xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c (revision d53b8e36925256097a08d7cb749198d85cbf9b2b)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  *
23  */
24 
25 #include <linux/device.h>
26 #include <linux/export.h>
27 #include <linux/err.h>
28 #include <linux/fs.h>
29 #include <linux/sched.h>
30 #include <linux/slab.h>
31 #include <linux/uaccess.h>
32 #include <linux/compat.h>
33 #include <uapi/linux/kfd_ioctl.h>
34 #include <linux/time.h>
35 #include "kfd_priv.h"
36 #include <linux/mm.h>
37 #include <linux/mman.h>
38 #include <linux/processor.h>
39 #include "amdgpu_vm.h"
40 
41 /*
42  * The primary memory I/O features being added for revisions of gfxip
43  * beyond 7.0 (Kaveri) are:
44  *
45  * Access to ATC/IOMMU mapped memory w/ associated extension of VA to 48b
46  *
47  * “Flat” shader memory access – These are new shader vector memory
48  * operations that do not reference a T#/V# so a “pointer” is what is
49  * sourced from the vector gprs for direct access to memory.
50  * This pointer space has the Shared(LDS) and Private(Scratch) memory
51  * mapped into this pointer space as apertures.
52  * The hardware then determines how to direct the memory request
53  * based on what apertures the request falls in.
54  *
55  * Unaligned support and alignment check
56  *
57  *
58  * System Unified Address - SUA
59  *
60  * The standard usage for GPU virtual addresses are that they are mapped by
61  * a set of page tables we call GPUVM and these page tables are managed by
62  * a combination of vidMM/driver software components.  The current virtual
63  * address (VA) range for GPUVM is 40b.
64  *
65  * As of gfxip7.1 and beyond we’re adding the ability for compute memory
66  * clients (CP/RLC, DMA, SHADER(ifetch, scalar, and vector ops)) to access
67  * the same page tables used by host x86 processors and that are managed by
68  * the operating system. This is via a technique and hardware called ATC/IOMMU.
69  * The GPU has the capability of accessing both the GPUVM and ATC address
70  * spaces for a given VMID (process) simultaneously and we call this feature
71  * system unified address (SUA).
72  *
73  * There are three fundamental address modes of operation for a given VMID
74  * (process) on the GPU:
75  *
76  *	HSA64 – 64b pointers and the default address space is ATC
77  *	HSA32 – 32b pointers and the default address space is ATC
78  *	GPUVM – 64b pointers and the default address space is GPUVM (driver
79  *		model mode)
80  *
81  *
82  * HSA64 - ATC/IOMMU 64b
83  *
84  * A 64b pointer in the AMD64/IA64 CPU architecture is not fully utilized
85  * by the CPU so an AMD CPU can only access the high area
86  * (VA[63:47] == 0x1FFFF) and low area (VA[63:47 == 0) of the address space
87  * so the actual VA carried to translation is 48b.  There is a “hole” in
88  * the middle of the 64b VA space.
89  *
90  * The GPU not only has access to all of the CPU accessible address space via
91  * ATC/IOMMU, but it also has access to the GPUVM address space.  The “system
92  * unified address” feature (SUA) is the mapping of GPUVM and ATC address
93  * spaces into a unified pointer space.  The method we take for 64b mode is
94  * to map the full 40b GPUVM address space into the hole of the 64b address
95  * space.
96 
97  * The GPUVM_Base/GPUVM_Limit defines the aperture in the 64b space where we
98  * direct requests to be translated via GPUVM page tables instead of the
99  * IOMMU path.
100  *
101  *
102  * 64b to 49b Address conversion
103  *
104  * Note that there are still significant portions of unused regions (holes)
105  * in the 64b address space even for the GPU.  There are several places in
106  * the pipeline (sw and hw), we wish to compress the 64b virtual address
107  * to a 49b address.  This 49b address is constituted of an “ATC” bit
108  * plus a 48b virtual address.  This 49b address is what is passed to the
109  * translation hardware.  ATC==0 means the 48b address is a GPUVM address
110  * (max of 2^40 – 1) intended to be translated via GPUVM page tables.
