/* Minimum number of sample-data-block-tables: * At least one table is required for the sampling buffer structure. * A single table contains up to 511 pointers to sample-data-blocks.
*/ #define CPUM_SF_MIN_SDBT 1
/* Number of sample-data-blocks per sample-data-block-table (SDBT): * A table contains SDB pointers (8 bytes) and one table-link entry * that points to the origin of the next SDBT.
*/ #define CPUM_SF_SDB_PER_TABLE ((PAGE_SIZE - 8) / 8)
/* Maximum page offset for an SDBT table-link entry: * If this page offset is reached, a table-link entry to the next SDBT * must be added.
*/ #define CPUM_SF_SDBT_TL_OFFSET (CPUM_SF_SDB_PER_TABLE * 8) staticinlineint require_table_link(constvoid *sdbt)
{ return ((unsignedlong)sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET;
}
/* Minimum and maximum sampling buffer sizes: * * This number represents the maximum size of the sampling buffer taking * the number of sample-data-block-tables into account. Note that these * numbers apply to the basic-sampling function only. * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if * the diagnostic-sampling function is active. * * Sampling buffer size Buffer characteristics * --------------------------------------------------- * 64KB == 16 pages (4KB per page) * 1 page for SDB-tables * 15 pages for SDBs * * 32MB == 8192 pages (4KB per page) * 16 pages for SDB-tables * 8176 pages for SDBs
*/ staticunsignedlong __read_mostly CPUM_SF_MIN_SDB = 15; staticunsignedlong __read_mostly CPUM_SF_MAX_SDB = 8176; staticunsignedlong __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1;
struct sf_buffer { unsignedlong *sdbt; /* Sample-data-block-table origin */ /* buffer characteristics (required for buffer increments) */ unsignedlong num_sdb; /* Number of sample-data-blocks */ unsignedlong num_sdbt; /* Number of sample-data-block-tables */ unsignedlong *tail; /* last sample-data-block-table */
};
struct aux_buffer { struct sf_buffer sfb; unsignedlong head; /* index of SDB of buffer head */ unsignedlong alert_mark; /* index of SDB of alert request position */ unsignedlong empty_mark; /* mark of SDB not marked full */ unsignedlong *sdb_index; /* SDB address for fast lookup */ unsignedlong *sdbt_index; /* SDBT address for fast lookup */
};
/* Return pointer to trailer entry of an sample data block */ staticinlinestruct hws_trailer_entry *trailer_entry_ptr(unsignedlong v)
{ void *ret;
ret = (void *)v;
ret += PAGE_SIZE;
ret -= sizeof(struct hws_trailer_entry);
return ret;
}
/* * Return true if the entry in the sample data block table (sdbt) * is a link to the next sdbt
*/ staticinlineint is_link_entry(unsignedlong *s)
{ return *s & 0x1UL ? 1 : 0;
}
/* Return pointer to the linked sdbt */ staticinlineunsignedlong *get_next_sdbt(unsignedlong *s)
{ return phys_to_virt(*s & ~0x1UL);
}
/* Allocate and initialize sample-data-block */
sdb = get_zeroed_page(gfp_flags); if (!sdb) return -ENOMEM;
te = trailer_entry_ptr(sdb);
te->header.a = 1;
/* Link SDB into the sample-data-block-table */
*sdbt = virt_to_phys((void *)sdb);
return 0;
}
/* * realloc_sampling_buffer() - extend sampler memory * * Allocates new sample-data-blocks and adds them to the specified sampling * buffer memory. * * Important: This modifies the sampling buffer and must be called when the * sampling facility is disabled. * * Returns zero on success, non-zero otherwise.
*/ staticint realloc_sampling_buffer(struct sf_buffer *sfb, unsignedlong num_sdb, gfp_t gfp_flags)
{ int i, rc; unsignedlong *new, *tail, *tail_prev = NULL;
if (!sfb->sdbt || !sfb->tail) return -EINVAL;
if (!is_link_entry(sfb->tail)) return -EINVAL;
/* Append to the existing sampling buffer, overwriting the table-link * register. * The tail variables always points to the "tail" (last and table-link) * entry in an SDB-table.
*/
tail = sfb->tail;
/* Do a sanity check whether the table-link entry points to * the sampling buffer origin.
*/ if (sfb->sdbt != get_next_sdbt(tail)) {
debug_sprintf_event(sfdbg, 3, "%s buffer not linked origin %#lx tail %#lx\n",
__func__, (unsignedlong)sfb->sdbt,
(unsignedlong)tail); return -EINVAL;
}
/* Allocate remaining SDBs */
rc = 0; for (i = 0; i < num_sdb; i++) { /* Allocate a new SDB-table if it is full. */ if (require_table_link(tail)) { new = (unsignedlong *)get_zeroed_page(gfp_flags); if (!new) {
rc = -ENOMEM; break;
}
sfb->num_sdbt++; /* Link current page to tail of chain */
*tail = virt_to_phys((void *)new) + 1;
tail_prev = tail;
tail = new;
}
/* Allocate a new sample-data-block. * If there is not enough memory, stop the realloc process * and simply use what was allocated. If this is a temporary * issue, a new realloc call (if required) might succeed.
*/
rc = alloc_sample_data_block(tail, gfp_flags); if (rc) { /* Undo last SDBT. An SDBT with no SDB at its first * entry but with an SDBT entry instead can not be * handled by the interrupt handler code. * Avoid this situation.
