amd/amdkfd/kfd_doorbell.c

0001 // SPDX-License-Identifier: GPL-2.0 OR MIT
0002 /*
0003  * Copyright 2014-2022 Advanced Micro Devices, Inc.
0004  *
0005  * Permission is hereby granted, free of charge, to any person obtaining a
0006  * copy of this software and associated documentation files (the "Software"),
0007  * to deal in the Software without restriction, including without limitation
0008  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
0009  * and/or sell copies of the Software, and to permit persons to whom the
0010  * Software is furnished to do so, subject to the following conditions:
0011  *
0012  * The above copyright notice and this permission notice shall be included in
0013  * all copies or substantial portions of the Software.
0014  *
0015  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
0016  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
0017  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
0018  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
0019  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
0020  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
0021  * OTHER DEALINGS IN THE SOFTWARE.
0022  */
0023 #include "kfd_priv.h"
0024 #include <linux/mm.h>
0025 #include <linux/mman.h>
0026 #include <linux/slab.h>
0027 #include <linux/io.h>
0028 #include <linux/idr.h>
0029
0030 /*
0031  * This extension supports a kernel level doorbells management for the
0032  * kernel queues using the first doorbell page reserved for the kernel.
0033  */
0034
0035 /*
0036  * Each device exposes a doorbell aperture, a PCI MMIO aperture that
0037  * receives 32-bit writes that are passed to queues as wptr values.
0038  * The doorbells are intended to be written by applications as part
0039  * of queueing work on user-mode queues.
0040  * We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks.
0041  * We map the doorbell address space into user-mode when a process creates
0042  * its first queue on each device.
0043  * Although the mapping is done by KFD, it is equivalent to an mmap of
0044  * the /dev/kfd with the particular device encoded in the mmap offset.
0045  * There will be other uses for mmap of /dev/kfd, so only a range of
0046  * offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells.
0047  */
0048
0049 /* # of doorbell bytes allocated for each process. */
0050 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd)
0051 {
0052     if (!kfd->shared_resources.enable_mes)
0053         return roundup(kfd->device_info.doorbell_size *
0054                 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
0055                 PAGE_SIZE);
0056     else
0057         return amdgpu_mes_doorbell_process_slice(
0058                     (struct amdgpu_device *)kfd->adev);
0059 }
0060
0061 /* Doorbell calculations for device init. */
0062 int kfd_doorbell_init(struct kfd_dev *kfd)
0063 {
0064     size_t doorbell_start_offset;
0065     size_t doorbell_aperture_size;
0066     size_t doorbell_process_limit;
0067
0068     /*
0069      * With MES enabled, just set the doorbell base as it is needed
0070      * to calculate doorbell physical address.
0071      */
0072     if (kfd->shared_resources.enable_mes) {
0073         kfd->doorbell_base =
0074             kfd->shared_resources.doorbell_physical_address;
0075         return 0;
0076     }
0077
0078     /*
0079      * We start with calculations in bytes because the input data might
0080      * only be byte-aligned.
0081      * Only after we have done the rounding can we assume any alignment.
0082      */
0083
0084     doorbell_start_offset =
0085             roundup(kfd->shared_resources.doorbell_start_offset,
0086                     kfd_doorbell_process_slice(kfd));
0087
0088     doorbell_aperture_size =
0089             rounddown(kfd->shared_resources.doorbell_aperture_size,
0090                     kfd_doorbell_process_slice(kfd));
0091
0092     if (doorbell_aperture_size > doorbell_start_offset)
0093         doorbell_process_limit =
0094             (doorbell_aperture_size - doorbell_start_offset) /
0095                         kfd_doorbell_process_slice(kfd);
0096     else
0097         return -ENOSPC;
0098
0099     if (!kfd->max_doorbell_slices ||
0100         doorbell_process_limit < kfd->max_doorbell_slices)
0101         kfd->max_doorbell_slices = doorbell_process_limit;
0102
0103     kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address +
0104                 doorbell_start_offset;
0105
0106     kfd->doorbell_base_dw_offset = doorbell_start_offset / sizeof(u32);
0107
0108     kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
0109                        kfd_doorbell_process_slice(kfd));
0110
0111     if (!kfd->doorbell_kernel_ptr)
0112         return -ENOMEM;
0113
0114     pr_debug("Doorbell initialization:\n");
0115     pr_debug("doorbell base           == 0x%08lX\n",
0116             (uintptr_t)kfd->doorbell_base);
0117
0118     pr_debug("doorbell_base_dw_offset      == 0x%08lX\n",
0119             kfd->doorbell_base_dw_offset);
0120
0121     pr_debug("doorbell_process_limit  == 0x%08lX\n",
0122             doorbell_process_limit);
0123
0124     pr_debug("doorbell_kernel_offset  == 0x%08lX\n",
0125             (uintptr_t)kfd->doorbell_base);
0126
0127     pr_debug("doorbell aperture size  == 0x%08lX\n",
0128             kfd->shared_resources.doorbell_aperture_size);
0129
0130     pr_debug("doorbell kernel address == %p\n", kfd->doorbell_kernel_ptr);
0131
0132     return 0;
0133 }
0134
0135 void kfd_doorbell_fini(struct kfd_dev *kfd)
0136 {
0137     if (kfd->doorbell_kernel_ptr)
0138         iounmap(kfd->doorbell_kernel_ptr);
0139 }
0140
0141 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
0142               struct vm_area_struct *vma)
0143 {
0144     phys_addr_t address;
0145     struct kfd_process_device *pdd;
0146
0147     /*
0148      * For simplicitly we only allow mapping of the entire doorbell
0149      * allocation of a single device & process.