111  * ATC==1 means the 48b address is intended to be translated via IOMMU
112  * page tables.
113  *
114  * A 64b pointer is compared to the apertures that are defined (Base/Limit), in
115  * this case the GPUVM aperture (red) is defined and if a pointer falls in this
116  * aperture, we subtract the GPUVM_Base address and set the ATC bit to zero
117  * as part of the 64b to 49b conversion.
118  *
119  * Where this 64b to 49b conversion is done is a function of the usage.
120  * Most GPU memory access is via memory objects where the driver builds
121  * a descriptor which consists of a base address and a memory access by
122  * the GPU usually consists of some kind of an offset or Cartesian coordinate
123  * that references this memory descriptor.  This is the case for shader
124  * instructions that reference the T# or V# constants, or for specified
125  * locations of assets (ex. the shader program location).  In these cases
126  * the driver is what handles the 64b to 49b conversion and the base
127  * address in the descriptor (ex. V# or T# or shader program location)
128  * is defined as a 48b address w/ an ATC bit.  For this usage a given
129  * memory object cannot straddle multiple apertures in the 64b address
130  * space. For example a shader program cannot jump in/out between ATC
131  * and GPUVM space.
132  *
133  * In some cases we wish to pass a 64b pointer to the GPU hardware and
134  * the GPU hw does the 64b to 49b conversion before passing memory
135  * requests to the cache/memory system.  This is the case for the
136  * S_LOAD and FLAT_* shader memory instructions where we have 64b pointers
137  * in scalar and vector GPRs respectively.
138  *
139  * In all cases (no matter where the 64b -> 49b conversion is done), the gfxip
140  * hardware sends a 48b address along w/ an ATC bit, to the memory controller
141  * on the memory request interfaces.
142  *
143  *	<client>_MC_rdreq_atc   // read request ATC bit
144  *
145  *		0 : <client>_MC_rdreq_addr is a GPUVM VA
146  *
147  *		1 : <client>_MC_rdreq_addr is a ATC VA
148  *
149  *
150  * “Spare” aperture (APE1)
151  *
152  * We use the GPUVM aperture to differentiate ATC vs. GPUVM, but we also use
153  * apertures to set the Mtype field for S_LOAD/FLAT_* ops which is input to the
154  * config tables for setting cache policies. The “spare” (APE1) aperture is
155  * motivated by getting a different Mtype from the default.
156  * The default aperture isn’t an actual base/limit aperture; it is just the
157  * address space that doesn’t hit any defined base/limit apertures.
158  * The following diagram is a complete picture of the gfxip7.x SUA apertures.
159  * The APE1 can be placed either below or above
160  * the hole (cannot be in the hole).
161  *
162  *
163  * General Aperture definitions and rules
164  *
165  * An aperture register definition consists of a Base, Limit, Mtype, and
166  * usually an ATC bit indicating which translation tables that aperture uses.
167  * In all cases (for SUA and DUA apertures discussed later), aperture base
168  * and limit definitions are 64KB aligned.
169  *
170  *	<ape>_Base[63:0] = { <ape>_Base_register[63:16], 0x0000 }
171  *
172  *	<ape>_Limit[63:0] = { <ape>_Limit_register[63:16], 0xFFFF }
173  *
174  * The base and limit are considered inclusive to an aperture so being
175  * inside an aperture means (address >= Base) AND (address <= Limit).
176  *
177  * In no case is a payload that straddles multiple apertures expected to work.
178  * For example a load_dword_x4 that starts in one aperture and ends in another,
179  * does not work.  For the vector FLAT_* ops we have detection capability in
180  * the shader for reporting a “memory violation” back to the
181  * SQ block for use in traps.
182  * A memory violation results when an op falls into the hole,
183  * or a payload straddles multiple apertures.  The S_LOAD instruction
184  * does not have this detection.