*/ if (tail_prev) {
sfb->num_sdbt--;
free_page((unsignedlong)new);
tail = tail_prev;
} break;
}
sfb->num_sdb++;
tail++;
tail_prev = new = NULL; /* Allocated at least one SBD */
}
/* Link sampling buffer to its origin */
*tail = virt_to_phys(sfb->sdbt) + 1;
sfb->tail = tail;
return rc;
}
/* * allocate_sampling_buffer() - allocate sampler memory * * Allocates and initializes a sampling buffer structure using the * specified number of sample-data-blocks (SDB). For each allocation, * a 4K page is used. The number of sample-data-block-tables (SDBT) * are calculated from SDBs. * Also set the ALERT_REQ mask in each SDBs trailer. * * Returns zero on success, non-zero otherwise.
*/ staticint alloc_sampling_buffer(struct sf_buffer *sfb, unsignedlong num_sdb)
{ int rc;
if (sfb->sdbt) return -EINVAL;
/* Allocate the sample-data-block-table origin */
sfb->sdbt = (unsignedlong *)get_zeroed_page(GFP_KERNEL); if (!sfb->sdbt) return -ENOMEM;
sfb->num_sdb = 0;
sfb->num_sdbt = 1;
/* Link the table origin to point to itself to prepare for * realloc_sampling_buffer() invocation.
*/
sfb->tail = sfb->sdbt;
*sfb->tail = virt_to_phys((void *)sfb->sdbt) + 1;
/* Allocate requested number of sample-data-blocks */
rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL); if (rc)
free_sampling_buffer(sfb); return rc;
}
staticvoid sfb_account_allocs(unsignedlong num, struct hw_perf_event *hwc)
{ /* Limit the number of SDBs to not exceed the maximum */
num = min_t(unsignedlong, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc)); if (num)
SFB_ALLOC_REG(hwc) += num;
}
/* Calculate sampling buffers using 4K pages * * 1. The sampling size is 32 bytes for basic sampling. This size * is the same for all machine types. Diagnostic * sampling uses auxlilary data buffer setup which provides the * memory for SDBs using linux common code auxiliary trace * setup. * * 2. Function alloc_sampling_buffer() sets the Alert Request * Control indicator to trigger a measurement-alert to harvest * sample-data-blocks (SDB). This is done per SDB. This * measurement alert interrupt fires quick enough to handle * one SDB, on very high frequency and work loads there might * be 2 to 3 SBDs available for sample processing. * Currently there is no need for setup alert request on every * n-th page. This is counterproductive as one IRQ triggers * a very high number of samples to be processed at one IRQ. * * 3. Use the sampling frequency as input. * Compute the number of SDBs and ensure a minimum * of CPUM_SF_MIN_SDB. Depending on frequency add some more * SDBs to handle a higher sampling rate. * Use a minimum of CPUM_SF_MIN_SDB and allow for 100 samples * (one SDB) for every 10000 HZ frequency increment. * * 4. Compute the number of sample-data-block-tables (SDBT) and * ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up * to 511 SDBs).
*/
freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
n_sdb = CPUM_SF_MIN_SDB + DIV_ROUND_UP(freq, 10000);
/* If there is already a sampling buffer allocated, it is very likely * that the sampling facility is enabled too. If the event to be * initialized requires a greater sampling buffer, the allocation must * be postponed. Changing the sampling buffer requires the sampling * facility to be in the disabled state. So, account the number of * required SDBs and let cpumsf_pmu_enable() resize the buffer just * before the event is started.
*/
sfb_init_allocs(n_sdb, hwc); if (sf_buffer_available(cpuhw)) return 0;
staticunsignedlong compute_sfb_extent(unsignedlong ratio, unsignedlong base)
{ /* Use a percentage-based approach to extend the sampling facility * buffer. Accept up to 5% sample data loss. * Vary the extents between 1% to 5% of the current number of * sample-data-blocks.
*/ if (ratio <= 5) return 0; if (ratio <= 25) return min_percent(1, base, 1); if (ratio <= 50) return min_percent(1, base, 1); if (ratio <= 75) return min_percent(2, base, 2); if (ratio <= 100) return min_percent(3, base, 3); if (ratio <= 250) return min_percent(4, base, 4);
/* The sample_overflow contains the average number of sample data * that has been lost because sample-data-blocks were full. * * Calculate the total number of sample data entries that has been * discarded. Then calculate the ratio of lost samples to total samples * per second in percent.
*/
ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb,
sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)));
/* Compute number of sample-data-blocks */
num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb); if (num)
sfb_account_allocs(num, hwc);
OVERFLOW_REG(hwc) = 0;
}
/* extend_sampling_buffer() - Extend sampling buffer * @sfb: Sampling buffer structure (for local CPU) * @hwc: Perf event hardware structure * * Use this function to extend the sampling buffer based on the overflow counter * and postponed allocation extents stored in the specified Perf event hardware. * * Important: This function disables the sampling facility in order to safely * change the sampling buffer structure. Do not call this function * when the PMU is active.
*/ staticvoid extend_sampling_buffer(struct sf_buffer *sfb, struct hw_perf_event *hwc)
{ unsignedlong num;
num = sfb_pending_allocs(sfb, hwc); if (!num) return;
/* Disable the sampling facility to reset any states and also * clear pending measurement alerts.
