diff --git a/scst/include/scst.h b/scst/include/scst.h
index bb566b720..efe2bf5ec 100644
--- a/scst/include/scst.h
+++ b/scst/include/scst.h
@@ -1947,14 +1947,6 @@ struct scst_cmd {
 	/* Set if cmd is queued as hw pending */
 	unsigned int cmd_hw_pending:1;
 
-	/*
-	 * Set, if for this cmd required to not have any IO or FS calls on
-	 * memory buffers allocations, at least for READ and WRITE commands.
-	 * Needed for cases like file systems mounted over scst_local's
-	 * devices.
-	 */
-	unsigned noio_mem_alloc:1;
-
 	/*
 	 * Set if the target driver wants to alloc data buffers on its own.
 	 * In this case tgt_alloc_data_buf() must be provided in the target
@@ -2039,6 +2031,15 @@ struct scst_cmd {
 	/* cmd's async flags */
 	unsigned long cmd_flags;
 
+	/*
+	 * GFP mask with which memory on READ or WRITE data path for this cmd
+	 * should be allocated, if the current context is not ATOMIC. Useful
+	 * for cases like if this cmd required to not have any IO or FS calls
+	 * on allocations, like for file systems mounted over scst_local's
+	 * devices.
+	 */
+	gfp_t cmd_gfp_mask;
+
 	/* Keeps status of cmd's status/data delivery to remote initiator */
 	int delivery_status;
 
@@ -2569,8 +2570,16 @@ struct scst_tgt_dev {
 	struct scst_device *dev; /* to save extra dereferences */
 	uint64_t lun;		 /* to save extra dereferences */
 
-	gfp_t gfp_mask;
+	/*
+	 * Extra flags in GFP mask for data buffers allocations of this
+	 * tgt_dev's cmds
+	 */
+	gfp_t tgt_dev_gfp_mask;
+
+	/* SGV pool from which buffers of this tgt_dev's cmds should be allocated */
 	struct sgv_pool *pool;
+
+	/* Max number of allowed in this tgt_dev SG segments */
 	int max_sg_cnt;
 
 	/*************************************************************
@@ -3545,16 +3554,13 @@ static inline void scst_cmd_set_tgt_sn(struct scst_cmd *cmd, uint32_t tgt_sn)
 }
 
 /*
- * Get/Set functions for noio_mem_alloc
+ * Forbids for this cmd any IO-causing allocations.
+ *
+ * !! Must be called before scst_cmd_init_done() !!
  */
-static inline bool scst_cmd_get_noio_mem_alloc(struct scst_cmd *cmd)
-{
-	return cmd->noio_mem_alloc;
-}
-
 static inline void scst_cmd_set_noio_mem_alloc(struct scst_cmd *cmd)
 {
-	cmd->noio_mem_alloc = 1;
+	cmd->cmd_gfp_mask = GFP_NOIO;
 }
 
 /*
@@ -4082,9 +4088,9 @@ int scst_get_buf_full(struct scst_cmd *cmd, uint8_t **buf);
 int scst_get_buf_full_sense(struct scst_cmd *cmd, uint8_t **buf);
 void scst_put_buf_full(struct scst_cmd *cmd, uint8_t *buf);
 
-static inline gfp_t scst_cmd_get_gfp_flags(struct scst_cmd *cmd)
+static inline gfp_t scst_cmd_get_gfp_mask(struct scst_cmd *cmd)
 {
-	return cmd->noio_mem_alloc ? GFP_NOIO : GFP_KERNEL;
+	return cmd->cmd_gfp_mask;
 }
 
