graphene/Pal/src/host/Linux-SGX/enclave_framework.c

#include <api.h>
#include <list.h>
#include <pal_crypto.h>
#include <pal_debug.h>
#include <pal_error.h>
#include <pal_internal.h>
#include <pal_linux.h>
#include <pal_linux_error.h>
#include <pal_security.h>
#include <spinlock.h>
#include <stdbool.h>

#include "enclave_pages.h"

__sgx_mem_aligned struct pal_enclave_state pal_enclave_state;

void * enclave_base, * enclave_top;

struct pal_enclave_config pal_enclave_config;

static int register_trusted_file(const char* uri, const char* checksum_str, bool check_duplicates);

bool sgx_is_completely_within_enclave (const void * addr, uint64_t size)
{
    if (((uint64_t) addr) > (UINT64_MAX - size)) {
        return false;
    }

    return enclave_base <= addr && addr + size <= enclave_top;
}

bool sgx_is_completely_outside_enclave(const void* addr, uint64_t size) {
    if (((uint64_t) addr) > (UINT64_MAX - size)) {
        return false;
    }

    return enclave_base >= addr + size || enclave_top <= addr;
}

void* sgx_alloc_on_ustack(uint64_t size) {
    void* ustack = GET_ENCLAVE_TLS(ustack) - size;
    if (!sgx_is_completely_outside_enclave(ustack, size)) {
        return NULL;
    }
    SET_ENCLAVE_TLS(ustack, ustack);
    return ustack;
}

void* sgx_copy_to_ustack(const void* ptr, uint64_t size) {
    if (!sgx_is_completely_within_enclave(ptr, size)) {
        return NULL;
    }
    void* uptr = sgx_alloc_on_ustack(size);
    if (uptr) {
        memcpy(uptr, ptr, size);
    }
    return uptr;
}

void sgx_reset_ustack(void) {
    SET_ENCLAVE_TLS(ustack, GET_ENCLAVE_TLS(ustack_top));
}

/* NOTE: Value from possibly untrusted uptr must be copied inside
 * CPU register or enclave stack (to prevent TOCTOU). Function call
 * achieves this. Attribute ensures no inline optimization. */
__attribute__((noinline))
bool sgx_copy_ptr_to_enclave(void** ptr, void* uptr, uint64_t size) {
    assert(ptr);
    if (!sgx_is_completely_outside_enclave(uptr, size)) {
        *ptr = NULL;
        return false;
    }
    *ptr = uptr;
    return true;
}

/* NOTE: Value from possibly untrusted uptr and usize must be copied
 * inside CPU registers or enclave stack (to prevent TOCTOU). Function
 * call achieves this. Attribute ensures no inline optimization. */
__attribute__((noinline))
uint64_t sgx_copy_to_enclave(const void* ptr, uint64_t maxsize, const void* uptr, uint64_t usize) {
    if (usize > maxsize ||
        !sgx_is_completely_outside_enclave(uptr, usize) ||
        !sgx_is_completely_within_enclave(ptr, usize)) {
        return 0;
    }
    memcpy((void*) ptr, uptr, usize);
    return usize;
}

static void print_report(sgx_report_t* r) {
    SGX_DBG(DBG_S, "  cpu_svn:     %s\n",     ALLOCA_BYTES2HEXSTR(r->body.cpu_svn.svn));
    SGX_DBG(DBG_S, "  mr_enclave:  %s\n",     ALLOCA_BYTES2HEXSTR(r->body.mr_enclave.m));
    SGX_DBG(DBG_S, "  mr_signer:   %s\n",     ALLOCA_BYTES2HEXSTR(r->body.mr_signer.m));
    SGX_DBG(DBG_S, "  attr.flags:  %016lx\n", r->body.attributes.flags);
    SGX_DBG(DBG_S, "  attr.xfrm:   %016lx\n", r->body.attributes.xfrm);
    SGX_DBG(DBG_S, "  isv_prod_id: %02x\n",   r->body.isv_prod_id);
    SGX_DBG(DBG_S, "  isv_svn:     %02x\n",   r->body.isv_svn);
    SGX_DBG(DBG_S, "  report_data: %s\n",     ALLOCA_BYTES2HEXSTR(r->body.report_data.d));
    SGX_DBG(DBG_S, "  key_id:      %s\n",     ALLOCA_BYTES2HEXSTR(r->key_id.id));
    SGX_DBG(DBG_S, "  mac:         %s\n",     ALLOCA_BYTES2HEXSTR(r->mac));
}

static sgx_key_128bit_t enclave_key;

/*
 * sgx_get_report() obtains a CPU-signed report for local attestation
 * @target_info:  the enclave target info
 * @data:         the data to be included and signed in the report
 * @report:       a buffer for storing the report
 */
static int sgx_get_report(sgx_target_info_t* target_info, sgx_sign_data_t* data,
                          sgx_report_t* report) {
    __sgx_mem_aligned struct pal_enclave_state state;
    memcpy(&state, &pal_enclave_state, sizeof(state));
    memcpy(&state.enclave_data, data, sizeof(*data));

    int ret = sgx_report(target_info, &state, report);
    if (ret) {
        SGX_DBG(DBG_E, "sgx_report failed: ret = %d)\n", ret);
        return -PAL_ERROR_DENIED;
    }

    print_report(report);
    return 0;
}

int sgx_verify_report (sgx_report_t* report)
{
    __sgx_mem_aligned sgx_key_request_t keyrequest;
    memset(&keyrequest, 0, sizeof(sgx_key_request_t));
    keyrequest.key_name = REPORT_KEY;
    memcpy(&keyrequest.key_id, &report->key_id, sizeof(keyrequest.key_id));

    sgx_key_128bit_t report_key __attribute__((aligned(sizeof(sgx_key_128bit_t))));
    memset(&report_key, 0, sizeof(report_key));

    int ret = sgx_getkey(&keyrequest, &report_key);
    if (ret) {
        SGX_DBG(DBG_E, "Can't get report key\n");
        return -PAL_ERROR_DENIED;
    }

    SGX_DBG(DBG_S, "Get report key for verification: %s\n", ALLOCA_BYTES2HEXSTR(report_key));

    sgx_mac_t check_mac;
    memset(&check_mac, 0, sizeof(check_mac));

