xref: /titanic_41/usr/src/uts/common/crypto/api/kcf_random.c (revision d89fccd8788afe1e920f842edd883fe192a1b8fe)

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#pragma ident	"%Z%%M%	%I%	%E% SMI"

/*
 * This file implements the interfaces that the /dev/random
 * driver uses for read(2), write(2) and poll(2) on /dev/random or
 * /dev/urandom. It also implements the kernel API - random_add_entropy(),
 * random_get_pseudo_bytes() and random_get_bytes().
 *
 * We periodically collect random bits from providers which are registered
 * with the Kernel Cryptographic Framework (kCF) as capable of random
 * number generation. The random bits are maintained in a cache and
 * it is used for high quality random numbers (/dev/random) requests.
 * We pick a provider and call its SPI routine, if the cache does not have
 * enough bytes to satisfy a request.
 *
 * /dev/urandom requests use a software-based generator algorithm that uses the
 * random bits in the cache as a seed. We create one pseudo-random generator
 * (for /dev/urandom) per possible CPU on the system, and use it,
 * kmem-magazine-style, to avoid cache line contention.
 *
 * LOCKING HIERARCHY:
 *	1) rmp->rm_lock protects the per-cpu pseudo-random generators.
 * 	2) rndpool_lock protects the high-quality randomness pool.
 *		It may be locked while a rmp->rm_lock is held.
 *
 * A history note: The kernel API and the software-based algorithms in this
 * file used to be part of the /dev/random driver.
 */

#include <sys/types.h>
#include <sys/conf.h>
#include <sys/sunddi.h>
#include <sys/disp.h>
#include <sys/modctl.h>
#include <sys/ddi.h>
#include <sys/crypto/common.h>
#include <sys/crypto/api.h>
#include <sys/crypto/impl.h>
#include <sys/crypto/sched_impl.h>
#include <sys/random.h>
#include <sys/sha1.h>
#include <sys/time.h>
#include <sys/sysmacros.h>
#include <sys/cpuvar.h>
#include <sys/taskq.h>

#define	RNDPOOLSIZE		1024	/* Pool size in bytes */
#define	MINEXTRACTBYTES		20
#define	MAXEXTRACTBYTES		1024
#define	PRNG_MAXOBLOCKS		1310720	/* Max output block per prng key */
#define	TIMEOUT_INTERVAL	5	/* Periodic mixing interval in secs */

typedef enum    extract_type {
	NONBLOCK_EXTRACT,
	BLOCKING_EXTRACT,
	ALWAYS_EXTRACT
} extract_type_t;

/*
 * Hash-algo generic definitions. For now, they are SHA1's. We use SHA1
 * routines directly instead of using k-API because we can't return any
 * error code in /dev/urandom case and we can get an error using k-API
 * if a mechanism is disabled.
 */
#define	HASHSIZE		20
#define	HASH_CTX		SHA1_CTX
#define	HashInit(ctx)		SHA1Init((ctx))
#define	HashUpdate(ctx, p, s)	SHA1Update((ctx), (p), (s))
#define	HashFinal(d, ctx)	SHA1Final((d), (ctx))

/* HMAC-SHA1 */
#define	HMAC_KEYSIZE			20
#define	HMAC_BLOCK_SIZE			64
#define	HMAC_KEYSCHED			sha1keysched_t
#define	SET_ENCRYPT_KEY(k, s, ks)	hmac_key((k), (s), (ks))
#define	HMAC_ENCRYPT(ks, p, s, d)	hmac_encr((ks), (uint8_t *)(p), s, d)

/* HMAC-SHA1 "keyschedule" */
typedef struct sha1keysched_s {
	SHA1_CTX ictx;
	SHA1_CTX octx;
} sha1keysched_t;

/*
 * Cache of random bytes implemented as a circular buffer. findex and rindex
 * track the front and back of the circular buffer.
 */
uint8_t rndpool[RNDPOOLSIZE];
static int findex, rindex;
static int rnbyte_cnt;		/* Number of bytes in the cache */

static kmutex_t rndpool_lock;	/* protects r/w accesses to the cache, */
				/* and the global variables */
static kcondvar_t rndpool_read_cv; /* serializes poll/read syscalls */
static int num_waiters;		/* #threads waiting to read from /dev/random */

static struct pollhead rnd_pollhead;
static timeout_id_t kcf_rndtimeout_id;
static crypto_mech_type_t rngmech_type = CRYPTO_MECH_INVALID;
rnd_stats_t rnd_stats;

