On 21/03/23 13:03, Ulf Hansson wrote: > On Tue, 21 Mar 2023 at 11:36, Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote: >> >> On 16/03/23 18:45, Ulf Hansson wrote: >>> REQ_FUA translates into so called "reliable writes" (atomic writes) for >>> eMMC cards, which is generally supported as it was introduced as a >>> mandatory feature already in the v4.3 (2007) of the eMMC spec. To fully >>> support the reliable writes (thus REQ_FUA), the mmc host driver needs to >>> support the CMD23 (MMC_CAP_CMD23) too, which is rather common nowadays. >>> >>> File systems typically uses REQ_FUA for writing their meta-data and other >>> important information. Ideally it should provide an increased protection >>> against data-corruption, during sudden power-failures. This said, file >>> systems have other ways to handle sudden power-failures too, like using >>> checksums to detect partly-written data, for example. >>> >>> It has been reported that the reliable writes are costly for some eMMCs, >>> leading to performance degradations. Exactly why, is in the implementation >>> details of the internals of the eMMC. >>> >>> Moreover, in the v4.5 (2011) of the eMMC spec, the cache-control was >>> introduced as an optional feature. It allows the host to trigger a flush of >>> the eMMC's internal write-cache. In the past, before the cache-control >>> feature was added, the reliable write acted as trigger for the eMMC, to >>> also flush its internal write-cache, even if that too remains as an >>> implementation detail of the eMMC. >>> >>> In a way to try to improve the situation with costly reliable writes and >>> REQ_FUA, let's add a new card quirk MMC_QUIRK_AVOID_REL_WRITE, which may be >>> set to avoid announcing the support for it. However, as mentioned above, >>> due to the specific relationship with the cache-control feature, we must >>> keep REQ_FUA unless that is supported too. >>> >>> Reported-by: Wenchao Chen <wenchao.chen666@xxxxxxxxx> >>> Acked-by: Bean Huo <beanhuo@xxxxxxxxxx> >>> Acked-by: Avri Altman <avri.altman@xxxxxxx> >>> Signed-off-by: Ulf Hansson <ulf.hansson@xxxxxxxxxx> >> >> Minor cosmetic suggestion below, but nevertheless: >> >> Acked-by: Adrian Hunter <adrian.hunter@xxxxxxxxx> > > Thanks! > >> >>> --- >>> >>> Updated since the RFC: >>> Added a card quirk to maintain the current behaviour. The quirk isn't >>> set for any cards yet, which is needed (a patch on top) to move forward >>> with this. >>> >>> --- >>> drivers/mmc/core/block.c | 16 ++++++++++++---- >>> drivers/mmc/core/card.h | 5 +++++ >>> include/linux/mmc/card.h | 1 + >>> 3 files changed, 18 insertions(+), 4 deletions(-) >>> >>> diff --git a/drivers/mmc/core/block.c b/drivers/mmc/core/block.c >>> index 672ab90c4b2d..35292e36a1fb 100644 >>> --- a/drivers/mmc/core/block.c >>> +++ b/drivers/mmc/core/block.c >>> @@ -2409,8 +2409,7 @@ static struct mmc_blk_data *mmc_blk_alloc_req(struct mmc_card *card, >>> struct mmc_blk_data *md; >>> int devidx, ret; >>> char cap_str[10]; >>> - bool cache_enabled = false; >>> - bool fua_enabled = false; >>> + bool cache_enabled, avoid_fua, fua_enabled = false; >>> >>> devidx = ida_simple_get(&mmc_blk_ida, 0, max_devices, GFP_KERNEL); >>> if (devidx < 0) { >>> @@ -2494,11 +2493,20 @@ static struct mmc_blk_data *mmc_blk_alloc_req(struct mmc_card *card, >>> ((card->ext_csd.rel_param & EXT_CSD_WR_REL_PARAM_EN) || >>> card->ext_csd.rel_sectors)) { >>> md->flags |= MMC_BLK_REL_WR; >>> + } >>> + >>> + /* >>> + * REQ_FUA is supported through eMMC reliable writes, which has been >>> + * reported to be a bit costly for some eMMCs. In these cases, let's >>> + * rely on the flush requests (REQ_OP_FLUSH) instead, if we can use the >>> + * cache-control feature too. >>> + */ >>> + cache_enabled = mmc_cache_enabled(card->host); >>> + avoid_fua = cache_enabled && mmc_card_avoid_rel_write(card); >>> + if (md->flags & MMC_BLK_REL_WR && !avoid_fua) { >>> fua_enabled = true; >>> cache_enabled = true; >>> } >> >> looks like this could be just: >> >> fua_enabled = (md->flags & MMC_BLK_REL_WR) && !avoid_fua; >> >> with fua_enabled no longer needing initialization > > Unless I misunderstand your point, that would work for fua_enabled, > but would not be sufficient for cache_enabled. > > cache_enabled should be set if fua_enabled is set - and no matter > whether mmc_cache_enabled() returns true or not. > > Did that make sense? Yes, you are right, sorry! > > [...] > > Kind regards > Uffe