1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019302030213022302330243025302630273028302930303031303230333034303530363037303830393040304130423043304430453046304730483049305030513052305330543055305630573058305930603061306230633064306530663067306830693070307130723073307430753076307730783079308030813082308330843085308630873088308930903091309230933094309530963097309830993100310131023103310431053106310731083109311031113112311331143115311631173118311931203121312231233124312531263127312831293130313131323133313431353136313731383139314031413142314331443145314631473148314931503151315231533154315531563157315831593160316131623163316431653166316731683169317031713172317331743175317631773178317931803181318231833184318531863187318831893190319131923193319431953196319731983199320032013202320332043205320632073208320932103211321232133214321532163217321832193220322132223223322432253226322732283229323032313232323332343235323632373238 |
- /*
- *
- * (C) COPYRIGHT 2011-2016 ARM Limited. All rights reserved.
- *
- * This program is free software and is provided to you under the terms of the
- * GNU General Public License version 2 as published by the Free Software
- * Foundation, and any use by you of this program is subject to the terms
- * of such GNU licence.
- *
- * A copy of the licence is included with the program, and can also be obtained
- * from Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
- * Boston, MA 02110-1301, USA.
- *
- */
- /*
- * Job Scheduler Implementation
- */
- #include <mali_kbase.h>
- #include <mali_kbase_js.h>
- #if defined(CONFIG_MALI_GATOR_SUPPORT)
- #include <mali_kbase_gator.h>
- #endif
- #include <mali_kbase_tlstream.h>
- #include <mali_kbase_hw.h>
- #include <mali_kbase_defs.h>
- #include <mali_kbase_config_defaults.h>
- #include "mali_kbase_jm.h"
- #include "mali_kbase_hwaccess_jm.h"
- /*
- * Private types
- */
- /* Bitpattern indicating the result of releasing a context */
- enum {
- /* The context was descheduled - caller should try scheduling in a new
- * one to keep the runpool full */
- KBASEP_JS_RELEASE_RESULT_WAS_DESCHEDULED = (1u << 0),
- /* Ctx attributes were changed - caller should try scheduling all
- * contexts */
- KBASEP_JS_RELEASE_RESULT_SCHED_ALL = (1u << 1)
- };
- typedef u32 kbasep_js_release_result;
- const int kbasep_js_atom_priority_to_relative[BASE_JD_NR_PRIO_LEVELS] = {
- KBASE_JS_ATOM_SCHED_PRIO_MED, /* BASE_JD_PRIO_MEDIUM */
- KBASE_JS_ATOM_SCHED_PRIO_HIGH, /* BASE_JD_PRIO_HIGH */
- KBASE_JS_ATOM_SCHED_PRIO_LOW /* BASE_JD_PRIO_LOW */
- };
- const base_jd_prio
- kbasep_js_relative_priority_to_atom[KBASE_JS_ATOM_SCHED_PRIO_COUNT] = {
- BASE_JD_PRIO_HIGH, /* KBASE_JS_ATOM_SCHED_PRIO_HIGH */
- BASE_JD_PRIO_MEDIUM, /* KBASE_JS_ATOM_SCHED_PRIO_MED */
- BASE_JD_PRIO_LOW /* KBASE_JS_ATOM_SCHED_PRIO_LOW */
- };
- /*
- * Private function prototypes
- */
- static kbasep_js_release_result kbasep_js_runpool_release_ctx_internal(
- struct kbase_device *kbdev, struct kbase_context *kctx,
- struct kbasep_js_atom_retained_state *katom_retained_state);
- static int kbase_js_get_slot(struct kbase_device *kbdev,
- struct kbase_jd_atom *katom);
- static void kbase_js_foreach_ctx_job(struct kbase_context *kctx,
- kbasep_js_policy_ctx_job_cb callback);
- /* Helper for trace subcodes */
- #if KBASE_TRACE_ENABLE
- static int kbasep_js_trace_get_refcnt(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- unsigned long flags;
- struct kbasep_js_device_data *js_devdata;
- int as_nr;
- int refcnt = 0;
- js_devdata = &kbdev->js_data;
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- as_nr = kctx->as_nr;
- if (as_nr != KBASEP_AS_NR_INVALID) {
- struct kbasep_js_per_as_data *js_per_as_data;
- js_per_as_data = &js_devdata->runpool_irq.per_as_data[as_nr];
- refcnt = js_per_as_data->as_busy_refcount;
- }
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- return refcnt;
- }
- static int kbasep_js_trace_get_refcnt_nolock(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- struct kbasep_js_device_data *js_devdata;
- int as_nr;
- int refcnt = 0;
- js_devdata = &kbdev->js_data;
- as_nr = kctx->as_nr;
- if (as_nr != KBASEP_AS_NR_INVALID) {
- struct kbasep_js_per_as_data *js_per_as_data;
- js_per_as_data = &js_devdata->runpool_irq.per_as_data[as_nr];
- refcnt = js_per_as_data->as_busy_refcount;
- }
- return refcnt;
- }
- #else /* KBASE_TRACE_ENABLE */
- static int kbasep_js_trace_get_refcnt(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- CSTD_UNUSED(kbdev);
- CSTD_UNUSED(kctx);
- return 0;
- }
- static int kbasep_js_trace_get_refcnt_nolock(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- CSTD_UNUSED(kbdev);
- CSTD_UNUSED(kctx);
- return 0;
- }
- #endif /* KBASE_TRACE_ENABLE */
- /*
- * Private types
- */
- enum {
- JS_DEVDATA_INIT_NONE = 0,
- JS_DEVDATA_INIT_CONSTANTS = (1 << 0),
- JS_DEVDATA_INIT_POLICY = (1 << 1),
- JS_DEVDATA_INIT_ALL = ((1 << 2) - 1)
- };
- enum {
- JS_KCTX_INIT_NONE = 0,
- JS_KCTX_INIT_CONSTANTS = (1 << 0),
- JS_KCTX_INIT_POLICY = (1 << 1),
- JS_KCTX_INIT_ALL = ((1 << 2) - 1)
- };
- /*
- * Private functions
- */
- /**
- * core_reqs_from_jsn_features - Convert JSn_FEATURES to core requirements
- * @features: JSn_FEATURE register value
- *
- * Given a JSn_FEATURE register value returns the core requirements that match
- *
- * Return: Core requirement bit mask
- */
- static base_jd_core_req core_reqs_from_jsn_features(u16 features)
- {
- base_jd_core_req core_req = 0u;
- if ((features & JS_FEATURE_SET_VALUE_JOB) != 0)
- core_req |= BASE_JD_REQ_V;
- if ((features & JS_FEATURE_CACHE_FLUSH_JOB) != 0)
- core_req |= BASE_JD_REQ_CF;
- if ((features & JS_FEATURE_COMPUTE_JOB) != 0)
- core_req |= BASE_JD_REQ_CS;
- if ((features & JS_FEATURE_TILER_JOB) != 0)
- core_req |= BASE_JD_REQ_T;
- if ((features & JS_FEATURE_FRAGMENT_JOB) != 0)
- core_req |= BASE_JD_REQ_FS;
- return core_req;
- }
- static void kbase_js_sync_timers(struct kbase_device *kbdev)
- {
- mutex_lock(&kbdev->js_data.runpool_mutex);
- kbase_backend_ctx_count_changed(kbdev);
- mutex_unlock(&kbdev->js_data.runpool_mutex);
- }
- /* Hold the kbasep_js_device_data::runpool_irq::lock for this */
- bool kbasep_js_runpool_retain_ctx_nolock(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- struct kbasep_js_device_data *js_devdata;
- struct kbasep_js_per_as_data *js_per_as_data;
- bool result = false;
- int as_nr;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(kctx != NULL);
- js_devdata = &kbdev->js_data;
- as_nr = kctx->as_nr;
- if (as_nr != KBASEP_AS_NR_INVALID) {
- int new_refcnt;
- KBASE_DEBUG_ASSERT(as_nr >= 0);
- js_per_as_data = &js_devdata->runpool_irq.per_as_data[as_nr];
- KBASE_DEBUG_ASSERT(js_per_as_data->kctx != NULL);
- new_refcnt = ++(js_per_as_data->as_busy_refcount);
- KBASE_TRACE_ADD_REFCOUNT(kbdev, JS_RETAIN_CTX_NOLOCK, kctx,
- NULL, 0u, new_refcnt);
- result = true;
- }
- return result;
- }
- /* Helper macros to access and modify jsctx_queue.indicies */
- #define JSCTX_GET(offset, var, mask) \
- ((var >> offset) & mask)
- /* This wraps around to correct integer size automatically. */
- #define JSCTX_SET(var, offset, value, mask) \
- (var = ((var & ~(mask << offset)) /*Clear old bits */ \
- | (((value) & mask) << offset))) /* Set (after masking) new bits */
- #define JSCTX_GET_WR_IDX(var) \
- JSCTX_GET(JSCTX_WR_OFFSET, var, JSCTX_RB_MASK_STORE)
- #define JSCTX_GET_RN_IDX(var) \
- JSCTX_GET(JSCTX_RN_OFFSET, var, JSCTX_RB_MASK_STORE)
- #define JSCTX_GET_RD_IDX(var) \
- JSCTX_GET(JSCTX_RD_OFFSET, var, JSCTX_RB_MASK_STORE)
- #define JSCTX_GET_IDX_DIFF(lower, upper) \
- ((upper >= lower) ? (upper - lower) : (upper+JSCTX_RB_SIZE_STORE-lower))
- #define JSCTX_SET_WR_IDX(var, value) \
- JSCTX_SET(var, JSCTX_WR_OFFSET, value, JSCTX_RB_MASK_STORE)
- #define JSCTX_SET_RN_IDX(var, value) \
- JSCTX_SET(var, JSCTX_RN_OFFSET, value, JSCTX_RB_MASK_STORE)
- #define JSCTX_SET_RD_IDX(var, value) \
- JSCTX_SET(var, JSCTX_RD_OFFSET, value, JSCTX_RB_MASK_STORE)
- /**
- * jsctx_rb_none_to_pull_prio(): - Check if there are no pullable atoms
- * @kctx: Pointer to kbase context with ring buffer.
- * @js: Job slot id to check.
- * @prio: Priority to check.
- *
- * Return true if there are no atoms to pull. There may be running atoms in the
- * ring buffer even if there are no atoms to pull. It is also possible for the
- * ring buffer to be full (with running atoms) when this functions returns
- * true.
- *
- * Return: true if there are no atoms to pull, false otherwise.
- */
- static inline bool
- jsctx_rb_none_to_pull_prio(struct kbase_context *kctx, int js, int prio)
- {
- struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
- unsigned int var = atomic_read(&rb->indicies);
- return JSCTX_GET_RD_IDX(var) == JSCTX_GET_WR_IDX(var);
- }
- /**
- * jsctx_rb_none_to_pull(): - Check if all priority ring buffers have no
- * pullable atoms
- * @kctx: Pointer to kbase context with ring buffer.
- * @js: Job slot id to check.
- *
- * Caller must hold runpool_irq.lock
- *
- * Return: true if the ring buffers for all priorities have no pullable atoms,
- * false otherwise.
- */
- static inline bool
- jsctx_rb_none_to_pull(struct kbase_context *kctx, int js)
- {
- int prio;
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- for (prio = 0; prio < KBASE_JS_ATOM_SCHED_PRIO_COUNT; prio++) {
- if (!jsctx_rb_none_to_pull_prio(kctx, js, prio))
- return false;
- }
- return true;
- }
- /**
- * jsctx_rb_is_full(): - Check if the given ringbuffer is full.
- * @queue: Pointer to the queue containing the ringbuffer.
- *
- * No locks explicitly required, result will always be consistent.
- * But depending on usage, the caller should consider jctx.lock,
- * for the result to remain correct.
- *
- * Return: true if the ringbuffer is full, false otherwise.
- */
- static inline bool
- jsctx_rb_is_full(struct jsctx_queue *queue)
- {
- unsigned int var = atomic_read(&queue->indicies);
- u16 rn_idx = JSCTX_GET_RN_IDX(var);
- u16 wr_idx = JSCTX_GET_WR_IDX(var);
- return JSCTX_GET_IDX_DIFF(rn_idx, wr_idx) >= JSCTX_RB_SIZE;
- }
- /**
- * jsctx_queue_foreach_prio(): - Execute callback for each entry in the queue.
- * @kctx: Pointer to kbase context with the queue.
- * @js: Job slot id to iterate.
- * @prio: Priority id to iterate.
- * @callback: Function pointer to callback.
- *
- * Iterate over a ring buffer and invoke @callback for each entry in buffer, and
- * remove the entry from the buffer.
- *
- * If entries are added to the ring buffer while this is running those entries
- * may, or may not be covered. To ensure that all entries in the buffer have
- * been enumerated when this function returns jsctx->lock must be held when
- * calling this function.
- *
- * The HW access lock, js_data.runpool_irq.lock, must always be held when
- * calling this function.
- */
- static void
- jsctx_queue_foreach_prio(struct kbase_context *kctx, int js, int prio,
- kbasep_js_policy_ctx_job_cb callback)
- {
- struct jsctx_queue *queue = &kctx->jsctx_queue[prio][js];
- struct kbase_jd_atom *katom;
- struct list_head *pos, *q;
- unsigned int var = atomic_read(&queue->indicies);
- u16 running_idx = JSCTX_GET_RN_IDX(var);
- u16 read_idx = JSCTX_GET_RD_IDX(var);
- u16 wr_idx = JSCTX_GET_WR_IDX(var);
- u16 i;
- const u16 count = JSCTX_GET_IDX_DIFF(running_idx, wr_idx);
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- /* There must be no jobs currently in HW access */
- WARN_ON(read_idx != JSCTX_GET_RN_IDX(var));
- /* Invoke callback on all kbase_jd_atoms in the ring buffer, and
- * removes them from the buffer */
- for (i = 0; i < count; i++) {
- int id = queue->entries[read_idx & JSCTX_RB_MASK].atom_id;
- katom = kbase_jd_atom_from_id(kctx, id);
- read_idx++;
- callback(kctx->kbdev, katom);
- }
- atomic_set(&queue->indicies, 0);
- list_for_each_safe(pos, q, &queue->queue_head) {
- struct kbase_jd_atom *entry;
- entry = list_entry(pos, struct kbase_jd_atom, queue);
- list_del(pos);
- callback(kctx->kbdev, entry);
- }
- }
- /**
- * jsctx_queue_foreach(): - Execute callback for each entry in every queue
- * @kctx: Pointer to kbase context with queue.
- * @js: Job slot id to iterate.
- * @callback: Function pointer to callback.
- *
- * Iterate over all the different priorities, and for each call
- * jsctx_queue_foreach_prio() to iterate over the queue and invoke @callback
- * for each entry, and remove the entry from the queue.
