gennvm: Generic NVM manager
authorMatias Bjørling <m@bjorling.me>
Wed, 28 Oct 2015 18:54:56 +0000 (19:54 +0100)
committerJens Axboe <axboe@fb.com>
Thu, 29 Oct 2015 07:21:42 +0000 (16:21 +0900)
The implementation for Open-Channel SSDs is divided into media
management and targets. This patch implements a generic media manager
for open-channel SSDs. After a media manager has been initialized,
single or multiple targets can be instantiated with the media managed as
the backend.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
drivers/lightnvm/Kconfig
drivers/lightnvm/Makefile
drivers/lightnvm/gennvm.c [new file with mode: 0644]
drivers/lightnvm/gennvm.h [new file with mode: 0644]

index d4f309f127cda114dffc596ac816e381f255e72d..2a27971eae1af84a9112b15d75bcc1f13c2c1a91 100644 (file)
@@ -25,4 +25,11 @@ config NVM_DEBUG
 
        It is required to create/remove targets without IOCTLs.
 
+config NVM_GENNVM
+       tristate "Generic NVM manager for Open-Channel SSDs"
+       ---help---
+       NVM media manager for Open-Channel SSDs that offload management
+       functionality to device, while keeping data placement and garbage
+       collection decisions on the host.
+
 endif # NVM
index 38185e990d5dfa039237cac7a51a8b9b7b830925..e2428e04e7b82cb46976de01862e7e9bd59d96e5 100644 (file)
@@ -3,3 +3,4 @@
 #
 
 obj-$(CONFIG_NVM)              := core.o
+obj-$(CONFIG_NVM_GENNVM)       += gennvm.o
diff --git a/drivers/lightnvm/gennvm.c b/drivers/lightnvm/gennvm.c
new file mode 100644 (file)
index 0000000..ae1fb2b
--- /dev/null
@@ -0,0 +1,485 @@
+/*
+ * Copyright (C) 2015 Matias Bjorling <m@bjorling.me>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version
+ * 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; see the file COPYING.  If not, write to
+ * the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139,
+ * USA.
+ *
+ * Implementation of a generic nvm manager for Open-Channel SSDs.
+ */
+
+#include "gennvm.h"
+
+static void gennvm_blocks_free(struct nvm_dev *dev)
+{
+       struct gen_nvm *gn = dev->mp;
+       struct gen_lun *lun;
+       int i;
+
+       gennvm_for_each_lun(gn, lun, i) {
+               if (!lun->vlun.blocks)
+                       break;
+               vfree(lun->vlun.blocks);
+       }
+}
+
+static void gennvm_luns_free(struct nvm_dev *dev)
+{
+       struct gen_nvm *gn = dev->mp;
+
+       kfree(gn->luns);
+}
+
+static int gennvm_luns_init(struct nvm_dev *dev, struct gen_nvm *gn)
+{
+       struct gen_lun *lun;
+       int i;
+
+       gn->luns = kcalloc(dev->nr_luns, sizeof(struct gen_lun), GFP_KERNEL);
+       if (!gn->luns)
+               return -ENOMEM;
+
+       gennvm_for_each_lun(gn, lun, i) {
+               spin_lock_init(&lun->vlun.lock);
+               INIT_LIST_HEAD(&lun->free_list);
+               INIT_LIST_HEAD(&lun->used_list);
+               INIT_LIST_HEAD(&lun->bb_list);
+
+               lun->reserved_blocks = 2; /* for GC only */
+               lun->vlun.id = i;
+               lun->vlun.lun_id = i % dev->luns_per_chnl;
+               lun->vlun.chnl_id = i / dev->luns_per_chnl;
+               lun->vlun.nr_free_blocks = dev->blks_per_lun;
+       }
+       return 0;
+}
+
+static int gennvm_block_bb(u32 lun_id, void *bb_bitmap, unsigned int nr_blocks,
+                                                               void *private)
+{
+       struct gen_nvm *gn = private;
+       struct gen_lun *lun = &gn->luns[lun_id];
+       struct nvm_block *blk;
+       int i;
+
