system/dev/misc/cpu-trace/intel-pt.cpp

// Copyright 2016 The Fuchsia Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

// See the README.md in this directory for documentation.

#include <assert.h>
#include <cpuid.h>
#include <inttypes.h>
#include <limits.h>
#include <memory>
#include <new>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <threads.h>

#include <ddk/binding.h>
#include <ddk/debug.h>
#include <ddk/device.h>
#include <ddk/driver.h>
#include <ddk/io-buffer.h>
#include <ddk/platform-defs.h>
#include <ddk/protocol/platform/device.h>
#include <ddktl/device.h>

#include <fbl/alloc_checker.h>

#include <lib/zircon-internal/device/cpu-trace/intel-pt.h>
#include <lib/zircon-internal/mtrace.h>
#include <lib/zx/bti.h>
#include <zircon/status.h>
#include <zircon/syscalls.h>
#include <zircon/syscalls/resource.h>
#include <zircon/types.h>

namespace insntrace {

typedef enum {
    // IPT_MODE_* are the public names, we don't assume the values identical
    IPT_TRACE_CPUS,
    IPT_TRACE_THREADS
} ipt_trace_mode_t;

typedef struct ipt_per_trace_state {
    // the cpu or thread this buffer is assigned to
    // Which value to use is determined by the trace mode.
    union {
        uint32_t cpu;
        zx_handle_t thread;
    } owner;

    // number of chunks, each 2^|chunk_order| pages in size
    uint32_t num_chunks;
    // log2 size of each chunk, in pages
    uint32_t chunk_order;
    // if true then the buffer is circular, otherwise tracing stops when the
    // buffer fills
    bool is_circular;
    // true if allocated
    bool allocated;
    // true if buffer is assigned to a cpu/thread
    bool assigned;
    // number of ToPA tables needed
    uint32_t num_tables;

    // msrs
    uint64_t ctl;
    uint64_t status;
    uint64_t output_base;
    uint64_t output_mask_ptrs;
    uint64_t cr3_match;
    struct {
        uint64_t a,b;
    } addr_ranges[IPT_MAX_NUM_ADDR_RANGES];

    // trace buffers and ToPA tables
    // ToPA: Table of Physical Addresses
    // A "trace buffer" is a set of N chunks.
    std::unique_ptr<io_buffer_t[]> chunks;
    std::unique_ptr<io_buffer_t[]> topas;
} ipt_per_trace_state_t;

// TODO(dje): add unbindable?
class InsntraceDevice;
using DeviceType = ddk::Device<InsntraceDevice,
                               ddk::Openable,
                               ddk::Closable,
                               ddk::Ioctlable>;

class InsntraceDevice : public DeviceType {
  public:
    explicit InsntraceDevice(zx_device_t* parent, zx::bti bti)
        : DeviceType(parent), bti_(std::move(bti)) {}
    ~InsntraceDevice() = default;

    void DdkRelease();

    // Device protocol implementation
    zx_status_t DdkOpen(zx_device_t** dev_out, uint32_t flags);
    zx_status_t DdkClose(uint32_t flags);
    zx_status_t DdkIoctl(uint32_t op, const void* in_buf, size_t in_len,
                         void* out_buf, size_t out_len, size_t* out_actual);

  private:
    // Low level routines
    void MakeTopa(ipt_per_trace_state_t* per_trace);
    uint32_t ComputeTopaEntryCount(ipt_per_trace_state_t* per_trace);
    size_t ComputeCaptureSize(const ipt_per_trace_state_t* per_trace);
    zx_status_t X86PtAllocBuffer1(ipt_per_trace_state_t* per_trace,
                                  uint32_t num, uint32_t order,
                                  bool is_circular);
    void X86PtFreeBuffer1(ipt_per_trace_state_t* per_trace);
    zx_status_t X86PtAllocBuffer(const ioctl_insntrace_buffer_config_t* config,
                                 zx_itrace_buffer_descriptor_t* out_descriptor);
    zx_status_t X86PtAssignThreadBuffer(zx_itrace_buffer_descriptor_t descriptor,
                                        zx_handle_t thread);
    zx_status_t X86PtReleaseThreadBuffer(zx_itrace_buffer_descriptor_t descriptor,
                                         zx_handle_t thread);
    zx_status_t X86PtFreeBuffer(zx_itrace_buffer_descriptor_t descriptor);
    zx_status_t X86PtStageTraceData(zx_handle_t resource,
                                    zx_itrace_buffer_descriptor_t descriptor);
    zx_status_t X86PtGetTraceData(zx_handle_t resource,
                                  zx_itrace_buffer_descriptor_t descriptor);

    // Ioctl handlers
    zx_status_t IptAllocTrace(const void* cmd, size_t cmdlen);
    zx_status_t IptFreeTrace();
    zx_status_t IptGetTraceConfig(void* reply, size_t replymax,
                                  size_t* out_actual);
    zx_status_t IptAllocBuffer(const void* cmd, size_t cmdlen,
                               void* reply, size_t replymax,
                               size_t* out_actual);
    zx_status_t IptAssignThreadBuffer(const void* cmd, size_t cmdlen);
    zx_status_t IptReleaseThreadBuffer(const void* cmd, size_t cmdlen);
    zx_status_t IptGetBufferConfig(const void* cmd, size_t cmdlen,
                                   void* reply, size_t replymax,
                                   size_t* out_actual);
    zx_status_t IptGetBufferInfo(const void* cmd, size_t cmdlen,
                                 void* reply, size_t replymax,
                                 size_t* out_actual);
    zx_status_t IptGetChunkHandle(const void* cmd, size_t cmdlen,
                                  void* reply, size_t replymax,
                                  size_t* out_actual);
    zx_status_t IptFreeBuffer(const void* cmd, size_t cmdlen);
    zx_status_t IptStart();
    zx_status_t IptStop();
    zx_status_t IoctlWorker(uint32_t op, const void* cmd, size_t cmdlen,
                            void* reply, size_t replymax,
                            size_t* out_actual);

    mtx_t lock_{};

    // Only one open of this device is supported at a time. KISS for now.
    bool opened_ = false;

    ipt_trace_mode_t mode_ = IPT_TRACE_CPUS;

    // # of entries in |per_trace_state|.
    // When tracing by cpu, this is the max number of cpus.
    // When tracing by thread, this is the max number of threads.
    // TODO(dje): Add support for dynamically growing the vector.
    uint32_t num_traces_ = 0;

