Commit 66a7393a3ba9685d1eddfbce72e3ef8f4848f19f ("mmc: block: ensure CMDQ
is empty before queuing cache flush") added a workaround for particular
vendor's eMMC devices. Workaround was to wait for all the outstanding
requests to finish up before queuing the flush request. Now detailed
root cause analysis from vendor shows that original issue can happen only
if DCMD command is sent to device too quickly (within less than 6
microseconds) after completion of previous small sector (less than 8
sectors) read operations. Hence with this change, we are fine tuning the
previous workaround such that it would almost have no impact on the storage
benchmark performance numbers.
Change-Id: I1df1c5d7bbcd7b526236651077b7dade2626cb30
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
[subhashj@codeaurora.org: fixed trivial merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
eMMC runtime suspend first puts the CMDQ to halt and then sets the card
state to suspended after sending out sleep command. If new requests get
queued after halting the CMDQ but before the card state has changed to
suspended then command queue thread will not be woken up again until
some new requests are queued to request queue. And if new request gets
queued, we will remain in this state forever. Fix this issue by kicking
the CMDQ thread after the suspend completion.
Change-Id: I37a03d9a75acf2ab3ebda57da87e246b449abf18
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
eMMC CMDQ is pretty latest feature and we may continue to see issues
due to HW & SW immaturity. To debug these issues, we should have the
host controller register dumps along with current SW state from rampdumps.
This change adds the BUG_ON to crash the system in case of eMMC related
command timeouts.
Change-Id: Ib832aef6eb9ef3941fbe9e73852929c835e97847
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
[subhashj@codeaurora.org: fixed trivial merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Some devices might stop responding to new commands if device cache flush
is queued to host controller while host controller is still processing
outstanding requests.
To workaround this issue, we are making the cmdq thread to wait for all
the oustanding requests to be finished before queuing cache flush command
on host controller.
Change-Id: I15387734f51ca4cadfc9e11270f14d8a0806a00f
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
[subhashj@codeaurora.org: fixed trivial merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
The current code has some compilation errors when compiling for
single core configurations. This fixes those errors.
Change-Id: Idd1d1bab343cf5cf1e35f229d913d16a7854f358
Signed-off-by: Krishna Konda <kkonda@codeaurora.org>
Signed-off-by: Ritesh Harjani <riteshh@codeaurora.org>
Disable power management features for SDIO as the current code is
not ready to enable it.
Change-Id: I234ebf6cc46d2a3b70bf9ba3f94b4b1abc1acd44
Signed-off-by: Pavan Anamula <pavana@codeaurora.org>
There could be 2 different race conditions possible with current shutdown
handler:
1. Shutdown handler sees that there is no request in queue but mmc-cmdqd
might have fetched the request from queue and about to issue it to
driver.
2. Shutdown handler keeps looping in the while(1) loop as it sees request
being pending in the request queue. But mmc-cmdqd thread doesn't wake up
to fetch the requests from the request queue. mmc-cmdqd thread isn't
waking up because shutdown path has stopped queue and request came into
the queue after that. Once queue is stopped, block layer won't invoke
the queue's ->request_fn() callback to notify driver about pending
request.
Remedy to fix both these race condition is simple. In shutdown handler,
make sure that we drain (& complete) all the outstanding requests from the
queue and then don't allow any new requests to be queued. Block layer API
blk_cleanup_queue() precisely does what we want and this change basically
use the same API in shutdown handler.
Change-Id: I761ba6a2e2974d955bb72ff993b1cc2c32c9ec29
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Some host controllers may need additional TLMM registers to be
configured to enable the IO lines connected to it.
Change-Id: Ic334ce653bf13fef5969f08e19f6202377b8fd2e
Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
ICE init & reset can be synchronous now because ICE does not need
to go to secure side for any ICE configuration. This would simplify
interface and make call more efficient.
Change-Id: I7aa4e2d3ba3383d25758b21b8ae261a0220f35f9
Signed-off-by: Dinesh K Garg <dineshg@codeaurora.org>
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: Gilad Broner <gbroner@codeaurora.org>
[subhashj@codeaurora.org: fixed merge conflicts, dropped changes
to ICE and UFS driver as they are already present]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
As SDIO spec doesn't allow advertisement of 1.8v support, some SDIO
devices advertise support of only 3.0v even though they support 1.8v
as well.
sdhc3 host controller only supports 1.8v and rejects the initialization
of SDIO devices that advertise 3.0v support.