- */
- static inline void
- jsctx_queue_foreach(struct kbase_context *kctx, int js,
- kbasep_js_policy_ctx_job_cb callback)
- {
- int prio;
- for (prio = 0; prio < KBASE_JS_ATOM_SCHED_PRIO_COUNT; prio++)
- jsctx_queue_foreach_prio(kctx, js, prio, callback);
- }
- /**
- * jsctx_rb_peek_prio(): - Check buffer and get next atom
- * @kctx: Pointer to kbase context with ring buffer.
- * @js: Job slot id to check.
- * @prio: Priority id to check.
- *
- * Check the ring buffer for the specified @js and @prio and return a pointer to
- * the next atom, unless the ring buffer is empty.
- *
- * Return: Pointer to next atom in buffer, or NULL if there is no atom.
- */
- static inline struct kbase_jd_atom *
- jsctx_rb_peek_prio(struct kbase_context *kctx, int js, int prio)
- {
- struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
- int id;
- unsigned int var = atomic_read(&rb->indicies);
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- if (JSCTX_GET_RD_IDX(var) == JSCTX_GET_WR_IDX(var))
- return NULL;
- id = rb->entries[JSCTX_GET_RD_IDX(var) & JSCTX_RB_MASK].atom_id;
- return kbase_jd_atom_from_id(kctx, id);
- }
- /**
- * jsctx_rb_peek(): - Check all priority buffers and get next atom
- * @kctx: Pointer to kbase context with ring buffer.
- * @js: Job slot id to check.
- *
- * Check the ring buffers for all priorities, starting from
- * KBASE_JS_ATOM_SCHED_PRIO_HIGH, for the specified @js and @prio and return a
- * pointer to the next atom, unless all the priority's ring buffers are empty.
- *
- * Caller must hold the runpool_irq.lock.
- *
- * Return: Pointer to next atom in buffer, or NULL if there is no atom.
- */
- static inline struct kbase_jd_atom *
- jsctx_rb_peek(struct kbase_context *kctx, int js)
- {
- int prio;
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- for (prio = 0; prio < KBASE_JS_ATOM_SCHED_PRIO_COUNT; prio++) {
- struct kbase_jd_atom *katom;
- katom = jsctx_rb_peek_prio(kctx, js, prio);
- if (katom)
- return katom;
- }
- return NULL;
- }
- /**
- * jsctx_rb_peek_last(): - Check a ring buffer and get the last atom
- * @kctx: Pointer to kbase context with ring buffer.
- * @js: Job slot id to check.
- * @prio: Priority id to check.
- *
- * Check the ring buffer for the specified @js and @prio and return a
- * pointer to the last atom, unless all the priority's ring buffers are empty.
- *
- * The last atom is the atom that was added using jsctx_rb_add() most recently.
- *
- * Return: Pointer to last atom in buffer, or NULL if there is no atom.
- */
- static inline struct kbase_jd_atom *
- jsctx_rb_peek_last(struct kbase_context *kctx, int js, int prio)
- {
- struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
- unsigned int var = atomic_read(&rb->indicies);
- int id;
- lockdep_assert_held(&kctx->jctx.lock);
- if (!list_empty(&rb->queue_head)) {
- return list_entry(rb->queue_head.prev,
- struct kbase_jd_atom, queue);
- }
- if (JSCTX_GET_RN_IDX(var) == JSCTX_GET_WR_IDX(var))
- return NULL;
- id = rb->entries[(JSCTX_GET_WR_IDX(var) - 1) & JSCTX_RB_MASK].atom_id;
- return kbase_jd_atom_from_id(kctx, id);
- }
- /**
- * jsctx_rb_pull(): - Mark atom in list as running
- * @kctx: Pointer to kbase context with ring buffer.
- * @katom: Pointer to katom to pull.
- *
- * Mark an atom previously obtained from jsctx_rb_peek() as running.
- *
- * @katom must currently be at the head of the ring buffer.
- */
- static inline void
- jsctx_rb_pull(struct kbase_context *kctx, struct kbase_jd_atom *katom)
- {
- int prio = katom->sched_priority;
- int js = katom->slot_nr;
- unsigned int oldvar, var;
- struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- /* Atoms must be pulled in the correct order. */
- WARN_ON(katom != jsctx_rb_peek_prio(kctx, js, prio));
- do {
- u16 rd_idx;
- oldvar = atomic_read(&rb->indicies);
- var = oldvar;
- rd_idx = JSCTX_GET_RD_IDX(var);
- JSCTX_SET_RD_IDX(var, rd_idx+1);
- } while (atomic_cmpxchg(&rb->indicies, oldvar, var) != oldvar);
- }
- /**
- * jsctx_rb_unpull(): - Undo marking of atom in list as running
- * @kctx: Pointer to kbase context with ring buffer.
- * @katom: Pointer to katom to unpull.
- *
- * Undo jsctx_rb_pull() and put @katom back in the queue.
- *
- * jsctx_rb_unpull() must be called on atoms in the same order the atoms were
- * pulled.
- */
- static inline void
- jsctx_rb_unpull(struct kbase_context *kctx, struct kbase_jd_atom *katom)
- {
- int prio = katom->sched_priority;
- int js = katom->slot_nr;
- unsigned int oldvar, var;
- struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- do {
- u16 rd_idx;
- oldvar = atomic_read(&rb->indicies);
- var = oldvar;
- rd_idx = JSCTX_GET_RD_IDX(var)-1;
- /* Atoms must be unpulled in correct order. */
- WARN_ON(rb->entries[rd_idx & JSCTX_RB_MASK].atom_id !=
- kbase_jd_atom_id(kctx, katom));
- JSCTX_SET_RD_IDX(var, rd_idx);
- } while (atomic_cmpxchg(&rb->indicies, oldvar, var) != oldvar);
- }
- /**
- * jsctx_rb_add(): - Add atom to ring buffer
- * @kctx: Pointer to kbase context with ring buffer.
- * @katom: Pointer to katom to add.
- *
- * Add @katom to the ring buffer determined by the atom's priority and job slot
- * number.
- *
- * If the ring buffer is full -EBUSY will be returned.
- *
- * Return: On success 0 is returned, on failure a negative error code.
- */
- static int
- jsctx_rb_add_atom(struct kbase_context *kctx, struct kbase_jd_atom *katom)
- {
- int prio = katom->sched_priority;
- int js = katom->slot_nr;
- struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
- unsigned int oldvar, var;
- u16 wr_idx, running_idx, count;
- lockdep_assert_held(&kctx->jctx.lock);
- oldvar = atomic_read(&rb->indicies);
- var = oldvar;
- running_idx = JSCTX_GET_RN_IDX(var);
- wr_idx = JSCTX_GET_WR_IDX(var);
- count = JSCTX_GET_IDX_DIFF(running_idx, wr_idx);
- /* Check if the ring buffer is full */
- if (count >= JSCTX_RB_SIZE)
- return -EBUSY;
- rb->entries[wr_idx & JSCTX_RB_MASK].atom_id =
- kbase_jd_atom_id(kctx, katom);
- wr_idx++;
- JSCTX_SET_WR_IDX(var, wr_idx);
- while (atomic_cmpxchg(&rb->indicies, oldvar, var) != oldvar) {
- oldvar = atomic_read(&rb->indicies);
- var = oldvar;
- wr_idx = JSCTX_GET_WR_IDX(var)+1;
- JSCTX_SET_WR_IDX(var, wr_idx);
- }
- return 0;
- }
- /**
- * jsctx_rb_remove(): - Remove atom from ring buffer
- * @kctx: Pointer to kbase context with ring buffer.
- * @katom: Pointer to katom to remove.
- *
- * Remove @katom from the ring buffer.
- *
- * @katom must have been pulled from the buffer earlier by jsctx_rb_pull(), and
- * atoms must be removed in the same order they were pulled from the ring
- * buffer.
- */
- static inline void
- jsctx_rb_remove(struct kbase_context *kctx, struct kbase_jd_atom *katom)
- {
- int prio = katom->sched_priority;
- int js = katom->slot_nr;
- unsigned int oldvar, var;
- struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
- lockdep_assert_held(&kctx->jctx.lock);
- do {
- unsigned int rn_idx;
- oldvar = atomic_read(&rb->indicies);
- var = oldvar;
- rn_idx = JSCTX_GET_RN_IDX(var);
- JSCTX_SET_RN_IDX(var, rn_idx+1);
- } while (atomic_cmpxchg(&rb->indicies, oldvar, var) != oldvar);
- }
- static void
- jsctx_ll_add(struct kbase_context *kctx, struct kbase_jd_atom *katom)
- {
- int prio = katom->sched_priority;
- int js = katom->slot_nr;
- struct jsctx_queue *queue = &kctx->jsctx_queue[prio][js];
- lockdep_assert_held(&kctx->jctx.lock);
- list_add_tail(&katom->queue, &queue->queue_head);
- }
- static bool kbase_js_ctx_pullable(struct kbase_context *kctx,
- int js,
- bool is_scheduled);
- static bool kbase_js_ctx_list_add_pullable(struct kbase_device *kbdev,
- struct kbase_context *kctx,
- int js);
- static bool kbase_js_ctx_list_add_unpullable(struct kbase_device *kbdev,
- struct kbase_context *kctx,
- int js);
- void
- jsctx_ll_flush_to_rb(struct kbase_context *kctx, int prio, int js)
- {
- unsigned long flags;
- struct list_head *pos, *q;
- struct jsctx_queue *queue = &kctx->jsctx_queue[prio][js];
- bool flushed_any = false;
- struct kbasep_js_device_data *js_devdata = &kctx->kbdev->js_data;
- bool enqueue_required = false;
- lockdep_assert_held(&kctx->jctx.lock);
- /* Early out for common case */
- if (list_empty(&queue->queue_head) || jsctx_rb_is_full(queue))
- return;
- mutex_lock(&js_devdata->queue_mutex);
- mutex_lock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- mutex_lock(&js_devdata->runpool_mutex);
- spin_lock_irqsave(&kctx->kbdev->js_data.runpool_irq.lock, flags);
- /* If slot will transition from unpullable to pullable then add to
- * pullable list */
- if (jsctx_rb_none_to_pull(kctx, js))
- enqueue_required = true;
- else
- enqueue_required = false;
- list_for_each_safe(pos, q, &queue->queue_head) {
- struct kbase_jd_atom *katom;
- katom = list_entry(pos, struct kbase_jd_atom, queue);
- KBASE_DEBUG_ASSERT(katom);
- if (jsctx_rb_add_atom(kctx, katom))
- break;
- katom->atom_flags &= ~KBASE_KATOM_FLAG_JSCTX_IN_LL;
- katom->atom_flags |= KBASE_KATOM_FLAG_JSCTX_RB_SUBMITTED;
- flushed_any = true;
- list_del(pos);
- }
- if (flushed_any) {
- bool timer_sync = false;
- if (enqueue_required) {
- if (kbase_js_ctx_pullable(kctx, js, false))
- timer_sync = kbase_js_ctx_list_add_pullable(
- kctx->kbdev, kctx, js);
- else
- timer_sync = kbase_js_ctx_list_add_unpullable(
- kctx->kbdev, kctx, js);
- /* If this context is active and the atom is the first
- * on its slot, kick the job manager to attempt to
- * fast-start the atom */
- if (kctx == kctx->kbdev->hwaccess.active_kctx)
- kbase_jm_try_kick(kctx->kbdev, 1 << js);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock,
- flags);
- if (timer_sync)
- kbase_backend_ctx_count_changed(kctx->kbdev);
- } else {
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock,
- flags);
- }
- } else {
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- }
- mutex_unlock(&js_devdata->runpool_mutex);
- mutex_unlock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- mutex_unlock(&js_devdata->queue_mutex);
- }
- /*
- * Functions private to KBase ('Protected' functions)
- */
- int kbasep_js_devdata_init(struct kbase_device * const kbdev)
- {
- struct kbasep_js_device_data *jsdd;
- int err;
- int i;
- u16 as_present;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- jsdd = &kbdev->js_data;
- KBASE_DEBUG_ASSERT(jsdd->init_status == JS_DEVDATA_INIT_NONE);
- /* These two must be recalculated if nr_hw_address_spaces changes
- * (e.g. for HW workarounds) */
- as_present = (1U << kbdev->nr_hw_address_spaces) - 1;
- kbdev->nr_user_address_spaces = kbdev->nr_hw_address_spaces;
- if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8987)) {
- bool use_workaround;
- use_workaround = DEFAULT_SECURE_BUT_LOSS_OF_PERFORMANCE;
- if (use_workaround) {
- dev_dbg(kbdev->dev, "GPU has HW ISSUE 8987, and driver configured for security workaround: 1 address space only");
- kbdev->nr_user_address_spaces = 1;
- }
- }
- #ifdef CONFIG_MALI_DEBUG
- /* Soft-stop will be disabled on a single context by default unless
- * softstop_always is set */
- jsdd->softstop_always = false;
- #endif /* CONFIG_MALI_DEBUG */
- jsdd->nr_all_contexts_running = 0;
- jsdd->nr_user_contexts_running = 0;
- jsdd->nr_contexts_pullable = 0;
- atomic_set(&jsdd->nr_contexts_runnable, 0);
- /* All ASs initially free */
- jsdd->as_free = as_present;
- /* No ctx allowed to submit */
- jsdd->runpool_irq.submit_allowed = 0u;
- memset(jsdd->runpool_irq.ctx_attr_ref_count, 0,
- sizeof(jsdd->runpool_irq.ctx_attr_ref_count));
- memset(jsdd->runpool_irq.slot_affinities, 0,
- sizeof(jsdd->runpool_irq.slot_affinities));
- memset(jsdd->runpool_irq.slot_affinity_refcount, 0,
- sizeof(jsdd->runpool_irq.slot_affinity_refcount));
- INIT_LIST_HEAD(&jsdd->suspended_soft_jobs_list);
- /* Config attributes */
- jsdd->scheduling_period_ns = DEFAULT_JS_SCHEDULING_PERIOD_NS;
- jsdd->soft_stop_ticks = DEFAULT_JS_SOFT_STOP_TICKS;
- jsdd->soft_stop_ticks_cl = DEFAULT_JS_SOFT_STOP_TICKS_CL;
