amdgpu_device.c 96 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963296429652966296729682969297029712972297329742975297629772978297929802981298229832984298529862987298829892990299129922993299429952996299729982999300030013002300330043005300630073008300930103011301230133014301530163017301830193020302130223023302430253026302730283029303030313032303330343035303630373038303930403041304230433044304530463047304830493050305130523053305430553056305730583059306030613062306330643065306630673068306930703071307230733074307530763077307830793080308130823083308430853086308730883089309030913092309330943095309630973098309931003101310231033104310531063107310831093110311131123113311431153116311731183119312031213122312331243125312631273128312931303131313231333134313531363137313831393140314131423143314431453146314731483149315031513152315331543155315631573158315931603161316231633164316531663167316831693170317131723173317431753176317731783179318031813182318331843185318631873188318931903191319231933194319531963197319831993200320132023203320432053206320732083209321032113212321332143215321632173218321932203221322232233224322532263227322832293230323132323233323432353236323732383239324032413242324332443245324632473248324932503251325232533254325532563257325832593260326132623263326432653266326732683269327032713272327332743275327632773278327932803281328232833284328532863287328832893290329132923293329432953296329732983299330033013302330333043305330633073308330933103311331233133314331533163317331833193320332133223323332433253326332733283329333033313332333333343335333633373338333933403341334233433344334533463347334833493350335133523353335433553356335733583359336033613362336333643365336633673368336933703371337233733374337533763377337833793380338133823383338433853386338733883389339033913392339333943395339633973398339934003401340234033404340534063407340834093410341134123413341434153416341734183419342034213422342334243425342634273428342934303431343234333434343534363437343834393440344134423443344434453446344734483449345034513452345334543455345634573458345934603461346234633464346534663467346834693470347134723473347434753476347734783479348034813482348334843485348634873488348934903491349234933494349534963497349834993500350135023503350435053506350735083509351035113512351335143515351635173518351935203521352235233524352535263527352835293530
  1. /*
  2. * Copyright 2008 Advanced Micro Devices, Inc.
  3. * Copyright 2008 Red Hat Inc.
  4. * Copyright 2009 Jerome Glisse.
  5. *
  6. * Permission is hereby granted, free of charge, to any person obtaining a
  7. * copy of this software and associated documentation files (the "Software"),
  8. * to deal in the Software without restriction, including without limitation
  9. * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10. * and/or sell copies of the Software, and to permit persons to whom the
  11. * Software is furnished to do so, subject to the following conditions:
  12. *
  13. * The above copyright notice and this permission notice shall be included in
  14. * all copies or substantial portions of the Software.
  15. *
  16. * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17. * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18. * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19. * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20. * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21. * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22. * OTHER DEALINGS IN THE SOFTWARE.
  23. *
  24. * Authors: Dave Airlie
  25. * Alex Deucher
  26. * Jerome Glisse
  27. */
  28. #include <linux/power_supply.h>
  29. #include <linux/kthread.h>
  30. #include <linux/console.h>
  31. #include <linux/slab.h>
  32. #include <drm/drmP.h>
  33. #include <drm/drm_crtc_helper.h>
  34. #include <drm/drm_atomic_helper.h>
  35. #include <drm/amdgpu_drm.h>
  36. #include <linux/vgaarb.h>
  37. #include <linux/vga_switcheroo.h>
  38. #include <linux/efi.h>
  39. #include "amdgpu.h"
  40. #include "amdgpu_trace.h"
  41. #include "amdgpu_i2c.h"
  42. #include "atom.h"
  43. #include "amdgpu_atombios.h"
  44. #include "amdgpu_atomfirmware.h"
  45. #include "amd_pcie.h"
  46. #ifdef CONFIG_DRM_AMDGPU_SI
  47. #include "si.h"
  48. #endif
  49. #ifdef CONFIG_DRM_AMDGPU_CIK
  50. #include "cik.h"
  51. #endif
  52. #include "vi.h"
  53. #include "soc15.h"
  54. #include "bif/bif_4_1_d.h"
  55. #include <linux/pci.h>
  56. #include <linux/firmware.h>
  57. #include "amdgpu_vf_error.h"
  58. #include "amdgpu_amdkfd.h"
  59. #include "amdgpu_pm.h"
  60. MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
  61. MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
  62. MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
  63. #define AMDGPU_RESUME_MS 2000
  64. static const char *amdgpu_asic_name[] = {
  65. "TAHITI",
  66. "PITCAIRN",
  67. "VERDE",
  68. "OLAND",
  69. "HAINAN",
  70. "BONAIRE",
  71. "KAVERI",
  72. "KABINI",
  73. "HAWAII",
  74. "MULLINS",
  75. "TOPAZ",
  76. "TONGA",
  77. "FIJI",
  78. "CARRIZO",
  79. "STONEY",
  80. "POLARIS10",
  81. "POLARIS11",
  82. "POLARIS12",
  83. "VEGAM",
  84. "VEGA10",
  85. "VEGA12",
  86. "VEGA20",
  87. "RAVEN",
  88. "LAST",
  89. };
  90. static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
  91. /**
  92. * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
  93. *
  94. * @dev: drm_device pointer
  95. *
  96. * Returns true if the device is a dGPU with HG/PX power control,
  97. * otherwise return false.
  98. */
  99. bool amdgpu_device_is_px(struct drm_device *dev)
  100. {
  101. struct amdgpu_device *adev = dev->dev_private;
  102. if (adev->flags & AMD_IS_PX)
  103. return true;
  104. return false;
  105. }
  106. /*
  107. * MMIO register access helper functions.
  108. */
  109. /**
  110. * amdgpu_mm_rreg - read a memory mapped IO register
  111. *
  112. * @adev: amdgpu_device pointer
  113. * @reg: dword aligned register offset
  114. * @acc_flags: access flags which require special behavior
  115. *
  116. * Returns the 32 bit value from the offset specified.
  117. */
  118. uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
  119. uint32_t acc_flags)
  120. {
  121. uint32_t ret;
  122. if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
  123. return amdgpu_virt_kiq_rreg(adev, reg);
  124. if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
  125. ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
  126. else {
  127. unsigned long flags;
  128. spin_lock_irqsave(&adev->mmio_idx_lock, flags);
  129. writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
  130. ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
  131. spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
  132. }
  133. trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
  134. return ret;
  135. }
  136. /*
  137. * MMIO register read with bytes helper functions
  138. * @offset:bytes offset from MMIO start
  139. *
  140. */
  141. /**
  142. * amdgpu_mm_rreg8 - read a memory mapped IO register
  143. *
  144. * @adev: amdgpu_device pointer
  145. * @offset: byte aligned register offset
  146. *
  147. * Returns the 8 bit value from the offset specified.
  148. */
  149. uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
  150. if (offset < adev->rmmio_size)
  151. return (readb(adev->rmmio + offset));
  152. BUG();
  153. }
  154. /*
  155. * MMIO register write with bytes helper functions
  156. * @offset:bytes offset from MMIO start
  157. * @value: the value want to be written to the register
  158. *
  159. */
  160. /**
  161. * amdgpu_mm_wreg8 - read a memory mapped IO register
  162. *
  163. * @adev: amdgpu_device pointer
  164. * @offset: byte aligned register offset
  165. * @value: 8 bit value to write
  166. *
  167. * Writes the value specified to the offset specified.
  168. */
  169. void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
  170. if (offset < adev->rmmio_size)
  171. writeb(value, adev->rmmio + offset);
  172. else
  173. BUG();
  174. }
  175. /**
  176. * amdgpu_mm_wreg - write to a memory mapped IO register
  177. *
  178. * @adev: amdgpu_device pointer
  179. * @reg: dword aligned register offset
  180. * @v: 32 bit value to write to the register
  181. * @acc_flags: access flags which require special behavior
  182. *
  183. * Writes the value specified to the offset specified.
  184. */
  185. void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
  186. uint32_t acc_flags)
  187. {
  188. trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
  189. if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
  190. adev->last_mm_index = v;
  191. }
  192. if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
  193. return amdgpu_virt_kiq_wreg(adev, reg, v);
  194. if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
  195. writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
  196. else {
  197. unsigned long flags;
  198. spin_lock_irqsave(&adev->mmio_idx_lock, flags);
  199. writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
  200. writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
  201. spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
  202. }
  203. if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
  204. udelay(500);
  205. }
  206. }
  207. /**
  208. * amdgpu_io_rreg - read an IO register
  209. *
  210. * @adev: amdgpu_device pointer
  211. * @reg: dword aligned register offset
  212. *
  213. * Returns the 32 bit value from the offset specified.
  214. */
  215. u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
  216. {
  217. if ((reg * 4) < adev->rio_mem_size)
  218. return ioread32(adev->rio_mem + (reg * 4));
  219. else {
  220. iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
  221. return ioread32(adev->rio_mem + (mmMM_DATA * 4));
  222. }
  223. }
  224. /**
  225. * amdgpu_io_wreg - write to an IO register
  226. *
  227. * @adev: amdgpu_device pointer
  228. * @reg: dword aligned register offset
  229. * @v: 32 bit value to write to the register
  230. *
  231. * Writes the value specified to the offset specified.
  232. */
  233. void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
  234. {
  235. if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
  236. adev->last_mm_index = v;
  237. }
  238. if ((reg * 4) < adev->rio_mem_size)
  239. iowrite32(v, adev->rio_mem + (reg * 4));
  240. else {
  241. iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
  242. iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
  243. }
  244. if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
  245. udelay(500);
  246. }
  247. }
  248. /**
  249. * amdgpu_mm_rdoorbell - read a doorbell dword
  250. *
  251. * @adev: amdgpu_device pointer
  252. * @index: doorbell index
  253. *
  254. * Returns the value in the doorbell aperture at the
  255. * requested doorbell index (CIK).
  256. */
  257. u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
  258. {
  259. if (index < adev->doorbell.num_doorbells) {
  260. return readl(adev->doorbell.ptr + index);
  261. } else {
  262. DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
  263. return 0;
  264. }
  265. }
  266. /**
  267. * amdgpu_mm_wdoorbell - write a doorbell dword
  268. *
  269. * @adev: amdgpu_device pointer
  270. * @index: doorbell index
  271. * @v: value to write
  272. *
  273. * Writes @v to the doorbell aperture at the
  274. * requested doorbell index (CIK).
  275. */
  276. void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
  277. {
  278. if (index < adev->doorbell.num_doorbells) {
  279. writel(v, adev->doorbell.ptr + index);
  280. } else {
  281. DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
  282. }
  283. }
  284. /**
  285. * amdgpu_mm_rdoorbell64 - read a doorbell Qword
  286. *
  287. * @adev: amdgpu_device pointer
  288. * @index: doorbell index
  289. *
  290. * Returns the value in the doorbell aperture at the
  291. * requested doorbell index (VEGA10+).
  292. */
  293. u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
  294. {
  295. if (index < adev->doorbell.num_doorbells) {
  296. return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
  297. } else {
  298. DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
  299. return 0;
  300. }
  301. }
  302. /**
  303. * amdgpu_mm_wdoorbell64 - write a doorbell Qword
  304. *
  305. * @adev: amdgpu_device pointer
  306. * @index: doorbell index
  307. * @v: value to write
  308. *
  309. * Writes @v to the doorbell aperture at the
  310. * requested doorbell index (VEGA10+).
  311. */
  312. void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
  313. {
  314. if (index < adev->doorbell.num_doorbells) {
  315. atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
  316. } else {
  317. DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
  318. }
  319. }
  320. /**
  321. * amdgpu_invalid_rreg - dummy reg read function
  322. *
  323. * @adev: amdgpu device pointer
  324. * @reg: offset of register
  325. *
  326. * Dummy register read function. Used for register blocks
  327. * that certain asics don't have (all asics).
  328. * Returns the value in the register.
  329. */
  330. static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
  331. {
  332. DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
  333. BUG();
  334. return 0;
  335. }
  336. /**
  337. * amdgpu_invalid_wreg - dummy reg write function
  338. *
  339. * @adev: amdgpu device pointer
  340. * @reg: offset of register
  341. * @v: value to write to the register
  342. *
  343. * Dummy register read function. Used for register blocks
  344. * that certain asics don't have (all asics).
  345. */
  346. static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
  347. {
  348. DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
  349. reg, v);
  350. BUG();
  351. }
  352. /**
  353. * amdgpu_block_invalid_rreg - dummy reg read function
  354. *
  355. * @adev: amdgpu device pointer
  356. * @block: offset of instance
  357. * @reg: offset of register
  358. *
  359. * Dummy register read function. Used for register blocks
  360. * that certain asics don't have (all asics).
  361. * Returns the value in the register.
  362. */
  363. static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
  364. uint32_t block, uint32_t reg)
  365. {
  366. DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
  367. reg, block);
  368. BUG();
  369. return 0;
  370. }
  371. /**
  372. * amdgpu_block_invalid_wreg - dummy reg write function
  373. *
  374. * @adev: amdgpu device pointer
  375. * @block: offset of instance
  376. * @reg: offset of register
  377. * @v: value to write to the register
  378. *
  379. * Dummy register read function. Used for register blocks
  380. * that certain asics don't have (all asics).
  381. */
  382. static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
  383. uint32_t block,
  384. uint32_t reg, uint32_t v)
  385. {
  386. DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
  387. reg, block, v);
  388. BUG();
  389. }
  390. /**
  391. * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
  392. *
  393. * @adev: amdgpu device pointer
  394. *
  395. * Allocates a scratch page of VRAM for use by various things in the
  396. * driver.
  397. */
  398. static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
  399. {
  400. return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
  401. PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
  402. &adev->vram_scratch.robj,
  403. &adev->vram_scratch.gpu_addr,
  404. (void **)&adev->vram_scratch.ptr);
  405. }
  406. /**
  407. * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
  408. *
  409. * @adev: amdgpu device pointer
  410. *
  411. * Frees the VRAM scratch page.
  412. */
  413. static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
  414. {
  415. amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
  416. }
  417. /**
  418. * amdgpu_device_program_register_sequence - program an array of registers.
  419. *
  420. * @adev: amdgpu_device pointer
  421. * @registers: pointer to the register array
  422. * @array_size: size of the register array
  423. *
  424. * Programs an array or registers with and and or masks.