    // one entry for each trace
    std::unique_ptr<ipt_per_trace_state_t[]> per_trace_state_;

    // Once tracing has started various things are not allowed until it stops.
    bool active_ = false;

    zx::bti bti_;
};

static uint32_t ipt_config_family;
static uint32_t ipt_config_model;
static uint32_t ipt_config_stepping;

static uint32_t ipt_config_addr_cfg_max = 0;
static uint32_t ipt_config_mtc_freq_mask = 0;
static uint32_t ipt_config_cyc_thresh_mask = 0;
static uint32_t ipt_config_psb_freq_mask = 0;
static uint32_t ipt_config_num_addr_ranges = 0;
static uint32_t ipt_config_bus_freq = 0;

static bool ipt_config_supported = false;

static bool ipt_config_cr3_filtering = false;
static bool ipt_config_psb = false;
static bool ipt_config_ip_filtering = false;
static bool ipt_config_mtc = false;
static bool ipt_config_ptwrite = false;
static bool ipt_config_power_events = false;
static bool ipt_config_output_topa = false;
static bool ipt_config_output_topa_multi = false;
static bool ipt_config_output_single = false;
static bool ipt_config_output_transport = false;
static bool ipt_config_lip = false;

// maximum space, in bytes, for trace buffers (per cpu)
// This isn't necessarily
// MAX_NUM_CHUNKS * (1 << (MAX_CHUNK_ORDER + PAGE_SIZE_SHIFT)).
// Buffers have to be naturally aligned contiguous pages, but we can have
// a lot of them. Supporting large buffers and/or lots of them is for
// experimentation.
#define MAX_PER_TRACE_SPACE (256 * 1024 * 1024)

// maximum number of buffers
#define MAX_NUM_CHUNKS 4096

// maximum size of each buffer, in pages (1MB)
#define MAX_CHUNK_ORDER 8

#if PAGE_SIZE == 4096
#define PAGE_SIZE_SHIFT 12
#else
#error "unsupported page size"
#endif

#define BIT(x, b) ((x) & (1u << (b)))


// The userspace side of the driver

static zx_status_t InsntraceInitOnce() {
    unsigned a, b, c, d, max_leaf;

    max_leaf = __get_cpuid_max(0, nullptr);
    if (max_leaf < 0x14) {
        zxlogf(INFO, "IntelPT: No PT support\n");
        return ZX_ERR_NOT_SUPPORTED;
    }

    __cpuid(1, a, b, c, d);
    ipt_config_stepping = a & 0xf;
    ipt_config_model = (a >> 4) & 0xf;
    ipt_config_family = (a >> 8) & 0xf;
    if (ipt_config_family == 0xf)
        ipt_config_family += (a >> 20) & 0xff;
    if (ipt_config_family == 6 || ipt_config_family == 0xf)
        ipt_config_model += ((a >> 16) & 0xf) << 4;

    __cpuid_count(0x07, 0, a, b, c, d);
    if (!BIT(b, 25)) {
        zxlogf(INFO, "IntelPT: No PT support\n");
        return ZX_ERR_NOT_SUPPORTED;
    }

    ipt_config_supported = true;

    __cpuid_count(0x14, 0, a, b, c, d);
    if (BIT(b, 2))
        ipt_config_addr_cfg_max = 2;
    if (BIT(b, 1) && a >= 1) {
        unsigned a1, b1, c1, d1;
        __cpuid_count(0x14, 1, a1, b1, c1, d1);
        ipt_config_mtc_freq_mask = (a1 >> 16) & 0xffff;
        ipt_config_cyc_thresh_mask = b1 & 0xffff;
        ipt_config_psb_freq_mask = (b1 >> 16) & 0xffff;
        ipt_config_num_addr_ranges = a1 & 0x7;
    }

    if (max_leaf >= 0x15) {
        unsigned a1 = 0, b1 = 0, c1 = 0, d1 = 0;
        __cpuid(0x15, a1, b1, c1, d1);
        if (a1 && b1)
            ipt_config_bus_freq = static_cast<uint32_t>(1. / ((float)a1 / (float)b1));
    }

    ipt_config_cr3_filtering = !!BIT(b, 0);
    ipt_config_psb = !!BIT(b, 1);
    ipt_config_ip_filtering = !!BIT(b, 2);
    ipt_config_mtc = !!BIT(b, 3);
    ipt_config_ptwrite = !!BIT(b, 4);
    ipt_config_power_events = !!BIT(b, 5);

    ipt_config_output_topa = !!BIT(c, 0);
    ipt_config_output_topa_multi = !!BIT(c, 1);
    ipt_config_output_single = !!BIT(c, 2);
    ipt_config_output_transport = !!BIT(c, 3);
    ipt_config_lip = !!BIT(c, 31);

    zxlogf(INFO, "Intel Processor Trace configuration for this chipset:\n");
    // No need to print everything, but these are useful.
    zxlogf(INFO, "mtc_freq_mask:   0x%x\n", ipt_config_mtc_freq_mask);
    zxlogf(INFO, "cyc_thresh_mask: 0x%x\n", ipt_config_cyc_thresh_mask);
    zxlogf(INFO, "psb_freq_mask:   0x%x\n", ipt_config_psb_freq_mask);
    zxlogf(INFO, "num addr ranges: %u\n", ipt_config_num_addr_ranges);

    return ZX_OK;
}

// Create the ToPA for the configured number of pages for |cpu|.
// A circular collection of buffers is set up, even if we're going to apply
// the stop bit to the last entry.
void InsntraceDevice::MakeTopa(ipt_per_trace_state_t* per_trace) {
    const size_t run_len_log2 = per_trace->chunk_order;
    assert(run_len_log2 + PAGE_SIZE_SHIFT <= IPT_TOPA_MAX_SHIFT);
    assert(run_len_log2 + PAGE_SIZE_SHIFT >= IPT_TOPA_MIN_SHIFT);

    uint32_t curr_table = 0;
    uint32_t curr_idx = 0;
    uint64_t* last_entry = nullptr;

    // Note: An early version of this patch auto-computed the desired grouping
    // of pages with sufficient alignment. If you find yourself needing this
    // functionality again, see change 9470.

