On Fri, 26 Apr 2024 at 17:11, Kamal Dasu <kamal.dasu@xxxxxxxxxxxx> wrote: > > On Fri, Apr 26, 2024 at 2:17 AM Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote: > > > > On 25/04/24 19:18, Ulf Hansson wrote: > > > + Wolfram, Adrian (to see if they have some input) > > > > > > On Tue, 23 Apr 2024 at 22:02, Kamal Dasu <kamal.dasu@xxxxxxxxxxxx> wrote: > > >> > > >> When erase/trim/discard completion was converted to mmc_poll_for_busy(), > > >> optional ->card_busy() host ops support was added. sdhci card->busy() > > >> could return busy for long periods to cause mmc_do_erase() to block during > > >> discard operation as shown below during mkfs.f2fs : > > >> > > >> Info: [/dev/mmcblk1p9] Discarding device > > >> [ 39.597258] sysrq: Show Blocked State > > >> [ 39.601183] task:mkfs.f2fs state:D stack:0 pid:1561 tgid:1561 ppid:1542 flags:0x0000000d > > >> [ 39.610609] Call trace: > > >> [ 39.613098] __switch_to+0xd8/0xf4 > > >> [ 39.616582] __schedule+0x440/0x4f4 > > >> [ 39.620137] schedule+0x2c/0x48 > > >> [ 39.623341] schedule_hrtimeout_range_clock+0xe0/0x114 > > >> [ 39.628562] schedule_hrtimeout_range+0x10/0x18 > > >> [ 39.633169] usleep_range_state+0x5c/0x90 > > >> [ 39.637253] __mmc_poll_for_busy+0xec/0x128 > > >> [ 39.641514] mmc_poll_for_busy+0x48/0x70 > > >> [ 39.645511] mmc_do_erase+0x1ec/0x210 > > >> [ 39.649237] mmc_erase+0x1b4/0x1d4 > > >> [ 39.652701] mmc_blk_mq_issue_rq+0x35c/0x6ac > > >> [ 39.657037] mmc_mq_queue_rq+0x18c/0x214 > > >> [ 39.661022] blk_mq_dispatch_rq_list+0x3a8/0x528 > > >> [ 39.665722] __blk_mq_sched_dispatch_requests+0x3a0/0x4ac > > >> [ 39.671198] blk_mq_sched_dispatch_requests+0x28/0x5c > > >> [ 39.676322] blk_mq_run_hw_queue+0x11c/0x12c > > >> [ 39.680668] blk_mq_flush_plug_list+0x200/0x33c > > >> [ 39.685278] blk_add_rq_to_plug+0x68/0xd8 > > >> [ 39.689365] blk_mq_submit_bio+0x3a4/0x458 > > >> [ 39.693539] __submit_bio+0x1c/0x80 > > >> [ 39.697096] submit_bio_noacct_nocheck+0x94/0x174 > > >> [ 39.701875] submit_bio_noacct+0x1b0/0x22c > > >> [ 39.706042] submit_bio+0xac/0xe8 > > >> [ 39.709424] blk_next_bio+0x4c/0x5c > > >> [ 39.712973] blkdev_issue_secure_erase+0x118/0x170 > > >> [ 39.717835] blkdev_common_ioctl+0x374/0x728 > > >> [ 39.722175] blkdev_ioctl+0x8c/0x2b0 > > >> [ 39.725816] vfs_ioctl+0x24/0x40 > > >> [ 39.729117] __arm64_sys_ioctl+0x5c/0x8c > > >> [ 39.733114] invoke_syscall+0x68/0xec > > >> [ 39.736839] el0_svc_common.constprop.0+0x70/0xd8 > > >> [ 39.741609] do_el0_svc+0x18/0x20 > > >> [ 39.744981] el0_svc+0x68/0x94 > > >> [ 39.748107] el0t_64_sync_handler+0x88/0x124 > > >> [ 39.752455] el0t_64_sync+0x168/0x16c > > > > > > Thanks for the detailed log! > > > > > >> > > >> Fix skips the card->busy() and uses MMC_SEND_STATUS and R1_STATUS > > >> check for MMC_ERASE_BUSY busy_cmd case in the mmc_busy_cb() function. > > >> > > >> Fixes: 0d84c3e6a5b2 ("mmc: core: Convert to mmc_poll_for_busy() for erase/trim/discard") > > >> Signed-off-by: Kamal Dasu <kamal.dasu@xxxxxxxxxxxx> > > >> --- > > >> drivers/mmc/core/mmc_ops.c | 3 ++- > > >> 1 file changed, 2 insertions(+), 1 deletion(-) > > >> > > >> diff --git a/drivers/mmc/core/mmc_ops.c b/drivers/mmc/core/mmc_ops.c > > >> index 3b3adbddf664..603fbd78c342 100644 > > >> --- a/drivers/mmc/core/mmc_ops.c > > >> +++ b/drivers/mmc/core/mmc_ops.c > > >> @@ -464,7 +464,8 @@ static int mmc_busy_cb(void *cb_data, bool *busy) > > >> u32 status = 0; > > >> int err; > > >> > > >> - if (data->busy_cmd != MMC_BUSY_IO && host->ops->card_busy) { > > >> + if (data->busy_cmd != MMC_BUSY_IO && > > >> + data->busy_cmd != MMC_BUSY_ERASE && host->ops->card_busy) { > > >> *busy = host->ops->card_busy(host); > > >> return 0; > > >> } > > > > > > So it seems like the ->card_busy() callback is broken in for your mmc > > > host-driver and platform. Can you perhaps provide the information > > > about what HW/driver you are using? > > > > > Using the sdhci-brcmstb driver on a Broadcom Settop based SoC. > > > > The point with using the ->card_busy() callback, is to avoid sending > > > the CMD13. Ideally it should be cheaper/faster and in most cases it > > > translates to a read of a register. For larger erases, we would > > > probably end up sending the CMD13 periodically every 32-64 ms, which > > > shouldn't be a problem. However, for smaller erases and discards, we > > > may want the benefit the ->card_busy() callback provides us. > > > > > I have tested two scenarios. One is like the mkfs.f2fs app that calls : > ioctl(fd, BLKSECDISCARD, &range); > > This has the following CMD and completion sequence: > {CMD35->CMD36->CMD38} and poll for DAT0 signal via card->busy . > CMD38 has a response of R1b. The DAT0 (Busy line) will be driven by the device. > Busy (DAT0 = 0) is asserted by a device for erasing blocks. Stays > busy in brcmstb sdhci controller. How big is the "range"? Just so I get this right, it stays busy and we are waiting for the timeout to fire? And ideally you think we should not be busy for that long, right? > > With the additional change followed by CMD13 (response of R1), which > returns the device status, the > DAT0 will be pulled-up and next time we read the BUSY status it will > indicate it is not busy. So you are referring to read the BUSY status with you ->card_busy() callback? Or did you actually verify that this is true from an electrical point of view, by monitoring the DAT0 signal? If the latter, perhaps it's the card that is failing and simply requires CMD13 to be used to poll for busy. What card is this? Have you tried different cards with the same platform/driver? > > Also I tried the mmc util and that does not show the same issue with > exactly the same ranges, however in that case there are some > differences in the way the CMD sequence is sent for the entire discard > operation. > # mmc erase discard 0x000087a4 0x002effff /dev/mmcblk1 > /* send erase cmd with multi-cmd */ > ret = ioctl(dev_fd, MMC_IOC_MULTI_CMD, multi_cmd); > > CMD35->CMD13->CMD36->CMD13->CMD38->CMD13 > I do not see any hang in all the erase options discard, legacy, trim, trim2, > secure-trim used here with mmc util . So CMD13 seems to do the trick for you. Although, I think we need to figure out if this a special "broken" card or if the problem is with the ->card_busy() implementation for your platform. > > Also looking at JEDEC Standard No. 84-B51 Page 276, 277 > "Once the erase groups are selected the host will send an ERASE > (CMD38) command. It is recom- > mended that the host terminates the sequence with a SEND_STATUS > (CMD13) to poll any additional > status information the Device may have (e.g., WP_ERASE_SKIP, etc.)." Isn't that exactly what is being done? After the card has stopped signaling busy, we send a CMD13 in mmc_busy_cb() to read the additional status information. I don't get it, why should the card stop signal busy, just because we send a CMD13. If so, the card should probably be considered broken. For broken cards, we can try to use a card-quirk instead - to enforce CMD13 polling. > > > > I would suggest that we first try to fix the implementation of the > > > ->card_busy() callback for your HW. If that isn't possible or fails, > > > then let's consider the approach you have taken in the $subject patch. > > > > I have verified this with hardware behavior and default > sdhci->card->busy() seems to work fine except > in the above sequence of CMD35->CMD36->CMD38 without any MMC_STATUS interleaved. > > Maybe we can do both, check the card busy and send CMD13. > __mmc_poll_for_busy() does have > throttling changes as well that would limit the rate at which CMD13 > will be sent. Polling with both wouldn't help, but make the polling-loop heaviers, we certainly don't want that. > > > Note, sdhci drivers can override host->ops. For example, > > sdhci-omap.c has: > > > > host->mmc_host_ops.card_busy = sdhci_omap_card_busy; > > > > Probably, if ->card_busy() cannot be supported, then setting > > it to NULL would work. > > > > host->mmc_host_ops.card_busy = NULL; /* Cannot detect card busy */ > > > > Yes, that's an option I have tested to be working and will have to > take in case mmc_busy_cb() change is not acceptable. > > Thanks > Kamal Kind regards Uffe