| Message ID | 20200921115424.4105-1-cezary.rojewski@intel.com |
|---|---|
| Headers | show |
| Series | ASoC: Intel: Catpt - Lynx and Wildcat point | expand |
On Mon, Sep 21, 2020 at 05:20:12PM +0300, Andy Shevchenko wrote: > On Mon, Sep 21, 2020 at 03:58:33PM +0200, Amadeusz Sławiński wrote: > > On 9/21/2020 2:59 PM, Andy Shevchenko wrote: > > > > +struct catpt_set_volume_input { > > > > + u32 channel; > > > > + u32 target_volume; > > > > + u64 curve_duration; > > > > + enum catpt_audio_curve_type curve_type __aligned(4); > > > > +} __packed; > > > How this __packed changes anything? In general __packed doesn't make sense for > > > in-kernel data structures. Otherwise you have to use proper (POD) types for > > > data. Ditto for all similar cases. > > > > All of __packed use in code is done on structures used to communicate with > > FW, which is binary interface, so it is not kernel only structure, as it is > > also FW one. While we can expect compiler to do the right thing, I consider > > it is better to be explicit about what kind of data we are handling, so > > there aren't any surprises. > > Size of enum is compiler defined. It may not be used in the ABIs. I have to elaborate that I'm talking in ABIs which implies different compilers and even may be run on different CPU architectures. > __uXX vs. uXX I dunno. And here I'm talking about FW <--> OS interface. It's not user visible ABI, but still some Ext <--> Int protocol. I saw uXX types in data structures of FW communication protocols.
On 2020-09-21 2:59 PM, Andy Shevchenko wrote: > On Mon, Sep 21, 2020 at 01:54:13PM +0200, Cezary Rojewski wrote: >> Declare global and stream IPC message handlers for all known message >> types. ... >> +int catpt_coredump(struct catpt_dev *cdev) >> +{ >> + struct catpt_dump_section_hdr *hdr; >> + size_t dump_size, regs_size; >> + u8 *dump, *pos; >> + int i, j; >> + >> + regs_size = CATPT_SHIM_REGS_SIZE; >> + regs_size += CATPT_DMA_COUNT * CATPT_DMA_REGS_SIZE; >> + regs_size += CATPT_SSP_COUNT * CATPT_SSP_REGS_SIZE; >> + dump_size = resource_size(&cdev->dram); >> + dump_size += resource_size(&cdev->iram); >> + dump_size += regs_size; > >> + dump_size += 4 * sizeof(*hdr) + 20; /* hdrs and fw hash */ > > Function is full of hard coded 20s. Can you provide descriptive macro? > Will declare CATPT_DUMP_HASH_SIZE instead of hardcodes, sure. >> + dump = vzalloc(dump_size); >> + if (!dump) >> + return -ENOMEM; >> + >> + pos = dump; >> + >> + hdr = (struct catpt_dump_section_hdr *)pos; >> + hdr->magic = CATPT_DUMP_MAGIC; >> + hdr->core_id = cdev->spec->core_id; >> + hdr->section_id = CATPT_DUMP_SECTION_ID_FILE; >> + hdr->size = dump_size - sizeof(*hdr); >> + pos += sizeof(*hdr); >> + >> + for (i = j = 0; i < FW_INFO_SIZE_MAX; i++) >> + if (cdev->ipc.config.fw_info[i] == ' ') >> + if (++j == 4) >> + break; > >> + for (j = ++i; j < FW_INFO_SIZE_MAX && j - i < 20; j++) { > > This should have static_assert() at the place where you define both constants > (2nd is mentioned above 20). > >> + if (cdev->ipc.config.fw_info[j] == ' ') >> + break; >> + *(pos + j - i) = cdev->ipc.config.fw_info[j]; >> + } >> + pos += 20; > > These two for-loops should have some comment to explain what's going on. > Actually, after poking my FW friends again I realized that it's just dumping 20chars from "hash" segment of fw_info (struct catpt_fw_ready, field: fw_info[]). So, this could be replaced by: /* navigate to fifth info segment (fw hash) */ for (i = j = 0; i < FW_INFO_SIZE_MAX; i++) /* info segments are separated by space each */ if (cdev->ipc.config.fw_info[i] == ' ') if (++j == 4) break; memcpy(pos, &cdev->ipc.config.fw_info[++i], CATPT_DUMP_HASH_SIZE); pos += CATPT_DUMP_HASH_SIZE; Existing for-loops were based on internal solution. Half of the code isn't needed afterall.. Czarek