+       if (unlikely(bitmap_empty(bb_bitmap, nr_blocks)))
+               return 0;
+
+       i = -1;
+       while ((i = find_next_bit(bb_bitmap, nr_blocks, i + 1)) < nr_blocks) {
+               blk = &lun->vlun.blocks[i];
+               if (!blk) {
+                       pr_err("gennvm: BB data is out of bounds.\n");
+                       return -EINVAL;
+               }
+
+               list_move_tail(&blk->list, &lun->bb_list);
+       }
+
+       return 0;
+}
+
+static int gennvm_block_map(u64 slba, u32 nlb, __le64 *entries, void *private)
+{
+       struct nvm_dev *dev = private;
+       struct gen_nvm *gn = dev->mp;
+       sector_t max_pages = dev->total_pages * (dev->sec_size >> 9);
+       u64 elba = slba + nlb;
+       struct gen_lun *lun;
+       struct nvm_block *blk;
+       u64 i;
+       int lun_id;
+
+       if (unlikely(elba > dev->total_pages)) {
+               pr_err("gennvm: L2P data from device is out of bounds!\n");
+               return -EINVAL;
+       }
+
+       for (i = 0; i < nlb; i++) {
+               u64 pba = le64_to_cpu(entries[i]);
+
+               if (unlikely(pba >= max_pages && pba != U64_MAX)) {
+                       pr_err("gennvm: L2P data entry is out of bounds!\n");
+                       return -EINVAL;
+               }
+
+               /* Address zero is a special one. The first page on a disk is
+                * protected. It often holds internal device boot
+                * information.
+                */
+               if (!pba)
+                       continue;
+
+               /* resolve block from physical address */
+               lun_id = div_u64(pba, dev->sec_per_lun);
+               lun = &gn->luns[lun_id];
+
+               /* Calculate block offset into lun */
+               pba = pba - (dev->sec_per_lun * lun_id);
+               blk = &lun->vlun.blocks[div_u64(pba, dev->sec_per_blk)];
+
+               if (!blk->type) {
+                       /* at this point, we don't know anything about the
+                        * block. It's up to the FTL on top to re-etablish the
+                        * block state
+                        */
+                       list_move_tail(&blk->list, &lun->used_list);
+                       blk->type = 1;
+                       lun->vlun.nr_free_blocks--;
+               }
+       }
+
+       return 0;
+}
+
+static int gennvm_blocks_init(struct nvm_dev *dev, struct gen_nvm *gn)
+{
+       struct gen_lun *lun;
+       struct nvm_block *block;
+       sector_t lun_iter, blk_iter, cur_block_id = 0;
+       int ret;
+
+       gennvm_for_each_lun(gn, lun, lun_iter) {
+               lun->vlun.blocks = vzalloc(sizeof(struct nvm_block) *
+                                                       dev->blks_per_lun);
+               if (!lun->vlun.blocks)
+                       return -ENOMEM;
+
+               for (blk_iter = 0; blk_iter < dev->blks_per_lun; blk_iter++) {
+                       block = &lun->vlun.blocks[blk_iter];
+
+                       INIT_LIST_HEAD(&block->list);
+
+                       block->lun = &lun->vlun;
+                       block->id = cur_block_id++;
+
+                       /* First block is reserved for device */
+                       if (unlikely(lun_iter == 0 && blk_iter == 0))
+                               continue;
+
+                       list_add_tail(&block->list, &lun->free_list);
+               }
+
+               if (dev->ops->get_bb_tbl) {
+                       ret = dev->ops->get_bb_tbl(dev->q, lun->vlun.id,