    // Generating the MAC with AES-CMAC using the report key. Only hash the part of the report
    // BEFORE the keyid field (hence the offsetof(...) trick). ENCLU[EREPORT] does not include
    // the MAC and the keyid fields when generating the MAC.
    lib_AESCMAC((uint8_t*)&report_key, sizeof(report_key),
                (uint8_t*)report, offsetof(sgx_report_t, key_id),
                (uint8_t*)&check_mac, sizeof(check_mac));

    // Clear the report key for security
    memset(&report_key, 0, sizeof(report_key));

    SGX_DBG(DBG_S, "Verify report:\n");
    print_report(report);
    SGX_DBG(DBG_S, "  verify:     %s\n", ALLOCA_BYTES2HEXSTR(check_mac));

    if (memcmp(&check_mac, &report->mac, sizeof(check_mac))) {
        SGX_DBG(DBG_E, "Report verification failed\n");
        return -PAL_ERROR_DENIED;
    }

    return 0;
}

int init_enclave_key (void)
{
    __sgx_mem_aligned sgx_key_request_t keyrequest;
    memset(&keyrequest, 0, sizeof(sgx_key_request_t));
    keyrequest.key_name = SEAL_KEY;

    int ret = sgx_getkey(&keyrequest, &enclave_key);
    if (ret) {
        SGX_DBG(DBG_E, "Can't get seal key\n");
        return -PAL_ERROR_DENIED;
    }

    SGX_DBG(DBG_S, "Seal key: %s\n", ALLOCA_BYTES2HEXSTR(enclave_key));
    return 0;
}

/*
 * The file integrity check is designed as follow:
 *
 * For each file that requires authentication (specified in the manifest
 * as "sgx.trusted_files.xxx"), a SHA256 checksum is generated and stored
 * in the manifest, signed and verified as part of the enclave's crypto
 * measurement. When user requests for opening the file, Graphene loads
 * the whole file, generate the SHA256 checksum, and check with the known
 * checksums listed in the manifest. If the checksum does not match, and
 * neither does the file is allowed for unauthenticated access, the file
 * access will be rejected.
 *
 * During the generation of the SHA256 checksum, a 128-bit hash is also
 * generated for each chunk in the file. The per-chunk hashes are used
 * for partial verification in future reads, to avoid re-verifying the
 * whole file again or the need of caching file contents. The per-chunk
 * hashes are stored as "stubs" for each file. For a performance reason,
 * each per-chunk hash is a 128-bit AES-CMAC hash value, using a secret
 * key generated at the beginning of the enclave.
 */

DEFINE_LIST(trusted_file);
struct trusted_file {
    LIST_TYPE(trusted_file) list;
    int64_t index;
    uint64_t size;
    size_t uri_len;
    char uri[URI_MAX];
    sgx_checksum_t checksum;
    sgx_stub_t * stubs;
};

DEFINE_LISTP(trusted_file);
static LISTP_TYPE(trusted_file) trusted_file_list = LISTP_INIT;
static spinlock_t trusted_file_lock = INIT_SPINLOCK_UNLOCKED;
static int trusted_file_indexes = 0;
static bool allow_file_creation = 0;
static int file_check_policy = FILE_CHECK_POLICY_STRICT;

/* Assumes `path` is normalized */
static bool path_is_equal_or_subpath(const struct trusted_file* tf,
                                     const char* path,
                                     size_t path_len) {
    if (tf->uri_len > path_len || memcmp(tf->uri, path, tf->uri_len)) {
        /* tf->uri is not prefix of `path` */
        return false;
    }
    if (tf->uri_len == path_len) {
        /* Both are equal */
        return true;
    }
    if (tf->uri[tf->uri_len - 1] == '/' || path[tf->uri_len] == '/') {
        /* tf->uri is a subpath of `path` */
        return true;
    }
    if (tf->uri_len == URI_PREFIX_FILE_LEN && !memcmp(tf->uri, URI_PREFIX_FILE, URI_PREFIX_FILE_LEN)) {
        /* Empty path is a prefix of everything */
        return true;
    }
    return false;
}

/*
 * 'load_trusted_file' checks if the file to be opened is trusted
 * or allowed for unauthenticated access, according to the manifest.
 *
 * file:     file handle to be opened
 * stubptr:  buffer for catching matched file stub.
 * sizeptr:  size pointer
 * create:   this file is newly created or not
 *
 * Returns 0 if succeeded, or an error code otherwise.
 */
int load_trusted_file (PAL_HANDLE file, sgx_stub_t ** stubptr,
                       uint64_t * sizeptr, int create, void** umem)
{
    *stubptr = NULL;
    *sizeptr = 0;
    *umem = NULL;

    struct trusted_file * tf = NULL, * tmp;
    char uri[URI_MAX];
    char normpath[URI_MAX];
    int ret, fd = file->file.fd;

    if (!(HANDLE_HDR(file)->flags & RFD(0)))
        return -PAL_ERROR_DENIED;

    ret = _DkStreamGetName(file, uri, URI_MAX);
    if (ret < 0) {
        return ret;
    }

    /* Allow to create the file when allow_file_creation is turned on;
       The created file is added to allowed_file list for later access */
    if (create && allow_file_creation) {
       register_trusted_file(uri, NULL, /*check_duplicates=*/true);
       return 0;
    }

    /* Normalize the uri */
    if (!strstartswith_static(uri, URI_PREFIX_FILE)) {
        SGX_DBG(DBG_E, "Invalid URI [%s]: Trusted files must start with 'file:'\n", uri);
        return -PAL_ERROR_INVAL;
    }
    assert(sizeof(normpath) > URI_PREFIX_FILE_LEN);
    memcpy(normpath, URI_PREFIX_FILE, URI_PREFIX_FILE_LEN);
    size_t len = sizeof(normpath) - URI_PREFIX_FILE_LEN;
    ret = get_norm_path(uri + URI_PREFIX_FILE_LEN, normpath + URI_PREFIX_FILE_LEN, &len);
    if (ret < 0) {
        SGX_DBG(DBG_E,
                "Path (%s) normalization failed: %s\n",
                uri + URI_PREFIX_FILE_LEN,
                pal_strerror(ret));
        return ret;
    }
    len += URI_PREFIX_FILE_LEN;

    spinlock_lock(&trusted_file_lock);