On 2020-09-21 8:41 PM, Andy Shevchenko wrote:> On Mon, Sep 21, 2020 at 06:13:59PM +0000, Rojewski, Cezary wrote: >> On 2020-09-21 2:59 PM, Andy Shevchenko wrote: >>> On Mon, Sep 21, 2020 at 01:54:13PM +0200, Cezary Rojewski wrote: ... >>>> + for (i = j = 0; i < FW_INFO_SIZE_MAX; i++) >>>> + if (cdev->ipc.config.fw_info[i] == ' ') >>>> + if (++j == 4) >>>> + break; >>> >>>> + for (j = ++i; j < FW_INFO_SIZE_MAX && j - i < 20; j++) { >>> >>> This should have static_assert() at the place where you define both constants >>> (2nd is mentioned above 20). >>> >>>> + if (cdev->ipc.config.fw_info[j] == ' ') >>>> + break; >>>> + *(pos + j - i) = cdev->ipc.config.fw_info[j]; >>>> + } >>>> + pos += 20; >>> >>> These two for-loops should have some comment to explain what's going on. >>> >> >> Actually, after poking my FW friends again I realized that it's just >> dumping 20chars from "hash" segment of fw_info (struct catpt_fw_ready, >> field: fw_info[]). >> >> So, this could be replaced by: >> > >> /* navigate to fifth info segment (fw hash) */ >> for (i = j = 0; i < FW_INFO_SIZE_MAX; i++) >> /* info segments are separated by space each */ >> if (cdev->ipc.config.fw_info[i] == ' ') >> if (++j == 4) >> break; > > ...and this is repeating strnchr() / strnchrnul(). > Indeed, will incorporate into above. > With the questions "what if...": > - nul in the middle of this? > - less than 4 spaces found? > While this should never happen (means user is somehow not making use of officially released firmware binary), coredumps are useful only if you have access to debug tools. In cases you'd mentioned, invalid hash would have been dumped to coredump and crash reader simply wouldn't have been able to navigate to actual build for it. The rest of the coredump is still vital though. memcpy() could be gated behind an 'if' for safety if needed: info = cdev->ipc.config.fw_info; eof = info + FW_INFO_SIZE_MAX; /* navigate to fifth info segment (fw hash) */ for (i = 0; i < 4 && info < eof; i++, info++) /* info segments are separated by space each */ if ((info = strnchr(info, eof - info, ' ')) == NULL) break; if (i == 4 && info < eof) memcpy(pos, info, min(eof - info, CATPT_DUMP_HASH_SIZE)); Didn't compile this, some typecheck fixes might be in order and so on. >> memcpy(pos, &cdev->ipc.config.fw_info[++i], CATPT_DUMP_HASH_SIZE); >> pos += CATPT_DUMP_HASH_SIZE; >> >> >> Existing for-loops were based on internal solution. Half of the code >> isn't needed afterall.. >
On 2020-09-22 11:04 AM, Andy Shevchenko wrote: > On Mon, Sep 21, 2020 at 08:48:12PM +0000, Rojewski, Cezary wrote: >> On 2020-09-21 8:41 PM, Andy Shevchenko wrote: ... >> While this should never happen (means user is somehow not making use of >> officially released firmware binary), coredumps are useful only if you >> have access to debug tools. In cases you'd mentioned, invalid hash would >> have been dumped to coredump and crash reader simply wouldn't have been >> able to navigate to actual build for it. The rest of the coredump is still >> vital though. >> >> memcpy() could be gated behind an 'if' for safety if needed: >> >> info = cdev->ipc.config.fw_info; >> eof = info + FW_INFO_SIZE_MAX; >> /* navigate to fifth info segment (fw hash) */ >> for (i = 0; i < 4 && info < eof; i++, info++) >> /* info segments are separated by space each */ >> if ((info = strnchr(info, eof - info, ' ')) == NULL) >> break; > >> if (i == 4 && info < eof) >> memcpy(pos, info, min(eof - info, CATPT_DUMP_HASH_SIZE)); > > And here basically enough check is info against NULL, right? > Just try to look at different possibilities how to make code simpler and neater. > >> Didn't compile this, some typecheck fixes might be in order and so on. > What you meant is: if (i == 4 && !info) // instead of 'info < eof' right? If 4th space is last char in this string then info would end up being non-NULL and equal to 'eof' and thus memcpy() would get invoked with size=eof-info=0. catpt_coredump() is here to gather debug info for Intel folks to analyze in case of critical error. In ideal world, it should not be required at all as when we get here, there are bigger problems on our head. Above solution is simpler than what is prevent in v7 while also maintaining good readability - variable names - plus comments which you suggested. Czarek
