Index: stable/9/sys/dev/cpuctl/cpuctl.c =================================================================== --- stable/9/sys/dev/cpuctl/cpuctl.c (revision 263357) +++ stable/9/sys/dev/cpuctl/cpuctl.c (revision 263358) @@ -1,541 +1,536 @@ /*- * Copyright (c) 2006-2008 Stanislav Sedov * 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 THE AUTHOR AND CONTRIBUTORS ``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 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. * */ #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 static d_open_t cpuctl_open; static d_ioctl_t cpuctl_ioctl; #define CPUCTL_VERSION 1 #ifdef DEBUG # define DPRINTF(format,...) printf(format, __VA_ARGS__); #else # define DPRINTF(...) #endif #define UCODE_SIZE_MAX (10 * 1024) static int cpuctl_do_msr(int cpu, cpuctl_msr_args_t *data, u_long cmd, struct thread *td); static int cpuctl_do_cpuid(int cpu, cpuctl_cpuid_args_t *data, struct thread *td); static int cpuctl_do_update(int cpu, cpuctl_update_args_t *data, struct thread *td); static int update_intel(int cpu, cpuctl_update_args_t *args, struct thread *td); static int update_amd(int cpu, cpuctl_update_args_t *args, struct thread *td); static int update_via(int cpu, cpuctl_update_args_t *args, struct thread *td); static struct cdev **cpuctl_devs; static MALLOC_DEFINE(M_CPUCTL, "cpuctl", "CPUCTL buffer"); static struct cdevsw cpuctl_cdevsw = { .d_version = D_VERSION, .d_open = cpuctl_open, .d_ioctl = cpuctl_ioctl, .d_name = "cpuctl", }; /* * This function checks if specified cpu enabled or not. */ static int cpu_enabled(int cpu) { return (pmc_cpu_is_disabled(cpu) == 0); } /* * Check if the current thread is bound to a specific cpu. */ static int cpu_sched_is_bound(struct thread *td) { int ret; thread_lock(td); ret = sched_is_bound(td); thread_unlock(td); return (ret); } /* * Switch to target cpu to run. */ static void set_cpu(int cpu, struct thread *td) { KASSERT(cpu >= 0 && cpu < mp_ncpus && cpu_enabled(cpu), ("[cpuctl,%d]: bad cpu number %d", __LINE__, cpu)); thread_lock(td); sched_bind(td, cpu); thread_unlock(td); KASSERT(td->td_oncpu == cpu, ("[cpuctl,%d]: cannot bind to target cpu %d", __LINE__, cpu)); } static void restore_cpu(int oldcpu, int is_bound, struct thread *td) { KASSERT(oldcpu >= 0 && oldcpu < mp_ncpus && cpu_enabled(oldcpu), ("[cpuctl,%d]: bad cpu number %d", __LINE__, oldcpu)); thread_lock(td); if (is_bound == 0) sched_unbind(td); else sched_bind(td, oldcpu); thread_unlock(td); } int cpuctl_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int flags, struct thread *td) { int ret; int cpu = dev2unit(dev); if (cpu >= mp_ncpus || !cpu_enabled(cpu)) { DPRINTF("[cpuctl,%d]: bad cpu number %d\n", __LINE__, cpu); return (ENXIO); } /* Require write flag for "write" requests. */ if ((cmd == CPUCTL_WRMSR || cmd == CPUCTL_UPDATE) && ((flags & FWRITE) == 0)) return (EPERM); switch (cmd) { case CPUCTL_RDMSR: ret = cpuctl_do_msr(cpu, (cpuctl_msr_args_t *)data, cmd, td); break; case CPUCTL_MSRSBIT: case CPUCTL_MSRCBIT: case CPUCTL_WRMSR: ret = priv_check(td, PRIV_CPUCTL_WRMSR); if (ret != 0) goto fail; ret = cpuctl_do_msr(cpu, (cpuctl_msr_args_t *)data, cmd, td); break; case CPUCTL_CPUID: ret = cpuctl_do_cpuid(cpu, (cpuctl_cpuid_args_t *)data, td); break; case CPUCTL_UPDATE: ret = priv_check(td, PRIV_CPUCTL_UPDATE); if (ret != 0) goto fail; ret = cpuctl_do_update(cpu, (cpuctl_update_args_t *)data, td); break; default: ret = EINVAL; break; } fail: return (ret); } /* * Actually perform cpuid operation. */ static int cpuctl_do_cpuid(int cpu, cpuctl_cpuid_args_t *data, struct thread *td) { int is_bound = 0; int oldcpu; KASSERT(cpu >= 0 && cpu < mp_ncpus, ("[cpuctl,%d]: bad cpu number %d", __LINE__, cpu)); /* Explicitly clear cpuid data to avoid returning stale info. */ bzero(data->data, sizeof(data->data)); DPRINTF("[cpuctl,%d]: retriving cpuid level %#0x for %d cpu\n", __LINE__, data->level, cpu); oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); cpuid_count(data->level, 0, data->data); restore_cpu(oldcpu, is_bound, td); return (0); } /* * Actually perform MSR operations. */ static int cpuctl_do_msr(int cpu, cpuctl_msr_args_t *data, u_long cmd, struct thread *td) { uint64_t reg; int is_bound = 0; int oldcpu; int ret; KASSERT(cpu >= 0 && cpu < mp_ncpus, ("[cpuctl,%d]: bad cpu number %d", __LINE__, cpu)); /* * Explicitly clear cpuid data to avoid returning stale * info */ DPRINTF("[cpuctl,%d]: operating on MSR %#0x for %d cpu\n", __LINE__, data->msr, cpu); oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); if (cmd == CPUCTL_RDMSR) { data->data = 0; ret = rdmsr_safe(data->msr, &data->data); } else if (cmd == CPUCTL_WRMSR) { ret = wrmsr_safe(data->msr, data->data); } else if (cmd == CPUCTL_MSRSBIT) { critical_enter(); ret = rdmsr_safe(data->msr, ®); if (ret == 0) ret = wrmsr_safe(data->msr, reg | data->data); critical_exit(); } else if (cmd == CPUCTL_MSRCBIT) { critical_enter(); ret = rdmsr_safe(data->msr, ®); if (ret == 0) ret = wrmsr_safe(data->msr, reg & ~data->data); critical_exit(); } else panic("[cpuctl,%d]: unknown operation requested: %lu", __LINE__, cmd); restore_cpu(oldcpu, is_bound, td); return (ret); } /* * Actually perform microcode update. */ static int cpuctl_do_update(int cpu, cpuctl_update_args_t *data, struct thread *td) { cpuctl_cpuid_args_t args = { .level = 0, }; char vendor[13]; int ret; KASSERT(cpu >= 0 && cpu < mp_ncpus, ("[cpuctl,%d]: bad cpu number %d", __LINE__, cpu)); DPRINTF("[cpuctl,%d]: XXX %d", __LINE__, cpu); ret = cpuctl_do_cpuid(cpu, &args, td); if (ret != 0) { DPRINTF("[cpuctl,%d]: cannot retrive cpuid info for cpu %d", __LINE__, cpu); return (ENXIO); } ((uint32_t *)vendor)[0] = args.data[1]; ((uint32_t *)vendor)[1] = args.data[3]; ((uint32_t *)vendor)[2] = args.data[2]; vendor[12] = '\0'; if (strncmp(vendor, INTEL_VENDOR_ID, sizeof(INTEL_VENDOR_ID)) == 0) ret = update_intel(cpu, data, td); else if(strncmp(vendor, AMD_VENDOR_ID, sizeof(AMD_VENDOR_ID)) == 0) ret = update_amd(cpu, data, td); else if(strncmp(vendor, CENTAUR_VENDOR_ID, sizeof(CENTAUR_VENDOR_ID)) == 0) ret = update_via(cpu, data, td); else ret = ENXIO; return (ret); } static int update_intel(int cpu, cpuctl_update_args_t *args, struct thread *td) { void *ptr; uint64_t rev0, rev1; uint32_t tmp[4]; int is_bound; int oldcpu; int ret; if (args->size == 0 || args->data == NULL) { DPRINTF("[cpuctl,%d]: zero-sized firmware image", __LINE__); return (EINVAL); } if (args->size > UCODE_SIZE_MAX) { DPRINTF("[cpuctl,%d]: firmware image too large", __LINE__); return (EINVAL); } /* * 16 byte alignment required. Rely on the fact that * malloc(9) always returns the pointer aligned at least on * the size of the allocation. */ ptr = malloc(args->size + 16, M_CPUCTL, M_WAITOK); if (copyin(args->data, ptr, args->size) != 0) { DPRINTF("[cpuctl,%d]: copyin %p->%p of %zd bytes failed", __LINE__, args->data, ptr, args->size); ret = EFAULT; goto fail; } oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); critical_enter(); rdmsr_safe(MSR_BIOS_SIGN, &rev0); /* Get current micorcode revision. */ /* * Perform update. */ wrmsr_safe(MSR_BIOS_UPDT_TRIG, (uintptr_t)(ptr)); wrmsr_safe(MSR_BIOS_SIGN, 0); /* * Serialize instruction flow. */ do_cpuid(0, tmp); critical_exit(); rdmsr_safe(MSR_BIOS_SIGN, &rev1); /* Get new micorcode revision. */ restore_cpu(oldcpu, is_bound, td); if (rev1 > rev0) ret = 0; else ret = EEXIST; fail: free(ptr, M_CPUCTL); return (ret); } static int update_amd(int cpu, cpuctl_update_args_t *args, struct thread *td) { void *ptr = NULL; uint32_t tmp[4]; int is_bound = 0; int oldcpu; int ret; if (args->size == 0 || args->data == NULL) { DPRINTF("[cpuctl,%d]: zero-sized firmware image", __LINE__); return (EINVAL); } if (args->size > UCODE_SIZE_MAX) { DPRINTF("[cpuctl,%d]: firmware image too large", __LINE__); return (EINVAL); } /* * XXX Might not require contignous address space - needs check */ ptr = contigmalloc(args->size, M_CPUCTL, 0, 0, 0xffffffff, 16, 0); if (ptr == NULL) { DPRINTF("[cpuctl,%d]: cannot allocate %zd bytes of memory", __LINE__, args->size); return (ENOMEM); } if (copyin(args->data, ptr, args->size) != 0) { DPRINTF("[cpuctl,%d]: copyin %p->%p of %zd bytes failed", __LINE__, args->data, ptr, args->size); ret = EFAULT; goto fail; } oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); critical_enter(); /* * Perform update. */ wrmsr_safe(MSR_K8_UCODE_UPDATE, (uintptr_t)ptr); /* * Serialize instruction flow. */ do_cpuid(0, tmp); critical_exit(); restore_cpu(oldcpu, is_bound, td); ret = 0; fail: if (ptr != NULL) contigfree(ptr, args->size, M_CPUCTL); return (ret); } static int update_via(int cpu, cpuctl_update_args_t *args, struct thread *td) { void *ptr; uint64_t rev0, rev1, res; uint32_t tmp[4]; int is_bound; int oldcpu; int ret; if (args->size == 0 || args->data == NULL) { DPRINTF("[cpuctl,%d]: zero-sized firmware image", __LINE__); return (EINVAL); } if (args->size > UCODE_SIZE_MAX) { DPRINTF("[cpuctl,%d]: firmware image too large", __LINE__); return (EINVAL); } /* * 4 byte alignment required. */ ptr = malloc(args->size, M_CPUCTL, M_WAITOK); if (copyin(args->data, ptr, args->size) != 0) { DPRINTF("[cpuctl,%d]: copyin %p->%p of %zd bytes failed", __LINE__, args->data, ptr, args->size); ret = EFAULT; goto fail; } oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); critical_enter(); rdmsr_safe(MSR_BIOS_SIGN, &rev0); /* Get current micorcode revision. */ /* * Perform update. */ wrmsr_safe(MSR_BIOS_UPDT_TRIG, (uintptr_t)(ptr)); do_cpuid(1, tmp); /* * Result are in low byte of MSR FCR5: * 0x00: No update has been attempted since RESET. * 0x01: The last attempted update was successful. * 0x02: The last attempted update was unsuccessful due to a bad * environment. No update was loaded and any preexisting * patches are still active. * 0x03: The last attempted update was not applicable to this processor. * No update was loaded and any preexisting patches are still * active. * 0x04: The last attempted update was not successful due to an invalid * update data block. No update was loaded and any preexisting * patches are still active */ rdmsr_safe(0x1205, &res); res &= 0xff; critical_exit(); rdmsr_safe(MSR_BIOS_SIGN, &rev1); /* Get new microcode revision. */ restore_cpu(oldcpu, is_bound, td); DPRINTF("[cpu,%d]: rev0=%x rev1=%x res=%x\n", __LINE__, (unsigned)(rev0 >> 32), (unsigned)(rev1 >> 32), (unsigned)res); if (res != 0x01) ret = EINVAL; else ret = 0; fail: free(ptr, M_CPUCTL); return (ret); } int cpuctl_open(struct cdev *dev, int flags, int fmt __unused, struct thread *td) { int ret = 0; int cpu; cpu = dev2unit(dev); if (cpu >= mp_ncpus || !cpu_enabled(cpu)) { DPRINTF("[cpuctl,%d]: incorrect cpu number %d\n", __LINE__, cpu); return (ENXIO); } if (flags & FWRITE) ret = securelevel_gt(td->td_ucred, 0); return (ret); } static int cpuctl_modevent(module_t mod __unused, int type, void *data __unused) { int cpu; switch(type) { case MOD_LOAD: if ((cpu_feature & CPUID_MSR) == 0) { if (bootverbose) printf("cpuctl: not available.\n"); return (ENODEV); } if (bootverbose) printf("cpuctl: access to MSR registers/cpuid info.\n"); - cpuctl_devs = (struct cdev **)malloc(sizeof(void *) * mp_ncpus, - M_CPUCTL, M_WAITOK | M_ZERO); - if (cpuctl_devs == NULL) { - DPRINTF("[cpuctl,%d]: cannot allocate memory\n", - __LINE__); - return (ENOMEM); - } + cpuctl_devs = malloc(sizeof(*cpuctl_devs) * mp_ncpus, M_CPUCTL, + M_WAITOK | M_ZERO); for (cpu = 0; cpu < mp_ncpus; cpu++) if (cpu_enabled(cpu)) cpuctl_devs[cpu] = make_dev(&cpuctl_cdevsw, cpu, UID_ROOT, GID_KMEM, 0640, "cpuctl%d", cpu); break; case MOD_UNLOAD: for (cpu = 0; cpu < mp_ncpus; cpu++) { if (cpuctl_devs[cpu] != NULL) destroy_dev(cpuctl_devs[cpu]); } free(cpuctl_devs, M_CPUCTL); break; case MOD_SHUTDOWN: break; default: return (EOPNOTSUPP); } return (0); } DEV_MODULE(cpuctl, cpuctl_modevent, NULL); MODULE_VERSION(cpuctl, CPUCTL_VERSION); Index: stable/9/sys/dev/hwpmc/hwpmc_piv.c =================================================================== --- stable/9/sys/dev/hwpmc/hwpmc_piv.c (revision 263357) +++ stable/9/sys/dev/hwpmc/hwpmc_piv.c (revision 263358) @@ -1,1698 +1,1697 @@ /*- * Copyright (c) 2003-2007 Joseph Koshy * Copyright (c) 2007 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by A. Joseph Koshy under * sponsorship from the FreeBSD Foundation and Google, 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 THE AUTHOR AND CONTRIBUTORS ``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 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. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * PENTIUM 4 SUPPORT * * The P4 has 18 PMCs, divided into 4 groups with 4,4,4 and 6 PMCs * respectively. Each PMC comprises of two model specific registers: * a counter configuration control register (CCCR) and a counter * register that holds the actual event counts. * * Configuring an event requires the use of one of 45 event selection * control registers (ESCR). Events are associated with specific * ESCRs. Each PMC group has a set of ESCRs it can use. * * - The BPU counter group (4 PMCs) can use the 16 ESCRs: * BPU_ESCR{0,1}, IS_ESCR{0,1}, MOB_ESCR{0,1}, ITLB_ESCR{0,1}, * PMH_ESCR{0,1}, IX_ESCR{0,1}, FSB_ESCR{0,}, BSU_ESCR{0,1}. * * - The MS counter group (4 PMCs) can use the 6 ESCRs: MS_ESCR{0,1}, * TC_ESCR{0,1}, TBPU_ESCR{0,1}. * * - The FLAME counter group (4 PMCs) can use the 10 ESCRs: * FLAME_ESCR{0,1}, FIRM_ESCR{0,1}, SAAT_ESCR{0,1}, U2L_ESCR{0,1}, * DAC_ESCR{0,1}. * * - The IQ counter group (6 PMCs) can use the 13 ESCRs: IQ_ESCR{0,1}, * ALF_ESCR{0,1}, RAT_ESCR{0,1}, SSU_ESCR0, CRU_ESCR{0,1,2,3,4,5}. * * Even-numbered ESCRs can be used with counters 0, 1 and 4 (if * present) of a counter group. Odd-numbers ESCRs can be used with * counters 2, 3 and 5 (if present) of a counter group. The * 'p4_escrs[]' table describes these restrictions in a form that * function 'p4_allocate()' uses for making allocation decisions. * * SYSTEM-MODE AND THREAD-MODE ALLOCATION * * In addition to remembering the state of PMC rows * ('FREE','STANDALONE', or 'THREAD'), we similar need to track the * state of ESCR rows. If an ESCR is allocated to a system-mode PMC * on a CPU we cannot allocate this to a thread-mode PMC. On a * multi-cpu (multiple physical CPUs) system, ESCR allocation on each * CPU is tracked by the pc_escrs[] array. * * Each system-mode PMC that is using an ESCR records its row-index in * the appropriate entry and system-mode allocation attempts check * that an ESCR is available using this array. Process-mode PMCs do * not use the pc_escrs[] array, since ESCR row itself would have been * marked as in 'THREAD' mode. * * HYPERTHREADING SUPPORT * * When HTT is enabled, the FreeBSD kernel treats the two 'logical' * cpus as independent CPUs and can schedule kernel threads on them * independently. However, the two logical CPUs share the same set of * PMC resources. We need to ensure that: * - PMCs that use the PMC_F_DESCENDANTS semantics are handled correctly, * and, * - Threads of multi-threaded processes that get scheduled on the same * physical CPU are handled correctly. * * HTT Detection * * Not all HTT capable systems will have HTT enabled. We detect the * presence of HTT by detecting if 'p4_init()' was called for a secondary * CPU in a HTT pair. * * Note that hwpmc(4) cannot currently deal with a change in HTT status once * loaded. * * Handling HTT READ / WRITE / START / STOP * * PMC resources are shared across the CPUs in an HTT pair. We * designate the lower numbered CPU in a HTT pair as the 'primary' * CPU. In each primary CPU's state we keep track of a 'runcount' * which reflects the number of PMC-using processes that have been * scheduled on its secondary CPU. Process-mode PMC operations will * actually 'start' or 'stop' hardware only if these are the first or * last processes respectively to use the hardware. PMC values * written by a 'write' operation are saved and are transferred to * hardware at PMC 'start' time if the runcount is 0. If the runcount * is greater than 0 at the time of a 'start' operation, we keep track * of the actual hardware value at the time of the 'start' operation * and use this to adjust the final readings at PMC 'stop' or 'read' * time. * * Execution sequences: * * Case 1: CPUx +...- (no overlap) * CPUy +...- * RC 0 1 0 1 0 * * Case 2: CPUx +........- (partial overlap) * CPUy +........- * RC 0 1 2 1 0 * * Case 3: CPUx +..............- (fully overlapped) * CPUy +.....- * RC 0 1 2 1 0 * * Key: * 'CPU[xy]' : one of the two logical processors on a HTT CPU. * 'RC' : run count (#threads per physical core). * '+' : point in time when a thread is put on a CPU. * '-' : point in time where a thread is taken off a CPU. * * Handling HTT CONFIG * * Different processes attached to the same PMC may get scheduled on * the two logical processors in the package. We keep track of config * and de-config operations using the CFGFLAGS fields of the per-physical * cpu state. */ #define P4_PMCS() \ P4_PMC(BPU_COUNTER0) \ P4_PMC(BPU_COUNTER1) \ P4_PMC(BPU_COUNTER2) \ P4_PMC(BPU_COUNTER3) \ P4_PMC(MS_COUNTER0) \ P4_PMC(MS_COUNTER1) \ P4_PMC(MS_COUNTER2) \ P4_PMC(MS_COUNTER3) \ P4_PMC(FLAME_COUNTER0) \ P4_PMC(FLAME_COUNTER1) \ P4_PMC(FLAME_COUNTER2) \ P4_PMC(FLAME_COUNTER3) \ P4_PMC(IQ_COUNTER0) \ P4_PMC(IQ_COUNTER1) \ P4_PMC(IQ_COUNTER2) \ P4_PMC(IQ_COUNTER3) \ P4_PMC(IQ_COUNTER4) \ P4_PMC(IQ_COUNTER5) \ P4_PMC(NONE) enum pmc_p4pmc { #undef P4_PMC #define P4_PMC(N) P4_PMC_##N , P4_PMCS() }; /* * P4 ESCR descriptors */ #define P4_ESCRS() \ P4_ESCR(BSU_ESCR0, 0x3A0, BPU_COUNTER0, BPU_COUNTER1, NONE) \ P4_ESCR(BSU_ESCR1, 0x3A1, BPU_COUNTER2, BPU_COUNTER3, NONE) \ P4_ESCR(FSB_ESCR0, 0x3A2, BPU_COUNTER0, BPU_COUNTER1, NONE) \ P4_ESCR(FSB_ESCR1, 0x3A3, BPU_COUNTER2, BPU_COUNTER3, NONE) \ P4_ESCR(FIRM_ESCR0, 0x3A4, FLAME_COUNTER0, FLAME_COUNTER1, NONE) \ P4_ESCR(FIRM_ESCR1, 0x3A5, FLAME_COUNTER2, FLAME_COUNTER3, NONE) \ P4_ESCR(FLAME_ESCR0, 0x3A6, FLAME_COUNTER0, FLAME_COUNTER1, NONE) \ P4_ESCR(FLAME_ESCR1, 0x3A7, FLAME_COUNTER2, FLAME_COUNTER3, NONE) \ P4_ESCR(DAC_ESCR0, 0x3A8, FLAME_COUNTER0, FLAME_COUNTER1, NONE) \ P4_ESCR(DAC_ESCR1, 0x3A9, FLAME_COUNTER2, FLAME_COUNTER3, NONE) \ P4_ESCR(MOB_ESCR0, 0x3AA, BPU_COUNTER0, BPU_COUNTER1, NONE) \ P4_ESCR(MOB_ESCR1, 0x3AB, BPU_COUNTER2, BPU_COUNTER3, NONE) \ P4_ESCR(PMH_ESCR0, 0x3AC, BPU_COUNTER0, BPU_COUNTER1, NONE) \ P4_ESCR(PMH_ESCR1, 0x3AD, BPU_COUNTER2, BPU_COUNTER3, NONE) \ P4_ESCR(SAAT_ESCR0, 0x3AE, FLAME_COUNTER0, FLAME_COUNTER1, NONE) \ P4_ESCR(SAAT_ESCR1, 0x3AF, FLAME_COUNTER2, FLAME_COUNTER3, NONE) \ P4_ESCR(U2L_ESCR0, 0x3B0, FLAME_COUNTER0, FLAME_COUNTER1, NONE) \ P4_ESCR(U2L_ESCR1, 0x3B1, FLAME_COUNTER2, FLAME_COUNTER3, NONE) \ P4_ESCR(BPU_ESCR0, 0x3B2, BPU_COUNTER0, BPU_COUNTER1, NONE) \ P4_ESCR(BPU_ESCR1, 0x3B3, BPU_COUNTER2, BPU_COUNTER3, NONE) \ P4_ESCR(IS_ESCR0, 0x3B4, BPU_COUNTER0, BPU_COUNTER1, NONE) \ P4_ESCR(IS_ESCR1, 0x3B5, BPU_COUNTER2, BPU_COUNTER3, NONE) \ P4_ESCR(ITLB_ESCR0, 0x3B6, BPU_COUNTER0, BPU_COUNTER1, NONE) \ P4_ESCR(ITLB_ESCR1, 0x3B7, BPU_COUNTER2, BPU_COUNTER3, NONE) \ P4_ESCR(CRU_ESCR0, 0x3B8, IQ_COUNTER0, IQ_COUNTER1, IQ_COUNTER4) \ P4_ESCR(CRU_ESCR1, 0x3B9, IQ_COUNTER2, IQ_COUNTER3, IQ_COUNTER5) \ P4_ESCR(IQ_ESCR0, 0x3BA, IQ_COUNTER0, IQ_COUNTER1, IQ_COUNTER4) \ P4_ESCR(IQ_ESCR1, 0x3BB, IQ_COUNTER1, IQ_COUNTER3, IQ_COUNTER5) \ P4_ESCR(RAT_ESCR0, 0x3BC, IQ_COUNTER0, IQ_COUNTER1, IQ_COUNTER4) \ P4_ESCR(RAT_ESCR1, 0x3BD, IQ_COUNTER2, IQ_COUNTER3, IQ_COUNTER5) \ P4_ESCR(SSU_ESCR0, 0x3BE, IQ_COUNTER0, IQ_COUNTER2, IQ_COUNTER4) \ P4_ESCR(MS_ESCR0, 0x3C0, MS_COUNTER0, MS_COUNTER1, NONE) \ P4_ESCR(MS_ESCR1, 0x3C1, MS_COUNTER2, MS_COUNTER3, NONE) \ P4_ESCR(TBPU_ESCR0, 0x3C2, MS_COUNTER0, MS_COUNTER1, NONE) \ P4_ESCR(TBPU_ESCR1, 0x3C3, MS_COUNTER2, MS_COUNTER3, NONE) \ P4_ESCR(TC_ESCR0, 0x3C4, MS_COUNTER0, MS_COUNTER1, NONE) \ P4_ESCR(TC_ESCR1, 0x3C5, MS_COUNTER2, MS_COUNTER3, NONE) \ P4_ESCR(IX_ESCR0, 0x3C8, BPU_COUNTER0, BPU_COUNTER1, NONE) \ P4_ESCR(IX_ESCR1, 0x3C9, BPU_COUNTER2, BPU_COUNTER3, NONE) \ P4_ESCR(ALF_ESCR0, 0x3CA, IQ_COUNTER0, IQ_COUNTER1, IQ_COUNTER4) \ P4_ESCR(ALF_ESCR1, 0x3CB, IQ_COUNTER2, IQ_COUNTER3, IQ_COUNTER5) \ P4_ESCR(CRU_ESCR2, 0x3CC, IQ_COUNTER0, IQ_COUNTER1, IQ_COUNTER4) \ P4_ESCR(CRU_ESCR3, 0x3CD, IQ_COUNTER2, IQ_COUNTER3, IQ_COUNTER5) \ P4_ESCR(CRU_ESCR4, 0x3E0, IQ_COUNTER0, IQ_COUNTER1, IQ_COUNTER4) \ P4_ESCR(CRU_ESCR5, 0x3E1, IQ_COUNTER2, IQ_COUNTER3, IQ_COUNTER5) \ P4_ESCR(NONE, ~0, NONE, NONE, NONE) enum pmc_p4escr { #define P4_ESCR(N, MSR, P1, P2, P3) P4_ESCR_##N , P4_ESCRS() #undef P4_ESCR }; struct pmc_p4escr_descr { const char pm_escrname[PMC_NAME_MAX]; u_short pm_escr_msr; const enum pmc_p4pmc pm_pmcs[P4_MAX_PMC_PER_ESCR]; }; static struct pmc_p4escr_descr p4_escrs[] = { #define P4_ESCR(N, MSR, P1, P2, P3) \ { \ .pm_escrname = #N, \ .pm_escr_msr = (MSR), \ .pm_pmcs = \ { \ P4_PMC_##P1, \ P4_PMC_##P2, \ P4_PMC_##P3 \ } \ } , P4_ESCRS() #undef P4_ESCR }; /* * P4 Event descriptor */ struct p4_event_descr { const enum pmc_event pm_event; const uint32_t pm_escr_eventselect; const uint32_t pm_cccr_select; const char pm_is_ti_event; enum pmc_p4escr pm_escrs[P4_MAX_ESCR_PER_EVENT]; }; static struct p4_event_descr p4_events[] = { #define P4_EVDESCR(NAME, ESCREVENTSEL, CCCRSEL, TI_EVENT, ESCR0, ESCR1) \ { \ .pm_event = PMC_EV_P4_##NAME, \ .pm_escr_eventselect = (ESCREVENTSEL), \ .pm_cccr_select = (CCCRSEL), \ .pm_is_ti_event = (TI_EVENT), \ .pm_escrs = \ { \ P4_ESCR_##ESCR0, \ P4_ESCR_##ESCR1 \ } \ } P4_EVDESCR(TC_DELIVER_MODE, 0x01, 0x01, TRUE, TC_ESCR0, TC_ESCR1), P4_EVDESCR(BPU_FETCH_REQUEST, 0x03, 0x00, FALSE, BPU_ESCR0, BPU_ESCR1), P4_EVDESCR(ITLB_REFERENCE, 0x18, 0x03, FALSE, ITLB_ESCR0, ITLB_ESCR1), P4_EVDESCR(MEMORY_CANCEL, 0x02, 0x05, FALSE, DAC_ESCR0, DAC_ESCR1), P4_EVDESCR(MEMORY_COMPLETE, 0x08, 0x02, FALSE, SAAT_ESCR0, SAAT_ESCR1), P4_EVDESCR(LOAD_PORT_REPLAY, 0x04, 0x02, FALSE, SAAT_ESCR0, SAAT_ESCR1), P4_EVDESCR(STORE_PORT_REPLAY, 0x05, 0x02, FALSE, SAAT_ESCR0, SAAT_ESCR1), P4_EVDESCR(MOB_LOAD_REPLAY, 0x03, 0x02, FALSE, MOB_ESCR0, MOB_ESCR1), P4_EVDESCR(PAGE_WALK_TYPE, 0x01, 0x04, TRUE, PMH_ESCR0, PMH_ESCR1), P4_EVDESCR(BSQ_CACHE_REFERENCE, 0x0C, 0x07, FALSE, BSU_ESCR0, BSU_ESCR1), P4_EVDESCR(IOQ_ALLOCATION, 0x03, 0x06, FALSE, FSB_ESCR0, FSB_ESCR1), P4_EVDESCR(IOQ_ACTIVE_ENTRIES, 0x1A, 0x06, FALSE, FSB_ESCR1, NONE), P4_EVDESCR(FSB_DATA_ACTIVITY, 0x17, 0x06, TRUE, FSB_ESCR0, FSB_ESCR1), P4_EVDESCR(BSQ_ALLOCATION, 0x05, 0x07, FALSE, BSU_ESCR0, NONE), P4_EVDESCR(BSQ_ACTIVE_ENTRIES, 0x06, 0x07, FALSE, BSU_ESCR1, NONE), /* BSQ_ACTIVE_ENTRIES inherits CPU specificity from BSQ_ALLOCATION */ P4_EVDESCR(SSE_INPUT_ASSIST, 0x34, 0x01, TRUE, FIRM_ESCR0, FIRM_ESCR1), P4_EVDESCR(PACKED_SP_UOP, 0x08, 0x01, TRUE, FIRM_ESCR0, FIRM_ESCR1), P4_EVDESCR(PACKED_DP_UOP, 0x0C, 0x01, TRUE, FIRM_ESCR0, FIRM_ESCR1), P4_EVDESCR(SCALAR_SP_UOP, 0x0A, 0x01, TRUE, FIRM_ESCR0, FIRM_ESCR1), P4_EVDESCR(SCALAR_DP_UOP, 0x0E, 0x01, TRUE, FIRM_ESCR0, FIRM_ESCR1), P4_EVDESCR(64BIT_MMX_UOP, 0x02, 0x01, TRUE, FIRM_ESCR0, FIRM_ESCR1), P4_EVDESCR(128BIT_MMX_UOP, 0x1A, 0x01, TRUE, FIRM_ESCR0, FIRM_ESCR1), P4_EVDESCR(X87_FP_UOP, 0x04, 0x01, TRUE, FIRM_ESCR0, FIRM_ESCR1), P4_EVDESCR(X87_SIMD_MOVES_UOP, 0x2E, 0x01, TRUE, FIRM_ESCR0, FIRM_ESCR1), P4_EVDESCR(GLOBAL_POWER_EVENTS, 0x13, 0x06, FALSE, FSB_ESCR0, FSB_ESCR1), P4_EVDESCR(TC_MS_XFER, 0x05, 0x00, FALSE, MS_ESCR0, MS_ESCR1), P4_EVDESCR(UOP_QUEUE_WRITES, 0x09, 0x00, FALSE, MS_ESCR0, MS_ESCR1), P4_EVDESCR(RETIRED_MISPRED_BRANCH_TYPE, 0x05, 0x02, FALSE, TBPU_ESCR0, TBPU_ESCR1), P4_EVDESCR(RETIRED_BRANCH_TYPE, 0x04, 0x02, FALSE, TBPU_ESCR0, TBPU_ESCR1), P4_EVDESCR(RESOURCE_STALL, 0x01, 0x01, FALSE, ALF_ESCR0, ALF_ESCR1), P4_EVDESCR(WC_BUFFER, 0x05, 0x05, TRUE, DAC_ESCR0, DAC_ESCR1), P4_EVDESCR(B2B_CYCLES, 0x16, 0x03, TRUE, FSB_ESCR0, FSB_ESCR1), P4_EVDESCR(BNR, 0x08, 0x03, TRUE, FSB_ESCR0, FSB_ESCR1), P4_EVDESCR(SNOOP, 0x06, 0x03, TRUE, FSB_ESCR0, FSB_ESCR1), P4_EVDESCR(RESPONSE, 0x04, 0x03, TRUE, FSB_ESCR0, FSB_ESCR1), P4_EVDESCR(FRONT_END_EVENT, 0x08, 0x05, FALSE, CRU_ESCR2, CRU_ESCR3), P4_EVDESCR(EXECUTION_EVENT, 0x0C, 0x05, FALSE, CRU_ESCR2, CRU_ESCR3), P4_EVDESCR(REPLAY_EVENT, 0x09, 0x05, FALSE, CRU_ESCR2, CRU_ESCR3), P4_EVDESCR(INSTR_RETIRED, 0x02, 0x04, FALSE, CRU_ESCR0, CRU_ESCR1), P4_EVDESCR(UOPS_RETIRED, 0x01, 0x04, FALSE, CRU_ESCR0, CRU_ESCR1), P4_EVDESCR(UOP_TYPE, 0x02, 0x02, FALSE, RAT_ESCR0, RAT_ESCR1), P4_EVDESCR(BRANCH_RETIRED, 0x06, 0x05, FALSE, CRU_ESCR2, CRU_ESCR3), P4_EVDESCR(MISPRED_BRANCH_RETIRED, 0x03, 0x04, FALSE, CRU_ESCR0, CRU_ESCR1), P4_EVDESCR(X87_ASSIST, 0x03, 0x05, FALSE, CRU_ESCR2, CRU_ESCR3), P4_EVDESCR(MACHINE_CLEAR, 0x02, 0x05, FALSE, CRU_ESCR2, CRU_ESCR3) #undef P4_EVDESCR }; #define P4_EVENT_IS_TI(E) ((E)->pm_is_ti_event == TRUE) #define P4_NEVENTS (PMC_EV_P4_LAST - PMC_EV_P4_FIRST + 1) /* * P4 PMC descriptors */ struct p4pmc_descr { struct pmc_descr pm_descr; /* common information */ enum pmc_p4pmc pm_pmcnum; /* PMC number */ uint32_t pm_pmc_msr; /* PERFCTR MSR address */ uint32_t pm_cccr_msr; /* CCCR MSR address */ }; static struct p4pmc_descr p4_pmcdesc[P4_NPMCS] = { #define P4_PMC_CAPS (PMC_CAP_INTERRUPT | PMC_CAP_USER | PMC_CAP_SYSTEM | \ PMC_CAP_EDGE | PMC_CAP_THRESHOLD | PMC_CAP_READ | PMC_CAP_WRITE | \ PMC_CAP_INVERT | PMC_CAP_QUALIFIER | PMC_CAP_PRECISE | \ PMC_CAP_TAGGING | PMC_CAP_CASCADE) #define P4_PMCDESCR(N, PMC, CCCR) \ { \ .pm_descr = \ { \ .pd_name = #N, \ .pd_class = PMC_CLASS_P4, \ .pd_caps = P4_PMC_CAPS, \ .pd_width = 40 \ }, \ .pm_pmcnum = P4_PMC_##N, \ .pm_cccr_msr = (CCCR), \ .pm_pmc_msr = (PMC) \ } P4_PMCDESCR(BPU_COUNTER0, 0x300, 0x360), P4_PMCDESCR(BPU_COUNTER1, 0x301, 0x361), P4_PMCDESCR(BPU_COUNTER2, 0x302, 0x362), P4_PMCDESCR(BPU_COUNTER3, 0x303, 0x363), P4_PMCDESCR(MS_COUNTER0, 0x304, 0x364), P4_PMCDESCR(MS_COUNTER1, 0x305, 0x365), P4_PMCDESCR(MS_COUNTER2, 0x306, 0x366), P4_PMCDESCR(MS_COUNTER3, 0x307, 0x367), P4_PMCDESCR(FLAME_COUNTER0, 0x308, 0x368), P4_PMCDESCR(FLAME_COUNTER1, 0x309, 0x369), P4_PMCDESCR(FLAME_COUNTER2, 0x30A, 0x36A), P4_PMCDESCR(FLAME_COUNTER3, 0x30B, 0x36B), P4_PMCDESCR(IQ_COUNTER0, 0x30C, 0x36C), P4_PMCDESCR(IQ_COUNTER1, 0x30D, 0x36D), P4_PMCDESCR(IQ_COUNTER2, 0x30E, 0x36E), P4_PMCDESCR(IQ_COUNTER3, 0x30F, 0x36F), P4_PMCDESCR(IQ_COUNTER4, 0x310, 0x370), P4_PMCDESCR(IQ_COUNTER5, 0x311, 0x371), #undef P4_PMCDESCR }; /* HTT support */ #define P4_NHTT 2 /* logical processors/chip */ static int p4_system_has_htt; /* * Per-CPU data structure for P4 class CPUs * * [19 struct pmc_hw structures] * [45 ESCRs status bytes] * [per-cpu spin mutex] * [19 flag fields for holding config flags and a runcount] * [19*2 hw value fields] (Thread mode PMC support) * or * [19*2 EIP values] (Sampling mode PMCs) * [19*2 pmc value fields] (Thread mode PMC support)) */ struct p4_cpu { struct pmc_hw pc_p4pmcs[P4_NPMCS]; char pc_escrs[P4_NESCR]; struct mtx pc_mtx; /* spin lock */ uint32_t pc_intrflag; /* NMI handler flags */ unsigned int pc_intrlock; /* NMI handler spin lock */ unsigned char pc_flags[P4_NPMCS]; /* 4 bits each: {cfg,run}count */ union { pmc_value_t pc_hw[P4_NPMCS * P4_NHTT]; uintptr_t pc_ip[P4_NPMCS * P4_NHTT]; } pc_si; pmc_value_t pc_pmc_values[P4_NPMCS * P4_NHTT]; }; static struct p4_cpu **p4_pcpu; #define P4_PCPU_PMC_VALUE(PC,RI,CPU) (PC)->pc_pmc_values[(RI)*((CPU) & 1)] #define P4_PCPU_HW_VALUE(PC,RI,CPU) (PC)->pc_si.pc_hw[(RI)*((CPU) & 1)] #define P4_PCPU_SAVED_IP(PC,RI,CPU) (PC)->pc_si.pc_ip[(RI)*((CPU) & 1)] #define P4_PCPU_GET_FLAGS(PC,RI,MASK) ((PC)->pc_flags[(RI)] & (MASK)) #define P4_PCPU_SET_FLAGS(PC,RI,MASK,VAL) do { \ char _tmp; \ _tmp = (PC)->pc_flags[(RI)]; \ _tmp &= ~(MASK); \ _tmp |= (VAL) & (MASK); \ (PC)->pc_flags[(RI)] = _tmp; \ } while (0) #define P4_PCPU_GET_RUNCOUNT(PC,RI) P4_PCPU_GET_FLAGS(PC,RI,0x0F) #define P4_PCPU_SET_RUNCOUNT(PC,RI,V) P4_PCPU_SET_FLAGS(PC,RI,0x0F,V) #define P4_PCPU_GET_CFGFLAGS(PC,RI) (P4_PCPU_GET_FLAGS(PC,RI,0xF0) >> 4) #define P4_PCPU_SET_CFGFLAGS(PC,RI,C) P4_PCPU_SET_FLAGS(PC,RI,0xF0,((C) <<4)) #define P4_CPU_TO_FLAG(C) (P4_CPU_IS_HTT_SECONDARY(cpu) ? 