    LISTP_FOR_EACH_ENTRY(tmp, &trusted_file_list, list) {
        if (tmp->stubs) {
            /* trusted files: must be exactly the same URI */
            if (tmp->uri_len == len && !memcmp(tmp->uri, normpath, len + 1)) {
                tf = tmp;
                break;
            }
        } else {
            /* allowed files: must be a subfolder or file */
            if (path_is_equal_or_subpath(tmp, normpath, len)) {
                tf = tmp;
                break;
            }
        }
    }

    spinlock_unlock(&trusted_file_lock);

    if (!tf || !tf->index) {
        if (!tf) {
            if (get_file_check_policy() != FILE_CHECK_POLICY_ALLOW_ALL_BUT_LOG)
                return -PAL_ERROR_DENIED;

            pal_printf("Allowing access to an unknown file due to "
                       "file_check_policy settings: %s\n", uri);
        }

        *stubptr = NULL;
        PAL_STREAM_ATTR attr;
        ret = _DkStreamAttributesQuery(normpath, &attr);
        if (!ret)
            *sizeptr = attr.pending_size;
        else
            *sizeptr = 0;

        return 0;
    }

    if (tf->index < 0)
        return tf->index;

    sgx_stub_t* stubs = NULL;
    /* mmap the whole trusted file in untrusted memory for future reads/writes; it is
     * caller's responsibility to unmap those areas after use */
    *sizeptr = tf->size;
    if (*sizeptr) {
        ret = ocall_mmap_untrusted(fd, 0, tf->size, PROT_READ, umem);
        if (IS_ERR(ret)) {
            *umem = NULL;
            ret = unix_to_pal_error(ERRNO(ret));
            goto failed;
        }
    }

#if CACHE_FILE_STUBS == 1
    if (tf->stubs) {
        *stubptr = tf->stubs;
        *sizeptr = tf->size;
        return 0;
    }
#endif

    int nstubs = tf->size / TRUSTED_STUB_SIZE +
                (tf->size % TRUSTED_STUB_SIZE ? 1 : 0);

    stubs = malloc(sizeof(sgx_stub_t) * nstubs);
    if (!stubs) {
        ret = -PAL_ERROR_NOMEM;
        goto failed;
    }

    sgx_stub_t * s = stubs; /* stubs is an array of 128bit values */
    uint64_t offset = 0;
    LIB_SHA256_CONTEXT sha;

    ret = lib_SHA256Init(&sha);
    if (ret < 0)
        goto failed;

    for (; offset < tf->size ; offset += TRUSTED_STUB_SIZE, s++) {
        /* For each stub, generate a 128bit hash of a file chunk with
         * AES-CMAC, and then update the SHA256 digest. */
        uint64_t mapping_size = MIN(tf->size - offset, TRUSTED_STUB_SIZE);
        LIB_AESCMAC_CONTEXT aes_cmac;
        ret = lib_AESCMACInit(&aes_cmac, (uint8_t*)&enclave_key, sizeof(enclave_key));
        if (ret < 0)
            goto failed;

        /*
         * To prevent TOCTOU attack when generating the file checksum, we
         * need to copy the file content into the enclave before hashing.
         * For optimization, we use a relatively small buffer (1024 byte) to
         * store the data for checksum generation.
         */

#define FILE_CHUNK_SIZE 1024UL

        uint8_t small_chunk[FILE_CHUNK_SIZE]; /* Buffer for hashing */
        size_t chunk_offset = 0;

        for (; chunk_offset < mapping_size; chunk_offset += FILE_CHUNK_SIZE) {
            uint64_t chunk_size = MIN(mapping_size - chunk_offset, FILE_CHUNK_SIZE);

            /* Any file content needs to be copied into the enclave before
             * checking and re-hashing */
            memcpy(small_chunk, *umem + offset + chunk_offset, chunk_size);

            /* Update the file checksum */
            ret = lib_SHA256Update(&sha, small_chunk, chunk_size);
            if (ret < 0)
                goto failed;

            /* Update the checksum for the file chunk */
            ret = lib_AESCMACUpdate(&aes_cmac, small_chunk, chunk_size);
            if (ret < 0)
                goto failed;
        }

        /* Store the checksum for one file chunk for checking */
        ret = lib_AESCMACFinish(&aes_cmac, (uint8_t *) s, sizeof *s);
        if (ret < 0)
            goto failed;
    }

    sgx_checksum_t hash;

    /* Finalize and checking if the checksum of the whole file matches
     * with record given in the manifest. */

    ret = lib_SHA256Final(&sha, (uint8_t *) hash.bytes);
    if (ret < 0)
        goto failed;

    if (memcmp(&hash, &tf->checksum, sizeof(sgx_checksum_t))) {
        ret = -PAL_ERROR_DENIED;
        goto failed;
    }

    spinlock_lock(&trusted_file_lock);
    if (tf->stubs || tf->index == -PAL_ERROR_DENIED)
        free(tf->stubs);
    *stubptr = tf->stubs = stubs;
    ret = tf->index;
    spinlock_unlock(&trusted_file_lock);
    return ret;

failed:
    if (*umem) {
        assert(*sizeptr > 0);
        ocall_munmap_untrusted(*umem, *sizeptr);
    }
    free(stubs);

    spinlock_lock(&trusted_file_lock);
    if (!tf->stubs) {
        tf->index = -PAL_ERROR_DENIED;
    }
    spinlock_unlock(&trusted_file_lock);

    return ret;
}

int get_file_check_policy ()
{
    return file_check_policy;
}

static void set_file_check_policy (int policy)
{
    file_check_policy = policy;
}