- if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8408))
- jsdd->hard_stop_ticks_ss = DEFAULT_JS_HARD_STOP_TICKS_SS_8408;
- else
- jsdd->hard_stop_ticks_ss = DEFAULT_JS_HARD_STOP_TICKS_SS;
- jsdd->hard_stop_ticks_cl = DEFAULT_JS_HARD_STOP_TICKS_CL;
- jsdd->hard_stop_ticks_dumping = DEFAULT_JS_HARD_STOP_TICKS_DUMPING;
- if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8408))
- jsdd->gpu_reset_ticks_ss = DEFAULT_JS_RESET_TICKS_SS_8408;
- else
- jsdd->gpu_reset_ticks_ss = DEFAULT_JS_RESET_TICKS_SS;
- jsdd->gpu_reset_ticks_cl = DEFAULT_JS_RESET_TICKS_CL;
- jsdd->gpu_reset_ticks_dumping = DEFAULT_JS_RESET_TICKS_DUMPING;
- jsdd->ctx_timeslice_ns = DEFAULT_JS_CTX_TIMESLICE_NS;
- jsdd->cfs_ctx_runtime_init_slices =
- DEFAULT_JS_CFS_CTX_RUNTIME_INIT_SLICES;
- jsdd->cfs_ctx_runtime_min_slices =
- DEFAULT_JS_CFS_CTX_RUNTIME_MIN_SLICES;
- atomic_set(&jsdd->soft_event_timeout_ms, DEFAULT_JS_SOFT_EVENT_TIMEOUT);
- dev_dbg(kbdev->dev, "JS Config Attribs: ");
- dev_dbg(kbdev->dev, "\tscheduling_period_ns:%u",
- jsdd->scheduling_period_ns);
- dev_dbg(kbdev->dev, "\tsoft_stop_ticks:%u",
- jsdd->soft_stop_ticks);
- dev_dbg(kbdev->dev, "\tsoft_stop_ticks_cl:%u",
- jsdd->soft_stop_ticks_cl);
- dev_dbg(kbdev->dev, "\thard_stop_ticks_ss:%u",
- jsdd->hard_stop_ticks_ss);
- dev_dbg(kbdev->dev, "\thard_stop_ticks_cl:%u",
- jsdd->hard_stop_ticks_cl);
- dev_dbg(kbdev->dev, "\thard_stop_ticks_dumping:%u",
- jsdd->hard_stop_ticks_dumping);
- dev_dbg(kbdev->dev, "\tgpu_reset_ticks_ss:%u",
- jsdd->gpu_reset_ticks_ss);
- dev_dbg(kbdev->dev, "\tgpu_reset_ticks_cl:%u",
- jsdd->gpu_reset_ticks_cl);
- dev_dbg(kbdev->dev, "\tgpu_reset_ticks_dumping:%u",
- jsdd->gpu_reset_ticks_dumping);
- dev_dbg(kbdev->dev, "\tctx_timeslice_ns:%u",
- jsdd->ctx_timeslice_ns);
- dev_dbg(kbdev->dev, "\tcfs_ctx_runtime_init_slices:%u",
- jsdd->cfs_ctx_runtime_init_slices);
- dev_dbg(kbdev->dev, "\tcfs_ctx_runtime_min_slices:%u",
- jsdd->cfs_ctx_runtime_min_slices);
- dev_dbg(kbdev->dev, "\tsoft_event_timeout:%i",
- atomic_read(&jsdd->soft_event_timeout_ms));
- if (!(jsdd->soft_stop_ticks < jsdd->hard_stop_ticks_ss &&
- jsdd->hard_stop_ticks_ss < jsdd->gpu_reset_ticks_ss &&
- jsdd->soft_stop_ticks < jsdd->hard_stop_ticks_dumping &&
- jsdd->hard_stop_ticks_dumping <
- jsdd->gpu_reset_ticks_dumping)) {
- dev_err(kbdev->dev, "Job scheduler timeouts invalid; soft/hard/reset tick counts should be in increasing order\n");
- return -EINVAL;
- }
- #if KBASE_DISABLE_SCHEDULING_SOFT_STOPS
- dev_dbg(kbdev->dev, "Job Scheduling Policy Soft-stops disabled, ignoring value for soft_stop_ticks==%u at %uns per tick. Other soft-stops may still occur.",
- jsdd->soft_stop_ticks,
- jsdd->scheduling_period_ns);
- #endif
- #if KBASE_DISABLE_SCHEDULING_HARD_STOPS
- dev_dbg(kbdev->dev, "Job Scheduling Policy Hard-stops disabled, ignoring values for hard_stop_ticks_ss==%d and hard_stop_ticks_dumping==%u at %uns per tick. Other hard-stops may still occur.",
- jsdd->hard_stop_ticks_ss,
- jsdd->hard_stop_ticks_dumping,
- jsdd->scheduling_period_ns);
- #endif
- #if KBASE_DISABLE_SCHEDULING_SOFT_STOPS && KBASE_DISABLE_SCHEDULING_HARD_STOPS
- dev_dbg(kbdev->dev, "Note: The JS policy's tick timer (if coded) will still be run, but do nothing.");
- #endif
- /* setup the number of irq throttle cycles base on given time */
- {
- int time_us = kbdev->gpu_props.irq_throttle_time_us;
- int cycles = kbasep_js_convert_us_to_gpu_ticks_max_freq(kbdev,
- time_us);
- atomic_set(&kbdev->irq_throttle_cycles, cycles);
- }
- /* Clear the AS data, including setting NULL pointers */
- memset(&jsdd->runpool_irq.per_as_data[0], 0,
- sizeof(jsdd->runpool_irq.per_as_data));
- for (i = 0; i < kbdev->gpu_props.num_job_slots; ++i)
- jsdd->js_reqs[i] = core_reqs_from_jsn_features(
- kbdev->gpu_props.props.raw_props.js_features[i]);
- jsdd->init_status |= JS_DEVDATA_INIT_CONSTANTS;
- /* On error, we could continue on: providing none of the below resources
- * rely on the ones above */
- mutex_init(&jsdd->runpool_mutex);
- mutex_init(&jsdd->queue_mutex);
- spin_lock_init(&jsdd->runpool_irq.lock);
- sema_init(&jsdd->schedule_sem, 1);
- err = kbasep_js_policy_init(kbdev);
- if (!err)
- jsdd->init_status |= JS_DEVDATA_INIT_POLICY;
- for (i = 0; i < kbdev->gpu_props.num_job_slots; ++i) {
- INIT_LIST_HEAD(&jsdd->ctx_list_pullable[i]);
- INIT_LIST_HEAD(&jsdd->ctx_list_unpullable[i]);
- }
- /* On error, do no cleanup; this will be handled by the caller(s), since
- * we've designed this resource to be safe to terminate on init-fail */
- if (jsdd->init_status != JS_DEVDATA_INIT_ALL)
- return -EINVAL;
- return 0;
- }
- void kbasep_js_devdata_halt(struct kbase_device *kbdev)
- {
- CSTD_UNUSED(kbdev);
- }
- void kbasep_js_devdata_term(struct kbase_device *kbdev)
- {
- struct kbasep_js_device_data *js_devdata;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- js_devdata = &kbdev->js_data;
- if ((js_devdata->init_status & JS_DEVDATA_INIT_CONSTANTS)) {
- s8 zero_ctx_attr_ref_count[KBASEP_JS_CTX_ATTR_COUNT] = { 0, };
- /* The caller must de-register all contexts before calling this
- */
- KBASE_DEBUG_ASSERT(js_devdata->nr_all_contexts_running == 0);
- KBASE_DEBUG_ASSERT(memcmp(
- js_devdata->runpool_irq.ctx_attr_ref_count,
- zero_ctx_attr_ref_count,
- sizeof(zero_ctx_attr_ref_count)) == 0);
- CSTD_UNUSED(zero_ctx_attr_ref_count);
- }
- if ((js_devdata->init_status & JS_DEVDATA_INIT_POLICY))
- kbasep_js_policy_term(&js_devdata->policy);
- js_devdata->init_status = JS_DEVDATA_INIT_NONE;
- }
- int kbasep_js_kctx_init(struct kbase_context * const kctx)
- {
- struct kbase_device *kbdev;
- struct kbasep_js_kctx_info *js_kctx_info;
- int err;
- int i, j;
- KBASE_DEBUG_ASSERT(kctx != NULL);
- kbdev = kctx->kbdev;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- for (i = 0; i < BASE_JM_MAX_NR_SLOTS; ++i)
- INIT_LIST_HEAD(&kctx->jctx.sched_info.ctx.ctx_list_entry[i]);
- js_kctx_info = &kctx->jctx.sched_info;
- KBASE_DEBUG_ASSERT(js_kctx_info->init_status == JS_KCTX_INIT_NONE);
- js_kctx_info->ctx.nr_jobs = 0;
- js_kctx_info->ctx.is_scheduled = false;
- js_kctx_info->ctx.is_dying = false;
- memset(js_kctx_info->ctx.ctx_attr_ref_count, 0,
- sizeof(js_kctx_info->ctx.ctx_attr_ref_count));
- /* Initially, the context is disabled from submission until the create
- * flags are set */
- js_kctx_info->ctx.flags = KBASE_CTX_FLAG_SUBMIT_DISABLED;
- js_kctx_info->init_status |= JS_KCTX_INIT_CONSTANTS;
- /* On error, we could continue on: providing none of the below resources
- * rely on the ones above */
- mutex_init(&js_kctx_info->ctx.jsctx_mutex);
- init_waitqueue_head(&js_kctx_info->ctx.is_scheduled_wait);
- err = kbasep_js_policy_init_ctx(kbdev, kctx);
- if (!err)
- js_kctx_info->init_status |= JS_KCTX_INIT_POLICY;
- /* On error, do no cleanup; this will be handled by the caller(s), since
- * we've designed this resource to be safe to terminate on init-fail */
- if (js_kctx_info->init_status != JS_KCTX_INIT_ALL)
- return -EINVAL;
- for (i = 0; i < KBASE_JS_ATOM_SCHED_PRIO_COUNT; i++) {
- for (j = 0; j < BASE_JM_MAX_NR_SLOTS; j++) {
- INIT_LIST_HEAD(&kctx->jsctx_queue[i][j].queue_head);
- atomic_set(&kctx->jsctx_queue[i][j].indicies, 0);
- }
- }
- return 0;
- }
- void kbasep_js_kctx_term(struct kbase_context *kctx)
- {
- struct kbase_device *kbdev;
- struct kbasep_js_kctx_info *js_kctx_info;
- union kbasep_js_policy *js_policy;
- int js;
- bool update_ctx_count = false;
- KBASE_DEBUG_ASSERT(kctx != NULL);
- kbdev = kctx->kbdev;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- js_policy = &kbdev->js_data.policy;
- js_kctx_info = &kctx->jctx.sched_info;
- if ((js_kctx_info->init_status & JS_KCTX_INIT_CONSTANTS)) {
- /* The caller must de-register all jobs before calling this */
- KBASE_DEBUG_ASSERT(!js_kctx_info->ctx.is_scheduled);
- KBASE_DEBUG_ASSERT(js_kctx_info->ctx.nr_jobs == 0);
- }
- mutex_lock(&kbdev->js_data.queue_mutex);
- mutex_lock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- for (js = 0; js < kbdev->gpu_props.num_job_slots; js++)
- list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
- if (kctx->ctx_runnable_ref) {
- WARN_ON(atomic_read(&kbdev->js_data.nr_contexts_runnable) <= 0);
- atomic_dec(&kbdev->js_data.nr_contexts_runnable);
- update_ctx_count = true;
- kctx->ctx_runnable_ref = false;
- }
- mutex_unlock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- mutex_unlock(&kbdev->js_data.queue_mutex);
- if ((js_kctx_info->init_status & JS_KCTX_INIT_POLICY))
- kbasep_js_policy_term_ctx(js_policy, kctx);
- js_kctx_info->init_status = JS_KCTX_INIT_NONE;
- if (update_ctx_count) {
- mutex_lock(&kbdev->js_data.runpool_mutex);
- kbase_backend_ctx_count_changed(kbdev);
- mutex_unlock(&kbdev->js_data.runpool_mutex);
- }
- }
- /**
- * kbase_js_ctx_list_add_pullable - Add context to the tail of the per-slot
- * pullable context queue
- * @kbdev: Device pointer
- * @kctx: Context to add to queue
- * @js: Job slot to use
- *
- * If the context is on either the pullable or unpullable queues, then it is
- * removed before being added to the tail.
- *
- * This function should be used when queueing a context for the first time, or
- * re-queueing a context that has been pulled from.
- *
- * Caller must hold kbasep_jd_device_data.queue_mutex
- *
- * Return: true if caller should call kbase_backend_ctx_count_changed()
- */
- static bool kbase_js_ctx_list_add_pullable(struct kbase_device *kbdev,
- struct kbase_context *kctx,
- int js)
- {
- bool ret = false;
- lockdep_assert_held(&kbdev->js_data.queue_mutex);
- lockdep_assert_held(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- if (!list_empty(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]))
- list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
- list_add_tail(&kctx->jctx.sched_info.ctx.ctx_list_entry[js],
- &kbdev->js_data.ctx_list_pullable[js]);
- if (!kctx->slots_pullable) {
- kbdev->js_data.nr_contexts_pullable++;
- ret = true;
- if (!atomic_read(&kctx->atoms_pulled)) {
- WARN_ON(kctx->ctx_runnable_ref);
- kctx->ctx_runnable_ref = true;
- atomic_inc(&kbdev->js_data.nr_contexts_runnable);
- }
- }
- kctx->slots_pullable |= (1 << js);
- return ret;
- }
- /**
- * kbase_js_ctx_list_add_pullable_head - Add context to the head of the
- * per-slot pullable context queue
- * @kbdev: Device pointer
- * @kctx: Context to add to queue
- * @js: Job slot to use
- *
- * If the context is on either the pullable or unpullable queues, then it is
- * removed before being added to the head.
- *
- * This function should be used when a context has been scheduled, but no jobs
- * can currently be pulled from it.
- *
- * Caller must hold kbasep_jd_device_data.queue_mutex
- *
- * Return: true if caller should call kbase_backend_ctx_count_changed()
- */
- static bool kbase_js_ctx_list_add_pullable_head(struct kbase_device *kbdev,
- struct kbase_context *kctx,
- int js)
- {
- bool ret = false;
- lockdep_assert_held(&kbdev->js_data.queue_mutex);
- lockdep_assert_held(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- if (!list_empty(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]))
- list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
- list_add(&kctx->jctx.sched_info.ctx.ctx_list_entry[js],
- &kbdev->js_data.ctx_list_pullable[js]);
- if (!kctx->slots_pullable) {
- kbdev->js_data.nr_contexts_pullable++;
- ret = true;
- if (!atomic_read(&kctx->atoms_pulled)) {
- WARN_ON(kctx->ctx_runnable_ref);
- kctx->ctx_runnable_ref = true;
- atomic_inc(&kbdev->js_data.nr_contexts_runnable);
- }
- }
- kctx->slots_pullable |= (1 << js);
- return ret;
- }
- /**
- * kbase_js_ctx_list_add_unpullable - Add context to the tail of the per-slot
- * unpullable context queue
- * @kbdev: Device pointer
- * @kctx: Context to add to queue
- * @js: Job slot to use
- *
- * The context must already be on the per-slot pullable queue. It will be
- * removed from the pullable queue before being added to the unpullable queue.
