diff options
Diffstat (limited to 'cachepc/uspt.c')
| -rw-r--r-- | cachepc/uspt.c | 488 |
1 files changed, 488 insertions, 0 deletions
diff --git a/cachepc/uspt.c b/cachepc/uspt.c new file mode 100644 index 0000000..9050a55 --- /dev/null +++ b/cachepc/uspt.c @@ -0,0 +1,488 @@ +#include "uspt.h" +#include "sevstep.h" +#include "cachepc.h" + +#include <linux/kvm.h> +#include <linux/timekeeping.h> +#include <linux/uaccess.h> +#include <linux/types.h> +#include <linux/vmalloc.h> +#include <linux/printk.h> +#include <linux/ratelimit.h> + +#define ARRLEN(x) (sizeof(x)/sizeof((x)[0])) + +typedef struct { + bool is_active; + int tracking_type; + bool retrack; + + int perf_cpu; + + uint64_t gfn_retrack_backlog[10]; + int gfn_retrack_backlog_next_idx; + + page_fault_event_t * events; + uint64_t event_next_idx; + uint64_t events_size; + + bool error_occured; +} batch_track_state_t; + +typedef struct { + uint64_t idx_for_last_perf_reading; + uint64_t last_perf_reading; + uint64_t delta_valid_idx; + uint64_t delta; +} perf_state_t; + +// crude sync mechanism. don't know a good way to act on errors yet. +static uint64_t last_sent_event_id = 1; +static uint64_t last_acked_event_id = 1; +DEFINE_RWLOCK(event_lock); + +static page_fault_event_t sent_event; +static int have_event = 0; + +static bool get_rip = true; + +static int inited = 0; + +DEFINE_SPINLOCK(batch_track_state_lock); +static batch_track_state_t batch_track_state; + +static perf_state_t perf_state; + +static uint64_t perf_state_update_and_get_delta(uint64_t current_event_idx); + +void +sevstep_uspt_clear(void) +{ + write_lock(&event_lock); + inited = 0; + last_sent_event_id = 1; + last_acked_event_id = 1; + have_event = 0; + get_rip = false; + write_unlock(&event_lock); +} + +int +sevstep_uspt_initialize(int pid, bool should_get_rip) +{ + write_lock(&event_lock); + inited = 1; + last_sent_event_id = 1; + last_acked_event_id = 1; + have_event = 0; + get_rip = should_get_rip; + write_unlock(&event_lock); + + return 0; +} + +int +sevstep_uspt_is_initialiized() +{ + return inited; +} + +bool +sevstep_uspt_should_get_rip() +{ + bool tmp; + + read_lock(&event_lock); + tmp = get_rip; + read_unlock(&event_lock); + + return tmp; +} + +int +sevstep_uspt_send_and_block(uint64_t faulted_gpa, uint32_t error_code, + bool have_rip, uint64_t rip) +{ + ktime_t abort_after; + page_fault_event_t message_for_user; + + read_lock(&event_lock); + if (!sevstep_uspt_is_initialiized()) { + pr_warn("sevstep_uspt_send_and_block: ctx not initialized!\n"); + read_unlock(&event_lock); + return 1; + } + read_unlock(&event_lock); + + write_lock(&event_lock); + if (last_sent_event_id != last_acked_event_id) { + pr_warn("sevstep_uspt_send_and_block: " + "event id_s out of sync, aborting. Fix this later\n"); + write_unlock(&event_lock); + return 1; + } else { + // TODO: handle overflow + last_sent_event_id++; + } + message_for_user.id = last_sent_event_id; + message_for_user.faulted_gpa = faulted_gpa; + message_for_user.error_code = error_code; + message_for_user.have_rip_info = have_rip; + message_for_user.rip = rip; + message_for_user.ns_timestamp = ktime_get_real_ns(); + message_for_user.have_retired_instructions = false; + + // for poll based system; + have_event = 1; + sent_event = message_for_user; + // printk("sevstep_uspt_send_and_block sending event %llu\n", sent_event.id); + + write_unlock(&event_lock); + + // wait for ack, but with timeout. Otherwise small bugs in userland + // easily lead to a kernel hang + abort_after = ktime_get() + 1000000000ULL; // 1 sec in nanosecond + while (!sevstep_uspt_is_event_done(sent_event.id)) { + if (ktime_get() > abort_after) { + pr_warn("sevstep_uspt_send_and_block: " + "Waiting for ack of event %llu timed out, " + "continuing\n",sent_event.id); + return 3; + } + } + + return 0; +} + +int +sevstep_uspt_is_event_done(uint64_t id) +{ + int res; + + read_lock(&event_lock); + res = last_acked_event_id >= id; + read_unlock(&event_lock); + + return res; +} + +int +sevstep_uspt_handle_poll_event(page_fault_event_t* userpace_mem) +{ + int err; + + // most of the time we won't have an event + read_lock(&event_lock); + if (!have_event) { + read_unlock(&event_lock); + return KVM_USPT_POLL_EVENT_NO_EVENT; + } + read_unlock(&event_lock); + + write_lock(&event_lock); + if (have_event) { + err = copy_to_user(userpace_mem, + &sent_event, sizeof(page_fault_event_t)); + have_event = 0; + } else { + err = KVM_USPT_POLL_EVENT_NO_EVENT; + } + write_unlock(&event_lock); + + return err; +} + +int +sevstep_uspt_handle_ack_event_ioctl(ack_event_t event) +{ + int err = 0; + + write_lock(&event_lock); + if (event.id == last_sent_event_id) { + last_acked_event_id = last_sent_event_id; + } else { + err = 1; + pr_warn("sevstep_uspt_handle_ack_event_ioctl: " + "last sent event id is %llu but received ack for %llu\n", + last_sent_event_id, event.id); + } + write_unlock(&event_lock); + + return err; +} + +// get retired instructions between current_event_idx-1 and current_event_idx +// value is cached for multiple calls to the same current_event_idx +uint64_t +perf_state_update_and_get_delta(uint64_t current_event_idx) +{ + uint64_t current_value; + + /* check if value is "cached" */ + if (perf_state.delta_valid_idx == current_event_idx) { + if (current_event_idx == 0) { + perf_state.idx_for_last_perf_reading = current_event_idx; + perf_state.last_perf_reading = cachepc_read_pmc(0); + } + return perf_state.delta; + } + + /* otherwise update, but logic is only valid for two consecutive events */ + if (current_event_idx != perf_state.idx_for_last_perf_reading+1) { + pr_warn("perf_state_update_and_get_delta: " + "last reading was for idx %llu but was queried for %llu\n", + perf_state.idx_for_last_perf_reading, current_event_idx); + } + + current_value = cachepc_read_pmc(0); + perf_state.delta = (current_value - perf_state.last_perf_reading); + perf_state.delta_valid_idx = current_event_idx; + + perf_state.idx_for_last_perf_reading = current_event_idx; + perf_state.last_perf_reading = current_value; + + return perf_state.delta; +} + +void +sevstep_uspt_batch_tracking_inc_event_idx(void) +{ + spin_lock(&batch_track_state_lock); + batch_track_state.event_next_idx++; + spin_unlock(&batch_track_state_lock); +} + +int +sevstep_uspt_batch_tracking_start(int tracking_type,uint64_t expected_events, + int perf_cpu, bool retrack) +{ + page_fault_event_t* events; + uint64_t buffer_size, i; + + spin_lock(&batch_track_state_lock); + if (batch_track_state.is_active) { + pr_warn("sevstep_uspt_batch_tracking_start: " + "overwriting active batch track config!\n"); + if (batch_track_state.events != NULL ) { + vfree(batch_track_state.events); + } + } + batch_track_state.is_active = false; + spin_unlock(&batch_track_state_lock); + + buffer_size = expected_events * sizeof(page_fault_event_t); + pr_warn("sevstep_uspt_batch_tracking_start: " + "trying to alloc %llu bytes buffer for events\n", + buffer_size); + events = vmalloc(buffer_size); + if (events == NULL) { + pr_warn("sevstep_uspt_batch_tracking_start: " + "faperf_cpuiled to alloc %llu bytes for event buffer\n", + buffer_size); + return 1; // note: lock not held here + } + + // access each element once to force them into memory, improving performance + // during tracking + for (i = 0; i < expected_events * sizeof(page_fault_event_t); i++) { + ((volatile uint8_t*)events)[i] = 0; + } + + perf_state.idx_for_last_perf_reading = 0; + perf_state.last_perf_reading = 0; + perf_state.delta_valid_idx = 0; + perf_state.delta = 0; + cachepc_init_pmc(0, 0xc0, 0x00, PMC_GUEST, PMC_KERNEL | PMC_USER); + + spin_lock(&batch_track_state_lock); + + batch_track_state.perf_cpu = perf_cpu; + batch_track_state.retrack = retrack; + + batch_track_state.events = events; + batch_track_state.event_next_idx = 0; + batch_track_state.events_size = expected_events; + + batch_track_state.gfn_retrack_backlog_next_idx = 0; + batch_track_state.tracking_type = tracking_type; + batch_track_state.error_occured = false; + + batch_track_state.is_active = true; + + spin_unlock(&batch_track_state_lock); + + return 0; +} + +void +sevstep_uspt_batch_tracking_handle_retrack(struct kvm_vcpu* vcpu, + uint64_t current_fault_gfn) +{ + uint64_t ret_instr_delta; + int i, next_idx; + + spin_lock(&batch_track_state_lock); + + if (!batch_track_state.retrack) { + spin_unlock(&batch_track_state_lock); + return; + } + + if (smp_processor_id() != batch_track_state.perf_cpu) { + pr_warn("sevstep_uspt_batch_tracking_handle_retrack: perf was " + "programmed on logical cpu %d but handler was called " + "on %d. Did you forget to pin the vcpu thread?