  425. * This is a helper for setting golden registers.
  426. */
  427. void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
  428. const u32 *registers,
  429. const u32 array_size)
  430. {
  431. u32 tmp, reg, and_mask, or_mask;
  432. int i;
  433. if (array_size % 3)
  434. return;
  435. for (i = 0; i < array_size; i +=3) {
  436. reg = registers[i + 0];
  437. and_mask = registers[i + 1];
  438. or_mask = registers[i + 2];
  439. if (and_mask == 0xffffffff) {
  440. tmp = or_mask;
  441. } else {
  442. tmp = RREG32(reg);
  443. tmp &= ~and_mask;
  444. tmp |= or_mask;
  445. }
  446. WREG32(reg, tmp);
  447. }
  448. }
  449. /**
  450. * amdgpu_device_pci_config_reset - reset the GPU
  451. *
  452. * @adev: amdgpu_device pointer
  453. *
  454. * Resets the GPU using the pci config reset sequence.
  455. * Only applicable to asics prior to vega10.
  456. */
  457. void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
  458. {
  459. pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
  460. }
  461. /*
  462. * GPU doorbell aperture helpers function.
  463. */
  464. /**
  465. * amdgpu_device_doorbell_init - Init doorbell driver information.
  466. *
  467. * @adev: amdgpu_device pointer
  468. *
  469. * Init doorbell driver information (CIK)
  470. * Returns 0 on success, error on failure.
  471. */
  472. static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
  473. {
  474. /* No doorbell on SI hardware generation */
  475. if (adev->asic_type < CHIP_BONAIRE) {
  476. adev->doorbell.base = 0;
  477. adev->doorbell.size = 0;
  478. adev->doorbell.num_doorbells = 0;
  479. adev->doorbell.ptr = NULL;
  480. return 0;
  481. }
  482. if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
  483. return -EINVAL;
  484. /* doorbell bar mapping */
  485. adev->doorbell.base = pci_resource_start(adev->pdev, 2);
  486. adev->doorbell.size = pci_resource_len(adev->pdev, 2);
  487. adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
  488. AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
  489. if (adev->doorbell.num_doorbells == 0)
  490. return -EINVAL;
  491. adev->doorbell.ptr = ioremap(adev->doorbell.base,
  492. adev->doorbell.num_doorbells *
  493. sizeof(u32));
  494. if (adev->doorbell.ptr == NULL)
  495. return -ENOMEM;
  496. return 0;
  497. }
  498. /**
  499. * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
  500. *
  501. * @adev: amdgpu_device pointer
  502. *
  503. * Tear down doorbell driver information (CIK)
  504. */
  505. static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
  506. {
  507. iounmap(adev->doorbell.ptr);
  508. adev->doorbell.ptr = NULL;
  509. }
  510. /*
  511. * amdgpu_device_wb_*()
  512. * Writeback is the method by which the GPU updates special pages in memory
  513. * with the status of certain GPU events (fences, ring pointers,etc.).
  514. */
  515. /**
  516. * amdgpu_device_wb_fini - Disable Writeback and free memory
  517. *
  518. * @adev: amdgpu_device pointer
  519. *
  520. * Disables Writeback and frees the Writeback memory (all asics).
  521. * Used at driver shutdown.
  522. */
  523. static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
  524. {
  525. if (adev->wb.wb_obj) {
  526. amdgpu_bo_free_kernel(&adev->wb.wb_obj,
  527. &adev->wb.gpu_addr,
  528. (void **)&adev->wb.wb);
  529. adev->wb.wb_obj = NULL;
  530. }
  531. }
  532. /**
  533. * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
  534. *
  535. * @adev: amdgpu_device pointer
  536. *
  537. * Initializes writeback and allocates writeback memory (all asics).
  538. * Used at driver startup.
  539. * Returns 0 on success or an -error on failure.
  540. */
  541. static int amdgpu_device_wb_init(struct amdgpu_device *adev)
  542. {
  543. int r;
  544. if (adev->wb.wb_obj == NULL) {
  545. /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
  546. r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
  547. PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
  548. &adev->wb.wb_obj, &adev->wb.gpu_addr,
  549. (void **)&adev->wb.wb);
  550. if (r) {
  551. dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
  552. return r;
  553. }
  554. adev->wb.num_wb = AMDGPU_MAX_WB;
  555. memset(&adev->wb.used, 0, sizeof(adev->wb.used));
  556. /* clear wb memory */
  557. memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
  558. }
  559. return 0;
  560. }
  561. /**
  562. * amdgpu_device_wb_get - Allocate a wb entry
  563. *
  564. * @adev: amdgpu_device pointer
  565. * @wb: wb index
  566. *
  567. * Allocate a wb slot for use by the driver (all asics).
  568. * Returns 0 on success or -EINVAL on failure.
  569. */
  570. int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
  571. {
  572. unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
  573. if (offset < adev->wb.num_wb) {
  574. __set_bit(offset, adev->wb.used);
  575. *wb = offset << 3; /* convert to dw offset */
  576. return 0;
  577. } else {
  578. return -EINVAL;
  579. }
  580. }
  581. /**
  582. * amdgpu_device_wb_free - Free a wb entry
  583. *
  584. * @adev: amdgpu_device pointer
  585. * @wb: wb index
  586. *
  587. * Free a wb slot allocated for use by the driver (all asics)
  588. */
  589. void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
  590. {
  591. wb >>= 3;
  592. if (wb < adev->wb.num_wb)
  593. __clear_bit(wb, adev->wb.used);
  594. }
  595. /**
  596. * amdgpu_device_vram_location - try to find VRAM location
  597. *
  598. * @adev: amdgpu device structure holding all necessary informations
  599. * @mc: memory controller structure holding memory informations
  600. * @base: base address at which to put VRAM
  601. *
  602. * Function will try to place VRAM at base address provided
  603. * as parameter.
  604. */
  605. void amdgpu_device_vram_location(struct amdgpu_device *adev,
  606. struct amdgpu_gmc *mc, u64 base)
  607. {
  608. uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
  609. mc->vram_start = base;
  610. mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
  611. if (limit && limit < mc->real_vram_size)
  612. mc->real_vram_size = limit;
  613. dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
  614. mc->mc_vram_size >> 20, mc->vram_start,
  615. mc->vram_end, mc->real_vram_size >> 20);
  616. }
  617. /**
  618. * amdgpu_device_gart_location - try to find GART location
  619. *
  620. * @adev: amdgpu device structure holding all necessary informations
  621. * @mc: memory controller structure holding memory informations
  622. *
  623. * Function will place try to place GART before or after VRAM.
  624. *
  625. * If GART size is bigger than space left then we ajust GART size.
  626. * Thus function will never fails.
  627. */
  628. void amdgpu_device_gart_location(struct amdgpu_device *adev,
  629. struct amdgpu_gmc *mc)
  630. {
  631. u64 size_af, size_bf;
  632. mc->gart_size += adev->pm.smu_prv_buffer_size;
  633. size_af = adev->gmc.mc_mask - mc->vram_end;
  634. size_bf = mc->vram_start;
  635. if (size_bf > size_af) {
  636. if (mc->gart_size > size_bf) {
  637. dev_warn(adev->dev, "limiting GART\n");
  638. mc->gart_size = size_bf;
  639. }
  640. mc->gart_start = 0;
  641. } else {
  642. if (mc->gart_size > size_af) {
  643. dev_warn(adev->dev, "limiting GART\n");
  644. mc->gart_size = size_af;
  645. }
  646. /* VCE doesn't like it when BOs cross a 4GB segment, so align
  647. * the GART base on a 4GB boundary as well.
  648. */
  649. mc->gart_start = ALIGN(mc->vram_end + 1, 0x100000000ULL);
  650. }
  651. mc->gart_end = mc->gart_start + mc->gart_size - 1;
  652. dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
  653. mc->gart_size >> 20, mc->gart_start, mc->gart_end);
  654. }
  655. /**
  656. * amdgpu_device_resize_fb_bar - try to resize FB BAR
  657. *
  658. * @adev: amdgpu_device pointer
  659. *
  660. * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
  661. * to fail, but if any of the BARs is not accessible after the size we abort
  662. * driver loading by returning -ENODEV.
  663. */
  664. int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
  665. {
  666. u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
  667. u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
  668. struct pci_bus *root;
  669. struct resource *res;
  670. unsigned i;
  671. u16 cmd;
  672. int r;
  673. /* Bypass for VF */
  674. if (amdgpu_sriov_vf(adev))
  675. return 0;
  676. /* Check if the root BUS has 64bit memory resources */
  677. root = adev->pdev->bus;
  678. while (root->parent)
  679. root = root->parent;
  680. pci_bus_for_each_resource(root, res, i) {
  681. if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
  682. res->start > 0x100000000ull)
  683. break;
  684. }
  685. /* Trying to resize is pointless without a root hub window above 4GB */
  686. if (!res)
  687. return 0;
  688. /* Disable memory decoding while we change the BAR addresses and size */
  689. pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
  690. pci_write_config_word(adev->pdev, PCI_COMMAND,
  691. cmd & ~PCI_COMMAND_MEMORY);
  692. /* Free the VRAM and doorbell BAR, we most likely need to move both. */
  693. amdgpu_device_doorbell_fini(adev);
  694. if (adev->asic_type >= CHIP_BONAIRE)
  695. pci_release_resource(adev->pdev, 2);
  696. pci_release_resource(adev->pdev, 0);
  697. r = pci_resize_resource(adev->pdev, 0, rbar_size);
  698. if (r == -ENOSPC)
  699. DRM_INFO("Not enough PCI address space for a large BAR.");
  700. else if (r && r != -ENOTSUPP)
  701. DRM_ERROR("Problem resizing BAR0 (%d).", r);
  702. pci_assign_unassigned_bus_resources(adev->pdev->bus);
  703. /* When the doorbell or fb BAR isn't available we have no chance of
  704. * using the device.
  705. */
  706. r = amdgpu_device_doorbell_init(adev);
  707. if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
  708. return -ENODEV;
  709. pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
  710. return 0;
  711. }
  712. /*
  713. * GPU helpers function.
  714. */
  715. /**
  716. * amdgpu_device_need_post - check if the hw need post or not
  717. *
  718. * @adev: amdgpu_device pointer
  719. *
  720. * Check if the asic has been initialized (all asics) at driver startup
  721. * or post is needed if hw reset is performed.
  722. * Returns true if need or false if not.
  723. */
  724. bool amdgpu_device_need_post(struct amdgpu_device *adev)
  725. {
  726. uint32_t reg;
  727. if (amdgpu_sriov_vf(adev))
  728. return false;
  729. if (amdgpu_passthrough(adev)) {
  730. /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
  731. * some old smc fw still need driver do vPost otherwise gpu hang, while
  732. * those smc fw version above 22.15 doesn't have this flaw, so we force
  733. * vpost executed for smc version below 22.15
  734. */
  735. if (adev->asic_type == CHIP_FIJI) {
  736. int err;
  737. uint32_t fw_ver;
  738. err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
  739. /* force vPost if error occured */
  740. if (err)
  741. return true;
  742. fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
  743. if (fw_ver < 0x00160e00)
  744. return true;
  745. }
  746. }
  747. if (adev->has_hw_reset) {
  748. adev->has_hw_reset = false;
  749. return true;
  750. }
  751. /* bios scratch used on CIK+ */
  752. if (adev->asic_type >= CHIP_BONAIRE)
  753. return amdgpu_atombios_scratch_need_asic_init(adev);
  754. /* check MEM_SIZE for older asics */
  755. reg = amdgpu_asic_get_config_memsize(adev);
  756. if ((reg != 0) && (reg != 0xffffffff))
  757. return false;
  758. return true;
  759. }
  760. /* if we get transitioned to only one device, take VGA back */
  761. /**
  762. * amdgpu_device_vga_set_decode - enable/disable vga decode
  763. *
  764. * @cookie: amdgpu_device pointer
  765. * @state: enable/disable vga decode
  766. *
  767. * Enable/disable vga decode (all asics).
  768. * Returns VGA resource flags.
  769. */
  770. static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
  771. {
  772. struct amdgpu_device *adev = cookie;
  773. amdgpu_asic_set_vga_state(adev, state);
  774. if (state)
  775. return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
  776. VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
  777. else
  778. return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
  779. }
  780. /**
  781. * amdgpu_device_check_block_size - validate the vm block size
  782. *
  783. * @adev: amdgpu_device pointer
  784. *
  785. * Validates the vm block size specified via module parameter.
  786. * The vm block size defines number of bits in page table versus page directory,
  787. * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
  788. * page table and the remaining bits are in the page directory.
  789. */
  790. static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
  791. {
  792. /* defines number of bits in page table versus page directory,
  793. * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
  794. * page table and the remaining bits are in the page directory */
  795. if (amdgpu_vm_block_size == -1)
  796. return;
  797. if (amdgpu_vm_block_size < 9) {
  798. dev_warn(adev->dev, "VM page table size (%d) too small\n",
  799. amdgpu_vm_block_size);
  800. amdgpu_vm_block_size = -1;
  801. }
  802. }
  803. /**
  804. * amdgpu_device_check_vm_size - validate the vm size
  805. *
  806. * @adev: amdgpu_device pointer
  807. *
  808. * Validates the vm size in GB specified via module parameter.
  809. * The VM size is the size of the GPU virtual memory space in GB.