+                                       dev->blks_per_lun, gennvm_block_bb, gn);
+                       if (ret)
+                               pr_err("gennvm: could not read BB table\n");
+               }
+       }
+
+       if (dev->ops->get_l2p_tbl) {
+               ret = dev->ops->get_l2p_tbl(dev->q, 0, dev->total_pages,
+                                                       gennvm_block_map, dev);
+               if (ret) {
+                       pr_err("gennvm: could not read L2P table.\n");
+                       pr_warn("gennvm: default block initialization");
+               }
+       }
+
+       return 0;
+}
+
+static int gennvm_register(struct nvm_dev *dev)
+{
+       struct gen_nvm *gn;
+       int ret;
+
+       gn = kzalloc(sizeof(struct gen_nvm), GFP_KERNEL);
+       if (!gn)
+               return -ENOMEM;
+
+       gn->nr_luns = dev->nr_luns;
+       dev->mp = gn;
+
+       ret = gennvm_luns_init(dev, gn);
+       if (ret) {
+               pr_err("gennvm: could not initialize luns\n");
+               goto err;
+       }
+
+       ret = gennvm_blocks_init(dev, gn);
+       if (ret) {
+               pr_err("gennvm: could not initialize blocks\n");
+               goto err;
+       }
+
+       return 1;
+err:
+       kfree(gn);
+       return ret;
+}
+
+static void gennvm_unregister(struct nvm_dev *dev)
+{
+       gennvm_blocks_free(dev);
+       gennvm_luns_free(dev);
+       kfree(dev->mp);
+       dev->mp = NULL;
+}
+
+static struct nvm_block *gennvm_get_blk(struct nvm_dev *dev,
+                               struct nvm_lun *vlun, unsigned long flags)
+{
+       struct gen_lun *lun = container_of(vlun, struct gen_lun, vlun);
+       struct nvm_block *blk = NULL;
+       int is_gc = flags & NVM_IOTYPE_GC;
+
+       spin_lock(&vlun->lock);
+
+       if (list_empty(&lun->free_list)) {
+               pr_err_ratelimited("gennvm: lun %u have no free pages available",
+                                                               lun->vlun.id);
+               spin_unlock(&vlun->lock);
+               goto out;
+       }
+
+       while (!is_gc && lun->vlun.nr_free_blocks < lun->reserved_blocks) {
+               spin_unlock(&vlun->lock);
+               goto out;
+       }
+
+       blk = list_first_entry(&lun->free_list, struct nvm_block, list);
+       list_move_tail(&blk->list, &lun->used_list);
+       blk->type = 1;
+
+       lun->vlun.nr_free_blocks--;
+
+       spin_unlock(&vlun->lock);
+out:
+       return blk;
+}
+
+static void gennvm_put_blk(struct nvm_dev *dev, struct nvm_block *blk)
+{
+       struct nvm_lun *vlun = blk->lun;
+       struct gen_lun *lun = container_of(vlun, struct gen_lun, vlun);
+
+       spin_lock(&vlun->lock);
+
+       switch (blk->type) {
+       case 1:
+               list_move_tail(&blk->list, &lun->free_list);
+               lun->vlun.nr_free_blocks++;
+               blk->type = 0;
+               break;
+       case 2:
+               list_move_tail(&blk->list, &lun->bb_list);
+               break;
+       default:
+               WARN_ON_ONCE(1);
+               pr_err("gennvm: erroneous block type (%lu -> %u)\n",
+                                                       blk->id, blk->type);
+               list_move_tail(&blk->list, &lun->bb_list);
+       }
+
+       spin_unlock(&vlun->lock);
+}
+
+static void gennvm_addr_to_generic_mode(struct nvm_dev *dev, struct nvm_rq *rqd)
+{
+       int i;
+
+       if (rqd->nr_pages > 1) {
+               for (i = 0; i < rqd->nr_pages; i++)
