(sc->vm, vmsegdesc->cpuid, vmsegdesc->regnum, &vmsegdesc->desc); break; case VM_GET_REGISTER_SET: vmregset = (struct vm_register_set *)data; if (vmregset->count > VM_REG_LAST) { error = EINVAL; break; } regvals = malloc(sizeof(regvals[0]) * vmregset->count, M_VMMDEV, M_WAITOK); regnums = malloc(sizeof(regnums[0]) * vmregset->count, M_VMMDEV, M_WAITOK); error = copyin(vmregset->regnums, regnums, sizeof(regnums[0]) * vmregset->count); if (error == 0) error = vm_get_register_set(sc->vm, vmregset->cpuid, vmregset->count, regnums, regvals); if (error == 0) error = copyout(regvals, vmregset->regvals, sizeof(regvals[0]) * vmregset->count); free(regvals, M_VMMDEV); free(regnums, M_VMMDEV); break; case VM_SET_REGISTER_SET: vmregset = (struct vm_register_set *)data; if (vmregset->count > VM_REG_LAST) { error = EINVAL; break; } regvals = malloc(sizeof(regvals[0]) * vmregset->count, M_VMMDEV, M_WAITOK); regnums = malloc(sizeof(regnums[0]) * vmregset->count, M_VMMDEV, M_WAITOK); error = copyin(vmregset->regnums, regnums, sizeof(regnums[0]) * vmregset->count); if (error == 0) error = copyin(vmregset->regvals, regvals, sizeof(regvals[0]) * vmregset->count); if (error == 0) error = vm_set_register_set(sc->vm, vmregset->cpuid, vmregset->count, regnums, regvals); free(regvals, M_VMMDEV); free(regnums, M_VMMDEV); break; case VM_GET_CAPABILITY: vmcap = (struct vm_capability *)data; error = vm_get_capability(sc->vm, vmcap->cpuid, vmcap->captype, &vmcap->capval); break; case VM_SET_CAPABILITY: vmcap = (struct vm_capability *)data; error = vm_set_capability(sc->vm, vmcap->cpuid, vmcap->captype, vmcap->capval); break; case VM_SET_X2APIC_STATE: x2apic = (struct vm_x2apic *)data; error = vm_set_x2apic_state(sc->vm, x2apic->cpuid, x2apic->state); break; case VM_GET_X2APIC_STATE: x2apic = (struct vm_x2apic *)data; error = vm_get_x2apic_state(sc->vm, x2apic->cpuid, &x2apic->state); break; case VM_GET_GPA_PMAP: gpapte = (struct vm_gpa_pte *)data; pmap_get_mapping(vmspace_pmap(vm_get_vmspace(sc->vm)), gpapte->gpa, gpapte->pte, &gpapte->ptenum); error = 0; break; case VM_GET_HPET_CAPABILITIES: error = vhpet_getcap((struct vm_hpet_cap *)data); break; case VM_GLA2GPA: { CTASSERT(PROT_READ == VM_PROT_READ); CTASSERT(PROT_WRITE == VM_PROT_WRITE); CTASSERT(PROT_EXEC == VM_PROT_EXECUTE); gg = (struct vm_gla2gpa *)data; error = vm_gla2gpa(sc->vm, gg->vcpuid, &gg->paging, gg->gla, gg->prot, &gg->gpa, &gg->fault); KASSERT(error == 0 || error == EFAULT, ("%s: vm_gla2gpa unknown error %d", __func__, error)); break; } case VM_GLA2GPA_NOFAULT: gg = (struct vm_gla2gpa *)data; error = vm_gla2gpa_nofault(sc->vm, gg->vcpuid, &gg->paging, gg->gla, gg->prot, &gg->gpa, &gg->fault); KASSERT(error == 0 || error == EFAULT, ("%s: vm_gla2gpa unknown error %d", __func__, error)); break; case VM_ACTIVATE_CPU: vac = (struct vm_activate_cpu *)data; error = vm_activate_cpu(sc->vm, vac->vcpuid); break; case VM_GET_CPUS: error = 0; vm_cpuset = (struct vm_cpuset *)data; size = vm_cpuset->cpusetsize; if (size < sizeof(cpuset_t) || size > CPU_MAXSIZE / NBBY) { error = ERANGE; break; } cpuset = malloc(size, M_TEMP, M_WAITOK | M_ZERO); if (vm_cpuset->which == VM_ACTIVE_CPUS) *cpuset = vm_active_cpus(sc->vm); else if (vm_cpuset->which == VM_SUSPENDED_CPUS) *cpuset = vm_suspended_cpus(sc->vm); else if (vm_cpuset->which == VM_DEBUG_CPUS) *cpuset = vm_debug_cpus(sc->vm); else error = EINVAL; if (error == 0) error = copyout(cpuset, vm_cpuset->cpus, size); free(cpuset, M_TEMP); break; case VM_SUSPEND_CPU: vac = (struct vm_activate_cpu *)data; error = vm_suspend_cpu(sc->vm, vac->vcpuid); break; case VM_RESUME_CPU: vac = (struct vm_activate_cpu *)data; error = vm_resume_cpu(sc->vm, vac->vcpuid); break; case VM_SET_INTINFO: vmii = (struct vm_intinfo *)data; error = vm_exit_intinfo(sc->vm, vmii->vcpuid, vmii->info1); break; case VM_GET_INTINFO: vmii = (struct vm_intinfo *)data; error = vm_get_intinfo(sc->vm, vmii->vcpuid, &vmii->info1, &vmii->info2); break; case VM_RTC_WRITE: rtcdata = (struct vm_rtc_data *)data; error = vrtc_nvram_write(sc->vm, rtcdata->offset, rtcdata->value); break; case VM_RTC_READ: rtcdata = (struct vm_rtc_data *)data; error = vrtc_nvram_read(sc->vm, rtcdata->offset, &rtcdata->value); break; case VM_RTC_SETTIME: rtctime = (struct vm_rtc_time *)data; error = vrtc_set_time(sc->vm, rtctime->secs); break; case VM_RTC_GETTIME: error = 0; rtctime = (struct vm_rtc_time *)data; rtctime->secs = vrtc_get_time(sc->vm); break; case VM_RESTART_INSTRUCTION: error = vm_restart_instruction(sc->vm, vcpu); break; case VM_SET_TOPOLOGY: topology = (struct vm_cpu_topology *)data; error = vm_set_topology(sc->vm, topology->sockets, topology->cores, topology->threads, topology->maxcpus); break; case VM_GET_TOPOLOGY: topology = (struct vm_cpu_topology *)data; vm_get_topology(sc->vm, &topology->sockets, &topology->cores, &topology->threads, &topology->maxcpus); error = 0; break; default: error = ENOTTY; break; } if (state_changed == 1) vcpu_unlock_one(sc, vcpu); else if (state_changed == 2) vcpu_unlock_all(sc); done: /* * Make sure that no handler returns a kernel-internal * error value to userspace. */ KASSERT(error == ERESTART || error >= 0, ("vmmdev_ioctl: invalid error return %d", error)); return (error); } static int vmmdev_mmap_single(struct cdev *cdev, vm_ooffset_t *offset, vm_size_t mapsize, struct vm_object **objp, int nprot) { struct vmmdev_softc *sc; vm_paddr_t gpa; size_t len; vm_ooffset_t segoff, first, last; int error, found, segid; uint16_t lastcpu; bool sysmem; error = vmm_priv_check(curthread->td_ucred); if (error) return (error); first = *offset; last = first + mapsize; if ((nprot & PROT_EXEC) || first < 0 || first >= last) return (EINVAL); sc = vmmdev_lookup2(cdev); if (sc == NULL) { /* virtual machine is in the process of being created */ return (EINVAL); } /* * Get a read lock on the guest memory map by freezing any vcpu. */ lastcpu = vm_get_maxcpus(sc->vm) - 1; error = vcpu_lock_one(sc, lastcpu); if (error) return (error); gpa = 0; found = 0; while (!found) { error = vm_mmap_getnext(sc->vm, &gpa, &segid, &segoff, &len, NULL, NULL); if (error) break; if (first >= gpa && last <= gpa + len) found = 1; else gpa += len; } if (found) { error = vm_get_memseg(sc->vm, segid, &len, &sysmem, objp); KASSERT(error == 0 && *objp != NULL, ("%s: invalid memory segment %d", __func__, segid)); if (sysmem) { vm_object_reference(*objp); *offset = segoff + (first - gpa); } else { error = EINVAL; } } vcpu_unlock_one(sc, lastcpu); return (error); } static void vmmdev_destroy(void *arg) { struct vmmdev_softc *sc = arg; struct devmem_softc *dsc; int error; error = vcpu_lock_all(sc); KASSERT(error == 0, ("%s: error %d freezing vcpus", __func__, error)); while ((dsc = SLIST_FIRST(&sc->devmem)) != NULL) { KASSERT(dsc->cdev == NULL, ("%s: devmem not free", __func__)); SLIST_REMOVE_HEAD(&sc->devmem, link); free(dsc->name, M_VMMDEV); free(dsc, M_VMMDEV); } if (sc->cdev != NULL) destroy_dev(sc->cdev); if (sc->vm != NULL) vm_destroy(sc->vm); if ((sc->flags & VSC_LINKED) != 0) { mtx_lock(&vmmdev_mtx); SLIST_REMOVE(&head, sc, vmmdev_softc, link); mtx_unlock(&vmmdev_mtx); } free(sc, M_VMMDEV); } static int sysctl_vmm_destroy(SYSCTL_HANDLER_ARGS) { struct devmem_softc *dsc; struct vmmdev_softc *sc; struct cdev *cdev; char *buf; int error, buflen; error = vmm_priv_check(req->td->td_ucred); if (error) return (error); buflen = VM_MAX_NAMELEN + 1; buf = malloc(buflen, M_VMMDEV, M_WAITOK | M_ZERO); strlcpy(buf, "beavis", buflen); error = sysctl_handle_string(oidp, buf, buflen, req); if (error != 0 || req->newptr == NULL) goto out; mtx_lock(&vmmdev_mtx); sc = vmmdev_lookup(buf); if (sc == NULL || sc->cdev == NULL) { mtx_unlock(&vmmdev_mtx); error = EINVAL; goto out; } /* * The 'cdev' will be destroyed asynchronously when 'si_threadcount' * goes down to 0 so we should not do it again in the callback. * * Setting 'sc->cdev' to NULL is also used to indicate that the VM * is scheduled for destruction. */ cdev = sc->cdev; sc->cdev = NULL; mtx_unlock(&vmmdev_mtx); /* * Schedule all cdevs to be destroyed: * * - any new operations on the 'cdev' will return an error (ENXIO). * * - when the 'si_threadcount' dwindles down to zero the 'cdev' will * be destroyed and the callback will be invoked in a taskqueue * context. * * - the 'devmem' cdevs are destroyed before the virtual machine 'cdev' */ SLIST_FOREACH(dsc, &sc->devmem, link) { KASSERT(dsc->cdev != NULL, ("devmem cdev already destroyed")); destroy_dev_sched_cb(dsc->cdev, devmem_destroy, dsc); } destroy_dev_sched_cb(cdev, vmmdev_destroy, sc); error = 0; out: free(buf, M_VMMDEV); return (error); } SYSCTL_PROC(_hw_vmm, OID_AUTO, destroy, - CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_PRISON, - NULL, 0, sysctl_vmm_destroy, "A", NULL); + CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_PRISON | CTLFLAG_MPSAFE, + NULL, 0, sysctl_vmm_destroy, "A", + NULL); static struct cdevsw vmmdevsw = { .d_name = "vmmdev", .d_version = D_VERSION, .d_ioctl = vmmdev_ioctl, .d_mmap_single = vmmdev_mmap_single, .d_read = vmmdev_rw, .d_write = vmmdev_rw, }; static int sysctl_vmm_create(SYSCTL_HANDLER_ARGS) { struct vm *vm; struct cdev *cdev; struct vmmdev_softc *sc, *sc2; char *buf; int error, buflen; error = vmm_priv_check(req->td->td_ucred); if (error) return (error); buflen = VM_MAX_NAMELEN + 1; buf = malloc(buflen, M_VMMDEV, M_WAITOK | M_ZERO); strlcpy(buf, "beavis", buflen); error = sysctl_handle_string(oidp, buf, buflen, req); if (error != 0 || req->newptr == NULL) goto out; mtx_lock(&vmmdev_mtx); sc = vmmdev_lookup(buf); mtx_unlock(&vmmdev_mtx); if (sc != NULL) { error = EEXIST; goto out; } error = vm_create(buf, &vm); if (error != 0) goto out; sc = malloc(sizeof(struct vmmdev_softc), M_VMMDEV, M_WAITOK | M_ZERO); sc->vm = vm; SLIST_INIT(&sc->devmem); /* * Lookup the name again just in case somebody sneaked in when we * dropped the lock. */ mtx_lock(&vmmdev_mtx); sc2 = vmmdev_lookup(buf); if (sc2 == NULL) { SLIST_INSERT_HEAD(&head, sc, link); sc->flags |= VSC_LINKED; } mtx_unlock(&vmmdev_mtx); if (sc2 != NULL) { vmmdev_destroy(sc); error = EEXIST; goto out; } error = make_dev_p(MAKEDEV_CHECKNAME, &cdev, &vmmdevsw, NULL, UID_ROOT, GID_WHEEL, 0600, "vmm/%s", buf); if (error != 0) { vmmdev_destroy(sc); goto out; } mtx_lock(&vmmdev_mtx); sc->cdev = cdev; sc->cdev->si_drv1 = sc; mtx_unlock(&vmmdev_mtx); out: free(buf, M_VMMDEV); return (error); } SYSCTL_PROC(_hw_vmm, OID_AUTO, create, - CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_PRISON, - NULL, 0, sysctl_vmm_create, "A", NULL); + CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_PRISON | CTLFLAG_MPSAFE, + NULL, 0, sysctl_vmm_create, "A", + NULL); void vmmdev_init(void) { mtx_init(&vmmdev_mtx, "vmm device mutex", NULL, MTX_DEF); pr_allow_flag = prison_add_allow(NULL, "vmm", NULL, "Allow use of vmm in a jail."); } int vmmdev_cleanup(void) { int error; if (SLIST_EMPTY(&head)) error = 0; else error = EBUSY; return (error); } static int devmem_mmap_single(struct cdev *cdev, vm_ooffset_t *offset, vm_size_t len, struct vm_object **objp, int nprot) { struct devmem_softc *dsc; vm_ooffset_t first, last; size_t seglen; int error; uint16_t lastcpu; bool sysmem; dsc = cdev->si_drv1; if (dsc == NULL) { /* 'cdev' has been created but is not ready for use */ return (ENXIO); } first = *offset; last = *offset + len; if ((nprot & PROT_EXEC) || first < 0 || first >= last) return (EINVAL); lastcpu = vm_get_maxcpus(dsc->sc->vm) - 1; error = vcpu_lock_one(dsc->sc, lastcpu); if (error) return (error); error = vm_get_memseg(dsc->sc->vm, dsc->segid, &seglen, &sysmem, objp); KASSERT(error == 0 && !sysmem && *objp != NULL, ("%s: invalid devmem segment %d", __func__, dsc->segid)); vcpu_unlock_one(dsc->sc, lastcpu); if (seglen >= last) { vm_object_reference(*objp); return (0); } else { return (EINVAL); } } static struct cdevsw devmemsw = { .d_name = "devmem", .d_version = D_VERSION, .d_mmap_single = devmem_mmap_single, }; static int devmem_create_cdev(const char *vmname, int segid, char *devname) { struct devmem_softc *dsc; struct vmmdev_softc *sc; struct cdev *cdev; int error; error = make_dev_p(MAKEDEV_CHECKNAME, &cdev, &devmemsw, NULL, UID_ROOT, GID_WHEEL, 0600, "vmm.io/%s.