  810. */
  811. static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
  812. {
  813. /* no need to check the default value */
  814. if (amdgpu_vm_size == -1)
  815. return;
  816. if (amdgpu_vm_size < 1) {
  817. dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
  818. amdgpu_vm_size);
  819. amdgpu_vm_size = -1;
  820. }
  821. }
  822. static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
  823. {
  824. struct sysinfo si;
  825. bool is_os_64 = (sizeof(void *) == 8) ? true : false;
  826. uint64_t total_memory;
  827. uint64_t dram_size_seven_GB = 0x1B8000000;
  828. uint64_t dram_size_three_GB = 0xB8000000;
  829. if (amdgpu_smu_memory_pool_size == 0)
  830. return;
  831. if (!is_os_64) {
  832. DRM_WARN("Not 64-bit OS, feature not supported\n");
  833. goto def_value;
  834. }
  835. si_meminfo(&si);
  836. total_memory = (uint64_t)si.totalram * si.mem_unit;
  837. if ((amdgpu_smu_memory_pool_size == 1) ||
  838. (amdgpu_smu_memory_pool_size == 2)) {
  839. if (total_memory < dram_size_three_GB)
  840. goto def_value1;
  841. } else if ((amdgpu_smu_memory_pool_size == 4) ||
  842. (amdgpu_smu_memory_pool_size == 8)) {
  843. if (total_memory < dram_size_seven_GB)
  844. goto def_value1;
  845. } else {
  846. DRM_WARN("Smu memory pool size not supported\n");
  847. goto def_value;
  848. }
  849. adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
  850. return;
  851. def_value1:
  852. DRM_WARN("No enough system memory\n");
  853. def_value:
  854. adev->pm.smu_prv_buffer_size = 0;
  855. }
  856. /**
  857. * amdgpu_device_check_arguments - validate module params
  858. *
  859. * @adev: amdgpu_device pointer
  860. *
  861. * Validates certain module parameters and updates
  862. * the associated values used by the driver (all asics).
  863. */
  864. static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
  865. {
  866. if (amdgpu_sched_jobs < 4) {
  867. dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
  868. amdgpu_sched_jobs);
  869. amdgpu_sched_jobs = 4;
  870. } else if (!is_power_of_2(amdgpu_sched_jobs)){
  871. dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
  872. amdgpu_sched_jobs);
  873. amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
  874. }
  875. if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
  876. /* gart size must be greater or equal to 32M */
  877. dev_warn(adev->dev, "gart size (%d) too small\n",
  878. amdgpu_gart_size);
  879. amdgpu_gart_size = -1;
  880. }
  881. if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
  882. /* gtt size must be greater or equal to 32M */
  883. dev_warn(adev->dev, "gtt size (%d) too small\n",
  884. amdgpu_gtt_size);
  885. amdgpu_gtt_size = -1;
  886. }
  887. /* valid range is between 4 and 9 inclusive */
  888. if (amdgpu_vm_fragment_size != -1 &&
  889. (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
  890. dev_warn(adev->dev, "valid range is between 4 and 9\n");
  891. amdgpu_vm_fragment_size = -1;
  892. }
  893. amdgpu_device_check_smu_prv_buffer_size(adev);
  894. amdgpu_device_check_vm_size(adev);
  895. amdgpu_device_check_block_size(adev);
  896. if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
  897. !is_power_of_2(amdgpu_vram_page_split))) {
  898. dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
  899. amdgpu_vram_page_split);
  900. amdgpu_vram_page_split = 1024;
  901. }
  902. if (amdgpu_lockup_timeout == 0) {
  903. dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
  904. amdgpu_lockup_timeout = 10000;
  905. }
  906. adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
  907. }
  908. /**
  909. * amdgpu_switcheroo_set_state - set switcheroo state
  910. *
  911. * @pdev: pci dev pointer
  912. * @state: vga_switcheroo state
  913. *
  914. * Callback for the switcheroo driver. Suspends or resumes the
  915. * the asics before or after it is powered up using ACPI methods.
  916. */
  917. static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
  918. {
  919. struct drm_device *dev = pci_get_drvdata(pdev);
  920. if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
  921. return;
  922. if (state == VGA_SWITCHEROO_ON) {
  923. pr_info("amdgpu: switched on\n");
  924. /* don't suspend or resume card normally */
  925. dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
  926. amdgpu_device_resume(dev, true, true);
  927. dev->switch_power_state = DRM_SWITCH_POWER_ON;
  928. drm_kms_helper_poll_enable(dev);
  929. } else {
  930. pr_info("amdgpu: switched off\n");
  931. drm_kms_helper_poll_disable(dev);
  932. dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
  933. amdgpu_device_suspend(dev, true, true);
  934. dev->switch_power_state = DRM_SWITCH_POWER_OFF;
  935. }
  936. }
  937. /**
  938. * amdgpu_switcheroo_can_switch - see if switcheroo state can change
  939. *
  940. * @pdev: pci dev pointer
  941. *
  942. * Callback for the switcheroo driver. Check of the switcheroo
  943. * state can be changed.
  944. * Returns true if the state can be changed, false if not.
  945. */
  946. static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
  947. {
  948. struct drm_device *dev = pci_get_drvdata(pdev);
  949. /*
  950. * FIXME: open_count is protected by drm_global_mutex but that would lead to
  951. * locking inversion with the driver load path. And the access here is
  952. * completely racy anyway. So don't bother with locking for now.
  953. */
  954. return dev->open_count == 0;
  955. }
  956. static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
  957. .set_gpu_state = amdgpu_switcheroo_set_state,
  958. .reprobe = NULL,
  959. .can_switch = amdgpu_switcheroo_can_switch,
  960. };
  961. /**
  962. * amdgpu_device_ip_set_clockgating_state - set the CG state
  963. *
  964. * @dev: amdgpu_device pointer
  965. * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
  966. * @state: clockgating state (gate or ungate)
  967. *
  968. * Sets the requested clockgating state for all instances of
  969. * the hardware IP specified.
  970. * Returns the error code from the last instance.
  971. */
  972. int amdgpu_device_ip_set_clockgating_state(void *dev,
  973. enum amd_ip_block_type block_type,
  974. enum amd_clockgating_state state)
  975. {
  976. struct amdgpu_device *adev = dev;
  977. int i, r = 0;
  978. for (i = 0; i < adev->num_ip_blocks; i++) {
  979. if (!adev->ip_blocks[i].status.valid)
  980. continue;
  981. if (adev->ip_blocks[i].version->type != block_type)
  982. continue;
  983. if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
  984. continue;
  985. r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
  986. (void *)adev, state);
  987. if (r)
  988. DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
  989. adev->ip_blocks[i].version->funcs->name, r);
  990. }
  991. return r;
  992. }
  993. /**
  994. * amdgpu_device_ip_set_powergating_state - set the PG state
  995. *
  996. * @dev: amdgpu_device pointer
  997. * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
  998. * @state: powergating state (gate or ungate)
  999. *
  1000. * Sets the requested powergating state for all instances of
  1001. * the hardware IP specified.
  1002. * Returns the error code from the last instance.
  1003. */
  1004. int amdgpu_device_ip_set_powergating_state(void *dev,
  1005. enum amd_ip_block_type block_type,
  1006. enum amd_powergating_state state)
  1007. {
  1008. struct amdgpu_device *adev = dev;
  1009. int i, r = 0;
  1010. for (i = 0; i < adev->num_ip_blocks; i++) {
  1011. if (!adev->ip_blocks[i].status.valid)
  1012. continue;
  1013. if (adev->ip_blocks[i].version->type != block_type)
  1014. continue;
  1015. if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
  1016. continue;
  1017. r = adev->ip_blocks[i].version->funcs->set_powergating_state(
  1018. (void *)adev, state);
  1019. if (r)
  1020. DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
  1021. adev->ip_blocks[i].version->funcs->name, r);
  1022. }
  1023. return r;
  1024. }
  1025. /**
  1026. * amdgpu_device_ip_get_clockgating_state - get the CG state
  1027. *
  1028. * @adev: amdgpu_device pointer
  1029. * @flags: clockgating feature flags
  1030. *
  1031. * Walks the list of IPs on the device and updates the clockgating
  1032. * flags for each IP.
  1033. * Updates @flags with the feature flags for each hardware IP where
  1034. * clockgating is enabled.
  1035. */
  1036. void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
  1037. u32 *flags)
  1038. {
  1039. int i;
  1040. for (i = 0; i < adev->num_ip_blocks; i++) {
  1041. if (!adev->ip_blocks[i].status.valid)
  1042. continue;
  1043. if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
  1044. adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
  1045. }
  1046. }
  1047. /**
  1048. * amdgpu_device_ip_wait_for_idle - wait for idle
  1049. *
  1050. * @adev: amdgpu_device pointer
  1051. * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
  1052. *
  1053. * Waits for the request hardware IP to be idle.
  1054. * Returns 0 for success or a negative error code on failure.
  1055. */
  1056. int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
  1057. enum amd_ip_block_type block_type)
  1058. {
  1059. int i, r;
  1060. for (i = 0; i < adev->num_ip_blocks; i++) {
  1061. if (!adev->ip_blocks[i].status.valid)
  1062. continue;
  1063. if (adev->ip_blocks[i].version->type == block_type) {
  1064. r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
  1065. if (r)
  1066. return r;
  1067. break;
  1068. }
  1069. }
  1070. return 0;
  1071. }
  1072. /**
  1073. * amdgpu_device_ip_is_idle - is the hardware IP idle
  1074. *
  1075. * @adev: amdgpu_device pointer
  1076. * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
  1077. *
  1078. * Check if the hardware IP is idle or not.
  1079. * Returns true if it the IP is idle, false if not.
  1080. */
  1081. bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
  1082. enum amd_ip_block_type block_type)
  1083. {
  1084. int i;
  1085. for (i = 0; i < adev->num_ip_blocks; i++) {
  1086. if (!adev->ip_blocks[i].status.valid)
  1087. continue;
  1088. if (adev->ip_blocks[i].version->type == block_type)
  1089. return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
  1090. }
  1091. return true;
  1092. }
  1093. /**
  1094. * amdgpu_device_ip_get_ip_block - get a hw IP pointer
  1095. *
  1096. * @adev: amdgpu_device pointer
  1097. * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
  1098. *
  1099. * Returns a pointer to the hardware IP block structure
  1100. * if it exists for the asic, otherwise NULL.
  1101. */
  1102. struct amdgpu_ip_block *
  1103. amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
  1104. enum amd_ip_block_type type)
  1105. {
  1106. int i;
  1107. for (i = 0; i < adev->num_ip_blocks; i++)
  1108. if (adev->ip_blocks[i].version->type == type)
  1109. return &adev->ip_blocks[i];
  1110. return NULL;
  1111. }
  1112. /**
  1113. * amdgpu_device_ip_block_version_cmp
  1114. *
  1115. * @adev: amdgpu_device pointer
  1116. * @type: enum amd_ip_block_type
  1117. * @major: major version
  1118. * @minor: minor version
  1119. *
  1120. * return 0 if equal or greater
  1121. * return 1 if smaller or the ip_block doesn't exist
  1122. */
  1123. int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
  1124. enum amd_ip_block_type type,
  1125. u32 major, u32 minor)
  1126. {
  1127. struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
  1128. if (ip_block && ((ip_block->version->major > major) ||
  1129. ((ip_block->version->major == major) &&
  1130. (ip_block->version->minor >= minor))))
  1131. return 0;
  1132. return 1;
  1133. }
  1134. /**
  1135. * amdgpu_device_ip_block_add
  1136. *
  1137. * @adev: amdgpu_device pointer
  1138. * @ip_block_version: pointer to the IP to add
  1139. *
  1140. * Adds the IP block driver information to the collection of IPs
  1141. * on the asic.
  1142. */
  1143. int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
  1144. const struct amdgpu_ip_block_version *ip_block_version)
  1145. {
  1146. if (!ip_block_version)
  1147. return -EINVAL;
  1148. DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
  1149. ip_block_version->funcs->name);
  1150. adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
  1151. return 0;
  1152. }
  1153. /**
  1154. * amdgpu_device_enable_virtual_display - enable virtual display feature
  1155. *
  1156. * @adev: amdgpu_device pointer
  1157. *
  1158. * Enabled the virtual display feature if the user has enabled it via
  1159. * the module parameter virtual_display. This feature provides a virtual
  1160. * display hardware on headless boards or in virtualized environments.
  1161. * This function parses and validates the configuration string specified by
  1162. * the user and configues the virtual display configuration (number of
  1163. * virtual connectors, crtcs, etc.) specified.
  1164. */
  1165. static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
  1166. {
  1167. adev->enable_virtual_display = false;
  1168. if (amdgpu_virtual_display) {
  1169. struct drm_device *ddev = adev->ddev;
  1170. const char *pci_address_name = pci_name(ddev->pdev);
  1171. char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
  1172. pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
  1173. pciaddstr_tmp = pciaddstr;
  1174. while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
  1175. pciaddname = strsep(&pciaddname_tmp, ",");
  1176. if (!strcmp("all", pciaddname)
  1177. || !strcmp(pci_address_name, pciaddname)) {
  1178. long num_crtc;
  1179. int res = -1;
  1180. adev->enable_virtual_display = true;
  1181. if (pciaddname_tmp)
  1182. res = kstrtol(pciaddname_tmp, 10,
  1183. &num_crtc);
  1184. if (!res) {
  1185. if (num_crtc < 1)
  1186. num_crtc = 1;
  1187. if (num_crtc > 6)
  1188. num_crtc = 6;
  1189. adev->mode_info.num_crtc = num_crtc;
  1190. } else {
  1191. adev->mode_info.num_crtc = 1;
  1192. }
  1193. break;
  1194. }
  1195. }
  1196. DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
  1197. amdgpu_virtual_display, pci_address_name,
  1198. adev->enable_virtual_display, adev->mode_info.num_crtc);
  1199. kfree(pciaddstr);
  1200. }
  1201. }
  1202. /**
  1203. * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
  1204. *
  1205. * @adev: amdgpu_device pointer
  1206. *
  1207. * Parses the asic configuration parameters specified in the gpu info
  1208. * firmware and makes them availale to the driver for use in configuring
  1209. * the asic.
  1210. * Returns 0 on success, -EINVAL on failure.