- *
- * This function should be used when a context has been pulled from, and there
- * are no jobs remaining on the specified slot.
- *
- * Caller must hold kbasep_jd_device_data.queue_mutex
- *
- * Return: true if caller should call kbase_backend_ctx_count_changed()
- */
- static bool kbase_js_ctx_list_add_unpullable(struct kbase_device *kbdev,
- struct kbase_context *kctx,
- int js)
- {
- bool ret = false;
- lockdep_assert_held(&kbdev->js_data.queue_mutex);
- lockdep_assert_held(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- list_move_tail(&kctx->jctx.sched_info.ctx.ctx_list_entry[js],
- &kbdev->js_data.ctx_list_unpullable[js]);
- if (kctx->slots_pullable == (1 << js)) {
- kbdev->js_data.nr_contexts_pullable--;
- ret = true;
- if (!atomic_read(&kctx->atoms_pulled)) {
- WARN_ON(!kctx->ctx_runnable_ref);
- kctx->ctx_runnable_ref = false;
- atomic_dec(&kbdev->js_data.nr_contexts_runnable);
- }
- }
- kctx->slots_pullable &= ~(1 << js);
- return ret;
- }
- /**
- * kbase_js_ctx_list_remove - Remove context from the per-slot pullable or
- * unpullable context queues
- * @kbdev: Device pointer
- * @kctx: Context to remove from queue
- * @js: Job slot to use
- *
- * The context must already be on one of the queues.
- *
- * This function should be used when a context has no jobs on the GPU, and no
- * jobs remaining for the specified slot.
- *
- * Caller must hold kbasep_jd_device_data.queue_mutex
- *
- * Return: true if caller should call kbase_backend_ctx_count_changed()
- */
- static bool kbase_js_ctx_list_remove(struct kbase_device *kbdev,
- struct kbase_context *kctx,
- int js)
- {
- bool ret = false;
- lockdep_assert_held(&kbdev->js_data.queue_mutex);
- lockdep_assert_held(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- WARN_ON(list_empty(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]));
- list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
- if (kctx->slots_pullable == (1 << js)) {
- kbdev->js_data.nr_contexts_pullable--;
- ret = true;
- if (!atomic_read(&kctx->atoms_pulled)) {
- WARN_ON(!kctx->ctx_runnable_ref);
- kctx->ctx_runnable_ref = false;
- atomic_dec(&kbdev->js_data.nr_contexts_runnable);
- }
- }
- kctx->slots_pullable &= ~(1 << js);
- return ret;
- }
- /**
- * kbase_js_ctx_list_pop_head - Pop the head context off the per-slot pullable
- * queue.
- * @kbdev: Device pointer
- * @js: Job slot to use
- *
- * Caller must hold kbasep_jd_device_data::queue_mutex
- *
- * Return: Context to use for specified slot.
- * NULL if no contexts present for specified slot
- */
- static struct kbase_context *kbase_js_ctx_list_pop_head(
- struct kbase_device *kbdev,
- int js)
- {
- struct kbase_context *kctx;
- lockdep_assert_held(&kbdev->js_data.queue_mutex);
- if (list_empty(&kbdev->js_data.ctx_list_pullable[js]))
- return NULL;
- kctx = list_entry(kbdev->js_data.ctx_list_pullable[js].next,
- struct kbase_context,
- jctx.sched_info.ctx.ctx_list_entry[js]);
- list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
- return kctx;
- }
- /**
- * kbase_js_ctx_pullable - Return if a context can be pulled from on the
- * specified slot
- * @kctx: Context pointer
- * @js: Job slot to use
- * @is_scheduled: true if the context is currently scheduled
- *
- * Caller must hold runpool_irq.lock
- *
- * Return: true if context can be pulled from on specified slot
- * false otherwise
- */
- static bool kbase_js_ctx_pullable(struct kbase_context *kctx, int js,
- bool is_scheduled)
- {
- struct kbasep_js_device_data *js_devdata;
- struct kbase_jd_atom *katom;
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- js_devdata = &kctx->kbdev->js_data;
- if (is_scheduled) {
- if (!kbasep_js_is_submit_allowed(js_devdata, kctx))
- return false;
- }
- katom = jsctx_rb_peek(kctx, js);
- if (!katom)
- return false; /* No pullable atoms */
- if (atomic_read(&katom->blocked))
- return false; /* next atom blocked */
- if (katom->atom_flags & KBASE_KATOM_FLAG_X_DEP_BLOCKED) {
- if (katom->x_pre_dep->gpu_rb_state ==
- KBASE_ATOM_GPU_RB_NOT_IN_SLOT_RB ||
- katom->x_pre_dep->will_fail_event_code)
- return false;
- if ((katom->atom_flags & KBASE_KATOM_FLAG_FAIL_BLOCKER) &&
- kbase_backend_nr_atoms_on_slot(kctx->kbdev, js))
- return false;
- }
- return true;
- }
- static bool kbase_js_dep_validate(struct kbase_context *kctx,
- struct kbase_jd_atom *katom)
- {
- struct kbase_device *kbdev = kctx->kbdev;
- bool ret = true;
- bool has_dep = false, has_x_dep = false;
- int js = kbase_js_get_slot(kbdev, katom);
- int prio = katom->sched_priority;
- int i;
- for (i = 0; i < 2; i++) {
- struct kbase_jd_atom *dep_atom = katom->dep[i].atom;
- if (dep_atom) {
- int dep_js = kbase_js_get_slot(kbdev, dep_atom);
- int dep_prio = dep_atom->sched_priority;
- /* Dependent atom must already have been submitted */
- if (!(dep_atom->atom_flags &
- (KBASE_KATOM_FLAG_JSCTX_IN_LL |
- KBASE_KATOM_FLAG_JSCTX_RB_SUBMITTED))){
- ret = false;
- break;
- }
- /* Dependencies with different priorities can't
- be represented in the ringbuffer */
- if (prio != dep_prio) {
- ret = false;
- break;
- }
- if (js == dep_js) {
- /* Only one same-slot dependency can be
- * represented in the ringbuffer */
- if (has_dep) {
- ret = false;
- break;
- }
- has_dep = true;
- } else {
- /* Only one cross-slot dependency can be
- * represented in the ringbuffer */
- if (has_x_dep) {
- ret = false;
- break;
- }
- /* Each dependee atom can only have one
- * cross-slot dependency */
- if (dep_atom->x_post_dep) {
- ret = false;
- break;
- }
- /* The dependee atom can not already be in the
- * HW access ringbuffer */
- if (dep_atom->gpu_rb_state !=
- KBASE_ATOM_GPU_RB_NOT_IN_SLOT_RB) {
- ret = false;
- break;
- }
- /* The dependee atom can not already have
- * completed */
- if (dep_atom->status !=
- KBASE_JD_ATOM_STATE_IN_JS) {
- ret = false;
- break;
- }
- /* Cross-slot dependencies must not violate
- * PRLAM-8987 affinity restrictions */
- if (kbase_hw_has_issue(kbdev,
- BASE_HW_ISSUE_8987) &&
- (js == 2 || dep_js == 2)) {
- ret = false;
- break;
- }
- has_x_dep = true;
- }
- if (kbase_jd_katom_dep_type(&katom->dep[i]) ==
- BASE_JD_DEP_TYPE_DATA &&
- js == dep_js) {
- struct kbase_jd_atom *last_atom =
- jsctx_rb_peek_last(kctx, js,
- prio);
- /* Last atom on slot must be pre-dep for this
- * atom */
- if (last_atom != dep_atom) {
- ret = false;
- break;
- }
- }
- /* Dependency can be represented in ringbuffers */
- }
- }
- /* If dependencies can be represented by ringbuffer then clear them from
- * atom structure */
- if (ret) {
- for (i = 0; i < 2; i++) {
- struct kbase_jd_atom *dep_atom = katom->dep[i].atom;
- if (dep_atom) {
- int dep_js = kbase_js_get_slot(kbdev, dep_atom);
- if ((js != dep_js) &&
- (dep_atom->status !=
- KBASE_JD_ATOM_STATE_COMPLETED)
- && (dep_atom->status !=
- KBASE_JD_ATOM_STATE_HW_COMPLETED)
- && (dep_atom->status !=
- KBASE_JD_ATOM_STATE_UNUSED)) {
- katom->atom_flags |=
- KBASE_KATOM_FLAG_X_DEP_BLOCKED;
- katom->x_pre_dep = dep_atom;
- dep_atom->x_post_dep = katom;
- if (kbase_jd_katom_dep_type(
- &katom->dep[i]) ==
- BASE_JD_DEP_TYPE_DATA)
- katom->atom_flags |=
- KBASE_KATOM_FLAG_FAIL_BLOCKER;
- }
- if ((kbase_jd_katom_dep_type(&katom->dep[i])
- == BASE_JD_DEP_TYPE_DATA) &&
- (js == dep_js))
- katom->atom_flags |=
- KBASE_KATOM_FLAG_FAIL_PREV;
- list_del(&katom->dep_item[i]);
- kbase_jd_katom_dep_clear(&katom->dep[i]);
- }
- }
- }
- return ret;
- }
- bool kbasep_js_add_job(struct kbase_context *kctx,
- struct kbase_jd_atom *atom)
- {
- unsigned long flags;
- struct kbasep_js_kctx_info *js_kctx_info;
- struct kbase_device *kbdev;
- struct kbasep_js_device_data *js_devdata;
- union kbasep_js_policy *js_policy;
- bool enqueue_required = false;
- bool timer_sync = false;
- KBASE_DEBUG_ASSERT(kctx != NULL);
- KBASE_DEBUG_ASSERT(atom != NULL);
- lockdep_assert_held(&kctx->jctx.lock);
- kbdev = kctx->kbdev;
- js_devdata = &kbdev->js_data;
- js_policy = &kbdev->js_data.policy;
- js_kctx_info = &kctx->jctx.sched_info;
- mutex_lock(&js_devdata->queue_mutex);
- mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
- /*
- * Begin Runpool transaction
- */
- mutex_lock(&js_devdata->runpool_mutex);
- /* Refcount ctx.nr_jobs */
- KBASE_DEBUG_ASSERT(js_kctx_info->ctx.nr_jobs < U32_MAX);
- ++(js_kctx_info->ctx.nr_jobs);
- /* Setup any scheduling information */
- kbasep_js_clear_job_retry_submit(atom);
- /* Lock for state available during IRQ */
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- if (!kbase_js_dep_validate(kctx, atom)) {
- /* Dependencies could not be represented */
- --(js_kctx_info->ctx.nr_jobs);
- /* Setting atom status back to queued as it still has unresolved
- * dependencies */
- atom->status = KBASE_JD_ATOM_STATE_QUEUED;
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- mutex_unlock(&js_devdata->runpool_mutex);
- goto out_unlock;
- }
- KBASE_TIMELINE_ATOM_READY(kctx, kbase_jd_atom_id(kctx, atom));
- enqueue_required = kbase_js_dep_resolved_submit(kctx, atom);
- KBASE_TRACE_ADD_REFCOUNT(kbdev, JS_ADD_JOB, kctx, atom, atom->jc,
- kbasep_js_trace_get_refcnt_nolock(kbdev, kctx));
- /* Context Attribute Refcounting */
- kbasep_js_ctx_attr_ctx_retain_atom(kbdev, kctx, atom);
- if (enqueue_required) {
- if (kbase_js_ctx_pullable(kctx, atom->slot_nr, false))
- timer_sync = kbase_js_ctx_list_add_pullable(kbdev, kctx,
- atom->slot_nr);
- else
- timer_sync = kbase_js_ctx_list_add_unpullable(kbdev,
- kctx, atom->slot_nr);
- }
- /* If this context is active and the atom is the first on its slot,
- * kick the job manager to attempt to fast-start the atom */
- if (enqueue_required && kctx == kbdev->hwaccess.active_kctx)
- kbase_jm_try_kick(kbdev, 1 << atom->slot_nr);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- if (timer_sync)
- kbase_backend_ctx_count_changed(kbdev);
- mutex_unlock(&js_devdata->runpool_mutex);
- /* End runpool transaction */
- if (!js_kctx_info->ctx.is_scheduled) {
- if (js_kctx_info->ctx.is_dying) {
- /* A job got added while/after kbase_job_zap_context()
- * was called on a non-scheduled context (e.g. KDS
- * dependency resolved). Kill that job by killing the
- * context. */
- kbasep_js_runpool_requeue_or_kill_ctx(kbdev, kctx,
- false);
- } else if (js_kctx_info->ctx.nr_jobs == 1) {
- /* Handle Refcount going from 0 to 1: schedule the
- * context on the Policy Queue */
- KBASE_DEBUG_ASSERT(!js_kctx_info->ctx.is_scheduled);
- dev_dbg(kbdev->dev, "JS: Enqueue Context %p", kctx);
- /* Policy Queue was updated - caller must try to
- * schedule the head context */
- WARN_ON(!enqueue_required);
- }
- }
- out_unlock:
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_unlock(&js_devdata->queue_mutex);
- return enqueue_required;
- }
- void kbasep_js_remove_job(struct kbase_device *kbdev,
- struct kbase_context *kctx, struct kbase_jd_atom *atom)
- {
- struct kbasep_js_kctx_info *js_kctx_info;
- struct kbasep_js_device_data *js_devdata;
- union kbasep_js_policy *js_policy;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(kctx != NULL);
- KBASE_DEBUG_ASSERT(atom != NULL);
- js_devdata = &kbdev->js_data;
- js_policy = &kbdev->js_data.policy;
- js_kctx_info = &kctx->jctx.sched_info;
- KBASE_TRACE_ADD_REFCOUNT(kbdev, JS_REMOVE_JOB, kctx, atom, atom->jc,
- kbasep_js_trace_get_refcnt(kbdev, kctx));
- /* De-refcount ctx.nr_jobs */
- KBASE_DEBUG_ASSERT(js_kctx_info->ctx.nr_jobs > 0);
- --(js_kctx_info->ctx.nr_jobs);
- }
- bool kbasep_js_remove_cancelled_job(struct kbase_device *kbdev,
- struct kbase_context *kctx, struct kbase_jd_atom *katom)
- {
- unsigned long flags;
- struct kbasep_js_atom_retained_state katom_retained_state;
- struct kbasep_js_device_data *js_devdata;
- bool attr_state_changed;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(kctx != NULL);
- KBASE_DEBUG_ASSERT(katom != NULL);
- js_devdata = &kbdev->js_data;
- kbasep_js_atom_retained_state_copy(&katom_retained_state, katom);
- kbasep_js_remove_job(kbdev, kctx, katom);
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- /* The atom has 'finished' (will not be re-run), so no need to call
- * kbasep_js_has_atom_finished().
