add support for setting algorithm/nfactor for each pool separately, and fix bugged OpenCL thread cleanup (thread shutdown)

algorithm.c

@@ -33,3 +33,8 @@ void set_algorithm_nfactor(algorithm_t* algo, const uint8_t nfactor) {
 
     return;
 }
+
+bool cmp_algorithm(algorithm_t* algo1, algorithm_t* algo2) {
+    return (strcmp(algo1->name, algo2->name) == 0) &&
+          (algo1->nfactor == algo2->nfactor);
+}

algorithm.h

@@ -2,6 +2,7 @@
 #define ALGORITHM_H
 
 #include <inttypes.h>
+#include <stdbool.h>
 
 /* Describes the Scrypt parameters and hashing functions used to mine
  * a specific coin.
@@ -18,4 +19,7 @@ void set_algorithm(algorithm_t* algo, const char* name);
 /* Set to specific N factor. */
 void set_algorithm_nfactor(algorithm_t* algo, const uint8_t nfactor);
 
+/* Compare two algorithm parameters */
+bool cmp_algorithm(algorithm_t* algo1, algorithm_t* algo2);
+
 #endif /* ALGORITHM_H */

api.c

@@ -2856,7 +2856,7 @@ static void minecoin(struct io_data *io_data, __maybe_unused SOCKETTYPE c, __may
 	message(io_data, MSG_MINECOIN, 0, NULL, isjson);
 	io_open = io_add(io_data, isjson ? COMSTR JSON_MINECOIN : _MINECOIN COMSTR);
 
-	root = api_add_string(root, "Hash Method", algorithm->name, false);
+	root = api_add_string(root, "Hash Method", get_devices(0)->algorithm.name, false);
 
 	cg_rlock(&ch_lock);
 	root = api_add_timeval(root, "Current Block Time", &block_timeval, true);

doc/configuration.md

@@ -23,6 +23,9 @@ cryptocurrencies.
 * everything else - currently defaults to `scrypt`, subject to change
   without warning.
 
+### pool-algorithm
+
+Allows choosing the algorithm for a specific pool. See `algorithm`.
 
 ### nfactor
 
@@ -33,6 +36,9 @@ Overrides the default scrypt parameter N, specified as the factor of 2
 
 *Default:* depends on `algorithm`; otherwise `10`.
 
+### pool-nfactor
+
+Overrides the default scrypt parameter N for a specific pool. See `nfactor`.
 
 ## CLI-only options
 

driver-opencl.c

@@ -1094,6 +1094,8 @@ select_cgpu:
 		cgtime(&thr->sick);
 		if (!pthread_cancel(thr->pth)) {
 			applog(LOG_WARNING, "Thread %d still exists, killing it off", thr_id);
+			pthread_join(thr->pth, NULL);
+			thr->cgpu->drv->thread_shutdown(thr);
 		} else
 			applog(LOG_WARNING, "Thread %d no longer exists", thr_id);
 	}
@@ -1120,7 +1122,7 @@ select_cgpu:
 		//free(clState);
 
 		applog(LOG_INFO, "Reinit GPU thread %d", thr_id);
-		clStates[thr_id] = initCl(virtual_gpu, name, sizeof(name));
+		clStates[thr_id] = initCl(virtual_gpu, name, sizeof(name), &cgpu->algorithm);
 		if (!clStates[thr_id]) {
 			applog(LOG_ERR, "Failed to reinit GPU thread %d", thr_id);
 			goto select_cgpu;
@@ -1186,6 +1188,7 @@ static void opencl_detect(bool hotplug)
 			cgpu->threads = 1;
 #endif
 		cgpu->virtual_gpu = i;
+		cgpu->algorithm = *default_algorithm;
 		add_cgpu(cgpu);
 	}
 
@@ -1261,7 +1264,7 @@ static bool opencl_thread_prepare(struct thr_info *thr)
 
 	strcpy(name, "");
 	applog(LOG_INFO, "Init GPU thread %i GPU %i virtual GPU %i", i, gpu, virtual_gpu);
-	clStates[i] = initCl(virtual_gpu, name, sizeof(name));
+	clStates[i] = initCl(virtual_gpu, name, sizeof(name), &cgpu->algorithm);
 	if (!clStates[i]) {
 #ifdef HAVE_CURSES
 		if (use_curses)
@@ -1443,15 +1446,25 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work,
 	return hashes;
 }
 