    for (uint32_t i = 0; i < per_trace->num_chunks; ++i) {
        io_buffer_t* buffer = &per_trace->chunks[i];
        io_buffer_t* topa = &per_trace->topas[curr_table];
        zx_paddr_t pa = io_buffer_phys(buffer);

        uint64_t val = IPT_TOPA_ENTRY_PHYS_ADDR(pa) |
            IPT_TOPA_ENTRY_SIZE(run_len_log2 + PAGE_SIZE_SHIFT);
        auto table = reinterpret_cast<uint64_t*>(io_buffer_virt(topa));
        table[curr_idx] = val;
        last_entry = &table[curr_idx];

        // Make sure we leave one at the end of the table for the END marker.
        if (unlikely(curr_idx >= IPT_TOPA_MAX_TABLE_ENTRIES - 2)) {
            curr_idx = 0;
            curr_table++;
        } else {
            curr_idx++;
        }
    }

    assert(curr_table + 1 == per_trace->num_tables ||
           // If the last table is full curr_table will be the next one.
           (curr_table == per_trace->num_tables && curr_idx == 0));

    // Populate END entries for completed tables
    // Assume the table is circular. We'll set the stop bit on the last
    // entry later.
    for (uint32_t i = 0; i < curr_table; ++i) {
        io_buffer_t* this_table = &per_trace->topas[i];
        io_buffer_t* next_table;
        if (i == per_trace->num_tables - 1) {
            next_table = &per_trace->topas[0];
        } else {
            next_table = &per_trace->topas[i + 1];
        }

        zx_paddr_t next_table_pa = io_buffer_phys(next_table);
        uint64_t val = IPT_TOPA_ENTRY_PHYS_ADDR(next_table_pa) | IPT_TOPA_ENTRY_END;
        auto table = reinterpret_cast<uint64_t*>(io_buffer_virt(this_table));
        table[IPT_TOPA_MAX_TABLE_ENTRIES - 1] = val;
    }

    // Populate the END entry for a possibly non-full last table
    if (curr_table < per_trace->num_tables) {
        io_buffer_t* this_table = &per_trace->topas[curr_table];
        io_buffer_t* first_table = &per_trace->topas[0];
        zx_paddr_t first_table_pa = io_buffer_phys(first_table);
        uint64_t val = IPT_TOPA_ENTRY_PHYS_ADDR(first_table_pa) | IPT_TOPA_ENTRY_END;
        auto table = reinterpret_cast<uint64_t*>(io_buffer_virt(this_table));
        table[curr_idx] = val;
    }

    // Add the STOP flag to the last non-END entry in the tables
    assert(last_entry);
    if (!per_trace->is_circular)
        *last_entry |= IPT_TOPA_ENTRY_STOP;
}

// Compute the number of ToPA entries needed for the configured number of
// buffers.
// The output count includes the END entries across all needed tables.
uint32_t InsntraceDevice::ComputeTopaEntryCount(ipt_per_trace_state_t* per_trace) {
    uint32_t num_entries = per_trace->num_chunks;
    uint32_t num_end_entries = (num_entries + IPT_TOPA_MAX_TABLE_ENTRIES - 2) /
        (IPT_TOPA_MAX_TABLE_ENTRIES - 1);
    uint32_t result = num_entries + num_end_entries;

    zxlogf(DEBUG1, "IPT: compute_topa_entry_count: num_entries: %u\n", num_entries);
    zxlogf(DEBUG1, "IPT: compute_topa_entry_count: num_end_entries: %u\n", num_end_entries);
    zxlogf(DEBUG1, "IPT: compute_topa_entry_count: total entries: %u\n", result);

    return result;
}

// Walk the tables to discover how much data has been captured for |per_trace|.
// Note: If this is a circular buffer this is just where tracing stopped.
size_t InsntraceDevice::ComputeCaptureSize(const ipt_per_trace_state_t* per_trace) {
    uint64_t curr_table_paddr = per_trace->output_base;
    uint32_t curr_table_entry_idx = (uint32_t)per_trace->output_mask_ptrs >> 7;
    uint32_t curr_entry_offset = (uint32_t)(per_trace->output_mask_ptrs >> 32);

    zxlogf(DEBUG1, "IPT: compute_capture_size: trace %tu\n", per_trace - per_trace_state_.get());
    zxlogf(DEBUG1, "IPT: curr_table_paddr 0x%" PRIx64 ", curr_table_entry_idx %u, curr_entry_offset %u\n",
           curr_table_paddr, curr_table_entry_idx, curr_entry_offset);

    size_t total_size = 0;
    for (uint32_t table = 0; table < per_trace->num_tables; ++table) {
        // Get the physical address so that we can compare it with the value
        // in output_base.
        zx_paddr_t table_paddr = io_buffer_phys(&per_trace->topas[table]);

        for (uint32_t entry = 0; entry < IPT_TOPA_MAX_TABLE_ENTRIES - 1; ++entry) {
            if (table_paddr == curr_table_paddr && entry >= curr_table_entry_idx) {
                total_size += curr_entry_offset;
                return total_size;
            }
            auto table_ptr = reinterpret_cast<uint64_t*>(io_buffer_virt(&per_trace->topas[table]));
            uint64_t topa_entry = table_ptr[entry];
            total_size += 1UL << IPT_TOPA_ENTRY_EXTRACT_SIZE(topa_entry);
        }
    }

    // Should be unreachable.
    // TODO(dje): Later flag state as broken.
    zxlogf(ERROR, "IPT: unexpectedly exited capture loop\n");
    return 0;
}

zx_status_t InsntraceDevice::X86PtAllocBuffer1(ipt_per_trace_state_t* per_trace,
                                               uint32_t num, uint32_t order,
                                               bool is_circular) {
    zx_status_t status;
    size_t chunk_pages = 1 << order;

    fbl::AllocChecker ac;
    per_trace->chunks = std::unique_ptr<io_buffer_t[]>(new (&ac) io_buffer_t[num]{});
    if (!ac.check()) {
        return ZX_ERR_NO_MEMORY;
    }

    for (uint32_t i = 0; i < num; ++i) {
        // ToPA entries of size N must be aligned to N, too.
        uint32_t alignment_log2 = PAGE_SIZE_SHIFT + order;
        status = io_buffer_init_aligned(&per_trace->chunks[i], bti_.get(),
                                        chunk_pages * PAGE_SIZE, alignment_log2,
                                        IO_BUFFER_RW | IO_BUFFER_CONTIG);
        if (status != ZX_OK)
            return status;
        // Keep track of allocated buffers as we go in case we later fail:
        // we want to be able to free those that got allocated.
        ++per_trace->num_chunks;
        // Catch bugs in io_buffer_init_aligned. If it doesn't give us a
        // properly aligned buffer we'll get an "operational error" later.
        // See Intel Vol3 36.2.6.2.
        zx_paddr_t pa = io_buffer_phys(&per_trace->chunks[i]);
        zx_paddr_t align_mask = (1ull << alignment_log2) - 1;
        if (pa & align_mask) {
            zxlogf(ERROR, "%s: WARNING: chunk has bad alignment: alignment %u, got 0x%" PRIx64 "\n",
                   __func__, alignment_log2, pa);
            return ZX_ERR_INTERNAL;
        }
    }
    assert(per_trace->num_chunks == num);

    per_trace->chunk_order = order;
    per_trace->is_circular = is_circular;

