Date:   Fri, 29 Jun 2018 13:28:40 +0200
From:   Matias Bjørling <mb@...htnvm.io>
To:     javier@...xlabs.com
Cc:     hans.holmberg@...xlabs.com, linux-block@...r.kernel.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH] lightnvm: pblk: recover chunk state on 1.2 devices

On 06/29/2018 01:22 PM, Javier Gonzalez wrote:
>> On 29 Jun 2018, at 13.14, Matias Bjørling <mb@...htnvm.io> wrote:
>>
>> On 06/28/2018 11:12 AM, Javier González wrote:
>>> The Open-Channel 1.2 spec does not define a mechanism for the host to
>>> recover the block (chunk) state. As a consequence, a newly format device
>>> will need to reconstruct the state. Currently, pblk assumes that blocks
>>> are not erased, which might cause double-erases in case that the device
>>> does not protect itself against them (which is not specified in the spec
>>> either).
>>
>> It should not be specified in the spec. It is up to the device to handle
>> double erases and not do it.
>>
>>> This patch, reconstructs the state based on read errors. If the first
>>> sector of a block returns and empty page (NVM_RSP_ERR_EMPTYPAGE), then
>>> the block s marked free, i.e., erased and ready to be used
>>> (NVM_CHK_ST_FREE). Otherwise, the block is marked as closed
>>> (NVM_CHK_ST_CLOSED). Note that even if a block is open and not fully
>>> written, it has to be erased in order to be used again.
>>
>> Should we extend it to do the scan, and update the write pointer as
>> well? I think this kind of feature already is baked into pblk?
>>
> 
> This is already in place: we scan until empty page and take it from
> there. This patch is only for the case in which we start a pblk instance
> form scratch. On a device already owned by pblk, we would not have the
> problem we are trying to solve here because we know the state.

Agree. What I meant was that when we anyway are recovering the state, we 
could just as well update ->wp and set to NVM_CHK_ST_OPEN and so forth 
for the initialization phase.

> 
>>> One caveat of this approach is that blocks that have been erased at a
>>> moment in time, will always be considered as erased. However, some media
>>> might become unstable if blocks are not erased before usage. Since pblk
>>> would follow this principle after the state of all blocks fall under
>>> pblk's domain, we can consider this as an initialization problem. The
>>> trade-off would be to fall back to the old behavior and risk premature
>>> media wearing.
>>
>> The above is up to the device implementation to handle. We cannot
>> expect users to understand the intrinsics of media.
>>
> 
> Of course. The point is that with this approach, erases are left a bit
> in the air and "preventable" write errors might happen, with the previous
> the burden was put on the device to deal with double erases. It's a
> tradeoff that I want to make clear before the path is taken.

Cool.