+// Cleanup OpenCL memory on the GPU
+// Note: This function is not thread-safe (clStates modification not atomic)
 static void opencl_thread_shutdown(struct thr_info *thr)
 {
 	const int thr_id = thr->id;
 	_clState *clState = clStates[thr_id];
+	clStates[thr_id] = NULL;
 
-	clReleaseKernel(clState->kernel);
-	clReleaseProgram(clState->program);
-	clReleaseCommandQueue(clState->commandQueue);
-	clReleaseContext(clState->context);
+	if (clState) {
+		clFinish(clState->commandQueue);
+		clReleaseMemObject(clState->outputBuffer);
+		clReleaseMemObject(clState->CLbuffer0);
+		clReleaseMemObject(clState->padbuffer8);
+		clReleaseKernel(clState->kernel);
+		clReleaseProgram(clState->program);
+		clReleaseCommandQueue(clState->commandQueue);
+		clReleaseContext(clState->context);
+		free(clState);
+	}
 }
 
 struct device_drv opencl_drv = {

miner.h

@@ -475,6 +475,7 @@ struct cgpu_info {
 	cl_uint vwidth;
 	size_t work_size;
 	cl_ulong max_alloc;
+	algorithm_t algorithm;
 
 	int opt_lg, lookup_gap;
 	size_t opt_tc, thread_concurrency;
@@ -548,7 +549,6 @@ struct thread_q {
 struct thr_info {
 	int		id;
 	int		device_thread;
-	bool		primary_thread;
 
 	pthread_t	pth;
 	cgsem_t		sem;
@@ -559,6 +559,7 @@ struct thr_info {
 	struct timeval sick;
 
 	bool	pause;
+	bool	paused;
 	bool	getwork;
 	double	rolling;
 
@@ -1014,8 +1015,7 @@ extern int opt_queue;
 extern int opt_scantime;
 extern int opt_expiry;
 
-extern char *opt_algorithm;
-extern algorithm_t *algorithm;
+extern algorithm_t *default_algorithm;
 
 extern cglock_t control_lock;
 extern pthread_mutex_t hash_lock;
@@ -1219,6 +1219,8 @@ struct pool {
 	proxytypes_t rpc_proxytype;
 	char *rpc_proxy;
 
+	algorithm_t algorithm;
+
 	pthread_mutex_t pool_lock;
 	cglock_t data_lock;
 

ocl.c

@@ -31,6 +31,7 @@
 #include <unistd.h>
 
 #include "findnonce.h"
+#include "algorithm.h"
 #include "ocl.h"
 
 /* FIXME: only here for global config vars, replace with configuration.h
@@ -218,7 +219,7 @@ void patch_opcodes(char *w, unsigned remaining)
 	applog(LOG_DEBUG, "Patched a total of %i BFI_INT instructions", patched);
 }
 
-_clState *initCl(unsigned int gpu, char *name, size_t nameSize)
+_clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *algorithm)
 {
 	_clState *clState = (_clState *)calloc(1, sizeof(_clState));
 	bool patchbfi = false, prog_built = false;
@@ -732,8 +733,9 @@ built:
 	free(binaries);
 	free(binary_sizes);
 
-	applog(LOG_NOTICE, "Initialising kernel %s with%s bitalign, %spatched BFI",
-	       filename, clState->hasBitAlign ? "" : "out", patchbfi ? "" : "un");
+	applog(LOG_NOTICE, "Initialising kernel %s with%s bitalign, %spatched BFI, nfactor %d, n %d",
+	       filename, clState->hasBitAlign ? "" : "out", patchbfi ? "" : "un",
+	       algorithm->nfactor, algorithm->n);
 
 	if (!prog_built) {
 		/* create a cl program executable for all the devices specified */

ocl.h

@@ -34,5 +34,5 @@ typedef struct {
 
 extern char *file_contents(const char *filename, int *length);
 extern int clDevicesNum(void);
-extern _clState *initCl(unsigned int gpu, char *name, size_t nameSize);
+extern _clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *algorithm);
 #endif /* __OCL_H__ */