This change adds fake support for 3.0v to sdhc host controller.
This will allow initialization of SDIO devices that supports 1.8v but
advertise 3.0v support.
Change-Id: I5a98c54ad4998e6439f83081628c9c083e95bbf0
Signed-off-by: Pavan Anamula <pavana@codeaurora.org>
In addition to dumping all the SDCC vendor specific registers,
also dump the ICE registers that helps in debugging crypto
related errors.
Change-Id: I804b0cfd77dec21d45016aafbbb001dafb050241
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
Clock gating logic maintains the reference count which gets incremented
with *hold* API call and get decremented with *release* API call.
If we want to make sure that clocks are enabled during some operation,
we should call *hold* API before the operation and then call *release*
API after the operation is finished. So *hold* and *release* should be
in pair and called exactly once per operation, any violation of these
rules may put the reference count out of sync and can cause issues
(like unclocked register access). This change fixes the places where
this can go out of sync.
Change-Id: I2d784cc31027a551a466beec63ee8cd0a1f67af7
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Keep track of task history for the last 32 tasks and dump it
along with registers in case of any error for debugging.
The log is of the following format:
[index] <DATA|DCMD> Task: <lower 32bits> | Args: <upper 32bits>
of the task descriptor structure. The values need to be decoded
accordingly depending on the data or dcmd task descriptor.
The last entry index denotes the latest entry in this list.
---- Circular Task History ----
cmdq-host: Last entry index: 1
cmdq-host: [00]DATA Task: 0x0400002f | Args: 0x00d37d18
cmdq-host: [01]DATA Task: 0x0400002f | Args: 0x00d38118
cmdq-host: [02]DCMD Task: 0x0186402f | Args: 0x03200101
Add a debugfs entry to enable/disable this dynamically. It is
disabled by default. This applies only to eMMC devices.
Usage: echo Y > /sys/kernel/debug/mmcX/cmdq_task_history
X - denotes the slot id
Change-Id: I6abf29aa928d3d8270405cfc104241043dadfe45
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
[subhashj@codeaurora.org: fixed compilation error]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
If the last access to eMMC before runtime/system suspend is an
RPMB access, the partition type within EXT_CSD[179] will be set
to RPMB. As per specification, the deselect CMD7 and sleep CMD5
are ignored by the card and are treated as illegal commands in
this state. This causes eMMC sleep command to timeout and thus
fails runtime/system suspend. Hence, switch to default partition
config before sending deselect CMD7 and sleep CMD5.
CRs-fixed: 630894
Change-Id: I40f3fb590aeba787de8ca3356a8eed5f2780bcc1
Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
[subhashj@codeaurora.org: fixed merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Fix the timeout issue with mmc_cmdq_halt() that happens in the
below scenario -
1. Let us say an RPMB request is processed just before the
runtime suspend. The RPMB request will put the controller in HALT
state and will put the card in legacy mode.
2. So during runtime suspend, the SW will check for card's state and
since it is in legacy mode, it won't call CQ disable host->op.
But CQ HW will get reset due to SW reset that gets issued during suspend.
3. If we get another RPMB request, we try to do the runtime resume.
But since SW did not invoke CQ disable host->op, it thinks that CQ is still
enabled and returns here without enabling CQ.
4. Now the RPMB request tries to put the controller in HALT state but
timesout since CQ is still not enabled.
Fix this issue by checking for host->card->cmdq_init to determine
if the controller is initialized in CQ mode instead of checking for
card's state in runtime suspend. This will ensure CQ will be disabled
as part of runtime suspend and will be enabled again during
runtime resume.
Change-Id: I26bf97962d31522a9cb25009c6a048dfff6e4535
Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
During RPMB operation, command queue mode would be disabled and we don't
exlicitly move back to command queue once the RPMB operation is completed.
We expect the command queue thread to switch operating mode (or partition)
to command queue before issuing the new transfers in command queue but
command queue thread isn't getting scheduled if we don't wake it up
explicitly or if some other new transfer get submitted to the queue.
Fix this issue by explicitly waking up the command queue thread after the
completion of RPMB operation.
