Re: Rock64 configuration fails to boot for main 22c4ab6cb015 but worked for main 06bd74e1e39c (Nov 21): e.MMC mishandled?
Date: Fri, 17 Dec 2021 15:49:30 UTC
On 2021-Dec-17, at 03:17, Kornel Dulęba <mindal@semihalf.com> wrote:
>>>>> [I've cut out the history: just presenting some new evidence.]
>>>>>
>>>>> First, a little context from getting to the db> prompt.
>>>>>
>>>>> db> ps
>>>>> pid ppid pgrp uid state wmesg wchan cmd
>>>>> 18 0 0 0 DL syncer 0xffff000000eca5a8 [syncer]
>>>>> 17 0 0 0 DL vlruwt 0xffffa00007d2ea60 [vnlru]
>>>>> 16 0 0 0 DL (threaded) [bufdaemon]
>>>>> 100089 D qsleep 0xffff000000ec9478 [bufdaemon]
>>>>> 100092 D - 0xffff000000c11100 [bufspacedaemon-0]
>>>>> 100093 D - 0xffff000000c21680 [bufspacedaemon-1]
>>>>> 9 0 0 0 DL psleep 0xffff000000ef0650 [vmdaemon]
>>>>> 8 0 0 0 DL (threaded) [pagedaemon]
>>>>> 100087 D psleep 0xffff000000ee2b38 [dom0]
>>>>> 100094 D launds 0xffff000000ee2b44 [laundry: dom0]
>>>>> 100095 D umarcl 0xffff0000007b38d8 [uma]
>>>>> 7 0 0 0 DL mmcsd d 0xffffa00007b72e00 [mmcsd0boot1: mmc/sd]
>>>>> 6 0 0 0 DL mmcsd d 0xffffa00007b71300 [mmcsd0boot0: mmc/sd]
>>>>> 5 0 0 0 DL mmcreq 0xffff00009b5d0710 [mmcsd0: mmc/sd card]
>>>>> 4 0 0 0 DL - 0xffff000000ccc020 [rand_harvestq]
>>>>> 15 0 0 0 DL (threaded) [usb]
>>>>> . . .
>>>>>
>>>>> and "mmcreq" is from the while loop in:
>>>>>
>>>>> static int
>>>>> mmc_wait_for_req(struct mmc_softc *sc, struct mmc_request *req)
>>>>> {
>>>>>
>>>>> req->done = mmc_wakeup;
>>>>> req->done_data = sc;
>>>>> if (__predict_false(mmc_debug > 1)) {
>>>>> device_printf(sc->dev, "REQUEST: CMD%d arg %#x flags %#x",
>>>>> req->cmd->opcode, req->cmd->arg, req->cmd->flags);
>>>>> if (req->cmd->data) {
>>>>> printf(" data %d\n", (int)req->cmd->data->len);
>>>>> } else
>>>>> printf("\n");
>>>>> }
>>>>> MMCBR_REQUEST(device_get_parent(sc->dev), sc->dev, req);
>>>>> MMC_LOCK(sc);
>>>>> while ((req->flags & MMC_REQ_DONE) == 0)
>>>>> msleep(req, &sc->sc_mtx, 0, "mmcreq", 0);
>>>>> MMC_UNLOCK(sc);
>>>>> if (__predict_false(mmc_debug > 2 || (mmc_debug > 0 &&
>>>>> req->cmd->error != MMC_ERR_NONE)))
>>>>> device_printf(sc->dev, "CMD%d RESULT: %d\n",
>>>>> req->cmd->opcode, req->cmd->error);
>>>>> return (0);
>>>>> }
>>>>>
>>>>> So it appears that the error report:
>>>>>
>>>>> mmcsd0: Error indicated: 4 Failed
>>>>>
>>>>> ends up associated with (req->flags & MMC_REQ_DONE) == 0 staying
>>>>> true in the above code: an unbounded loop with MMC_LOCK(sc) active.