    // TODO(dje): No need to allocate the max on the last table.
    uint32_t entry_count = ComputeTopaEntryCount(per_trace);
    uint32_t table_count = (entry_count + IPT_TOPA_MAX_TABLE_ENTRIES - 1) /
            IPT_TOPA_MAX_TABLE_ENTRIES;

    if (entry_count < 2) {
        zxlogf(INFO, "IPT: INVALID ENTRY COUNT: %u\n", entry_count);
        return ZX_ERR_INVALID_ARGS;
    }

    // Some early Processor Trace implementations only supported having a
    // table with a single real entry and an END.
    if (!ipt_config_output_topa_multi && entry_count > 2)
        return ZX_ERR_NOT_SUPPORTED;

    // Allocate Table(s) of Physical Addresses (ToPA) for each cpu.

    per_trace->topas = std::unique_ptr<io_buffer_t[]>(new (&ac) io_buffer_t[table_count]{});
    if (!ac.check()) {
        return ZX_ERR_NO_MEMORY;
    }

    for (uint32_t i = 0; i < table_count; ++i) {
        status = io_buffer_init(&per_trace->topas[i], bti_.get(),
                                sizeof(uint64_t) * IPT_TOPA_MAX_TABLE_ENTRIES,
                                IO_BUFFER_RW | IO_BUFFER_CONTIG);
        if (status != ZX_OK)
            return ZX_ERR_NO_MEMORY;
        // Keep track of allocated tables as we go in case we later fail:
        // we want to be able to free those that got allocated.
        ++per_trace->num_tables;
    }
    assert(per_trace->num_tables == table_count);

    MakeTopa(per_trace);

    return ZX_OK;
}

void InsntraceDevice::X86PtFreeBuffer1(ipt_per_trace_state_t* per_trace) {
    assert(!per_trace->assigned);

    if (per_trace->chunks) {
        for (uint32_t i = 0; i < per_trace->num_chunks; ++i) {
            io_buffer_release(&per_trace->chunks[i]);
        }
    }
    per_trace->chunks.reset();

    if (per_trace->topas) {
        for (uint32_t i = 0; i < per_trace->num_tables; ++i) {
            io_buffer_release(&per_trace->topas[i]);
        }
    }
    per_trace->topas.reset();

    per_trace->allocated = false;
}

zx_status_t InsntraceDevice::X86PtAllocBuffer(const ioctl_insntrace_buffer_config_t* config,
                                              zx_itrace_buffer_descriptor_t* out_descriptor) {
    zxlogf(DEBUG1, "%s: num_chunks %u, chunk_order %u\n",
           __func__, config->num_chunks, config->chunk_order);

    if (config->num_chunks == 0 || config->num_chunks > MAX_NUM_CHUNKS)
        return ZX_ERR_INVALID_ARGS;
    if (config->chunk_order > MAX_CHUNK_ORDER)
        return ZX_ERR_INVALID_ARGS;
    size_t chunk_pages = 1 << config->chunk_order;
    size_t nr_pages = config->num_chunks * chunk_pages;
    size_t total_per_trace = nr_pages * PAGE_SIZE;
    if (total_per_trace > MAX_PER_TRACE_SPACE)
        return ZX_ERR_INVALID_ARGS;

    uint64_t settable_ctl_mask = (
        IPT_CTL_OS_ALLOWED_MASK |
        IPT_CTL_USER_ALLOWED_MASK |
        IPT_CTL_TSC_EN_MASK |
        IPT_CTL_DIS_RETC_MASK |
        IPT_CTL_BRANCH_EN_MASK
        );
    if (ipt_config_ptwrite)
        settable_ctl_mask |= IPT_CTL_PTW_EN_MASK | IPT_CTL_FUP_ON_PTW_MASK;
    if (ipt_config_cr3_filtering)
        settable_ctl_mask |= IPT_CTL_CR3_FILTER_MASK;
    if (ipt_config_mtc)
        settable_ctl_mask |= IPT_CTL_MTC_EN_MASK | IPT_CTL_MTC_FREQ_MASK;
    if (ipt_config_power_events)
        settable_ctl_mask |= IPT_CTL_POWER_EVENT_EN_MASK;
    if (ipt_config_ip_filtering) {
        if (ipt_config_num_addr_ranges >= 1)
            settable_ctl_mask |= IPT_CTL_ADDR0_MASK;
        if (ipt_config_num_addr_ranges >= 2)
            settable_ctl_mask |= IPT_CTL_ADDR1_MASK;
        if (ipt_config_num_addr_ranges >= 3)
            settable_ctl_mask |= IPT_CTL_ADDR2_MASK;
        if (ipt_config_num_addr_ranges >= 4)
            settable_ctl_mask |= IPT_CTL_ADDR3_MASK;
    }
    if (ipt_config_psb)
        settable_ctl_mask |= (IPT_CTL_CYC_EN_MASK |
                              IPT_CTL_PSB_FREQ_MASK |
                              IPT_CTL_CYC_THRESH_MASK);
    if ((config->ctl & ~settable_ctl_mask) != 0) {
        zxlogf(ERROR, "bad ctl, requested 0x%" PRIx64 ", valid 0x%" PRIx64 "\n",
               config->ctl, settable_ctl_mask);
        return ZX_ERR_INVALID_ARGS;
    }