*/
sf_disable();
/* Extend the sampling buffer. * This memory allocation typically happens in an atomic context when * called by perf. Because this is a reallocation, it is fine if the * new SDB-request cannot be satisfied immediately.
*/
realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
}
/* Number of perf events counting hardware events */ static refcount_t num_events; /* Used to avoid races in calling reserve/release_cpumf_hardware */ static DEFINE_MUTEX(pmc_reserve_mutex);
staticvoid hw_perf_event_destroy(struct perf_event *event)
{ /* Release PMC if this is the last perf event */ if (refcount_dec_and_mutex_lock(&num_events, &pmc_reserve_mutex)) {
release_pmc_hardware();
mutex_unlock(&pmc_reserve_mutex);
}
}
tsk = find_task_by_pid_ns(pid, &init_pid_ns);
pid = -1; if (tsk) { /* * Only top level events contain the pid namespace in which * they are created.
*/ if (event->parent)
event = event->parent;
pid = __task_pid_nr_ns(tsk, type, event->ns); /* * See also 1d953111b648 * "perf/core: Don't report zero PIDs for exiting tasks".
*/ if (!pid && !pid_alive(tsk))
pid = -1;
}
out: return pid;
}
/* Update the process ID (see also kernel/events/core.c) */
data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID);
data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID);
if (freq) {
rate = freq_to_sample_rate(si, sample);
rate = hw_limit_rate(si, rate);
} else { /* The min/max sampling rates specifies the valid range * of sample periods. If the specified sample period is * out of range, limit the period to the range boundary.
*/
rate = hw_limit_rate(si, sample);
/* The perf core maintains a maximum sample rate that is * configurable through the sysctl interface. Ensure the * sampling rate does not exceed this value. This also helps * to avoid throttling when pushing samples with * perf_event_overflow().
*/ if (sample_rate_to_freq(si, rate) >
sysctl_perf_event_sample_rate) {
rate = 0;
}
} return rate;
}
/* The sampling information (si) contains information about the * min/max sampling intervals and the CPU speed. So calculate the * correct sampling interval and avoid the whole period adjust * feedback loop. * * Since the CPU Measurement sampling facility can not handle frequency * calculate the sampling interval when frequency is specified using * this formula: * interval := cpu_speed * 1000000 / sample_freq * * Returns errno on bad input and zero on success with parameter interval * set to the correct sampling rate. * * Note: This function turns off freq bit to avoid calling function * perf_adjust_period(). This causes frequency adjustment in the common * code part which causes tremendous variations in the counter values.
*/ staticint __hw_perf_event_init_rate(struct perf_event *event, struct hws_qsi_info_block *si)
{ struct perf_event_attr *attr = &event->attr; struct hw_perf_event *hwc = &event->hw; unsignedlong rate;
/* Access per-CPU sampling information (query sampling info) */ /* * The event->cpu value can be -1 to count on every CPU, for example, * when attaching to a task. If this is specified, use the query * sampling info from the current CPU, otherwise use event->cpu to * retrieve the per-CPU information. * Later, cpuhw indicates whether to allocate sampling buffers for a * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
*/
memset(&si, 0, sizeof(si));
cpuhw = NULL; if (event->cpu == -1) {
qsi(&si);
} else { /* Event is pinned to a particular CPU, retrieve the per-CPU * sampling structure for accessing the CPU-specific QSI.
*/
cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
si = cpuhw->qsi;
}
/* Check sampling facility authorization and, if not authorized, * fall back to other PMUs. It is safe to check any CPU because * the authorization is identical for all configured CPUs.
*/ if (!si.as) {
err = -ENOENT; goto out;
}
if (si.ribm & CPU_MF_SF_RIBM_NOTAV) {
pr_warn("CPU Measurement Facility sampling is temporarily not available\n");
err = -EBUSY; goto out;
}
/* Check if diagnostic sampling is requested. Deny if the required * sampling authorization is missing.
*/ if (attr->config == PERF_EVENT_CPUM_SF_DIAG) { if (!si.ad) {
err = -EPERM; goto out;
}
SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
}
err = __hw_perf_event_init_rate(event, &si); if (err) goto out;
/* Use AUX buffer. No need to allocate it by ourself */ if (attr->config == PERF_EVENT_CPUM_SF_DIAG) goto out;
/* Allocate the per-CPU sampling buffer using the CPU information * from the event. If the event is not pinned to a particular * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling * buffers for each online CPU.
*/ if (cpuhw) /* Event is pinned to a particular CPU */
err = allocate_buffers(cpuhw, hwc); else { /* Event is not pinned, allocate sampling buffer on * each online CPU
*/
for_each_online_cpu(cpu) {
cpuhw = &per_cpu(cpu_hw_sf, cpu);
err = allocate_buffers(cpuhw, hwc); if (err) break;
}
}
/* If PID/TID sampling is active, replace the default overflow * handler to extract and resolve the PIDs from the basic-sampling * data entries.