Change-Id: I68b2f7989d68b51b4810346458e1966d45aee5a2
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Currently we are not releasing the runtime PM reference count after
halt/unhalt completion which basically keeps the runtime PM state to
active forever. Fix this by releasing the PM reference count on
completion of halt/unhalt operation.
Change-Id: I28a37917b49acde7f5d75bf9a639d0eb67a1c169
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Currently we are printing the cmdq debug ram contents during the
error handling but this code path shouldn't be executed for non cmdq
hosts.
Change-Id: Ic397e4378c290b604037e69b7df58200810e981c
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Currently we are crashing the system if PM QoS request counter goes
negative but this doesn't seem to be right way to handle. We should
instead skip decrementing this counter once it reaches 0 and just
print a dump stack to know the callstack. This change fixes this
as mentioned above.
Change-Id: I36fb03b1ddf8e04ecc9fe449496b656db84e77d2
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Currently we might end up calling ->post_req() 2 times if we failed to
submit the request to host driver. This change fixes this issue.
Change-Id: I10b681ab955e0a9126b22df68d3601b75328c949
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Currently all functions which calls cmdq_halt relies upon HALT
to pass other wise considers it as a FATAL error (like clk scaling).
So add retry mechanism in halt - retry if halt completion
timesout & HALT also did not complete(by doing register read).
Change-Id: I59e199681e46085f804636f4c07249e6f21de1d5
Signed-off-by: Ritesh Harjani <riteshh@codeaurora.org>
CQ irq handler first calls completion path for every notified tag and then
clears CQTCN register for all completed tags. This approach could cause
following problem:
1) issue context (unblocked by tag completion) issues next
request, CQE successfully completes it and notifies task completion
through CQTCN again for the same tag.
2) CQ irq handler proceeds to clear CQTCN for the original request.
In the above scenario clear of CQTCN register will mask next
request completion and issue context will be never unblocked.
Hence clear CQTCN first and then notify the request completion.
Change-Id: Ie644e24279ca30de42bbc9f8e1ae4326609d38a5
Signed-off-by: Konstantin Dorfman <kdorfman@codeaurora.org>
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
Enable partial init on eMMC devices for using card Sleep/Awake
sequence (CMD5).
Card will be sent to Sleep state during runtime/system suspend
and will be woken up during runtime/system resume.
By using this sequence the card doesn't need full initialization
which gives time reduction in system/runtime resume path.
Change-Id: Ide2b8050b56e858cda7bb22b4acf43b5bd3db786
Signed-off-by: Talel Shenhar <tatias@codeaurora.org>
Signed-off-by: Maya Erez <merez@codeaurora.org>
We make sure that the doorbell bit for any regular slot is not
set before ringing the doorbell for that slot, make sure of the
same for DCMD slot too.
Change-Id: Ia96e8d6ae0c28aad21f4a3cf46b27c7a5c878971
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
Keep a reference to the sdhc host instance to retrieve all the
necessary data structures in case of a crash for debugging.
Change-Id: Ic8d696d53cf48290dbcc92f1567650989dd45755
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
If request has to be requeued (due to any DCMD commmand pending or cmdq
being halted) and if we change the task status to interruptible before
going to sleep then thread may not wakeup again. Note that
blk_requeue_request() doesn't trigger ->request_fn() again to wakeup
the thread.
Fix this issue by making cmdq thread wait for the completion of DCMD
or until the cmdq is unhalted. This change also simplifies the
cmdq thread function.
Change-Id: Iebffc993241e5fadb2962fedc44576566dc66e9c
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
The SDHC spec allows to force raise errors that is useful for
debugging error handler routines. Add debugfs entry force_error
to trigger host errors from userspace. Check SDHCI_SET_INT_ERROR
register for error bitmask info.
Usage: echo 0x1 > /sys/kernel/debug/mmcX/force_error
X - denotes the slot id
Change-Id: I9f67442a79b2645cbdc3020d1a10c0b32840ce32
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
[subhashj@codeaurora.org: fixed trivial merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
To trigger queue status command after sending dcmd, we need
to set bit 31 of CQ_VENDOR_CFG register. But the current code
incorrectly sets bit 31 of CQCTL register.