/*
 * A common helper function for copying and checking the file contents
 * from a buffer mapped outside the enclaves into an in-enclave buffer.
 * If needed, regions at either the beginning or the end of the copied regions
 * are copied into a scratch buffer to avoid a TOCTTOU race.
 *
 * * Note that it must be done this way to avoid the following TOCTTOU race
 * * condition with the untrusted host as an adversary:
 *       *  Adversary: put good contents in buffer
 *       *  Enclave: buffer check passes
 *       *  Adversary: put bad contents in buffer
 *       *  Enclave: copies in bad buffer contents
 *
 * * For optimization, we verify the memory in place, as the application code
 *   should not use the memory before return.  There can be subtle interactions
 *   at the edges of a region with ELF loading.  Namely, the ELF loader will
 *   want to map several file chunks that are not aligned to TRUSTED_STUB_SIZE
 *   next to each other, sometimes overlapping.  There is probably room to
 *   improve load time with more smarts around ELF loading, but for now, just
 *   make things work.
 *
 * 'umem' is the untrusted file memory mapped outside the enclave (should
 * already be mapped up by the caller). 'umem_start' and 'umem_end' are
 * the offset _within the file_ of 'umem'.  'umem_start' should be aligned
 * to the file checking chunk size (TRUSTED_STUB_SIZE). 'umem_end' can be
 * either aligned, or equal to 'total_size'. 'buffer' is the in-enclave
 * buffer for copying the file content. 'offset' is the offset within the file
 * for copying into the buffer. 'size' is the size of the in-enclave buffer.
 * 'stubs' contain the checksums of all the chunks in a file.
 */
int copy_and_verify_trusted_file (const char * path, const void * umem,
                    uint64_t umem_start, uint64_t umem_end,
                    void * buffer, uint64_t offset, uint64_t size,
                    sgx_stub_t * stubs, uint64_t total_size)
{
    /* Check that the untrusted mapping is aligned to TRUSTED_STUB_SIZE
     * and includes the range for copying into the buffer */
    assert(IS_ALIGNED(umem_start, TRUSTED_STUB_SIZE));
    assert(offset >= umem_start && offset + size <= umem_end);

    /* Start copying and checking at umem_start. The checked content may or
     * may not be copied into the file content, depending on the offset of
     * the content within the file. */
    uint64_t checking = umem_start;
    /* The stubs is an array of 128-bit hash values of the file chunks.
     * from the beginning of the file. 's' points to the stub that needs to
     * be checked for the current offset. */
    sgx_stub_t * s = stubs + checking / TRUSTED_STUB_SIZE;
    int ret = 0;

    for (; checking < umem_end ; checking += TRUSTED_STUB_SIZE, s++) {
        /* Check one chunk at a time. */
        uint64_t checking_size = MIN(total_size - checking, TRUSTED_STUB_SIZE);
        uint64_t checking_end = checking + checking_size;
        sgx_checksum_t hash;

        if (checking >= offset && checking_end <= offset + size) {
            /* If the checking chunk completely overlaps with the region
             * needed for copying into the buffer, simplying use the buffer
             * for checking */
            memcpy(buffer + checking - offset, umem + checking - umem_start,
                   checking_size);

            /* Storing the checksum (using AES-CMAC) inside hash. */
            ret = lib_AESCMAC((uint8_t*)&enclave_key, sizeof(enclave_key),
                              buffer + checking - offset, checking_size,
                              (uint8_t*)&hash, sizeof(hash));
        } else {
            /* If the checking chunk only partially overlaps with the region,
             * read the file content in smaller chunks and only copy the part
             * needed by the caller. */
            LIB_AESCMAC_CONTEXT aes_cmac;
            ret = lib_AESCMACInit(&aes_cmac, (uint8_t*)&enclave_key, sizeof(enclave_key));
            if (ret < 0)
                goto failed;

            uint8_t small_chunk[FILE_CHUNK_SIZE]; /* A small buffer */
            uint64_t chunk_offset = checking;

            for (; chunk_offset < checking_end
                 ; chunk_offset += FILE_CHUNK_SIZE) {
                uint64_t chunk_size = MIN(checking_end - chunk_offset,
                                          FILE_CHUNK_SIZE);

                /* Copy into the small buffer before hashing the content */
                memcpy(small_chunk, umem + (chunk_offset - umem_start),
                       chunk_size);

                /* Update the hash for the current chunk */
                ret = lib_AESCMACUpdate(&aes_cmac, small_chunk, chunk_size);
                if (ret < 0)
                    goto failed;

                /* Determine if the part just copied and checked is needed
                 * by the caller. If so, copy it into the user buffer. */
                uint64_t copy_start = chunk_offset;
                uint64_t copy_end = copy_start + chunk_size;

                if (copy_start < offset)
                    copy_start = offset;
                if (copy_end > offset + size)
                    copy_end = offset + size;

                if (copy_end > copy_start)
                    memcpy(buffer + (copy_start - offset),
                           small_chunk + (copy_start - chunk_offset),
                           copy_end - copy_start);
            }

            /* Storing the checksum (using AES-CMAC) inside hash. */
            ret = lib_AESCMACFinish(&aes_cmac, (uint8_t*)&hash, sizeof(hash));
        }

        if (ret < 0)
            goto failed;

        /*
         * Check if the hash matches with the checksum of current chunk.
         * If not, return with access denied. Note: some file content may
         * still be in the buffer (including the corrupted part).
         * We assume the user won't use the content if this function
         * returns with failures.
         *
         * XXX: Maybe we should zero the buffer after denying the access?
         */
        if (memcmp(s, &hash, sizeof(sgx_stub_t))) {
            SGX_DBG(DBG_E, "Accesing file:%s is denied. Does not match with MAC"
                    " at chunk starting at %lu-%lu.\n",
                    path, checking, checking_end);
            return -PAL_ERROR_DENIED;
        }
    }

    return 0;

failed:
    return -PAL_ERROR_DENIED;
}

static int register_trusted_file(const char* uri, const char* checksum_str, bool check_duplicates) {
    struct trusted_file * tf = NULL, * new;
    size_t uri_len = strlen(uri);
    int ret;

    if (check_duplicates) {
        /* this check is only done during runtime (when creating a new file) and not needed during
         * initialization (because manifest is assumed to have no duplicates); skipping this check
         * significantly improves startup time */
        spinlock_lock(&trusted_file_lock);
        LISTP_FOR_EACH_ENTRY(tf, &trusted_file_list, list) {
            if (tf->uri_len == uri_len && !memcmp(tf->uri, uri, uri_len)) {
                spinlock_unlock(&trusted_file_lock);
                return 0;
            }
        }
        spinlock_unlock(&trusted_file_lock);
    }

    new = malloc(sizeof(struct trusted_file));
    if (!new)
        return -PAL_ERROR_NOMEM;