  1211. */
  1212. static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
  1213. {
  1214. const char *chip_name;
  1215. char fw_name[30];
  1216. int err;
  1217. const struct gpu_info_firmware_header_v1_0 *hdr;
  1218. adev->firmware.gpu_info_fw = NULL;
  1219. switch (adev->asic_type) {
  1220. case CHIP_TOPAZ:
  1221. case CHIP_TONGA:
  1222. case CHIP_FIJI:
  1223. case CHIP_POLARIS10:
  1224. case CHIP_POLARIS11:
  1225. case CHIP_POLARIS12:
  1226. case CHIP_VEGAM:
  1227. case CHIP_CARRIZO:
  1228. case CHIP_STONEY:
  1229. #ifdef CONFIG_DRM_AMDGPU_SI
  1230. case CHIP_VERDE:
  1231. case CHIP_TAHITI:
  1232. case CHIP_PITCAIRN:
  1233. case CHIP_OLAND:
  1234. case CHIP_HAINAN:
  1235. #endif
  1236. #ifdef CONFIG_DRM_AMDGPU_CIK
  1237. case CHIP_BONAIRE:
  1238. case CHIP_HAWAII:
  1239. case CHIP_KAVERI:
  1240. case CHIP_KABINI:
  1241. case CHIP_MULLINS:
  1242. #endif
  1243. case CHIP_VEGA20:
  1244. default:
  1245. return 0;
  1246. case CHIP_VEGA10:
  1247. chip_name = "vega10";
  1248. break;
  1249. case CHIP_VEGA12:
  1250. chip_name = "vega12";
  1251. break;
  1252. case CHIP_RAVEN:
  1253. chip_name = "raven";
  1254. break;
  1255. }
  1256. snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
  1257. err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
  1258. if (err) {
  1259. dev_err(adev->dev,
  1260. "Failed to load gpu_info firmware \"%s\"\n",
  1261. fw_name);
  1262. goto out;
  1263. }
  1264. err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
  1265. if (err) {
  1266. dev_err(adev->dev,
  1267. "Failed to validate gpu_info firmware \"%s\"\n",
  1268. fw_name);
  1269. goto out;
  1270. }
  1271. hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
  1272. amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
  1273. switch (hdr->version_major) {
  1274. case 1:
  1275. {
  1276. const struct gpu_info_firmware_v1_0 *gpu_info_fw =
  1277. (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
  1278. le32_to_cpu(hdr->header.ucode_array_offset_bytes));
  1279. adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
  1280. adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
  1281. adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
  1282. adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
  1283. adev->gfx.config.max_texture_channel_caches =
  1284. le32_to_cpu(gpu_info_fw->gc_num_tccs);
  1285. adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
  1286. adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
  1287. adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
  1288. adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
  1289. adev->gfx.config.double_offchip_lds_buf =
  1290. le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
  1291. adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
  1292. adev->gfx.cu_info.max_waves_per_simd =
  1293. le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
  1294. adev->gfx.cu_info.max_scratch_slots_per_cu =
  1295. le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
  1296. adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
  1297. break;
  1298. }
  1299. default:
  1300. dev_err(adev->dev,
  1301. "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
  1302. err = -EINVAL;
  1303. goto out;
  1304. }
  1305. out:
  1306. return err;
  1307. }
  1308. /**
  1309. * amdgpu_device_ip_early_init - run early init for hardware IPs
  1310. *
  1311. * @adev: amdgpu_device pointer
  1312. *
  1313. * Early initialization pass for hardware IPs. The hardware IPs that make
  1314. * up each asic are discovered each IP's early_init callback is run. This
  1315. * is the first stage in initializing the asic.
  1316. * Returns 0 on success, negative error code on failure.
  1317. */
  1318. static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
  1319. {
  1320. int i, r;
  1321. amdgpu_device_enable_virtual_display(adev);
  1322. switch (adev->asic_type) {
  1323. case CHIP_TOPAZ:
  1324. case CHIP_TONGA:
  1325. case CHIP_FIJI:
  1326. case CHIP_POLARIS10:
  1327. case CHIP_POLARIS11:
  1328. case CHIP_POLARIS12:
  1329. case CHIP_VEGAM:
  1330. case CHIP_CARRIZO:
  1331. case CHIP_STONEY:
  1332. if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
  1333. adev->family = AMDGPU_FAMILY_CZ;
  1334. else
  1335. adev->family = AMDGPU_FAMILY_VI;
  1336. r = vi_set_ip_blocks(adev);
  1337. if (r)
  1338. return r;
  1339. break;
  1340. #ifdef CONFIG_DRM_AMDGPU_SI
  1341. case CHIP_VERDE:
  1342. case CHIP_TAHITI:
  1343. case CHIP_PITCAIRN:
  1344. case CHIP_OLAND:
  1345. case CHIP_HAINAN:
  1346. adev->family = AMDGPU_FAMILY_SI;
  1347. r = si_set_ip_blocks(adev);
  1348. if (r)
  1349. return r;
  1350. break;
  1351. #endif
  1352. #ifdef CONFIG_DRM_AMDGPU_CIK
  1353. case CHIP_BONAIRE:
  1354. case CHIP_HAWAII:
  1355. case CHIP_KAVERI:
  1356. case CHIP_KABINI:
  1357. case CHIP_MULLINS:
  1358. if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
  1359. adev->family = AMDGPU_FAMILY_CI;
  1360. else
  1361. adev->family = AMDGPU_FAMILY_KV;
  1362. r = cik_set_ip_blocks(adev);
  1363. if (r)
  1364. return r;
  1365. break;
  1366. #endif
  1367. case CHIP_VEGA10:
  1368. case CHIP_VEGA12:
  1369. case CHIP_VEGA20:
  1370. case CHIP_RAVEN:
  1371. if (adev->asic_type == CHIP_RAVEN)
  1372. adev->family = AMDGPU_FAMILY_RV;
  1373. else
  1374. adev->family = AMDGPU_FAMILY_AI;
  1375. r = soc15_set_ip_blocks(adev);
  1376. if (r)
  1377. return r;
  1378. break;
  1379. default:
  1380. /* FIXME: not supported yet */
  1381. return -EINVAL;
  1382. }
  1383. r = amdgpu_device_parse_gpu_info_fw(adev);
  1384. if (r)
  1385. return r;
  1386. amdgpu_amdkfd_device_probe(adev);
  1387. if (amdgpu_sriov_vf(adev)) {
  1388. r = amdgpu_virt_request_full_gpu(adev, true);
  1389. if (r)
  1390. return -EAGAIN;
  1391. }
  1392. adev->powerplay.pp_feature = amdgpu_pp_feature_mask;
  1393. for (i = 0; i < adev->num_ip_blocks; i++) {
  1394. if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
  1395. DRM_ERROR("disabled ip block: %d <%s>\n",
  1396. i, adev->ip_blocks[i].version->funcs->name);
  1397. adev->ip_blocks[i].status.valid = false;
  1398. } else {
  1399. if (adev->ip_blocks[i].version->funcs->early_init) {
  1400. r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
  1401. if (r == -ENOENT) {
  1402. adev->ip_blocks[i].status.valid = false;
  1403. } else if (r) {
  1404. DRM_ERROR("early_init of IP block <%s> failed %d\n",
  1405. adev->ip_blocks[i].version->funcs->name, r);
  1406. return r;
  1407. } else {
  1408. adev->ip_blocks[i].status.valid = true;
  1409. }
  1410. } else {
  1411. adev->ip_blocks[i].status.valid = true;
  1412. }
  1413. }
  1414. }
  1415. adev->cg_flags &= amdgpu_cg_mask;
  1416. adev->pg_flags &= amdgpu_pg_mask;
  1417. return 0;
  1418. }
  1419. /**
  1420. * amdgpu_device_ip_init - run init for hardware IPs
  1421. *
  1422. * @adev: amdgpu_device pointer
  1423. *
  1424. * Main initialization pass for hardware IPs. The list of all the hardware
  1425. * IPs that make up the asic is walked and the sw_init and hw_init callbacks
  1426. * are run. sw_init initializes the software state associated with each IP
  1427. * and hw_init initializes the hardware associated with each IP.
  1428. * Returns 0 on success, negative error code on failure.
  1429. */
  1430. static int amdgpu_device_ip_init(struct amdgpu_device *adev)
  1431. {
  1432. int i, r;
  1433. for (i = 0; i < adev->num_ip_blocks; i++) {
  1434. if (!adev->ip_blocks[i].status.valid)
  1435. continue;
  1436. r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
  1437. if (r) {
  1438. DRM_ERROR("sw_init of IP block <%s> failed %d\n",
  1439. adev->ip_blocks[i].version->funcs->name, r);
  1440. return r;
  1441. }
  1442. adev->ip_blocks[i].status.sw = true;
  1443. /* need to do gmc hw init early so we can allocate gpu mem */
  1444. if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
  1445. r = amdgpu_device_vram_scratch_init(adev);
  1446. if (r) {
  1447. DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
  1448. return r;
  1449. }
  1450. r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
  1451. if (r) {
  1452. DRM_ERROR("hw_init %d failed %d\n", i, r);
  1453. return r;
  1454. }
  1455. r = amdgpu_device_wb_init(adev);
  1456. if (r) {
  1457. DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
  1458. return r;
  1459. }
  1460. adev->ip_blocks[i].status.hw = true;
  1461. /* right after GMC hw init, we create CSA */
  1462. if (amdgpu_sriov_vf(adev)) {
  1463. r = amdgpu_allocate_static_csa(adev);
  1464. if (r) {
  1465. DRM_ERROR("allocate CSA failed %d\n", r);
  1466. return r;
  1467. }
  1468. }
  1469. }
  1470. }
  1471. for (i = 0; i < adev->num_ip_blocks; i++) {
  1472. if (!adev->ip_blocks[i].status.sw)
  1473. continue;
  1474. if (adev->ip_blocks[i].status.hw)
  1475. continue;
  1476. r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
  1477. if (r) {
  1478. DRM_ERROR("hw_init of IP block <%s> failed %d\n",
  1479. adev->ip_blocks[i].version->funcs->name, r);
  1480. return r;
  1481. }
  1482. adev->ip_blocks[i].status.hw = true;
  1483. }
  1484. amdgpu_amdkfd_device_init(adev);
  1485. if (amdgpu_sriov_vf(adev))
  1486. amdgpu_virt_release_full_gpu(adev, true);
  1487. return 0;
  1488. }
  1489. /**
  1490. * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
  1491. *
  1492. * @adev: amdgpu_device pointer
  1493. *
  1494. * Writes a reset magic value to the gart pointer in VRAM. The driver calls
  1495. * this function before a GPU reset. If the value is retained after a
  1496. * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
  1497. */
  1498. static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
  1499. {
  1500. memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
  1501. }
  1502. /**
  1503. * amdgpu_device_check_vram_lost - check if vram is valid
  1504. *
  1505. * @adev: amdgpu_device pointer
  1506. *
  1507. * Checks the reset magic value written to the gart pointer in VRAM.
  1508. * The driver calls this after a GPU reset to see if the contents of
  1509. * VRAM is lost or now.
  1510. * returns true if vram is lost, false if not.
  1511. */
  1512. static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
  1513. {
  1514. return !!memcmp(adev->gart.ptr, adev->reset_magic,
  1515. AMDGPU_RESET_MAGIC_NUM);
  1516. }
  1517. /**
  1518. * amdgpu_device_set_cg_state - set clockgating for amdgpu device
  1519. *
  1520. * @adev: amdgpu_device pointer
  1521. *
  1522. * The list of all the hardware IPs that make up the asic is walked and the
  1523. * set_clockgating_state callbacks are run.
  1524. * Late initialization pass enabling clockgating for hardware IPs.
  1525. * Fini or suspend, pass disabling clockgating for hardware IPs.
  1526. * Returns 0 on success, negative error code on failure.
  1527. */
  1528. static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
  1529. enum amd_clockgating_state state)
  1530. {
  1531. int i, j, r;
  1532. if (amdgpu_emu_mode == 1)
  1533. return 0;
  1534. for (j = 0; j < adev->num_ip_blocks; j++) {
  1535. i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
  1536. if (!adev->ip_blocks[i].status.valid)
  1537. continue;
  1538. /* skip CG for VCE/UVD, it's handled specially */
  1539. if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
  1540. adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
  1541. adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
  1542. adev->ip_blocks[i].version->funcs->set_clockgating_state) {
  1543. /* enable clockgating to save power */
  1544. r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
  1545. state);
  1546. if (r) {
  1547. DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
  1548. adev->ip_blocks[i].version->funcs->name, r);
  1549. return r;
  1550. }
  1551. }
  1552. }
  1553. return 0;
  1554. }
  1555. static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
  1556. {
  1557. int i, j, r;
  1558. if (amdgpu_emu_mode == 1)
  1559. return 0;
  1560. for (j = 0; j < adev->num_ip_blocks; j++) {
  1561. i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
  1562. if (!adev->ip_blocks[i].status.valid)
  1563. continue;
  1564. /* skip CG for VCE/UVD, it's handled specially */
  1565. if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
  1566. adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
  1567. adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
  1568. adev->ip_blocks[i].version->funcs->set_powergating_state) {
  1569. /* enable powergating to save power */
  1570. r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
  1571. state);
  1572. if (r) {
  1573. DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
  1574. adev->ip_blocks[i].version->funcs->name, r);
  1575. return r;
  1576. }
  1577. }
  1578. }
  1579. return 0;
  1580. }
  1581. /**
  1582. * amdgpu_device_ip_late_init - run late init for hardware IPs
  1583. *
  1584. * @adev: amdgpu_device pointer
  1585. *
  1586. * Late initialization pass for hardware IPs. The list of all the hardware
  1587. * IPs that make up the asic is walked and the late_init callbacks are run.
  1588. * late_init covers any special initialization that an IP requires
  1589. * after all of the have been initialized or something that needs to happen
  1590. * late in the init process.
  1591. * Returns 0 on success, negative error code on failure.
  1592. */
  1593. static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
  1594. {
  1595. int i = 0, r;
  1596. for (i = 0; i < adev->num_ip_blocks; i++) {
  1597. if (!adev->ip_blocks[i].status.valid)
  1598. continue;
  1599. if (adev->ip_blocks[i].version->funcs->late_init) {
  1600. r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
  1601. if (r) {
  1602. DRM_ERROR("late_init of IP block <%s> failed %d\n",
  1603. adev->ip_blocks[i].version->funcs->name, r);
  1604. return r;
  1605. }
  1606. adev->ip_blocks[i].status.late_initialized = true;
  1607. }
  1608. }
  1609. amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
  1610. amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
  1611. queue_delayed_work(system_wq, &adev->late_init_work,
  1612. msecs_to_jiffies(AMDGPU_RESUME_MS));
  1613. amdgpu_device_fill_reset_magic(adev);
  1614. return 0;
  1615. }
  1616. /**
  1617. * amdgpu_device_ip_fini - run fini for hardware IPs
  1618. *
  1619. * @adev: amdgpu_device pointer
  1620. *
  1621. * Main teardown pass for hardware IPs. The list of all the hardware
  1622. * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
  1623. * are run. hw_fini tears down the hardware associated with each IP
  1624. * and sw_fini tears down any software state associated with each IP.