185  *
186  * Apertures cannot overlap.
187  *
188  *
189  *
190  * HSA32 - ATC/IOMMU 32b
191  *
192  * For HSA32 mode, the pointers are interpreted as 32 bits and use a single GPR
193  * instead of two for the S_LOAD and FLAT_* ops. The entire GPUVM space of 40b
194  * will not fit so there is only partial visibility to the GPUVM
195  * space (defined by the aperture) for S_LOAD and FLAT_* ops.
196  * There is no spare (APE1) aperture for HSA32 mode.
197  *
198  *
199  * GPUVM 64b mode (driver model)
200  *
201  * This mode is related to HSA64 in that the difference really is that
202  * the default aperture is GPUVM (ATC==0) and not ATC space.
203  * We have gfxip7.x hardware that has FLAT_* and S_LOAD support for
204  * SUA GPUVM mode, but does not support HSA32/HSA64.
205  *
206  *
207  * Device Unified Address - DUA
208  *
209  * Device unified address (DUA) is the name of the feature that maps the
210  * Shared(LDS) memory and Private(Scratch) memory into the overall address
211  * space for use by the new FLAT_* vector memory ops.  The Shared and
212  * Private memories are mapped as apertures into the address space,
213  * and the hardware detects when a FLAT_* memory request is to be redirected
214  * to the LDS or Scratch memory when it falls into one of these apertures.
215  * Like the SUA apertures, the Shared/Private apertures are 64KB aligned and
216  * the base/limit is “in” the aperture. For both HSA64 and GPUVM SUA modes,
217  * the Shared/Private apertures are always placed in a limited selection of
218  * options in the hole of the 64b address space. For HSA32 mode, the
219  * Shared/Private apertures can be placed anywhere in the 32b space
220  * except at 0.
221  *
222  *
223  * HSA64 Apertures for FLAT_* vector ops
224  *
225  * For HSA64 SUA mode, the Shared and Private apertures are always placed
226  * in the hole w/ a limited selection of possible locations. The requests
227  * that fall in the private aperture are expanded as a function of the
228  * work-item id (tid) and redirected to the location of the
229  * “hidden private memory”. The hidden private can be placed in either GPUVM
230  * or ATC space. The addresses that fall in the shared aperture are
231  * re-directed to the on-chip LDS memory hardware.
232  *
233  *
234  * HSA32 Apertures for FLAT_* vector ops
235  *
236  * In HSA32 mode, the Private and Shared apertures can be placed anywhere
237  * in the 32b space except at 0 (Private or Shared Base at zero disables
238  * the apertures). If the base address of the apertures are non-zero
239  * (ie apertures exists), the size is always 64KB.
240  *
241  *
242  * GPUVM Apertures for FLAT_* vector ops
243  *
244  * In GPUVM mode, the Shared/Private apertures are specified identically
245  * to HSA64 mode where they are always in the hole at a limited selection
246  * of locations.
247  *
248  *
249  * Aperture Definitions for SUA and DUA
250  *
251  * The interpretation of the aperture register definitions for a given
252  * VMID is a function of the “SUA Mode” which is one of HSA64, HSA32, or
253  * GPUVM64 discussed in previous sections. The mode is first decoded, and
254  * then the remaining register decode is a function of the mode.
255  *
256  *
257  * SUA Mode Decode
258  *
259  * For the S_LOAD and FLAT_* shader operations, the SUA mode is decoded from
260  * the COMPUTE_DISPATCH_INITIATOR:DATA_ATC bit and
261  * the SH_MEM_CONFIG:PTR32 bits.
262  *
263  * COMPUTE_DISPATCH_INITIATOR:DATA_ATC    SH_MEM_CONFIG:PTR32        Mode
264  *
265  * 1                                              0                  HSA64
266  *
267  * 1                                              1                  HSA32
268  *
269  * 0                                              X                 GPUVM64
270  *
271  * In general the hardware will ignore the PTR32 bit and treat
272  * as “0” whenever DATA_ATC = “0”, but sw should set PTR32=0
273  * when DATA_ATC=0.