0x2 : 0x1) #define P4_PCPU_GET_INTRFLAG(PC,I) ((PC)->pc_intrflag & (1 << (I))) #define P4_PCPU_SET_INTRFLAG(PC,I,V) do { \ uint32_t __mask; \ __mask = 1 << (I); \ if ((V)) \ (PC)->pc_intrflag |= __mask; \ else \ (PC)->pc_intrflag &= ~__mask; \ } while (0) /* * A minimal spin lock implementation for use inside the NMI handler. * * We don't want to use a regular spin lock here, because curthread * may not be consistent at the time the handler is invoked. */ #define P4_PCPU_ACQ_INTR_SPINLOCK(PC) do { \ while (!atomic_cmpset_acq_int(&pc->pc_intrlock, 0, 1)) \ ia32_pause(); \ } while (0) #define P4_PCPU_REL_INTR_SPINLOCK(PC) \ atomic_store_rel_int(&pc->pc_intrlock, 0); /* ESCR row disposition */ static int p4_escrdisp[P4_NESCR]; #define P4_ESCR_ROW_DISP_IS_THREAD(E) (p4_escrdisp[(E)] > 0) #define P4_ESCR_ROW_DISP_IS_STANDALONE(E) (p4_escrdisp[(E)] < 0) #define P4_ESCR_ROW_DISP_IS_FREE(E) (p4_escrdisp[(E)] == 0) #define P4_ESCR_MARK_ROW_STANDALONE(E) do { \ KASSERT(p4_escrdisp[(E)] <= 0, ("[p4,%d] row disposition error",\ __LINE__)); \ atomic_add_int(&p4_escrdisp[(E)], -1); \ KASSERT(p4_escrdisp[(E)] >= (-pmc_cpu_max_active()), \ ("[p4,%d] row disposition error", __LINE__)); \ } while (0) #define P4_ESCR_UNMARK_ROW_STANDALONE(E) do { \ atomic_add_int(&p4_escrdisp[(E)], 1); \ KASSERT(p4_escrdisp[(E)] <= 0, ("[p4,%d] row disposition error",\ __LINE__)); \ } while (0) #define P4_ESCR_MARK_ROW_THREAD(E) do { \ KASSERT(p4_escrdisp[(E)] >= 0, ("[p4,%d] row disposition error", \ __LINE__)); \ atomic_add_int(&p4_escrdisp[(E)], 1); \ } while (0) #define P4_ESCR_UNMARK_ROW_THREAD(E) do { \ atomic_add_int(&p4_escrdisp[(E)], -1); \ KASSERT(p4_escrdisp[(E)] >= 0, ("[p4,%d] row disposition error", \ __LINE__)); \ } while (0) #define P4_PMC_IS_STOPPED(cccr) ((rdmsr(cccr) & P4_CCCR_ENABLE) == 0) #define P4_CPU_IS_HTT_SECONDARY(cpu) \ (p4_system_has_htt ? ((cpu) & 1) : 0) #define P4_TO_HTT_PRIMARY(cpu) \ (p4_system_has_htt ? ((cpu) & ~1) : (cpu)) #define P4_CCCR_Tx_MASK (~(P4_CCCR_OVF_PMI_T0|P4_CCCR_OVF_PMI_T1| \ P4_CCCR_ENABLE|P4_CCCR_OVF)) #define P4_ESCR_Tx_MASK (~(P4_ESCR_T0_OS|P4_ESCR_T0_USR|P4_ESCR_T1_OS| \ P4_ESCR_T1_USR)) /* * support routines */ static struct p4_event_descr * p4_find_event(enum pmc_event ev) { int n; for (n = 0; n < P4_NEVENTS; n++) if (p4_events[n].pm_event == ev) break; if (n == P4_NEVENTS) return (NULL); return (&p4_events[n]); } /* * Initialize per-cpu state */ static int p4_pcpu_init(struct pmc_mdep *md, int cpu) { char *pescr; int n, first_ri, phycpu; struct pmc_hw *phw; struct p4_cpu *p4c; struct pmc_cpu *pc, *plc; KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[p4,%d] insane cpu number %d", __LINE__, cpu)); PMCDBG(MDP,INI,0, "p4-init cpu=%d is-primary=%d", cpu, pmc_cpu_is_primary(cpu) != 0); first_ri = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_P4].pcd_ri; /* * The two CPUs in an HT pair share their per-cpu state. * * For HT capable CPUs, we assume that the two logical * processors in the HT pair get two consecutive CPU ids * starting with an even id #. * * The primary CPU (the even numbered CPU of the pair) would * have been initialized prior to the initialization for the * secondary. */ if (!pmc_cpu_is_primary(cpu) && (cpu & 1)) { p4_system_has_htt = 1; phycpu = P4_TO_HTT_PRIMARY(cpu); pc = pmc_pcpu[phycpu]; plc = pmc_pcpu[cpu]; KASSERT(plc != pc, ("[p4,%d] per-cpu config error", __LINE__)); PMCDBG(MDP,INI,1, "p4-init cpu=%d phycpu=%d pc=%p", cpu, phycpu, pc); KASSERT(pc, ("[p4,%d] Null Per-Cpu state cpu=%d phycpu=%d", __LINE__, cpu, phycpu)); /* PMCs are shared with the physical CPU. */ for (n = 0; n < P4_NPMCS; n++) plc->pc_hwpmcs[n + first_ri] = pc->pc_hwpmcs[n + first_ri]; return (0); } p4c = malloc(sizeof(struct p4_cpu), M_PMC, M_WAITOK|M_ZERO); if (p4c == NULL) return (ENOMEM); pc = pmc_pcpu[cpu]; KASSERT(pc != NULL, ("[p4,%d] cpu %d null per-cpu", __LINE__, cpu)); p4_pcpu[cpu] = p4c; phw = p4c->pc_p4pmcs; for (n = 0; n < P4_NPMCS; n++, phw++) { phw->phw_state = PMC_PHW_FLAG_IS_ENABLED | PMC_PHW_CPU_TO_STATE(cpu) | PMC_PHW_INDEX_TO_STATE(n); phw->phw_pmc = NULL; pc->pc_hwpmcs[n + first_ri] = phw; } pescr = p4c->pc_escrs; for (n = 0; n < P4_NESCR; n++) *pescr++ = P4_INVALID_PMC_INDEX; mtx_init(&p4c->pc_mtx, "p4-pcpu", "pmc-leaf", MTX_SPIN); return (0); } /* * Destroy per-cpu state. */ static int p4_pcpu_fini(struct pmc_mdep *md, int cpu) { int first_ri, i; struct p4_cpu *p4c; struct pmc_cpu *pc; PMCDBG(MDP,INI,0, "p4-cleanup cpu=%d", cpu); pc = pmc_pcpu[cpu]; first_ri = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_P4].pcd_ri; for (i = 0; i < P4_NPMCS; i++) pc->pc_hwpmcs[i + first_ri] = NULL; if (!pmc_cpu_is_primary(cpu) && (cpu & 1)) return (0); p4c = p4_pcpu[cpu]; KASSERT(p4c != NULL, ("[p4,%d] NULL pcpu", __LINE__)); /* Turn off all PMCs on this CPU */ for (i = 0; i < P4_NPMCS - 1; i++) wrmsr(P4_CCCR_MSR_FIRST + i, rdmsr(P4_CCCR_MSR_FIRST + i) & ~P4_CCCR_ENABLE); mtx_destroy(&p4c->pc_mtx); free(p4c, M_PMC); p4_pcpu[cpu] = NULL; return (0); } /* * Read a PMC */ static int p4_read_pmc(int cpu, int ri, pmc_value_t *v) { struct pmc *pm; pmc_value_t tmp; struct p4_cpu *pc; enum pmc_mode mode; struct p4pmc_descr *pd; KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[p4,%d] illegal CPU value %d", __LINE__, cpu)); KASSERT(ri >= 0 && ri < P4_NPMCS, ("[p4,%d] illegal row-index %d", __LINE__, ri)); pc = p4_pcpu[P4_TO_HTT_PRIMARY(cpu)]; pm = pc->pc_p4pmcs[ri].phw_pmc; pd = &p4_pmcdesc[ri]; KASSERT(pm != NULL, ("[p4,%d] No owner for HWPMC [cpu%d,pmc%d]", __LINE__, cpu, ri)); KASSERT(pd->pm_descr.pd_class == PMC_TO_CLASS(pm), ("[p4,%d] class mismatch pd %d != id class %d", __LINE__, pd->pm_descr.pd_class, PMC_TO_CLASS(pm))); mode = PMC_TO_MODE(pm); PMCDBG(MDP,REA,1, "p4-read cpu=%d ri=%d mode=%d", cpu, ri, mode); KASSERT(pd->pm_descr.pd_class == PMC_CLASS_P4, ("[p4,%d] unknown PMC class %d", __LINE__, pd->pm_descr.pd_class)); tmp = rdmsr(p4_pmcdesc[ri].pm_pmc_msr); if (PMC_IS_VIRTUAL_MODE(mode)) { if (tmp < P4_PCPU_HW_VALUE(pc,ri,cpu)) /* 40 bit overflow */ tmp += (P4_PERFCTR_MASK + 1) - P4_PCPU_HW_VALUE(pc,ri,cpu); else tmp -= P4_PCPU_HW_VALUE(pc,ri,cpu); tmp += P4_PCPU_PMC_VALUE(pc,ri,cpu); } if (PMC_IS_SAMPLING_MODE(mode)) /* undo transformation */ *v = P4_PERFCTR_VALUE_TO_RELOAD_COUNT(tmp); else *v = tmp; PMCDBG(MDP,REA,2, "p4-read -> %jx", *v); return (0); } /* * Write a PMC */ static int p4_write_pmc(int cpu, int ri, pmc_value_t v) { enum pmc_mode mode; struct pmc *pm; struct p4_cpu *pc; const struct pmc_hw *phw; const struct p4pmc_descr *pd; KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[amd,%d] illegal CPU value %d", __LINE__, cpu)); KASSERT(ri >= 0 && ri < P4_NPMCS, ("[amd,%d] illegal row-index %d", __LINE__, ri)); pc = p4_pcpu[P4_TO_HTT_PRIMARY(cpu)]; phw = &pc->pc_p4pmcs[ri]; pm = phw->phw_pmc; pd = &p4_pmcdesc[ri]; KASSERT(pm != NULL, ("[p4,%d] No owner for HWPMC [cpu%d,pmc%d]", __LINE__, cpu, ri)); mode = PMC_TO_MODE(pm); PMCDBG(MDP,WRI,1, "p4-write cpu=%d ri=%d mode=%d v=%jx", cpu, ri, mode, v); /* * write the PMC value to the register/saved value: for * sampling mode PMCs, the value to be programmed into the PMC * counter is -(C+1) where 'C' is the requested sample rate. */ if (PMC_IS_SAMPLING_MODE(mode)) v = P4_RELOAD_COUNT_TO_PERFCTR_VALUE(v); if (PMC_IS_SYSTEM_MODE(mode)) wrmsr(pd->pm_pmc_msr, v); else P4_PCPU_PMC_VALUE(pc,ri,cpu) = v; return (0); } /* * Configure a PMC 'pm' on the given CPU and row-index. * * 'pm' may be NULL to indicate de-configuration. * * On HTT systems, a PMC may get configured twice, once for each * "logical" CPU. We track this using the CFGFLAGS field of the * per-cpu state; this field is a bit mask with one bit each for * logical CPUs 0 & 1. */ static int p4_config_pmc(int cpu, int ri, struct pmc *pm) { struct pmc_hw *phw; struct p4_cpu *pc; int cfgflags, cpuflag; KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[p4,%d] illegal CPU %d", __LINE__, cpu)); KASSERT(ri >= 0 && ri < P4_NPMCS, ("[p4,%d] illegal row-index %d", __LINE__, ri)); PMCDBG(MDP,CFG,1, "cpu=%d ri=%d pm=%p", cpu, ri, pm); pc = p4_pcpu[P4_TO_HTT_PRIMARY(cpu)]; phw = &pc->pc_p4pmcs[ri]; KASSERT(pm == NULL || phw->phw_pmc == NULL || (p4_system_has_htt && phw->phw_pmc == pm), ("[p4,%d] hwpmc not unconfigured before re-config", __LINE__)); mtx_lock_spin(&pc->pc_mtx); cfgflags = P4_PCPU_GET_CFGFLAGS(pc,ri); KASSERT(cfgflags >= 0 || cfgflags <= 3, ("[p4,%d] illegal cfgflags cfg=%d on cpu=%d ri=%d", __LINE__, cfgflags, cpu, ri)); KASSERT(cfgflags == 0 || phw->phw_pmc, ("[p4,%d] cpu=%d ri=%d pmc configured with zero cfg count", __LINE__, cpu, ri)); cpuflag = P4_CPU_TO_FLAG(cpu); if (pm) { /* config */ if (cfgflags == 0) phw->phw_pmc = pm; KASSERT(phw->phw_pmc == pm, ("[p4,%d] cpu=%d ri=%d config %p != hw %p", __LINE__, cpu, ri, pm, phw->phw_pmc)); cfgflags |= cpuflag; } else { /* unconfig */ cfgflags &= ~cpuflag; if (cfgflags == 0) phw->phw_pmc = NULL; } KASSERT(cfgflags >= 0 || cfgflags <= 3, ("[p4,%d] illegal runcount cfg=%d on cpu=%d ri=%d", __LINE__, cfgflags, cpu, ri)); P4_PCPU_SET_CFGFLAGS(pc,ri,cfgflags); mtx_unlock_spin(&pc->pc_mtx); return (0); } /* * Retrieve a configured PMC pointer from hardware state. */ static int p4_get_config(int cpu, int ri, struct pmc **ppm) { int cfgflags; struct p4_cpu *pc; KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[p4,%d] illegal CPU %d", __LINE__, cpu)); KASSERT(ri >= 0 && ri < P4_NPMCS, ("[p4,%d] illegal row-index %d", __LINE__, ri)); pc = p4_pcpu[P4_TO_HTT_PRIMARY(cpu)]; mtx_lock_spin(&pc->pc_mtx); cfgflags = P4_PCPU_GET_CFGFLAGS(pc,ri); mtx_unlock_spin(&pc->pc_mtx); if (cfgflags & P4_CPU_TO_FLAG(cpu)) *ppm = pc->pc_p4pmcs[ri].phw_pmc; /* PMC config'ed on this CPU */ else *ppm = NULL; return 0; } /* * Allocate a PMC. * * The allocation strategy differs between HTT and non-HTT systems. * * The non-HTT case: * - Given the desired event and the PMC row-index, lookup the * list of valid ESCRs for the event. * - For each valid ESCR: * - Check if the ESCR is free and the ESCR row is in a compatible * mode (i.e., system or process)) * - Check if the ESCR is usable with a P4 PMC at the desired row-index. * If everything matches, we determine the appropriate bit values for the * ESCR and CCCR registers. * * The HTT case: * * - Process mode PMCs require special care. The FreeBSD scheduler could * schedule any two processes on the same physical CPU. We need to ensure * that a given PMC row-index is never allocated to two different * PMCs owned by different user-processes. * This is ensured by always allocating a PMC from a 'FREE' PMC row * if the system has HTT active. * - A similar check needs to be done for ESCRs; we do not want two PMCs * using the same ESCR to be scheduled at the same time. Thus ESCR * allocation is also restricted to FREE rows if the system has HTT * enabled. * - Thirdly, some events are 'thread-independent' terminology, i.e., * the PMC hardware cannot distinguish between events caused by * different logical CPUs. This makes it impossible to assign events * to a given thread of execution. If the system has HTT enabled, * these events are not allowed for process-mode PMCs. */ static int p4_allocate_pmc(int cpu, int ri, struct pmc *pm, const struct pmc_op_pmcallocate *a) { int found, n, m; uint32_t caps, cccrvalue, escrvalue, tflags; enum pmc_p4escr escr; struct p4_cpu *pc; struct p4_event_descr *pevent; const struct p4pmc_descr *pd; KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[p4,%d] illegal CPU %d", __LINE__, cpu)); KASSERT(ri >= 0 && ri < P4_NPMCS, ("[p4,%d] illegal row-index value %d", __LINE__, ri)); pd = &p4_pmcdesc[ri]; PMCDBG(MDP,ALL,1, "p4-allocate ri=%d class=%d pmccaps=0x%x " "reqcaps=0x%x", ri, pd->pm_descr.pd_class, pd->pm_descr.pd_caps, pm->pm_caps); /* check class */ if (pd->pm_descr.pd_class != a->pm_class) return (EINVAL); /* check requested capabilities */ caps = a->pm_caps; if ((pd->pm_descr.pd_caps & caps) != caps) return (EPERM); /* * If the system has HTT enabled, and the desired allocation * mode is process-private, and the PMC row disposition is not * FREE (0), decline the allocation. */ if (p4_system_has_htt && PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) && pmc_getrowdisp(ri) != 0) return (EBUSY); KASSERT(pd->pm_descr.pd_class == PMC_CLASS_P4, ("[p4,%d] unknown