 #if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 23) && !defined(BACKPORT_LINUX_WORKQUEUE_TO_2_6_19)
diff --git a/scst/src/dev_handlers/scst_vdisk.c b/scst/src/dev_handlers/scst_vdisk.c
index 5bb919a15..b96707b67 100644
--- a/scst/src/dev_handlers/scst_vdisk.c
+++ b/scst/src/dev_handlers/scst_vdisk.c
@@ -1034,14 +1034,12 @@ static enum compl_status_e vdisk_synchronize_cache(struct vdisk_cmd_params *p)
 		cmd->completed = 1;
 		cmd->scst_cmd_done(cmd, SCST_CMD_STATE_DEFAULT,
 				   SCST_CONTEXT_SAME);
-		vdisk_fsync(p, loff, data_len, dev,
-			    scst_cmd_get_gfp_flags(cmd), NULL, true);
+		vdisk_fsync(p, loff, data_len, dev, cmd->cmd_gfp_mask, NULL, true);
 		/* ToDo: vdisk_fsync() error processing */
 		scst_cmd_put(cmd);
 		res = RUNNING_ASYNC;
 	} else {
-		vdisk_fsync(p, loff, data_len, dev,
-			    scst_cmd_get_gfp_flags(cmd), cmd, true);
+		vdisk_fsync(p, loff, data_len, dev, cmd->cmd_gfp_mask, cmd, true);
 		res = RUNNING_ASYNC;
 	}
 
@@ -1057,8 +1055,7 @@ static enum compl_status_e vdisk_exec_start_stop(struct vdisk_cmd_params *p)
 
 	TRACE_ENTRY();
 
-	vdisk_fsync(p, 0, virt_dev->file_size, dev, scst_cmd_get_gfp_flags(cmd),
-		cmd, false);
+	vdisk_fsync(p, 0, virt_dev->file_size, dev, cmd->cmd_gfp_mask, cmd, false);
 
 	TRACE_EXIT();
 	return CMD_SUCCEEDED;
@@ -1472,7 +1469,7 @@ static int fileio_alloc_and_parse(struct scst_cmd *cmd)
 
 	TRACE_ENTRY();
 
-	p = kmem_cache_zalloc(vdisk_cmd_param_cachep, GFP_KERNEL);
+	p = kmem_cache_zalloc(vdisk_cmd_param_cachep, cmd->cmd_gfp_mask);
 	if (!p) {
 		scst_set_busy(cmd);
 		goto out_err;
@@ -2136,7 +2133,7 @@ static int vdisk_unmap_range(struct scst_cmd *cmd,
 	if (virt_dev->blockio) {
 #if LINUX_VERSION_CODE > KERNEL_VERSION(2, 6, 27)
 		struct inode *inode = fd->f_dentry->d_inode;
-		gfp_t gfp = scst_cmd_get_gfp_flags(cmd);
+		gfp_t gfp = cmd->cmd_gfp_mask;
 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2, 6, 31)
 		err = blkdev_issue_discard(inode->i_bdev, start_lba, blocks, gfp);
 #elif LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 35)       \
@@ -2373,7 +2370,7 @@ static enum compl_status_e vdisk_exec_inquiry(struct vdisk_cmd_params *p)
 
 	TRACE_ENTRY();
 
-	buf = kzalloc(INQ_BUF_SZ, GFP_KERNEL);
+	buf = kzalloc(INQ_BUF_SZ, cmd->cmd_gfp_mask);
 	if (buf == NULL) {
 		scst_set_busy(cmd);
 		goto out;
@@ -2934,7 +2931,7 @@ static enum compl_status_e vdisk_exec_mode_sense(struct vdisk_cmd_params *p)
 
 	TRACE_ENTRY();
 
-	buf = kzalloc(MSENSE_BUF_SZ, GFP_KERNEL);
+	buf = kzalloc(MSENSE_BUF_SZ, cmd->cmd_gfp_mask);
 	if (buf == NULL) {
 		scst_set_busy(cmd);
 		goto out;
@@ -3651,7 +3648,7 @@ static struct iovec *vdisk_alloc_iv(struct scst_cmd *cmd,
 		p->iv_count = 0;
 		/* It can't be called in atomic context */
 		p->iv = (iv_count <= ARRAY_SIZE(p->small_iv)) ? p->small_iv :
-			kmalloc(sizeof(*p->iv) * iv_count, GFP_KERNEL);
+			kmalloc(sizeof(*p->iv) * iv_count, cmd->cmd_gfp_mask);
 		if (p->iv == NULL) {
 			PRINT_ERROR("Unable to allocate iv (%d)", iv_count);
 			goto out;
@@ -3924,7 +3921,7 @@ out_sync:
 	/* O_DSYNC flag is used for WT devices */
 	if (p->fua)
 		vdisk_fsync(p, loff, scst_cmd_get_data_len(cmd), cmd->dev,
-			    scst_cmd_get_gfp_flags(cmd), cmd, false);
+			    cmd->cmd_gfp_mask, cmd, false);
 out:
 	TRACE_EXIT();
 	return CMD_SUCCEEDED;
@@ -4027,7 +4024,7 @@ static void blockio_exec_rw(struct vdisk_cmd_params *p, bool write, bool fua)
 	int need_new_bio;
 	struct scst_blockio_work *blockio_work;
 	int bios = 0;
-	gfp_t gfp_mask = scst_cmd_get_gfp_flags(cmd);
+	gfp_t gfp_mask = cmd->cmd_gfp_mask;
 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 39)
 	struct blk_plug plug;
 #endif
@@ -4244,7 +4241,7 @@ static int vdisk_blockio_flush(struct block_device *bdev, gfp_t gfp_mask,
 