>>>>> The "4" in the error report seems to be from:
>>>>>
>>>>> #define MMC_ERR_FAILED 4
>>>>>
>>>>> It looks like there are some problems with handling errors, problems
>>>>> such that it gets stuck looping (no panic, no progress).
>>>>>
>>>>> That seems to be separate from why the MMC_ERR_FAILED was generated
>>>>> in the first place. So: 2 problems, not just one. Thus it may be a
>>>>> good context for tackling the looping problem with a known example
>>>>> failure to look at.
>>>>>
>>>>>
>>>>>
>>>>> Just for reference, I tried "boot -v" with debug.verbose_sysinit=1 in place,
>>>>> just to capture and report the tail of the output for the boot failure.
>>>>>
>>>>> . . .
>>>>> subsystem f000000
>>>>> release_aps(0)... Release APs...done
>>>>> done.
>>>>> intr_irq_shuffle(0)... Trying to mount root from ufs:/dev/gpt/Rock64root []...
>>>>> done.
>>>>> netisr_start(0)... done.
>>>>> taskqgroup_bind_softirq(0)... done.
>>>>> GEOM: new disk mmcsd0
>>>>> GEOM: new disk mmcsd0boot0
>>>>> GEOM: new disk mmcsd0boot1
>>>>> smp_after_idle_runnable(0)... done.
>>>>> taskqgroup_bind_if_config_tqg(0)... done.
>>>>> taskqgroup_bind_if_io_tqg(0)... done.
>>>>> tmr_setup_user_access(0)... done.
>>>>> subsystem f000001
>>>>> mmcsd0: Error indicated: 4 Failed
>>>>> epoch_init_smp(0)... done.
>>>>> subsystem f100000
>>>>> racctd_init(0)... done.
>>>>> subsystem fffffff
>>>>> start_periodic_resettodr(0)... done.
>>>>> oktousecallout(0)... done.
>>>>> clknode_finish(0)... Unresolved linked clock found: hdmi_phy
>>>>> Unresolved linked clock found: usb480m_phy
>>>>> done.
>>>>> regulator_constraint(0)... done.
>>>>> regulator_shutdown(0)... regulator: shutting down unused regulators
>>>>> regulator: shutting down vcc_sd... busy
>>>>> done.
>>>>> uhub0: 1 port with 1 removable, self powered
>>>>> uhub2: 2 ports with 2 removable, self powered
>>>>> uhub3: 1 port with 1 removable, self powered
>>>>> uhub1: 1 port with 1 removable, self powered
>>>>> ugen4.2: <Samsung PSSD T7 Touch> at usbus4
>>>>> umass0 on uhub2
>>>>> umass0: <Samsung PSSD T7 Touch, class 0/0, rev 3.20/1.00, addr 1> on usbus4
>>>>> umass0: SCSI over Bulk-Only; quirks = 0x0000
>>>>> umass0:0:0: Attached to scbus0
>>>>> pass0 at umass-sim0 bus 0 scbus0 target 0 lun 0
>>>>> pass0: <Samsung PSSD T7 Touch 0> Fixed Direct Access SPC-4 SCSI device
>>>>> pass0: Serial Number REPLACED
>>>>> pass0: 400.000MB/s transfers
>>>>> da0 at umass-sim0 bus 0 scbus0 target 0 lun 0
>>>>> da0: <Samsung PSSD T7 Touch 0> Fixed Direct Access SPC-4 SCSI device
>>>>> da0: Serial Number REPLACED
>>>>> da0: 400.000MB/s transfers
>>>>> da0: 953869MB (1953525168 512 byte sectors)
>>>>> da0: quirks=0x2<NO_6_BYTE>
>>>>> da0: Delete methods: <NONE(*),ZERO>
>>>>> random: unblocking device.
>>>>>
>>>>> No more output after that.
>>>>
>>>> As for why MMC_ERR_FAILED results, the following code diff is
>>>> intended to suggest what I think may be incomplete about sticking
>>>> to what the device-specific code supports vs. does not support
>>>> (not supporting HS200 here). The code does compile in my context
>>>> but is untested.