    INIT_LIST_HEAD(new, list);
    new->uri_len = uri_len;
    memcpy(new->uri, uri, uri_len + 1);
    new->size = 0;
    new->stubs = NULL;

    if (checksum_str) {
        PAL_STREAM_ATTR attr;
        ret = _DkStreamAttributesQuery(uri, &attr);
        if (!ret)
            new->size = attr.pending_size;

        char checksum_text[sizeof(sgx_checksum_t) * 2 + 1] = "\0";
        size_t nbytes = 0;
        for (; nbytes < sizeof(sgx_checksum_t) ; nbytes++) {
            char byte1 = checksum_str[nbytes * 2];
            char byte2 = checksum_str[nbytes * 2 + 1];
            unsigned char val = 0;

            if (byte1 == 0 || byte2 == 0) {
                break;
            }
            if (!(byte1 >= '0' && byte1 <= '9') &&
                !(byte1 >= 'a' && byte1 <= 'f')) {
                break;
            }
            if (!(byte2 >= '0' && byte2 <= '9') &&
                !(byte2 >= 'a' && byte2 <= 'f')) {
                break;
            }

            if (byte1 >= '0' && byte1 <= '9')
                val = byte1 - '0';
            if (byte1 >= 'a' && byte1 <= 'f')
                val = byte1 - 'a' + 10;
            val *= 16;
            if (byte2 >= '0' && byte2 <= '9')
                val += byte2 - '0';
            if (byte2 >= 'a' && byte2 <= 'f')
                val += byte2 - 'a' + 10;

            new->checksum.bytes[nbytes] = val;
            snprintf(checksum_text + nbytes * 2, 3, "%02x", val);
        }

        if (nbytes < sizeof(sgx_checksum_t)) {
            free(new);
            return -PAL_ERROR_INVAL;
        }

        new->index = (++trusted_file_indexes);
        SGX_DBG(DBG_S, "trusted: [%ld] %s %s\n", new->index,
                checksum_text, new->uri);
    } else {
        memset(&new->checksum, 0, sizeof(sgx_checksum_t));
        new->index = 0;
        SGX_DBG(DBG_S, "allowed: %s\n", new->uri);
    }

    spinlock_lock(&trusted_file_lock);

    if (check_duplicates) {
        /* this check is only done during runtime and not needed during initialization (see above);
         * we check again because same file could have been added by another thread in meantime */
        LISTP_FOR_EACH_ENTRY(tf, &trusted_file_list, list) {
            if (tf->uri_len == uri_len && !memcmp(tf->uri, uri, uri_len)) {
                spinlock_unlock(&trusted_file_lock);
                free(new);
                return 0;
            }
        }
    }

    LISTP_ADD_TAIL(new, &trusted_file_list, list);
    spinlock_unlock(&trusted_file_lock);

    return 0;
}

static int init_trusted_file (const char * key, const char * uri)
{
    char cskey[URI_MAX], * tmp;
    char checksum[URI_MAX];
    char normpath[URI_MAX];

    tmp = strcpy_static(cskey, "sgx.trusted_checksum.", URI_MAX);
    memcpy(tmp, key, strlen(key) + 1);

    ssize_t ret = get_config(pal_state.root_config, cskey, checksum, sizeof(checksum));
    if (ret < 0)
        return 0;

    /* Normalize the uri */
    if (!strstartswith_static(uri, URI_PREFIX_FILE)) {
        SGX_DBG(DBG_E, "Invalid URI [%s]: Trusted files must start with 'file:'\n", uri);
        return -PAL_ERROR_INVAL;
    }
    assert(sizeof(normpath) > URI_PREFIX_FILE_LEN);
    memcpy(normpath, URI_PREFIX_FILE, URI_PREFIX_FILE_LEN);
    size_t len = sizeof(normpath) - URI_PREFIX_FILE_LEN;
    ret = get_norm_path(uri + URI_PREFIX_FILE_LEN, normpath + URI_PREFIX_FILE_LEN, &len);
    if (ret < 0) {
        SGX_DBG(DBG_E,
                "Path (%s) normalization failed: %s\n",
                uri + URI_PREFIX_FILE_LEN,
                pal_strerror(ret));
        return ret;
    }

    return register_trusted_file(normpath, checksum, /*check_duplicates=*/false);
}

int init_trusted_files (void) {
    struct config_store* store = pal_state.root_config;
    char* cfgbuf = NULL;
    ssize_t cfgsize;
    int nuris, ret;
    char key[CONFIG_MAX];
    char uri[CONFIG_MAX];
    char* k;
    char* tmp;

    if (pal_sec.exec_name[0] != '\0') {
        ret = init_trusted_file("exec", pal_sec.exec_name);
        if (ret < 0)
            goto out;
    }

    cfgbuf = malloc(CONFIG_MAX);
    if (!cfgbuf) {
        ret = -PAL_ERROR_NOMEM;
        goto out;
    }

    ssize_t len = get_config(store, "loader.preload", cfgbuf, CONFIG_MAX);
    if (len > 0) {
        int npreload = 0;
        char key[10];
        const char * start, * end;

        for (start = cfgbuf ; start < cfgbuf + len ; start = end + 1) {
            for (end = start ; end < cfgbuf + len && *end && *end != ',' ; end++);
            if (end > start) {
                char uri[end - start + 1];
                memcpy(uri, start, end - start);
                uri[end - start] = 0;
                snprintf(key, 10, "preload%d", npreload++);

                ret = init_trusted_file(key, uri);
                if (ret < 0)
                    goto out;
            }
        }
    }

    cfgsize = get_config_entries_size(store, "sgx.trusted_files");
    if (cfgsize <= 0)
        goto no_trusted;

    free(cfgbuf);
    cfgbuf = malloc(cfgsize);
    if (!cfgbuf) {
        ret = -PAL_ERROR_NOMEM;
        goto out;
    }


    nuris = get_config_entries(store, "sgx.trusted_files", cfgbuf, cfgsize);
    if (nuris <= 0)
        goto no_trusted;

    tmp = strcpy_static(key, "sgx.trusted_files.", sizeof(key));

    k = cfgbuf;

    for (int i = 0 ; i < nuris ; i++) {
        len = strlen(k);
        memcpy(tmp, k, len + 1);
        k += len + 1;
        len = get_config(store, key, uri, sizeof(uri));
        if (len > 0) {
            ret = init_trusted_file(key + static_strlen("sgx.trusted_files."), uri);
            if (ret < 0)
                goto out;
        }
    }

no_trusted:

    cfgsize = get_config_entries_size(store, "sgx.allowed_files");
    if (cfgsize <= 0)
        goto no_allowed;

    free(cfgbuf);
    cfgbuf = malloc(cfgsize);
    if (!cfgbuf) {
        ret = -PAL_ERROR_NOMEM;
        goto out;
    }

    nuris = get_config_entries(store, "sgx.allowed_files", cfgbuf, cfgsize);
    if (nuris <= 0)
        goto no_allowed;

    tmp = strcpy_static(key, "sgx.allowed_files.", sizeof(key));

    k = cfgbuf;

    for (int i = 0 ; i < nuris ; i++) {
        len = strlen(k);
        memcpy(tmp, k, len + 1);
        k += len + 1;
        len = get_config(store, key, uri, sizeof(uri));
        if (len <= 0) {
            continue;
        }