 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 37)
 	if (async) {
-		struct bio *bio = bio_alloc(GFP_KERNEL, 0);
+		struct bio *bio = bio_alloc(gfp_mask, 0);
 		if (bio == NULL) {
 			res = -ENOMEM;
 			goto out_rep;
@@ -4308,7 +4305,7 @@ static enum compl_status_e fileio_exec_verify(struct vdisk_cmd_params *p)
 	sBUG_ON(virt_dev->blockio);
 
 	if (vdisk_fsync(p, loff, data_len, cmd->dev,
-			scst_cmd_get_gfp_flags(cmd), cmd, false) != 0)
+			cmd->cmd_gfp_mask, cmd, false) != 0)
 		goto out;
 
 	/*
diff --git a/scst/src/scst_lib.c b/scst/src/scst_lib.c
index 14c7558ab..9cbe8c743 100644
--- a/scst/src/scst_lib.c
+++ b/scst/src/scst_lib.c
@@ -1343,7 +1343,7 @@ static inline void tm_dbg_deinit_tgt_dev(struct scst_tgt_dev *tgt_dev) {}
 int scst_alloc_sense(struct scst_cmd *cmd, int atomic)
 {
 	int res = 0;
-	gfp_t gfp_mask = atomic ? GFP_ATOMIC : (GFP_KERNEL|__GFP_NOFAIL);
+	gfp_t gfp_mask = atomic ? GFP_ATOMIC : (cmd->cmd_gfp_mask|__GFP_NOFAIL);
 
 	TRACE_ENTRY();
 
@@ -4482,7 +4482,7 @@ static struct scst_cmd *scst_create_prepare_internal_cmd(
 {
 	struct scst_cmd *res;
 	int rc;
-	gfp_t gfp_mask = scst_cmd_atomic(orig_cmd) ? GFP_ATOMIC : GFP_KERNEL;
+	gfp_t gfp_mask = scst_cmd_atomic(orig_cmd) ? GFP_ATOMIC : orig_cmd->cmd_gfp_mask;
 	unsigned long flags;
 
 	TRACE_ENTRY();
@@ -5318,6 +5318,7 @@ int scst_pre_init_cmd(struct scst_cmd *cmd, const uint8_t *cdb,
 	cmd->start_time = jiffies;
 	atomic_set(&cmd->cmd_ref, 1);
 	cmd->cmd_threads = &scst_main_cmd_threads;
+	cmd->cmd_gfp_mask = GFP_KERNEL;
 	INIT_LIST_HEAD(&cmd->mgmt_cmd_list);
 	cmd->cdb = cmd->cdb_buf;
 	cmd->queue_type = SCST_CMD_QUEUE_SIMPLE;
@@ -5652,7 +5653,7 @@ int scst_alloc_space(struct scst_cmd *cmd)
 
 	TRACE_ENTRY();
 