*/ if (event->attr.sample_type & PERF_SAMPLE_TID) if (is_default_overflow_handler(event))
event->overflow_handler = cpumsf_output_event_pid;
out:
mutex_unlock(&pmc_reserve_mutex); return err;
}
staticint cpumsf_pmu_event_init(struct perf_event *event)
{ int err;
/* No support for taken branch sampling */ /* No support for callchain, stacks and registers */ if (has_branch_stack(event) || is_callchain_event(event)) return -EOPNOTSUPP;
switch (event->attr.type) { case PERF_TYPE_RAW: if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
(event->attr.config != PERF_EVENT_CPUM_SF_DIAG)) return -ENOENT; break; case PERF_TYPE_HARDWARE: /* Support sampling of CPU cycles in addition to the * counter facility. However, the counter facility * is more precise and, hence, restrict this PMU to * sampling events only.
*/ if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES) return -ENOENT; if (!is_sampling_event(event)) return -ENOENT; break; default: return -ENOENT;
}
/* Force reset of idle/hv excludes regardless of what the * user requested.
*/ if (event->attr.exclude_hv)
event->attr.exclude_hv = 0; if (event->attr.exclude_idle)
event->attr.exclude_idle = 0;
/* * Event must be * - added/started on this CPU (PMU_F_IN_USE set) * - and CPU must be available (PMU_F_RESERVED set) * - and not already enabled (PMU_F_ENABLED not set) * - and not in error condition (PMU_F_ERR_MASK not set)
*/ if (cpuhw->flags != (PMU_F_IN_USE | PMU_F_RESERVED)) return;
/* Check whether to extent the sampling buffer. * * Two conditions trigger an increase of the sampling buffer for a * perf event: * 1. Postponed buffer allocations from the event initialization. * 2. Sampling overflows that contribute to pending allocations. * * Note that the extend_sampling_buffer() function disables the sampling * facility, but it can be fully re-enabled using sampling controls that * have been saved in cpumsf_pmu_disable().
*/
hwc = &cpuhw->event->hw; if (!(SAMPL_DIAG_MODE(hwc))) { /* * Account number of overflow-designated buffer extents
*/
sfb_account_overflows(cpuhw, hwc);
extend_sampling_buffer(&cpuhw->sfb, hwc);
} /* Rate may be adjusted with ioctl() */
cpuhw->lsctl.interval = SAMPL_RATE(hwc);
/* (Re)enable the PMU and sampling facility */
err = lsctl(&cpuhw->lsctl); if (err) {
pr_err("Loading sampling controls failed: op 1 err %i\n", err); return;
}
/* Load current program parameter */
lpp(&get_lowcore()->lpp);
cpuhw->flags |= PMU_F_ENABLED;
}
/* Switch off sampling activation control */
inactive = cpuhw->lsctl;
inactive.cs = 0;
inactive.cd = 0;
err = lsctl(&inactive); if (err) {
pr_err("Loading sampling controls failed: op 2 err %i\n", err); return;
}
/* * Save state of TEAR and DEAR register contents. * TEAR/DEAR values are valid only if the sampling facility is * enabled. Note that cpumsf_pmu_disable() might be called even * for a disabled sampling facility because cpumsf_pmu_enable() * controls the enable/disable state.
*/
qsi(&si); if (si.es) {
cpuhw->lsctl.tear = si.tear;
cpuhw->lsctl.dear = si.dear;
}
cpuhw->flags &= ~PMU_F_ENABLED;
}
/* perf_event_exclude() - Filter event * @event: The perf event * @regs: pt_regs structure * @sde_regs: Sample-data-entry (sde) regs structure * * Filter perf events according to their exclude specification. * * Return non-zero if the event shall be excluded.
*/ staticint perf_event_exclude(struct perf_event *event, struct pt_regs *regs, struct perf_sf_sde_regs *sde_regs)
{ if (event->attr.exclude_user && user_mode(regs)) return 1; if (event->attr.exclude_kernel && !user_mode(regs)) return 1; if (event->attr.exclude_guest && sde_regs->in_guest) return 1; if (event->attr.exclude_host && !sde_regs->in_guest) return 1; return 0;
}
/* perf_push_sample() - Push samples to perf * @event: The perf event * @sample: Hardware sample data * * Use the hardware sample data to create perf event sample. The sample * is the pushed to the event subsystem and the function checks for * possible event overflows. If an event overflow occurs, the PMU is * stopped. * * Return non-zero if an event overflow occurred.
*/ staticint perf_push_sample(struct perf_event *event, struct hws_basic_entry *basic)
{ int overflow; struct pt_regs regs; struct perf_sf_sde_regs *sde_regs; struct perf_sample_data data;
/* Setup pt_regs to look like an CPU-measurement external interrupt * using the Program Request Alert code. The regs.int_parm_long * field which is unused contains additional sample-data-entry related * indicators.
*/
memset(®s, 0, sizeof(regs));
regs.int_code = 0x1407;
regs.int_parm = CPU_MF_INT_SF_PRA;
sde_regs = (struct perf_sf_sde_regs *) ®s.int_parm_long;
/* * Use the hardware provided configuration level to decide if the * sample belongs to a guest or host. If that is not available, * fall back to the following heuristics: * A non-zero guest program parameter always indicates a guest * sample. Some early samples or samples from guests without * lpp usage would be misaccounted to the host. We use the asn * value as an addon heuristic to detect most of these guest samples. * If the value differs from 0xffff (the host value), we assume to * be a KVM guest.
*/ switch (basic->CL) { case 1: /* logical partition */
sde_regs->in_guest = 0; break; case 2: /* virtual machine */
sde_regs->in_guest = 1; break; default: /* old machine, use heuristics */ if (basic->gpp || basic->prim_asn != 0xffff)
sde_regs->in_guest = 1; break;
}
/* * Store the PID value from the sample-data-entry to be * processed and resolved by cpumsf_output_event_pid().