PMC class %d", __LINE__, pd->pm_descr.pd_class)); if (pm->pm_event < PMC_EV_P4_FIRST || pm->pm_event > PMC_EV_P4_LAST) return (EINVAL); if ((pevent = p4_find_event(pm->pm_event)) == NULL) return (ESRCH); PMCDBG(MDP,ALL,2, "pevent={ev=%d,escrsel=0x%x,cccrsel=0x%x,isti=%d}", pevent->pm_event, pevent->pm_escr_eventselect, pevent->pm_cccr_select, pevent->pm_is_ti_event); /* * Some PMC events are 'thread independent'and therefore * cannot be used for process-private modes if HTT is being * used. */ if (P4_EVENT_IS_TI(pevent) && PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) && p4_system_has_htt) return (EINVAL); pc = p4_pcpu[P4_TO_HTT_PRIMARY(cpu)]; found = 0; /* look for a suitable ESCR for this event */ for (n = 0; n < P4_MAX_ESCR_PER_EVENT && !found; n++) { if ((escr = pevent->pm_escrs[n]) == P4_ESCR_NONE) break; /* out of ESCRs */ /* * Check ESCR row disposition. * * If the request is for a system-mode PMC, then the * ESCR row should not be in process-virtual mode, and * should also be free on the current CPU. */ if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { if (P4_ESCR_ROW_DISP_IS_THREAD(escr) || pc->pc_escrs[escr] != P4_INVALID_PMC_INDEX) continue; } /* * If the request is for a process-virtual PMC, and if * HTT is not enabled, we can use an ESCR row that is * either FREE or already in process mode. * * If HTT is enabled, then we need to ensure that a * given ESCR is never allocated to two PMCS that * could run simultaneously on the two logical CPUs of * a CPU package. We ensure this be only allocating * ESCRs from rows marked as 'FREE'. */ if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { if (p4_system_has_htt) { if (!P4_ESCR_ROW_DISP_IS_FREE(escr)) continue; } else if (P4_ESCR_ROW_DISP_IS_STANDALONE(escr)) continue; } /* * We found a suitable ESCR for this event. Now check if * this escr can work with the PMC at row-index 'ri'. */ for (m = 0; m < P4_MAX_PMC_PER_ESCR; m++) if (p4_escrs[escr].pm_pmcs[m] == pd->pm_pmcnum) { found = 1; break; } } if (found == 0) return (ESRCH); KASSERT((int) escr >= 0 && escr < P4_NESCR, ("[p4,%d] illegal ESCR value %d", __LINE__, escr)); /* mark ESCR row mode */ if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { pc->pc_escrs[escr] = ri; /* mark ESCR as in use on this cpu */ P4_ESCR_MARK_ROW_STANDALONE(escr); } else { KASSERT(pc->pc_escrs[escr] == P4_INVALID_PMC_INDEX, ("[p4,%d] escr[%d] already in use", __LINE__, escr)); P4_ESCR_MARK_ROW_THREAD(escr); } pm->pm_md.pm_p4.pm_p4_escrmsr = p4_escrs[escr].pm_escr_msr; pm->pm_md.pm_p4.pm_p4_escr = escr; cccrvalue = P4_CCCR_TO_ESCR_SELECT(pevent->pm_cccr_select); escrvalue = P4_ESCR_TO_EVENT_SELECT(pevent->pm_escr_eventselect); /* CCCR fields */ if (caps & PMC_CAP_THRESHOLD) cccrvalue |= (a->pm_md.pm_p4.pm_p4_cccrconfig & P4_CCCR_THRESHOLD_MASK) | P4_CCCR_COMPARE; if (caps & PMC_CAP_EDGE) cccrvalue |= P4_CCCR_EDGE; if (caps & PMC_CAP_INVERT) cccrvalue |= P4_CCCR_COMPLEMENT; if (p4_system_has_htt) cccrvalue |= a->pm_md.pm_p4.pm_p4_cccrconfig & P4_CCCR_ACTIVE_THREAD_MASK; else /* no HTT; thread field should be '11b' */ cccrvalue |= P4_CCCR_TO_ACTIVE_THREAD(0x3); if (caps & PMC_CAP_CASCADE) cccrvalue |= P4_CCCR_CASCADE; /* On HTT systems the PMI T0 field may get moved to T1 at pmc start */ if (caps & PMC_CAP_INTERRUPT) cccrvalue |= P4_CCCR_OVF_PMI_T0; /* ESCR fields */ if (caps & PMC_CAP_QUALIFIER) escrvalue |= a->pm_md.pm_p4.pm_p4_escrconfig & P4_ESCR_EVENT_MASK_MASK; if (caps & PMC_CAP_TAGGING) escrvalue |= (a->pm_md.pm_p4.pm_p4_escrconfig & P4_ESCR_TAG_VALUE_MASK) | P4_ESCR_TAG_ENABLE; if (caps & PMC_CAP_QUALIFIER) escrvalue |= (a->pm_md.pm_p4.pm_p4_escrconfig & P4_ESCR_EVENT_MASK_MASK); /* HTT: T0_{OS,USR} bits may get moved to T1 at pmc start */ tflags = 0; if (caps & PMC_CAP_SYSTEM) tflags |= P4_ESCR_T0_OS; if (caps & PMC_CAP_USER) tflags |= P4_ESCR_T0_USR; if (tflags == 0) tflags = (P4_ESCR_T0_OS|P4_ESCR_T0_USR); escrvalue |= tflags; pm->pm_md.pm_p4.pm_p4_cccrvalue = cccrvalue; pm->pm_md.pm_p4.pm_p4_escrvalue = escrvalue; PMCDBG(MDP,ALL,2, "p4-allocate cccrsel=0x%x cccrval=0x%x " "escr=%d escrmsr=0x%x escrval=0x%x", pevent->pm_cccr_select, cccrvalue, escr, pm->pm_md.pm_p4.pm_p4_escrmsr, escrvalue); return (0); } /* * release a PMC. */ static int p4_release_pmc(int cpu, int ri, struct pmc *pm) { enum pmc_p4escr escr; struct p4_cpu *pc; KASSERT(ri >= 0 && ri < P4_NPMCS, ("[p4,%d] illegal row-index %d", __LINE__, ri)); escr = pm->pm_md.pm_p4.pm_p4_escr; PMCDBG(MDP,REL,1, "p4-release cpu=%d ri=%d escr=%d", cpu, ri, escr); if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { pc = p4_pcpu[P4_TO_HTT_PRIMARY(cpu)]; KASSERT(pc->pc_p4pmcs[ri].phw_pmc == NULL, ("[p4,%d] releasing configured PMC ri=%d", __LINE__, ri)); P4_ESCR_UNMARK_ROW_STANDALONE(escr); KASSERT(pc->pc_escrs[escr] == ri, ("[p4,%d] escr[%d] not allocated to ri %d", __LINE__, escr, ri)); pc->pc_escrs[escr] = P4_INVALID_PMC_INDEX; /* mark as free */ } else P4_ESCR_UNMARK_ROW_THREAD(escr); return (0); } /* * Start a PMC */ static int p4_start_pmc(int cpu, int ri) { int rc; struct pmc *pm; struct p4_cpu *pc; struct p4pmc_descr *pd; uint32_t cccrvalue, cccrtbits, escrvalue, escrmsr, escrtbits; KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[p4,%d] illegal CPU value %d", __LINE__, cpu)); KASSERT(ri >= 0 && ri < P4_NPMCS, ("[p4,%d] illegal row-index %d", __LINE__, ri)); pc = p4_pcpu[P4_TO_HTT_PRIMARY(cpu)]; pm = pc->pc_p4pmcs[ri].phw_pmc; pd = &p4_pmcdesc[ri]; KASSERT(pm != NULL, ("[p4,%d] starting cpu%d,pmc%d with null pmc", __LINE__, cpu, ri)); PMCDBG(MDP,STA,1, "p4-start cpu=%d ri=%d", cpu, ri); KASSERT(pd->pm_descr.pd_class == PMC_CLASS_P4, ("[p4,%d] wrong PMC class %d", __LINE__, pd->pm_descr.pd_class)); /* retrieve the desired CCCR/ESCR values from the PMC */ cccrvalue = pm->pm_md.pm_p4.pm_p4_cccrvalue; escrvalue = pm->pm_md.pm_p4.pm_p4_escrvalue; escrmsr = pm->pm_md.pm_p4.pm_p4_escrmsr; /* extract and zero the logical processor selection bits */ cccrtbits = cccrvalue & P4_CCCR_OVF_PMI_T0; escrtbits = escrvalue & (P4_ESCR_T0_OS|P4_ESCR_T0_USR); cccrvalue &= ~P4_CCCR_OVF_PMI_T0; escrvalue &= ~(P4_ESCR_T0_OS|P4_ESCR_T0_USR); if (P4_CPU_IS_HTT_SECONDARY(cpu)) { /* shift T0 bits to T1 position */ cccrtbits <<= 1; escrtbits >>= 2; } /* start system mode PMCs directly */ if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { wrmsr(escrmsr, escrvalue | escrtbits); wrmsr(pd->pm_cccr_msr, cccrvalue | cccrtbits | P4_CCCR_ENABLE); return 0; } /* * Thread mode PMCs * * On HTT machines, the same PMC could be scheduled on the * same physical CPU twice (once for each logical CPU), for * example, if two threads of a multi-threaded process get * scheduled on the same CPU. * */ mtx_lock_spin(&pc->pc_mtx); rc = P4_PCPU_GET_RUNCOUNT(pc,ri); KASSERT(rc == 0 || rc == 1, ("[p4,%d] illegal runcount cpu=%d ri=%d rc=%d", __LINE__, cpu, ri, rc)); if (rc == 0) { /* 1st CPU and the non-HTT case */ KASSERT(P4_PMC_IS_STOPPED(pd->pm_cccr_msr), ("[p4,%d] cpu=%d ri=%d cccr=0x%x not stopped", __LINE__, cpu, ri, pd->pm_cccr_msr)); /* write out the low 40 bits of the saved value to hardware */ wrmsr(pd->pm_pmc_msr, P4_PCPU_PMC_VALUE(pc,ri,cpu) & P4_PERFCTR_MASK); } else if (rc == 1) { /* 2nd CPU */ /* * Stop the PMC and retrieve the CCCR and ESCR values * from their MSRs, and turn on the additional T[0/1] * bits for the 2nd CPU. */ cccrvalue = rdmsr(pd->pm_cccr_msr); wrmsr(pd->pm_cccr_msr, cccrvalue & ~P4_CCCR_ENABLE); /* check that the configuration bits read back match the PMC */ KASSERT((cccrvalue & P4_CCCR_Tx_MASK) == (pm->pm_md.pm_p4.pm_p4_cccrvalue & P4_CCCR_Tx_MASK), ("[p4,%d] Extra CCCR bits cpu=%d rc=%d ri=%d " "cccr=0x%x PMC=0x%x", __LINE__, cpu, rc, ri, cccrvalue & P4_CCCR_Tx_MASK, pm->pm_md.pm_p4.pm_p4_cccrvalue & P4_CCCR_Tx_MASK)); KASSERT(cccrvalue & P4_CCCR_ENABLE, ("[p4,%d] 2nd cpu rc=%d cpu=%d ri=%d not running", __LINE__, rc, cpu, ri)); KASSERT((cccrvalue & cccrtbits) == 0, ("[p4,%d] CCCR T0/T1 mismatch rc=%d cpu=%d ri=%d" "cccrvalue=0x%x tbits=0x%x", __LINE__, rc, cpu, ri, cccrvalue, cccrtbits)); escrvalue = rdmsr(escrmsr); KASSERT((escrvalue & P4_ESCR_Tx_MASK) == (pm->pm_md.pm_p4.pm_p4_escrvalue & P4_ESCR_Tx_MASK), ("[p4,%d] Extra ESCR bits cpu=%d rc=%d ri=%d " "escr=0x%x pm=0x%x", __LINE__, cpu, rc, ri, escrvalue & P4_ESCR_Tx_MASK, pm->pm_md.pm_p4.pm_p4_escrvalue & P4_ESCR_Tx_MASK)); KASSERT((escrvalue & escrtbits) == 0, ("[p4,%d] ESCR T0/T1 mismatch rc=%d cpu=%d ri=%d " "escrmsr=0x%x escrvalue=0x%x tbits=0x%x", __LINE__, rc, cpu, ri, escrmsr, escrvalue, escrtbits)); } /* Enable the correct bits for this CPU. */ escrvalue |= escrtbits; cccrvalue |= cccrtbits | P4_CCCR_ENABLE; /* Save HW value at the time of starting hardware */ P4_PCPU_HW_VALUE(pc,ri,cpu) = rdmsr(pd->pm_pmc_msr); /* Program the ESCR and CCCR and start the PMC */ wrmsr(escrmsr, escrvalue); wrmsr(pd->pm_cccr_msr, cccrvalue); ++rc; P4_PCPU_SET_RUNCOUNT(pc,ri,rc); mtx_unlock_spin(&pc->pc_mtx); PMCDBG(MDP,STA,2,"p4-start cpu=%d rc=%d ri=%d escr=%d " "escrmsr=0x%x escrvalue=0x%x cccr_config=0x%x v=%jx", cpu, rc, ri, pm->pm_md.pm_p4.pm_p4_escr, escrmsr, escrvalue, cccrvalue, P4_PCPU_HW_VALUE(pc,ri,cpu)); return (0); } /* * Stop a PMC. */ static int p4_stop_pmc(int cpu, int ri) { int rc; uint32_t cccrvalue, cccrtbits, escrvalue, escrmsr, escrtbits; struct pmc *pm; struct p4_cpu *pc; struct p4pmc_descr *pd; pmc_value_t tmp; KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[p4,%d] illegal CPU value %d", __LINE__, cpu)); KASSERT(ri >= 0 && ri < P4_NPMCS, ("[p4,%d] illegal row index %d", __LINE__, ri)); pd = &p4_pmcdesc[ri]; pc = p4_pcpu[P4_TO_HTT_PRIMARY(cpu)]; pm = pc->pc_p4pmcs[ri].phw_pmc; KASSERT(pm != NULL, ("[p4,%d] null pmc for cpu%d, ri%d", __LINE__, cpu, ri)); PMCDBG(MDP,STO,1, "p4-stop cpu=%d ri=%d", cpu, ri); if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { wrmsr(pd->pm_cccr_msr, pm->pm_md.pm_p4.pm_p4_cccrvalue & ~P4_CCCR_ENABLE); return (0); } /* * Thread mode PMCs. * * On HTT machines, this PMC may be in use by two threads * running on two logical CPUS. Thus we look at the * 'runcount' field and only turn off the appropriate TO/T1 * bits (and keep the PMC running) if two logical CPUs were * using the PMC. * */ /* bits to mask */ cccrtbits = P4_CCCR_OVF_PMI_T0; escrtbits = P4_ESCR_T0_OS | P4_ESCR_T0_USR; if (P4_CPU_IS_HTT_SECONDARY(cpu)) { cccrtbits <<= 1; escrtbits >>= 2; } mtx_lock_spin(&pc->pc_mtx); rc = P4_PCPU_GET_RUNCOUNT(pc,ri); KASSERT(rc == 2 || rc == 1, ("[p4,%d] illegal runcount cpu=%d ri=%d rc=%d", __LINE__, cpu, ri, rc)); --rc; P4_PCPU_SET_RUNCOUNT(pc,ri,rc); /* Stop this PMC */ cccrvalue = rdmsr(pd->pm_cccr_msr); wrmsr(pd->pm_cccr_msr, cccrvalue & ~P4_CCCR_ENABLE); escrmsr = pm->pm_md.pm_p4.pm_p4_escrmsr; escrvalue = rdmsr(escrmsr); /* The current CPU should be running on this PMC */ KASSERT(escrvalue & escrtbits, ("[p4,%d] ESCR T0/T1 mismatch cpu=%d rc=%d ri=%d escrmsr=0x%x " "escrvalue=0x%x tbits=0x%x", __LINE__, cpu, rc, ri, escrmsr, escrvalue, escrtbits)); KASSERT(PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)) || (cccrvalue & cccrtbits), ("[p4,%d] CCCR T0/T1 mismatch cpu=%d ri=%d cccrvalue=0x%x " "tbits=0x%x", __LINE__, cpu, ri, cccrvalue, cccrtbits)); /* get the current hardware reading */ tmp = rdmsr(pd->pm_pmc_msr); if (rc == 1) { /* need to keep the PMC running */ escrvalue &= ~escrtbits; cccrvalue &= ~cccrtbits; wrmsr(escrmsr, escrvalue); wrmsr(pd->pm_cccr_msr, cccrvalue); } mtx_unlock_spin(&pc->pc_mtx); PMCDBG(MDP,STO,2, "p4-stop cpu=%d rc=%d ri=%d escrmsr=0x%x " "escrval=0x%x cccrval=0x%x v=%jx", cpu, rc, ri, escrmsr, escrvalue, cccrvalue, tmp); if (tmp < P4_PCPU_HW_VALUE(pc,ri,cpu)) /* 40 bit counter overflow */ tmp += (P4_PERFCTR_MASK + 1) - P4_PCPU_HW_VALUE(pc,ri,cpu); else tmp -= P4_PCPU_HW_VALUE(pc,ri,cpu); P4_PCPU_PMC_VALUE(pc,ri,cpu) += tmp; return 0; } /* * Handle an interrupt. * * The hardware sets the CCCR_OVF whenever a counter overflow occurs, * so the handler examines all the 18 CCCR registers, processing the * counters that have overflowed. * * On HTT machines, the CCCR register is shared and will interrupt * both logical processors if so configured. Thus multiple logical * CPUs could enter the NMI service routine at the same time. These * will get serialized using a per-cpu spinlock dedicated for use in * the NMI handler. */ static int p4_intr(int cpu, struct trapframe *tf) { uint32_t cccrval, ovf_mask, ovf_partner; int did_interrupt, error, ri; struct p4_cpu *pc; struct pmc *pm; pmc_value_t v; PMCDBG(MDP,INT, 1, "cpu=%d tf=0x%p um=%d", cpu, (void *) tf, TRAPF_USERMODE(tf)); pc = p4_pcpu[P4_TO_HTT_PRIMARY(cpu)]; ovf_mask = P4_CPU_IS_HTT_SECONDARY(cpu) ? P4_CCCR_OVF_PMI_T1 : P4_CCCR_OVF_PMI_T0; ovf_mask |= P4_CCCR_OVF; if (p4_system_has_htt) ovf_partner = P4_CPU_IS_HTT_SECONDARY(cpu) ? P4_CCCR_OVF_PMI_T0 : P4_CCCR_OVF_PMI_T1; else ovf_partner = 0; did_interrupt = 0; if (p4_system_has_htt) P4_PCPU_ACQ_INTR_SPINLOCK(pc); /* * Loop through all CCCRs, looking for ones that have * interrupted this CPU. */ for (ri = 0; ri < P4_NPMCS; ri++) { /* * Check if our partner logical CPU has already marked * this PMC has having interrupted it. If so, reset * the flag and process the interrupt, but leave the * hardware alone. */ if (p4_system_has_htt && P4_PCPU_GET_INTRFLAG(pc,ri)) { P4_PCPU_SET_INTRFLAG(pc,ri,0); did_interrupt = 1; /* * Ignore de-configured or stopped PMCs. * Ignore PMCs not in sampling mode. */ pm = pc->pc_p4pmcs[ri].phw_pmc; if (pm == NULL || pm->pm_state != PMC_STATE_RUNNING || !