static void rndc_addbytes(uint8_t *, size_t);
static void rndc_getbytes(uint8_t *ptr, size_t len);
static void rnd_handler(void *);
static void rnd_alloc_magazines();
static void hmac_key(uint8_t *, size_t, void *);
static void hmac_encr(void *, uint8_t *, size_t, uint8_t *);


void
kcf_rnd_init()
{
	hrtime_t ts;
	time_t now;

	mutex_init(&rndpool_lock, NULL, MUTEX_DEFAULT, NULL);
	cv_init(&rndpool_read_cv, NULL, CV_DEFAULT, NULL);

	/*
	 * Add bytes to the cache using
	 * . 2 unpredictable times: high resolution time since the boot-time,
	 *   and the current time-of-the day.
	 * This is used only to make the timeout value in the timer
	 * unpredictable.
	 */
	ts = gethrtime();
	rndc_addbytes((uint8_t *)&ts, sizeof (ts));

	(void) drv_getparm(TIME, &now);
	rndc_addbytes((uint8_t *)&now, sizeof (now));

	rnbyte_cnt = 0;
	findex = rindex = 0;
	num_waiters = 0;
	rngmech_type = KCF_MECHID(KCF_MISC_CLASS, 0);

	rnd_alloc_magazines();
}

/*
 * Return TRUE if at least one provider exists that can
 * supply random numbers.
 */
boolean_t
kcf_rngprov_check(void)
{
	int rv;
	kcf_provider_desc_t *pd;

	if ((pd = kcf_get_mech_provider(rngmech_type, NULL, &rv,
	    NULL, CRYPTO_FG_RANDOM, B_FALSE, 0)) != NULL) {
		KCF_PROV_REFRELE(pd);
		return (B_TRUE);
	} else
		return (B_FALSE);
}

/*
 * Pick a software-based provider and submit a request to seed
 * its random number generator.
 */
static void
rngprov_seed(uint8_t *buf, int len)
{
	kcf_provider_desc_t *pd = NULL;
	kcf_req_params_t params;

	if (kcf_get_sw_prov(rngmech_type, &pd, B_FALSE) == CRYPTO_SUCCESS) {
		KCF_WRAP_RANDOM_OPS_PARAMS(&params, KCF_OP_RANDOM_SEED,
		    pd->pd_sid, buf, len);
		(void) kcf_submit_request(pd, NULL, NULL, &params, B_FALSE);
		KCF_PROV_REFRELE(pd);
	}
}

/* Boot-time tunable for experimentation. */
int kcf_limit_hwrng = 1;


/*
 * This routine is called for blocking reads.
 *
 * The argument from_user_api indicates whether the caller is
 * from userland coming via the /dev/random driver.
 *
 * The argument is_taskq_thr indicates whether the caller is
 * the taskq thread dispatched by the timeout handler routine.
 * In this case, we cycle through all the providers
 * submitting a request to each provider to generate random numbers.
 *
 * For other cases, we pick a provider and submit a request to generate
 * random numbers. We retry using another provider if we get an error.
 *
 * Returns the number of bytes that are written to 'ptr'. Returns -1
 * if no provider is found. ptr and need are unchanged.
 */
static int
rngprov_getbytes(uint8_t *ptr, size_t need, boolean_t from_user_api,
    boolean_t is_taskq_thr)
{
	int rv;
	int prov_cnt = 0;
	int total_bytes = 0;
	kcf_provider_desc_t *pd;
	kcf_req_params_t params;
	kcf_prov_tried_t *list = NULL;

	while ((pd = kcf_get_mech_provider(rngmech_type, NULL, &rv,
	    list, CRYPTO_FG_RANDOM, B_FALSE, 0)) != NULL) {

		prov_cnt++;
		/*
		 * Typically a hardware RNG is a multi-purpose
		 * crypto card and hence we do not want to overload the card
		 * just for random numbers. The following check is to prevent
		 * a user process from hogging the hardware RNG. Note that we
		 * still use the hardware RNG from the periodically run
		 * taskq thread.
		 */
		if (pd->pd_prov_type == CRYPTO_HW_PROVIDER && from_user_api &&
		    kcf_limit_hwrng == 1) {
			ASSERT(is_taskq_thr == B_FALSE);
			goto try_next;
		}