- *
- * This is because it returns false for soft-stopped atoms, but we
- * want to override that, because we're cancelling an atom regardless of
- * whether it was soft-stopped or not */
- attr_state_changed = kbasep_js_ctx_attr_ctx_release_atom(kbdev, kctx,
- &katom_retained_state);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- return attr_state_changed;
- }
- bool kbasep_js_runpool_retain_ctx(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- unsigned long flags;
- struct kbasep_js_device_data *js_devdata;
- bool result;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- js_devdata = &kbdev->js_data;
- /* KBASE_TRACE_ADD_REFCOUNT( kbdev, JS_RETAIN_CTX, kctx, NULL, 0,
- kbasep_js_trace_get_refcnt(kbdev, kctx)); */
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- result = kbasep_js_runpool_retain_ctx_nolock(kbdev, kctx);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- return result;
- }
- struct kbase_context *kbasep_js_runpool_lookup_ctx(struct kbase_device *kbdev,
- int as_nr)
- {
- unsigned long flags;
- struct kbasep_js_device_data *js_devdata;
- struct kbase_context *found_kctx = NULL;
- struct kbasep_js_per_as_data *js_per_as_data;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(0 <= as_nr && as_nr < BASE_MAX_NR_AS);
- js_devdata = &kbdev->js_data;
- js_per_as_data = &js_devdata->runpool_irq.per_as_data[as_nr];
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- found_kctx = js_per_as_data->kctx;
- if (found_kctx != NULL)
- ++(js_per_as_data->as_busy_refcount);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- return found_kctx;
- }
- struct kbase_context *kbasep_js_runpool_lookup_ctx_nolock(
- struct kbase_device *kbdev, int as_nr)
- {
- struct kbasep_js_device_data *js_devdata;
- struct kbase_context *found_kctx = NULL;
- struct kbasep_js_per_as_data *js_per_as_data;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(0 <= as_nr && as_nr < BASE_MAX_NR_AS);
- lockdep_assert_held(&kbdev->js_data.runpool_irq.lock);
- js_devdata = &kbdev->js_data;
- js_per_as_data = &js_devdata->runpool_irq.per_as_data[as_nr];
- found_kctx = js_per_as_data->kctx;
- if (found_kctx != NULL)
- ++(js_per_as_data->as_busy_refcount);
- return found_kctx;
- }
- /**
- * kbasep_js_release_result - Try running more jobs after releasing a context
- * and/or atom
- *
- * @kbdev: The kbase_device to operate on
- * @kctx: The kbase_context to operate on
- * @katom_retained_state: Retained state from the atom
- * @runpool_ctx_attr_change: True if the runpool context attributes have changed
- *
- * This collates a set of actions that must happen whilst
- * kbasep_js_device_data.runpool_irq.lock is held.
- *
- * This includes running more jobs when:
- * - The previously released kctx caused a ctx attribute change,
- * - The released atom caused a ctx attribute change,
- * - Slots were previously blocked due to affinity restrictions,
- * - Submission during IRQ handling failed.
- *
- * Return: %KBASEP_JS_RELEASE_RESULT_SCHED_ALL if context attributes were
- * changed. The caller should try scheduling all contexts
- */
- static kbasep_js_release_result kbasep_js_run_jobs_after_ctx_and_atom_release(
- struct kbase_device *kbdev,
- struct kbase_context *kctx,
- struct kbasep_js_atom_retained_state *katom_retained_state,
- bool runpool_ctx_attr_change)
- {
- struct kbasep_js_device_data *js_devdata;
- kbasep_js_release_result result = 0;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(kctx != NULL);
- KBASE_DEBUG_ASSERT(katom_retained_state != NULL);
- js_devdata = &kbdev->js_data;
- lockdep_assert_held(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- lockdep_assert_held(&js_devdata->runpool_mutex);
- lockdep_assert_held(&js_devdata->runpool_irq.lock);
- if (js_devdata->nr_user_contexts_running != 0) {
- bool retry_submit = false;
- int retry_jobslot = 0;
- if (katom_retained_state)
- retry_submit = kbasep_js_get_atom_retry_submit_slot(
- katom_retained_state, &retry_jobslot);
- if (runpool_ctx_attr_change || retry_submit) {
- /* A change in runpool ctx attributes might mean we can
- * run more jobs than before */
- result = KBASEP_JS_RELEASE_RESULT_SCHED_ALL;
- KBASE_TRACE_ADD_SLOT(kbdev, JD_DONE_TRY_RUN_NEXT_JOB,
- kctx, NULL, 0u, retry_jobslot);
- }
- }
- return result;
- }
- /*
- * Internal function to release the reference on a ctx and an atom's "retained
- * state", only taking the runpool and as transaction mutexes
- *
- * This also starts more jobs running in the case of an ctx-attribute state
- * change
- *
- * This does none of the followup actions for scheduling:
- * - It does not schedule in a new context
- * - It does not requeue or handle dying contexts
- *
- * For those tasks, just call kbasep_js_runpool_release_ctx() instead
- *
- * Requires:
- * - Context is scheduled in, and kctx->as_nr matches kctx_as_nr
- * - Context has a non-zero refcount
- * - Caller holds js_kctx_info->ctx.jsctx_mutex
- * - Caller holds js_devdata->runpool_mutex
- */
- static kbasep_js_release_result kbasep_js_runpool_release_ctx_internal(
- struct kbase_device *kbdev,
- struct kbase_context *kctx,
- struct kbasep_js_atom_retained_state *katom_retained_state)
- {
- unsigned long flags;
- struct kbasep_js_device_data *js_devdata;
- struct kbasep_js_kctx_info *js_kctx_info;
- union kbasep_js_policy *js_policy;
- struct kbasep_js_per_as_data *js_per_as_data;
- kbasep_js_release_result release_result = 0u;
- bool runpool_ctx_attr_change = false;
- int kctx_as_nr;
- struct kbase_as *current_as;
- int new_ref_count;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(kctx != NULL);
- js_kctx_info = &kctx->jctx.sched_info;
- js_devdata = &kbdev->js_data;
- js_policy = &kbdev->js_data.policy;
- /* Ensure context really is scheduled in */
- KBASE_DEBUG_ASSERT(js_kctx_info->ctx.is_scheduled);
- /* kctx->as_nr and js_per_as_data are only read from here. The caller's
- * js_ctx_mutex provides a barrier that ensures they are up-to-date.
- *
- * They will not change whilst we're reading them, because the refcount
- * is non-zero (and we ASSERT on that last fact).
- */
- kctx_as_nr = kctx->as_nr;
- KBASE_DEBUG_ASSERT(kctx_as_nr != KBASEP_AS_NR_INVALID);
- js_per_as_data = &js_devdata->runpool_irq.per_as_data[kctx_as_nr];
- KBASE_DEBUG_ASSERT(js_per_as_data->as_busy_refcount > 0);
- /*
- * Transaction begins on AS and runpool_irq
- *
- * Assert about out calling contract
- */
- current_as = &kbdev->as[kctx_as_nr];
- mutex_lock(&kbdev->pm.lock);
- mutex_lock(¤t_as->transaction_mutex);
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- KBASE_DEBUG_ASSERT(kctx_as_nr == kctx->as_nr);
- KBASE_DEBUG_ASSERT(js_per_as_data->as_busy_refcount > 0);
- /* Update refcount */
- new_ref_count = --(js_per_as_data->as_busy_refcount);
- /* Release the atom if it finished (i.e. wasn't soft-stopped) */
- if (kbasep_js_has_atom_finished(katom_retained_state))
- runpool_ctx_attr_change |= kbasep_js_ctx_attr_ctx_release_atom(
- kbdev, kctx, katom_retained_state);
- KBASE_TRACE_ADD_REFCOUNT(kbdev, JS_RELEASE_CTX, kctx, NULL, 0u,
- new_ref_count);
- if (new_ref_count == 1 && kctx->jctx.sched_info.ctx.flags &
- KBASE_CTX_FLAG_PRIVILEGED &&
- !kbase_pm_is_suspending(kbdev)) {
- /* Context is kept scheduled into an address space even when
- * there are no jobs, in this case we have to handle the
- * situation where all jobs have been evicted from the GPU and
- * submission is disabled.
- *
- * At this point we re-enable submission to allow further jobs
- * to be executed
- */
- kbasep_js_set_submit_allowed(js_devdata, kctx);
- }
- /* Make a set of checks to see if the context should be scheduled out */
- if (new_ref_count == 0 &&
- (!kbasep_js_is_submit_allowed(js_devdata, kctx) ||
- kbdev->pm.suspending)) {
- /* Last reference, and we've been told to remove this context
- * from the Run Pool */
- dev_dbg(kbdev->dev, "JS: RunPool Remove Context %p because as_busy_refcount=%d, jobs=%d, allowed=%d",
- kctx, new_ref_count, js_kctx_info->ctx.nr_jobs,
- kbasep_js_is_submit_allowed(js_devdata, kctx));
- #if defined(CONFIG_MALI_GATOR_SUPPORT)
- kbase_trace_mali_mmu_as_released(kctx->as_nr);
- #endif
- kbase_tlstream_tl_nret_as_ctx(&kbdev->as[kctx->as_nr], kctx);
- kbase_backend_release_ctx_irq(kbdev, kctx);
- if (kbdev->hwaccess.active_kctx == kctx)
- kbdev->hwaccess.active_kctx = NULL;
- /* Ctx Attribute handling
- *
- * Releasing atoms attributes must either happen before this, or
- * after 'is_scheduled' is changed, otherwise we double-decount
- * the attributes */
- runpool_ctx_attr_change |=
- kbasep_js_ctx_attr_runpool_release_ctx(kbdev, kctx);
- /* Releasing the context and katom retained state can allow
- * more jobs to run */
- release_result |=
- kbasep_js_run_jobs_after_ctx_and_atom_release(kbdev,
- kctx, katom_retained_state,
- runpool_ctx_attr_change);
- /*
- * Transaction ends on AS and runpool_irq:
- *
- * By this point, the AS-related data is now clear and ready
- * for re-use.
- *
- * Since releases only occur once for each previous successful
- * retain, and no more retains are allowed on this context, no
- * other thread will be operating in this
- * code whilst we are
- */
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- kbase_backend_release_ctx_noirq(kbdev, kctx);
- mutex_unlock(¤t_as->transaction_mutex);
- mutex_unlock(&kbdev->pm.lock);
- /* Note: Don't reuse kctx_as_nr now */
- /* Synchronize with any policy timers */
- kbase_backend_ctx_count_changed(kbdev);
- /* update book-keeping info */
- js_kctx_info->ctx.is_scheduled = false;
- /* Signal any waiter that the context is not scheduled, so is
- * safe for termination - once the jsctx_mutex is also dropped,
- * and jobs have finished. */
- wake_up(&js_kctx_info->ctx.is_scheduled_wait);
- /* Queue an action to occur after we've dropped the lock */
- release_result |= KBASEP_JS_RELEASE_RESULT_WAS_DESCHEDULED;
- } else {
- kbasep_js_run_jobs_after_ctx_and_atom_release(kbdev, kctx,
- katom_retained_state, runpool_ctx_attr_change);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- mutex_unlock(¤t_as->transaction_mutex);
- mutex_unlock(&kbdev->pm.lock);
- }
- return release_result;
- }
- void kbasep_js_runpool_release_ctx_nolock(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- struct kbasep_js_atom_retained_state katom_retained_state;
- /* Setup a dummy katom_retained_state */
- kbasep_js_atom_retained_state_init_invalid(&katom_retained_state);
- kbasep_js_runpool_release_ctx_internal(kbdev, kctx,
- &katom_retained_state);
- }
- void kbasep_js_runpool_requeue_or_kill_ctx(struct kbase_device *kbdev,
- struct kbase_context *kctx, bool has_pm_ref)
- {
- struct kbasep_js_device_data *js_devdata;
- union kbasep_js_policy *js_policy;
- struct kbasep_js_kctx_info *js_kctx_info;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(kctx != NULL);
- js_kctx_info = &kctx->jctx.sched_info;
- js_policy = &kbdev->js_data.policy;
- js_devdata = &kbdev->js_data;
- /* This is called if and only if you've you've detached the context from
- * the Runpool or the Policy Queue, and not added it back to the Runpool
- */
- KBASE_DEBUG_ASSERT(!js_kctx_info->ctx.is_scheduled);
- if (js_kctx_info->ctx.is_dying) {
- /* Dying: don't requeue, but kill all jobs on the context. This
- * happens asynchronously */
- dev_dbg(kbdev->dev,
- "JS: ** Killing Context %p on RunPool Remove **", kctx);
- kbase_js_foreach_ctx_job(kctx, &kbase_jd_cancel);
- }
- }
- void kbasep_js_runpool_release_ctx_and_katom_retained_state(
- struct kbase_device *kbdev, struct kbase_context *kctx,
- struct kbasep_js_atom_retained_state *katom_retained_state)
- {
- struct kbasep_js_device_data *js_devdata;
- struct kbasep_js_kctx_info *js_kctx_info;
- base_jd_event_code event_code;
- kbasep_js_release_result release_result;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(kctx != NULL);
- js_kctx_info = &kctx->jctx.sched_info;
- js_devdata = &kbdev->js_data;
- event_code = katom_retained_state->event_code;
- mutex_lock(&js_devdata->queue_mutex);
- mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_lock(&js_devdata->runpool_mutex);
- release_result = kbasep_js_runpool_release_ctx_internal(kbdev, kctx,
- katom_retained_state);
- /* Drop the runpool mutex to allow requeing kctx */
- mutex_unlock(&js_devdata->runpool_mutex);
- if ((release_result & KBASEP_JS_RELEASE_RESULT_WAS_DESCHEDULED) != 0u)
- kbasep_js_runpool_requeue_or_kill_ctx(kbdev, kctx, true);
- /* Drop the jsctx_mutex to allow scheduling in a new context */
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_unlock(&js_devdata->queue_mutex);
- if (release_result & KBASEP_JS_RELEASE_RESULT_SCHED_ALL)
- kbase_js_sched_all(kbdev);
- }
- void kbasep_js_runpool_release_ctx(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- struct kbasep_js_atom_retained_state katom_retained_state;
- kbasep_js_atom_retained_state_init_invalid(&katom_retained_state);
- kbasep_js_runpool_release_ctx_and_katom_retained_state(kbdev, kctx,
- &katom_retained_state);
- }
- /* Variant of kbasep_js_runpool_release_ctx() that doesn't call into
- * kbase_js_sched_all() */
- static void kbasep_js_runpool_release_ctx_no_schedule(
- struct kbase_device *kbdev, struct kbase_context *kctx)
- {
- struct kbasep_js_device_data *js_devdata;
- struct kbasep_js_kctx_info *js_kctx_info;
- kbasep_js_release_result release_result;
- struct kbasep_js_atom_retained_state katom_retained_state_struct;
- struct kbasep_js_atom_retained_state *katom_retained_state =
- &katom_retained_state_struct;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(kctx != NULL);
- js_kctx_info = &kctx->jctx.sched_info;
- js_devdata = &kbdev->js_data;
- kbasep_js_atom_retained_state_init_invalid(katom_retained_state);
- mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_lock(&js_devdata->runpool_mutex);
- release_result = kbasep_js_runpool_release_ctx_internal(kbdev, kctx,
- katom_retained_state);
- /* Drop the runpool mutex to allow requeing kctx */
- mutex_unlock(&js_devdata->runpool_mutex);
- if ((release_result & KBASEP_JS_RELEASE_RESULT_WAS_DESCHEDULED) != 0u)
- kbasep_js_runpool_requeue_or_kill_ctx(kbdev, kctx, true);
- /* Drop the jsctx_mutex to allow scheduling in a new context */
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- /* NOTE: could return release_result if the caller would like to know
- * whether it should schedule a new context, but currently no callers do
- */
- }
- /**
- * kbase_js_set_timeouts - update all JS timeouts with user specified data
- * @kbdev: Device pointer
- *
- * Timeouts are specified through the 'js_timeouts' sysfs file. If a timeout is
- * set to a positive number then that becomes the new value used, if a timeout
- * is negative then the default is set.