scrypt.c

@@ -358,7 +358,7 @@ salsa20_8(uint32_t B[16], const uint32_t Bx[16])
  * scratchpad size needs to be at least (bytes):
  * 63 + (128 * r * p) + (256 * r + 64) + (128 * r * N)
  */
-static void scrypt_n_1_1_256_sp(const uint32_t* input, char* scratchpad, uint32_t *ostate)
+static void scrypt_n_1_1_256_sp(const uint32_t* input, char* scratchpad, uint32_t *ostate, uint32_t n)
 {
 	uint32_t * V;
 	uint32_t X[32];
@@ -372,7 +372,7 @@ static void scrypt_n_1_1_256_sp(const uint32_t* input, char* scratchpad, uint32_
 
 	PBKDF2_SHA256_80_128(input, X);
 
-	for (i = 0; i < algorithm->n; i += 2) {
+	for (i = 0; i < n; i += 2) {
 		memcpy(&V[i * 32], X, 128);
 
 		salsa20_8(&X[0], &X[16]);
@@ -383,8 +383,8 @@ static void scrypt_n_1_1_256_sp(const uint32_t* input, char* scratchpad, uint32_
 		salsa20_8(&X[0], &X[16]);
 		salsa20_8(&X[16], &X[0]);
 	}
-	for (i = 0; i < algorithm->n; i += 2) {
-		j = X[16] & (algorithm->n-1);
+	for (i = 0; i < n; i += 2) {
+		j = X[16] & (n-1);
 		p2 = (uint64_t *)(&V[j * 32]);
 		for(k = 0; k < 16; k++)
 			p1[k] ^= p2[k];
@@ -392,7 +392,7 @@ static void scrypt_n_1_1_256_sp(const uint32_t* input, char* scratchpad, uint32_
 		salsa20_8(&X[0], &X[16]);
 		salsa20_8(&X[16], &X[0]);
 
-		j = X[16] & (algorithm->n-1);
+		j = X[16] & (n-1);
 		p2 = (uint64_t *)(&V[j * 32]);
 		for(k = 0; k < 16; k++)
 			p1[k] ^= p2[k];
@@ -404,7 +404,7 @@ static void scrypt_n_1_1_256_sp(const uint32_t* input, char* scratchpad, uint32_
 	PBKDF2_SHA256_80_128_32(input, X, ostate);
 }
 
-void scrypt_regenhash(struct work *work)
+void scrypt_regenhash(struct work *work, uint32_t n)
 {
 	uint32_t data[20];
 	char *scratchbuf;
@@ -414,81 +414,9 @@ void scrypt_regenhash(struct work *work)
 	be32enc_vect(data, (const uint32_t *)work->data, 19);
 	data[19] = htobe32(*nonce);
 
-	scratchbuf = (char *)alloca(algorithm->n * 128 + 512);
-	scrypt_n_1_1_256_sp(data, scratchbuf, ohash);
+	scratchbuf = (char *)alloca(n * 128 + 512);
+	scrypt_n_1_1_256_sp(data, scratchbuf, ohash, n);
 	flip32(ohash, ohash);
 }
 