On Tue, Sep 22, 2020 at 02:29:10PM +0300, Andy Shevchenko wrote: > On Tue, Sep 22, 2020 at 11:04:31AM +0000, Rojewski, Cezary wrote: > > On 2020-09-22 11:04 AM, Andy Shevchenko wrote: > > > On Mon, Sep 21, 2020 at 08:48:12PM +0000, Rojewski, Cezary wrote: ... > > > And here basically enough check is info against NULL, right? > > > Just try to look at different possibilities how to make code simpler and neater. > > > > > >> Didn't compile this, some typecheck fixes might be in order and so on. > > > > > > > What you meant is: > > if (i == 4 && !info) // instead of 'info < eof' > > > > right? > > Simply if (!info)... if (info) memcpy(); of course, otherwise it will crash. > > If 4th space is last char in this string then info would end up being > > non-NULL and equal to 'eof' and thus memcpy() would get invoked with > > size=eof-info=0. > > ...which is not a problem.
Implement support for Lynxpoint and Wildcat Point AudioDSP. Catpt solution deprecates existing sound/soc/intel/haswell which is removed in the following series. Due to high range of errors and desynchronization from recommendations set by Windows solution, re-write came as a lower-cost solution compared to refactoring /haswell/ with several series of patches. Series is dependent on linux-spi change: spi: pxa2xx: Add SSC2 and SSPSP2 SSP registers https://www.spinics.net/lists/linux-spi/msg23885.html which has been already merged and is now part of linux-spi tree. Bulk of series content is device driver core code - everything up to patch 7/14 - with fs entries and trace macros introduced right after. While each core patch is shaped in such a way that no unavailable members are ever called, until patch 14/14 is applied, no code compilation can occur as no Makefile is present. Once said patch is added, Makefile and Kconfig are implemented and driver module compiles as expected. Special thanks go to Marcin Barlik and Piotr Papierkowski for sharing their LPT/WPT AudioDSP architecture expertise as well as helping backtrack its historical background. My thanks go to Amadeusz Slawinski for reviews and improvements proposed on and off the internal list. Most of internal diff below is his contribution. Krzysztof Hejmowski helped me setup my own Xtensa environment and recompile LPT/WPT FW binary sources what sped up the development greatly. This would not have been possible without help from these champions, especially considering how quickly the catpt was written: 2 weeks features, 3 weeks optimizations. Thank you. Userspace-exposed members are compatible with what is exposed by deprecated solution as well as FW binary being re-used thus no harm is done. The only visible differences are: the newly added 'Loopback Mute' kcontrol and volume support extending to quad from stereo. On top of fixing erros and design flows, catpt also adds module reload, dynamic SRAM memory allocation during PCM runtime and exposes missing userspace API: 'Loopback Mute' kcontrol, quad volume controls and sysfs fw-version entries. Event tracing is provided to easy solution debugging. Following are not included in this update and are scheduled as later addition: - fw logging - module (library) support Note: LPT power up/down sequences might get aligned with WPT once enough testing is done as capabilities are shared for both DSPs. Note #2: Both LPT and WPT power up/down sequences may get optimized in future updates as thanks to help from the Windows team, most of nuances behind why/what/when in regard to hw registers have been backtracked and reviewed again. Link to developer's deep dive message: https://www.spinics.net/lists/alsa-devel/msg113563.html Changes in v7: - fixed licence header for fs.c - renamed fs.c to sysfs.c to better match its purpose - added documentation within Documentation/ABI/testing for entries exposed by catpt - bin_attribute fw_build replaced by attribute fw_info: fw_info contains full FW information and after successful handshake, it's always available (stored in