  1625. * Returns 0 on success, negative error code on failure.
  1626. */
  1627. static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
  1628. {
  1629. int i, r;
  1630. amdgpu_amdkfd_device_fini(adev);
  1631. amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
  1632. /* need to disable SMC first */
  1633. for (i = 0; i < adev->num_ip_blocks; i++) {
  1634. if (!adev->ip_blocks[i].status.hw)
  1635. continue;
  1636. if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC &&
  1637. adev->ip_blocks[i].version->funcs->set_clockgating_state) {
  1638. /* ungate blocks before hw fini so that we can shutdown the blocks safely */
  1639. r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
  1640. AMD_CG_STATE_UNGATE);
  1641. if (r) {
  1642. DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
  1643. adev->ip_blocks[i].version->funcs->name, r);
  1644. return r;
  1645. }
  1646. r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
  1647. /* XXX handle errors */
  1648. if (r) {
  1649. DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
  1650. adev->ip_blocks[i].version->funcs->name, r);
  1651. }
  1652. adev->ip_blocks[i].status.hw = false;
  1653. break;
  1654. }
  1655. }
  1656. for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
  1657. if (!adev->ip_blocks[i].status.hw)
  1658. continue;
  1659. if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
  1660. adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
  1661. adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
  1662. adev->ip_blocks[i].version->funcs->set_clockgating_state) {
  1663. /* ungate blocks before hw fini so that we can shutdown the blocks safely */
  1664. r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
  1665. AMD_CG_STATE_UNGATE);
  1666. if (r) {
  1667. DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
  1668. adev->ip_blocks[i].version->funcs->name, r);
  1669. return r;
  1670. }
  1671. }
  1672. r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
  1673. /* XXX handle errors */
  1674. if (r) {
  1675. DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
  1676. adev->ip_blocks[i].version->funcs->name, r);
  1677. }
  1678. adev->ip_blocks[i].status.hw = false;
  1679. }
  1680. for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
  1681. if (!adev->ip_blocks[i].status.sw)
  1682. continue;
  1683. if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
  1684. amdgpu_free_static_csa(adev);
  1685. amdgpu_device_wb_fini(adev);
  1686. amdgpu_device_vram_scratch_fini(adev);
  1687. }
  1688. r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
  1689. /* XXX handle errors */
  1690. if (r) {
  1691. DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
  1692. adev->ip_blocks[i].version->funcs->name, r);
  1693. }
  1694. adev->ip_blocks[i].status.sw = false;
  1695. adev->ip_blocks[i].status.valid = false;
  1696. }
  1697. for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
  1698. if (!adev->ip_blocks[i].status.late_initialized)
  1699. continue;
  1700. if (adev->ip_blocks[i].version->funcs->late_fini)
  1701. adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
  1702. adev->ip_blocks[i].status.late_initialized = false;
  1703. }
  1704. if (amdgpu_sriov_vf(adev))
  1705. if (amdgpu_virt_release_full_gpu(adev, false))
  1706. DRM_ERROR("failed to release exclusive mode on fini\n");
  1707. return 0;
  1708. }
  1709. /**
  1710. * amdgpu_device_ip_late_init_func_handler - work handler for ib test
  1711. *
  1712. * @work: work_struct.
  1713. */
  1714. static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
  1715. {
  1716. struct amdgpu_device *adev =
  1717. container_of(work, struct amdgpu_device, late_init_work.work);
  1718. int r;
  1719. r = amdgpu_ib_ring_tests(adev);
  1720. if (r)
  1721. DRM_ERROR("ib ring test failed (%d).\n", r);
  1722. }
  1723. static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
  1724. {
  1725. struct amdgpu_device *adev =
  1726. container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
  1727. mutex_lock(&adev->gfx.gfx_off_mutex);
  1728. if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
  1729. if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
  1730. adev->gfx.gfx_off_state = true;
  1731. }
  1732. mutex_unlock(&adev->gfx.gfx_off_mutex);
  1733. }
  1734. /**
  1735. * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
  1736. *
  1737. * @adev: amdgpu_device pointer
  1738. *
  1739. * Main suspend function for hardware IPs. The list of all the hardware
  1740. * IPs that make up the asic is walked, clockgating is disabled and the
  1741. * suspend callbacks are run. suspend puts the hardware and software state
  1742. * in each IP into a state suitable for suspend.
  1743. * Returns 0 on success, negative error code on failure.
  1744. */
  1745. static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
  1746. {
  1747. int i, r;
  1748. if (amdgpu_sriov_vf(adev))
  1749. amdgpu_virt_request_full_gpu(adev, false);
  1750. amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
  1751. for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
  1752. if (!adev->ip_blocks[i].status.valid)
  1753. continue;
  1754. /* displays are handled separately */
  1755. if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
  1756. /* ungate blocks so that suspend can properly shut them down */
  1757. if (adev->ip_blocks[i].version->funcs->set_clockgating_state) {
  1758. r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
  1759. AMD_CG_STATE_UNGATE);
  1760. if (r) {
  1761. DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
  1762. adev->ip_blocks[i].version->funcs->name, r);
  1763. }
  1764. }
  1765. /* XXX handle errors */
  1766. r = adev->ip_blocks[i].version->funcs->suspend(adev);
  1767. /* XXX handle errors */
  1768. if (r) {
  1769. DRM_ERROR("suspend of IP block <%s> failed %d\n",
  1770. adev->ip_blocks[i].version->funcs->name, r);
  1771. }
  1772. }
  1773. }
  1774. if (amdgpu_sriov_vf(adev))
  1775. amdgpu_virt_release_full_gpu(adev, false);
  1776. return 0;
  1777. }
  1778. /**
  1779. * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
  1780. *
  1781. * @adev: amdgpu_device pointer
  1782. *
  1783. * Main suspend function for hardware IPs. The list of all the hardware
  1784. * IPs that make up the asic is walked, clockgating is disabled and the
  1785. * suspend callbacks are run. suspend puts the hardware and software state
  1786. * in each IP into a state suitable for suspend.
  1787. * Returns 0 on success, negative error code on failure.
  1788. */
  1789. static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
  1790. {
  1791. int i, r;
  1792. if (amdgpu_sriov_vf(adev))
  1793. amdgpu_virt_request_full_gpu(adev, false);
  1794. /* ungate SMC block first */
  1795. r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
  1796. AMD_CG_STATE_UNGATE);
  1797. if (r) {
  1798. DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r);
  1799. }
  1800. for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
  1801. if (!adev->ip_blocks[i].status.valid)
  1802. continue;
  1803. /* displays are handled in phase1 */
  1804. if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
  1805. continue;
  1806. /* ungate blocks so that suspend can properly shut them down */
  1807. if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_SMC &&
  1808. adev->ip_blocks[i].version->funcs->set_clockgating_state) {
  1809. r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
  1810. AMD_CG_STATE_UNGATE);
  1811. if (r) {
  1812. DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
  1813. adev->ip_blocks[i].version->funcs->name, r);
  1814. }
  1815. }
  1816. /* XXX handle errors */
  1817. r = adev->ip_blocks[i].version->funcs->suspend(adev);
  1818. /* XXX handle errors */
  1819. if (r) {
  1820. DRM_ERROR("suspend of IP block <%s> failed %d\n",
  1821. adev->ip_blocks[i].version->funcs->name, r);
  1822. }
  1823. }
  1824. if (amdgpu_sriov_vf(adev))
  1825. amdgpu_virt_release_full_gpu(adev, false);
  1826. return 0;
  1827. }
  1828. /**
  1829. * amdgpu_device_ip_suspend - run suspend for hardware IPs
  1830. *
  1831. * @adev: amdgpu_device pointer
  1832. *
  1833. * Main suspend function for hardware IPs. The list of all the hardware
  1834. * IPs that make up the asic is walked, clockgating is disabled and the
  1835. * suspend callbacks are run. suspend puts the hardware and software state
  1836. * in each IP into a state suitable for suspend.
  1837. * Returns 0 on success, negative error code on failure.
  1838. */
  1839. int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
  1840. {
  1841. int r;
  1842. r = amdgpu_device_ip_suspend_phase1(adev);
  1843. if (r)
  1844. return r;
  1845. r = amdgpu_device_ip_suspend_phase2(adev);
  1846. return r;
  1847. }
  1848. static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
  1849. {
  1850. int i, r;
  1851. static enum amd_ip_block_type ip_order[] = {
  1852. AMD_IP_BLOCK_TYPE_GMC,
  1853. AMD_IP_BLOCK_TYPE_COMMON,
  1854. AMD_IP_BLOCK_TYPE_IH,
  1855. };
  1856. for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
  1857. int j;
  1858. struct amdgpu_ip_block *block;
  1859. for (j = 0; j < adev->num_ip_blocks; j++) {
  1860. block = &adev->ip_blocks[j];
  1861. if (block->version->type != ip_order[i] ||
  1862. !block->status.valid)
  1863. continue;
  1864. r = block->version->funcs->hw_init(adev);
  1865. DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
  1866. if (r)
  1867. return r;
  1868. }
  1869. }
  1870. return 0;
  1871. }
  1872. static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
  1873. {
  1874. int i, r;
  1875. static enum amd_ip_block_type ip_order[] = {
  1876. AMD_IP_BLOCK_TYPE_SMC,
  1877. AMD_IP_BLOCK_TYPE_PSP,
  1878. AMD_IP_BLOCK_TYPE_DCE,
  1879. AMD_IP_BLOCK_TYPE_GFX,
  1880. AMD_IP_BLOCK_TYPE_SDMA,
  1881. AMD_IP_BLOCK_TYPE_UVD,
  1882. AMD_IP_BLOCK_TYPE_VCE
  1883. };
  1884. for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
  1885. int j;
  1886. struct amdgpu_ip_block *block;
  1887. for (j = 0; j < adev->num_ip_blocks; j++) {
  1888. block = &adev->ip_blocks[j];
  1889. if (block->version->type != ip_order[i] ||
  1890. !block->status.valid)
  1891. continue;
  1892. r = block->version->funcs->hw_init(adev);
  1893. DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
  1894. if (r)
  1895. return r;
  1896. }
  1897. }
  1898. return 0;
  1899. }
  1900. /**
  1901. * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
  1902. *
  1903. * @adev: amdgpu_device pointer
  1904. *
  1905. * First resume function for hardware IPs. The list of all the hardware
  1906. * IPs that make up the asic is walked and the resume callbacks are run for
  1907. * COMMON, GMC, and IH. resume puts the hardware into a functional state
  1908. * after a suspend and updates the software state as necessary. This
  1909. * function is also used for restoring the GPU after a GPU reset.
  1910. * Returns 0 on success, negative error code on failure.
  1911. */
  1912. static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
  1913. {
  1914. int i, r;
  1915. for (i = 0; i < adev->num_ip_blocks; i++) {
  1916. if (!adev->ip_blocks[i].status.valid)
  1917. continue;
  1918. if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
  1919. adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
  1920. adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
  1921. r = adev->ip_blocks[i].version->funcs->resume(adev);
  1922. if (r) {
  1923. DRM_ERROR("resume of IP block <%s> failed %d\n",
  1924. adev->ip_blocks[i].version->funcs->name, r);
  1925. return r;
  1926. }
  1927. }
  1928. }
  1929. return 0;
  1930. }
  1931. /**
  1932. * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
  1933. *
  1934. * @adev: amdgpu_device pointer
  1935. *
  1936. * First resume function for hardware IPs. The list of all the hardware
  1937. * IPs that make up the asic is walked and the resume callbacks are run for
  1938. * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
  1939. * functional state after a suspend and updates the software state as
  1940. * necessary. This function is also used for restoring the GPU after a GPU
  1941. * reset.
  1942. * Returns 0 on success, negative error code on failure.
  1943. */
  1944. static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
  1945. {
  1946. int i, r;
  1947. for (i = 0; i < adev->num_ip_blocks; i++) {
  1948. if (!adev->ip_blocks[i].status.valid)
  1949. continue;
  1950. if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
  1951. adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
  1952. adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)
  1953. continue;
  1954. r = adev->ip_blocks[i].version->funcs->resume(adev);
  1955. if (r) {
  1956. DRM_ERROR("resume of IP block <%s> failed %d\n",
  1957. adev->ip_blocks[i].version->funcs->name, r);
  1958. return r;
  1959. }
  1960. }
  1961. return 0;
  1962. }
  1963. /**
  1964. * amdgpu_device_ip_resume - run resume for hardware IPs
  1965. *
  1966. * @adev: amdgpu_device pointer
  1967. *
  1968. * Main resume function for hardware IPs. The hardware IPs
  1969. * are split into two resume functions because they are
  1970. * are also used in in recovering from a GPU reset and some additional
  1971. * steps need to be take between them. In this case (S3/S4) they are
  1972. * run sequentially.
  1973. * Returns 0 on success, negative error code on failure.
  1974. */
  1975. static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
  1976. {
  1977. int r;
  1978. r = amdgpu_device_ip_resume_phase1(adev);
  1979. if (r)
  1980. return r;
  1981. r = amdgpu_device_ip_resume_phase2(adev);
  1982. return r;
  1983. }
  1984. /**
  1985. * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
  1986. *
  1987. * @adev: amdgpu_device pointer
  1988. *
  1989. * Query the VBIOS data tables to determine if the board supports SR-IOV.
  1990. */
  1991. static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
  1992. {
  1993. if (amdgpu_sriov_vf(adev)) {
  1994. if (adev->is_atom_fw) {
  1995. if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
  1996. adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
  1997. } else {
  1998. if (amdgpu_atombios_has_gpu_virtualization_table(adev))
  1999. adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
  2000. }
  2001. if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
  2002. amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
  2003. }
  2004. }
  2005. /**
  2006. * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
  2007. *
  2008. * @asic_type: AMD asic type
  2009. *
  2010. * Check if there is DC (new modesetting infrastructre) support for an asic.