    uint32_t mtc_freq = (uint32_t) ((config->ctl & IPT_CTL_MTC_FREQ_MASK) >> IPT_CTL_MTC_FREQ_SHIFT);
    if (mtc_freq != 0 && ((1 << mtc_freq) & ipt_config_mtc_freq_mask) == 0) {
        zxlogf(ERROR, "bad mtc_freq value, requested 0x%x, valid mask 0x%x\n",
               mtc_freq, ipt_config_mtc_freq_mask);
        return ZX_ERR_INVALID_ARGS;
    }
    uint32_t cyc_thresh = (uint32_t) ((config->ctl & IPT_CTL_CYC_THRESH_MASK) >> IPT_CTL_CYC_THRESH_SHIFT);
    if (cyc_thresh != 0 && ((1 << cyc_thresh) & ipt_config_cyc_thresh_mask) == 0) {
        zxlogf(ERROR, "bad cyc_thresh value, requested 0x%x, valid mask 0x%x\n",
               cyc_thresh, ipt_config_cyc_thresh_mask);
        return ZX_ERR_INVALID_ARGS;
    }
    uint32_t psb_freq = (uint32_t) ((config->ctl & IPT_CTL_PSB_FREQ_MASK) >> IPT_CTL_PSB_FREQ_SHIFT);
    if (psb_freq != 0 && ((1 << psb_freq) & ipt_config_psb_freq_mask) == 0) {
        zxlogf(ERROR, "bad psb_freq value, requested 0x%x, valid mask 0x%x\n",
               psb_freq, ipt_config_psb_freq_mask);
        return ZX_ERR_INVALID_ARGS;
    }

    // Find an unallocated buffer entry.
    zx_itrace_buffer_descriptor_t descriptor;
    for (descriptor = 0; descriptor < num_traces_; ++descriptor) {
        if (!per_trace_state_[descriptor].allocated)
            break;
    }
    if (descriptor == num_traces_)
        return ZX_ERR_NO_RESOURCES;

    ipt_per_trace_state_t* per_trace = &per_trace_state_[descriptor];
    zx_status_t status = X86PtAllocBuffer1(per_trace, config->num_chunks,
                                           config->chunk_order, config->is_circular);
    if (status != ZX_OK) {
        X86PtFreeBuffer1(per_trace);
        return status;
    }

    per_trace->ctl = config->ctl;
    per_trace->status = 0;
    per_trace->output_base = io_buffer_phys(&per_trace->topas[0]);
    per_trace->output_mask_ptrs = 0;
    per_trace->cr3_match = config->cr3_match;
    static_assert(sizeof(per_trace->addr_ranges) == sizeof(config->addr_ranges),
                  "addr range size mismatch");
    memcpy(per_trace->addr_ranges, config->addr_ranges, sizeof(config->addr_ranges));
    per_trace->allocated = true;
    *out_descriptor = descriptor;
    return ZX_OK;
}

zx_status_t InsntraceDevice::X86PtAssignThreadBuffer(zx_itrace_buffer_descriptor_t descriptor,
                                                     zx_handle_t thread) {
    zx_handle_close(thread);
    // TODO(dje): Thread support is still work-in-progress.
    return ZX_ERR_NOT_SUPPORTED;
}

zx_status_t InsntraceDevice::X86PtReleaseThreadBuffer(zx_itrace_buffer_descriptor_t descriptor,
                                                      zx_handle_t thread) {
    zx_handle_close(thread);
    // TODO(dje): Thread support is still work-in-progress.
    return ZX_ERR_NOT_SUPPORTED;
}

zx_status_t InsntraceDevice::X86PtFreeBuffer(zx_itrace_buffer_descriptor_t descriptor) {
    if (active_)
        return ZX_ERR_BAD_STATE;
    if (descriptor >= num_traces_)
        return ZX_ERR_INVALID_ARGS;
    assert(per_trace_state_);
    ipt_per_trace_state_t* per_trace = &per_trace_state_[descriptor];

    if (!per_trace->allocated)
        return ZX_ERR_INVALID_ARGS;
    if (per_trace->assigned)
        return ZX_ERR_BAD_STATE;
    X86PtFreeBuffer1(per_trace);
    return ZX_OK;
}

zx_status_t InsntraceDevice::X86PtStageTraceData(zx_handle_t resource,
                                                 zx_itrace_buffer_descriptor_t descriptor) {
    if (descriptor >= num_traces_)
        return ZX_ERR_INVALID_ARGS;
    assert(per_trace_state_);
    const ipt_per_trace_state_t* per_trace = &per_trace_state_[descriptor];

    zx_x86_pt_regs_t regs;
    regs.ctl = per_trace->ctl;
    regs.ctl |= IPT_CTL_TOPA_MASK | IPT_CTL_TRACE_EN_MASK;
    regs.status = per_trace->status;
    regs.output_base = per_trace->output_base;
    regs.output_mask_ptrs = per_trace->output_mask_ptrs;
    regs.cr3_match = per_trace->cr3_match;
    static_assert(sizeof(regs.addr_ranges) == sizeof(per_trace->addr_ranges),
                  "addr range size mismatch");
    memcpy(regs.addr_ranges, per_trace->addr_ranges, sizeof(per_trace->addr_ranges));

    return zx_mtrace_control(resource, MTRACE_KIND_INSNTRACE,
                             MTRACE_INSNTRACE_STAGE_TRACE_DATA,
                             descriptor, &regs, sizeof(regs));
}

zx_status_t InsntraceDevice::X86PtGetTraceData(zx_handle_t resource,
                                               zx_itrace_buffer_descriptor_t descriptor) {
    if (descriptor >= num_traces_)
        return ZX_ERR_INVALID_ARGS;
    assert(per_trace_state_);
    ipt_per_trace_state_t* per_trace = &per_trace_state_[descriptor];

    zx_x86_pt_regs_t regs;
    zx_status_t status = zx_mtrace_control(resource, MTRACE_KIND_INSNTRACE,
                                           MTRACE_INSNTRACE_GET_TRACE_DATA,
                                           descriptor, &regs, sizeof(regs));
    if (status != ZX_OK)
        return status;
    per_trace->ctl = regs.ctl;
    per_trace->status = regs.status;
    per_trace->output_base = regs.output_base;
    per_trace->output_mask_ptrs = regs.output_mask_ptrs;
    per_trace->cr3_match = regs.cr3_match;
    static_assert(sizeof(per_trace->addr_ranges) == sizeof(regs.addr_ranges),
                  "addr range size mismatch");
    memcpy(per_trace->addr_ranges, regs.addr_ranges, sizeof(regs.addr_ranges));

    return ZX_OK;
}


// ioctl handlers

zx_status_t InsntraceDevice::IptAllocTrace(const void* cmd, size_t cmdlen) {
    if (!ipt_config_supported)
        return ZX_ERR_NOT_SUPPORTED;
    // For now we only support ToPA, though there are no current plans to
    // support anything else.
    if (!ipt_config_output_topa)
        return ZX_ERR_NOT_SUPPORTED;
    if (per_trace_state_)
        return ZX_ERR_BAD_STATE;

    ioctl_insntrace_trace_config_t config;
    if (cmdlen != sizeof(config))
        return ZX_ERR_INVALID_ARGS;
    memcpy(&config, cmd, sizeof(config));

