On Wed, 16 Oct 2024 at 17:21, Avri Altman <Avri.Altman@xxxxxxx> wrote: > > > On Tue, 15 Oct 2024 at 11:44, Adrian Hunter <adrian.hunter@xxxxxxxxx> > > wrote: > > > > > > On 14/10/24 14:44, Avri Altman wrote: > > > > While reviewing the SDUC series, Adrian made a comment concerning > > > > the memory allocation code in mmc_sd_num_wr_blocks() - see [1]. > > > > Prevent memory allocations from triggering I/O operations while > > > > ACMD22 is in progress. > > > > > > > > [1] https://www.spinics.net/lists/linux-mmc/msg82199.html > > > > > > > > Suggested-by: Adrian Hunter <adrian.hunter@xxxxxxxxx> > > > > Signed-off-by: Avri Altman <avri.altman@xxxxxxx> > > > > Cc: stable@xxxxxxxxxxxxxxx > > > > --- > > > > drivers/mmc/core/block.c | 10 +++++++++- > > > > 1 file changed, 9 insertions(+), 1 deletion(-) > > > > > > > > diff --git a/drivers/mmc/core/block.c b/drivers/mmc/core/block.c > > > > index 04f3165cf9ae..042b0147d47e 100644 > > > > --- a/drivers/mmc/core/block.c > > > > +++ b/drivers/mmc/core/block.c > > > > @@ -995,6 +995,8 @@ static int mmc_sd_num_wr_blocks(struct > > mmc_card *card, u32 *written_blocks) > > > > u32 result; > > > > __be32 *blocks; > > > > u8 resp_sz = mmc_card_ult_capacity(card) ? 8 : 4; > > > > + unsigned int noio_flag; > > > > + > > > > struct mmc_request mrq = {}; > > > > struct mmc_command cmd = {}; > > > > struct mmc_data data = {}; > > > > @@ -1018,9 +1020,13 @@ static int mmc_sd_num_wr_blocks(struct > > mmc_card *card, u32 *written_blocks) > > > > mrq.cmd = &cmd; > > > > mrq.data = &data; > > > > > > > > + noio_flag = memalloc_noio_save(); > > > > + > > > > blocks = kmalloc(resp_sz, GFP_KERNEL); > > > > > > Could have memalloc_noio_restore() here: > > > > > > memalloc_noio_restore(noio_flag); > > > > > > but I feel maybe adding something like: > > > > > > u64 __aligned(8) tiny_io_buf; > > > > > > to either struct mmc_card or struct mmc_host is better? > > > Ulf, any thoughts? > > > > > > > I have no strong opinion. > Then I would vote to stay with Adrian's original NOIO suggestion, because: > 1) My testing shows that mmc_sd_num_wr_blocks() is hardly being hit, and > 2) that allocation is within the write timeout anyway > > So unless you want it otherwise, will remove the redundant macro call and re-spin. Sounds good to me! [...] Kind regards Uffe