        /* Normalize the uri */
        char norm_path[URI_MAX];

        if (!strstartswith_static(uri, URI_PREFIX_FILE)) {
            SGX_DBG(DBG_E, "Invalid URI [%s]: Allowed files must start with 'file:'\n", uri);
            ret = -PAL_ERROR_INVAL;
            goto out;
        }
        assert(sizeof(norm_path) > URI_PREFIX_FILE_LEN);
        memcpy(norm_path, URI_PREFIX_FILE, URI_PREFIX_FILE_LEN);

        size_t norm_path_len = sizeof(norm_path) - URI_PREFIX_FILE_LEN;

        ret = get_norm_path(uri + URI_PREFIX_FILE_LEN,
                            norm_path + URI_PREFIX_FILE_LEN,
                            &norm_path_len);
        if (ret < 0) {
            SGX_DBG(DBG_E,
                    "Path (%s) normalization failed: %s\n",
                    uri + URI_PREFIX_FILE_LEN,
                    pal_strerror(ret));
            goto out;
        }

        register_trusted_file(norm_path, NULL, /*check_duplicates=*/false);
    }

no_allowed:
    ret = 0;

    if (get_config(store, "sgx.allow_file_creation", cfgbuf, cfgsize) > 0 && cfgbuf[0] == '1')
        allow_file_creation = true;
    else
        allow_file_creation = false;

out:
    free(cfgbuf);
    return ret;
}

int init_trusted_children (void)
{
    struct config_store * store = pal_state.root_config;

    char key[CONFIG_MAX], mrkey[CONFIG_MAX];
    char uri[CONFIG_MAX], mr_enclave[CONFIG_MAX];

    char * tmp1 = strcpy_static(key, "sgx.trusted_children.", sizeof(key));
    char * tmp2 = strcpy_static(mrkey, "sgx.trusted_mrenclave.", sizeof(mrkey));

    ssize_t cfgsize = get_config_entries_size(store, "sgx.trusted_mrenclave");
    if (cfgsize <= 0)
        return 0;

    char * cfgbuf = malloc(cfgsize);
    if (!cfgbuf)
        return -PAL_ERROR_NOMEM;

    int nuris = get_config_entries(store, "sgx.trusted_mrenclave",
                                   cfgbuf, cfgsize);
    if (nuris > 0) {
        char * k = cfgbuf;
        for (int i = 0 ; i < nuris ; i++) {
            int len = strlen(k);
            memcpy(tmp1, k, len + 1);
            memcpy(tmp2, k, len + 1);
            k += len + 1;

            ssize_t ret = get_config(store, key, uri, sizeof(uri));
            if (ret < 0)
                continue;

            ret = get_config(store, mrkey, mr_enclave, sizeof(mr_enclave));
            if (ret > 0)
                register_trusted_child(uri, mr_enclave);
        }
    }
    free(cfgbuf);
    return 0;
}

int init_file_check_policy (void)
{
    char cfgbuf[CONFIG_MAX];
    ssize_t ret = get_config(pal_state.root_config, "sgx.file_check_policy",
                             cfgbuf, sizeof(cfgbuf));

    if (ret > 0) {
        if (!strcmp_static(cfgbuf, "strict"))
            set_file_check_policy(FILE_CHECK_POLICY_STRICT);
        else if (!strcmp_static(cfgbuf, "allow_all_but_log"))
            set_file_check_policy(FILE_CHECK_POLICY_ALLOW_ALL_BUT_LOG);
        else
            INIT_FAIL(PAL_ERROR_INVAL, "unknown file check policy");

        SGX_DBG(DBG_S, "File check policy: %s\n", cfgbuf);
    }

    return 0;
}

int init_enclave (void)
{
    // Get report to initialize info (MR_ENCLAVE, etc.) about this enclave from
    // a trusted source.

    // Since this report is only read by ourselves we can
    // leave targetinfo zeroed.
    __sgx_mem_aligned sgx_target_info_t targetinfo = {0};
    __sgx_mem_aligned struct pal_enclave_state reportdata = {0};
    __sgx_mem_aligned sgx_report_t report;

    assert(sizeof(reportdata) == sizeof(sgx_report_data_t));
    int ret = sgx_report(&targetinfo, (sgx_report_data_t*)&reportdata, &report);
    if (ret) {
        SGX_DBG(DBG_E, "failed to get self report: %d\n", ret);
        return -PAL_ERROR_INVAL;
    }

    memcpy(&pal_sec.mr_enclave, &report.body.mr_enclave, sizeof(pal_sec.mr_enclave));
    memcpy(&pal_sec.mr_signer, &report.body.mr_signer, sizeof(pal_sec.mr_signer));
    pal_sec.enclave_attributes = report.body.attributes;

    /*
     * The enclave id is uniquely created for each enclave as a token
     * for authenticating the enclave as the sender of attestation.
     * See 'host/Linux-SGX/db_process.c' for further explanation.
     */
    ret = _DkRandomBitsRead(&pal_enclave_state.enclave_id,
                            sizeof(pal_enclave_state.enclave_id));
    if (ret < 0) {
        SGX_DBG(DBG_E, "Failed to generate a random id: %d\n", ret);
        return ret;
    }

    return 0;
}

int _DkStreamKeyExchange(PAL_HANDLE stream, PAL_SESSION_KEY* key) {
    uint8_t pub[DH_SIZE]   __attribute__((aligned(DH_SIZE)));
    uint8_t agree[DH_SIZE] __attribute__((aligned(DH_SIZE)));
    PAL_NUM pubsz, agreesz;
    LIB_DH_CONTEXT context;
    int64_t bytes;
    int64_t ret;

    ret = lib_DhInit(&context);
    if (ret < 0) {
        SGX_DBG(DBG_E, "Key Exchange: DH Init failed: %ld\n", ret);
        goto out_no_final;
    }