Change-Id: Ic5b914cf6a5237ac51b2104453caba2c49c1efbc
Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
With newer mmc drivers, max_discard_to has been remvoed from the sdhci
driver. So instead of incorrectly using max_busy_timeout for
calculating max_discard to be used by mmc queue, use the mmc cap to
indicate that the max discard should be used for msm sdhci driver.
Change-Id: I424cd0a5ee9ffd7199be58a5a091984c5fcda52f
Signed-off-by: Krishna Konda <kkonda@codeaurora.org>
[subhashj@codeaurora.org: fixed merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
max_busy_timeout is used to decide whether R1B response should be used or a
R1 response should be used. This is also used to decide what the discard
size of mmc queue (registered with block layer) can be set to. In order to
keep both the features in place, this change will allow for hosts to
specify a larger discard size while still specifying max_busy_timeout.
Change-Id: I1e607329c4377897a7cb4086db02cbc150bd02b7
Signed-off-by: Krishna Konda <kkonda@codeaurora.org>
"downdifferential" parameter of devfreq's simple ondemand governor supposed
to report expected difference between upthreshold and downthreshold value.
In other words, downdifferential = upthreshold - downthreshold. But
currently downdifferential is set to same as downthreshold, this change
fixes this issue.
Change-Id: Ic2e762d192f1fed8f94d4d2579d6a4b5d4c2c8b5
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
The controller triggers an ADMA error when ADMA engine is configured
and used in 32-bit mode. This is because the current code always
writes 64-bit address to 32-bit address field of a transfer
descriptor (bits [63:32]). This corrupts the first 32-bit value
of the next transfer descriptor.
Below scenario describes how ADMA error can happen -
1. Req#1 - uses slot 1, prepares it's descriptors, queues to the controller
2. Req#2 - uses slot 0, prepares max descriptors (cq_host->mmc->max_segs).
3. Req#1 gets ADMA error from the controller.
At step 2, when it prepares the last transfer descriptor (max_segs), it
overwrites the 32-bit address field with a 64-bit address and thus corrupts
the first entry of slot 1 transfer descriptor.
Change-Id: I3eb2dbb40c76ec77626f647d6ec24df4a0858fcb
Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
Add sysfs entries to allow getting the current status of
PM QoS voting and enable or disable voting.
Change-Id: If5b8e4b155090343112916c9c57a766bb2104e10
Signed-off-by: Gilad Broner <gbroner@codeaurora.org>
Starting from MCI_VERSION 4.2.0 use a different offset for the following
registers:
PERIPH_SS_SDC1_SDCC_HC_CMDQ_COMMAND_DEBUG_RAM_n
PERIPH_SS_SDC1_SDCC_HC_CMDQ_COMMAND_DEBUG_RAM_WRAPAROUND
PERIPH_SS_SDC1_SDCC_HC_CMDQ_COMMAND_DEBUG_RAM_OVERLAP
In addition, move dump debug ram functionality to the MSM specific file.
Change-Id: I3b0f5101150de9f2c190ce69b306bd151cbb65ae
Signed-off-by: Konstantin Dorfman <kdorfman@codeaurora.org>
IOCTL and RPMB commands can be issued while the device is busy
with background Operations handling.
Stop the device BKOPs before handling the RPMB / IOCTL command.
Change-Id: I088f74c77026ccd901276e1214e4466ac7815bf1
Signed-off-by: Maya Erez <merez@codeaurora.org>
[subhashj@codeaurora.org: fixed merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
DCMD requests don't have data, check for the data pointer before
accessing it to prevent null pointer dereference in case of DCMD
timeout err. Also signal a completion event for non flush requests
like discard that wait for the completion of DCMD request.
Change-Id: Ia71a5f1e278a039ba22f6ac42614d9ae79dba7e9
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
Read the firmware version from ext_csd register and print it
for debugging purpose.
Change-Id: I4c1fefd5bff753915c9858fb35c958335986c778
Signed-off-by: Asutosh Das <asutoshd@codeaurora.org>
Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
[subhashj@codeaurora.org: fixed trivial merge conflicts and
compilation errors]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Dump crypto related register information during error for
debugging purpose.
Change-Id: I8976e8c0b5e9bda910634464202578dbacd7666e
Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
Clock scaling commands are not allowed to be sent when the device
performs manual BKOPS.
In case the device is busy with manual BKOPS, skip the clock
scaling and allow the device to continue the BKOPS.