>>>
>>> It is now tested (at least to be a useful hack): no longer am I
>>> running an older 1400042 kernel. For reference,
>>>
>>> # uname -apKU
>>> FreeBSD Rock64_RPi_4_3_2v1p2 14.0-CURRENT FreeBSD 14.0-CURRENT #18 main-n251456-22c4ab6cb015-dirty: Sun Dec 12 00:34:53 PST 2021 root@CA72_16Gp_ZFS:/usr/obj/BUILDs/main-CA53-nodbg-clang/usr/main-src/arm64.aarch64/sys/GENERIC-NODBG-CA53 arm64 aarch64 1400043 1400043
>>>
>>> And it reports during the boot (other than the "REPLACED"):
>>>
>>> mmcsd0: 125GB <MMCHC DJNB4R 0.7 SN REPLACED MFG 06/2016 by 21 0x0000> at mmc0 52.0MHz/8bit/1016-block
>>>
>>> So it no longer sets up a mode that the rk3328-specific-code does not
>>> actually support.
>>>
>>> (Nothing that I've done here deals with the looping issue when there
>>> is a MMC_ERR_FAILED or the like.)
>>>
>>>> The email handling may mess up some leading
>>>> whitespace --but, again, I'm only trying to suggest a type of
>>>> change.
>>>>
>>>> # git -C /usr/main-src/ diff /usr/main-src/sys/dev/mmc
>>>> diff --git a/sys/dev/mmc/mmc.c b/sys/dev/mmc/mmc.c
>>>> index 9c73dfd57ce0..dffd1c382684 100644
>>>> --- a/sys/dev/mmc/mmc.c
>>>> +++ b/sys/dev/mmc/mmc.c
>>>> @@ -59,6 +59,7 @@ __FBSDID("$FreeBSD$");
>>>> #include <sys/param.h>
>>>> #include <sys/systm.h>
>>>> #include <sys/kernel.h>
>>>> +#include <sys/kobj.h>
>>>> #include <sys/malloc.h>
>>>> #include <sys/lock.h>
>>>> #include <sys/module.h>
>>>> @@ -1512,6 +1513,8 @@ mmc_timing_to_string(enum mmc_bus_timing timing)
>>>> static bool
>>>> mmc_host_timing(device_t dev, enum mmc_bus_timing timing)
>>>> {
>>>> + kobjop_desc_t kobj_desc;
>>>> + kobj_method_t *kobj_method;
>>>> int host_caps;
>>>>
>>>> host_caps = mmcbr_get_caps(dev);
>>>> @@ -1543,14 +1546,37 @@ mmc_host_timing(device_t dev, enum mmc_bus_timing timing)
>>>> case bus_timing_mmc_ddr52:
>>>> return (HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_DDR52));
>>>> case bus_timing_mmc_hs200:
>>>> - return (HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS200_120) ||
>>>> - HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS200_180));
>>>> case bus_timing_mmc_hs400:
>>>> - return (HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS400_120) ||
>>>> - HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS400_180));
>>>> case bus_timing_mmc_hs400es:
>>>> - return (HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS400 |
>>>> - MMC_CAP_MMC_ENH_STROBE));
>>>> + /*
>>>> + * Disable eMMC modes that require use of
>>>> + * MMC_SEND_TUNING_BLOCK_HS200 to set things up if either the
>>>> + * tune or re-tune method is the default NULL implementation.
>>>> + */
>>>> + kobj_desc = &mmcbr_tune_desc;
>>>> + kobj_method = kobj_lookup_method(((kobj_t)dev)->ops->cls, NULL,
>>>> + kobj_desc);
>>>> + if (kobj_method == &kobj_desc->deflt)
>>>> + return (false);
>>>> + kobj_desc = &mmcbr_retune_desc;
>>>> + kobj_method = kobj_lookup_method(((kobj_t)dev)->ops->cls, NULL,
>>>> + kobj_desc);
>>>> + if (kobj_method == &kobj_desc->deflt) {
>>>> + return (false);
>>>> + }
>>>> +
>>>> + /*
>>>> + * Otherwise track the host capabilities.