 static const uint32_t diff1targ = 0x0000ffff;
-
-/* Used externally as confirmation of correct OCL code */
-/* FIXME: find reference in git blame and see why it was present, remove if obsolete */
-/*
-int scrypt_test(unsigned char *pdata, const unsigned char *ptarget, uint32_t nonce)
-{
-	uint32_t tmp_hash7, Htarg = le32toh(((const uint32_t *)ptarget)[7]);
-	uint32_t data[20], ohash[8];
-	char *scratchbuf;
-
-	be32enc_vect(data, (const uint32_t *)pdata, 19);
-	data[19] = htobe32(nonce);
-	scratchbuf = (char *)alloca(SCRATCHBUF_SIZE);
-	scrypt_n_1_1_256_sp(data, scratchbuf, ohash, algorithm->n);
-	tmp_hash7 = be32toh(ohash[7]);
-
-	applog(LOG_DEBUG, "htarget %08lx diff1 %08lx hash %08lx",
-				(long unsigned int)Htarg,
-				(long unsigned int)diff1targ,
-				(long unsigned int)tmp_hash7);
-	if (tmp_hash7 > diff1targ)
-		return -1;
-	if (tmp_hash7 > Htarg)
-		return 0;
-	return 1;
-}
-
-bool scanhash_scrypt(struct thr_info *thr, const unsigned char __maybe_unused *pmidstate,
-		     unsigned char *pdata, unsigned char __maybe_unused *phash1,
-		     unsigned char __maybe_unused *phash, const unsigned char *ptarget,
-		     uint32_t max_nonce, uint32_t *last_nonce, uint32_t n)
-{
-	uint32_t *nonce = (uint32_t *)(pdata + 76);
-	char *scratchbuf;
-	uint32_t data[20];
-	uint32_t tmp_hash7;
-	uint32_t Htarg = le32toh(((const uint32_t *)ptarget)[7]);
-	bool ret = false;
-
-	be32enc_vect(data, (const uint32_t *)pdata, 19);
-
-	scratchbuf = (char *)malloc(SCRATCHBUF_SIZE);
-	if (unlikely(!scratchbuf)) {
-		applog(LOG_ERR, "Failed to malloc scratchbuf in scanhash_scrypt");
-		return ret;
-	}
-
-	while(1) {
-		uint32_t ostate[8];
-
-		*nonce = ++n;
-		data[19] = htobe32(n);
-		scrypt_n_1_1_256_sp(data, scratchbuf, ostate, algorithm->n);
-		tmp_hash7 = be32toh(ostate[7]);
-
-		if (unlikely(tmp_hash7 <= Htarg)) {
-			((uint32_t *)pdata)[19] = htobe32(n);
-			*last_nonce = n;
-			ret = true;
-			break;
-		}
-
-		if (unlikely((n >= max_nonce) || thr->work_restart)) {
-			*last_nonce = n;
-			break;
-		}
-	}
-
-	free(scratchbuf);;
-	return ret;
-}
-*/

scrypt.h

@@ -5,6 +5,6 @@
 
 /* extern int scrypt_test(unsigned char *pdata, const unsigned char *ptarget, */
 /* 			uint32_t nonce); */
-extern void scrypt_regenhash(struct work *work);
+extern void scrypt_regenhash(struct work *work, uint32_t n);
 
 #endif /* SCRYPT_H */

sgminer.c

@@ -98,8 +98,7 @@ int opt_queue = 1;
 int opt_scantime = 7;
 int opt_expiry = 28;
 
-char *opt_algorithm;
-algorithm_t *algorithm;
+algorithm_t *default_algorithm;
 
 static const bool opt_time = true;
 unsigned long long global_hashrate;
@@ -199,6 +198,9 @@ pthread_rwlock_t devices_lock;
 static pthread_mutex_t lp_lock;
 static pthread_cond_t lp_cond;
 
+static pthread_mutex_t algo_switch_lock;
+static pthread_t algo_switch_thr = 0;
+
 pthread_mutex_t restart_lock;
 pthread_cond_t restart_cond;
 
@@ -536,6 +538,9 @@ struct pool *add_pool(void)
 	buf[0] = '\0';
 	pool->name = strdup(buf);
 
+	/* Algorithm */
+	pool->algorithm = *default_algorithm;
+
 	pools = (struct pool **)realloc(pools, sizeof(struct pool *) * (total_pools + 2));
 	pools[total_pools++] = pool;
 	mutex_init(&pool->pool_lock);
@@ -769,6 +774,35 @@ static char *set_url(char *arg)
 	return NULL;
 }
 