+                       rqd->ppa_list[i] = addr_to_generic_mode(dev,
+                                                       rqd->ppa_list[i]);
+       } else {
+               rqd->ppa_addr = addr_to_generic_mode(dev, rqd->ppa_addr);
+       }
+}
+
+static void gennvm_generic_to_addr_mode(struct nvm_dev *dev, struct nvm_rq *rqd)
+{
+       int i;
+
+       if (rqd->nr_pages > 1) {
+               for (i = 0; i < rqd->nr_pages; i++)
+                       rqd->ppa_list[i] = generic_to_addr_mode(dev,
+                                                       rqd->ppa_list[i]);
+       } else {
+               rqd->ppa_addr = generic_to_addr_mode(dev, rqd->ppa_addr);
+       }
+}
+
+static int gennvm_submit_io(struct nvm_dev *dev, struct nvm_rq *rqd)
+{
+       if (!dev->ops->submit_io)
+               return 0;
+
+       /* Convert address space */
+       gennvm_generic_to_addr_mode(dev, rqd);
+
+       rqd->dev = dev;
+       return dev->ops->submit_io(dev->q, rqd);
+}
+
+static void gennvm_blk_set_type(struct nvm_dev *dev, struct ppa_addr *ppa,
+                                                               int type)
+{
+       struct gen_nvm *gn = dev->mp;
+       struct gen_lun *lun;
+       struct nvm_block *blk;
+
+       if (unlikely(ppa->g.ch > dev->nr_chnls ||
+                                       ppa->g.lun > dev->luns_per_chnl ||
+                                       ppa->g.blk > dev->blks_per_lun)) {
+               WARN_ON_ONCE(1);
+               pr_err("gennvm: ppa broken (ch: %u > %u lun: %u > %u blk: %u > %u",
+                               ppa->g.ch, dev->nr_chnls,
+                               ppa->g.lun, dev->luns_per_chnl,
+                               ppa->g.blk, dev->blks_per_lun);
+               return;
+       }
+
+       lun = &gn->luns[ppa->g.lun * ppa->g.ch];
+       blk = &lun->vlun.blocks[ppa->g.blk];
+
+       /* will be moved to bb list on put_blk from target */
+       blk->type = type;
+}
+
+/* mark block bad. It is expected the target recover from the error. */
+static void gennvm_mark_blk_bad(struct nvm_dev *dev, struct nvm_rq *rqd)
+{
+       int i;
+
+       if (!dev->ops->set_bb)
+               return;
+
+       if (dev->ops->set_bb(dev->q, rqd, 1))
+               return;
+
+       gennvm_addr_to_generic_mode(dev, rqd);
+
+       /* look up blocks and mark them as bad */
+       if (rqd->nr_pages > 1)
+               for (i = 0; i < rqd->nr_pages; i++)
+                       gennvm_blk_set_type(dev, &rqd->ppa_list[i], 2);
+       else
+               gennvm_blk_set_type(dev, &rqd->ppa_addr, 2);
+}
+
+static int gennvm_end_io(struct nvm_rq *rqd, int error)
+{
+       struct nvm_tgt_instance *ins = rqd->ins;
+       int ret = 0;
+
+       switch (error) {
+       case NVM_RSP_SUCCESS:
+               break;
+       case NVM_RSP_ERR_EMPTYPAGE:
+               break;
+       case NVM_RSP_ERR_FAILWRITE:
+               gennvm_mark_blk_bad(rqd->dev, rqd);
+       default:
+               ret++;
+       }
+
+       ret += ins->tt->end_io(rqd, error);
+
+       return ret;
+}
+
+static int gennvm_erase_blk(struct nvm_dev *dev, struct nvm_block *blk,
+                                                       unsigned long flags)
+{
+       int plane_cnt = 0, pl_idx, ret;
+       struct ppa_addr addr;
+       struct nvm_rq rqd;
+
+       if (!dev->ops->erase_block)
+               return 0;
+
+       addr = block_to_ppa(dev, blk);
+
+       if (dev->plane_mode == NVM_PLANE_SINGLE) {