>>>> + */
>>>> + if (timing == bus_timing_mmc_hs200)
>>>> + return (HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS200_120) ||
>>>> + HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS200_180));
>>>> + if (timing == bus_timing_mmc_hs400)
>>>> + return (HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS400_120) ||
>>>> + HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS400_180));
>>>> + if (timing == bus_timing_mmc_hs400es)
>>>> + return (HOST_TIMING_CAP(host_caps, MMC_CAP_MMC_HS400 |
>>>> + MMC_CAP_MMC_ENH_STROBE));
>>>> }
>>>>
>>>> #undef HOST_TIMING_CAP
>>>>
>>>>
>>>> In other words: have mmc_host_timing avoid returning true for some
>>>> combinations that definitely do not have sufficient software support
>>>> present at the time. (So far as I can tell, the rk3328's get the
>>>> NULL-implementations as things are.)
>>>>
>>>> I expect that this sort of thing would go back to using
>>>> MMC_CAP_MMC_DDR52 for the rk3328's, as an example. Working, but in a
>>>> slower mode, the same mode as FreeBSD was previously using.
>>>>
>>>> A possible incompleteness in the suggestion is that there is also a
>>>> drive-strength setting involved. If that also had "kobj" interfacing
>>>> and NULL-implementation possibilities, then in the future there would
>>>> be more to test for possibly forcing return-false than I did above.
>>>>
>>>> Hopefully this sort of thing would help, possibly more than just for
>>>> rk3328's.
>>>
>>>
>>
>> As for what was happening without my patch . . .
>>
>> sys/dev/mmc/mmcbr_if.m defines:
>>
>> static int
>> null_retune(device_t brdev __unused, device_t reqdev __unused,
>> bool reset __unused)
>> {
>>
>> return (0);
>> }
>>
>> static int
>> null_tune(device_t brdev __unused, device_t reqdev __unused,
>> bool hs400 __unused)
>> {
>>
>> return (0);
>> }
>> . . .
>> #
>> # Called by the mmcbus with the bridge claimed to execute initial tuning.
>> #
>> METHOD int tune {
>> device_t brdev;
>> device_t reqdev;
>> bool hs400;
>> } DEFAULT null_tune;
>>
>> #
>> # Called by the mmcbus with the bridge claimed to execute re-tuning.
>> #
>> METHOD int retune {
>> device_t brdev;
>> device_t reqdev;
>> bool reset;
>> } DEFAULT null_retune;
>> . . .
>>
>> It is these success-reporting no-op routines that were being
>> used to attempt the tuning: so there was no tuning done.
>>
>> The code that I added detects that these routines would be
>> used and avoids allowing contexts that would involve putting
>> them to use with HS200 mode.
>>
>> I'll note that there is another such null_* routine that the
>> code (even with my patch) does not deal with avoiding the use
>> of:
>>
>> . . .
>> static int
>> null_switch_vccq(device_t brdev __unused, device_t reqdev __unused)
>> {
>>
>> return (0);
>> }
>> . . .
>> #
>> # Called by the mmcbus to switch the signaling voltage (VCCQ).
>> #
>> METHOD int switch_vccq {
>> device_t brdev;
>> device_t reqdev;
>> } DEFAULT null_switch_vccq;
>> . . .
>>
>> /usr/main-src/sys/dev/sdhci/sdhci.c has somewhat analogous code for
>> somewhat analogous null_* routines. null_set_uhs_timing for that is
>> from sys/dev/sdhci/sdhci_if.m (but the other two are again the above
>> null_tune and null_retune routines, so not repeated here):
>>
>> . . .
>> static void
>> null_set_uhs_timing(device_t brdev __unused,
>> struct sdhci_slot *slot __unused)
>> {
>>
>> }
>> . . .