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { continue; } (void) pmc_process_interrupt(cpu, PMC_HR, pm, tf, TRAPF_USERMODE(tf)); continue; } /* * Fresh interrupt. Look for the CCCR_OVF bit * and the OVF_Tx bit for this logical * processor being set. */ cccrval = rdmsr(P4_CCCR_MSR_FIRST + ri); if ((cccrval & ovf_mask) != ovf_mask) continue; /* * If the other logical CPU would also have been * interrupted due to the PMC being shared, record * this fact in the per-cpu saved interrupt flag * bitmask. */ if (p4_system_has_htt && (cccrval & ovf_partner)) P4_PCPU_SET_INTRFLAG(pc, ri, 1); v = rdmsr(P4_PERFCTR_MSR_FIRST + ri); PMCDBG(MDP,INT, 2, "ri=%d v=%jx", ri, v); /* Stop the counter, and reset the overflow bit */ cccrval &= ~(P4_CCCR_OVF | P4_CCCR_ENABLE); wrmsr(P4_CCCR_MSR_FIRST + ri, cccrval); did_interrupt = 1; /* * Ignore de-configured or stopped PMCs. Ignore PMCs * not in sampling mode. */ pm = pc->pc_p4pmcs[ri].phw_pmc; if (pm == NULL || pm->pm_state != PMC_STATE_RUNNING || !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { continue; } /* * Process the interrupt. Re-enable the PMC if * processing was successful. */ error = pmc_process_interrupt(cpu, PMC_HR, pm, tf, TRAPF_USERMODE(tf)); /* * Only the first processor executing the NMI handler * in a HTT pair will restart a PMC, and that too * only if there were no errors. */ v = P4_RELOAD_COUNT_TO_PERFCTR_VALUE( pm->pm_sc.pm_reloadcount); wrmsr(P4_PERFCTR_MSR_FIRST + ri, v); if (error == 0) wrmsr(P4_CCCR_MSR_FIRST + ri, cccrval | P4_CCCR_ENABLE); } /* allow the other CPU to proceed */ if (p4_system_has_htt) P4_PCPU_REL_INTR_SPINLOCK(pc); /* * On Intel P4 CPUs, the PMC 'pcint' entry in the LAPIC gets * masked when a PMC interrupts the CPU. We need to unmask * the interrupt source explicitly. */ if (did_interrupt) lapic_reenable_pmc(); atomic_add_int(did_interrupt ? &pmc_stats.pm_intr_processed : &pmc_stats.pm_intr_ignored, 1); return (did_interrupt); } /* * Describe a CPU's PMC state. */ static int p4_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc) { int error; size_t copied; const struct p4pmc_descr *pd; KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[p4,%d] illegal CPU %d", __LINE__, cpu)); KASSERT(ri >= 0 && ri < P4_NPMCS, ("[p4,%d] row-index %d out of range", __LINE__, ri)); PMCDBG(MDP,OPS,1,"p4-describe cpu=%d ri=%d", cpu, ri); if (P4_CPU_IS_HTT_SECONDARY(cpu)) return (EINVAL); pd = &p4_pmcdesc[ri]; if ((error = copystr(pd->pm_descr.pd_name, pi->pm_name, PMC_NAME_MAX, &copied)) != 0) return (error); pi->pm_class = pd->pm_descr.pd_class; if (p4_pcpu[cpu]->pc_p4pmcs[ri].phw_state & PMC_PHW_FLAG_IS_ENABLED) { pi->pm_enabled = TRUE; *ppmc = p4_pcpu[cpu]->pc_p4pmcs[ri].phw_pmc; } else { pi->pm_enabled = FALSE; *ppmc = NULL; } return (0); } /* * Get MSR# for use with RDPMC. */ static int p4_get_msr(int ri, uint32_t *msr) { KASSERT(ri >= 0 && ri < P4_NPMCS, ("[p4,%d] ri %d out of range", __LINE__, ri)); *msr = p4_pmcdesc[ri].pm_pmc_msr - P4_PERFCTR_MSR_FIRST; PMCDBG(MDP,OPS, 1, "ri=%d getmsr=0x%x", ri, *msr); return 0; } int pmc_p4_initialize(struct pmc_mdep *md, int ncpus) { struct pmc_classdep *pcd; struct p4_event_descr *pe; KASSERT(md != NULL, ("[p4,%d] md is NULL", __LINE__)); KASSERT(cpu_vendor_id == CPU_VENDOR_INTEL, ("[p4,%d] Initializing non-intel processor", __LINE__)); PMCDBG(MDP,INI,1, "%s", "p4-initialize"); /* Allocate space for pointers to per-cpu descriptors. */ - p4_pcpu = malloc(sizeof(struct p4_cpu **) * ncpus, M_PMC, - M_ZERO|M_WAITOK); + p4_pcpu = malloc(sizeof(*p4_pcpu) * ncpus, M_PMC, M_ZERO | M_WAITOK); /* Fill in the class dependent descriptor. */ pcd = &md->pmd_classdep[PMC_MDEP_CLASS_INDEX_P4]; switch (md->pmd_cputype) { case PMC_CPU_INTEL_PIV: pcd->pcd_caps = P4_PMC_CAPS; pcd->pcd_class = PMC_CLASS_P4; pcd->pcd_num = P4_NPMCS; pcd->pcd_ri = md->pmd_npmc; pcd->pcd_width = 40; pcd->pcd_allocate_pmc = p4_allocate_pmc; pcd->pcd_config_pmc = p4_config_pmc; pcd->pcd_describe = p4_describe; pcd->pcd_get_config = p4_get_config; pcd->pcd_get_msr = p4_get_msr; pcd->pcd_pcpu_fini = p4_pcpu_fini; pcd->pcd_pcpu_init = p4_pcpu_init; pcd->pcd_read_pmc = p4_read_pmc; pcd->pcd_release_pmc = p4_release_pmc; pcd->pcd_start_pmc = p4_start_pmc; pcd->pcd_stop_pmc = p4_stop_pmc; pcd->pcd_write_pmc = p4_write_pmc; md->pmd_pcpu_fini = NULL; md->pmd_pcpu_init = NULL; md->pmd_intr = p4_intr; md->pmd_npmc += P4_NPMCS; /* model specific configuration */ if ((cpu_id & 0xFFF) < 0xF27) { /* * On P4 and Xeon with CPUID < (Family 15, * Model 2, Stepping 7), only one ESCR is * available for the IOQ_ALLOCATION event. */ pe = p4_find_event(PMC_EV_P4_IOQ_ALLOCATION); pe->pm_escrs[1] = P4_ESCR_NONE; } break; default: KASSERT(0,("[p4,%d] Unknown CPU type", __LINE__)); return ENOSYS; } return (0); } void pmc_p4_finalize(struct pmc_mdep *md) { #if defined(INVARIANTS) int i, ncpus; #endif KASSERT(p4_pcpu != NULL, ("[p4,%d] NULL p4_pcpu", __LINE__)); #if defined(INVARIANTS) ncpus = pmc_cpu_max(); for (i = 0; i < ncpus; i++) KASSERT(p4_pcpu[i] == NULL, ("[p4,%d] non-null pcpu %d", __LINE__, i)); #endif free(p4_pcpu, M_PMC); p4_pcpu = NULL; } Index: stable/9/sys/dev =================================================================== --- stable/9/sys/dev (revision 263357) +++ stable/9/sys/dev (revision 263358) Property changes on: stable/9/sys/dev ___________________________________________________________________ Modified: svn:mergeinfo ## -0,0 +0,1 ## Merged /head/sys/dev:r263080 Index: stable/9/sys/kern/kern_linker.c =================================================================== --- stable/9/sys/kern/kern_linker.c (revision 263357) +++ stable/9/sys/kern/kern_linker.c (revision 263358) @@ -1,2166 +1,2163 @@ /*- * Copyright (c) 1997-2000 Doug Rabson * 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 THE AUTHOR AND CONTRIBUTORS ``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 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. */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_hwpmc_hooks.h" #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 "linker_if.h" #ifdef HWPMC_HOOKS #include #endif #ifdef KLD_DEBUG int kld_debug = 0; SYSCTL_INT(_debug, OID_AUTO, kld_debug, CTLFLAG_RW, &kld_debug, 0, "Set various levels of KLD debug"); #endif /* * static char *linker_search_path(const char *name, struct mod_depend * *verinfo); */ static const char *linker_basename(const char *path); /* * Find a currently loaded file given its filename. */ static linker_file_t linker_find_file_by_name(const char* _filename); /* * Find a currently loaded file given its file id. */ static linker_file_t linker_find_file_by_id(int _fileid); /* Metadata from the static kernel */ SET_DECLARE(modmetadata_set, struct mod_metadata); MALLOC_DEFINE(M_LINKER, "linker", "kernel linker"); linker_file_t linker_kernel_file; static struct sx kld_sx; /* kernel linker lock */ /* * Load counter used by clients to determine if a linker file has been * re-loaded. This counter is incremented for each file load. */ static int loadcnt; static linker_class_list_t classes; static linker_file_list_t linker_files; static int next_file_id = 1; static int linker_no_more_classes = 0; #define LINKER_GET_NEXT_FILE_ID(a) do { \ linker_file_t lftmp; \ \ if (!cold) \ sx_assert(&kld_sx, SA_XLOCKED); \ retry: \ TAILQ_FOREACH(lftmp, &linker_files, link) { \ if (next_file_id == lftmp->id) { \ next_file_id++; \ goto retry; \ } \ } \ (a) = next_file_id; \ } while(0) /* XXX wrong name; we're looking at version provision tags here, not modules */ typedef TAILQ_HEAD(, modlist) modlisthead_t; struct modlist { TAILQ_ENTRY(modlist) link; /* chain together all modules */ linker_file_t container; const char *name; int version; }; typedef struct modlist *modlist_t; static modlisthead_t found_modules; static int linker_file_add_dependency(linker_file_t file, linker_file_t dep); static caddr_t linker_file_lookup_symbol_internal(linker_file_t file, const char* name, int deps); static int linker_load_module(const char *kldname, const char *modname, struct linker_file *parent, struct mod_depend *verinfo, struct linker_file **lfpp); static modlist_t modlist_lookup2(const char *name, struct mod_depend *verinfo); static char * linker_strdup(const char *str) { char *result; if ((result = malloc((strlen(str) + 1), M_LINKER, M_WAITOK)) != NULL) strcpy(result, str); return (result); } static void linker_init(void *arg) { sx_init(&kld_sx, "kernel linker"); TAILQ_INIT(&classes); TAILQ_INIT(&linker_files); } SYSINIT(linker, SI_SUB_KLD, SI_ORDER_FIRST, linker_init, 0); static void linker_stop_class_add(void *arg) { linker_no_more_classes = 1; } SYSINIT(linker_class, SI_SUB_KLD, SI_ORDER_ANY, linker_stop_class_add, NULL); int linker_add_class(linker_class_t lc) { /* * We disallow any class registration past SI_ORDER_ANY * of SI_SUB_KLD. We bump the reference count to keep the * ops from being freed. */ if (linker_no_more_classes == 1) return (EPERM); kobj_class_compile((kobj_class_t) lc); ((kobj_class_t)lc)->refs++; /* XXX: kobj_mtx */ TAILQ_INSERT_TAIL(&classes, lc, link); return (0); } static void linker_file_sysinit(linker_file_t lf) { struct sysinit **start, **stop, **sipp, **xipp, *save; KLD_DPF(FILE, ("linker_file_sysinit: calling SYSINITs for %s\n", lf->filename)); sx_assert(&kld_sx, SA_XLOCKED); if (linker_file_lookup_set(lf, "sysinit_set", &start, &stop, NULL) != 0) return; /* * Perform a bubble sort of the system initialization objects by * their subsystem (primary key) and order (secondary key). * * Since some things care about execution order, this is the operation * which ensures continued function. */ for (sipp = start; sipp < stop; sipp++) { for (xipp = sipp + 1; xipp < stop; xipp++) { if ((*sipp)->subsystem < (*xipp)->subsystem || ((*sipp)->subsystem == (*xipp)->subsystem && (*sipp)->order <= (*xipp)->order)) continue; /* skip */ save = *sipp; *sipp = *xipp; *xipp = save; } } /* * Traverse the (now) ordered list of system initialization tasks. * Perform each task, and continue on to the next task. */ sx_xunlock(&kld_sx); mtx_lock(&Giant); for (sipp = start; sipp < stop; sipp++) { if ((*sipp)->subsystem == SI_SUB_DUMMY) continue; /* skip dummy task(s) */ /* Call function */ (*((*sipp)->func)) ((*sipp)->udata); } mtx_unlock(&Giant); sx_xlock(&kld_sx); } static void linker_file_sysuninit(linker_file_t lf) { struct sysinit **start, **stop, **sipp, **xipp, *save; KLD_DPF(FILE, ("linker_file_sysuninit: calling SYSUNINITs for %s\n", lf->filename)); sx_assert(&kld_sx, SA_XLOCKED); if (linker_file_lookup_set(lf, "sysuninit_set", &start, &stop, NULL) != 0) return; /* * Perform a reverse bubble sort of the system initialization objects * by their subsystem (primary key) and order (secondary key). * * Since some things care about execution order, this is the operation * which ensures continued function. */ for (sipp = start; sipp < stop; sipp++) { for (xipp = sipp + 1; xipp < stop; xipp++) { if ((*sipp)->subsystem > (*xipp)->subsystem || ((*sipp)->subsystem == (*xipp)->subsystem && (*sipp)->order >= (*xipp)->order)) continue; /* skip */ save = *sipp; *sipp = *xipp; *xipp = save; } } /* * Traverse the (now) ordered list of system initialization tasks. * Perform each task, and continue on to the next task. */ sx_xunlock(&kld_sx); mtx_lock(&Giant); for (sipp = start; sipp < stop; sipp++) { if ((*sipp)->subsystem == SI_SUB_DUMMY) continue; /* skip dummy task(s) */ /* Call function */ (*((*sipp)->func)) ((*sipp)->udata); } mtx_unlock(&Giant); sx_xlock(&kld_sx); } static void linker_file_register_sysctls(linker_file_t lf) { struct sysctl_oid **start, **stop, **oidp; KLD_DPF(FILE, ("linker_file_register_sysctls: registering SYSCTLs for %s\n", lf->filename)); sx_assert(&kld_sx, SA_XLOCKED); if (linker_file_lookup_set(lf, "sysctl_set", &start, &stop, NULL) != 0) return; sx_xunlock(&kld_sx); sysctl_lock(); for (oidp = start; oidp < stop; oidp++) sysctl_register_oid(*oidp); sysctl_unlock(); sx_xlock(&kld_sx); } static void linker_file_unregister_sysctls(linker_file_t lf) { struct sysctl_oid **start, **stop, **oidp; KLD_DPF(FILE, ("linker_file_unregister_sysctls: registering SYSCTLs" " for %s\n", lf->filename)); sx_assert(&kld_sx, SA_XLOCKED); if (linker_file_lookup_set(lf, "sysctl_set", &start, &stop, NULL) != 0) return; sx_xunlock(&kld_sx); sysctl_lock(); for (oidp = start; oidp < stop; oidp++) sysctl_unregister_oid(*oidp); sysctl_unlock(); sx_xlock(&kld_sx); } static int linker_file_register_modules(linker_file_t lf) { struct mod_metadata **start, **stop, **mdp; const moduledata_t *moddata; int first_error, error; KLD_DPF(FILE, ("linker_file_register_modules: registering modules" " in %s\n", lf->filename)); sx_assert(&kld_sx, SA_XLOCKED); if (linker_file_lookup_set(lf, "modmetadata_set", &start, &stop, NULL) != 0) { /* * This fallback should be unnecessary, but if we get booted * from boot2 instead of loader and we are missing our * metadata then we have to try the best we can. */ if (lf == linker_kernel_file) { start = SET_BEGIN(modmetadata_set); stop = SET_LIMIT(modmetadata_set); } else return (0); } first_error = 0; for (mdp = start; mdp < stop; mdp++) { if ((*mdp)->md_type != MDT_MODULE) continue; moddata = (*mdp)->md_data; KLD_DPF(FILE, ("Registering module %s in %s\n", moddata->name, lf->filename)); error = module_register(moddata, lf); if (error) { printf("Module %s failed to register: %d\n", moddata->name, error); if (first_error == 0) first_error = error; } } return (first_error); } static void linker_init_kernel_modules(void) { sx_xlock(&kld_sx); linker_file_register_modules(linker_kernel_file); sx_xunlock(&kld_sx); } SYSINIT(linker_kernel, SI_SUB_KLD, SI_ORDER_ANY, linker_init_kernel_modules, 0); static int linker_load_file(const char *filename, linker_file_t *result) { linker_class_t lc; linker_file_t lf; int foundfile, error, modules; /* Refuse to load modules if securelevel raised */ if (prison0.pr_securelevel > 0) return (EPERM); sx_assert(&kld_sx, SA_XLOCKED); lf = linker_find_file_by_name(filename); if (lf) { KLD_DPF(FILE, ("linker_load_file: file %s is already loaded," " incrementing refs\n", filename)); *result = lf; lf->refs++; return (0); } foundfile = 0; error = 0; /* * We do not need to protect (lock) classes here because there is * no class registration past startup (SI_SUB_KLD, SI_ORDER_ANY) * and there is no class deregistration mechanism at this time. */ TAILQ_FOREACH(lc, &classes, link) { KLD_DPF(FILE, ("linker_load_file: trying to load %s\n", filename)); error = LINKER_LOAD_FILE(lc, filename, &lf); /* * If we got something other than ENOENT, then it exists but * we cannot load it for some other reason. */ if (error != ENOENT) foundfile = 1; if (lf) { error = linker_file_register_modules(lf); if (error == EEXIST) { linker_file_unload(lf, LINKER_UNLOAD_FORCE); return (error); } modules = !TAILQ_EMPTY(&lf->modules); linker_file_register_sysctls(lf); linker_file_sysinit(lf); lf->flags |= LINKER_FILE_LINKED; /* * If all of the modules in this file failed * to load, unload the file and return an * error of ENOEXEC. */ if (modules && TAILQ_EMPTY(&lf->modules)) { linker_file_unload(lf, LINKER_UNLOAD_FORCE); return (ENOEXEC); } EVENTHANDLER_INVOKE(kld_load, lf); *result = lf; return (0); } } /* * Less than ideal, but tells the user whether it failed to load or * the module was not found. */ if (foundfile) { /* * If the file type has not been recognized by the last try * printout a message before to fail. */ if (error == ENOSYS) printf("linker_load_file: Unsupported file type\n"); /* * Format not recognized or otherwise unloadable. * When loading a module that is statically built into * the kernel EEXIST percolates back up as the return * value. Preserve this so that apps like sysinstall * can recognize this special case and not post bogus * dialog boxes. */ if (error != EEXIST) error = ENOEXEC; } else error = ENOENT; /* Nothing found */ return (error); } int linker_reference_module(const char *modname, struct mod_depend *verinfo, linker_file_t *result) { modlist_t mod; int error; sx_xlock(&kld_sx); if ((mod = modlist_lookup2(modname, verinfo)) != NULL) { *result = mod->container; (*result)->refs++; sx_xunlock(&kld_sx); return (0); } error = linker_load_module(NULL, modname, NULL, verinfo, result); sx_xunlock(&kld_sx); return (error); } int linker_release_module(const char *modname, struct mod_depend *verinfo, linker_file_t lf) { modlist_t mod; int error; sx_xlock(&kld_sx); if (lf == NULL) { KASSERT(modname != NULL, ("linker_release_module: no file or name")); mod = modlist_lookup2(modname, verinfo); if (mod == NULL) { sx_xunlock(&kld_sx); return (ESRCH); } lf = mod->container; } else KASSERT(modname == NULL && verinfo == NULL, ("linker_release_module: both file and name")); error = linker_file_unload(lf, LINKER_UNLOAD_NORMAL); sx_xunlock(&kld_sx); return (error); } static linker_file_t linker_find_file_by_name(const char *filename) { linker_file_t lf; char *koname; koname = malloc(strlen(filename) + 4, M_LINKER, M_WAITOK); sprintf(koname, "%s.ko", filename); sx_assert(&kld_sx, SA_XLOCKED); TAILQ_FOREACH(lf, &linker_files, link) { if (strcmp(lf->filename, koname) == 0) break; if (strcmp(lf->filename, filename) == 0) break; } free(koname, M_LINKER); return (lf); } static linker_file_t linker_find_file_by_id(int fileid) { linker_file_t lf; sx_assert(&kld_sx, SA_XLOCKED); TAILQ_FOREACH(lf, &linker_files, link) if (lf->id == fileid && lf->flags & LINKER_FILE_LINKED) break; return (lf); } int linker_file_foreach(linker_predicate_t *predicate, void *context) { linker_file_t lf; int retval = 0; sx_xlock(&kld_sx); TAILQ_FOREACH(lf, &linker_files, link) { retval = predicate(lf, context); if (retval != 0) break; } sx_xunlock(&kld_sx); return (retval); } linker_file_t linker_make_file(const char *pathname, linker_class_t lc) { linker_file_t lf; const char *filename; if (!cold) sx_assert(&kld_sx, SA_XLOCKED); filename = linker_basename(pathname); KLD_DPF(FILE, ("linker_make_file: new file, filename='%s' for pathname='%s'\n", filename, pathname)); lf = (linker_file_t)kobj_create((kobj_class_t)lc, M_LINKER, M_WAITOK); if (lf == NULL) return (NULL); lf->refs = 1; lf->userrefs = 0; lf->flags = 0; lf->filename = linker_strdup(filename); lf->pathname = linker_strdup(pathname); LINKER_GET_NEXT_FILE_ID(lf->id); lf->ndeps = 0; lf->deps = NULL; lf->loadcnt = ++loadcnt; STAILQ_INIT(&lf->common); TAILQ_INIT(&lf->modules); TAILQ_INSERT_TAIL(&linker_files, lf, link); return (lf); } int linker_file_unload(linker_file_t file, int flags) { module_t mod, next; modlist_t ml, nextml; struct common_symbol *cp; int error, i; /* Refuse to unload modules if securelevel raised. */ if (prison0.pr_securelevel > 0) return (EPERM); sx_assert(&kld_sx, SA_XLOCKED); KLD_DPF(FILE, ("linker_file_unload: lf->refs=%d\n", file->refs)); /* Easy case of just dropping a reference. */ if (file->refs > 1) { file->refs--; return (0); } /* Give eventhandlers a chance to prevent the unload. */ error = 0; EVENTHANDLER_INVOKE(kld_unload_try, file, &error); if (error != 0) return (EBUSY); KLD_DPF(FILE, ("linker_file_unload: file is unloading," " informing modules\n")); /* * Quiesce all the modules to give them a chance to veto the unload. */ MOD_SLOCK; for (mod = TAILQ_FIRST(&file->modules); mod; mod = module_getfnext(mod)) { error = module_quiesce(mod); if (error != 0 && flags != LINKER_UNLOAD_FORCE) { KLD_DPF(FILE, ("linker_file_unload: module %s" " vetoed unload\n", module_getname(mod))); /* * XXX: Do we need to tell all the quiesced modules * that they can resume work now via a new module * event? */ MOD_SUNLOCK; return (error); } } MOD_SUNLOCK; /* * Inform any modules associated with this file that they are * being unloaded. */ MOD_XLOCK; for (mod = TAILQ_FIRST(&file->modules); mod; mod = next) { next = module_getfnext(mod); MOD_XUNLOCK; /* * Give the module a chance to veto the unload. */ if ((error = module_unload(mod)) != 0) { #ifdef KLD_DEBUG MOD_SLOCK; KLD_DPF(FILE, ("linker_file_unload: module %s" " failed unload\n", module_getname(mod))); MOD_SUNLOCK; #endif return (error); } MOD_XLOCK; module_release(mod); } MOD_XUNLOCK; TAILQ_FOREACH_SAFE(ml, &found_modules, link, nextml) { if (ml->container == file) { TAILQ_REMOVE(&found_modules, ml, link); free(ml, M_LINKER); } } /* * Don't try to run SYSUNINITs if we are unloaded due to a * link error. */ if (file->flags & LINKER_FILE_LINKED) { file->flags &= ~LINKER_FILE_LINKED; linker_file_sysuninit(file); linker_file_unregister_sysctls(file); } TAILQ_REMOVE(&linker_files, file, link); if (file->deps) { for (i = 0; i < file->ndeps; i++) linker_file_unload(file->deps[i], flags); free(file->deps, M_LINKER); file->deps = NULL; } while ((cp = STAILQ_FIRST(&file->common)) != NULL) { STAILQ_REMOVE_HEAD(&file->common, link); free(cp, M_LINKER); } LINKER_UNLOAD(file); EVENTHANDLER_INVOKE(kld_unload, file->filename, file->address, file->size); if (file->filename) { free(file->filename, M_LINKER); file->filename = NULL; } if (file->pathname) { free(file->pathname, M_LINKER); file->pathname = NULL; } kobj_delete((kobj_t) file, M_LINKER); return (0); } int linker_ctf_get(linker_file_t file, linker_ctf_t *lc) { return (LINKER_CTF_GET(file, lc)); } static int linker_file_add_dependency(linker_file_t file, linker_file_t dep) { linker_file_t *newdeps; sx_assert(&kld_sx, SA_XLOCKED); - newdeps = malloc((file->ndeps + 1) * sizeof(linker_file_t *), - M_LINKER, M_WAITOK | M_ZERO); - if (newdeps == NULL) - return (ENOMEM); + newdeps = malloc((file->ndeps + 1) * sizeof(*newdeps), M_LINKER, + M_WAITOK | M_ZERO); if (file->deps) { - bcopy(file->deps, newdeps, - file->ndeps * sizeof(linker_file_t *)); + bcopy(file->deps, newdeps, file->ndeps * sizeof(*newdeps)); free(file->deps, M_LINKER); } file->deps = newdeps; file->deps[file->ndeps] = dep; file->ndeps++; KLD_DPF(FILE, ("linker_file_add_dependency:" " adding %s as dependency for %s\n", dep->filename, file->filename)); return (0); } /* * Locate a linker set and its contents. This is a helper function to avoid * linker_if.h exposure elsewhere. Note: firstp and lastp are really void **. * This function is used in this file so we can avoid having lots of (void **) * casts. */ int linker_file_lookup_set(linker_file_t file, const char *name, void *firstp, void *lastp, int *countp) { sx_assert(&kld_sx, SA_LOCKED); return (LINKER_LOOKUP_SET(file, name, firstp, lastp, countp)); } /* * List all functions in a file. */ int linker_file_function_listall(linker_file_t lf, linker_function_nameval_callback_t callback_func, void *arg) { return (LINKER_EACH_FUNCTION_NAMEVAL(lf, callback_func, arg)); } caddr_t linker_file_lookup_symbol(linker_file_t file, const char *name, int deps) { caddr_t sym; int locked; locked = sx_xlocked(&kld_sx); if (!locked) sx_xlock(&kld_sx); sym = linker_file_lookup_symbol_internal(file, name, deps); if (!locked) sx_xunlock(&kld_sx); return (sym); } static caddr_t linker_file_lookup_symbol_internal(linker_file_t file, const char *name, int deps) { c_linker_sym_t sym; linker_symval_t symval; caddr_t address; size_t common_size = 0; int i; sx_assert(&kld_sx, SA_XLOCKED); KLD_DPF(SYM, ("linker_file_lookup_symbol: file=%p, name=%s, deps=%d\n", file, name, deps)); if (LINKER_LOOKUP_SYMBOL(file, name, &sym) == 0) { LINKER_SYMBOL_VALUES(file, sym, &symval); if (symval.value == 0) /* * For commons, first look them up in the * dependencies and only allocate space if not found * there. */ common_size = symval.size; else { KLD_DPF(SYM, ("linker_file_lookup_symbol: symbol" ".value=%p\n", symval.value)); return (symval.value); } } if (deps) { for (i = 0; i < file->ndeps; i++) { address = linker_file_lookup_symbol_internal( file->deps[i], name, 0); if (address) { KLD_DPF(SYM, ("linker_file_lookup_symbol:" " deps value=%p\n", address)); return (address); } } } if (common_size > 0) { /* * This is a common symbol which was not found in the * dependencies. We maintain a simple common symbol table in * the file object. */ struct common_symbol *cp; STAILQ_FOREACH(cp, &file->common, link) { if (strcmp(cp->name, name) == 0) { KLD_DPF(SYM, ("linker_file_lookup_symbol:" " old common value=%p\n", cp->address)); return (cp->address); } } /* * Round the symbol size up to align. */ common_size = (common_size + sizeof(int) - 1) & -sizeof(int); cp = malloc(sizeof(struct common_symbol) + common_size + strlen(name) + 1, M_LINKER, M_WAITOK | M_ZERO); cp->address = (caddr_t)(cp + 1); cp->name = cp->address + common_size; strcpy(cp->name, name); bzero(cp->address, common_size); STAILQ_INSERT_TAIL(&file->common, cp, link); KLD_DPF(SYM, ("linker_file_lookup_symbol: new common" " value=%p\n", cp->address)); return (cp->address); } KLD_DPF(SYM, ("linker_file_lookup_symbol: fail\n")); return (0); } /* * Both DDB and stack(9) rely on the kernel linker to provide forward and * backward lookup of symbols. However, DDB and sometimes stack(9) need to * do this in a lockfree manner. We provide a set of internal helper * routines to perform these operations without locks, and then wrappers that * optionally lock. * * linker_debug_lookup() is ifdef DDB as currently it's only used by DDB. */ #ifdef DDB static int linker_debug_lookup(const char *symstr, c_linker_sym_t *sym) { linker_file_t lf; TAILQ_FOREACH(lf, &linker_files, link) { if (LINKER_LOOKUP_SYMBOL(lf, symstr, sym) == 0) return (0); } return (ENOENT); } #endif static int linker_debug_search_symbol(caddr_t value, c_linker_sym_t *sym, long *diffp) { linker_file_t lf; c_linker_sym_t best, es; u_long diff, bestdiff, off; best = 0; off = (uintptr_t)value; bestdiff = off; TAILQ_FOREACH(lf, &linker_files, link) { if (LINKER_SEARCH_SYMBOL(lf, value, &es, &diff) != 0) continue; if (es != 0 && diff < bestdiff) { best = es; bestdiff = diff; } if (bestdiff == 0) break; } if (best) { *sym = best; *diffp = bestdiff; return (0); } else { *sym = 0; *diffp = off; return (ENOENT); } } static int linker_debug_symbol_values(c_linker_sym_t sym, linker_symval_t *symval) { linker_file_t lf; TAILQ_FOREACH(lf, &linker_files, link) { if (LINKER_SYMBOL_VALUES(lf, sym, symval) == 0) return (0); } return (ENOENT); } static int linker_debug_search_symbol_name(caddr_t value, char *buf, u_int buflen, long *offset) { linker_symval_t symval; c_linker_sym_t sym; int error; *offset = 0; error = linker_debug_search_symbol(value, &sym, offset); if (error) return (error); error = linker_debug_symbol_values(sym, &symval); if (error) return (error); strlcpy(buf, symval.name, buflen); return (0); } /* * DDB Helpers. DDB has to look across multiple files with their own symbol * tables and string tables. * * Note that we do not obey list locking protocols here. We really don't need * DDB to hang because somebody's got the lock held. We'll take the chance * that the files list is inconsistant instead. */ #ifdef DDB int linker_ddb_lookup(const char *symstr, c_linker_sym_t *sym) { return (linker_debug_lookup(symstr, sym)); } #endif int linker_ddb_search_symbol(caddr_t value, c_linker_sym_t *sym, long *diffp) { return (linker_debug_search_symbol(value, sym, diffp)); } int linker_ddb_symbol_values(c_linker_sym_t sym, linker_symval_t *symval) { return (linker_debug_symbol_values(sym, symval)); } int linker_ddb_search_symbol_name(caddr_t value, char *buf, u_int buflen, long *offset) { return (linker_debug_search_symbol_name(value, buf, buflen, offset)); } /* * stack(9) helper for non-debugging environemnts. Unlike DDB helpers, we do * obey locking protocols, and offer a significantly less complex interface. */ int linker_search_symbol_name(caddr_t value, char *buf, u_int buflen, long *offset) { int error; sx_xlock(&kld_sx); error = linker_debug_search_symbol_name(value, buf, buflen, offset); sx_xunlock(&kld_sx); return (error); } /* * Syscalls. */ int kern_kldload(struct thread *td, const char *file, int *fileid) { const char *kldname, *modname; linker_file_t lf; int error; if ((error = securelevel_gt(td->td_ucred, 0)) != 0) return (error); if ((error = priv_check(td, PRIV_KLD_LOAD)) != 0) return (error); /* * It is possible that kldloaded module will attach a new ifnet, * so vnet context must be set when this ocurs. */ CURVNET_SET(TD_TO_VNET(td)); /* * If file does not contain a qualified name or any dot in it * (kldname.ko, or kldname.ver.ko) treat it as an interface * name. */ if (index(file, '/') || index(file, '.')) { kldname = file; modname = NULL; } else { kldname = NULL; modname = file; } sx_xlock(&kld_sx); error = linker_load_module(kldname, modname, NULL, NULL, &lf); if (error) { sx_xunlock(&kld_sx); goto done; } lf->userrefs++; if (fileid != NULL) *fileid = lf->id; sx_xunlock(&kld_sx); done: CURVNET_RESTORE(); return (error); } int sys_kldload(struct thread *td, struct kldload_args *uap) { char *pathname = NULL; int error, fileid; td->td_retval[0] = -1; pathname = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); error = copyinstr(uap->file, pathname, MAXPATHLEN, NULL); if (error == 0) { error = kern_kldload(td, pathname, &fileid); if (error == 0) td->td_retval[0] = fileid; } free(pathname, M_TEMP); return (error); } int kern_kldunload(struct thread *td, int fileid, int flags) { linker_file_t lf; int error = 0; if ((error = securelevel_gt(td->td_ucred, 0)) != 0) return (error); if ((error = priv_check(td, PRIV_KLD_UNLOAD)) != 0) return (error); CURVNET_SET(TD_TO_VNET(td)); sx_xlock(&kld_sx); lf = linker_find_file_by_id(fileid); if (lf) { KLD_DPF(FILE, ("kldunload: lf->userrefs=%d\n", lf->userrefs)); if (lf->userrefs == 0) { /* * XXX: maybe LINKER_UNLOAD_FORCE should override ? */ printf("kldunload: attempt to unload file that was" " loaded by the kernel\n"); error = EBUSY; } else { lf->userrefs--; error = linker_file_unload(lf, flags); if (error) lf->userrefs++; } } else error = ENOENT; sx_xunlock(&kld_sx); CURVNET_RESTORE(); return (error); } int sys_kldunload(struct thread *td, struct kldunload_args *uap) { return (kern_kldunload(td, uap->fileid, LINKER_UNLOAD_NORMAL)); } int sys_kldunloadf(struct thread *td, struct kldunloadf_args *uap) { if (uap->flags != LINKER_UNLOAD_NORMAL && uap->flags != LINKER_UNLOAD_FORCE) return (EINVAL); return (kern_kldunload(td, uap->fileid, uap->flags)); } int sys_kldfind(struct thread *td, struct kldfind_args *uap) { char *pathname; const char *filename; linker_file_t lf; int error; #ifdef MAC error = mac_kld_check_stat(td->td_ucred); if (error) return (error); #endif td->td_retval[0] = -1; pathname = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); if ((error = copyinstr(uap->file, pathname, MAXPATHLEN, NULL)) != 0) goto out; filename = linker_basename(pathname); sx_xlock(&kld_sx); lf = linker_find_file_by_name(filename); if (lf) td->td_retval[0] = lf->id; else error = ENOENT; sx_xunlock(&kld_sx); out: free(pathname, M_TEMP); return (error); } int sys_kldnext(struct thread *td, struct kldnext_args *uap) { linker_file_t lf; int error = 0; #ifdef MAC error = mac_kld_check_stat(td->td_ucred); if (error) return (error); #endif sx_xlock(&kld_sx); if (uap->fileid == 0) lf = TAILQ_FIRST(&linker_files); else { lf = linker_find_file_by_id(uap->fileid); if (lf == NULL) { error = ENOENT; goto out; } lf = TAILQ_NEXT(lf, link); } /* Skip partially loaded files. */ while (lf != NULL && !(lf->flags & LINKER_FILE_LINKED)) lf = TAILQ_NEXT(lf, link); if (lf) td->td_retval[0] = lf->id; else td->td_retval[0] = 0; out: sx_xunlock(&kld_sx); return (error); } int sys_kldstat(struct thread *td, struct kldstat_args *uap) { struct kld_file_stat stat; int error, version; /* * Check the version of the user's structure. */ if ((error = copyin(&uap->stat->version, &version, sizeof(version))) != 0) return (error); if (version != sizeof(struct kld_file_stat_1) && version != sizeof(struct kld_file_stat)) return (EINVAL); error = kern_kldstat(td, uap->fileid, &stat); if (error != 0) return (error); return (copyout(&stat, uap->stat, version)); } int kern_kldstat(struct thread *td, int fileid, struct kld_file_stat *stat) { linker_file_t lf; int namelen; #ifdef MAC int error; error = mac_kld_check_stat(td->td_ucred); if (error) return (error); #endif sx_xlock(&kld_sx); lf = linker_find_file_by_id(fileid); if (lf == NULL) { sx_xunlock(&kld_sx); return (ENOENT); } /* Version 1 fields: */ namelen = strlen(lf->filename) + 1; if (namelen > MAXPATHLEN) namelen = MAXPATHLEN; bcopy(lf->filename, &stat->name[0], namelen); stat->refs = lf->refs; stat->id = lf->id; stat->address = lf->address; stat->size = lf->size; /* Version 2 fields: */ namelen = strlen(lf->pathname) + 1; if (namelen > MAXPATHLEN) namelen = MAXPATHLEN; bcopy(lf->pathname, &stat->pathname[0], namelen); sx_xunlock(&kld_sx); td->td_retval[0] = 0; return (0); } int sys_kldfirstmod(struct thread *td, struct kldfirstmod_args *uap) { linker_file_t lf; module_t mp; int error = 0; #ifdef MAC error = mac_kld_check_stat(td->td_ucred); if (error) return (error); #endif sx_xlock(&kld_sx); lf = linker_find_file_by_id(uap->fileid); if (lf) { MOD_SLOCK; mp = TAILQ_FIRST(&lf->modules); if (mp != NULL) td->td_retval[0] = module_getid(mp); else td->td_retval[0] = 0; MOD_SUNLOCK; } else error = ENOENT; sx_xunlock(&kld_sx); return (error); } int sys_kldsym(struct thread *td, struct kldsym_args *uap) { char *symstr = NULL; c_linker_sym_t sym; linker_symval_t symval; linker_file_t lf; struct kld_sym_lookup lookup; int error = 0; #ifdef MAC error = mac_kld_check_stat(td->td_ucred); if (error) return (error); #endif if ((error = copyin(uap->data, &lookup, sizeof(lookup))) != 0) return (error); if (lookup.version != sizeof(lookup) || uap->cmd != KLDSYM_LOOKUP) return (EINVAL); symstr = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); if ((error = copyinstr(lookup.symname, symstr, MAXPATHLEN, NULL)) != 0) goto out; sx_xlock(&kld_sx); if (uap->fileid != 0) { lf = linker_find_file_by_id(uap->fileid); if (lf == NULL) error = ENOENT; else if (LINKER_LOOKUP_SYMBOL(lf, symstr, &sym) == 0 && LINKER_SYMBOL_VALUES(lf, sym, &symval) == 0) { lookup.symvalue = (uintptr_t) symval.value; lookup.symsize = symval.size; error = copyout(&lookup, uap->data, sizeof(lookup)); } else error = ENOENT; } else { TAILQ_FOREACH(lf, &linker_files, link) { if (LINKER_LOOKUP_SYMBOL(lf, symstr, &sym) == 0 && LINKER_SYMBOL_VALUES(lf, sym, &symval) == 0) { lookup.symvalue = (uintptr_t)symval.value; lookup.symsize = symval.size; error = copyout(&lookup, uap->data, sizeof(lookup)); break; } } if (lf == NULL) error = ENOENT; } sx_xunlock(&kld_sx); out: free(symstr, M_TEMP); return (error); } /* * Preloaded module support */ static modlist_t modlist_lookup(const char *name, int ver) { modlist_t mod; TAILQ_FOREACH(mod, &found_modules, link) { if (strcmp(mod->name, name) == 0 && (ver == 0 || mod->version == ver)) return (mod); } return (NULL); } static modlist_t modlist_lookup2(const char *name, struct mod_depend *verinfo) { modlist_t mod, bestmod; int ver; if (verinfo == NULL) return (modlist_lookup(name, 0)); bestmod = NULL; TAILQ_FOREACH(mod, &found_modules, link) { if (strcmp(mod->name, name) != 0) continue; ver = mod->version; if (ver == verinfo->md_ver_preferred) return (mod); if (ver >= verinfo->md_ver_minimum && ver <= verinfo->md_ver_maximum && (bestmod == NULL || ver > bestmod->version)) bestmod = mod; } return (bestmod); } static modlist_t modlist_newmodule(const char *modname, int version, linker_file_t container) { modlist_t mod; mod = malloc(sizeof(struct modlist), M_LINKER, M_NOWAIT | M_ZERO); if (mod == NULL) panic("no memory for module list"); mod->container = container; mod->name = modname; mod->version = version; TAILQ_INSERT_TAIL(&found_modules, mod, link); return (mod); } static void linker_addmodules(linker_file_t lf, struct mod_metadata **start, struct mod_metadata **stop, int preload) { struct mod_metadata *mp, **mdp; const char *modname; int ver; for (mdp = start; mdp < stop; mdp++) { mp = *mdp; if (mp->md_type != MDT_VERSION) continue; modname = mp->md_cval; ver = ((struct mod_version *)mp->md_data)->mv_version; if (modlist_lookup(modname, ver) != NULL) { printf("module %s already present!\n", modname); /* XXX what can we do? this is a build error. :-( */ continue; } modlist_newmodule(modname, ver, lf); } } static void linker_preload(void *arg) { caddr_t modptr; const char *modname, *nmodname; char *modtype; linker_file_t lf, nlf; linker_class_t lc; int error; linker_file_list_t loaded_files; linker_file_list_t depended_files; struct mod_metadata *mp, *nmp; struct mod_metadata **start, **stop, **mdp, **nmdp; struct mod_depend *verinfo; int nver; int resolves; modlist_t mod; struct sysinit **si_start, **si_stop; TAILQ_INIT(&loaded_files); TAILQ_INIT(&depended_files); TAILQ_INIT(&found_modules); error = 0; modptr = NULL; sx_xlock(&kld_sx); while ((modptr = preload_search_next_name(modptr)) != NULL) { modname = (char *)preload_search_info(modptr, MODINFO_NAME); modtype = (char *)preload_search_info(modptr, MODINFO_TYPE); if (modname == NULL) { printf("Preloaded module at %p does not have a" " name!\n", modptr); continue; } if (modtype == NULL) { printf("Preloaded module at %p does not have a type!\n", modptr); continue; } if (bootverbose) printf("Preloaded %s \"%s\" at %p.\n", modtype, modname, modptr); lf = NULL; TAILQ_FOREACH(lc, &classes, link) { error = LINKER_LINK_PRELOAD(lc, modname, &lf); if (!error) break; lf = NULL; } if (lf) TAILQ_INSERT_TAIL(&loaded_files, lf, loaded); } /* * First get a list of stuff in the kernel. */ if (linker_file_lookup_set(linker_kernel_file, MDT_SETNAME, &start, &stop, NULL) == 0) linker_addmodules(linker_kernel_file, start, stop, 1); /* * This is a once-off kinky bubble sort to resolve relocation * dependency requirements. */ restart: TAILQ_FOREACH(lf, &loaded_files, loaded) { error = linker_file_lookup_set(lf, MDT_SETNAME, &start, &stop, NULL); /* * First, look to see if we would successfully link with this * stuff. */ resolves = 1; /* unless we know otherwise */ if (!error) { for (mdp = start; mdp < stop; mdp++) { mp = *mdp; if (mp->md_type != MDT_DEPEND) continue; modname = mp->md_cval; verinfo = mp->md_data; for (nmdp = start; nmdp < stop; nmdp++) { nmp = *nmdp; if (nmp->md_type != MDT_VERSION) continue; nmodname = nmp->md_cval; if (strcmp(modname, nmodname) == 0) break; } if (nmdp < stop) /* it's a self reference */ continue; /* * ok, the module isn't here yet, we * are not finished */ if (modlist_lookup2(modname, verinfo) == NULL) resolves = 0; } } /* * OK, if we found our modules, we can link. So, "provide" * the modules inside and add it to the end of the link order * list. */ if (resolves) { if (!error) { for (mdp = start; mdp < stop; mdp++) { mp = *mdp; if (mp->md_type != MDT_VERSION) continue; modname = mp->md_cval; nver = ((struct mod_version *) mp->md_data)->mv_version; if (modlist_lookup(modname, nver) != NULL) { printf("module %s already" " present!