		KCF_WRAP_RANDOM_OPS_PARAMS(&params, KCF_OP_RANDOM_GENERATE,
		    pd->pd_sid, ptr, need);
		rv = kcf_submit_request(pd, NULL, NULL, &params, B_FALSE);
		ASSERT(rv != CRYPTO_QUEUED);

		if (rv == CRYPTO_SUCCESS) {
			total_bytes += need;
			if (is_taskq_thr)
				rndc_addbytes(ptr, need);
			else {
				KCF_PROV_REFRELE(pd);
				break;
			}
		}

		if (is_taskq_thr || rv != CRYPTO_SUCCESS) {
try_next:
			/* Add pd to the linked list of providers tried. */
			if (kcf_insert_triedlist(&list, pd, KM_SLEEP) == NULL) {
				KCF_PROV_REFRELE(pd);
				break;
			}
		}

	}

	if (list != NULL)
		kcf_free_triedlist(list);

	if (prov_cnt == 0) { /* no provider could be found. */
		return (-1);
	}

	return (total_bytes);
}

static void
notify_done(void *arg, int rv)
{
	uchar_t *rndbuf = arg;

	if (rv == CRYPTO_SUCCESS)
		rndc_addbytes(rndbuf, MINEXTRACTBYTES);

	bzero(rndbuf, MINEXTRACTBYTES);
	kmem_free(rndbuf, MINEXTRACTBYTES);
}

/*
 * Cycle through all the providers submitting a request to each provider
 * to generate random numbers. This is called for the modes - NONBLOCK_EXTRACT
 * and ALWAYS_EXTRACT.
 *
 * Returns the number of bytes that are written to 'ptr'. Returns -1
 * if no provider is found. ptr and len are unchanged.
 */
static int
rngprov_getbytes_nblk(uint8_t *ptr, size_t len, boolean_t from_user_api)
{
	int rv, blen, total_bytes;
	uchar_t *rndbuf;
	kcf_provider_desc_t *pd;
	kcf_req_params_t params;
	crypto_call_req_t req;
	kcf_prov_tried_t *list = NULL;
	int prov_cnt = 0;

	blen = 0;
	total_bytes = 0;
	req.cr_flag = CRYPTO_SKIP_REQID;
	req.cr_callback_func = notify_done;

	while ((pd = kcf_get_mech_provider(rngmech_type, NULL, &rv,
	    list, CRYPTO_FG_RANDOM, CHECK_RESTRICT(&req), 0)) != NULL) {

		prov_cnt ++;
		switch (pd->pd_prov_type) {
		case CRYPTO_HW_PROVIDER:
			/* See comments in rngprov_getbytes() */
			if (from_user_api && kcf_limit_hwrng == 1)
				goto try_next;

			/*
			 * We have to allocate a buffer here as we can not
			 * assume that the input buffer will remain valid
			 * when the callback comes. We use a fixed size buffer
			 * to simplify the book keeping.
			 */
			rndbuf = kmem_alloc(MINEXTRACTBYTES, KM_NOSLEEP);
			if (rndbuf == NULL) {
				KCF_PROV_REFRELE(pd);
				if (list != NULL)
					kcf_free_triedlist(list);
				return (total_bytes);
			}
			req.cr_callback_arg = rndbuf;
			KCF_WRAP_RANDOM_OPS_PARAMS(&params,
			    KCF_OP_RANDOM_GENERATE,
			    pd->pd_sid, rndbuf, MINEXTRACTBYTES);
			break;

		case CRYPTO_SW_PROVIDER:
			/*
			 * We do not need to allocate a buffer in the software
			 * provider case as there is no callback involved. We
			 * avoid any extra data copy by directly passing 'ptr'.
			 */
			KCF_WRAP_RANDOM_OPS_PARAMS(&params,
			    KCF_OP_RANDOM_GENERATE,
			    pd->pd_sid, ptr, len);
			break;
		}

		rv = kcf_submit_request(pd, NULL, &req, &params, B_FALSE);
		if (rv == CRYPTO_SUCCESS) {
			switch (pd->pd_prov_type) {
			case CRYPTO_HW_PROVIDER:
				/*
				 * Since we have the input buffer handy,
				 * we directly copy to it rather than
				 * adding to the pool.
				 */
				blen = min(MINEXTRACTBYTES, len);
				bcopy(rndbuf, ptr, blen);
				if (len < MINEXTRACTBYTES)
					rndc_addbytes(rndbuf + len,
					    MINEXTRACTBYTES - len);
				ptr += blen;
				len -= blen;
				total_bytes += blen;
				break;

			case CRYPTO_SW_PROVIDER:
				total_bytes += len;
				len = 0;
				break;
			}
		}

