| Message ID | 20210420134641.57343-1-avri.altman@wdc.com |
|---|---|
| Headers | show |
| Series | Do not flush cache when it is disabled | expand |
Ulf - Would you consider picking this sires up now? I was hoping to close this bug. Thanks, Avri > -----Original Message----- > From: Avri Altman <avri.altman@wdc.com> > Sent: Tuesday, 20 April 2021 16:47 > To: Ulf Hansson <ulf.hansson@linaro.org>; linux-mmc@vger.kernel.org; Adrian > Hunter <adrian.hunter@intel.com> > Cc: linux-kernel@vger.kernel.org; Avri Altman <Avri.Altman@wdc.com> > Subject: [PATCH v4 0/2] Do not flush cache when it is disabled > > v3 -> v4 > - Attend Adrian's comments > - Add Acked-by tag > > v2 -> v3: > - rebase onto recent cache changes > > v1 -> v2: > - Attend Adrian's comments > > Cache is a temporary storage space in an eMMC device. Volatile by > nature, the cache should in typical case reduce the access time compared > to an access to the main nonvolatile storage. > > The cache function can be turned ON and OFF. Once OFF, the host is not > expected to issue a flush-cache command to the device. > > Avri Altman (2): > mmc: block: Issue flush only if allowed > mmc: block: Update ext_csd.cache_ctrl if it was written > > drivers/mmc/core/block.c | 21 +++++++++++++++++++++ > drivers/mmc/core/mmc.c | 2 +- > drivers/mmc/core/mmc_ops.h | 5 +++++ > 3 files changed, 27 insertions(+), 1 deletion(-) > > -- > 2.25.1
On Tue, 20 Apr 2021 at 15:46, Avri Altman <avri.altman@wdc.com> wrote: > > The cache may be flushed to the nonvolatile storage by writing to > FLUSH_CACHE byte (EXT_CSD byte [32]). When in command queueing mode, the > cache may be flushed by issuing a CMDQ_TASK_ DEV_MGMT (CMD48) with a > FLUSH_CACHE op-code. Either way, verify that The cache function is > turned ON before doing so. > > fixes: 1e8e55b67030 (mmc: block: Add CQE support) > > Reported-by: Brendan Peter <bpeter@lytx.com> > Tested-by: Brendan Peter <bpeter@lytx.com> > Signed-off-by: Avri Altman <avri.altman@wdc.com> > --- > drivers/mmc/core/block.c | 9 +++++++++ > drivers/mmc/core/mmc.c | 2 +- > drivers/mmc/core/mmc_ops.h | 5 +++++ > 3 files changed, 15 insertions(+), 1 deletion(-) > > diff --git a/drivers/mmc/core/block.c b/drivers/mmc/core/block.c > index 8bfd4d95b386..24e1ecbdd510 100644 > --- a/drivers/mmc/core/block.c > +++ b/drivers/mmc/core/block.c > @@ -2186,6 +2186,11 @@ static int mmc_blk_wait_for_idle(struct mmc_queue *mq, struct mmc_host *host) > return mmc_blk_rw_wait(mq, NULL); > } > > +static bool mmc_blk_cache_disabled(struct mmc_card *card) > +{ > + return mmc_card_mmc(card) && !mmc_flush_allowed(card); It's these kinds of use with mmc_card_mmc() that I think we need to strive towards avoiding when going forward. For example, newer SD cards support both cache and command queueing nowadays, which means that we need to make the code in the mmc block layer more agnostic. To do that, I think we should use the bus_ops callbacks. That's why I started out by adding the ->flush_cache() callback in the other patch. > +} > + > enum mmc_issued mmc_blk_mq_issue_rq(struct mmc_queue *mq, struct request *req) > { > struct mmc_blk_data *md = mq->blkdata; > @@ -2225,6 +2230,10 @@ enum mmc_issued mmc_blk_mq_issue_rq(struct