> 
>>> Signed-off-by: Javier González <javier@...xlabs.com>
>>> ---
>>>   drivers/lightnvm/pblk-init.c | 138 ++++++++++++++++++++++++++++++++++++++-----
>>>   1 file changed, 124 insertions(+), 14 deletions(-)
>>> diff --git a/drivers/lightnvm/pblk-init.c b/drivers/lightnvm/pblk-init.c
>>> index 3b8aa4a64cac..ce25f1473d8e 100644
>>> --- a/drivers/lightnvm/pblk-init.c
>>> +++ b/drivers/lightnvm/pblk-init.c
>>> @@ -697,47 +697,138 @@ static void pblk_set_provision(struct pblk *pblk, long nr_free_blks)
>>>   	atomic_set(&pblk->rl.free_user_blocks, nr_free_blks);
>>>   }
>>>   +static void pblk_state_complete(struct kref *ref)
>>> +{
>>> +	struct pblk_pad_rq *pad_rq = container_of(ref, struct pblk_pad_rq, ref);
>>> +
>>> +	complete(&pad_rq->wait);
>>> +}
>>> +
>>> +static void pblk_end_io_state(struct nvm_rq *rqd)
>>> +{
>>> +	struct pblk_pad_rq *pad_rq = rqd->private;
>>> +	struct pblk *pblk = pad_rq->pblk;
>>> +	struct nvm_tgt_dev *dev = pblk->dev;
>>> +	struct nvm_geo *geo = &dev->geo;
>>> +	struct pblk_line *line;
>>> +	struct nvm_chk_meta *chunk;
>>> +	int pos;
>>> +
>>> +	line = &pblk->lines[pblk_ppa_to_line(rqd->ppa_addr)];
>>> +	pos = pblk_ppa_to_pos(geo, rqd->ppa_addr);
>>> +
>>> +	chunk = &line->chks[pos];
>>> +
>>> +	if (rqd->error == NVM_RSP_ERR_EMPTYPAGE)
>>> +		chunk->state = NVM_CHK_ST_FREE;
>>> +	else
>>> +		chunk->state = NVM_CHK_ST_CLOSED;
>>> +
>>> +	bio_put(rqd->bio);
>>> +	pblk_free_rqd(pblk, rqd, PBLK_READ);
>>> +	kref_put(&pad_rq->ref, pblk_state_complete);
>>> +}
>>> +
>>> +static int pblk_check_chunk_state(struct pblk *pblk, struct nvm_chk_meta *chunk,
>>> +				struct ppa_addr ppa, struct pblk_pad_rq *pad_rq)
>>> +{
>>> +	struct nvm_rq *rqd;
>>> +	struct bio *bio;
>>> +	int ret;
>>> +
>>> +	bio = bio_alloc(GFP_KERNEL, 1);
>>> +
>>> +	if (pblk_bio_add_pages(pblk, bio, GFP_KERNEL, 1))
>>> +		goto fail_free_bio;
>>> +
>>> +	rqd = pblk_alloc_rqd(pblk, PBLK_READ);
>>> +
>>> +	rqd->bio = bio;
>>> +	rqd->opcode = NVM_OP_PREAD;
>>> +	rqd->flags = pblk_set_read_mode(pblk, PBLK_READ_SEQUENTIAL);
>>> +	rqd->nr_ppas = 1;
>>> +	rqd->ppa_addr = ppa;
>>> +	rqd->end_io = pblk_end_io_state;
>>> +	rqd->private = pad_rq;
>>> +
>>> +	kref_get(&pad_rq->ref);
>>> +
>>> +	ret = pblk_submit_io(pblk, rqd);
>>> +	if (ret) {
>>> +		pr_err("pblk: I/O submissin failed: %d\n", ret);
>>> +		goto fail_free_rqd;
>>> +	}
>>> +
>>> +	return NVM_IO_OK;
>>> +
>>> +fail_free_rqd:
>>> +	pblk_free_rqd(pblk, rqd, PBLK_READ);
>>> +	pblk_bio_free_pages(pblk, bio, 0, bio->bi_vcnt);
>>> +fail_free_bio:
>>> +	bio_put(bio);
>>> +
>>> +	return NVM_IO_ERR;
>>> +}
>>> +
>>>   static int pblk_setup_line_meta_12(struct pblk *pblk, struct pblk_line *line,
>>>   				   void *chunk_meta)
>>>   {
>>>   	struct nvm_tgt_dev *dev = pblk->dev;
>>>   	struct nvm_geo *geo = &dev->geo;
>>>   	struct pblk_line_meta *lm = &pblk->lm;
>>> +	struct pblk_pad_rq *pad_rq;
>>>   	int i, chk_per_lun, nr_bad_chks = 0;
>>>   +	pad_rq = kmalloc(sizeof(struct pblk_pad_rq), GFP_KERNEL);
>>> +	if (!pad_rq)
>>> +		return -1;
>>> +
>>> +	pad_rq->pblk = pblk;
>>> +	init_completion(&pad_rq->wait);