		/*
		 * We free the buffer in the callback routine
		 * for the CRYPTO_QUEUED case.
		 */
		if (pd->pd_prov_type == CRYPTO_HW_PROVIDER &&
		    rv != CRYPTO_QUEUED) {
			bzero(rndbuf, MINEXTRACTBYTES);
			kmem_free(rndbuf, MINEXTRACTBYTES);
		}

		if (len == 0) {
			KCF_PROV_REFRELE(pd);
			break;
		}

		if (rv != CRYPTO_SUCCESS) {
try_next:
			/* Add pd to the linked list of providers tried. */
			if (kcf_insert_triedlist(&list, pd, KM_NOSLEEP) ==
			    NULL) {
				KCF_PROV_REFRELE(pd);
				break;
			}
		}
	}

	if (list != NULL) {
		kcf_free_triedlist(list);
	}

	if (prov_cnt == 0) { /* no provider could be found. */
		return (-1);
	}

	return (total_bytes);
}

static void
rngprov_task(void *arg)
{
	int len = (int)(uintptr_t)arg;
	uchar_t tbuf[MAXEXTRACTBYTES];

	ASSERT(len <= MAXEXTRACTBYTES);
	if (rngprov_getbytes(tbuf, len, B_FALSE, B_TRUE) == -1) {
		cmn_err(CE_WARN, "No randomness provider enabled for "
		    "/dev/random. Use cryptoadm(1M) to enable a provider.");
	}
}

/*
 * Returns "len" random or pseudo-random bytes in *ptr.
 * Will block if not enough random bytes are available and the
 * call is blocking.
 *
 * Called with rndpool_lock held (allowing caller to do optimistic locking;
 * releases the lock before return).
 */
static int
rnd_get_bytes(uint8_t *ptr, size_t len, extract_type_t how,
    boolean_t from_user_api)
{
	int bytes;
	size_t got;

	ASSERT(mutex_owned(&rndpool_lock));
	/*
	 * Check if the request can be satisfied from the cache
	 * of random bytes.
	 */
	if (len <= rnbyte_cnt) {
		rndc_getbytes(ptr, len);
		mutex_exit(&rndpool_lock);
		return (0);
	}
	mutex_exit(&rndpool_lock);

	switch (how) {
	case BLOCKING_EXTRACT:
		if ((got = rngprov_getbytes(ptr, len, from_user_api,
		    B_FALSE)) == -1)
			break;	/* No provider found */

		if (got == len)
			return (0);
		len -= got;
		ptr += got;
		break;

	case NONBLOCK_EXTRACT:
	case ALWAYS_EXTRACT:
		if ((got = rngprov_getbytes_nblk(ptr, len,
		    from_user_api)) == -1) {
			/* No provider found */
			if (how == NONBLOCK_EXTRACT) {
				return (EAGAIN);
			}
		} else {
			if (got == len)
				return (0);
			len -= got;
			ptr += got;
		}
		if (how == NONBLOCK_EXTRACT && (rnbyte_cnt < len))
			return (EAGAIN);
		break;
	}

	mutex_enter(&rndpool_lock);
	while (len > 0) {
		if (how == BLOCKING_EXTRACT) {
			/* Check if there is enough */
			while (rnbyte_cnt < MINEXTRACTBYTES) {
				num_waiters++;
				if (cv_wait_sig(&rndpool_read_cv,
				    &rndpool_lock) == 0) {
					num_waiters--;
					mutex_exit(&rndpool_lock);
					return (EINTR);
				}
				num_waiters--;
			}
		}

		/* Figure out how many bytes to extract */
		bytes = min(len, rnbyte_cnt);
		rndc_getbytes(ptr, bytes);

		len -= bytes;
		ptr += bytes;

		if (len > 0 && how == ALWAYS_EXTRACT) {
			/*
			 * There are not enough bytes, but we can not block.
			 * This only happens in the case of /dev/urandom which
			 * runs an additional generation algorithm. So, there
			 * is no problem.
			 */
			while (len > 0) {
				*ptr = rndpool[findex];
				ptr++; len--;
				rindex = findex = (findex + 1) &
				    (RNDPOOLSIZE - 1);
			}
			break;
		}
	}

	mutex_exit(&rndpool_lock);
	return (0);
}

int
kcf_rnd_get_bytes(uint8_t *ptr, size_t len, boolean_t noblock,
    boolean_t from_user_api)
{
	extract_type_t how;
	int error;

	how = noblock ? NONBLOCK_EXTRACT : BLOCKING_EXTRACT;
	mutex_enter(&rndpool_lock);
	if ((error = rnd_get_bytes(ptr, len, how, from_user_api)) != 0)
		return (error);