\n", + batch_track_state.perf_cpu, smp_processor_id()); + } + ret_instr_delta = perf_state_update_and_get_delta(batch_track_state.event_next_idx); + + // faulting instructions is probably the same as on last fault + // try to add current fault to retrack log and return + // for first event idx we do not have a valid ret_instr_delta. + // Retracking for the frist time is fine, if we loop, we end up here + // again but with a valid delta on one of the next event + if ((ret_instr_delta < 2) && ( batch_track_state.event_next_idx != 0)) { + next_idx = batch_track_state.gfn_retrack_backlog_next_idx; + if (next_idx >= ARRLEN(batch_track_state.gfn_retrack_backlog)) { + pr_warn("sevstep_uspt_batch_tracking_handle_retrack: " + "retrack backlog full, dropping retrack for fault " + "at 0x%llx\n", current_fault_gfn); + } else { + batch_track_state.gfn_retrack_backlog[next_idx] = current_fault_gfn; + batch_track_state.gfn_retrack_backlog_next_idx++; + } + + spin_unlock(&batch_track_state_lock); + return; + } + + /* made progress, retrack everything in backlog and reset idx */ + for (i = 0; i < batch_track_state.gfn_retrack_backlog_next_idx; i++) { + sevstep_track_single_page(vcpu, + batch_track_state.gfn_retrack_backlog[i], + batch_track_state.tracking_type); + } + + /* add current fault to list */ + batch_track_state.gfn_retrack_backlog[0] = current_fault_gfn; + batch_track_state.gfn_retrack_backlog_next_idx = 1; + + spin_unlock(&batch_track_state_lock); + +} + +int +sevstep_uspt_batch_tracking_save(uint64_t faulted_gpa, uint32_t error_code, + bool have_rip, uint64_t rip) +{ + uint64_t ret_instr_delta; + page_fault_event_t* event; + + spin_lock(&batch_track_state_lock); + + if (!batch_track_state.is_active) { + pr_warn("sevstep_uspt_batch_tracking_save: " + "got save but batch tracking is not active!\n"); + batch_track_state.error_occured = true; + spin_unlock(&batch_track_state_lock); + return 1; + } + + + if (batch_track_state.event_next_idx >= batch_track_state.events_size) { + pr_warn("sevstep_uspt_batch_tracking_save: events buffer is full!\n"); + batch_track_state.error_occured = true; + spin_unlock(&batch_track_state_lock); + return 1; + } + + if (smp_processor_id() != batch_track_state.perf_cpu) { + pr_warn("sevstep_uspt_batch_tracking_save: perf was " + "programmed on logical cpu %d but handler was called " + "on %d. Did you forget to pin the vcpu thread?\n", + batch_track_state.perf_cpu, smp_processor_id()); + } + ret_instr_delta = perf_state_update_and_get_delta(batch_track_state.event_next_idx); + + + if (batch_track_state.events == NULL) { + pr_warn("sevstep_uspt_batch_tracking_save: events buf was " + "NULL but \"is_active\" was set! This should never happen!!!\n"); + spin_unlock(&batch_track_state_lock); + return 1; + } + + event = &batch_track_state.events[batch_track_state.event_next_idx]; + event->id = batch_track_state.event_next_idx; + event->faulted_gpa = faulted_gpa; + event->error_code = error_code; + event->have_rip_info = have_rip; + event->rip = rip; + event->ns_timestamp = ktime_get_real_ns(); + event->have_retired_instructions = true; + event->retired_instructions = ret_instr_delta; + + // old inc was here + + if (batch_track_state.gfn_retrack_backlog_next_idx + > ARRLEN(batch_track_state.gfn_retrack_backlog)) { + pr_warn("sevstep_uspt_batch_tracking_save: " + "gfn retrack backlog overflow!\n"); + batch_track_state.error_occured = true; + spin_unlock(&batch_track_state_lock); + return 1; + } + + spin_unlock(&batch_track_state_lock); + + return 0; +} + +int +sevstep_uspt_batch_tracking_stop(page_fault_event_t* results, + uint64_t len, __u8* error_occured) +{ + spin_lock(&batch_track_state_lock); + if (!batch_track_state.is_active) { + pr_warn("sevstep_uspt: batch tracking not active\n"); + spin_unlock(&batch_track_state_lock); + return 1; + + } + batch_track_state.is_active = false; + + if (len > batch_track_state.event_next_idx) { + pr_warn("sevstep_uspt_batch_tracking_stop: " + "requested %llu events but got only %llu\n", + len, batch_track_state.event_next_idx); + spin_unlock(&batch_track_state_lock); + return 1; + } + + memcpy(results,batch_track_state.events, len*sizeof(page_fault_event_t)); + vfree(batch_track_state.events); + + *error_occured = batch_track_state.error_occured; + + spin_unlock(&batch_track_state_lock); + + return 0; +} + +uint64_t +sevstep_uspt_batch_tracking_get_events_count() +{ + uint64_t buf; + + spin_lock(&batch_track_state_lock); + buf = batch_track_state.event_next_idx; + spin_unlock(&batch_track_state_lock); + + return buf; +} + +bool +sevstep_uspt_batch_tracking_in_progress() +{ + return batch_track_state.is_active; +} |