+
+static char *set_pool_algorithm(char *arg)
+{
+	struct pool *pool;
+
+	while ((json_array_index + 1) > total_pools)
+		add_pool();
+	pool = pools[json_array_index];
+
+	applog(LOG_DEBUG, "Setting pool %i algorithm to %s", pool->pool_no, arg);
+	set_algorithm(&pool->algorithm, arg);
+
+	return NULL;
+}
+
+static char *set_pool_nfactor(char *arg)
+{
+	struct pool *pool;
+
+	while ((json_array_index + 1) > total_pools)
+		add_pool();
+	pool = pools[json_array_index];
+
+	applog(LOG_DEBUG, "Setting pool %i N-factor to %s", pool->pool_no, arg);
+	set_algorithm_nfactor(&pool->algorithm, (const uint8_t) atoi(arg));
+
+	return NULL;
+}
+
 static char *set_poolname(char *arg)
 {
 	struct pool *pool = get_current_pool();
@@ -1030,17 +1064,17 @@ static void load_temp_cutoffs()
 
 static char *set_algo(const char *arg)
 {
-	set_algorithm(algorithm, arg);
-	applog(LOG_INFO, "Set algorithm to %s", algorithm->name);
+	set_algorithm(default_algorithm, arg);
+	applog(LOG_INFO, "Set default algorithm to %s", default_algorithm->name);
 
 	return NULL;
 }
 
 static char *set_nfactor(const char *arg)
 {
-	set_algorithm_nfactor(algorithm, (uint8_t)atoi(arg));
+	set_algorithm_nfactor(default_algorithm, (const uint8_t) atoi(arg));
 	applog(LOG_INFO, "Set algorithm N-factor to %d (N to %d)",
-	       algorithm->nfactor, algorithm->n);
+	       default_algorithm->nfactor, default_algorithm->n);
 
 	return NULL;
 }
@@ -1392,6 +1426,12 @@ static struct opt_table opt_config_table[] = {
 	OPT_WITH_ARG("--url|-o",
 		     set_url, NULL, NULL,
 		     "URL for bitcoin JSON-RPC server"),
+	OPT_WITH_ARG("--pool-algorithm",
+				 set_pool_algorithm, NULL, NULL,
+				 "Set algorithm for pool"),
+	OPT_WITH_ARG("--pool-nfactor",
+				 set_pool_nfactor, NULL, NULL,
+         "Set N-factor for pool"),
 	OPT_WITH_ARG("--user|-u",
 		     set_user, NULL, NULL,
 		     "Username for bitcoin JSON-RPC server"),
@@ -3779,7 +3819,6 @@ void switch_pools(struct pool *selected)
 	mutex_lock(&lp_lock);
 	pthread_cond_broadcast(&lp_cond);
 	mutex_unlock(&lp_lock);
-
 }
 
 void discard_work(struct work *work)
@@ -4411,8 +4450,8 @@ void write_config(FILE *fcfg)
 	}
 	if (opt_removedisabled)
 		fprintf(fcfg, ",\n\"remove-disabled\" : true");
-	if (strcmp(algorithm->name, "scrypt") != 0)
-		fprintf(fcfg, ",\n\"algorithm\" : \"%s\"", json_escape(algorithm->name));
+	if (strcmp(default_algorithm->name, "scrypt") != 0)
+		fprintf(fcfg, ",\n\"algorithm\" : \"%s\"", json_escape(default_algorithm->name));
 	if (opt_api_allow)
 		fprintf(fcfg, ",\n\"api-allow\" : \"%s\"", json_escape(opt_api_allow));
 	if (strcmp(opt_api_mcast_addr, API_MCAST_ADDR) != 0)
@@ -5984,6 +6023,69 @@ static void gen_stratum_work(struct pool *pool, struct work *work)
 	cgtime(&work->tv_staged);
 }
 