	BUMP_RND_STATS(rs_rndOut, len);
	return (0);
}

/*
 * Revisit this if the structs grow or we come up with a better way
 * of cache-line-padding structures.
 */
#define	RND_CPU_CACHE_SIZE	64
#define	RND_CPU_PAD_SIZE	RND_CPU_CACHE_SIZE*5
#define	RND_CPU_PAD (RND_CPU_PAD_SIZE - \
	(sizeof (kmutex_t) + 3*sizeof (uint8_t *) + sizeof (HMAC_KEYSCHED) + \
	sizeof (uint64_t) + 3*sizeof (uint32_t) + sizeof (rnd_stats_t)))

/*
 * Per-CPU random state.  Somewhat like like kmem's magazines, this provides
 * a per-CPU instance of the pseudo-random generator.  We have it much easier
 * than kmem, as we can afford to "leak" random bits if a CPU is DR'ed out.
 *
 * Note that this usage is preemption-safe; a thread
 * entering a critical section remembers which generator it locked
 * and unlocks the same one; should it be preempted and wind up running on
 * a different CPU, there will be a brief period of increased contention
 * before it exits the critical section but nothing will melt.
 */
typedef struct rndmag_s
{
	kmutex_t	rm_lock;
	uint8_t 	*rm_buffer;	/* Start of buffer */
	uint8_t		*rm_eptr;	/* End of buffer */
	uint8_t		*rm_rptr;	/* Current read pointer */
	HMAC_KEYSCHED 	rm_ks;		/* seed */
	uint64_t 	rm_counter;	/* rotating counter for extracting */
	uint32_t	rm_oblocks;	/* time to rekey? */
	uint32_t	rm_ofuzz;	/* Rekey backoff state */
	uint32_t	rm_olimit;	/* Hard rekey limit */
	rnd_stats_t	rm_stats;	/* Per-CPU Statistics */
	uint8_t		rm_pad[RND_CPU_PAD];
} rndmag_t;

/*
 * Generate random bytes for /dev/urandom by encrypting a
 * rotating counter with a key created from bytes extracted
 * from the pool.  A maximum of PRNG_MAXOBLOCKS output blocks
 * is generated before a new key is obtained.
 *
 * Note that callers to this routine are likely to assume it can't fail.
 *
 * Called with rmp locked; releases lock.
 */
static int
rnd_generate_pseudo_bytes(rndmag_t *rmp, uint8_t *ptr, size_t len)
{
	size_t bytes = len;
	int nblock, size;
	uint32_t oblocks;
	uint8_t digest[HASHSIZE];

	ASSERT(mutex_owned(&rmp->rm_lock));

	/* Nothing is being asked */
	if (len == 0) {
		mutex_exit(&rmp->rm_lock);
		return (0);
	}

	nblock = howmany(len, HASHSIZE);

	rmp->rm_oblocks += nblock;
	oblocks = rmp->rm_oblocks;

	do {
		if (oblocks >= rmp->rm_olimit) {
			hrtime_t timestamp;
			uint8_t key[HMAC_KEYSIZE];

			/*
			 * Contention-avoiding rekey: see if
			 * the pool is locked, and if so, wait a bit.
			 * Do an 'exponential back-in' to ensure we don't
			 * run too long without rekey.
			 */
			if (rmp->rm_ofuzz) {
				/*
				 * Decaying exponential back-in for rekey.
				 */
				if ((rnbyte_cnt < MINEXTRACTBYTES) ||
				    (!mutex_tryenter(&rndpool_lock))) {
					rmp->rm_olimit += rmp->rm_ofuzz;
					rmp->rm_ofuzz >>= 1;
					goto punt;
				}
			} else {
				mutex_enter(&rndpool_lock);
			}

			/* Get a new chunk of entropy */
			(void) rnd_get_bytes(key, HMAC_KEYSIZE,
			    ALWAYS_EXTRACT, B_FALSE);