%s", vmname, devname); if (error) return (error); dsc = malloc(sizeof(struct devmem_softc), M_VMMDEV, M_WAITOK | M_ZERO); mtx_lock(&vmmdev_mtx); sc = vmmdev_lookup(vmname); KASSERT(sc != NULL, ("%s: vm %s softc not found", __func__, vmname)); if (sc->cdev == NULL) { /* virtual machine is being created or destroyed */ mtx_unlock(&vmmdev_mtx); free(dsc, M_VMMDEV); destroy_dev_sched_cb(cdev, NULL, 0); return (ENODEV); } dsc->segid = segid; dsc->name = devname; dsc->cdev = cdev; dsc->sc = sc; SLIST_INSERT_HEAD(&sc->devmem, dsc, link); mtx_unlock(&vmmdev_mtx); /* The 'cdev' is ready for use after 'si_drv1' is initialized */ cdev->si_drv1 = dsc; return (0); } static void devmem_destroy(void *arg) { struct devmem_softc *dsc = arg; KASSERT(dsc->cdev, ("%s: devmem cdev already destroyed", __func__)); dsc->cdev = NULL; dsc->sc = NULL; } Index: head/sys/amd64/vmm/x86.c =================================================================== --- head/sys/amd64/vmm/x86.c (revision 357973) +++ head/sys/amd64/vmm/x86.c (revision 357974) @@ -1,618 +1,619 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include "vmm_host.h" #include "vmm_ktr.h" #include "vmm_util.h" #include "x86.h" SYSCTL_DECL(_hw_vmm); -static SYSCTL_NODE(_hw_vmm, OID_AUTO, topology, CTLFLAG_RD, 0, NULL); +static SYSCTL_NODE(_hw_vmm, OID_AUTO, topology, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, + NULL); #define CPUID_VM_HIGH 0x40000000 static const char bhyve_id[12] = "bhyve bhyve "; static uint64_t bhyve_xcpuids; SYSCTL_ULONG(_hw_vmm, OID_AUTO, bhyve_xcpuids, CTLFLAG_RW, &bhyve_xcpuids, 0, "Number of times an unknown cpuid leaf was accessed"); #if __FreeBSD_version < 1200060 /* Remove after 11 EOL helps MFCing */ extern u_int threads_per_core; SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, threads_per_core, CTLFLAG_RDTUN, &threads_per_core, 0, NULL); extern u_int cores_per_package; SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, cores_per_package, CTLFLAG_RDTUN, &cores_per_package, 0, NULL); #endif static int cpuid_leaf_b = 1; SYSCTL_INT(_hw_vmm_topology, OID_AUTO, cpuid_leaf_b, CTLFLAG_RDTUN, &cpuid_leaf_b, 0, NULL); /* * Round up to the next power of two, if necessary, and then take log2. * Returns -1 if argument is zero. */ static __inline int log2(u_int x) { return (fls(x << (1 - powerof2(x))) - 1); } int x86_emulate_cpuid(struct vm *vm, int vcpu_id, uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx) { const struct xsave_limits *limits; uint64_t cr4; int error, enable_invpcid, level, width, x2apic_id; unsigned int func, regs[4], logical_cpus; enum x2apic_state x2apic_state; uint16_t cores, maxcpus, sockets, threads; VCPU_CTR2(vm, vcpu_id, "cpuid %#x,%#x", *eax, *ecx); /* * Requests for invalid CPUID levels should map to the highest * available level instead. */ if (cpu_exthigh != 0 && *eax >= 0x80000000) { if (*eax > cpu_exthigh) *eax = cpu_exthigh; } else if (*eax >= 0x40000000) { if (*eax > CPUID_VM_HIGH) *eax = CPUID_VM_HIGH; } else if (*eax > cpu_high) { *eax = cpu_high; } func = *eax; /* * In general the approach used for CPU topology is to * advertise a flat topology where all CPUs are packages with * no multi-core or SMT. */ switch (func) { /* * Pass these through to the guest */ case CPUID_0000_0000: case CPUID_0000_0002: case CPUID_0000_0003: case CPUID_8000_0000: case CPUID_8000_0002: case CPUID_8000_0003: case CPUID_8000_0004: case CPUID_8000_0006: cpuid_count(*eax, *ecx, regs); break; case CPUID_8000_0008: cpuid_count(*eax, *ecx, regs); if (vmm_is_svm()) { /* * As on Intel (0000_0007:0, EDX), mask out * unsupported or unsafe AMD extended features * (8000_0008 EBX). */ regs[1] &= (AMDFEID_CLZERO | AMDFEID_IRPERF | AMDFEID_XSAVEERPTR); vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); /* * Here, width is ApicIdCoreIdSize, present on * at least Family 15h and newer. It * represents the "number of bits in the * initial apicid that indicate thread id * within a package." * * Our topo_probe_amd() uses it for * pkg_id_shift and other OSes may rely on it. */ width = MIN(0xF, log2(threads * cores)); if (width < 0x4) width = 0; logical_cpus = MIN(0xFF, threads * cores - 1); regs[2] = (width << AMDID_COREID_SIZE_SHIFT) | logical_cpus; } break; case CPUID_8000_0001: cpuid_count(*eax, *ecx, regs); /* * Hide SVM from guest. */ regs[2] &= ~AMDID2_SVM; /* * Don't advertise extended performance counter MSRs * to the guest. */ regs[2] &= ~AMDID2_PCXC; regs[2] &= ~AMDID2_PNXC; regs[2] &= ~AMDID2_PTSCEL2I; /* * Don't advertise Instruction Based Sampling feature. */ regs[2] &= ~AMDID2_IBS; /* NodeID MSR not available */ regs[2] &= ~AMDID2_NODE_ID; /* Don't advertise the OS visible workaround feature */ regs[2] &= ~AMDID2_OSVW; /* Hide mwaitx/monitorx capability from the guest */ regs[2] &= ~AMDID2_MWAITX; /* * Hide rdtscp/ia32_tsc_aux until we know how * to deal with them. */ regs[3] &= ~AMDID_RDTSCP; break; case CPUID_8000_0007: /* * AMD uses this leaf to advertise the processor's * power monitoring and RAS capabilities. These * features are hardware-specific and exposing * them to a guest doesn't make a lot of sense. * * Intel uses this leaf only to advertise the * "Invariant TSC" feature with all other bits * being reserved (set to zero). */ regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; /* * "Invariant TSC" can be advertised to the guest if: * - host TSC frequency is invariant * - host TSCs are synchronized across physical cpus * * XXX This still falls short because the vcpu * can observe the TSC moving backwards as it * migrates across physical cpus. But at least * it should discourage the guest from using the * TSC to keep track of time. */ if (tsc_is_invariant && smp_tsc) regs[3] |= AMDPM_TSC_INVARIANT; break; case CPUID_8000_001D: /* AMD Cache topology, like 0000_0004 for Intel. */ if (!vmm_is_svm()) goto default_leaf; /* * Similar to Intel, generate a ficticious cache * topology for the guest with L3 shared by the * package, and L1 and L2 local to a core. */ vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); switch (*ecx) { case 0: logical_cpus = threads; level = 1; func = 1; /* data cache */ break; case 1: logical_cpus = threads; level = 2; func = 3; /* unified cache */ break; case 2: logical_cpus = threads * cores; level = 3; func = 3; /* unified cache */ break; default: logical_cpus = 0; level = 0; func = 0; break; } logical_cpus = MIN(0xfff, logical_cpus - 1); regs[0] = (logical_cpus << 14) | (1 << 8) | (level << 5) | func; regs[1] = (func > 0) ? (CACHE_LINE_SIZE - 1) : 0; regs[2] = 0; regs[3] = 0; break; case CPUID_8000_001E: /* * AMD Family 16h+ and Hygon Family 18h additional * identifiers. */ if (!vmm_is_svm() || CPUID_TO_FAMILY(cpu_id) < 0x16) goto default_leaf; vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); regs[0] = vcpu_id; threads = MIN(0xFF, threads - 1); regs[1] = (threads << 8) | (vcpu_id >> log2(threads + 1)); /* * XXX Bhyve topology cannot yet represent >1 node per * processor. */ regs[2] = 0; regs[3] = 0; break; case CPUID_0000_0001: do_cpuid(1, regs); error = vm_get_x2apic_state(vm, vcpu_id, &x2apic_state); if (error) { panic("x86_emulate_cpuid: error %d " "fetching x2apic state", error); } /* * Override the APIC ID only in ebx */ regs[1] &= ~(CPUID_LOCAL_APIC_ID); regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT); /* * Don't expose VMX, SpeedStep, TME or SMX capability. * Advertise x2APIC capability and Hypervisor guest. */ regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2); regs[2] &= ~(CPUID2_SMX); regs[2] |= CPUID2_HV; if (x2apic_state != X2APIC_DISABLED) regs[2] |= CPUID2_X2APIC; else regs[2] &= ~CPUID2_X2APIC; /* * Only advertise CPUID2_XSAVE in the guest if * the host is using XSAVE. */ if (!(regs[2] & CPUID2_OSXSAVE)) regs[2] &= ~CPUID2_XSAVE; /* * If CPUID2_XSAVE is being advertised and the * guest has set CR4_XSAVE, set * CPUID2_OSXSAVE. */ regs[2] &= ~CPUID2_OSXSAVE; if (regs[2] & CPUID2_XSAVE) { error = vm_get_register(vm, vcpu_id, VM_REG_GUEST_CR4, &cr4); if (error) panic("x86_emulate_cpuid: error %d " "fetching %%cr4", error); if (cr4 & CR4_XSAVE) regs[2] |= CPUID2_OSXSAVE; } /* * Hide monitor/mwait until we know how to deal with * these instructions. */ regs[2] &= ~CPUID2_MON; /* * Hide the performance and debug features. */ regs[2] &= ~CPUID2_PDCM; /* * No TSC deadline support in the APIC yet */ regs[2] &= ~CPUID2_TSCDLT; /* * Hide thermal monitoring */ regs[3] &= ~(CPUID_ACPI | CPUID_TM); /* * Hide the debug store capability. */ regs[3] &= ~CPUID_DS; /* * Advertise the Machine Check and MTRR capability. * * Some guest OSes (e.g. Windows) will not boot if * these features are absent. */ regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR); vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); logical_cpus = threads * cores; regs[1] &= ~CPUID_HTT_CORES; regs[1] |= (logical_cpus & 0xff) << 16; regs[3] |= CPUID_HTT; break; case CPUID_0000_0004: cpuid_count(*eax, *ecx, regs); if (regs[0] || regs[1] || regs[2] || regs[3]) { vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); regs[0] &= 0x3ff; regs[0] |= (cores - 1) << 26; /* * Cache topology: * - L1 and L2 are shared only by the logical * processors in a single core. * - L3 and above are shared by all logical * processors in the package. */ logical_cpus = threads; level = (regs[0] >> 5) & 0x7; if (level >= 3) logical_cpus *= cores; regs[0] |= (logical_cpus - 1) << 14; } break; case CPUID_0000_0007: regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; /* leaf 0 */ if (*ecx == 0) { cpuid_count(*eax, *ecx, regs); /* Only leaf 0 is supported */ regs[0] = 0; /* * Expose known-safe features. */ regs[1] &= (CPUID_STDEXT_FSGSBASE | CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE | CPUID_STDEXT_AVX2 | CPUID_STDEXT_BMI2 | CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM | CPUID_STDEXT_AVX512F | CPUID_STDEXT_RDSEED | CPUID_STDEXT_AVX512PF | CPUID_STDEXT_AVX512ER | CPUID_STDEXT_AVX512CD | CPUID_STDEXT_SHA); regs[2] = 0; regs[3] &= CPUID_STDEXT3_MD_CLEAR; /* Advertise INVPCID if it is enabled. */ error = vm_get_capability(vm, vcpu_id, VM_CAP_ENABLE_INVPCID, &enable_invpcid); if (error == 0 && enable_invpcid) regs[1] |= CPUID_STDEXT_INVPCID; } break; case CPUID_0000_0006: regs[0] = CPUTPM1_ARAT; regs[1] = 0; regs[2] = 0; regs[3] = 0; break; case CPUID_0000_000A: /* * Handle the access, but report 0 for * all options */ regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; break; case CPUID_0000_000B: /* * Intel processor topology enumeration */ if (vmm_is_intel()) { vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); if (*ecx == 0) { logical_cpus = threads; width = log2(logical_cpus); level = CPUID_TYPE_SMT; x2apic_id = vcpu_id; } if (*ecx == 1) { logical_cpus = threads * cores; width = log2(logical_cpus); level = CPUID_TYPE_CORE; x2apic_id = vcpu_id; } if (!cpuid_leaf_b || *ecx >= 2) { width = 0; logical_cpus = 0; level = 0; x2apic_id = 0; } regs[0] = width & 0x1f; regs[1] = logical_cpus & 0xffff; regs[2] = (level << 8) | (*ecx & 0xff); regs[3] = x2apic_id; } else { regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; } break; case CPUID_0000_000D: limits = vmm_get_xsave_limits(); if (!limits->xsave_enabled) { regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; break; } cpuid_count(*eax, *ecx, regs); switch (*ecx) { case 0: /* * Only permit the guest to use bits * that are active in the host in * %xcr0. Also, claim that the * maximum save area size is * equivalent to the host's current * save area size. Since this runs * "inside" of vmrun(), it runs with * the guest's xcr0, so the current * save area size is correct as-is. */ regs[0] &= limits->xcr0_allowed; regs[2] = limits->xsave_max_size; regs[3] &= (limits->xcr0_allowed >> 32); break; case 1: /* Only permit XSAVEOPT. */ regs[0] &= CPUID_EXTSTATE_XSAVEOPT; regs[1] = 0; regs[2] = 0; regs[3] = 0; break; default: /* * If the leaf is for a permitted feature, * pass through as-is, otherwise return * all zeroes. */ if (!(limits->xcr0_allowed & (1ul << *ecx))) { regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; } break; } break; case 0x40000000: regs[0] = CPUID_VM_HIGH; bcopy(bhyve_id, ®s[1], 4); bcopy(bhyve_id + 4, ®s[2], 4); bcopy(bhyve_id + 8, ®s[3], 4); break; default: default_leaf: /* * The leaf value has already been clamped so * simply pass this through, keeping count of * how many unhandled leaf values have been seen. */ atomic_add_long(&bhyve_xcpuids, 1); cpuid_count(*eax, *ecx, regs); break; } *eax = regs[0]; *ebx = regs[1]; *ecx = regs[2]; *edx = regs[3]; return (1); } bool vm_cpuid_capability(struct vm *vm, int vcpuid, enum vm_cpuid_capability cap) { bool rv; KASSERT(cap > 0 && cap < VCC_LAST, ("%s: invalid vm_cpu_capability %d", __func__, cap)); /* * Simply passthrough the capabilities of the host cpu for now. */ rv = false; switch (cap) { case VCC_NO_EXECUTE: if (amd_feature & AMDID_NX) rv = true; break; case VCC_FFXSR: if (amd_feature & AMDID_FFXSR) rv = true; break; case VCC_TCE: if (amd_feature2 & AMDID2_TCE) rv = true; break; default: panic("%s: unknown vm_cpu_capability %d", __func__, cap); } return (rv); }