-	gfp_mask = tgt_dev->gfp_mask | (atomic ? GFP_ATOMIC : GFP_KERNEL);
+	gfp_mask = tgt_dev->tgt_dev_gfp_mask | (atomic ? GFP_ATOMIC : cmd->cmd_gfp_mask);
 
 	flags = atomic ? SGV_POOL_NO_ALLOC_ON_CACHE_MISS : 0;
 	if (cmd->no_sgv)
@@ -5933,7 +5934,7 @@ int scst_scsi_exec_async(struct scst_cmd *cmd, void *data,
 	struct request *rq;
 	struct scsi_io_context *sioc;
 	int write = (cmd->data_direction & SCST_DATA_WRITE) ? WRITE : READ;
-	gfp_t gfp = scst_cmd_get_gfp_flags(cmd);
+	gfp_t gfp = cmd->cmd_gfp_mask;
 	int cmd_len = cmd->cdb_len;
 
 	sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
diff --git a/scst/src/scst_mem.c b/scst/src/scst_mem.c
index 46e51185f..3b5dba47b 100644
--- a/scst/src/scst_mem.c
+++ b/scst/src/scst_mem.c
@@ -98,7 +98,7 @@ static inline bool sgv_pool_clustered(const struct sgv_pool *pool)
 
 void scst_sgv_pool_use_norm(struct scst_tgt_dev *tgt_dev)
 {
-	tgt_dev->gfp_mask = __GFP_NOWARN;
+	tgt_dev->tgt_dev_gfp_mask = __GFP_NOWARN;
 	tgt_dev->pool = sgv_norm_pool;
 	tgt_dev->tgt_dev_clust_pool = 0;
 }
@@ -106,7 +106,7 @@ void scst_sgv_pool_use_norm(struct scst_tgt_dev *tgt_dev)
 void scst_sgv_pool_use_norm_clust(struct scst_tgt_dev *tgt_dev)
 {
 	TRACE_MEM("%s", "Use clustering");
-	tgt_dev->gfp_mask = __GFP_NOWARN;
+	tgt_dev->tgt_dev_gfp_mask = __GFP_NOWARN;
 	tgt_dev->pool = sgv_norm_clust_pool;
 	tgt_dev->tgt_dev_clust_pool = 1;
 }
@@ -114,7 +114,7 @@ void scst_sgv_pool_use_norm_clust(struct scst_tgt_dev *tgt_dev)
 void scst_sgv_pool_use_dma(struct scst_tgt_dev *tgt_dev)
 {
 	TRACE_MEM("%s", "Use ISA DMA memory");
-	tgt_dev->gfp_mask = __GFP_NOWARN | GFP_DMA;
+	tgt_dev->tgt_dev_gfp_mask = __GFP_NOWARN | GFP_DMA;
 	tgt_dev->pool = sgv_dma_pool;
 	tgt_dev->tgt_dev_clust_pool = 0;
 }
diff --git a/scst/src/scst_targ.c b/scst/src/scst_targ.c
index 7a2c2fcb7..8c447e256 100644
--- a/scst/src/scst_targ.c
+++ b/scst/src/scst_targ.c
@@ -226,7 +226,7 @@ int scst_rx_cmd_prealloced(struct scst_cmd *cmd, struct scst_session *sess,
 	unsigned int cdb_len, bool atomic)
 {
 	int res;
-	gfp_t gfp_mask = atomic ? GFP_ATOMIC : GFP_KERNEL;
+	gfp_t gfp_mask = atomic ? GFP_ATOMIC : cmd->cmd_gfp_mask;
 
 	TRACE_ENTRY();
 
@@ -273,7 +273,7 @@ struct scst_cmd *scst_rx_cmd(struct scst_session *sess,
 	unsigned int cdb_len, bool atomic)
 {
 	struct scst_cmd *cmd;
-	gfp_t gfp_mask = atomic ? GFP_ATOMIC : GFP_KERNEL;
+	gfp_t gfp_mask = atomic ? GFP_ATOMIC : GFP_NOIO;
 
 	TRACE_ENTRY();
 
@@ -2708,7 +2708,7 @@ static int scst_do_real_exec(struct scst_cmd *cmd)
 	rc = scst_exec_req(scsi_dev, cmd->cdb, cmd->cdb_len,
 			cmd->data_direction, cmd->sg, cmd->bufflen,
 			cmd->sg_cnt, cmd->timeout, cmd->retries, cmd,
-			scst_pass_through_cmd_done, GFP_KERNEL);
+			scst_pass_through_cmd_done, cmd->cmd_gfp_mask);
 #else
 	rc = scst_scsi_exec_async(cmd, cmd, scst_pass_through_cmd_done);
 #endif