*/
data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
/* hw_collect_samples() - Walk through a sample-data-block and collect samples * @event: The perf event * @sdbt: Sample-data-block table * @overflow: Event overflow counter * * Walks through a sample-data-block and collects sampling data entries that are * then pushed to the perf event subsystem. Depending on the sampling function, * there can be either basic-sampling or combined-sampling data entries. A * combined-sampling data entry consists of a basic- and a diagnostic-sampling * data entry. The sampling function is determined by the flags in the perf * event hardware structure. The function always works with a combined-sampling * data entry but ignores the the diagnostic portion if it is not available. * * Note that the implementation focuses on basic-sampling data entries and, if * such an entry is not valid, the entire combined-sampling data entry is * ignored. * * The overflow variables counts the number of samples that has been discarded * due to a perf event overflow.
*/ staticvoid hw_collect_samples(struct perf_event *event, unsignedlong *sdbt, unsignedlonglong *overflow)
{ struct hws_trailer_entry *te; struct hws_basic_entry *sample;
te = trailer_entry_ptr((unsignedlong)sdbt);
sample = (struct hws_basic_entry *)sdbt; while ((unsignedlong *)sample < (unsignedlong *)te) { /* Check for an empty sample */ if (!sample->def || sample->LS) break;
/* Update perf event period */
perf_event_count_update(event, SAMPL_RATE(&event->hw));
/* Check whether sample is valid */ if (sample->def == 0x0001) { /* If an event overflow occurred, the PMU is stopped to * throttle event delivery. Remaining sample data is * discarded.
*/ if (!*overflow) { /* Check whether sample is consistent */ if (sample->I == 0 && sample->W == 0) { /* Deliver sample data to perf */
*overflow = perf_push_sample(event,
sample);
}
} else /* Count discarded samples */
*overflow += 1;
} else { /* Sample slot is not yet written or other record. * * This condition can occur if the buffer was reused * from a combined basic- and diagnostic-sampling. * If only basic-sampling is then active, entries are * written into the larger diagnostic entries. * This is typically the case for sample-data-blocks * that are not full. Stop processing if the first * invalid format was detected.
*/ if (!te->header.f) break;
}
/* Reset sample slot and advance to next sample */
sample->def = 0;
sample++;
}
}
/* hw_perf_event_update() - Process sampling buffer * @event: The perf event * @flush_all: Flag to also flush partially filled sample-data-blocks * * Processes the sampling buffer and create perf event samples. * The sampling buffer position are retrieved and saved in the TEAR_REG * register of the specified perf event. * * Only full sample-data-blocks are processed. Specify the flush_all flag * to also walk through partially filled sample-data-blocks.
*/ staticvoid hw_perf_event_update(struct perf_event *event, int flush_all)
{ unsignedlonglong event_overflow, sampl_overflow, num_sdb; struct hw_perf_event *hwc = &event->hw; union hws_trailer_header prev, new; struct hws_trailer_entry *te; unsignedlong *sdbt, sdb; int done;
/* * AUX buffer is used when in diagnostic sampling mode. * No perf events/samples are created.
*/ if (SAMPL_DIAG_MODE(hwc)) return;
sdbt = (unsignedlong *)TEAR_REG(hwc);
done = event_overflow = sampl_overflow = num_sdb = 0; while (!done) { /* Get the trailer entry of the sample-data-block */
sdb = (unsignedlong)phys_to_virt(*sdbt);
te = trailer_entry_ptr(sdb);
/* Leave loop if no more work to do (block full indicator) */ if (!te->header.f) {
done = 1; if (!flush_all) break;
}
/* Check the sample overflow count */ if (te->header.overflow) /* Account sample overflows and, if a particular limit * is reached, extend the sampling buffer. * For details, see sfb_account_overflows().
*/
sampl_overflow += te->header.overflow;
/* Collect all samples from a single sample-data-block and * flag if an (perf) event overflow happened. If so, the PMU * is stopped and remaining samples will be discarded.
*/
hw_collect_samples(event, (unsignedlong *)sdb, &event_overflow);
num_sdb++;
/* Stop processing sample-data if all samples of the current * sample-data-block were flushed even if it was not full.
*/ if (flush_all && done) break;
}
/* Account sample overflows in the event hardware structure */ if (sampl_overflow)
OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
sampl_overflow, 1 + num_sdb);
/* Perf_event_overflow() and perf_event_account_interrupt() limit * the interrupt rate to an upper limit. Roughly 1000 samples per * task tick. * Hitting this limit results in a large number * of throttled REF_REPORT_THROTTLE entries and the samples * are dropped. * Slightly increase the interval to avoid hitting this limit.
*/ if (event_overflow)
SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10);
}
/* * Get trailer entry by index of SDB.
*/ staticstruct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux, unsignedlong index)
{ unsignedlong sdb;
index = aux_sdb_index(aux, index);
sdb = aux->sdb_index[index]; return trailer_entry_ptr(sdb);
}
/* * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu * disabled. Collect the full SDBs in AUX buffer which have not reached * the point of alert indicator. And ignore the SDBs which are not * full. * * 1. Scan SDBs to see how much data is there and consume them. * 2. Remove alert indicator in the buffer.