    pubsz = sizeof pub;
    ret = lib_DhCreatePublic(&context, pub, &pubsz);
    if (ret < 0) {
        SGX_DBG(DBG_E, "Key Exchange: DH CreatePublic failed: %ld\n", ret);
        goto out;
    }

    assert(pubsz > 0 && pubsz <= DH_SIZE);
    if (pubsz < DH_SIZE) {
        /* Insert leading zero bytes if necessary. These values are big-
         * endian, so we either need to know the length of the bignum or
         * zero-pad at the beginning instead of the end. This code chooses
         * to do the latter. */
        memmove(pub + (DH_SIZE - pubsz), pub, pubsz);
        memset(pub, 0, DH_SIZE - pubsz);
    }

    for (bytes = 0, ret = 0; bytes < DH_SIZE; bytes += ret) {
        ret = _DkStreamWrite(stream, 0, DH_SIZE - bytes, pub + bytes, NULL, 0);
        if (ret < 0) {
            if (ret == -PAL_ERROR_INTERRUPTED || ret == -PAL_ERROR_TRYAGAIN) {
                ret = 0;
                continue;
            }
            SGX_DBG(DBG_E, "Failed to exchange the secret key via RPC: %ld\n", ret);
            goto out;
        }
    }

    for (bytes = 0, ret = 0 ; bytes < DH_SIZE ; bytes += ret) {
        ret = _DkStreamRead(stream, 0, DH_SIZE - bytes, pub + bytes, NULL, 0);
        if (ret < 0) {
            if (ret == -PAL_ERROR_INTERRUPTED || ret == -PAL_ERROR_TRYAGAIN) {
                ret = 0;
                continue;
            }
            SGX_DBG(DBG_E, "Failed to exchange the secret key via RPC: %ld\n", ret);
            goto out;
        }
    }

    agreesz = sizeof agree;
    ret = lib_DhCalcSecret(&context, pub, DH_SIZE, agree, &agreesz);
    if (ret < 0) {
        SGX_DBG(DBG_E, "Key Exchange: DH CalcSecret failed: %ld\n", ret);
        goto out;
    }

    assert(agreesz > 0 && agreesz <= sizeof agree);

    /*
     * Using SHA256 as a KDF to convert the 128-byte DH secret to a 256-bit AES key.
     * According to the NIST recommendation:
     * https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-56Cr1.pdf,
     * a key derivation function (KDF) can be a secure hash function (e.g., SHA-256),
     * HMAC, or KMAC.
     */
    LIB_SHA256_CONTEXT sha;
    if ((ret = lib_SHA256Init(&sha)) < 0 ||
        (ret = lib_SHA256Update(&sha, agree, agreesz)) < 0 ||
        (ret = lib_SHA256Final(&sha, (uint8_t*)key)) < 0) {
        SGX_DBG(DBG_E, "Failed to derive the session key: %ld\n", ret);
        goto out;
    }


    SGX_DBG(DBG_S, "Key exchange succeeded: %s\n", ALLOCA_BYTES2HEXSTR(*key));
    ret = 0;
out:
    lib_DhFinal(&context);
out_no_final:
    return ret;
}

/*
 * Initalize the request of local report exchange.
 *
 * We refer to this enclave as A and to the other enclave as B, e.g., A is this
 * parent enclave and B is the child enclave in the fork case (for more info,
 * see comments in db_process.c).
 */
int _DkStreamReportRequest(PAL_HANDLE stream, sgx_sign_data_t* data,
                           check_mr_enclave_t check_mr_enclave) {
    __sgx_mem_aligned sgx_target_info_t target_info;
    __sgx_mem_aligned sgx_report_t report;
    uint64_t bytes;
    int64_t ret;

    /* A -> B: targetinfo[A] */
    memset(&target_info, 0, sizeof(target_info));
    memcpy(&target_info.mr_enclave,  &pal_sec.mr_enclave, sizeof(sgx_measurement_t));
    memcpy(&target_info.attributes, &pal_sec.enclave_attributes, sizeof(sgx_attributes_t));

    for (bytes = 0, ret = 0; bytes < SGX_TARGETINFO_FILLED_SIZE; bytes += ret) {
        ret = _DkStreamWrite(stream, 0, SGX_TARGETINFO_FILLED_SIZE - bytes,
                             ((void*)&target_info) + bytes, NULL, 0);
        if (ret < 0) {
            if (ret == -PAL_ERROR_INTERRUPTED || ret == -PAL_ERROR_TRYAGAIN) {
                ret = 0;
                continue;
            }
            SGX_DBG(DBG_E, "Failed to send target info via RPC: %ld\n", ret);
            goto out;
        }
    }

    /* B -> A: report[B -> A] */
    for (bytes = 0, ret = 0 ; bytes < sizeof(report) ; bytes += ret) {
        ret = _DkStreamRead(stream, 0, sizeof(report) - bytes,
                            ((void*)&report) + bytes, NULL, 0);
        if (ret < 0) {
            if (ret == -PAL_ERROR_INTERRUPTED || ret == -PAL_ERROR_TRYAGAIN) {
                ret = 0;
                continue;
            }
            SGX_DBG(DBG_E, "Failed to receive local report via RPC: %ld\n", ret);
            goto out;
        }
    }

    SGX_DBG(DBG_S, "Received local report (mr_enclave = %s)\n",
            ALLOCA_BYTES2HEXSTR(report.body.mr_enclave.m));

    /* Verify report[B -> A] */
    ret = sgx_verify_report(&report);
    if (ret < 0) {
        SGX_DBG(DBG_E, "Failed to verify local report: %ld\n", ret);
        goto out;
    }

    struct pal_enclave_state* remote_state = (void*)&report.body.report_data;
    ret = check_mr_enclave(stream, &report.body.mr_enclave, remote_state);
    if (ret < 0) {
        SGX_DBG(DBG_E, "Failed to check local report: %ld\n", ret);
        goto out;
    }

    if (ret == 1) {
        SGX_DBG(DBG_E, "Not an allowed enclave (mr_enclave = %s). Maybe missing 'sgx.trusted_children' in the manifest file?\n",
                ALLOCA_BYTES2HEXSTR(report.body.mr_enclave.m));
        ret = -PAL_ERROR_DENIED;
        goto out;
    }

    SGX_DBG(DBG_S, "Local attestation succeeded!\n");