+static void *switch_algo_thread(void *arg)
+{
+	algorithm_t *new_algo = (algorithm_t *) arg;
+	int i;
+
+	pthread_detach(pthread_self());
+
+	applog(LOG_WARNING, "Switching algorithm to %s (%d)",
+		new_algo->name, new_algo->nfactor);
+
+	// TODO: When threads are canceled, they may leak memory, for example
+	// the "work" variable in get_work.
+	rd_lock(&devices_lock);
+	for (i = 0; i < total_devices; i++) {
+		devices[i]->algorithm = *new_algo;
+		reinit_device(devices[i]);
+	}
+	rd_unlock(&devices_lock);
+
+	// Wait for reinit_gpu to finish
+	while (42) {
+		struct thread_q *tq = control_thr[gpur_thr_id].q;
+		bool stop = false;
+		mutex_lock(&tq->mutex);
+		stop = list_empty(&tq->q);
+		mutex_unlock(&tq->mutex);
+		if (stop) break;
+		usleep(50000);
+	}
+
+	mutex_lock(&algo_switch_lock);
+	algo_switch_thr = 0;
+	mutex_unlock(&algo_switch_lock);
+
+	return NULL;
+}
+
+static void wait_to_die(void)
+{
+	mutex_unlock(&algo_switch_lock);
+	pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, NULL);
+	sleep(60);
+	applog(LOG_ERR, "Thread not canceled within 60 seconds");
+}
+
+static void get_work_prepare_thread(struct thr_info *mythr, struct work *work)
+{
+	struct cgpu_info *cgpu = mythr->cgpu;
+
+	mutex_lock(&algo_switch_lock);
+	pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, NULL);
+	if (!cmp_algorithm(&work->pool->algorithm, &cgpu->algorithm)) {
+		// stage work back to queue, we cannot process it yet
+		stage_work(work);
+		if (algo_switch_thr == 0)
+			pthread_create(&algo_switch_thr, NULL, &switch_algo_thread, &work->pool->algorithm);
+		wait_to_die();
+	} else {
+		mutex_unlock(&algo_switch_lock);
+		pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, NULL);
+	}
+}
+
 struct work *get_work(struct thr_info *thr, const int thr_id)
 {
 	struct work *work = NULL;
@@ -6000,6 +6102,9 @@ struct work *get_work(struct thr_info *thr, const int thr_id)
 			wake_gws();
 		}
 	}
+
+	get_work_prepare_thread(thr, work);
+
 	diff_t = time(NULL) - diff_t;
 	/* Since this is a blocking function, we need to add grace time to
 	 * the device's last valid work to not make outages appear to be
@@ -6080,7 +6185,7 @@ static void rebuild_nonce(struct work *work, uint32_t nonce)
 
 	*work_nonce = htole32(nonce);
 
-	scrypt_regenhash(work);
+	scrypt_regenhash(work, work->pool->algorithm.n);
 }
 
 /* For testing a nonce against diff 1 */
@@ -6208,8 +6313,10 @@ static void mt_disable(struct thr_info *mythr, const int thr_id,
 	applog(LOG_WARNING, "Thread %d being disabled", thr_id);
 	mythr->rolling = mythr->cgpu->rolling = 0;
 	applog(LOG_DEBUG, "Waiting on sem in miner thread");
+	mythr->paused = true;
 	cgsem_wait(&mythr->sem);
 	applog(LOG_WARNING, "Thread %d being re-enabled", thr_id);
+	mythr->paused = false;
 	drv->thread_enable(mythr);
 }
 
@@ -6226,7 +6333,7 @@ static void hash_sole_work(struct thr_info *mythr)
 	struct sgminer_stats *pool_stats;
 	/* Try to cycle approximately 5 times before each log update */
 	const long cycle = opt_log_interval / 5 ? 5 : 1;
-	const bool primary = (!mythr->device_thread) || mythr->primary_thread;
+	const bool primary = mythr->device_thread == 0;
 	struct timeval diff, sdiff, wdiff = {0, 0};
 	uint32_t max_nonce = drv->can_limit_work(mythr);
 	int64_t hashes_done = 0;
@@ -7831,6 +7938,7 @@ int main(int argc, char *argv[])
 	rwlock_init(&netacc_lock);
 	rwlock_init(&mining_thr_lock);
 	rwlock_init(&devices_lock);
+	mutex_init(&algo_switch_lock);
 
 	mutex_init(&lp_lock);
 	if (unlikely(pthread_cond_init(&lp_cond, NULL)))
@@ -7883,8 +7991,8 @@ int main(int argc, char *argv[])
 #endif
 
 	/* Default algorithm specified in algorithm.c ATM */
-	algorithm = (algorithm_t *)alloca(sizeof(algorithm_t));
-	set_algorithm(algorithm, "scrypt");
+	default_algorithm = (algorithm_t *)alloca(sizeof(algorithm_t));
+	set_algorithm(default_algorithm, "scrypt");
 
 	devcursor = 8;
 	logstart = devcursor + 1;