  2011. * returns true if DC has support, false if not.
  2012. */
  2013. bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
  2014. {
  2015. switch (asic_type) {
  2016. #if defined(CONFIG_DRM_AMD_DC)
  2017. case CHIP_BONAIRE:
  2018. case CHIP_KAVERI:
  2019. case CHIP_KABINI:
  2020. case CHIP_MULLINS:
  2021. /*
  2022. * We have systems in the wild with these ASICs that require
  2023. * LVDS and VGA support which is not supported with DC.
  2024. *
  2025. * Fallback to the non-DC driver here by default so as not to
  2026. * cause regressions.
  2027. */
  2028. return amdgpu_dc > 0;
  2029. case CHIP_HAWAII:
  2030. case CHIP_CARRIZO:
  2031. case CHIP_STONEY:
  2032. case CHIP_POLARIS10:
  2033. case CHIP_POLARIS11:
  2034. case CHIP_POLARIS12:
  2035. case CHIP_VEGAM:
  2036. case CHIP_TONGA:
  2037. case CHIP_FIJI:
  2038. case CHIP_VEGA10:
  2039. case CHIP_VEGA12:
  2040. case CHIP_VEGA20:
  2041. #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
  2042. case CHIP_RAVEN:
  2043. #endif
  2044. return amdgpu_dc != 0;
  2045. #endif
  2046. default:
  2047. return false;
  2048. }
  2049. }
  2050. /**
  2051. * amdgpu_device_has_dc_support - check if dc is supported
  2052. *
  2053. * @adev: amdgpu_device_pointer
  2054. *
  2055. * Returns true for supported, false for not supported
  2056. */
  2057. bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
  2058. {
  2059. if (amdgpu_sriov_vf(adev))
  2060. return false;
  2061. return amdgpu_device_asic_has_dc_support(adev->asic_type);
  2062. }
  2063. /**
  2064. * amdgpu_device_init - initialize the driver
  2065. *
  2066. * @adev: amdgpu_device pointer
  2067. * @ddev: drm dev pointer
  2068. * @pdev: pci dev pointer
  2069. * @flags: driver flags
  2070. *
  2071. * Initializes the driver info and hw (all asics).
  2072. * Returns 0 for success or an error on failure.
  2073. * Called at driver startup.
  2074. */
  2075. int amdgpu_device_init(struct amdgpu_device *adev,
  2076. struct drm_device *ddev,
  2077. struct pci_dev *pdev,
  2078. uint32_t flags)
  2079. {
  2080. int r, i;
  2081. bool runtime = false;
  2082. u32 max_MBps;
  2083. adev->shutdown = false;
  2084. adev->dev = &pdev->dev;
  2085. adev->ddev = ddev;
  2086. adev->pdev = pdev;
  2087. adev->flags = flags;
  2088. adev->asic_type = flags & AMD_ASIC_MASK;
  2089. adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
  2090. if (amdgpu_emu_mode == 1)
  2091. adev->usec_timeout *= 2;
  2092. adev->gmc.gart_size = 512 * 1024 * 1024;
  2093. adev->accel_working = false;
  2094. adev->num_rings = 0;
  2095. adev->mman.buffer_funcs = NULL;
  2096. adev->mman.buffer_funcs_ring = NULL;
  2097. adev->vm_manager.vm_pte_funcs = NULL;
  2098. adev->vm_manager.vm_pte_num_rqs = 0;
  2099. adev->gmc.gmc_funcs = NULL;
  2100. adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
  2101. bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
  2102. adev->smc_rreg = &amdgpu_invalid_rreg;
  2103. adev->smc_wreg = &amdgpu_invalid_wreg;
  2104. adev->pcie_rreg = &amdgpu_invalid_rreg;
  2105. adev->pcie_wreg = &amdgpu_invalid_wreg;
  2106. adev->pciep_rreg = &amdgpu_invalid_rreg;
  2107. adev->pciep_wreg = &amdgpu_invalid_wreg;
  2108. adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
  2109. adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
  2110. adev->didt_rreg = &amdgpu_invalid_rreg;
  2111. adev->didt_wreg = &amdgpu_invalid_wreg;
  2112. adev->gc_cac_rreg = &amdgpu_invalid_rreg;
  2113. adev->gc_cac_wreg = &amdgpu_invalid_wreg;
  2114. adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
  2115. adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
  2116. DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
  2117. amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
  2118. pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
  2119. /* mutex initialization are all done here so we
  2120. * can recall function without having locking issues */
  2121. atomic_set(&adev->irq.ih.lock, 0);
  2122. mutex_init(&adev->firmware.mutex);
  2123. mutex_init(&adev->pm.mutex);
  2124. mutex_init(&adev->gfx.gpu_clock_mutex);
  2125. mutex_init(&adev->srbm_mutex);
  2126. mutex_init(&adev->gfx.pipe_reserve_mutex);
  2127. mutex_init(&adev->gfx.gfx_off_mutex);
  2128. mutex_init(&adev->grbm_idx_mutex);
  2129. mutex_init(&adev->mn_lock);
  2130. mutex_init(&adev->virt.vf_errors.lock);
  2131. hash_init(adev->mn_hash);
  2132. mutex_init(&adev->lock_reset);
  2133. amdgpu_device_check_arguments(adev);
  2134. spin_lock_init(&adev->mmio_idx_lock);
  2135. spin_lock_init(&adev->smc_idx_lock);
  2136. spin_lock_init(&adev->pcie_idx_lock);
  2137. spin_lock_init(&adev->uvd_ctx_idx_lock);
  2138. spin_lock_init(&adev->didt_idx_lock);
  2139. spin_lock_init(&adev->gc_cac_idx_lock);
  2140. spin_lock_init(&adev->se_cac_idx_lock);
  2141. spin_lock_init(&adev->audio_endpt_idx_lock);
  2142. spin_lock_init(&adev->mm_stats.lock);
  2143. INIT_LIST_HEAD(&adev->shadow_list);
  2144. mutex_init(&adev->shadow_list_lock);
  2145. INIT_LIST_HEAD(&adev->ring_lru_list);
  2146. spin_lock_init(&adev->ring_lru_list_lock);
  2147. INIT_DELAYED_WORK(&adev->late_init_work,
  2148. amdgpu_device_ip_late_init_func_handler);
  2149. INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
  2150. amdgpu_device_delay_enable_gfx_off);
  2151. adev->gfx.gfx_off_req_count = 1;
  2152. adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
  2153. /* Registers mapping */
  2154. /* TODO: block userspace mapping of io register */
  2155. if (adev->asic_type >= CHIP_BONAIRE) {
  2156. adev->rmmio_base = pci_resource_start(adev->pdev, 5);
  2157. adev->rmmio_size = pci_resource_len(adev->pdev, 5);
  2158. } else {
  2159. adev->rmmio_base = pci_resource_start(adev->pdev, 2);
  2160. adev->rmmio_size = pci_resource_len(adev->pdev, 2);
  2161. }
  2162. adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
  2163. if (adev->rmmio == NULL) {
  2164. return -ENOMEM;
  2165. }
  2166. DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
  2167. DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
  2168. /* doorbell bar mapping */
  2169. amdgpu_device_doorbell_init(adev);
  2170. /* io port mapping */
  2171. for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
  2172. if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
  2173. adev->rio_mem_size = pci_resource_len(adev->pdev, i);
  2174. adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
  2175. break;
  2176. }
  2177. }
  2178. if (adev->rio_mem == NULL)
  2179. DRM_INFO("PCI I/O BAR is not found.\n");
  2180. amdgpu_device_get_pcie_info(adev);
  2181. /* early init functions */
  2182. r = amdgpu_device_ip_early_init(adev);
  2183. if (r)
  2184. return r;
  2185. /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
  2186. /* this will fail for cards that aren't VGA class devices, just
  2187. * ignore it */
  2188. vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
  2189. if (amdgpu_device_is_px(ddev))
  2190. runtime = true;
  2191. if (!pci_is_thunderbolt_attached(adev->pdev))
  2192. vga_switcheroo_register_client(adev->pdev,
  2193. &amdgpu_switcheroo_ops, runtime);
  2194. if (runtime)
  2195. vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
  2196. if (amdgpu_emu_mode == 1) {
  2197. /* post the asic on emulation mode */
  2198. emu_soc_asic_init(adev);
  2199. goto fence_driver_init;
  2200. }
  2201. /* Read BIOS */
  2202. if (!amdgpu_get_bios(adev)) {
  2203. r = -EINVAL;
  2204. goto failed;
  2205. }
  2206. r = amdgpu_atombios_init(adev);
  2207. if (r) {
  2208. dev_err(adev->dev, "amdgpu_atombios_init failed\n");
  2209. amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
  2210. goto failed;
  2211. }
  2212. /* detect if we are with an SRIOV vbios */
  2213. amdgpu_device_detect_sriov_bios(adev);
  2214. /* Post card if necessary */
  2215. if (amdgpu_device_need_post(adev)) {
  2216. if (!adev->bios) {
  2217. dev_err(adev->dev, "no vBIOS found\n");
  2218. r = -EINVAL;
  2219. goto failed;
  2220. }
  2221. DRM_INFO("GPU posting now...\n");
  2222. r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
  2223. if (r) {
  2224. dev_err(adev->dev, "gpu post error!\n");
  2225. goto failed;
  2226. }
  2227. }
  2228. if (adev->is_atom_fw) {
  2229. /* Initialize clocks */
  2230. r = amdgpu_atomfirmware_get_clock_info(adev);
  2231. if (r) {
  2232. dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
  2233. amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
  2234. goto failed;
  2235. }
  2236. } else {
  2237. /* Initialize clocks */
  2238. r = amdgpu_atombios_get_clock_info(adev);
  2239. if (r) {
  2240. dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
  2241. amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
  2242. goto failed;
  2243. }
  2244. /* init i2c buses */
  2245. if (!amdgpu_device_has_dc_support(adev))
  2246. amdgpu_atombios_i2c_init(adev);
  2247. }
  2248. fence_driver_init:
  2249. /* Fence driver */
  2250. r = amdgpu_fence_driver_init(adev);
  2251. if (r) {
  2252. dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
  2253. amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
  2254. goto failed;
  2255. }
  2256. /* init the mode config */
  2257. drm_mode_config_init(adev->ddev);
  2258. r = amdgpu_device_ip_init(adev);
  2259. if (r) {
  2260. /* failed in exclusive mode due to timeout */
  2261. if (amdgpu_sriov_vf(adev) &&
  2262. !amdgpu_sriov_runtime(adev) &&
  2263. amdgpu_virt_mmio_blocked(adev) &&
  2264. !amdgpu_virt_wait_reset(adev)) {
  2265. dev_err(adev->dev, "VF exclusive mode timeout\n");
  2266. /* Don't send request since VF is inactive. */
  2267. adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
  2268. adev->virt.ops = NULL;
  2269. r = -EAGAIN;
  2270. goto failed;
  2271. }
  2272. dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
  2273. amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
  2274. goto failed;
  2275. }
  2276. adev->accel_working = true;
  2277. amdgpu_vm_check_compute_bug(adev);
  2278. /* Initialize the buffer migration limit. */
  2279. if (amdgpu_moverate >= 0)
  2280. max_MBps = amdgpu_moverate;
  2281. else
  2282. max_MBps = 8; /* Allow 8 MB/s. */
  2283. /* Get a log2 for easy divisions. */
  2284. adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
  2285. r = amdgpu_ib_pool_init(adev);
  2286. if (r) {
  2287. dev_err(adev->dev, "IB initialization failed (%d).\n", r);
  2288. amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
  2289. goto failed;
  2290. }
  2291. if (amdgpu_sriov_vf(adev))
  2292. amdgpu_virt_init_data_exchange(adev);
  2293. amdgpu_fbdev_init(adev);
  2294. r = amdgpu_pm_sysfs_init(adev);
  2295. if (r)
  2296. DRM_ERROR("registering pm debugfs failed (%d).\n", r);
  2297. r = amdgpu_debugfs_gem_init(adev);
  2298. if (r)
  2299. DRM_ERROR("registering gem debugfs failed (%d).\n", r);
  2300. r = amdgpu_debugfs_regs_init(adev);
  2301. if (r)
  2302. DRM_ERROR("registering register debugfs failed (%d).\n", r);
  2303. r = amdgpu_debugfs_firmware_init(adev);
  2304. if (r)
  2305. DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
  2306. r = amdgpu_debugfs_init(adev);
  2307. if (r)
  2308. DRM_ERROR("Creating debugfs files failed (%d).\n", r);
  2309. if ((amdgpu_testing & 1)) {
  2310. if (adev->accel_working)
  2311. amdgpu_test_moves(adev);
  2312. else
  2313. DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
  2314. }
  2315. if (amdgpu_benchmarking) {
  2316. if (adev->accel_working)
  2317. amdgpu_benchmark(adev, amdgpu_benchmarking);
  2318. else
  2319. DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
  2320. }
  2321. /* enable clockgating, etc. after ib tests, etc. since some blocks require
  2322. * explicit gating rather than handling it automatically.
  2323. */
  2324. r = amdgpu_device_ip_late_init(adev);
  2325. if (r) {
  2326. dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
  2327. amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
  2328. goto failed;
  2329. }
  2330. return 0;
  2331. failed:
  2332. amdgpu_vf_error_trans_all(adev);
  2333. if (runtime)
  2334. vga_switcheroo_fini_domain_pm_ops(adev->dev);
  2335. return r;
  2336. }
  2337. /**
  2338. * amdgpu_device_fini - tear down the driver
  2339. *
  2340. * @adev: amdgpu_device pointer
  2341. *
  2342. * Tear down the driver info (all asics).
  2343. * Called at driver shutdown.