    /* A -> B: report[A -> B] */
    memcpy(&target_info.mr_enclave , &report.body.mr_enclave,  sizeof(sgx_measurement_t));
    memcpy(&target_info.attributes, &report.body.attributes, sizeof(sgx_attributes_t));

    ret = sgx_get_report(&target_info, data, &report);
    if (ret < 0) {
        SGX_DBG(DBG_E, "Failed to get local report from CPU: %ld\n", ret);
        goto out;
    }

    for (bytes = 0, ret = 0 ; bytes < sizeof(report) ; bytes += ret) {
        ret = _DkStreamWrite(stream, 0, sizeof(report) - bytes,
                             ((void*)&report) + bytes, NULL, 0);
        if (ret < 0) {
            if (ret == -PAL_ERROR_INTERRUPTED || ret == -PAL_ERROR_TRYAGAIN) {
                ret = 0;
                continue;
            }
            SGX_DBG(DBG_E, "Failed to send local report via RPC: %ld\n", ret);
            goto out;
        }
    }

    return 0;

out:
    DkStreamDelete(stream, 0);
    return ret;
}

/*
 * Respond to the request of local report exchange.
 *
 * We refer to this enclave as B and to the other enclave as A, e.g., B is this
 * child enclave and A is the parent enclave in the fork case (for more info,
 * see comments in db_process.c).
 */
int _DkStreamReportRespond(PAL_HANDLE stream, sgx_sign_data_t* data,
                           check_mr_enclave_t check_mr_enclave) {
    __sgx_mem_aligned sgx_target_info_t target_info;
    __sgx_mem_aligned sgx_report_t report;
    uint64_t bytes;
    int64_t ret;
    memset(&target_info, 0, sizeof(target_info));

    /* A -> B: targetinfo[A] */
    for (bytes = 0, ret = 0 ; bytes < SGX_TARGETINFO_FILLED_SIZE ; bytes += ret) {
        ret = _DkStreamRead(stream, 0, SGX_TARGETINFO_FILLED_SIZE - bytes,
                            ((void*)&target_info) + bytes, NULL, 0);
        if (ret < 0) {
            if (ret == -PAL_ERROR_INTERRUPTED || ret == -PAL_ERROR_TRYAGAIN) {
                ret = 0;
                continue;
            }
            SGX_DBG(DBG_E, "Failed to receive target info via RPC: %ld\n", ret);
            goto out;
        }
    }

    /* B -> A: report[B -> A] */
    ret = sgx_get_report(&target_info, data, &report);
    if (ret < 0) {
        SGX_DBG(DBG_E, "Failed to get local report from CPU: %ld\n", ret);
        goto out;
    }

    for (bytes = 0, ret = 0 ; bytes < sizeof(report) ; bytes += ret) {
        ret = _DkStreamWrite(stream, 0, sizeof(report) - bytes,
                             ((void*)&report) + bytes, NULL, 0);
        if (ret < 0) {
            if (ret == -PAL_ERROR_INTERRUPTED || ret == -PAL_ERROR_TRYAGAIN) {
                ret = 0;
                continue;
            }
            SGX_DBG(DBG_E, "Failed to send local report via PRC: %ld\n", ret);
            goto out;
        }
    }

    /* A -> B: report[A -> B] */
    for (bytes = 0, ret = 0 ; bytes < sizeof(report) ; bytes += ret) {
        ret = _DkStreamRead(stream, 0, sizeof(report) - bytes,
                            ((void*)&report) + bytes, NULL, 0);
        if (ret < 0) {
            if (ret == -PAL_ERROR_INTERRUPTED || ret == -PAL_ERROR_TRYAGAIN) {
                ret = 0;
                continue;
            }
            SGX_DBG(DBG_E, "Failed to receive local report via RPC: %ld\n", ret);
            goto out;
        }
    }

    SGX_DBG(DBG_S, "Received local report (mr_enclave = %s)\n",
            ALLOCA_BYTES2HEXSTR(report.body.mr_enclave.m));

    /* Verify report[A -> B] */
    ret = sgx_verify_report(&report);
    if (ret < 0) {
        SGX_DBG(DBG_E, "Failed to verify local report: %ld\n", ret);
        goto out;
    }

    struct pal_enclave_state* remote_state = (void*)&report.body.report_data;
    ret = check_mr_enclave(stream, &report.body.mr_enclave, remote_state);
    if (ret < 0) {
        SGX_DBG(DBG_E, "Failed to check mr_enclave: %ld\n", ret);
        goto out;
    }

    if (ret == 1) {
        SGX_DBG(DBG_E, "Not an allowed enclave (mr_enclave = %s). Maybe missing 'sgx.trusted_children' in the manifest file?\n",
                ALLOCA_BYTES2HEXSTR(report.body.mr_enclave.m));
        ret = -PAL_ERROR_DENIED;
        goto out;
    }

    SGX_DBG(DBG_S, "Local attestation succeeded!\n");
    return 0;

out:
    DkStreamDelete(stream, 0);
    return ret;
}

int _DkStreamSecureInit(PAL_HANDLE stream, bool is_server, PAL_SESSION_KEY* session_key,
                        LIB_SSL_CONTEXT** out_ssl_ctx) {
    int stream_fd;

    if (IS_HANDLE_TYPE(stream, process))
        stream_fd = stream->process.stream;
    else
        return -PAL_ERROR_BADHANDLE;


    LIB_SSL_CONTEXT* ssl_ctx = malloc(sizeof(*ssl_ctx));
    if (!ssl_ctx)
        return -PAL_ERROR_NOMEM;

    int ret = lib_SSLInit(ssl_ctx, stream_fd, is_server,
                          (const uint8_t*)session_key, sizeof(*session_key),
                          ocall_read, ocall_write);

    if (ret != 0) {
        free(ssl_ctx);
        return ret;
    }

    *out_ssl_ctx = ssl_ctx;
    return 0;
}

int _DkStreamSecureFree(LIB_SSL_CONTEXT* ssl_ctx) {
    lib_SSLFree(ssl_ctx);
    free(ssl_ctx);
    return 0;
}

int _DkStreamSecureRead(LIB_SSL_CONTEXT* ssl_ctx, uint8_t* buf, size_t len) {
    return lib_SSLRead(ssl_ctx, buf, len);
}

int _DkStreamSecureWrite(LIB_SSL_CONTEXT* ssl_ctx, const uint8_t* buf, size_t len) {
    return lib_SSLWrite(ssl_ctx, buf, len);
}