    // TODO(dje): Until thread tracing is supported.
    if (config.mode == IPT_MODE_THREADS)
        return ZX_ERR_NOT_SUPPORTED;

    ipt_trace_mode_t internal_mode;
    switch (config.mode) {
    case IPT_MODE_CPUS:
        internal_mode = IPT_TRACE_CPUS;
        break;
    case IPT_MODE_THREADS:
        internal_mode = IPT_TRACE_THREADS;
        break;
    default:
        return ZX_ERR_INVALID_ARGS;
    }

    if (config.num_traces > IPT_MAX_NUM_TRACES)
        return ZX_ERR_INVALID_ARGS;
    if (config.mode == IPT_MODE_CPUS) {
        // TODO(dje): KISS. No point in allowing anything else for now.
        if (config.num_traces != zx_system_get_num_cpus())
            return ZX_ERR_INVALID_ARGS;
    }

    num_traces_ = config.num_traces;
    fbl::AllocChecker ac;
    auto per_trace_state = new (&ac) ipt_per_trace_state_t[num_traces_]{};
    if (!ac.check()) {
        return ZX_ERR_NO_MEMORY;
    }
    per_trace_state_.reset(per_trace_state);

    zx_handle_t resource = get_root_resource();
    zx_status_t status =
        zx_mtrace_control(resource, MTRACE_KIND_INSNTRACE, MTRACE_INSNTRACE_ALLOC_TRACE, 0,
                          &config, sizeof(config));
    if (status != ZX_OK) {
        per_trace_state_.reset();
        return status;
    }

    mode_ = internal_mode;
    return ZX_OK;
}

zx_status_t InsntraceDevice::IptFreeTrace() {
    if (active_)
        return ZX_ERR_BAD_STATE;

    // Don't make any changes until we know it's going to work.
    for (uint32_t i = 0; i < num_traces_; ++i) {
        ipt_per_trace_state_t* per_trace = &per_trace_state_[i];
        if (per_trace->assigned)
            return ZX_ERR_BAD_STATE;
    }

    for (uint32_t i = 0; i < num_traces_; ++i) {
        ipt_per_trace_state_t* per_trace = &per_trace_state_[i];
        if (per_trace->allocated)
            X86PtFreeBuffer1(per_trace);
    }

    zx_handle_t resource = get_root_resource();
    zx_status_t status =
        zx_mtrace_control(resource, MTRACE_KIND_INSNTRACE, MTRACE_INSNTRACE_FREE_TRACE, 0, nullptr, 0);
    // TODO(dje): This really shouldn't fail. What to do?
    // For now flag things as busted and prevent further use.
    if (status != ZX_OK)
        return ZX_OK;

    per_trace_state_.reset();
    return ZX_OK;
}

zx_status_t InsntraceDevice::IptGetTraceConfig(void* reply, size_t replymax,
                                               size_t* out_actual) {
    ioctl_insntrace_trace_config_t config;
    if (replymax < sizeof(config))
        return ZX_ERR_BUFFER_TOO_SMALL;

    switch (mode_) {
        case IPT_TRACE_CPUS:
            config.mode = IPT_MODE_CPUS;
            break;
        case IPT_TRACE_THREADS:
            config.mode = IPT_MODE_THREADS;
            break;
        default:
            __UNREACHABLE;
    }
    memcpy(reply, &config, sizeof(config));
    *out_actual = sizeof(config);
    return ZX_OK;
}

zx_status_t InsntraceDevice::IptAllocBuffer(const void* cmd, size_t cmdlen,
                                            void* reply, size_t replymax,
                                            size_t* out_actual) {
    ioctl_insntrace_buffer_config_t config;
    if (cmdlen != sizeof(config))
        return ZX_ERR_INVALID_ARGS;
    memcpy(&config, cmd, sizeof(config));
    zx_itrace_buffer_descriptor_t descriptor;
    if (replymax < sizeof(descriptor))
        return ZX_ERR_BUFFER_TOO_SMALL;

    zx_status_t status = X86PtAllocBuffer(&config, &descriptor);
    if (status != ZX_OK)
        return status;
    memcpy(reply, &descriptor, sizeof(descriptor));
    *out_actual = sizeof(descriptor);
    return ZX_OK;
}

zx_status_t InsntraceDevice::IptAssignThreadBuffer(const void* cmd, size_t cmdlen) {
    ioctl_insntrace_assign_thread_buffer_t assign;
    if (cmdlen != sizeof(assign))
        return ZX_ERR_INVALID_ARGS;

    memcpy(&assign, cmd, sizeof(assign));
    return X86PtAssignThreadBuffer(assign.descriptor, assign.thread);
}

zx_status_t InsntraceDevice::IptReleaseThreadBuffer(const void* cmd, size_t cmdlen) {
    ioctl_insntrace_assign_thread_buffer_t assign;
    if (cmdlen != sizeof(assign))
        return ZX_ERR_INVALID_ARGS;

    memcpy(&assign, cmd, sizeof(assign));
    return X86PtReleaseThreadBuffer(assign.descriptor, assign.thread);
}

zx_status_t InsntraceDevice::IptGetBufferConfig(const void* cmd, size_t cmdlen,
                                                void* reply, size_t replymax,
                                                size_t* out_actual) {
    zx_itrace_buffer_descriptor_t descriptor;
    ioctl_insntrace_buffer_config_t config;

    if (cmdlen != sizeof(descriptor))
        return ZX_ERR_INVALID_ARGS;
    if (replymax < sizeof(config))
        return ZX_ERR_BUFFER_TOO_SMALL;

    memcpy(&descriptor, cmd, sizeof(descriptor));
    if (descriptor >= num_traces_)
        return ZX_ERR_INVALID_ARGS;
    const ipt_per_trace_state_t* per_trace = &per_trace_state_[descriptor];
    if (!per_trace->allocated)
        return ZX_ERR_INVALID_ARGS;