- */
- static void kbase_js_set_timeouts(struct kbase_device *kbdev)
- {
- struct kbasep_js_device_data *js_data = &kbdev->js_data;
- if (kbdev->js_scheduling_period_ns < 0)
- js_data->scheduling_period_ns = DEFAULT_JS_SCHEDULING_PERIOD_NS;
- else if (kbdev->js_scheduling_period_ns > 0)
- js_data->scheduling_period_ns = kbdev->js_scheduling_period_ns;
- if (kbdev->js_soft_stop_ticks < 0)
- js_data->soft_stop_ticks = DEFAULT_JS_SOFT_STOP_TICKS;
- else if (kbdev->js_soft_stop_ticks > 0)
- js_data->soft_stop_ticks = kbdev->js_soft_stop_ticks;
- if (kbdev->js_soft_stop_ticks_cl < 0)
- js_data->soft_stop_ticks_cl = DEFAULT_JS_SOFT_STOP_TICKS_CL;
- else if (kbdev->js_soft_stop_ticks_cl > 0)
- js_data->soft_stop_ticks_cl = kbdev->js_soft_stop_ticks_cl;
- if (kbdev->js_hard_stop_ticks_ss < 0) {
- if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8408))
- js_data->hard_stop_ticks_ss =
- DEFAULT_JS_HARD_STOP_TICKS_SS_8408;
- else
- js_data->hard_stop_ticks_ss =
- DEFAULT_JS_HARD_STOP_TICKS_SS;
- } else if (kbdev->js_hard_stop_ticks_ss > 0) {
- js_data->hard_stop_ticks_ss = kbdev->js_hard_stop_ticks_ss;
- }
- if (kbdev->js_hard_stop_ticks_cl < 0)
- js_data->hard_stop_ticks_cl = DEFAULT_JS_HARD_STOP_TICKS_CL;
- else if (kbdev->js_hard_stop_ticks_cl > 0)
- js_data->hard_stop_ticks_cl = kbdev->js_hard_stop_ticks_cl;
- if (kbdev->js_hard_stop_ticks_dumping < 0)
- js_data->hard_stop_ticks_dumping =
- DEFAULT_JS_HARD_STOP_TICKS_DUMPING;
- else if (kbdev->js_hard_stop_ticks_dumping > 0)
- js_data->hard_stop_ticks_dumping =
- kbdev->js_hard_stop_ticks_dumping;
- if (kbdev->js_reset_ticks_ss < 0) {
- if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8408))
- js_data->gpu_reset_ticks_ss =
- DEFAULT_JS_RESET_TICKS_SS_8408;
- else
- js_data->gpu_reset_ticks_ss = DEFAULT_JS_RESET_TICKS_SS;
- } else if (kbdev->js_reset_ticks_ss > 0) {
- js_data->gpu_reset_ticks_ss = kbdev->js_reset_ticks_ss;
- }
- if (kbdev->js_reset_ticks_cl < 0)
- js_data->gpu_reset_ticks_cl = DEFAULT_JS_RESET_TICKS_CL;
- else if (kbdev->js_reset_ticks_cl > 0)
- js_data->gpu_reset_ticks_cl = kbdev->js_reset_ticks_cl;
- if (kbdev->js_reset_ticks_dumping < 0)
- js_data->gpu_reset_ticks_dumping =
- DEFAULT_JS_RESET_TICKS_DUMPING;
- else if (kbdev->js_reset_ticks_dumping > 0)
- js_data->gpu_reset_ticks_dumping =
- kbdev->js_reset_ticks_dumping;
- }
- static bool kbasep_js_schedule_ctx(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- struct kbasep_js_device_data *js_devdata;
- struct kbasep_js_kctx_info *js_kctx_info;
- union kbasep_js_policy *js_policy;
- struct kbase_as *new_address_space = NULL;
- unsigned long flags;
- bool kctx_suspended = false;
- int as_nr;
- js_devdata = &kbdev->js_data;
- js_policy = &kbdev->js_data.policy;
- js_kctx_info = &kctx->jctx.sched_info;
- /* Pick available address space for this context */
- as_nr = kbase_backend_find_free_address_space(kbdev, kctx);
- if (as_nr == KBASEP_AS_NR_INVALID)
- return false; /* No address spaces currently available */
- new_address_space = &kbdev->as[as_nr];
- /*
- * Atomic transaction on the Context and Run Pool begins
- */
- mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_lock(&js_devdata->runpool_mutex);
- /* Check to see if context is dying due to kbase_job_zap_context() */
- if (js_kctx_info->ctx.is_dying) {
- /* Roll back the transaction so far and return */
- kbase_backend_release_free_address_space(kbdev, as_nr);
- mutex_unlock(&js_devdata->runpool_mutex);
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- return false;
- }
- KBASE_TRACE_ADD_REFCOUNT(kbdev, JS_TRY_SCHEDULE_HEAD_CTX, kctx, NULL,
- 0u,
- kbasep_js_trace_get_refcnt(kbdev, kctx));
- if (js_devdata->nr_user_contexts_running == 0 &&
- kbdev->js_timeouts_updated) {
- /* Only when there are no other contexts submitting jobs:
- * Latch in run-time job scheduler timeouts that were set
- * through js_timeouts sysfs file */
- kbase_js_set_timeouts(kbdev);
- kbdev->js_timeouts_updated = false;
- }
- js_kctx_info->ctx.is_scheduled = true;
- mutex_lock(&new_address_space->transaction_mutex);
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- /* Assign context to previously chosen address space */
- if (!kbase_backend_use_ctx(kbdev, kctx, as_nr)) {
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- mutex_unlock(&new_address_space->transaction_mutex);
- /* If address space is not pending, then kbase_backend_use_ctx()
- * failed. Roll back the transaction so far and return */
- if (!kctx->as_pending) {
- js_kctx_info->ctx.is_scheduled = false;
- kbase_backend_release_free_address_space(kbdev, as_nr);
- }
- mutex_unlock(&js_devdata->runpool_mutex);
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- return false;
- }
- kbdev->hwaccess.active_kctx = kctx;
- #if defined(CONFIG_MALI_GATOR_SUPPORT)
- kbase_trace_mali_mmu_as_in_use(kctx->as_nr);
- #endif
- kbase_tlstream_tl_ret_as_ctx(&kbdev->as[kctx->as_nr], kctx);
- /* Cause any future waiter-on-termination to wait until the context is
- * descheduled */
- wake_up(&js_kctx_info->ctx.is_scheduled_wait);
- /* Re-check for suspending: a suspend could've occurred, and all the
- * contexts could've been removed from the runpool before we took this
- * lock. In this case, we don't want to allow this context to run jobs,
- * we just want it out immediately.
- *
- * The DMB required to read the suspend flag was issued recently as part
- * of the runpool_irq locking. If a suspend occurs *after* that lock was
- * taken (i.e. this condition doesn't execute), then the
- * kbasep_js_suspend() code will cleanup this context instead (by virtue
- * of it being called strictly after the suspend flag is set, and will
- * wait for this lock to drop) */
- if (kbase_pm_is_suspending(kbdev)) {
- /* Cause it to leave at some later point */
- bool retained;
- retained = kbasep_js_runpool_retain_ctx_nolock(kbdev, kctx);
- KBASE_DEBUG_ASSERT(retained);
- kbasep_js_clear_submit_allowed(js_devdata, kctx);
- kctx_suspended = true;
- }
- /* Transaction complete */
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- mutex_unlock(&new_address_space->transaction_mutex);
- /* Synchronize with any policy timers */
- kbase_backend_ctx_count_changed(kbdev);
- mutex_unlock(&js_devdata->runpool_mutex);
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- /* Note: after this point, the context could potentially get scheduled
- * out immediately */
- if (kctx_suspended) {
- /* Finishing forcing out the context due to a suspend. Use a
- * variant of kbasep_js_runpool_release_ctx() that doesn't
- * schedule a new context, to prevent a risk of recursion back
- * into this function */
- kbasep_js_runpool_release_ctx_no_schedule(kbdev, kctx);
- return false;
- }
- return true;
- }
- static bool kbase_js_use_ctx(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- struct kbasep_js_device_data *js_devdata = &kbdev->js_data;
- unsigned long flags;
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- if (kctx->as_pending) {
- /* Context waiting for AS to be assigned */
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- return false;
- }
- if (kbase_backend_use_ctx_sched(kbdev, kctx)) {
- /* Context already has ASID - mark as active */
- kbdev->hwaccess.active_kctx = kctx;
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- return true; /* Context already scheduled */
- }
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- return kbasep_js_schedule_ctx(kbdev, kctx);
- }
- void kbasep_js_schedule_privileged_ctx(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- struct kbasep_js_kctx_info *js_kctx_info;
- struct kbasep_js_device_data *js_devdata;
- bool is_scheduled;
- KBASE_DEBUG_ASSERT(kbdev != NULL);
- KBASE_DEBUG_ASSERT(kctx != NULL);
- js_devdata = &kbdev->js_data;
- js_kctx_info = &kctx->jctx.sched_info;
- /* This must never be attempted whilst suspending - i.e. it should only
- * happen in response to a syscall from a user-space thread */
- BUG_ON(kbase_pm_is_suspending(kbdev));
- mutex_lock(&js_devdata->queue_mutex);
- mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
- /* Mark the context as privileged */
- js_kctx_info->ctx.flags |= KBASE_CTX_FLAG_PRIVILEGED;
- is_scheduled = js_kctx_info->ctx.is_scheduled;
- if (!is_scheduled) {
- /* Add the context to the pullable list */
- if (kbase_js_ctx_list_add_pullable(kbdev, kctx, 0))
- kbase_js_sync_timers(kbdev);
- /* Fast-starting requires the jsctx_mutex to be dropped,
- * because it works on multiple ctxs */
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_unlock(&js_devdata->queue_mutex);
- /* Try to schedule the context in */
- kbase_js_sched_all(kbdev);
- /* Wait for the context to be scheduled in */
- wait_event(kctx->jctx.sched_info.ctx.is_scheduled_wait,
- kctx->jctx.sched_info.ctx.is_scheduled);
- } else {
- /* Already scheduled in - We need to retain it to keep the
- * corresponding address space */
- kbasep_js_runpool_retain_ctx(kbdev, kctx);
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_unlock(&js_devdata->queue_mutex);
- }
- }
- KBASE_EXPORT_TEST_API(kbasep_js_schedule_privileged_ctx);
- void kbasep_js_release_privileged_ctx(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- struct kbasep_js_kctx_info *js_kctx_info;
- bool pending;
- KBASE_DEBUG_ASSERT(kctx != NULL);
- js_kctx_info = &kctx->jctx.sched_info;
- /* We don't need to use the address space anymore */
- mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
- js_kctx_info->ctx.flags &= (~KBASE_CTX_FLAG_PRIVILEGED);
- pending = kctx->as_pending;
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- /* Release the context - it will be scheduled out if there is no
- * pending job */
- if (!pending)
- kbasep_js_runpool_release_ctx(kbdev, kctx);
- kbase_js_sched_all(kbdev);
- }
- KBASE_EXPORT_TEST_API(kbasep_js_release_privileged_ctx);
- void kbasep_js_suspend(struct kbase_device *kbdev)
- {
- unsigned long flags;
- struct kbasep_js_device_data *js_devdata;
- int i;
- u16 retained = 0u;
- int nr_privileged_ctx = 0;
- KBASE_DEBUG_ASSERT(kbdev);
- KBASE_DEBUG_ASSERT(kbase_pm_is_suspending(kbdev));
- js_devdata = &kbdev->js_data;
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- /* Prevent all contexts from submitting */
- js_devdata->runpool_irq.submit_allowed = 0;
- /* Retain each of the contexts, so we can cause it to leave even if it
- * had no refcount to begin with */
- for (i = BASE_MAX_NR_AS - 1; i >= 0; --i) {
- struct kbasep_js_per_as_data *js_per_as_data =
- &js_devdata->runpool_irq.per_as_data[i];
- struct kbase_context *kctx = js_per_as_data->kctx;
- retained = retained << 1;
- if (kctx) {
- ++(js_per_as_data->as_busy_refcount);
- retained |= 1u;
- /* We can only cope with up to 1 privileged context -
- * the instrumented context. It'll be suspended by
- * disabling instrumentation */
- if (kctx->jctx.sched_info.ctx.flags &
- KBASE_CTX_FLAG_PRIVILEGED)
- KBASE_DEBUG_ASSERT(++nr_privileged_ctx == 1);
- }
- }
- CSTD_UNUSED(nr_privileged_ctx);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- /* De-ref the previous retain to ensure each context gets pulled out
- * sometime later. */
- for (i = 0;
- i < BASE_MAX_NR_AS;
- ++i, retained = retained >> 1) {
- struct kbasep_js_per_as_data *js_per_as_data =
- &js_devdata->runpool_irq.per_as_data[i];
- struct kbase_context *kctx = js_per_as_data->kctx;
- if (retained & 1u)
- kbasep_js_runpool_release_ctx(kbdev, kctx);
- }
- /* Caller must wait for all Power Manager active references to be
- * dropped */
- }
- void kbasep_js_resume(struct kbase_device *kbdev)
- {
- struct kbasep_js_device_data *js_devdata;
- int js;
- KBASE_DEBUG_ASSERT(kbdev);
- js_devdata = &kbdev->js_data;
- KBASE_DEBUG_ASSERT(!kbase_pm_is_suspending(kbdev));
- mutex_lock(&js_devdata->queue_mutex);
- for (js = 0; js < kbdev->gpu_props.num_job_slots; js++) {
- struct kbase_context *kctx, *n;
- list_for_each_entry_safe(kctx, n,
- &kbdev->js_data.ctx_list_unpullable[js],
- jctx.sched_info.ctx.ctx_list_entry[js]) {
- struct kbasep_js_kctx_info *js_kctx_info;
- unsigned long flags;
- bool timer_sync = false;
- js_kctx_info = &kctx->jctx.sched_info;
- mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_lock(&js_devdata->runpool_mutex);
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- if (!js_kctx_info->ctx.is_scheduled &&
- kbase_js_ctx_pullable(kctx, js, false))
- timer_sync = kbase_js_ctx_list_add_pullable(
- kbdev, kctx, js);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock,
- flags);
- if (timer_sync)
- kbase_backend_ctx_count_changed(kbdev);
- mutex_unlock(&js_devdata->runpool_mutex);
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- }
- }
- mutex_unlock(&js_devdata->queue_mutex);
- /* Restart atom processing */
- kbase_js_sched_all(kbdev);
- /* JS Resume complete */
- }
- bool kbase_js_is_atom_valid(struct kbase_device *kbdev,
- struct kbase_jd_atom *katom)
- {
- if ((katom->core_req & BASE_JD_REQ_FS) &&
- (katom->core_req & (BASE_JD_REQ_CS | BASE_JD_REQ_ONLY_COMPUTE |
- BASE_JD_REQ_T)))
- return false;
- if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8987) &&
- (katom->core_req & BASE_JD_REQ_ONLY_COMPUTE) &&
- (katom->core_req & (BASE_JD_REQ_CS | BASE_JD_REQ_T)))
- return false;
- return true;
- }
- static int kbase_js_get_slot(struct kbase_device *kbdev,
- struct kbase_jd_atom *katom)
- {
- if (katom->core_req & BASE_JD_REQ_FS)
- return 0;
- if (katom->core_req & BASE_JD_REQ_ONLY_COMPUTE) {
- if (katom->device_nr == 1 &&
- kbdev->gpu_props.num_core_groups == 2)
- return 2;
- if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8987))
- return 2;
- }
- return 1;
- }
- bool kbase_js_dep_resolved_submit(struct kbase_context *kctx,
- struct kbase_jd_atom *katom)
- {
- bool enqueue_required;
- katom->slot_nr = kbase_js_get_slot(kctx->kbdev, katom);
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- lockdep_assert_held(&kctx->jctx.lock);
- /* If slot will transition from unpullable to pullable then add to
- * pullable list */
- if (jsctx_rb_none_to_pull(kctx, katom->slot_nr)) {
- enqueue_required = true;
- } else {
- enqueue_required = false;
- }
- /* Check if there are lower priority jobs to soft stop */
- kbase_job_slot_ctx_priority_check_locked(kctx, katom);
- /* Add atom to ring buffer. */
- if (jsctx_rb_add_atom(kctx, katom)) {
- jsctx_ll_add(kctx, katom);
- enqueue_required = false;
- katom->atom_flags |= KBASE_KATOM_FLAG_JSCTX_IN_LL;
- } else {
- katom->atom_flags |= KBASE_KATOM_FLAG_JSCTX_RB_SUBMITTED;
- }
- return enqueue_required;
- }
- /**
- * kbase_js_evict_deps - Evict dependencies of a failed atom.