*/ staticvoid aux_output_end(struct perf_output_handle *handle)
{ unsignedlong i, range_scan, idx; struct aux_buffer *aux; struct hws_trailer_entry *te;
aux = perf_get_aux(handle); if (!aux) return;
range_scan = aux_sdb_num_alert(aux); for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
te = aux_sdb_trailer(aux, idx); if (!te->header.f) break;
} /* i is num of SDBs which are full */
perf_aux_output_end(handle, i << PAGE_SHIFT);
/* Remove alert indicators in the buffer */
te = aux_sdb_trailer(aux, aux->alert_mark);
te->header.a = 0;
}
/* * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event * is first added to the CPU or rescheduled again to the CPU. It is called * with pmu disabled. * * 1. Reset the trailer of SDBs to get ready for new data. * 2. Tell the hardware where to put the data by reset the SDBs buffer * head(tear/dear).
*/ staticint aux_output_begin(struct perf_output_handle *handle, struct aux_buffer *aux, struct cpu_hw_sf *cpuhw)
{ unsignedlong range, i, range_scan, idx, head, base, offset; struct hws_trailer_entry *te;
if (handle->head & ~PAGE_MASK) return -EINVAL;
aux->head = handle->head >> PAGE_SHIFT;
range = (handle->size + 1) >> PAGE_SHIFT; if (range <= 1) return -ENOMEM;
/* * SDBs between aux->head and aux->empty_mark are already ready * for new data. range_scan is num of SDBs not within them.
*/ if (range > aux_sdb_num_empty(aux)) {
range_scan = range - aux_sdb_num_empty(aux);
idx = aux->empty_mark + 1; for (i = 0; i < range_scan; i++, idx++) {
te = aux_sdb_trailer(aux, idx);
te->header.f = 0;
te->header.a = 0;
te->header.overflow = 0;
} /* Save the position of empty SDBs */
aux->empty_mark = aux->head + range - 1;
}
/* Set alert indicator */
aux->alert_mark = aux->head + range/2 - 1;
te = aux_sdb_trailer(aux, aux->alert_mark);
te->header.a = 1;
/* Reset hardware buffer head */
head = aux_sdb_index(aux, aux->head);
base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
offset = head % CPUM_SF_SDB_PER_TABLE;
cpuhw->lsctl.tear = virt_to_phys((void *)base) + offset * sizeof(unsignedlong);
cpuhw->lsctl.dear = virt_to_phys((void *)aux->sdb_index[head]);
return 0;
}
/* * Set alert indicator on SDB at index @alert_index while sampler is running. * * Return true if successfully. * Return false if full indicator is already set by hardware sampler.
*/ staticbool aux_set_alert(struct aux_buffer *aux, unsignedlong alert_index, unsignedlonglong *overflow)
{ union hws_trailer_header prev, new; struct hws_trailer_entry *te;
te = aux_sdb_trailer(aux, alert_index);
prev.val = READ_ONCE_ALIGNED_128(te->header.val); do { new.val = prev.val;
*overflow = prev.overflow; if (prev.f) { /* * SDB is already set by hardware. * Abort and try to set somewhere * behind.
*/ returnfalse;
} new.a = 1; new.overflow = 0;
} while (!try_cmpxchg128(&te->header.val, &prev.val, new.val)); returntrue;
}
/* * aux_reset_buffer() - Scan and setup SDBs for new samples * @aux: The AUX buffer to set * @range: The range of SDBs to scan started from aux->head * @overflow: Set to overflow count * * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is * marked as empty, check if it is already set full by the hardware sampler. * If yes, that means new data is already there before we can set an alert * indicator. Caller should try to set alert indicator to some position behind. * * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used * previously and have already been consumed by user space. Reset these SDBs * (clear full indicator and alert indicator) for new data. * If aux->alert_mark fall in this area, just set it. Overflow count is * recorded while scanning. * * SDBs between aux->head and aux->empty_mark are already reset at last time. * and ready for new samples. So scanning on this area could be skipped. * * Return true if alert indicator is set successfully and false if not.
*/ staticbool aux_reset_buffer(struct aux_buffer *aux, unsignedlong range, unsignedlonglong *overflow)
{ union hws_trailer_header prev, new; unsignedlong i, range_scan, idx; unsignedlonglong orig_overflow; struct hws_trailer_entry *te;
if (range <= aux_sdb_num_empty(aux)) /* * No need to scan. All SDBs in range are marked as empty. * Just set alert indicator. Should check race with hardware * sampler.
*/ return aux_set_alert(aux, aux->alert_mark, overflow);
if (aux->alert_mark <= aux->empty_mark) /* * Set alert indicator on empty SDB. Should check race * with hardware sampler.
*/ if (!aux_set_alert(aux, aux->alert_mark, overflow)) returnfalse;
/* * Scan the SDBs to clear full and alert indicator used previously. * Start scanning from one SDB behind empty_mark. If the new alert * indicator fall into this range, set it.