>>> +	kref_init(&pad_rq->ref);
>>> +
>>>   	chk_per_lun = geo->num_chk * geo->pln_mode;
>>>     	for (i = 0; i < lm->blk_per_line; i++) {
>>>   		struct pblk_lun *rlun = &pblk->luns[i];
>>>   		struct nvm_chk_meta *chunk;
>>> -		int pos = pblk_ppa_to_pos(geo, rlun->bppa);
>>> +		struct ppa_addr ppa = rlun->bppa;
>>> +		int pos = pblk_ppa_to_pos(geo, ppa);
>>>   		u8 *lun_bb_meta = chunk_meta + pos * chk_per_lun;
>>>     		chunk = &line->chks[pos];
>>>   -		/*
>>> -		 * In 1.2 spec. chunk state is not persisted by the device. Thus
>>> -		 * some of the values are reset each time pblk is instantiated,
>>> -		 * so we have to assume that the block is closed.
>>> -		 */
>>> -		if (lun_bb_meta[line->id] == NVM_BLK_T_FREE)
>>> -			chunk->state =  NVM_CHK_ST_CLOSED;
>>> -		else
>>> -			chunk->state = NVM_CHK_ST_OFFLINE;
>>> -
>>>   		chunk->type = NVM_CHK_TP_W_SEQ;
>>>   		chunk->wi = 0;
>>>   		chunk->slba = -1;
>>>   		chunk->cnlb = geo->clba;
>>>   		chunk->wp = 0;
>>>   -		if (!(chunk->state & NVM_CHK_ST_OFFLINE))
>>> +		if (lun_bb_meta[line->id] != NVM_BLK_T_FREE) {
>>> +			chunk->state = NVM_CHK_ST_OFFLINE;
>>> +			set_bit(pos, line->blk_bitmap);
>>> +			nr_bad_chks++;
>>> +
>>>   			continue;
>>> +		}
>>>   -		set_bit(pos, line->blk_bitmap);
>>> -		nr_bad_chks++;
>>> +		/*
>>> +		 * In 1.2 spec. chunk state is not persisted by the device.
>>> +		 * Recover the state based on media response.
>>> +		 */
>>> +		ppa.g.blk = line->id;
>>> +		pblk_check_chunk_state(pblk, chunk, ppa, pad_rq);
>>>   	}
>>>   +	kref_put(&pad_rq->ref, pblk_state_complete);
>>> +
>>> +	if (!wait_for_completion_io_timeout(&pad_rq->wait,
>>> +				msecs_to_jiffies(PBLK_COMMAND_TIMEOUT_MS))) {
>>> +		pr_err("pblk: state recovery timed out\n");
>>> +		return -1;
>>> +	}
>>> +
>>> +	kfree(pad_rq);
>>>   	return nr_bad_chks;
>>>   }
>>>   @@ -1036,6 +1127,23 @@ static int pblk_line_meta_init(struct pblk *pblk)
>>>   	return 0;
>>>   }
>>>   +static void check_meta(struct pblk *pblk, struct pblk_line *line)
>>> +{
>>> +	struct nvm_tgt_dev *dev = pblk->dev;
>>> +	struct nvm_geo *geo = &dev->geo;
>>> +	struct pblk_line_meta *lm = &pblk->lm;
>>> +	int i;
>>> +
>>> +	for (i = 0; i < lm->blk_per_line; i++) {
>>> +		struct pblk_lun *rlun = &pblk->luns[i];
>>> +		struct nvm_chk_meta *chunk;
>>> +		struct ppa_addr ppa = rlun->bppa;
>>> +		int pos = pblk_ppa_to_pos(geo, ppa);
>>> +
>>> +		chunk = &line->chks[pos];
>>> +	}
>>> +}
>>> +
>>>   static int pblk_lines_init(struct pblk *pblk)
>>>   {
>>>   	struct pblk_line_mgmt *l_mg = &pblk->l_mg;
>>> @@ -1077,6 +1185,8 @@ static int pblk_lines_init(struct pblk *pblk)
>>>   			goto fail_free_lines;
>>>     		nr_free_chks += pblk_setup_line_meta(pblk, line, chunk_meta, i);
>>> +
>>> +		check_meta(pblk, line);
>>>   	}
>>>     	if (!nr_free_chks) {
>>
>> I'm okay with us doing this in pblk for now. Over time, someone may do
>> the work move this (and other specific only-1.2/2.0 stuff) into the
>> lightnvm subsystem. I don't think pblk should need to care about
>> either 1.2 or 2.0.
> 
> That would be ideal.
> 
> Thanks!
> 

Powered by blists - more mailing lists