274  *
275  * The DATA_ATC bit is only set for compute dispatches.
276  * All “Draw” dispatches are hardcoded to GPUVM64 mode
277  * for FLAT_* / S_LOAD operations.
278  */
279 
280 #define MAKE_GPUVM_APP_BASE_VI(gpu_num) \
281 	(((uint64_t)(gpu_num) << 61) + 0x1000000000000L)
282 
283 #define MAKE_GPUVM_APP_LIMIT(base, size) \
284 	(((uint64_t)(base) & 0xFFFFFF0000000000UL) + (size) - 1)
285 
286 #define MAKE_SCRATCH_APP_BASE_VI() \
287 	(((uint64_t)(0x1UL) << 61) + 0x100000000L)
288 
289 #define MAKE_SCRATCH_APP_LIMIT(base) \
290 	(((uint64_t)base & 0xFFFFFFFF00000000UL) | 0xFFFFFFFF)
291 
292 #define MAKE_LDS_APP_BASE_VI() \
293 	(((uint64_t)(0x1UL) << 61) + 0x0)
294 #define MAKE_LDS_APP_LIMIT(base) \
295 	(((uint64_t)(base) & 0xFFFFFFFF00000000UL) | 0xFFFFFFFF)
296 
297 /* On GFXv9 the LDS and scratch apertures are programmed independently
298  * using the high 16 bits of the 64-bit virtual address. They must be
299  * in the hole, which will be the case as long as the high 16 bits are
300  * not 0.
301  *
302  * The aperture sizes are still 4GB implicitly.
303  *
304  * A GPUVM aperture is not applicable on GFXv9.
305  */
306 #define MAKE_LDS_APP_BASE_V9() ((uint64_t)(0x1UL) << 48)
307 #define MAKE_SCRATCH_APP_BASE_V9() ((uint64_t)(0x2UL) << 48)
308 
309 /* User mode manages most of the SVM aperture address space. The low
310  * 16MB are reserved for kernel use (CWSR trap handler and kernel IB
311  * for now).
312  */
313 #define SVM_USER_BASE (u64)(KFD_CWSR_TBA_TMA_SIZE + 2*PAGE_SIZE)
314 #define SVM_CWSR_BASE (SVM_USER_BASE - KFD_CWSR_TBA_TMA_SIZE)
315 #define SVM_IB_BASE   (SVM_CWSR_BASE - PAGE_SIZE)
316 
317 static void kfd_init_apertures_vi(struct kfd_process_device *pdd, uint8_t id)
318 {
319 	/*
320 	 * node id couldn't be 0 - the three MSB bits of
321 	 * aperture shouldn't be 0
322 	 */
323 	pdd->lds_base = MAKE_LDS_APP_BASE_VI();
324 	pdd->lds_limit = MAKE_LDS_APP_LIMIT(pdd->lds_base);
325 
326 	/* dGPUs: SVM aperture starting at 0
327 	 * with small reserved space for kernel.
328 	 * Set them to CANONICAL addresses.
329 	 */
330 	pdd->gpuvm_base = max(SVM_USER_BASE, AMDGPU_VA_RESERVED_BOTTOM);
331 	pdd->gpuvm_limit =
332 		pdd->dev->kfd->shared_resources.gpuvm_size - 1;
333 
334 	/* dGPUs: the reserved space for kernel
335 	 * before SVM
336 	 */
337 	pdd->qpd.cwsr_base = SVM_CWSR_BASE;
338 	pdd->qpd.ib_base = SVM_IB_BASE;
339 
340 	pdd->scratch_base = MAKE_SCRATCH_APP_BASE_VI();
341 	pdd->scratch_limit = MAKE_SCRATCH_APP_LIMIT(pdd->scratch_base);
342 }
343 
344 static void kfd_init_apertures_v9(struct kfd_process_device *pdd, uint8_t id)
345 {
346 	pdd->lds_base = MAKE_LDS_APP_BASE_V9();
347 	pdd->lds_limit = MAKE_LDS_APP_LIMIT(pdd->lds_base);
348 
349 	pdd->gpuvm_base = AMDGPU_VA_RESERVED_BOTTOM;
350 	pdd->gpuvm_limit =
351 		pdd->dev->kfd->shared_resources.gpuvm_size - 1;
352 
353 	pdd->scratch_base = MAKE_SCRATCH_APP_BASE_V9();
354 	pdd->scratch_limit = MAKE_SCRATCH_APP_LIMIT(pdd->scratch_base);
355 
356 	/*
357 	 * Place TBA/TMA on opposite side of VM hole to prevent
358 	 * stray faults from triggering SVM on these pages.