*/
range_scan = range - aux_sdb_num_empty(aux);
idx = aux->empty_mark + 1; for (i = 0; i < range_scan; i++, idx++) {
te = aux_sdb_trailer(aux, idx);
prev.val = READ_ONCE_ALIGNED_128(te->header.val); do { new.val = prev.val;
orig_overflow = prev.overflow; new.f = 0; new.overflow = 0; if (idx == aux->alert_mark) new.a = 1; else new.a = 0;
} while (!try_cmpxchg128(&te->header.val, &prev.val, new.val));
*overflow += orig_overflow;
}
/* Update empty_mark to new position */
aux->empty_mark = aux->head + range - 1;
/* Inform user space new data arrived */
size = aux_sdb_num_alert(aux) << PAGE_SHIFT;
debug_sprintf_event(sfdbg, 6, "%s #alert %ld\n", __func__,
size >> PAGE_SHIFT);
perf_aux_output_end(handle, size);
num_sdb = aux->sfb.num_sdb; while (!done) { /* Get an output handle */
aux = perf_aux_output_begin(handle, cpuhw->event); if (handle->size == 0) {
pr_err("The AUX buffer with %lu pages for the " "diagnostic-sampling mode is full\n",
num_sdb); break;
} if (!aux) return;
/* Update head and alert_mark to new position */
aux->head = handle->head >> PAGE_SHIFT;
range = (handle->size + 1) >> PAGE_SHIFT; if (range == 1)
aux->alert_mark = aux->head; else
aux->alert_mark = aux->head + range/2 - 1;
if (aux_reset_buffer(aux, range, &overflow)) { if (!overflow) {
done = 1; break;
}
size = range << PAGE_SHIFT;
perf_aux_output_end(&cpuhw->handle, size);
pr_err("Sample data caused the AUX buffer with %lu " "pages to overflow\n", aux->sfb.num_sdb);
} else {
size = aux_sdb_num_alert(aux) << PAGE_SHIFT;
perf_aux_output_end(&cpuhw->handle, size);
}
}
}
/* * Callback when freeing AUX buffers.
*/ staticvoid aux_buffer_free(void *data)
{ struct aux_buffer *aux = data; unsignedlong i, num_sdbt;
if (!aux) return;
/* Free SDBT. SDB is freed by the caller */
num_sdbt = aux->sfb.num_sdbt; for (i = 0; i < num_sdbt; i++)
free_page(aux->sdbt_index[i]);
/* Save clock base */
te->clock_base = 1;
te->progusage2 = tod_clock_base.tod;
}
/* * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling * @event: Event the buffer is setup for, event->cpu == -1 means current * @pages: Array of pointers to buffer pages passed from perf core * @nr_pages: Total pages * @snapshot: Flag for snapshot mode * * This is the callback when setup an event using AUX buffer. Perf tool can * trigger this by an additional mmap() call on the event. Unlike the buffer * for basic samples, AUX buffer belongs to the event. It is scheduled with * the task among online cpus when it is a per-thread event. * * Return the private AUX buffer structure if success or NULL if fails.
*/ staticvoid *aux_buffer_setup(struct perf_event *event, void **pages, int nr_pages, bool snapshot)
{ struct sf_buffer *sfb; struct aux_buffer *aux; unsignedlong *new, *tail; int i, n_sdbt;
if (!nr_pages || !pages) return NULL;
if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
pr_err("AUX buffer size (%i pages) is larger than the " "maximum sampling buffer limit\n",
nr_pages); return NULL;
} elseif (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
pr_err("AUX buffer size (%i pages) is less than the " "minimum sampling buffer limit\n",
nr_pages); return NULL;
}
/* Allocate aux_buffer struct for the event */
aux = kzalloc(sizeof(struct aux_buffer), GFP_KERNEL); if (!aux) goto no_aux;
sfb = &aux->sfb;
/* Allocate sdbt_index for fast reference */
n_sdbt = DIV_ROUND_UP(nr_pages, CPUM_SF_SDB_PER_TABLE);
aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL); if (!aux->sdbt_index) goto no_sdbt_index;
/* Allocate sdb_index for fast reference */
aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL); if (!aux->sdb_index) goto no_sdb_index;
/* Allocate the first SDBT */
sfb->num_sdbt = 0;
sfb->sdbt = (unsignedlong *)get_zeroed_page(GFP_KERNEL); if (!sfb->sdbt) goto no_sdbt;
aux->sdbt_index[sfb->num_sdbt++] = (unsignedlong)sfb->sdbt;
tail = sfb->tail = sfb->sdbt;
/* * Link the provided pages of AUX buffer to SDBT. * Allocate SDBT if needed.
*/ for (i = 0; i < nr_pages; i++, tail++) { if (require_table_link(tail)) { new = (unsignedlong *)get_zeroed_page(GFP_KERNEL); if (!new) goto no_sdbt;
aux->sdbt_index[sfb->num_sdbt++] = (unsignedlong)new; /* Link current page to tail of chain */
*tail = virt_to_phys(new) + 1;
tail = new;
} /* Tail is the entry in a SDBT */
*tail = virt_to_phys(pages[i]);
aux->sdb_index[i] = (unsignedlong)pages[i];
aux_sdb_init((unsignedlong)pages[i]);
}
sfb->num_sdb = nr_pages;
/* Link the last entry in the SDBT to the first SDBT */
*tail = virt_to_phys(sfb->sdbt) + 1;
sfb->tail = tail;
/* * Initial all SDBs are zeroed. Mark it as empty. * So there is no need to clear the full indicator * when this event is first added.