0150      */
0151     if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev))
0152         return -EINVAL;
0153
0154     pdd = kfd_get_process_device_data(dev, process);
0155     if (!pdd)
0156         return -EINVAL;
0157
0158     /* Calculate physical address of doorbell */
0159     address = kfd_get_process_doorbells(pdd);
0160     vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
0161                 VM_DONTDUMP | VM_PFNMAP;
0162
0163     vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
0164
0165     pr_debug("Mapping doorbell page\n"
0166          "     target user address == 0x%08llX\n"
0167          "     physical address    == 0x%08llX\n"
0168          "     vm_flags            == 0x%04lX\n"
0169          "     size                == 0x%04lX\n",
0170          (unsigned long long) vma->vm_start, address, vma->vm_flags,
0171          kfd_doorbell_process_slice(dev));
0172
0173
0174     return io_remap_pfn_range(vma,
0175                 vma->vm_start,
0176                 address >> PAGE_SHIFT,
0177                 kfd_doorbell_process_slice(dev),
0178                 vma->vm_page_prot);
0179 }
0180
0181
0182 /* get kernel iomem pointer for a doorbell */
0183 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
0184                     unsigned int *doorbell_off)
0185 {
0186     u32 inx;
0187
0188     mutex_lock(&kfd->doorbell_mutex);
0189     inx = find_first_zero_bit(kfd->doorbell_available_index,
0190                     KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
0191
0192     __set_bit(inx, kfd->doorbell_available_index);
0193     mutex_unlock(&kfd->doorbell_mutex);
0194
0195     if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
0196         return NULL;
0197
0198     inx *= kfd->device_info.doorbell_size / sizeof(u32);
0199
0200     /*
0201      * Calculating the kernel doorbell offset using the first
0202      * doorbell page.
0203      */
0204     *doorbell_off = kfd->doorbell_base_dw_offset + inx;
0205
0206     pr_debug("Get kernel queue doorbell\n"
0207             "     doorbell offset   == 0x%08X\n"
0208             "     doorbell index    == 0x%x\n",
0209         *doorbell_off, inx);
0210
0211     return kfd->doorbell_kernel_ptr + inx;
0212 }
0213
0214 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
0215 {
0216     unsigned int inx;
0217
0218     inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr)
0219         * sizeof(u32) / kfd->device_info.doorbell_size;
0220
0221     mutex_lock(&kfd->doorbell_mutex);
0222     __clear_bit(inx, kfd->doorbell_available_index);
0223     mutex_unlock(&kfd->doorbell_mutex);
0224 }
0225
0226 void write_kernel_doorbell(void __iomem *db, u32 value)
0227 {
0228     if (db) {
0229         writel(value, db);
0230         pr_debug("Writing %d to doorbell address %p\n", value, db);
0231     }
0232 }
0233
0234 void write_kernel_doorbell64(void __iomem *db, u64 value)
0235 {
0236     if (db) {
0237         WARN(((unsigned long)db & 7) != 0,
0238              "Unaligned 64-bit doorbell");
0239         writeq(value, (u64 __iomem *)db);
0240         pr_debug("writing %llu to doorbell address %p\n", value, db);
0241     }
0242 }
0243
0244 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
0245                     struct kfd_process_device *pdd,
0246                     unsigned int doorbell_id)
0247 {
0248     /*
0249      * doorbell_base_dw_offset accounts for doorbells taken by KGD.
0250      * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to
0251      * the process's doorbells. The offset returned is in dword
0252      * units regardless of the ASIC-dependent doorbell size.
0253      */
0254     if (!kfd->shared_resources.enable_mes)
0255         return kfd->doorbell_base_dw_offset +
0256             pdd->doorbell_index
0257             * kfd_doorbell_process_slice(kfd) / sizeof(u32) +
0258             doorbell_id *
0259             kfd->device_info.doorbell_size / sizeof(u32);
0260     else
0261         return amdgpu_mes_get_doorbell_dw_offset_in_bar(
0262                 (struct amdgpu_device *)kfd->adev,
0263                 pdd->doorbell_index, doorbell_id);
0264 }
0265
0266 uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
0267 {
0268     uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
0269                 kfd->shared_resources.doorbell_start_offset) /
0270                     kfd_doorbell_process_slice(kfd) + 1;
0271
0272     return num_of_elems;
0273
0274 }
0275
0276 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd)
0277 {
0278     return pdd->dev->doorbell_base +
0279         pdd->doorbell_index * kfd_doorbell_process_slice(pdd->dev);
0280 }
0281
0282 int kfd_alloc_process_doorbells(struct kfd_dev *kfd, unsigned int *doorbell_index)
0283 {
0284     int r = 0;
0285
0286     if (!kfd->shared_resources.enable_mes)
0287         r = ida_simple_get(&kfd->doorbell_ida, 1,
0288                    kfd->max_doorbell_slices, GFP_KERNEL);
0289     else
0290         r = amdgpu_mes_alloc_process_doorbells(
0291                 (struct amdgpu_device *)kfd->adev,
0292                 doorbell_index);
0293
0294     if (r > 0)
0295         *doorbell_index = r;
0296
0297     return r;
0298 }
0299
0300 void kfd_free_process_doorbells(struct kfd_dev *kfd, unsigned int doorbell_index)
0301 {
0302     if (doorbell_index) {
0303         if (!kfd->shared_resources.enable_mes)
0304             ida_simple_remove(&kfd->doorbell_ida, doorbell_index);
0305         else
0306             amdgpu_mes_free_process_doorbells(
0307                     (struct amdgpu_device *)kfd->adev,
0308                     doorbell_index);
0309     }
0310 }