>> METHOD void set_uhs_timing {
>> device_t brdev;
>> struct sdhci_slot *slot;
>> } DEFAULT null_set_uhs_timing;
>> . . .
>>
>> sdhci_init_slot(device_t dev, struct sdhci_slot *slot, int num)
>> in sdhci.c looks like (in part):
>>
>> . . .
>> /*
>> * Disable UHS-I and eMMC modes if the set_uhs_timing method is the
>> * default NULL implementation.
>> */
>> kobj_desc = &sdhci_set_uhs_timing_desc;
>> kobj_method = kobj_lookup_method(((kobj_t)dev)->ops->cls, NULL,
>> kobj_desc);
>> if (kobj_method == &kobj_desc->deflt)
>> host_caps &= ~(MMC_CAP_UHS_SDR12 | MMC_CAP_UHS_SDR25 |
>> MMC_CAP_UHS_SDR50 | MMC_CAP_UHS_DDR50 | MMC_CAP_UHS_SDR104 |
>> MMC_CAP_MMC_DDR52 | MMC_CAP_MMC_HS200 | MMC_CAP_MMC_HS400);
>>
>> #define SDHCI_CAP_MODES_TUNING(caps2) \
>> (((caps2) & SDHCI_TUNE_SDR50 ? MMC_CAP_UHS_SDR50 : 0) | \
>> MMC_CAP_UHS_DDR50 | MMC_CAP_UHS_SDR104 | MMC_CAP_MMC_HS200 | \
>> MMC_CAP_MMC_HS400)
>>
>> /*
>> * Disable UHS-I and eMMC modes that require (re-)tuning if either
>> * the tune or re-tune method is the default NULL implementation.
>> */
>> kobj_desc = &mmcbr_tune_desc;
>> kobj_method = kobj_lookup_method(((kobj_t)dev)->ops->cls, NULL,
>> kobj_desc);
>> if (kobj_method == &kobj_desc->deflt)
>> goto no_tuning;
>> kobj_desc = &mmcbr_retune_desc;
>> kobj_method = kobj_lookup_method(((kobj_t)dev)->ops->cls, NULL,
>> kobj_desc);
>> if (kobj_method == &kobj_desc->deflt) {
>> no_tuning:
>> host_caps &= ~(SDHCI_CAP_MODES_TUNING(caps2));
>> }
>> . . .
>>
>> What I've done in my patch is analogous to what the the code shown
>> after the #define SDHCI_CAP_MODES_TUNING above does, translated to
>> fit the mmc's pre-existing code structure.
>
> Good catch. For some reason mmcbr_tune/mmcbr_retune are not
> implemented in sdhci_fdt.c.
> This looks like a separate bug/issue.
> Note that pretty much all other SDHCI drivers (sdhci_pci.c,
> sdhci_xenon.c, sdhci_fsl_fdt.c, sdhci_acpi.c) provide some
> implementation of mmcbr_tune/mmcbr_retune.
> I agree that the logic in sdhci.c should disable HS200 if those
> methods are implemented.
> Could you try hooking mmcbr_tune and mmcbr_retune to their generic
> sdhci implementations?
What valid generic sdhci implementations? There is no such
thing as far as I can tell. JESD84-B51 (e.MMC v5.1's standard)
does nothing to standardize how adjustments are made during
HS200 tuning, beyond how to get a known-pattern to test against
(via: CMD21). That, of itself, does not make any adjustments.
Also, as far as I can tell, the mmc that I changed and the
sdhci code quoted above are independent at run-time. One or
the other is used, not both. Otherwise the above quoted
code would have prevented HS200 from being attempted.
> If we're lucky this could be enough to make HS200 work on rock64,
> though it's unlikely.
I do not see that there is any generic sdhci tuning code to
enable the use of: just avoiding use of code that is a no-op
that falsely indicates success by avoiding needing HS200
tuning.
===
Mark Millard
marklmi at yahoo.com