*/
aux->empty_mark = sfb->num_sdb - 1;
return aux;
no_sdbt: /* SDBs (AUX buffer pages) are freed by caller */ for (i = 0; i < sfb->num_sdbt; i++)
free_page(aux->sdbt_index[i]);
kfree(aux->sdb_index);
no_sdb_index:
kfree(aux->sdbt_index);
no_sdbt_index:
kfree(aux);
no_aux: return NULL;
}
staticvoid cpumsf_pmu_read(struct perf_event *event)
{ /* Nothing to do ... updates are interrupt-driven */
}
/* Check if the new sampling period/frequency is appropriate. * * Return non-zero on error and zero on passed checks.
*/ staticint cpumsf_pmu_check_period(struct perf_event *event, u64 value)
{ struct hws_qsi_info_block si; unsignedlong rate; bool do_freq;
memset(&si, 0, sizeof(si)); if (event->cpu == -1) {
qsi(&si);
} else { /* Event is pinned to a particular CPU, retrieve the per-CPU * sampling structure for accessing the CPU-specific QSI.
*/ struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
/* Set up sampling controls. Always program the sampling register * using the SDB-table start. Reset TEAR_REG event hardware register * that is used by hw_perf_event_update() to store the sampling buffer * position after samples have been flushed.
*/
cpuhw->lsctl.s = 0;
cpuhw->lsctl.h = 1;
cpuhw->lsctl.interval = SAMPL_RATE(&event->hw); if (!SAMPL_DIAG_MODE(&event->hw)) {
cpuhw->lsctl.tear = virt_to_phys(cpuhw->sfb.sdbt);
cpuhw->lsctl.dear = *(unsignedlong *)cpuhw->sfb.sdbt;
TEAR_REG(&event->hw) = (unsignedlong)cpuhw->sfb.sdbt;
}
/* Ensure sampling functions are in the disabled state. If disabled,
* switch on sampling enable control. */ if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
err = -EAGAIN; goto out;
} if (SAMPL_DIAG_MODE(&event->hw)) {
aux = perf_aux_output_begin(&cpuhw->handle, event); if (!aux) {
err = -EINVAL; goto out;
}
err = aux_output_begin(&cpuhw->handle, aux, cpuhw); if (err) goto out;
cpuhw->lsctl.ed = 1;
}
cpuhw->lsctl.es = 1;
/* Set in_use flag and store event */
cpuhw->event = event;
cpuhw->flags |= PMU_F_IN_USE;
/* Attribute list for CPU_SF. * * The availablitiy depends on the CPU_MF sampling facility authorization * for basic + diagnositic samples. This is determined at initialization * time by the sampling facility device driver. * If the authorization for basic samples is turned off, it should be * also turned off for diagnostic sampling. * * During initialization of the device driver, check the authorization * level for diagnostic sampling and installs the attribute * file for diagnostic sampling if necessary. * * For now install a placeholder to reference all possible attributes: * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG. * Add another entry for the final NULL pointer.
*/ enum {
SF_CYCLES_BASIC_ATTR_IDX = 0,
SF_CYCLES_BASIC_DIAG_ATTR_IDX,
SF_CYCLES_ATTR_MAX
};
if (!(alert & CPU_MF_INT_SF_MASK)) return;
inc_irq_stat(IRQEXT_CMS);
cpuhw = this_cpu_ptr(&cpu_hw_sf);
/* Measurement alerts are shared and might happen when the PMU
* is not reserved. Ignore these alerts in this case. */ if (!(cpuhw->flags & PMU_F_RESERVED)) return;
/* The processing below must take care of multiple alert events that
* might be indicated concurrently. */
/* Program alert request */ if (alert & CPU_MF_INT_SF_PRA) { if (cpuhw->flags & PMU_F_IN_USE) { if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
hw_collect_aux(cpuhw); else
hw_perf_event_update(cpuhw->event, 0);
}
}
/* Report measurement alerts only for non-PRA codes */ if (alert != CPU_MF_INT_SF_PRA)
debug_sprintf_event(sfdbg, 6, "%s alert %#x\n", __func__,
alert);
/* Loss of sample data due to high-priority machine activities */ if (alert & CPU_MF_INT_SF_LSDA) {
pr_err("Sample data was lost\n");
cpuhw->flags |= PMU_F_ERR_LSDA;
sf_disable();
}
staticint cpusf_pmu_setup(unsignedint cpu, int flags)
{ /* Ignore the notification if no events are scheduled on the PMU. * This might be racy...
*/ if (!refcount_read(&num_events)) return 0;
staticvoid __init pr_cpumsf_err(unsignedint reason)
{
pr_err("Sampling facility support for perf is not available: " "reason %#x\n", reason);
}
staticint __init init_cpum_sampling_pmu(void)
{ struct hws_qsi_info_block si; int err;
if (!cpum_sf_avail()) return -ENODEV;
memset(&si, 0, sizeof(si));
qsi(&si); if (!si.as && !si.ad) return -ENODEV;
if (si.bsdes != sizeof(struct hws_basic_entry)) {
pr_cpumsf_err(RS_INIT_FAILURE_BSDES); return -EINVAL;
}
if (si.ad) {
sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB); /* Sampling of diagnostic data authorized, * install event into attribute list of PMU device.
*/
cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] =
CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
}
sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80); if (!sfdbg) {
pr_err("Registering for s390dbf failed\n"); return -ENOMEM;
}
debug_register_view(sfdbg, &debug_sprintf_view);
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.