			/* Set up key */
			SET_ENCRYPT_KEY(key, HMAC_KEYSIZE, &rmp->rm_ks);

			/* Get new counter value by encrypting timestamp */
			timestamp = gethrtime();
			HMAC_ENCRYPT(&rmp->rm_ks, &timestamp,
			    sizeof (timestamp), digest);
			rmp->rm_olimit = PRNG_MAXOBLOCKS/2;
			rmp->rm_ofuzz = PRNG_MAXOBLOCKS/4;
			bcopy(digest, &rmp->rm_counter, sizeof (uint64_t));
			oblocks = 0;
			rmp->rm_oblocks = nblock;
		}
punt:
		/* Hash counter to produce prn stream */
		if (bytes >= HASHSIZE) {
			size = HASHSIZE;
			HMAC_ENCRYPT(&rmp->rm_ks, &rmp->rm_counter,
			    sizeof (rmp->rm_counter), ptr);
		} else {
			size = min(bytes, HASHSIZE);
			HMAC_ENCRYPT(&rmp->rm_ks, &rmp->rm_counter,
			    sizeof (rmp->rm_counter), digest);
			bcopy(digest, ptr, size);
		}
		ptr += size;
		bytes -= size;
		rmp->rm_counter++;
		oblocks++;
		nblock--;
	} while (bytes > 0);

	mutex_exit(&rmp->rm_lock);
	return (0);
}

/*
 * Per-CPU Random magazines.
 */
static rndmag_t *rndmag;
static uint8_t	*rndbuf;
static size_t 	rndmag_total;
/*
 * common/os/cpu.c says that platform support code can shrinkwrap
 * max_ncpus.  On the off chance that we get loaded very early, we
 * read it exactly once, to copy it here.
 */
static uint32_t	random_max_ncpus = 0;

/*
 * Boot-time tunables, for experimentation.
 */
size_t	rndmag_threshold = 2560;
size_t	rndbuf_len = 5120;
size_t	rndmag_size = 1280;


int
kcf_rnd_get_pseudo_bytes(uint8_t *ptr, size_t len)
{
	rndmag_t *rmp;
	uint8_t *cptr, *eptr;

	/*
	 * Anyone who asks for zero bytes of randomness should get slapped.
	 */
	ASSERT(len > 0);

	/*
	 * Fast path.
	 */
	for (;;) {
		rmp = &rndmag[CPU->cpu_seqid];
		mutex_enter(&rmp->rm_lock);

		/*
		 * Big requests bypass buffer and tail-call the
		 * generate routine directly.
		 */
		if (len > rndmag_threshold) {
			BUMP_CPU_RND_STATS(rmp, rs_urndOut, len);
			return (rnd_generate_pseudo_bytes(rmp, ptr, len));
		}

		cptr = rmp->rm_rptr;
		eptr = cptr + len;

		if (eptr <= rmp->rm_eptr) {
			rmp->rm_rptr = eptr;
			bcopy(cptr, ptr, len);
			BUMP_CPU_RND_STATS(rmp, rs_urndOut, len);
			mutex_exit(&rmp->rm_lock);

			return (0);
		}
		/*
		 * End fast path.
		 */
		rmp->rm_rptr = rmp->rm_buffer;
		/*
		 * Note:  We assume the generate routine always succeeds
		 * in this case (because it does at present..)
		 * It also always releases rm_lock.
		 */
		(void) rnd_generate_pseudo_bytes(rmp, rmp->rm_buffer,
		    rndbuf_len);
	}
}

/*
 * We set up (empty) magazines for all of max_ncpus, possibly wasting a
 * little memory on big systems that don't have the full set installed.
 * See above;  "empty" means "rptr equal to eptr"; this will trigger the
 * refill path in rnd_get_pseudo_bytes above on the first call for each CPU.
 *
 * TODO: make rndmag_size tunable at run time!
 */
static void
rnd_alloc_magazines()
{
	rndmag_t *rmp;
	int i;

	rndbuf_len = roundup(rndbuf_len, HASHSIZE);
	if (rndmag_size < rndbuf_len)
		rndmag_size = rndbuf_len;
	rndmag_size = roundup(rndmag_size, RND_CPU_CACHE_SIZE);

	random_max_ncpus = max_ncpus;
	rndmag_total = rndmag_size * random_max_ncpus;