- * @kctx: Context pointer
- * @katom: Pointer to the atom that has failed.
- * @js: The job slot the katom was run on.
- * @prio: Priority of the katom.
- *
- * Remove all post dependencies of an atom from the context ringbuffers.
- *
- * The original atom's event_code will be propogated to all dependent atoms.
- *
- * Context: Caller must hold the HW access lock
- */
- static void kbase_js_evict_deps(struct kbase_context *kctx,
- struct kbase_jd_atom *katom, int js, int prio)
- {
- struct kbase_jd_atom *x_dep = katom->x_post_dep;
- struct kbase_jd_atom *next_katom = jsctx_rb_peek_prio(kctx, js, prio);
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- if (next_katom &&
- (next_katom->atom_flags & KBASE_KATOM_FLAG_FAIL_PREV)) {
- KBASE_DEBUG_ASSERT(next_katom->status !=
- KBASE_JD_ATOM_STATE_HW_COMPLETED);
- next_katom->will_fail_event_code = katom->event_code;
- }
- /* Has cross slot depenency. */
- if (x_dep && (x_dep->atom_flags & (KBASE_KATOM_FLAG_JSCTX_IN_LL |
- KBASE_KATOM_FLAG_JSCTX_RB_SUBMITTED))) {
- /* Remove dependency.*/
- x_dep->atom_flags &= ~KBASE_KATOM_FLAG_X_DEP_BLOCKED;
- /* Fail if it had a data dependency. */
- if (x_dep->atom_flags & KBASE_KATOM_FLAG_FAIL_BLOCKER) {
- x_dep->will_fail_event_code = katom->event_code;
- }
- }
- }
- struct kbase_jd_atom *kbase_js_pull(struct kbase_context *kctx, int js)
- {
- struct kbase_jd_atom *katom;
- struct kbasep_js_device_data *js_devdata;
- int pulled;
- KBASE_DEBUG_ASSERT(kctx);
- js_devdata = &kctx->kbdev->js_data;
- lockdep_assert_held(&js_devdata->runpool_irq.lock);
- if (!kbasep_js_is_submit_allowed(js_devdata, kctx))
- return NULL;
- if (kbase_pm_is_suspending(kctx->kbdev))
- return NULL;
- katom = jsctx_rb_peek(kctx, js);
- if (!katom)
- return NULL;
- if (atomic_read(&katom->blocked))
- return NULL;
- /* Due to ordering restrictions when unpulling atoms on failure, we do
- * not allow multiple runs of fail-dep atoms from the same context to be
- * present on the same slot */
- if ((katom->atom_flags & KBASE_KATOM_FLAG_FAIL_PREV) &&
- atomic_read(&kctx->atoms_pulled_slot[js])) {
- struct kbase_jd_atom *prev_atom =
- kbase_backend_inspect_tail(kctx->kbdev, js);
- if (prev_atom && prev_atom->kctx != kctx)
- return NULL;
- }
- if (katom->atom_flags & KBASE_KATOM_FLAG_X_DEP_BLOCKED) {
- if (katom->x_pre_dep->gpu_rb_state ==
- KBASE_ATOM_GPU_RB_NOT_IN_SLOT_RB ||
- katom->x_pre_dep->will_fail_event_code)
- return NULL;
- if ((katom->atom_flags & KBASE_KATOM_FLAG_FAIL_BLOCKER) &&
- kbase_backend_nr_atoms_on_slot(kctx->kbdev, js))
- return NULL;
- }
- kctx->pulled = true;
- pulled = atomic_inc_return(&kctx->atoms_pulled);
- if (pulled == 1 && !kctx->slots_pullable) {
- WARN_ON(kctx->ctx_runnable_ref);
- kctx->ctx_runnable_ref = true;
- atomic_inc(&kctx->kbdev->js_data.nr_contexts_runnable);
- }
- atomic_inc(&kctx->atoms_pulled_slot[katom->slot_nr]);
- jsctx_rb_pull(kctx, katom);
- kbasep_js_runpool_retain_ctx_nolock(kctx->kbdev, kctx);
- katom->atom_flags |= KBASE_KATOM_FLAG_HOLDING_CTX_REF;
- katom->sched_info.cfs.ticks = 0;
- return katom;
- }
- static void js_return_worker(struct work_struct *data)
- {
- struct kbase_jd_atom *katom = container_of(data, struct kbase_jd_atom,
- work);
- struct kbase_context *kctx = katom->kctx;
- struct kbase_device *kbdev = kctx->kbdev;
- struct kbasep_js_device_data *js_devdata = &kbdev->js_data;
- struct kbasep_js_kctx_info *js_kctx_info = &kctx->jctx.sched_info;
- struct kbasep_js_atom_retained_state retained_state;
- int js = katom->slot_nr;
- bool timer_sync = false;
- bool context_idle = false;
- unsigned long flags;
- base_jd_core_req core_req = katom->core_req;
- u64 affinity = katom->affinity;
- enum kbase_atom_coreref_state coreref_state = katom->coreref_state;
- kbase_tlstream_aux_job_softstop_ex(katom);
- kbase_backend_complete_wq(kbdev, katom);
- if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8316))
- kbase_as_poking_timer_release_atom(kbdev, kctx, katom);
- kbasep_js_atom_retained_state_copy(&retained_state, katom);
- mutex_lock(&js_devdata->queue_mutex);
- mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
- atomic_dec(&kctx->atoms_pulled);
- atomic_dec(&kctx->atoms_pulled_slot[js]);
- atomic_dec(&katom->blocked);
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- if (!atomic_read(&kctx->atoms_pulled_slot[js]) &&
- jsctx_rb_none_to_pull(kctx, js))
- timer_sync |= kbase_js_ctx_list_remove(kbdev, kctx, js);
- if (!atomic_read(&kctx->atoms_pulled)) {
- if (!kctx->slots_pullable) {
- WARN_ON(!kctx->ctx_runnable_ref);
- kctx->ctx_runnable_ref = false;
- atomic_dec(&kbdev->js_data.nr_contexts_runnable);
- timer_sync = true;
- }
- if (kctx->as_nr != KBASEP_AS_NR_INVALID &&
- !js_kctx_info->ctx.is_dying) {
- int num_slots = kbdev->gpu_props.num_job_slots;
- int slot;
- if (!kbasep_js_is_submit_allowed(js_devdata, kctx))
- kbasep_js_set_submit_allowed(js_devdata, kctx);
- for (slot = 0; slot < num_slots; slot++) {
- if (kbase_js_ctx_pullable(kctx, slot, true))
- timer_sync |=
- kbase_js_ctx_list_add_pullable(
- kbdev, kctx, slot);
- }
- }
- kbase_jm_idle_ctx(kbdev, kctx);
- context_idle = true;
- }
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- if (context_idle) {
- WARN_ON(!kctx->ctx_active);
- kctx->ctx_active = false;
- kbase_pm_context_idle(kbdev);
- }
- if (timer_sync)
- kbase_js_sync_timers(kbdev);
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_unlock(&js_devdata->queue_mutex);
- katom->atom_flags &= ~KBASE_KATOM_FLAG_HOLDING_CTX_REF;
- kbasep_js_runpool_release_ctx_and_katom_retained_state(kbdev, kctx,
- &retained_state);
- kbase_js_sched_all(kbdev);
- kbase_backend_complete_wq_post_sched(kbdev, core_req, affinity,
- coreref_state);
- }
- void kbase_js_unpull(struct kbase_context *kctx, struct kbase_jd_atom *katom)
- {
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- jsctx_rb_unpull(kctx, katom);
- WARN_ON(work_pending(&katom->work));
- /* Block re-submission until workqueue has run */
- atomic_inc(&katom->blocked);
- kbase_job_check_leave_disjoint(kctx->kbdev, katom);
- KBASE_DEBUG_ASSERT(0 == object_is_on_stack(&katom->work));
- INIT_WORK(&katom->work, js_return_worker);
- queue_work(kctx->jctx.job_done_wq, &katom->work);
- }
- bool kbase_js_complete_atom_wq(struct kbase_context *kctx,
- struct kbase_jd_atom *katom)
- {
- struct kbasep_js_kctx_info *js_kctx_info;
- struct kbasep_js_device_data *js_devdata;
- struct kbase_device *kbdev;
- unsigned long flags;
- bool timer_sync = false;
- int atom_slot;
- bool context_idle = false;
- kbdev = kctx->kbdev;
- atom_slot = katom->slot_nr;
- js_kctx_info = &kctx->jctx.sched_info;
- js_devdata = &kbdev->js_data;
- lockdep_assert_held(&js_kctx_info->ctx.jsctx_mutex);
- mutex_lock(&js_devdata->runpool_mutex);
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- if (katom->atom_flags & KBASE_KATOM_FLAG_JSCTX_RB_SUBMITTED) {
- jsctx_rb_remove(kctx, katom);
- context_idle = !atomic_dec_return(&kctx->atoms_pulled);
- atomic_dec(&kctx->atoms_pulled_slot[atom_slot]);
- if (!atomic_read(&kctx->atoms_pulled) &&
- !kctx->slots_pullable) {
- WARN_ON(!kctx->ctx_runnable_ref);
- kctx->ctx_runnable_ref = false;
- atomic_dec(&kbdev->js_data.nr_contexts_runnable);
- timer_sync = true;
- }
- }
- WARN_ON(katom->atom_flags & KBASE_KATOM_FLAG_JSCTX_IN_LL);
- if (!atomic_read(&kctx->atoms_pulled_slot[atom_slot]) &&
- jsctx_rb_none_to_pull(kctx, atom_slot))
- timer_sync |= kbase_js_ctx_list_remove(kctx->kbdev, kctx,
- atom_slot);
- /*
- * If submission is disabled on this context (most likely due to an
- * atom failure) and there are now no atoms left in the system then
- * re-enable submission so that context can be scheduled again.
- */
- if (!kbasep_js_is_submit_allowed(js_devdata, kctx) &&
- !atomic_read(&kctx->atoms_pulled) &&
- !js_kctx_info->ctx.is_dying) {
- int js;
- kbasep_js_set_submit_allowed(js_devdata, kctx);
- for (js = 0; js < kbdev->gpu_props.num_job_slots; js++) {
- if (kbase_js_ctx_pullable(kctx, js, true))
- timer_sync |= kbase_js_ctx_list_add_pullable(
- kbdev, kctx, js);
- }
- } else if (katom->x_post_dep &&
- kbasep_js_is_submit_allowed(js_devdata, kctx)) {
- int js;
- for (js = 0; js < kbdev->gpu_props.num_job_slots; js++) {
- if (kbase_js_ctx_pullable(kctx, js, true))
- timer_sync |= kbase_js_ctx_list_add_pullable(
- kbdev, kctx, js);
- }
- }
- /* Mark context as inactive. The pm reference will be dropped later in
- * jd_done_worker().