359 	 */
360 	pdd->qpd.cwsr_base = AMDGPU_VA_RESERVED_TRAP_START(pdd->dev->adev);
361 }
362 
363 int kfd_init_apertures(struct kfd_process *process)
364 {
365 	uint8_t id  = 0;
366 	struct kfd_node *dev;
367 	struct kfd_process_device *pdd;
368 
369 	/*Iterating over all devices*/
370 	while (kfd_topology_enum_kfd_devices(id, &dev) == 0) {
371 		if (!dev || kfd_devcgroup_check_permission(dev)) {
372 			/* Skip non GPU devices and devices to which the
373 			 * current process have no access to. Access can be
374 			 * limited by placing the process in a specific
375 			 * cgroup hierarchy
376 			 */
377 			id++;
378 			continue;
379 		}
380 
381 		pdd = kfd_create_process_device_data(dev, process);
382 		if (!pdd) {
383 			dev_err(dev->adev->dev,
384 				"Failed to create process device data\n");
385 			return -ENOMEM;
386 		}
387 		/*
388 		 * For 64 bit process apertures will be statically reserved in
389 		 * the x86_64 non canonical process address space
390 		 * amdkfd doesn't currently support apertures for 32 bit process
391 		 */
392 		if (process->is_32bit_user_mode) {
393 			pdd->lds_base = pdd->lds_limit = 0;
394 			pdd->gpuvm_base = pdd->gpuvm_limit = 0;
395 			pdd->scratch_base = pdd->scratch_limit = 0;
396 		} else {
397 			switch (dev->adev->asic_type) {
398 			case CHIP_KAVERI:
399 			case CHIP_HAWAII:
400 			case CHIP_CARRIZO:
401 			case CHIP_TONGA:
402 			case CHIP_FIJI:
403 			case CHIP_POLARIS10:
404 			case CHIP_POLARIS11:
405 			case CHIP_POLARIS12:
406 			case CHIP_VEGAM:
407 				kfd_init_apertures_vi(pdd, id);
408 				break;
409 			default:
410 				if (KFD_GC_VERSION(dev) >= IP_VERSION(9, 0, 1))
411 					kfd_init_apertures_v9(pdd, id);
412 				else {
413 					WARN(1, "Unexpected ASIC family %u",
414 					     dev->adev->asic_type);
415 					return -EINVAL;
416 				}
417 			}
418 		}
419 
420 		dev_dbg(kfd_device, "node id %u\n", id);
421 		dev_dbg(kfd_device, "gpu id %u\n", pdd->dev->id);
422 		dev_dbg(kfd_device, "lds_base %llX\n", pdd->lds_base);
423 		dev_dbg(kfd_device, "lds_limit %llX\n", pdd->lds_limit);
424 		dev_dbg(kfd_device, "gpuvm_base %llX\n", pdd->gpuvm_base);
425 		dev_dbg(kfd_device, "gpuvm_limit %llX\n", pdd->gpuvm_limit);
426 		dev_dbg(kfd_device, "scratch_base %llX\n", pdd->scratch_base);
427 		dev_dbg(kfd_device, "scratch_limit %llX\n", pdd->scratch_limit);
428 
429 		id++;
430 	}
431 
432 	return 0;
433 }
434