	rndbuf = kmem_alloc(rndmag_total, KM_SLEEP);
	rndmag = kmem_zalloc(sizeof (rndmag_t) * random_max_ncpus, KM_SLEEP);

	for (i = 0; i < random_max_ncpus; i++) {
		uint8_t *buf;

		rmp = &rndmag[i];
		mutex_init(&rmp->rm_lock, NULL, MUTEX_DRIVER, NULL);

		buf = rndbuf + i * rndmag_size;

		rmp->rm_buffer = buf;
		rmp->rm_eptr = buf + rndbuf_len;
		rmp->rm_rptr = buf + rndbuf_len;
		rmp->rm_oblocks = 1;
	}
}

void
kcf_rnd_schedule_timeout(boolean_t do_mech2id)
{
	clock_t ut;	/* time in microseconds */

	if (do_mech2id)
		rngmech_type = crypto_mech2id(SUN_RANDOM);

	/*
	 * The new timeout value is taken from the buffer of random bytes.
	 * We're merely reading the first 32 bits from the buffer here, not
	 * consuming any random bytes.
	 * The timeout multiplier value is a random value between 0.5 sec and
	 * 1.544480 sec (0.5 sec + 0xFF000 microseconds).
	 * The new timeout is TIMEOUT_INTERVAL times that multiplier.
	 */
	ut = 500000 + (clock_t)((((uint32_t)rndpool[findex]) << 12) & 0xFF000);
	kcf_rndtimeout_id = timeout(rnd_handler, NULL,
	    TIMEOUT_INTERVAL * drv_usectohz(ut));
}

/*
 * &rnd_pollhead is passed in *phpp in order to indicate the calling thread
 * will block. When enough random bytes are available, later, the timeout
 * handler routine will issue the pollwakeup() calls.
 */
void
kcf_rnd_chpoll(int anyyet, short *reventsp, struct pollhead **phpp)
{
	/*
	 * Sampling of rnbyte_cnt is an atomic
	 * operation. Hence we do not need any locking.
	 */
	if (rnbyte_cnt >= MINEXTRACTBYTES) {
		*reventsp |= (POLLIN | POLLRDNORM);
	} else {
		*reventsp = 0;
		if (!anyyet)
			*phpp = &rnd_pollhead;
	}
}

/*ARGSUSED*/
static void
rnd_handler(void *arg)
{
	int len = 0;

	if (num_waiters > 0)
		len = MAXEXTRACTBYTES;
	else if (rnbyte_cnt < RNDPOOLSIZE)
		len = MINEXTRACTBYTES;

	if (len > 0) {
		(void) taskq_dispatch(system_taskq, rngprov_task,
		    (void *)(uintptr_t)len, TQ_NOSLEEP);
	} else if (!kcf_rngprov_check()) {
		cmn_err(CE_WARN, "No randomness provider enabled for "
		    "/dev/random. Use cryptoadm(1M) to enable a provider.");
	}

	mutex_enter(&rndpool_lock);
	/*
	 * Wake up threads waiting in poll() or for enough accumulated
	 * random bytes to read from /dev/random. In case a poll() is
	 * concurrent with a read(), the polling process may be woken up
	 * indicating that enough randomness is now available for reading,
	 * and another process *steals* the bits from the pool, causing the
	 * subsequent read() from the first process to block. It is acceptable
	 * since the blocking will eventually end, after the timeout
	 * has expired enough times to honor the read.
	 *
	 * Note - Since we hold the rndpool_lock across the pollwakeup() call
	 * we MUST NOT grab the rndpool_lock in kcf_rndchpoll().
	 */
	if (rnbyte_cnt >= MINEXTRACTBYTES)
		pollwakeup(&rnd_pollhead, POLLIN | POLLRDNORM);

	if (num_waiters > 0)
		cv_broadcast(&rndpool_read_cv);
	mutex_exit(&rndpool_lock);

	kcf_rnd_schedule_timeout(B_FALSE);
}

/* Hashing functions */

static void
hmac_key(uint8_t *key, size_t keylen, void *buf)
{
	uint32_t *ip, *op;
	uint32_t ipad[HMAC_BLOCK_SIZE/sizeof (uint32_t)];
	uint32_t opad[HMAC_BLOCK_SIZE/sizeof (uint32_t)];
	HASH_CTX *icontext, *ocontext;
	int i;
	int nints;

	icontext = buf;
	ocontext = (SHA1_CTX *)((uint8_t *)buf + sizeof (HASH_CTX));

	bzero((uchar_t *)ipad, HMAC_BLOCK_SIZE);
	bzero((uchar_t *)opad, HMAC_BLOCK_SIZE);
	bcopy(key, (uchar_t *)ipad, keylen);
	bcopy(key, (uchar_t *)opad, keylen);