  2344. */
  2345. void amdgpu_device_fini(struct amdgpu_device *adev)
  2346. {
  2347. int r;
  2348. DRM_INFO("amdgpu: finishing device.\n");
  2349. adev->shutdown = true;
  2350. /* disable all interrupts */
  2351. amdgpu_irq_disable_all(adev);
  2352. if (adev->mode_info.mode_config_initialized){
  2353. if (!amdgpu_device_has_dc_support(adev))
  2354. drm_crtc_force_disable_all(adev->ddev);
  2355. else
  2356. drm_atomic_helper_shutdown(adev->ddev);
  2357. }
  2358. amdgpu_ib_pool_fini(adev);
  2359. amdgpu_fence_driver_fini(adev);
  2360. amdgpu_pm_sysfs_fini(adev);
  2361. amdgpu_fbdev_fini(adev);
  2362. r = amdgpu_device_ip_fini(adev);
  2363. if (adev->firmware.gpu_info_fw) {
  2364. release_firmware(adev->firmware.gpu_info_fw);
  2365. adev->firmware.gpu_info_fw = NULL;
  2366. }
  2367. adev->accel_working = false;
  2368. cancel_delayed_work_sync(&adev->late_init_work);
  2369. /* free i2c buses */
  2370. if (!amdgpu_device_has_dc_support(adev))
  2371. amdgpu_i2c_fini(adev);
  2372. if (amdgpu_emu_mode != 1)
  2373. amdgpu_atombios_fini(adev);
  2374. kfree(adev->bios);
  2375. adev->bios = NULL;
  2376. if (!pci_is_thunderbolt_attached(adev->pdev))
  2377. vga_switcheroo_unregister_client(adev->pdev);
  2378. if (adev->flags & AMD_IS_PX)
  2379. vga_switcheroo_fini_domain_pm_ops(adev->dev);
  2380. vga_client_register(adev->pdev, NULL, NULL, NULL);
  2381. if (adev->rio_mem)
  2382. pci_iounmap(adev->pdev, adev->rio_mem);
  2383. adev->rio_mem = NULL;
  2384. iounmap(adev->rmmio);
  2385. adev->rmmio = NULL;
  2386. amdgpu_device_doorbell_fini(adev);
  2387. amdgpu_debugfs_regs_cleanup(adev);
  2388. }
  2389. /*
  2390. * Suspend & resume.
  2391. */
  2392. /**
  2393. * amdgpu_device_suspend - initiate device suspend
  2394. *
  2395. * @dev: drm dev pointer
  2396. * @suspend: suspend state
  2397. * @fbcon : notify the fbdev of suspend
  2398. *
  2399. * Puts the hw in the suspend state (all asics).
  2400. * Returns 0 for success or an error on failure.
  2401. * Called at driver suspend.
  2402. */
  2403. int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
  2404. {
  2405. struct amdgpu_device *adev;
  2406. struct drm_crtc *crtc;
  2407. struct drm_connector *connector;
  2408. int r;
  2409. if (dev == NULL || dev->dev_private == NULL) {
  2410. return -ENODEV;
  2411. }
  2412. adev = dev->dev_private;
  2413. if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
  2414. return 0;
  2415. drm_kms_helper_poll_disable(dev);
  2416. if (fbcon)
  2417. amdgpu_fbdev_set_suspend(adev, 1);
  2418. cancel_delayed_work_sync(&adev->late_init_work);
  2419. if (!amdgpu_device_has_dc_support(adev)) {
  2420. /* turn off display hw */
  2421. drm_modeset_lock_all(dev);
  2422. list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
  2423. drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
  2424. }
  2425. drm_modeset_unlock_all(dev);
  2426. /* unpin the front buffers and cursors */
  2427. list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
  2428. struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
  2429. struct drm_framebuffer *fb = crtc->primary->fb;
  2430. struct amdgpu_bo *robj;
  2431. if (amdgpu_crtc->cursor_bo) {
  2432. struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
  2433. r = amdgpu_bo_reserve(aobj, true);
  2434. if (r == 0) {
  2435. amdgpu_bo_unpin(aobj);
  2436. amdgpu_bo_unreserve(aobj);
  2437. }
  2438. }
  2439. if (fb == NULL || fb->obj[0] == NULL) {
  2440. continue;
  2441. }
  2442. robj = gem_to_amdgpu_bo(fb->obj[0]);
  2443. /* don't unpin kernel fb objects */
  2444. if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
  2445. r = amdgpu_bo_reserve(robj, true);
  2446. if (r == 0) {
  2447. amdgpu_bo_unpin(robj);
  2448. amdgpu_bo_unreserve(robj);
  2449. }
  2450. }
  2451. }
  2452. }
  2453. amdgpu_amdkfd_suspend(adev);
  2454. r = amdgpu_device_ip_suspend_phase1(adev);
  2455. /* evict vram memory */
  2456. amdgpu_bo_evict_vram(adev);
  2457. amdgpu_fence_driver_suspend(adev);
  2458. r = amdgpu_device_ip_suspend_phase2(adev);
  2459. /* evict remaining vram memory
  2460. * This second call to evict vram is to evict the gart page table
  2461. * using the CPU.
  2462. */
  2463. amdgpu_bo_evict_vram(adev);
  2464. pci_save_state(dev->pdev);
  2465. if (suspend) {
  2466. /* Shut down the device */
  2467. pci_disable_device(dev->pdev);
  2468. pci_set_power_state(dev->pdev, PCI_D3hot);
  2469. } else {
  2470. r = amdgpu_asic_reset(adev);
  2471. if (r)
  2472. DRM_ERROR("amdgpu asic reset failed\n");
  2473. }
  2474. return 0;
  2475. }
  2476. /**
  2477. * amdgpu_device_resume - initiate device resume
  2478. *
  2479. * @dev: drm dev pointer
  2480. * @resume: resume state
  2481. * @fbcon : notify the fbdev of resume
  2482. *
  2483. * Bring the hw back to operating state (all asics).
  2484. * Returns 0 for success or an error on failure.
  2485. * Called at driver resume.
  2486. */
  2487. int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
  2488. {
  2489. struct drm_connector *connector;
  2490. struct amdgpu_device *adev = dev->dev_private;
  2491. struct drm_crtc *crtc;
  2492. int r = 0;
  2493. if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
  2494. return 0;
  2495. if (resume) {
  2496. pci_set_power_state(dev->pdev, PCI_D0);
  2497. pci_restore_state(dev->pdev);
  2498. r = pci_enable_device(dev->pdev);
  2499. if (r)
  2500. return r;
  2501. }
  2502. /* post card */
  2503. if (amdgpu_device_need_post(adev)) {
  2504. r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
  2505. if (r)
  2506. DRM_ERROR("amdgpu asic init failed\n");
  2507. }
  2508. r = amdgpu_device_ip_resume(adev);
  2509. if (r) {
  2510. DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
  2511. return r;
  2512. }
  2513. amdgpu_fence_driver_resume(adev);
  2514. r = amdgpu_device_ip_late_init(adev);
  2515. if (r)
  2516. return r;
  2517. if (!amdgpu_device_has_dc_support(adev)) {
  2518. /* pin cursors */
  2519. list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
  2520. struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
  2521. if (amdgpu_crtc->cursor_bo) {
  2522. struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
  2523. r = amdgpu_bo_reserve(aobj, true);
  2524. if (r == 0) {
  2525. r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
  2526. if (r != 0)
  2527. DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
  2528. amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
  2529. amdgpu_bo_unreserve(aobj);
  2530. }
  2531. }
  2532. }
  2533. }
  2534. r = amdgpu_amdkfd_resume(adev);
  2535. if (r)
  2536. return r;
  2537. /* Make sure IB tests flushed */
  2538. flush_delayed_work(&adev->late_init_work);
  2539. /* blat the mode back in */
  2540. if (fbcon) {
  2541. if (!amdgpu_device_has_dc_support(adev)) {
  2542. /* pre DCE11 */
  2543. drm_helper_resume_force_mode(dev);
  2544. /* turn on display hw */
  2545. drm_modeset_lock_all(dev);
  2546. list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
  2547. drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
  2548. }
  2549. drm_modeset_unlock_all(dev);
  2550. }
  2551. amdgpu_fbdev_set_suspend(adev, 0);
  2552. }
  2553. drm_kms_helper_poll_enable(dev);
  2554. /*
  2555. * Most of the connector probing functions try to acquire runtime pm
  2556. * refs to ensure that the GPU is powered on when connector polling is
  2557. * performed. Since we're calling this from a runtime PM callback,
  2558. * trying to acquire rpm refs will cause us to deadlock.
  2559. *
  2560. * Since we're guaranteed to be holding the rpm lock, it's safe to
  2561. * temporarily disable the rpm helpers so this doesn't deadlock us.
  2562. */
  2563. #ifdef CONFIG_PM
  2564. dev->dev->power.disable_depth++;
  2565. #endif
  2566. if (!amdgpu_device_has_dc_support(adev))
  2567. drm_helper_hpd_irq_event(dev);
  2568. else
  2569. drm_kms_helper_hotplug_event(dev);
  2570. #ifdef CONFIG_PM
  2571. dev->dev->power.disable_depth--;
  2572. #endif
  2573. return 0;
  2574. }
  2575. /**
  2576. * amdgpu_device_ip_check_soft_reset - did soft reset succeed
  2577. *
  2578. * @adev: amdgpu_device pointer
  2579. *
  2580. * The list of all the hardware IPs that make up the asic is walked and
  2581. * the check_soft_reset callbacks are run. check_soft_reset determines
  2582. * if the asic is still hung or not.
  2583. * Returns true if any of the IPs are still in a hung state, false if not.
  2584. */
  2585. static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
  2586. {
  2587. int i;
  2588. bool asic_hang = false;
  2589. if (amdgpu_sriov_vf(adev))
  2590. return true;
  2591. if (amdgpu_asic_need_full_reset(adev))
  2592. return true;
  2593. for (i = 0; i < adev->num_ip_blocks; i++) {
  2594. if (!adev->ip_blocks[i].status.valid)
  2595. continue;
  2596. if (adev->ip_blocks[i].version->funcs->check_soft_reset)
  2597. adev->ip_blocks[i].status.hang =
  2598. adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
  2599. if (adev->ip_blocks[i].status.hang) {
  2600. DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
  2601. asic_hang = true;
  2602. }
  2603. }
  2604. return asic_hang;
  2605. }
  2606. /**
  2607. * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
  2608. *
  2609. * @adev: amdgpu_device pointer
  2610. *
  2611. * The list of all the hardware IPs that make up the asic is walked and the
  2612. * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
  2613. * handles any IP specific hardware or software state changes that are
  2614. * necessary for a soft reset to succeed.
  2615. * Returns 0 on success, negative error code on failure.
  2616. */
  2617. static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
  2618. {
  2619. int i, r = 0;
  2620. for (i = 0; i < adev->num_ip_blocks; i++) {
  2621. if (!adev->ip_blocks[i].status.valid)
  2622. continue;
  2623. if (adev->ip_blocks[i].status.hang &&
  2624. adev->ip_blocks[i].version->funcs->pre_soft_reset) {
  2625. r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
  2626. if (r)
  2627. return r;
  2628. }
  2629. }
  2630. return 0;
  2631. }
  2632. /**
  2633. * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
  2634. *
  2635. * @adev: amdgpu_device pointer
  2636. *
  2637. * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
  2638. * reset is necessary to recover.
  2639. * Returns true if a full asic reset is required, false if not.
  2640. */
  2641. static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
  2642. {
  2643. int i;
  2644. if (amdgpu_asic_need_full_reset(adev))
  2645. return true;
  2646. for (i = 0; i < adev->num_ip_blocks; i++) {
  2647. if (!adev->ip_blocks[i].status.valid)
  2648. continue;
  2649. if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
  2650. (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
  2651. (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
  2652. (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
  2653. adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
  2654. if (adev->ip_blocks[i].status.hang) {
  2655. DRM_INFO("Some block need full reset!\n");
  2656. return true;
  2657. }
  2658. }
  2659. }
  2660. return false;
  2661. }
  2662. /**
  2663. * amdgpu_device_ip_soft_reset - do a soft reset
  2664. *
  2665. * @adev: amdgpu_device pointer
  2666. *
  2667. * The list of all the hardware IPs that make up the asic is walked and the
  2668. * soft_reset callbacks are run if the block is hung. soft_reset handles any
  2669. * IP specific hardware or software state changes that are necessary to soft
  2670. * reset the IP.
  2671. * Returns 0 on success, negative error code on failure.
  2672. */
  2673. static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
  2674. {
  2675. int i, r = 0;
  2676. for (i = 0; i < adev->num_ip_blocks; i++) {
  2677. if (!adev->ip_blocks[i].status.valid)
  2678. continue;
  2679. if (adev->ip_blocks[i].status.hang &&
  2680. adev->ip_blocks[i].version->funcs->soft_reset) {
  2681. r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
  2682. if (r)
  2683. return r;
  2684. }
  2685. }
  2686. return 0;
  2687. }
  2688. /**
  2689. * amdgpu_device_ip_post_soft_reset - clean up from soft reset
  2690. *
  2691. * @adev: amdgpu_device pointer
  2692. *
  2693. * The list of all the hardware IPs that make up the asic is walked and the
  2694. * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
  2695. * handles any IP specific hardware or software state changes that are
  2696. * necessary after the IP has been soft reset.
  2697. * Returns 0 on success, negative error code on failure.
  2698. */
  2699. static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
  2700. {
  2701. int i, r = 0;
  2702. for (i = 0; i < adev->num_ip_blocks; i++) {
  2703. if (!adev->ip_blocks[i].status.valid)
  2704. continue;
  2705. if (adev->ip_blocks[i].status.hang &&
  2706. adev->ip_blocks[i].version->funcs->post_soft_reset)
  2707. r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
  2708. if (r)
  2709. return r;
  2710. }
  2711. return 0;
  2712. }
  2713. /**
  2714. * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers
  2715. *
  2716. * @adev: amdgpu_device pointer
  2717. * @ring: amdgpu_ring for the engine handling the buffer operations
  2718. * @bo: amdgpu_bo buffer whose shadow is being restored
  2719. * @fence: dma_fence associated with the operation
  2720. *
  2721. * Restores the VRAM buffer contents from the shadow in GTT. Used to
  2722. * restore things like GPUVM page tables after a GPU reset where
  2723. * the contents of VRAM might be lost.
  2724. * Returns 0 on success, negative error code on failure.