- */
- if (context_idle)
- kctx->ctx_active = false;
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- if (timer_sync)
- kbase_backend_ctx_count_changed(kbdev);
- mutex_unlock(&js_devdata->runpool_mutex);
- return context_idle;
- }
- void kbase_js_complete_atom(struct kbase_jd_atom *katom, ktime_t *end_timestamp)
- {
- u64 microseconds_spent = 0;
- struct kbase_device *kbdev;
- struct kbase_context *kctx = katom->kctx;
- union kbasep_js_policy *js_policy;
- struct kbasep_js_device_data *js_devdata;
- kbdev = kctx->kbdev;
- js_policy = &kbdev->js_data.policy;
- js_devdata = &kbdev->js_data;
- lockdep_assert_held(&kctx->kbdev->js_data.runpool_irq.lock);
- if (katom->will_fail_event_code)
- katom->event_code = katom->will_fail_event_code;
- katom->status = KBASE_JD_ATOM_STATE_HW_COMPLETED;
- if (katom->event_code != BASE_JD_EVENT_DONE) {
- kbase_js_evict_deps(kctx, katom, katom->slot_nr,
- katom->sched_priority);
- }
- #if defined(CONFIG_MALI_GATOR_SUPPORT)
- kbase_trace_mali_job_slots_event(GATOR_MAKE_EVENT(GATOR_JOB_SLOT_STOP,
- katom->slot_nr), NULL, 0);
- #endif
- kbase_tlstream_tl_nret_atom_lpu(
- katom,
- &kbdev->gpu_props.props.raw_props.js_features[
- katom->slot_nr]);
- kbase_tlstream_tl_nret_atom_as(katom, &kbdev->as[kctx->as_nr]);
- kbase_tlstream_tl_nret_ctx_lpu(
- kctx,
- &kbdev->gpu_props.props.raw_props.js_features[
- katom->slot_nr]);
- /* Calculate the job's time used */
- if (end_timestamp != NULL) {
- /* Only calculating it for jobs that really run on the HW (e.g.
- * removed from next jobs never actually ran, so really did take
- * zero time) */
- ktime_t tick_diff = ktime_sub(*end_timestamp,
- katom->start_timestamp);
- microseconds_spent = ktime_to_ns(tick_diff);
- do_div(microseconds_spent, 1000);
- /* Round up time spent to the minimum timer resolution */
- if (microseconds_spent < KBASEP_JS_TICK_RESOLUTION_US)
- microseconds_spent = KBASEP_JS_TICK_RESOLUTION_US;
- }
- /* Log the result of the job (completion status, and time spent). */
- kbasep_js_policy_log_job_result(js_policy, katom, microseconds_spent);
- kbase_jd_done(katom, katom->slot_nr, end_timestamp, 0);
- /* Unblock cross dependency if present */
- if (katom->x_post_dep && (katom->event_code == BASE_JD_EVENT_DONE ||
- !(katom->x_post_dep->atom_flags &
- KBASE_KATOM_FLAG_FAIL_BLOCKER)))
- katom->x_post_dep->atom_flags &=
- ~KBASE_KATOM_FLAG_X_DEP_BLOCKED;
- }
- void kbase_js_sched(struct kbase_device *kbdev, int js_mask)
- {
- struct kbasep_js_device_data *js_devdata;
- union kbasep_js_policy *js_policy;
- bool timer_sync = false;
- js_devdata = &kbdev->js_data;
- js_policy = &js_devdata->policy;
- down(&js_devdata->schedule_sem);
- mutex_lock(&js_devdata->queue_mutex);
- while (js_mask) {
- int js;
- js = ffs(js_mask) - 1;
- while (1) {
- struct kbase_context *kctx;
- unsigned long flags;
- bool context_idle = false;
- kctx = kbase_js_ctx_list_pop_head(kbdev, js);
- if (!kctx) {
- js_mask &= ~(1 << js);
- break; /* No contexts on pullable list */
- }
- if (!kctx->ctx_active) {
- context_idle = true;
- if (kbase_pm_context_active_handle_suspend(
- kbdev,
- KBASE_PM_SUSPEND_HANDLER_DONT_INCREASE)) {
- /* Suspend pending - return context to
- * queue and stop scheduling */
- mutex_lock(
- &kctx->jctx.sched_info.ctx.jsctx_mutex);
- if (kbase_js_ctx_list_add_pullable_head(
- kctx->kbdev, kctx, js))
- kbase_js_sync_timers(kbdev);
- mutex_unlock(
- &kctx->jctx.sched_info.ctx.jsctx_mutex);
- mutex_unlock(&js_devdata->queue_mutex);
- up(&js_devdata->schedule_sem);
- return;
- }
- kctx->ctx_active = true;
- }
- if (!kbase_js_use_ctx(kbdev, kctx)) {
- mutex_lock(
- &kctx->jctx.sched_info.ctx.jsctx_mutex);
- /* Context can not be used at this time */
- spin_lock_irqsave(&js_devdata->runpool_irq.lock,
- flags);
- if (kctx->as_pending ||
- kbase_js_ctx_pullable(kctx, js, false)
- || (kctx->jctx.sched_info.ctx.flags &
- KBASE_CTX_FLAG_PRIVILEGED))
- timer_sync |=
- kbase_js_ctx_list_add_pullable_head(
- kctx->kbdev, kctx, js);
- else
- timer_sync |=
- kbase_js_ctx_list_add_unpullable(
- kctx->kbdev, kctx, js);
- spin_unlock_irqrestore(
- &js_devdata->runpool_irq.lock, flags);
- mutex_unlock(
- &kctx->jctx.sched_info.ctx.jsctx_mutex);
- if (context_idle) {
- WARN_ON(!kctx->ctx_active);
- kctx->ctx_active = false;
- kbase_pm_context_idle(kbdev);
- }
- /* No more jobs can be submitted on this slot */
- js_mask &= ~(1 << js);
- break;
- }
- mutex_lock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- kctx->pulled = false;
- if (!kbase_jm_kick(kbdev, 1 << js))
- /* No more jobs can be submitted on this slot */
- js_mask &= ~(1 << js);
- if (!kctx->pulled) {
- /* Failed to pull jobs - push to head of list */
- if (kbase_js_ctx_pullable(kctx, js, true))
- timer_sync |=
- kbase_js_ctx_list_add_pullable_head(
- kctx->kbdev,
- kctx, js);
- else
- timer_sync |=
- kbase_js_ctx_list_add_unpullable(
- kctx->kbdev,
- kctx, js);
- if (context_idle) {
- kbase_jm_idle_ctx(kbdev, kctx);
- spin_unlock_irqrestore(
- &js_devdata->runpool_irq.lock,
- flags);
- WARN_ON(!kctx->ctx_active);
- kctx->ctx_active = false;
- kbase_pm_context_idle(kbdev);
- } else {
- spin_unlock_irqrestore(
- &js_devdata->runpool_irq.lock,
- flags);
- }
- mutex_unlock(
- &kctx->jctx.sched_info.ctx.jsctx_mutex);
- js_mask &= ~(1 << js);
- break; /* Could not run atoms on this slot */
- }
- /* Push to back of list */
- if (kbase_js_ctx_pullable(kctx, js, true))
- timer_sync |= kbase_js_ctx_list_add_pullable(
- kctx->kbdev, kctx, js);
- else
- timer_sync |= kbase_js_ctx_list_add_unpullable(
- kctx->kbdev, kctx, js);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock,
- flags);
- mutex_unlock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
- }
- }
- if (timer_sync)
- kbase_js_sync_timers(kbdev);
- mutex_unlock(&js_devdata->queue_mutex);
- up(&js_devdata->schedule_sem);
- }
- void kbase_js_zap_context(struct kbase_context *kctx)
- {
- struct kbase_device *kbdev = kctx->kbdev;
- struct kbasep_js_device_data *js_devdata = &kbdev->js_data;
- struct kbasep_js_kctx_info *js_kctx_info = &kctx->jctx.sched_info;
- int js;
- /*
- * Critical assumption: No more submission is possible outside of the
- * workqueue. This is because the OS *must* prevent U/K calls (IOCTLs)
- * whilst the struct kbase_context is terminating.
- */
- /* First, atomically do the following:
- * - mark the context as dying
- * - try to evict it from the policy queue */
- mutex_lock(&js_devdata->queue_mutex);
- mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
- js_kctx_info->ctx.is_dying = true;
- dev_dbg(kbdev->dev, "Zap: Try Evict Ctx %p", kctx);
- /*
- * At this point we know:
- * - If eviction succeeded, it was in the policy queue, but now no
- * longer is
- * - We must cancel the jobs here. No Power Manager active reference to
- * release.
- * - This happens asynchronously - kbase_jd_zap_context() will wait for
- * those jobs to be killed.
- * - If eviction failed, then it wasn't in the policy queue. It is one
- * of the following:
- * - a. it didn't have any jobs, and so is not in the Policy Queue or
- * the Run Pool (not scheduled)
- * - Hence, no more work required to cancel jobs. No Power Manager
- * active reference to release.
- * - b. it was in the middle of a scheduling transaction (and thus must
- * have at least 1 job). This can happen from a syscall or a
- * kernel thread. We still hold the jsctx_mutex, and so the thread
- * must be waiting inside kbasep_js_try_schedule_head_ctx(),
- * before checking whether the runpool is full. That thread will
- * continue after we drop the mutex, and will notice the context
- * is dying. It will rollback the transaction, killing all jobs at
- * the same time. kbase_jd_zap_context() will wait for those jobs
- * to be killed.
- * - Hence, no more work required to cancel jobs, or to release the
- * Power Manager active reference.
- * - c. it is scheduled, and may or may not be running jobs
- * - We must cause it to leave the runpool by stopping it from
- * submitting any more jobs. When it finally does leave,
- * kbasep_js_runpool_requeue_or_kill_ctx() will kill all remaining jobs
- * (because it is dying), release the Power Manager active reference,
- * and will not requeue the context in the policy queue.
- * kbase_jd_zap_context() will wait for those jobs to be killed.
- * - Hence, work required just to make it leave the runpool. Cancelling
- * jobs and releasing the Power manager active reference will be
- * handled when it leaves the runpool.
- */
- if (!js_kctx_info->ctx.is_scheduled) {
- for (js = 0; js < kbdev->gpu_props.num_job_slots; js++) {
- if (!list_empty(
- &kctx->jctx.sched_info.ctx.ctx_list_entry[js]))
- list_del_init(
- &kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
- }
- /* The following events require us to kill off remaining jobs
- * and update PM book-keeping:
- * - we evicted it correctly (it must have jobs to be in the
- * Policy Queue)
- *
- * These events need no action, but take this path anyway:
- * - Case a: it didn't have any jobs, and was never in the Queue
- * - Case b: scheduling transaction will be partially rolled-
- * back (this already cancels the jobs)
- */
- KBASE_TRACE_ADD(kbdev, JM_ZAP_NON_SCHEDULED, kctx, NULL, 0u,
- js_kctx_info->ctx.is_scheduled);
- dev_dbg(kbdev->dev, "Zap: Ctx %p scheduled=0", kctx);
- /* Only cancel jobs when we evicted from the policy
- * queue. No Power Manager active reference was held.
- *
- * Having is_dying set ensures that this kills, and
- * doesn't requeue */
- kbasep_js_runpool_requeue_or_kill_ctx(kbdev, kctx, false);
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_unlock(&js_devdata->queue_mutex);
- } else {
- unsigned long flags;
- bool was_retained;
- /* Case c: didn't evict, but it is scheduled - it's in the Run
- * Pool */
- KBASE_TRACE_ADD(kbdev, JM_ZAP_SCHEDULED, kctx, NULL, 0u,
- js_kctx_info->ctx.is_scheduled);
- dev_dbg(kbdev->dev, "Zap: Ctx %p is in RunPool", kctx);
- /* Disable the ctx from submitting any more jobs */
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- kbasep_js_clear_submit_allowed(js_devdata, kctx);
- /* Retain and (later) release the context whilst it is is now
- * disallowed from submitting jobs - ensures that someone
- * somewhere will be removing the context later on */
- was_retained = kbasep_js_runpool_retain_ctx_nolock(kbdev, kctx);
- /* Since it's scheduled and we have the jsctx_mutex, it must be
- * retained successfully */
- KBASE_DEBUG_ASSERT(was_retained);
- dev_dbg(kbdev->dev, "Zap: Ctx %p Kill Any Running jobs", kctx);
- /* Cancel any remaining running jobs for this kctx - if any.
- * Submit is disallowed which takes effect immediately, so no
- * more new jobs will appear after we do this. */
- for (js = 0; js < kbdev->gpu_props.num_job_slots; js++)
- kbase_job_slot_hardstop(kctx, js, NULL);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
- mutex_unlock(&js_devdata->queue_mutex);
- dev_dbg(kbdev->dev, "Zap: Ctx %p Release (may or may not schedule out immediately)",
- kctx);
- kbasep_js_runpool_release_ctx(kbdev, kctx);
- }
- KBASE_TRACE_ADD(kbdev, JM_ZAP_DONE, kctx, NULL, 0u, 0u);
- /* After this, you must wait on both the
- * kbase_jd_context::zero_jobs_wait and the
- * kbasep_js_kctx_info::ctx::is_scheduled_waitq - to wait for the jobs
- * to be destroyed, and the context to be de-scheduled (if it was on the
- * runpool).
- *
- * kbase_jd_zap_context() will do this. */
- }
- static inline int trace_get_refcnt(struct kbase_device *kbdev,
- struct kbase_context *kctx)
- {
- struct kbasep_js_device_data *js_devdata;
- int as_nr;
- int refcnt = 0;
- js_devdata = &kbdev->js_data;
- as_nr = kctx->as_nr;
- if (as_nr != KBASEP_AS_NR_INVALID) {
- struct kbasep_js_per_as_data *js_per_as_data;
- js_per_as_data = &js_devdata->runpool_irq.per_as_data[as_nr];
- refcnt = js_per_as_data->as_busy_refcount;
- }
- return refcnt;
- }
- /**
- * kbase_js_foreach_ctx_job(): - Call a function on all jobs in context
- * @kctx: Pointer to context.
- * @callback: Pointer to function to call for each job.
- *
- * Call a function on all jobs belonging to a non-queued, non-running
- * context, and detach the jobs from the context as it goes.
- *
- * Due to the locks that might be held at the time of the call, the callback
- * may need to defer work on a workqueue to complete its actions (e.g. when
- * cancelling jobs)
- *
- * Atoms will be removed from the queue, so this must only be called when
- * cancelling jobs (which occurs as part of context destruction).
- *
- * The locking conditions on the caller are as follows:
- * - it will be holding kbasep_js_kctx_info::ctx::jsctx_mutex.
- */
- static void kbase_js_foreach_ctx_job(struct kbase_context *kctx,
- kbasep_js_policy_ctx_job_cb callback)
- {
- struct kbase_device *kbdev;
- struct kbasep_js_device_data *js_devdata;
- unsigned long flags;
- u32 js;
- kbdev = kctx->kbdev;
- js_devdata = &kbdev->js_data;
- spin_lock_irqsave(&js_devdata->runpool_irq.lock, flags);
- KBASE_TRACE_ADD_REFCOUNT(kbdev, JS_POLICY_FOREACH_CTX_JOBS, kctx, NULL,
- 0u, trace_get_refcnt(kbdev, kctx));
- /* Invoke callback on jobs on each slot in turn */
- for (js = 0; js < kbdev->gpu_props.num_job_slots; js++)
- jsctx_queue_foreach(kctx, js, callback);
- spin_unlock_irqrestore(&js_devdata->runpool_irq.lock, flags);
- }
|