	/*
	 * XOR key with ipad (0x36) and opad (0x5c) as defined
	 * in RFC 2104.
	 */
	ip = ipad;
	op = opad;
	nints = HMAC_BLOCK_SIZE/sizeof (uint32_t);

	for (i = 0; i < nints; i++) {
		ip[i] ^= 0x36363636;
		op[i] ^= 0x5c5c5c5c;
	}

	/* Perform hash with ipad */
	HashInit(icontext);
	HashUpdate(icontext, (uchar_t *)ipad, HMAC_BLOCK_SIZE);

	/* Perform hash with opad */
	HashInit(ocontext);
	HashUpdate(ocontext, (uchar_t *)opad, HMAC_BLOCK_SIZE);
}

static void
hmac_encr(void *ctx, uint8_t *ptr, size_t len, uint8_t *digest)
{
	HASH_CTX *saved_contexts;
	HASH_CTX icontext;
	HASH_CTX ocontext;

	saved_contexts = (HASH_CTX *)ctx;
	icontext = saved_contexts[0];
	ocontext = saved_contexts[1];

	HashUpdate(&icontext, ptr, len);
	HashFinal(digest, &icontext);

	/*
	 * Perform Hash(K XOR OPAD, DIGEST), where DIGEST is the
	 * Hash(K XOR IPAD, DATA).
	 */
	HashUpdate(&ocontext, digest, HASHSIZE);
	HashFinal(digest, &ocontext);
}


static void
rndc_addbytes(uint8_t *ptr, size_t len)
{
	ASSERT(ptr != NULL && len > 0);
	ASSERT(rnbyte_cnt <= RNDPOOLSIZE);

	mutex_enter(&rndpool_lock);
	while ((len > 0) && (rnbyte_cnt < RNDPOOLSIZE)) {
		rndpool[rindex] ^= *ptr;
		ptr++; len--;
		rindex = (rindex + 1) & (RNDPOOLSIZE - 1);
		rnbyte_cnt++;
	}

	/* Handle buffer full case */
	while (len > 0) {
		rndpool[rindex] ^= *ptr;
		ptr++; len--;
		findex = rindex = (rindex + 1) & (RNDPOOLSIZE - 1);
	}
	mutex_exit(&rndpool_lock);
}

/*
 * Caller should check len <= rnbyte_cnt under the
 * rndpool_lock before calling.
 */
static void
rndc_getbytes(uint8_t *ptr, size_t len)
{
	ASSERT(MUTEX_HELD(&rndpool_lock));
	ASSERT(len <= rnbyte_cnt && rnbyte_cnt <= RNDPOOLSIZE);

	BUMP_RND_STATS(rs_rndcOut, len);

	while (len > 0) {
		*ptr = rndpool[findex];
		ptr++; len--;
		findex = (findex + 1) & (RNDPOOLSIZE - 1);
		rnbyte_cnt--;
	}
}

/* Random number exported entry points */

/*
 * Mix the supplied bytes into the entropy pool of a kCF
 * RNG provider.
 */
/* ARGSUSED */
int
random_add_entropy(uint8_t *ptr, size_t len, uint16_t entropy_est)
{
	if (len < 1)
		return (-1);

	rngprov_seed(ptr, len);

	return (0);
}

/*
 * Get bytes from the /dev/urandom generator. This function
 * always succeeds. Returns 0.
 */
int
random_get_pseudo_bytes(uint8_t *ptr, size_t len)
{
	ASSERT(!mutex_owned(&rndpool_lock));

	if (len < 1)
		return (0);
	return (kcf_rnd_get_pseudo_bytes(ptr, len));
}

/*
 * Get bytes from the /dev/random generator. Returns 0
 * on success. Returns EAGAIN if there is insufficient entropy.
 */
int
random_get_bytes(uint8_t *ptr, size_t len)
{
	ASSERT(!mutex_owned(&rndpool_lock));

	if (len < 1)
		return (0);
	return (kcf_rnd_get_bytes(ptr, len, B_TRUE, B_FALSE));
}