+               rqd.nr_pages = 1;
+               rqd.ppa_addr = addr;
+       } else {
+               plane_cnt = (1 << dev->plane_mode);
+               rqd.nr_pages = plane_cnt;
+
+               rqd.ppa_list = nvm_dev_dma_alloc(dev, GFP_KERNEL,
+                                                       &rqd.dma_ppa_list);
+               if (!rqd.ppa_list) {
+                       pr_err("gennvm: failed to allocate dma memory\n");
+                       return -ENOMEM;
+               }
+
+               for (pl_idx = 0; pl_idx < plane_cnt; pl_idx++) {
+                       addr.g.pl = pl_idx;
+                       rqd.ppa_list[pl_idx] = addr;
+               }
+       }
+
+       gennvm_generic_to_addr_mode(dev, &rqd);
+
+       ret = dev->ops->erase_block(dev->q, &rqd);
+
+       if (plane_cnt)
+               nvm_dev_dma_free(dev, rqd.ppa_list, rqd.dma_ppa_list);
+
+       return ret;
+}
+
+static struct nvm_lun *gennvm_get_lun(struct nvm_dev *dev, int lunid)
+{
+       struct gen_nvm *gn = dev->mp;
+
+       return &gn->luns[lunid].vlun;
+}
+
+static void gennvm_free_blocks_print(struct nvm_dev *dev)
+{
+       struct gen_nvm *gn = dev->mp;
+       struct gen_lun *lun;
+       unsigned int i;
+
+       gennvm_for_each_lun(gn, lun, i)
+               pr_info("%s: lun%8u\t%u\n",
+                                       dev->name, i, lun->vlun.nr_free_blocks);
+}
+
+static struct nvmm_type gennvm = {
+       .name           = "gennvm",
+       .version        = {0, 1, 0},
+
+       .register_mgr   = gennvm_register,
+       .unregister_mgr = gennvm_unregister,
+
+       .get_blk        = gennvm_get_blk,
+       .put_blk        = gennvm_put_blk,
+
+       .submit_io      = gennvm_submit_io,
+       .end_io         = gennvm_end_io,
+       .erase_blk      = gennvm_erase_blk,
+
+       .get_lun        = gennvm_get_lun,
+       .free_blocks_print = gennvm_free_blocks_print,
+};
+
+static int __init gennvm_module_init(void)
+{
+       return nvm_register_mgr(&gennvm);
+}
+
+static void gennvm_module_exit(void)
+{
+       nvm_unregister_mgr(&gennvm);
+}
+
+module_init(gennvm_module_init);
+module_exit(gennvm_module_exit);
+MODULE_LICENSE("GPL v2");
+MODULE_DESCRIPTION("Generic media manager for Open-Channel SSDs");
diff --git a/drivers/lightnvm/gennvm.h b/drivers/lightnvm/gennvm.h
new file mode 100644 (file)
index 0000000..d23bd35
--- /dev/null
@@ -0,0 +1,46 @@
+/*
+ * Copyright: Matias Bjorling <mb@bjorling.me>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version
+ * 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ */
+
+#ifndef GENNVM_H_
+#define GENNVM_H_
+
+#include <linux/module.h>
+#include <linux/vmalloc.h>
+
+#include <linux/lightnvm.h>
+
+struct gen_lun {
+       struct nvm_lun vlun;
+
+       int reserved_blocks;
+       /* lun block lists */
+       struct list_head used_list;     /* In-use blocks */
+       struct list_head free_list;     /* Not used blocks i.e. released
+                                        * and ready for use
+                                        */
+       struct list_head bb_list;       /* Bad blocks. Mutually exclusive with
+                                        * free_list and used_list
+                                        */
+};
+
+struct gen_nvm {
+       int nr_luns;
+       struct gen_lun *luns;
+};
+
+#define gennvm_for_each_lun(bm, lun, i) \
+               for ((i) = 0, lun = &(bm)->luns[0]; \
+                       (i) < (bm)->nr_luns; (i)++, lun = &(bm)->luns[(i)])
+
+#endif /* GENNVM_H_ */