mmc_queue *mq, struct request *req) > case MMC_ISSUE_ASYNC: > switch (req_op(req)) { > case REQ_OP_FLUSH: > + if (mmc_blk_cache_disabled(mq->card)) { I suggest that you add a new bus ops callback ->cache_enabled() and implement it for the mmc bus type. From the mmc block layer point of view, it would then mean that if the callback isn't implemented, the cache ctrl isn't supported (which would be the case for SD then) > + blk_mq_end_request(req, BLK_STS_OK); > + return MMC_REQ_FINISHED; > + } > ret = mmc_blk_cqe_issue_flush(mq, req); > break; > case REQ_OP_READ: > diff --git a/drivers/mmc/core/mmc.c b/drivers/mmc/core/mmc.c > index 9ad4aa537867..e3da62ffcb5e 100644 > --- a/drivers/mmc/core/mmc.c > +++ b/drivers/mmc/core/mmc.c > @@ -2037,7 +2037,7 @@ static int _mmc_flush_cache(struct mmc_card *card) > { > int err = 0; > > - if (card->ext_csd.cache_size > 0 && card->ext_csd.cache_ctrl & 1) { > + if (mmc_flush_allowed(card)) { > err = mmc_switch(card, EXT_CSD_CMD_SET_NORMAL, > EXT_CSD_FLUSH_CACHE, 1, > CACHE_FLUSH_TIMEOUT_MS); > diff --git a/drivers/mmc/core/mmc_ops.h b/drivers/mmc/core/mmc_ops.h > index 5782fdf4e8e9..2682bf66708a 100644 > --- a/drivers/mmc/core/mmc_ops.h > +++ b/drivers/mmc/core/mmc_ops.h > @@ -19,6 +19,11 @@ enum mmc_busy_cmd { > struct mmc_host; > struct mmc_card; > > +static inline bool mmc_flush_allowed(struct mmc_card *card) > +{ > + return card->ext_csd.cache_size > 0 && card->ext_csd.cache_ctrl & 1; > +} > + > int mmc_select_card(struct mmc_card *card); > int mmc_deselect_cards(struct mmc_host *host); > int mmc_set_dsr(struct mmc_host *host); > -- > 2.25.1 > Having said the above, we clearly want to tag $subject patch for stable kernels as well, which means we need a simple patch as possible. Clearly $subject patch should have come first, prior to my patch where I added the ->flush_cache() bus ops callback, as it messes things up. Therefore, I decided to rebase my next branch to drop the patch adding the ->flush_cache() bus ops (I will re-post the patch after we have got your changes in). Can you please re-base $subject patch and address my comments? My apologies for the mess! Kind regards Uffe
> On Tue, 20 Apr 2021 at 15:46, Avri Altman <avri.altman@wdc.com> wrote: > > > > The cache may be flushed to the nonvolatile storage by writing to > > FLUSH_CACHE byte (EXT_CSD byte [32]). When in command queueing mode, > the > > cache may be flushed by issuing a CMDQ_TASK_ DEV_MGMT (CMD48) with a > > FLUSH_CACHE op-code. Either way, verify that The cache function is > > turned ON before doing so. > > > > fixes: 1e8e55b67030 (mmc: block: Add CQE support) > > > > Reported-by: Brendan Peter <bpeter@lytx.com> > > Tested-by: Brendan Peter <bpeter@lytx.com> > > Signed-off-by: Avri Altman <avri.altman@wdc.com> > > --- > > drivers/mmc/core/block.c | 9 +++++++++ > > drivers/mmc/core/mmc.c | 2 +- > > drivers/mmc/core/mmc_ops.h | 5 +++++ > > 3 files changed, 15 insertions(+), 1 deletion(-) > > > > diff --git a/drivers/mmc/core/block.c b/drivers/mmc/core/block.c > > index 8bfd4d95b386..24e1ecbdd510 100644 > > --- a/drivers/mmc/core/block.c > > +++ b/drivers/mmc/core/block.c > > @@ -2186,6 +2186,11 @@ static int mmc_blk_wait_for_idle(struct > mmc_queue *mq, struct mmc_host *host) > > return mmc_blk_rw_wait(mq, NULL); > > } > > > > +static bool mmc_blk_cache_disabled(struct mmc_card *card) > > +{ > > + return mmc_card_mmc(card) && !mmc_flush_allowed(card); > > It's these kinds of use with mmc_card_mmc() that I think we need to > strive towards avoiding when going forward. > > For example, newer SD cards support both cache and command queueing > nowadays, which means that we need to make the code in the mmc block > layer more agnostic. To do that, I think we should use the bus_ops > callbacks. That's why I started out by adding the ->flush_cache() > callback in the other patch. Understood. > > > +} > > + > > enum mmc_issued mmc_blk_mq_issue_rq(struct mmc_queue *mq, struct > request *req) > > { > > struct mmc_blk_data *md = mq->blkdata; > > @@ -2225,6 +2230,10 @@ enum mmc_issued mmc_blk_mq_issue_rq(struct > mmc_queue *mq, struct request *req) > > case MMC_ISSUE_ASYNC: > > switch (req_op(req)) { > > case REQ_OP_FLUSH: > > + if (mmc_blk_cache_disabled(mq->card)) { > > I suggest that you add a new bus ops callback ->cache_enabled() and > implement it for the mmc bus type. > > From the mmc block layer point of view, it would then mean that if the > callback isn't implemented, the cache ctrl isn't supported (which > would be the case for SD then) Done. > > > + blk_mq_end_request(req, BLK_STS_OK); > > + return MMC_REQ_FINISHED; > > + } > > ret = mmc_blk_cqe_issue_flush(mq, req); > > > break; > > case REQ_OP_READ: > > diff --git a/drivers/mmc/core/mmc.c b/drivers/mmc/core/mmc.c > > index 9ad4aa537867..e3da62ffcb5e 100644 > > --- a/drivers/mmc/core/mmc.c > > +++ b/drivers/mmc/core/mmc.c > > @@ -2037,7 +2037,7 @@ static int _mmc_flush_cache(struct mmc_card > *card) > > { > > int err = 0; > > > > - if (card->ext_csd.cache_size > 0 && card->ext_csd.cache_ctrl & 1) { > > + if (mmc_flush_allowed(card)) { > > err = mmc_switch(card, EXT_CSD_CMD_SET_NORMAL, > > EXT_CSD_FLUSH_CACHE, 1, > > CACHE_FLUSH_TIMEOUT_MS); > > diff --git a/drivers/mmc/core/mmc_ops.h b/drivers/mmc/core/mmc_ops.h > > index 5782fdf4e8e9..2682bf66708a 100644 > > --- a/drivers/mmc/core/mmc_ops.h > > +++ b/drivers/mmc/core/mmc_ops.h > > @@ -19,6 +19,11 @@ enum mmc_busy_cmd { > > struct mmc_host; > > struct mmc_card; > > > > +static inline bool mmc_flush_allowed(struct mmc_card *card) > > +{ > > + return card->ext_csd.cache_size > 0 && card->ext_csd.cache_ctrl & 1; > > +} > > + > > int mmc_select_card(struct mmc_card *card); > > int mmc_deselect_cards(struct mmc_host *host); > > int mmc_set_dsr(struct mmc_host *host); > > -- > > 2.25.1 > > > > Having said the above, we clearly want to tag $subject patch for > stable kernels as well, which means we need a simple patch as > possible. > > Clearly $subject patch should have come first, prior to my patch where > I added the ->flush_cache() bus ops callback, as it messes things up. > > Therefore, I decided to rebase my next branch to drop the patch adding > the ->flush_cache() bus ops (I will re-post the patch after we have > got your changes in). > > Can you please re-base $subject patch and address my comments? My > apologies for the mess! Done. Thanks, Avri > > Kind regards > Uffe