    config.num_chunks = per_trace->num_chunks;
    config.chunk_order = per_trace->chunk_order;
    config.is_circular = per_trace->is_circular;
    config.ctl = per_trace->ctl;
    config.cr3_match = per_trace->cr3_match;
    static_assert(sizeof(config.addr_ranges) == sizeof(per_trace->addr_ranges),
                  "addr range size mismatch");
    memcpy(config.addr_ranges, per_trace->addr_ranges, sizeof(per_trace->addr_ranges));
    memcpy(reply, &config, sizeof(config));
    *out_actual = sizeof(config);
    return ZX_OK;
}

zx_status_t InsntraceDevice::IptGetBufferInfo(const void* cmd, size_t cmdlen,
                                              void* reply, size_t replymax,
                                              size_t* out_actual) {
    zx_itrace_buffer_descriptor_t descriptor;
    ioctl_insntrace_buffer_info_t data;

    if (cmdlen != sizeof(descriptor))
        return ZX_ERR_INVALID_ARGS;
    if (replymax < sizeof(data))
        return ZX_ERR_BUFFER_TOO_SMALL;

    // In thread-mode we need to get buffer info while tracing is active.
    if (mode_ == IPT_TRACE_CPUS && active_)
        return ZX_ERR_BAD_STATE;

    memcpy(&descriptor, cmd, sizeof(descriptor));
    if (descriptor >= num_traces_)
        return ZX_ERR_INVALID_ARGS;
    const ipt_per_trace_state_t* per_trace = &per_trace_state_[descriptor];
    if (!per_trace->allocated)
        return ZX_ERR_INVALID_ARGS;

    // Note: If this is a circular buffer this is just where tracing stopped.
    data.capture_end = ComputeCaptureSize(per_trace);
    memcpy(reply, &data, sizeof(data));
    *out_actual = sizeof(data);
    return ZX_OK;
}

zx_status_t InsntraceDevice::IptGetChunkHandle(const void* cmd, size_t cmdlen,
                                               void* reply, size_t replymax,
                                               size_t* out_actual) {
    ioctl_insntrace_chunk_handle_req_t req;
    zx_handle_t h;

    if (cmdlen != sizeof(req))
        return ZX_ERR_INVALID_ARGS;
    if (replymax < sizeof(h))
        return ZX_ERR_BUFFER_TOO_SMALL;

    memcpy(&req, cmd, sizeof(req));
    if (req.descriptor >= num_traces_)
        return ZX_ERR_INVALID_ARGS;
    const ipt_per_trace_state_t* per_trace = &per_trace_state_[req.descriptor];
    if (!per_trace->allocated)
        return ZX_ERR_INVALID_ARGS;
    if (req.chunk_num >= per_trace->num_chunks)
        return ZX_ERR_INVALID_ARGS;

    zx_handle_t vmo_handle = per_trace->chunks[req.chunk_num].vmo_handle;
    zx_info_handle_basic_t handle_info;
    zx_status_t status = zx_object_get_info(vmo_handle, ZX_INFO_HANDLE_BASIC,
                                            &handle_info, sizeof(handle_info),
                                            nullptr, nullptr);
    if (status != ZX_OK) {
        // This could only fail if vmo_handle is invalid.
        printf("%s: WARNING: unexpected error reading vmo handle rights: %d/%s\n",
               __func__, status, zx_status_get_string(status));
        return status;
    }
    zx_rights_t allowed_rights =
        (ZX_RIGHT_TRANSFER | ZX_RIGHT_WAIT | ZX_RIGHT_INSPECT |
         ZX_RIGHT_GET_PROPERTY | ZX_RIGHT_READ | ZX_RIGHT_MAP);
    status = zx_handle_duplicate(vmo_handle, handle_info.rights & allowed_rights, &h);
    if (status != ZX_OK)
        return status;
    memcpy(reply, &h, sizeof(h));
    *out_actual = sizeof(h);
    return ZX_OK;
}

zx_status_t InsntraceDevice::IptFreeBuffer(const void* cmd, size_t cmdlen) {
    zx_itrace_buffer_descriptor_t descriptor;
    if (cmdlen != sizeof(descriptor))
        return ZX_ERR_INVALID_ARGS;
    memcpy(&descriptor, cmd, sizeof(descriptor));

    return X86PtFreeBuffer(descriptor);
}

// Begin tracing.
// This is basically a nop in thread mode, it is still used for thread-mode
// for consistency and in case we some day need it to do something.

zx_status_t InsntraceDevice::IptStart() {
    if (active_)
        return ZX_ERR_BAD_STATE;
    if (mode_ != IPT_TRACE_CPUS)
        return ZX_ERR_BAD_STATE;

    zx_handle_t resource = get_root_resource();
    zx_status_t status;

    // In cpu-mode, until we support tracing particular cpus, auto-assign
    // buffers to each cpu.
    if (mode_ == IPT_TRACE_CPUS) {
        // First verify a buffer has been allocated for each cpu,
        // and not yet assigned.
        for (uint32_t cpu = 0; cpu < num_traces_; ++cpu) {
            const ipt_per_trace_state_t* per_trace = &per_trace_state_[cpu];
            if (!per_trace->allocated)
                return ZX_ERR_BAD_STATE;
            if (per_trace->assigned)
                return ZX_ERR_BAD_STATE;
        }

        for (uint32_t cpu = 0; cpu < num_traces_; ++cpu) {
            status = X86PtStageTraceData(resource, cpu);
            if (status != ZX_OK) {
                // TODO(dje): Unstage ones already done.
                return status;
            }
            ipt_per_trace_state_t* per_trace = &per_trace_state_[cpu];
            per_trace->owner.cpu = cpu;
            per_trace->assigned = true;
        }
    }

    status = zx_mtrace_control(resource, MTRACE_KIND_INSNTRACE,
                               MTRACE_INSNTRACE_START,
                               0, nullptr, 0);
    if (status != ZX_OK)
        return status;
    active_ = true;
    return ZX_OK;
}

// Stop tracing.
// In thread-mode all buffers must be released first. That is how we know that
// if we return ZX_OK then all threads are no longer being traced. Otherwise,
// this is basically a nop in thread-mode.