  2725. */
  2726. static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
  2727. struct amdgpu_ring *ring,
  2728. struct amdgpu_bo *bo,
  2729. struct dma_fence **fence)
  2730. {
  2731. uint32_t domain;
  2732. int r;
  2733. if (!bo->shadow)
  2734. return 0;
  2735. r = amdgpu_bo_reserve(bo, true);
  2736. if (r)
  2737. return r;
  2738. domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
  2739. /* if bo has been evicted, then no need to recover */
  2740. if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
  2741. r = amdgpu_bo_validate(bo->shadow);
  2742. if (r) {
  2743. DRM_ERROR("bo validate failed!\n");
  2744. goto err;
  2745. }
  2746. r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
  2747. NULL, fence, true);
  2748. if (r) {
  2749. DRM_ERROR("recover page table failed!\n");
  2750. goto err;
  2751. }
  2752. }
  2753. err:
  2754. amdgpu_bo_unreserve(bo);
  2755. return r;
  2756. }
  2757. /**
  2758. * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents
  2759. *
  2760. * @adev: amdgpu_device pointer
  2761. *
  2762. * Restores the contents of VRAM buffers from the shadows in GTT. Used to
  2763. * restore things like GPUVM page tables after a GPU reset where
  2764. * the contents of VRAM might be lost.
  2765. * Returns 0 on success, 1 on failure.
  2766. */
  2767. static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
  2768. {
  2769. struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
  2770. struct amdgpu_bo *bo, *tmp;
  2771. struct dma_fence *fence = NULL, *next = NULL;
  2772. long r = 1;
  2773. int i = 0;
  2774. long tmo;
  2775. if (amdgpu_sriov_runtime(adev))
  2776. tmo = msecs_to_jiffies(8000);
  2777. else
  2778. tmo = msecs_to_jiffies(100);
  2779. DRM_INFO("recover vram bo from shadow start\n");
  2780. mutex_lock(&adev->shadow_list_lock);
  2781. list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
  2782. next = NULL;
  2783. amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
  2784. if (fence) {
  2785. r = dma_fence_wait_timeout(fence, false, tmo);
  2786. if (r == 0)
  2787. pr_err("wait fence %p[%d] timeout\n", fence, i);
  2788. else if (r < 0)
  2789. pr_err("wait fence %p[%d] interrupted\n", fence, i);
  2790. if (r < 1) {
  2791. dma_fence_put(fence);
  2792. fence = next;
  2793. break;
  2794. }
  2795. i++;
  2796. }
  2797. dma_fence_put(fence);
  2798. fence = next;
  2799. }
  2800. mutex_unlock(&adev->shadow_list_lock);
  2801. if (fence) {
  2802. r = dma_fence_wait_timeout(fence, false, tmo);
  2803. if (r == 0)
  2804. pr_err("wait fence %p[%d] timeout\n", fence, i);
  2805. else if (r < 0)
  2806. pr_err("wait fence %p[%d] interrupted\n", fence, i);
  2807. }
  2808. dma_fence_put(fence);
  2809. if (r > 0)
  2810. DRM_INFO("recover vram bo from shadow done\n");
  2811. else
  2812. DRM_ERROR("recover vram bo from shadow failed\n");
  2813. return (r > 0) ? 0 : 1;
  2814. }
  2815. /**
  2816. * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
  2817. *
  2818. * @adev: amdgpu device pointer
  2819. *
  2820. * attempt to do soft-reset or full-reset and reinitialize Asic
  2821. * return 0 means succeeded otherwise failed
  2822. */
  2823. static int amdgpu_device_reset(struct amdgpu_device *adev)
  2824. {
  2825. bool need_full_reset, vram_lost = 0;
  2826. int r;
  2827. need_full_reset = amdgpu_device_ip_need_full_reset(adev);
  2828. if (!need_full_reset) {
  2829. amdgpu_device_ip_pre_soft_reset(adev);
  2830. r = amdgpu_device_ip_soft_reset(adev);
  2831. amdgpu_device_ip_post_soft_reset(adev);
  2832. if (r || amdgpu_device_ip_check_soft_reset(adev)) {
  2833. DRM_INFO("soft reset failed, will fallback to full reset!\n");
  2834. need_full_reset = true;
  2835. }
  2836. }
  2837. if (need_full_reset) {
  2838. r = amdgpu_device_ip_suspend(adev);
  2839. retry:
  2840. r = amdgpu_asic_reset(adev);
  2841. /* post card */
  2842. amdgpu_atom_asic_init(adev->mode_info.atom_context);
  2843. if (!r) {
  2844. dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
  2845. r = amdgpu_device_ip_resume_phase1(adev);
  2846. if (r)
  2847. goto out;
  2848. vram_lost = amdgpu_device_check_vram_lost(adev);
  2849. if (vram_lost) {
  2850. DRM_ERROR("VRAM is lost!\n");
  2851. atomic_inc(&adev->vram_lost_counter);
  2852. }
  2853. r = amdgpu_gtt_mgr_recover(
  2854. &adev->mman.bdev.man[TTM_PL_TT]);
  2855. if (r)
  2856. goto out;
  2857. r = amdgpu_device_ip_resume_phase2(adev);
  2858. if (r)
  2859. goto out;
  2860. if (vram_lost)
  2861. amdgpu_device_fill_reset_magic(adev);
  2862. }
  2863. }
  2864. out:
  2865. if (!r) {
  2866. amdgpu_irq_gpu_reset_resume_helper(adev);
  2867. r = amdgpu_ib_ring_tests(adev);
  2868. if (r) {
  2869. dev_err(adev->dev, "ib ring test failed (%d).\n", r);
  2870. r = amdgpu_device_ip_suspend(adev);
  2871. need_full_reset = true;
  2872. goto retry;
  2873. }
  2874. }
  2875. if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
  2876. r = amdgpu_device_handle_vram_lost(adev);
  2877. return r;
  2878. }
  2879. /**
  2880. * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
  2881. *
  2882. * @adev: amdgpu device pointer
  2883. * @from_hypervisor: request from hypervisor
  2884. *
  2885. * do VF FLR and reinitialize Asic
  2886. * return 0 means succeeded otherwise failed
  2887. */
  2888. static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
  2889. bool from_hypervisor)
  2890. {
  2891. int r;
  2892. if (from_hypervisor)
  2893. r = amdgpu_virt_request_full_gpu(adev, true);
  2894. else
  2895. r = amdgpu_virt_reset_gpu(adev);
  2896. if (r)
  2897. return r;
  2898. /* Resume IP prior to SMC */
  2899. r = amdgpu_device_ip_reinit_early_sriov(adev);
  2900. if (r)
  2901. goto error;
  2902. /* we need recover gart prior to run SMC/CP/SDMA resume */
  2903. amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
  2904. /* now we are okay to resume SMC/CP/SDMA */
  2905. r = amdgpu_device_ip_reinit_late_sriov(adev);
  2906. if (r)
  2907. goto error;
  2908. amdgpu_irq_gpu_reset_resume_helper(adev);
  2909. r = amdgpu_ib_ring_tests(adev);
  2910. error:
  2911. amdgpu_virt_release_full_gpu(adev, true);
  2912. if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
  2913. atomic_inc(&adev->vram_lost_counter);
  2914. r = amdgpu_device_handle_vram_lost(adev);
  2915. }
  2916. return r;
  2917. }
  2918. /**
  2919. * amdgpu_device_gpu_recover - reset the asic and recover scheduler
  2920. *
  2921. * @adev: amdgpu device pointer
  2922. * @job: which job trigger hang
  2923. * @force: forces reset regardless of amdgpu_gpu_recovery
  2924. *
  2925. * Attempt to reset the GPU if it has hung (all asics).
  2926. * Returns 0 for success or an error on failure.
  2927. */
  2928. int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
  2929. struct amdgpu_job *job, bool force)
  2930. {
  2931. int i, r, resched;
  2932. if (!force && !amdgpu_device_ip_check_soft_reset(adev)) {
  2933. DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
  2934. return 0;
  2935. }
  2936. if (!force && (amdgpu_gpu_recovery == 0 ||
  2937. (amdgpu_gpu_recovery == -1 && !amdgpu_sriov_vf(adev)))) {
  2938. DRM_INFO("GPU recovery disabled.\n");
  2939. return 0;
  2940. }
  2941. dev_info(adev->dev, "GPU reset begin!\n");
  2942. mutex_lock(&adev->lock_reset);
  2943. atomic_inc(&adev->gpu_reset_counter);
  2944. adev->in_gpu_reset = 1;
  2945. /* Block kfd */
  2946. amdgpu_amdkfd_pre_reset(adev);
  2947. /* block TTM */
  2948. resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
  2949. /* block all schedulers and reset given job's ring */
  2950. for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
  2951. struct amdgpu_ring *ring = adev->rings[i];
  2952. if (!ring || !ring->sched.thread)
  2953. continue;
  2954. kthread_park(ring->sched.thread);
  2955. if (job && job->base.sched == &ring->sched)
  2956. continue;
  2957. drm_sched_hw_job_reset(&ring->sched, job ? &job->base : NULL);
  2958. /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
  2959. amdgpu_fence_driver_force_completion(ring);
  2960. }
  2961. if (amdgpu_sriov_vf(adev))
  2962. r = amdgpu_device_reset_sriov(adev, job ? false : true);
  2963. else
  2964. r = amdgpu_device_reset(adev);
  2965. for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
  2966. struct amdgpu_ring *ring = adev->rings[i];
  2967. if (!ring || !ring->sched.thread)
  2968. continue;
  2969. /* only need recovery sched of the given job's ring
  2970. * or all rings (in the case @job is NULL)
  2971. * after above amdgpu_reset accomplished
  2972. */
  2973. if ((!job || job->base.sched == &ring->sched) && !r)
  2974. drm_sched_job_recovery(&ring->sched);
  2975. kthread_unpark(ring->sched.thread);
  2976. }
  2977. if (!amdgpu_device_has_dc_support(adev)) {
  2978. drm_helper_resume_force_mode(adev->ddev);
  2979. }
  2980. ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
  2981. if (r) {
  2982. /* bad news, how to tell it to userspace ? */
  2983. dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
  2984. amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
  2985. } else {
  2986. dev_info(adev->dev, "GPU reset(%d) succeeded!\n",atomic_read(&adev->gpu_reset_counter));
  2987. }
  2988. /*unlock kfd */
  2989. amdgpu_amdkfd_post_reset(adev);
  2990. amdgpu_vf_error_trans_all(adev);
  2991. adev->in_gpu_reset = 0;
  2992. mutex_unlock(&adev->lock_reset);
  2993. return r;
  2994. }
  2995. /**
  2996. * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
  2997. *
  2998. * @adev: amdgpu_device pointer
  2999. *
  3000. * Fetchs and stores in the driver the PCIE capabilities (gen speed
  3001. * and lanes) of the slot the device is in. Handles APUs and
  3002. * virtualized environments where PCIE config space may not be available.
  3003. */
  3004. static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
  3005. {
  3006. struct pci_dev *pdev;
  3007. enum pci_bus_speed speed_cap;
  3008. enum pcie_link_width link_width;
  3009. if (amdgpu_pcie_gen_cap)
  3010. adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
  3011. if (amdgpu_pcie_lane_cap)
  3012. adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
  3013. /* covers APUs as well */
  3014. if (pci_is_root_bus(adev->pdev->bus)) {
  3015. if (adev->pm.pcie_gen_mask == 0)
  3016. adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
  3017. if (adev->pm.pcie_mlw_mask == 0)
  3018. adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
  3019. return;
  3020. }
  3021. if (adev->pm.pcie_gen_mask == 0) {
  3022. /* asic caps */
  3023. pdev = adev->pdev;
  3024. speed_cap = pcie_get_speed_cap(pdev);
  3025. if (speed_cap == PCI_SPEED_UNKNOWN) {
  3026. adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
  3027. CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
  3028. CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
  3029. } else {
  3030. if (speed_cap == PCIE_SPEED_16_0GT)
  3031. adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
  3032. CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
  3033. CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
  3034. CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
  3035. else if (speed_cap == PCIE_SPEED_8_0GT)
  3036. adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
  3037. CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
  3038. CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
  3039. else if (speed_cap == PCIE_SPEED_5_0GT)
  3040. adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
  3041. CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
  3042. else
  3043. adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
  3044. }
  3045. /* platform caps */
  3046. pdev = adev->ddev->pdev->bus->self;
  3047. speed_cap = pcie_get_speed_cap(pdev);
  3048. if (speed_cap == PCI_SPEED_UNKNOWN) {
  3049. adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
  3050. CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
  3051. } else {
  3052. if (speed_cap == PCIE_SPEED_16_0GT)
  3053. adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
  3054. CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
  3055. CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
  3056. CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
  3057. else if (speed_cap == PCIE_SPEED_8_0GT)
  3058. adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
  3059. CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
  3060. CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
  3061. else if (speed_cap == PCIE_SPEED_5_0GT)
  3062. adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
  3063. CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
  3064. else
  3065. adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
  3066. }
  3067. }
  3068. if (adev->pm.pcie_mlw_mask == 0) {
  3069. pdev = adev->ddev->pdev->bus->self;
  3070. link_width = pcie_get_width_cap(pdev);
  3071. if (link_width == PCIE_LNK_WIDTH_UNKNOWN) {
  3072. adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
  3073. } else {
  3074. switch (link_width) {
  3075. case PCIE_LNK_X32:
  3076. adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
  3077. CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
  3078. CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
  3079. CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
  3080. CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
  3081. CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
  3082. CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
  3083. break;
  3084. case PCIE_LNK_X16:
  3085. adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
  3086. CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
  3087. CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
  3088. CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
  3089. CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
  3090. CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
  3091. break;
  3092. case PCIE_LNK_X12:
  3093. adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
  3094. CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
  3095. CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
  3096. CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
  3097. CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
  3098. break;
  3099. case PCIE_LNK_X8:
  3100. adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
  3101. CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
  3102. CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
  3103. CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
  3104. break;
  3105. case PCIE_LNK_X4:
  3106. adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
  3107. CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
  3108. CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
  3109. break;
  3110. case PCIE_LNK_X2:
  3111. adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
  3112. CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
  3113. break;
  3114. case PCIE_LNK_X1:
  3115. adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
  3116. break;
  3117. default:
  3118. break;
  3119. }
  3120. }
  3121. }
  3122. }