Clock scaling will be invoked again later by dev freq.
Change-Id: I78b505326d245b1ddcd9d6f1cdd4294850889d45
Signed-off-by: Maya Erez <merez@codeaurora.org>
In case there are no pending requests, runtime idle API can be
invoked and start manual BKOPS.
We need to check if manual BKOPS is enabled on the device and stop
it before serving new incoming requests in CQ mode.
Change-Id: I870eff40ea9fa91eedb4c0d2600c32d8534a3868
Signed-off-by: Maya Erez <merez@codeaurora.org>
This change adds the ability to partially initialize
the MMC card by using card Sleep/Awake sequence (CMD5).
Card will be sent to Sleep state during runtime/system suspend
and will be woken up during runtime/system resume.
By using this sequence the card doesn't need full initialization
which gives time reduction in system/runtime resume path.
Change-Id: Id8dadf03ef4226f7c4675fadbacac7bb15c0289e
Signed-off-by: Talel Shenhar <tatias@codeaurora.org>
Signed-off-by: Maya Erez <merez@codeaurora.org>
[subhashj@codeaurora.org: fixed merge conflicts & compilation errors]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Add PM QoS voting mechanism to sdhci-msm driver for legacy
eMMC.
Two types of voting schemes are supported:
1) Vote for HW IRQ
2) Vote for a cpu group according to the request's designated cpu
Using PM QoS voting should benefit performance.
Change-Id: I5d2b71fc4eabfa5060f343634fbc7363f2ee1344
Signed-off-by: Konstantin Dorfman <kdorfman@codeaurora.org>
Signed-off-by: Gilad Broner <gbroner@codeaurora.org>
[subhashj@codeaurora.org: fixed merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Add PM QoS voting mechanism to sdhci-msm driver for command queueing.
Two types of voting schemes are supported:
1) Vote for HW IRQ
2) Vote for a cpu group according to the request's designated cpu
Using PM QoS voting should benefit performance.
Change-Id: I8a20653eeb6348d5b442c846708d92c8fb64a8e9
Signed-off-by: Gilad Broner <gbroner@codeaurora.org>
[subhashj@codeaurora.org: fixed trivial merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
When encryption/decryption is enabled in CQ mode, the
legacy commands that are sent in HALT state will use
different slot other than slot 0 for crypto configuration
information. The slot that is selected depends on the last
slot that was used when it is in CQ mode. This is causing
the data of legacy commands to be encrypted/decrypted based
on the wrong slot usage for crypto config details. Hence,
clear the crypto configuration of the slot used in CQ mode
whenever it gets completed.
Change-Id: I6817de46d895b61f410dd732be57ba60efb58fb2
Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
There can be a race condition between runtime idle and incoming
requests. In such a race condition, we can send requests
while the queue is halted.
Claiming the host at the beginning of runtime idle will prevent
such a case.
Change-Id: I8e55f0798543b2de44b37da13f7770889e6fbe5f
Signed-off-by: Maya Erez <merez@codeaurora.org>
[subhashj@codeaurora.org: fixed merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
Add the necessary device tree properties and parsing in the driver
to support PM QoS voting for IRQ and CPU groups for CMDQ / legacy modes.
Change-Id: I1a94978ca66823d2ce78ee230cf36b4ebb72e6d8
Signed-off-by: Gilad Broner <gbroner@codeaurora.org>
[subhashj@codeaurora.org: fixed trivial merge conflicts]
Signed-off-by: Subhash Jadavani <subhashj@codeaurora.org>
pm_qos is causing race conditions in CQ mode with
power management. Removing the feature in order to
allow power management.
Change-Id: I340cd784829f389f18df6bff664337aca0f3c867
Signed-off-by: Dov Levenglick <dovl@codeaurora.org>
Signed-off-by: Konstantin Dorfman <kdorfman@codeaurora.org>
Any CMD/DAT errors raised using the error interrupt status in
CMDQ mode also needs to be cleared. If this is not cleared,
any error in supported CMDQ CMD's like CMD 44/45/46/47/48 will
cause an interrupt storm.
Change-Id: Ie725bbf1c859a2dc91b64274e05e71328dfeb1b2
Signed-off-by: Venkat Gopalakrishnan <venkatg@codeaurora.org>