\n", modname); TAILQ_REMOVE(&loaded_files, lf, loaded); linker_file_unload(lf, LINKER_UNLOAD_FORCE); /* we changed tailq next ptr */ goto restart; } modlist_newmodule(modname, nver, lf); } } TAILQ_REMOVE(&loaded_files, lf, loaded); TAILQ_INSERT_TAIL(&depended_files, lf, loaded); /* * Since we provided modules, we need to restart the * sort so that the previous files that depend on us * have a chance. Also, we've busted the tailq next * pointer with the REMOVE. */ goto restart; } } /* * At this point, we check to see what could not be resolved.. */ while ((lf = TAILQ_FIRST(&loaded_files)) != NULL) { TAILQ_REMOVE(&loaded_files, lf, loaded); printf("KLD file %s is missing dependencies\n", lf->filename); linker_file_unload(lf, LINKER_UNLOAD_FORCE); } /* * We made it. Finish off the linking in the order we determined. */ TAILQ_FOREACH_SAFE(lf, &depended_files, loaded, nlf) { if (linker_kernel_file) { linker_kernel_file->refs++; error = linker_file_add_dependency(lf, linker_kernel_file); if (error) panic("cannot add dependency"); } lf->userrefs++; /* so we can (try to) kldunload it */ error = linker_file_lookup_set(lf, MDT_SETNAME, &start, &stop, NULL); if (!error) { for (mdp = start; mdp < stop; mdp++) { mp = *mdp; if (mp->md_type != MDT_DEPEND) continue; modname = mp->md_cval; verinfo = mp->md_data; mod = modlist_lookup2(modname, verinfo); if (mod == NULL) { printf("KLD file %s - cannot find " "dependency \"%s\"\n", lf->filename, modname); goto fail; } /* Don't count self-dependencies */ if (lf == mod->container) continue; mod->container->refs++; error = linker_file_add_dependency(lf, mod->container); if (error) panic("cannot add dependency"); } } /* * Now do relocation etc using the symbol search paths * established by the dependencies */ error = LINKER_LINK_PRELOAD_FINISH(lf); if (error) { printf("KLD file %s - could not finalize loading\n", lf->filename); goto fail; } linker_file_register_modules(lf); if (linker_file_lookup_set(lf, "sysinit_set", &si_start, &si_stop, NULL) == 0) sysinit_add(si_start, si_stop); linker_file_register_sysctls(lf); lf->flags |= LINKER_FILE_LINKED; continue; fail: TAILQ_REMOVE(&depended_files, lf, loaded); linker_file_unload(lf, LINKER_UNLOAD_FORCE); } sx_xunlock(&kld_sx); /* woohoo! we made it! */ } SYSINIT(preload, SI_SUB_KLD, SI_ORDER_MIDDLE, linker_preload, 0); /* * Search for a not-loaded module by name. * * Modules may be found in the following locations: * * - preloaded (result is just the module name) - on disk (result is full path * to module) * * If the module name is qualified in any way (contains path, etc.) the we * simply return a copy of it. * * The search path can be manipulated via sysctl. Note that we use the ';' * character as a separator to be consistent with the bootloader. */ static char linker_hintfile[] = "linker.hints"; static char linker_path[MAXPATHLEN] = "/boot/kernel;/boot/modules"; SYSCTL_STRING(_kern, OID_AUTO, module_path, CTLFLAG_RW, linker_path, sizeof(linker_path), "module load search path"); TUNABLE_STR("module_path", linker_path, sizeof(linker_path)); static char *linker_ext_list[] = { "", ".ko", NULL }; /* * Check if file actually exists either with or without extension listed in * the linker_ext_list. (probably should be generic for the rest of the * kernel) */ static char * linker_lookup_file(const char *path, int pathlen, const char *name, int namelen, struct vattr *vap) { struct nameidata nd; struct thread *td = curthread; /* XXX */ char *result, **cpp, *sep; int error, len, extlen, reclen, flags, vfslocked; enum vtype type; extlen = 0; for (cpp = linker_ext_list; *cpp; cpp++) { len = strlen(*cpp); if (len > extlen) extlen = len; } extlen++; /* trailing '\0' */ sep = (path[pathlen - 1] != '/') ? "/" : ""; reclen = pathlen + strlen(sep) + namelen + extlen + 1; result = malloc(reclen, M_LINKER, M_WAITOK); for (cpp = linker_ext_list; *cpp; cpp++) { snprintf(result, reclen, "%.*s%s%.*s%s", pathlen, path, sep, namelen, name, *cpp); /* * Attempt to open the file, and return the path if * we succeed and it's a regular file. */ NDINIT(&nd, LOOKUP, FOLLOW | MPSAFE, UIO_SYSSPACE, result, td); flags = FREAD; error = vn_open(&nd, &flags, 0, NULL); if (error == 0) { vfslocked = NDHASGIANT(&nd); NDFREE(&nd, NDF_ONLY_PNBUF); type = nd.ni_vp->v_type; if (vap) VOP_GETATTR(nd.ni_vp, vap, td->td_ucred); VOP_UNLOCK(nd.ni_vp, 0); vn_close(nd.ni_vp, FREAD, td->td_ucred, td); VFS_UNLOCK_GIANT(vfslocked); if (type == VREG) return (result); } } free(result, M_LINKER); return (NULL); } #define INT_ALIGN(base, ptr) ptr = \ (base) + (((ptr) - (base) + sizeof(int) - 1) & ~(sizeof(int) - 1)) /* * Lookup KLD which contains requested module in the "linker.hints" file. If * version specification is available, then try to find the best KLD. * Otherwise just find the latest one. */ static char * linker_hints_lookup(const char *path, int pathlen, const char *modname, int modnamelen, struct mod_depend *verinfo) { struct thread *td = curthread; /* XXX */ struct ucred *cred = td ? td->td_ucred : NULL; struct nameidata nd; struct vattr vattr, mattr; u_char *hints = NULL; u_char *cp, *recptr, *bufend, *result, *best, *pathbuf, *sep; int error, ival, bestver, *intp, found, flags, clen, blen; ssize_t reclen; int vfslocked = 0; result = NULL; bestver = found = 0; sep = (path[pathlen - 1] != '/') ? "/" : ""; reclen = imax(modnamelen, strlen(linker_hintfile)) + pathlen + strlen(sep) + 1; pathbuf = malloc(reclen, M_LINKER, M_WAITOK); snprintf(pathbuf, reclen, "%.*s%s%s", pathlen, path, sep, linker_hintfile); NDINIT(&nd, LOOKUP, NOFOLLOW | MPSAFE, UIO_SYSSPACE, pathbuf, td); flags = FREAD; error = vn_open(&nd, &flags, 0, NULL); if (error) goto bad; vfslocked = NDHASGIANT(&nd); NDFREE(&nd, NDF_ONLY_PNBUF); if (nd.ni_vp->v_type != VREG) goto bad; best = cp = NULL; error = VOP_GETATTR(nd.ni_vp, &vattr, cred); if (error) goto bad; /* * XXX: we need to limit this number to some reasonable value */ if (vattr.va_size > 100 * 1024) { printf("hints file too large %ld\n", (long)vattr.va_size); goto bad; } hints = malloc(vattr.va_size, M_TEMP, M_WAITOK); if (hints == NULL) goto bad; error = vn_rdwr(UIO_READ, nd.ni_vp, (caddr_t)hints, vattr.va_size, 0, UIO_SYSSPACE, IO_NODELOCKED, cred, NOCRED, &reclen, td); if (error) goto bad; VOP_UNLOCK(nd.ni_vp, 0); vn_close(nd.ni_vp, FREAD, cred, td); VFS_UNLOCK_GIANT(vfslocked); nd.ni_vp = NULL; if (reclen != 0) { printf("can't read %zd\n", reclen); goto bad; } intp = (int *)hints; ival = *intp++; if (ival != LINKER_HINTS_VERSION) { printf("hints file version mismatch %d\n", ival); goto bad; } bufend = hints + vattr.va_size; recptr = (u_char *)intp; clen = blen = 0; while (recptr < bufend && !found) { intp = (int *)recptr; reclen = *intp++; ival = *intp++; cp = (char *)intp; switch (ival) { case MDT_VERSION: clen = *cp++; if (clen != modnamelen || bcmp(cp, modname, clen) != 0) break; cp += clen; INT_ALIGN(hints, cp); ival = *(int *)cp; cp += sizeof(int); clen = *cp++; if (verinfo == NULL || ival == verinfo->md_ver_preferred) { found = 1; break; } if (ival >= verinfo->md_ver_minimum && ival <= verinfo->md_ver_maximum && ival > bestver) { bestver = ival; best = cp; blen = clen; } break; default: break; } recptr += reclen + sizeof(int); } /* * Finally check if KLD is in the place */ if (found) result = linker_lookup_file(path, pathlen, cp, clen, &mattr); else if (best) result = linker_lookup_file(path, pathlen, best, blen, &mattr); /* * KLD is newer than hints file. What we should do now? */ if (result && timespeccmp(&mattr.va_mtime, &vattr.va_mtime, >)) printf("warning: KLD '%s' is newer than the linker.hints" " file\n", result); bad: free(pathbuf, M_LINKER); if (hints) free(hints, M_TEMP); if (nd.ni_vp != NULL) { VOP_UNLOCK(nd.ni_vp, 0); vn_close(nd.ni_vp, FREAD, cred, td); VFS_UNLOCK_GIANT(vfslocked); } /* * If nothing found or hints is absent - fallback to the old * way by using "kldname[.ko]" as module name. */ if (!found && !bestver && result == NULL) result = linker_lookup_file(path, pathlen, modname, modnamelen, NULL); return (result); } /* * Lookup KLD which contains requested module in the all directories. */ static char * linker_search_module(const char *modname, int modnamelen, struct mod_depend *verinfo) { char *cp, *ep, *result; /* * traverse the linker path */ for (cp = linker_path; *cp; cp = ep + 1) { /* find the end of this component */ for (ep = cp; (*ep != 0) && (*ep != ';'); ep++); result = linker_hints_lookup(cp, ep - cp, modname, modnamelen, verinfo); if (result != NULL) return (result); if (*ep == 0) break; } return (NULL); } /* * Search for module in all directories listed in the linker_path. */ static char * linker_search_kld(const char *name) { char *cp, *ep, *result; int len; /* qualified at all? */ if (index(name, '/')) return (linker_strdup(name)); /* traverse the linker path */ len = strlen(name); for (ep = linker_path; *ep; ep++) { cp = ep; /* find the end of this component */ for (; *ep != 0 && *ep != ';'; ep++); result = linker_lookup_file(cp, ep - cp, name, len, NULL); if (result != NULL) return (result); } return (NULL); } static const char * linker_basename(const char *path) { const char *filename; filename = rindex(path, '/'); if (filename == NULL) return path; if (filename[1]) filename++; return (filename); } #ifdef HWPMC_HOOKS /* * Inform hwpmc about the set of kernel modules currently loaded. */ void * linker_hwpmc_list_objects(void) { linker_file_t lf; struct pmckern_map_in *kobase; int i, nmappings; nmappings = 0; sx_slock(&kld_sx); TAILQ_FOREACH(lf, &linker_files, link) nmappings++; /* Allocate nmappings + 1 entries. */ kobase = malloc((nmappings + 1) * sizeof(struct pmckern_map_in), M_LINKER, M_WAITOK | M_ZERO); i = 0; TAILQ_FOREACH(lf, &linker_files, link) { /* Save the info for this linker file. */ kobase[i].pm_file = lf->filename; kobase[i].pm_address = (uintptr_t)lf->address; i++; } sx_sunlock(&kld_sx); KASSERT(i > 0, ("linker_hpwmc_list_objects: no kernel objects?")); /* The last entry of the malloced area comprises of all zeros. */ KASSERT(kobase[i].pm_file == NULL, ("linker_hwpmc_list_objects: last object not NULL")); return ((void *)kobase); } #endif /* * Find a file which contains given module and load it, if "parent" is not * NULL, register a reference to it. */ static int linker_load_module(const char *kldname, const char *modname, struct linker_file *parent, struct mod_depend *verinfo, struct linker_file **lfpp) { linker_file_t lfdep; const char *filename; char *pathname; int error; sx_assert(&kld_sx, SA_XLOCKED); if (modname == NULL) { /* * We have to load KLD */ KASSERT(verinfo == NULL, ("linker_load_module: verinfo" " is not NULL")); pathname = linker_search_kld(kldname); } else { if (modlist_lookup2(modname, verinfo) != NULL) return (EEXIST); if (kldname != NULL) pathname = linker_strdup(kldname); else if (rootvnode == NULL) pathname = NULL; else /* * Need to find a KLD with required module */ pathname = linker_search_module(modname, strlen(modname), verinfo); } if (pathname == NULL) return (ENOENT); /* * Can't load more than one file with the same basename XXX: * Actually it should be possible to have multiple KLDs with * the same basename but different path because they can * provide different versions of the same modules. */ filename = linker_basename(pathname); if (linker_find_file_by_name(filename)) error = EEXIST; else do { error = linker_load_file(pathname, &lfdep); if (error) break; if (modname && verinfo && modlist_lookup2(modname, verinfo) == NULL) { linker_file_unload(lfdep, LINKER_UNLOAD_FORCE); error = ENOENT; break; } if (parent) { error = linker_file_add_dependency(parent, lfdep); if (error) break; } if (lfpp) *lfpp = lfdep; } while (0); free(pathname, M_LINKER); return (error); } /* * This routine is responsible for finding dependencies of userland initiated * kldload(2)'s of files. */ int linker_load_dependencies(linker_file_t lf) { linker_file_t lfdep; struct mod_metadata **start, **stop, **mdp, **nmdp; struct mod_metadata *mp, *nmp; struct mod_depend *verinfo; modlist_t mod; const char *modname, *nmodname; int ver, error = 0, count; /* * All files are dependant on /kernel. */ sx_assert(&kld_sx, SA_XLOCKED); if (linker_kernel_file) { linker_kernel_file->refs++; error = linker_file_add_dependency(lf, linker_kernel_file); if (error) return (error); } if (linker_file_lookup_set(lf, MDT_SETNAME, &start, &stop, &count) != 0) return (0); for (mdp = start; mdp < stop; mdp++) { mp = *mdp; if (mp->md_type != MDT_VERSION) continue; modname = mp->md_cval; ver = ((struct mod_version *)mp->md_data)->mv_version; mod = modlist_lookup(modname, ver); if (mod != NULL) { printf("interface %s.%d already present in the KLD" " '%s'!\n", modname, ver, mod->container->filename); return (EEXIST); } } for (mdp = start; mdp < stop; mdp++) { mp = *mdp; if (mp->md_type != MDT_DEPEND) continue; modname = mp->md_cval; verinfo = mp->md_data; nmodname = NULL; for (nmdp = start; nmdp < stop; nmdp++) { nmp = *nmdp; if (nmp->md_type != MDT_VERSION) continue; nmodname = nmp->md_cval; if (strcmp(modname, nmodname) == 0) break; } if (nmdp < stop)/* early exit, it's a self reference */ continue; mod = modlist_lookup2(modname, verinfo); if (mod) { /* woohoo, it's loaded already */ lfdep = mod->container; lfdep->refs++; error = linker_file_add_dependency(lf, lfdep); if (error) break; continue; } error = linker_load_module(NULL, modname, lf, verinfo, NULL); if (error) { printf("KLD %s: depends on %s - not available or" " version mismatch\n", lf->filename, modname); break; } } if (error) return (error); linker_addmodules(lf, start, stop, 0); return (error); } static int sysctl_kern_function_list_iterate(const char *name, void *opaque) { struct sysctl_req *req; req = opaque; return (SYSCTL_OUT(req, name, strlen(name) + 1)); } /* * Export a nul-separated, double-nul-terminated list of all function names * in the kernel. */ static int sysctl_kern_function_list(SYSCTL_HANDLER_ARGS) { linker_file_t lf; int error; #ifdef MAC error = mac_kld_check_stat(req->td->td_ucred); if (error) return (error); #endif error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sx_xlock(&kld_sx); TAILQ_FOREACH(lf, &linker_files, link) { error = LINKER_EACH_FUNCTION_NAME(lf, sysctl_kern_function_list_iterate, req); if (error) { sx_xunlock(&kld_sx); return (error); } } sx_xunlock(&kld_sx); return (SYSCTL_OUT(req, "", 1)); } SYSCTL_PROC(_kern, OID_AUTO, function_list, CTLTYPE_OPAQUE | CTLFLAG_RD, NULL, 0, sysctl_kern_function_list, "", "kernel function list"); Index: stable/9/sys =================================================================== --- stable/9/sys (revision 263357) +++ stable/9/sys (revision 263358) Property changes on: stable/9/sys ___________________________________________________________________ Modified: svn:mergeinfo ## -0,0 +0,1 ## Merged /head/sys:r263080