zx_status_t InsntraceDevice::IptStop() {
    if (!active_)
        return ZX_ERR_BAD_STATE;

    zx_handle_t resource = get_root_resource();

    zx_status_t status = zx_mtrace_control(resource, MTRACE_KIND_INSNTRACE,
                                           MTRACE_INSNTRACE_STOP,
                                           0, nullptr, 0);
    if (status != ZX_OK)
        return status;
    active_ = false;

    // Until we support tracing individual cpus, auto-unassign the buffers
    // in cpu-mode.
    if (mode_ == IPT_TRACE_CPUS) {
        for (uint32_t cpu = 0; cpu < num_traces_; ++cpu) {
            status = X86PtGetTraceData(resource, cpu);
            if (status != ZX_OK)
                return status;
            ipt_per_trace_state_t* per_trace = &per_trace_state_[cpu];
            per_trace->assigned = false;
            per_trace->owner.cpu = 0;
            // If there was an operational error, report it.
            if (per_trace->status & IPT_STATUS_ERROR_MASK) {
                printf("%s: WARNING: operational error detected on cpu %d\n",
                       __func__, cpu);
            }
        }
    }

    return ZX_OK;
}

zx_status_t InsntraceDevice::IoctlWorker(uint32_t op,
                                         const void* cmd, size_t cmdlen,
                                         void* reply, size_t replymax,
                                         size_t* out_actual) {
    assert(IOCTL_FAMILY(op) == IOCTL_FAMILY_INSNTRACE);

    if (op != IOCTL_INSNTRACE_ALLOC_TRACE) {
        if (!per_trace_state_)
            return ZX_ERR_BAD_STATE;
    }

    switch (op) {
    case IOCTL_INSNTRACE_ALLOC_TRACE:
        if (replymax != 0)
            return ZX_ERR_INVALID_ARGS;
        return IptAllocTrace(cmd, cmdlen);

    case IOCTL_INSNTRACE_FREE_TRACE:
        if (cmdlen != 0 || replymax != 0)
            return ZX_ERR_INVALID_ARGS;
        return IptFreeTrace();

    case IOCTL_INSNTRACE_GET_TRACE_CONFIG:
        if (cmdlen != 0)
            return ZX_ERR_INVALID_ARGS;
        return IptGetTraceConfig(reply, replymax, out_actual);

    case IOCTL_INSNTRACE_ALLOC_BUFFER:
        return IptAllocBuffer(cmd, cmdlen, reply, replymax, out_actual);

    case IOCTL_INSNTRACE_ASSIGN_THREAD_BUFFER:
        if (replymax != 0)
            return ZX_ERR_INVALID_ARGS;
        return IptAssignThreadBuffer(cmd, cmdlen);

    case IOCTL_INSNTRACE_RELEASE_THREAD_BUFFER:
        if (replymax != 0)
            return ZX_ERR_INVALID_ARGS;
        return IptReleaseThreadBuffer(cmd, cmdlen);

    case IOCTL_INSNTRACE_GET_BUFFER_CONFIG:
        return IptGetBufferConfig(cmd, cmdlen, reply, replymax, out_actual);

    case IOCTL_INSNTRACE_GET_BUFFER_INFO:
        return IptGetBufferInfo(cmd, cmdlen, reply, replymax, out_actual);

    case IOCTL_INSNTRACE_GET_CHUNK_HANDLE:
        return IptGetChunkHandle(cmd, cmdlen, reply, replymax, out_actual);

    case IOCTL_INSNTRACE_FREE_BUFFER:
        if (replymax != 0)
            return ZX_ERR_INVALID_ARGS;
        return IptFreeBuffer(cmd, cmdlen);

    case IOCTL_INSNTRACE_START:
        if (cmdlen != 0 || replymax != 0)
            return ZX_ERR_INVALID_ARGS;
        return IptStart();

    case IOCTL_INSNTRACE_STOP:
        if (cmdlen != 0 || replymax != 0)
            return ZX_ERR_INVALID_ARGS;
        return IptStop();

    default:
        return ZX_ERR_INVALID_ARGS;
    }
}

// Devhost interface.

zx_status_t InsntraceDevice::DdkOpen(zx_device_t** dev_out, uint32_t flags) {
    if (opened_)
        return ZX_ERR_ALREADY_BOUND;

    opened_ = true;
    return ZX_OK;
}

zx_status_t InsntraceDevice::DdkClose(uint32_t flags) {
    opened_ = false;
    return ZX_OK;
}

zx_status_t InsntraceDevice::DdkIoctl(uint32_t op,
                                      const void* cmd, size_t cmdlen,
                                      void* reply, size_t replymax,
                                      size_t* out_actual) {
    mtx_lock(&lock_);

    ssize_t result;
    switch (IOCTL_FAMILY(op)) {
        case IOCTL_FAMILY_INSNTRACE:
            result = IoctlWorker(op, cmd, cmdlen,
                                 reply, replymax, out_actual);
            break;
        default:
            result = ZX_ERR_INVALID_ARGS;
            break;
    }

    mtx_unlock(&lock_);

    return static_cast<zx_status_t>(result);
}

void InsntraceDevice::DdkRelease() {
    // TODO(dje): None of these should fail. What to do?
    // For now flag things as busted and prevent further use.
    IptStop();
    IptFreeTrace();

    delete this;
}

} // namespace insntrace

zx_status_t insntrace_bind(void* ctx, zx_device_t* parent) {
    zx_status_t status = insntrace::InsntraceInitOnce();
    if (status != ZX_OK) {
        return status;
    }

    pdev_protocol_t pdev;
    status = device_get_protocol(parent, ZX_PROTOCOL_PDEV, &pdev);
    if (status != ZX_OK) {
        return status;
    }

    zx::bti bti;
    status = pdev_get_bti(&pdev, 0, bti.reset_and_get_address());
    if (status != ZX_OK) {
        return status;
    }

    fbl::AllocChecker ac;
    auto dev = std::unique_ptr<insntrace::InsntraceDevice>(
        new (&ac) insntrace::InsntraceDevice(parent, std::move(bti)));
    if (!ac.check()) {
        return ZX_ERR_NO_MEMORY;
    }

    status = dev->DdkAdd("insntrace");
    if (status != ZX_OK) {
        zxlogf(ERROR, "%s: could not add device: %d\n", __func__, status);
    } else {
        // devmgr owns the memory now
        __UNUSED auto ptr = dev.release();
    }
    return status;
}