driver data) so no need to invoke GET_FW_VERSION IPC again, just dump the stored information - rather than manually creating and removing sysfs files, now makes use of dev_groups member of struct device_driver - patch: 10/14 'ASoC: Intel: Select catpt and deprecate haswell' has been moved to the back of the list: enable catpt after machine boards have been prepared for it first - improved readability of several goto labels Changes in v6: https://www.spinics.net/lists/alsa-devel/msg115765.html - reordered and reorganized code for patches 1/13 - 8/13 of v5, so each patches makes use of no member or function which is unavailable to it. Series size increased from 13 to 14 patches: addition of base members e.g.: registers has been split from addition of device.c file which describes acpi device behavior Changes in v5: https://www.spinics.net/lists/alsa-devel/msg115621.html Basically everything below is result of Andy's review. Thank you Andy for taking time into this detailed review - catpt now makes use of common linux/pxa2xx_ssp.h header file, removing redundant SSP register declarations in the process. As stated in the opening, this is dependent upon linux-spi change: spi: pxa2xx: Add SSC2 and SSPSP2 SSP registers - updated Kconfig by removing DMADEVICES and adding COMPILE_TEST as optional depends-on - updated all register macros definitions to be more safe against common arithmetics when specifying macro's parameters - removed CONFIG_PM and CONFIG_PM_SLEEP usage in favor of __maybe_unused - all 'if (ret < 0)' converted to simple 'if (ret)' whenever possible - fixed erroneous check for platform_device_register_data within catpt_register_board() - _SLAVE/_MASTER replaced with more inclusive _CONSUMER/_PROVIDER for enum catpt_ssp_mode - catpt_acpi_probe() is now making use of high-level wrappers for ioremapping and resource assignment, reducing function's code size - due to improved catpt_acpi_probe() behavior, catpt_acpi_remove() needs not to cast dma_free_coherent() any longer - DMA source and destrination maxburst now of value 16, see: https://www.spinics.net/lists/alsa-devel/msg114394.html - simplified catpt_dsp_update_lpclock() as list_for_each_entry() is empty-safe by default - dropped '_SSP_' from all names of all CATPT_SSP_SSXXX_DEFAULT macros - catpt_updatel_pci now makes use of linux/pci.h and uapi/linux/pci.h constants such as: PCI_PM_CTRL_STATE_MASK and PCI_D3hot Changes in v4: https://www.spinics.net/lists/alsa-devel/msg113762.html - fixed compilation with i386 kconfig (conflicting names) - streamlined naming for SHIM and PCI registers to match SSP ones (SHIM_REG -> SHIM) - catpt_component_probe removed and kcontrols again initializzed statically via snd_kcontrol_new array: this is to remove kctl->id.device shenanigans - renamed catpt_set_ctlvol to catpt_set_dspvol - function name wasn't matching its purpose Changes in v3: - fixed IRAM mask usage in lpt_dsp_power_up (dsp.c) - updated dbg message formatting in catpt_restore_fwimage as suggested by Andy - fixed alignment for struct catpt_ssp_device_format - catpt_set_ctlvol now verifies all-equal scenario based on all channels rather than just first two as requested by Amadeo - fixed SPDX for registers.h Changes in v2: https://www.spinics.net/lists/alsa-devel/msg113660.html - fixed SPDX formatting for all header files as well as pcm.c - fixed size provided to memcpy() in fw_build_read() as reported by Mark - renamed struct catpt_pdata to struct catpt_spec (cosmetic) - fixed erroneous path in catpt_load_block: region is properly released - trace.h events for updating registers have been removed and usages replaced by dev_dbg (SRAMPGE/ LPCS) - as requested by Andy, struct resource has replaced struct catpt_mbank and struct catpt_mregion. This change cascaded into: - catpt_mbank_size and catpt_mregion_size replaced by resource_size - catpt_mregion_intersects replaced by resource_overlaps - all catpt_mbank_ and catpt_mregion_ handlers found in loader.c (_request, _reserve, _release, _extract, _split, _join) have been removed - __request_region and __release_region have been enlisted in their place - catpt_mregion_intersecting renamed to catpt_resource_overlapping - catpt_request_region helper has been provided to deal with -size based requests o haven't found direct replacements in resource.c/ ioport.h for both functions - catpt_mbank_create and catpt_mbank_remove renamed to catpt_sram_init and catpt_sram_free respectively - catpt_sram_init now returns void instead of int and has been converted to simple initialized. This change ultimately cascaded into: o both SRAM banks initialization being moved to catpt_dev_init from catpt_acpi_probe (device.c) o catpt_dev::spec is now initialized first, with catpt_dev_init following it soon after o catpt_acpi_probe erroneous path has been simplified as SRAM banks no longer need to be freed - catpt_sram_free now frees all resources via child -> sibling enumeration rather than region_list iteration - catpt_dsp_update_srampge and catpt_dsp_set_srampge now accept new argument: unsigned long mask. Caused by removal of catpt_mbank - mask is taken directly from catpt_dev::spec::d/iram_mask - trace.h events for catpt_mbank and catpt_mregion have been removed Diff against last drop on internal list: https://www.spinics.net/lists/alsa-devel/msg113549.html - replaced spinlock with mutex for mregion allocation and release to address sleeping in atomic context warnings - fixed coredump fw_hash dumping - kcontrol values are now always stored regardless of stream of interest is running or not - kcontrol values are now applied after stream is prepared instead of ignoring what has been set by user initially - catpt_pdata instances have been renamed from hsw_ and bdw_ to lpt_ and wpt_ respectively - reordered Makefile .o(s) (cosmetic) Cezary Rojewski (14): ASoC: Intel: Add catpt base members ASoC: Intel: catpt: Implement IPC protocol ASoC: Intel: catpt: Add IPC message handlers ASoC: Intel: catpt: Define DSP operations ASoC: Intel: catpt: Firmware loading and context restore ASoC: Intel: catpt: PCM operations ASoC: Intel: catpt: Device driver lifecycle ASoC: Intel: catpt: Event tracing ASoC: Intel: catpt: Simple sysfs attributes ASoC: Intel: haswell: Remove haswell-solution specific code ASoC: Intel: broadwell: Remove haswell-solution specific code ASoC: Intel: bdw-5650: Remove haswell-solution specific code ASoC: Intel: bdw-5677: Remove haswell-solution specific code ASoC: Intel: Select catpt and deprecate haswell .../ABI/testing/sysfs-bus-pci-devices-catpt | 16 + sound/soc/intel/Kconfig | 24 +- sound/soc/intel/Makefile | 2 +- sound/soc/intel/boards/Kconfig | 8 +- sound/soc/intel/boards/bdw-rt5650.c | 36 - sound/soc/intel/boards/bdw-rt5677.c | 33 - sound/soc/intel/boards/broadwell.c | 33 - sound/soc/intel/boards/haswell.c | 28 +- sound/soc/intel/catpt/Makefile | 6 + sound/soc/intel/catpt/core.h | 187 +++ sound/soc/intel/catpt/device.c | 348 +++++ sound/soc/intel/catpt/dsp.c | 571 ++++++++ sound/soc/intel/catpt/ipc.c | 298 ++++ sound/soc/intel/catpt/loader.c | 671 +++++++++ sound/soc/intel/catpt/messages.c | 313 +++++ sound/soc/intel/catpt/messages.h | 401 ++++++ sound/soc/intel/catpt/pcm.c | 1210 +++++++++++++++++ sound/soc/intel/catpt/registers.h | 178 +++ sound/soc/intel/catpt/sysfs.c | 58 + sound/soc/intel/catpt/trace.h | 83 ++ 20 files changed, 4361 insertions(+), 143 deletions(-) create mode 100644 Documentation/ABI/testing/sysfs-bus-pci-devices-catpt create mode 100644 sound/soc/intel/catpt/Makefile create mode 100644 sound/soc/intel/catpt/core.h create mode 100644 sound/soc/intel/catpt/device.c create mode 100644 sound/soc/intel/catpt/dsp.c create mode 100644 sound/soc/intel/catpt/ipc.c create mode 100644 sound/soc/intel/catpt/loader.c create mode 100644 sound/soc/intel/catpt/messages.c create mode 100644 sound/soc/intel/catpt/messages.h create mode 100644 sound/soc/intel/catpt/pcm.c create mode 100644 sound/soc/intel/catpt/registers.h create mode 100644 sound/soc/intel/catpt/sysfs.c create mode 100644 sound/soc/intel/catpt/trace.h