1 /* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
2 /* Copyright (c) 2021, Microsoft Corporation. */
7 #include <linux/dma-mapping.h>
8 #include <linux/netdevice.h>
10 #include "shm_channel.h"
12 #define GDMA_STATUS_MORE_ENTRIES 0x00000105
14 /* Structures labeled with "HW DATA" are exchanged with the hardware. All of
15 * them are naturally aligned and hence don't need __packed.
18 enum gdma_request_type {
19 GDMA_VERIFY_VF_DRIVER_VERSION = 1,
20 GDMA_QUERY_MAX_RESOURCES = 2,
21 GDMA_LIST_DEVICES = 3,
22 GDMA_REGISTER_DEVICE = 4,
23 GDMA_DEREGISTER_DEVICE = 5,
24 GDMA_GENERATE_TEST_EQE = 10,
25 GDMA_CREATE_QUEUE = 12,
26 GDMA_DISABLE_QUEUE = 13,
27 GDMA_ALLOCATE_RESOURCE_RANGE = 22,
28 GDMA_DESTROY_RESOURCE_RANGE = 24,
29 GDMA_CREATE_DMA_REGION = 25,
30 GDMA_DMA_REGION_ADD_PAGES = 26,
31 GDMA_DESTROY_DMA_REGION = 27,
36 GDMA_QUERY_HWC_TIMEOUT = 84, /* 0x54 */
39 #define GDMA_RESOURCE_DOORBELL_PAGE 27
41 enum gdma_queue_type {
49 enum gdma_work_request_flags {
51 GDMA_WR_OOB_IN_SGL = BIT(0),
52 GDMA_WR_PAD_BY_SGE0 = BIT(1),
56 GDMA_EQE_COMPLETION = 3,
57 GDMA_EQE_TEST_EVENT = 64,
58 GDMA_EQE_HWC_INIT_EQ_ID_DB = 129,
59 GDMA_EQE_HWC_INIT_DATA = 130,
60 GDMA_EQE_HWC_INIT_DONE = 131,
61 GDMA_EQE_HWC_SOC_RECONFIG = 132,
62 GDMA_EQE_HWC_SOC_RECONFIG_DATA = 133,
63 GDMA_EQE_RNIC_QP_FATAL = 176,
70 GDMA_DEVICE_MANA_IB = 3,
73 struct gdma_resource {
74 /* Protect the bitmap */
77 /* The bitmap size in bits. */
80 /* The bitmap tracks the resources. */
84 union gdma_doorbell_entry {
114 struct gdma_msg_hdr {
133 struct gdma_req_hdr {
134 struct gdma_msg_hdr req;
135 struct gdma_msg_hdr resp; /* The expected response */
136 struct gdma_dev_id dev_id;
140 struct gdma_resp_hdr {
141 struct gdma_msg_hdr response;
142 struct gdma_dev_id dev_id;
148 struct gdma_general_req {
149 struct gdma_req_hdr hdr;
152 #define GDMA_MESSAGE_V1 1
153 #define GDMA_MESSAGE_V2 2
154 #define GDMA_MESSAGE_V3 3
156 struct gdma_general_resp {
157 struct gdma_resp_hdr hdr;
160 #define GDMA_STANDARD_HEADER_TYPE 0
162 static inline void mana_gd_init_req_hdr(struct gdma_req_hdr *hdr, u32 code,
163 u32 req_size, u32 resp_size)
165 hdr->req.hdr_type = GDMA_STANDARD_HEADER_TYPE;
166 hdr->req.msg_type = code;
167 hdr->req.msg_version = GDMA_MESSAGE_V1;
168 hdr->req.msg_size = req_size;
170 hdr->resp.hdr_type = GDMA_STANDARD_HEADER_TYPE;
171 hdr->resp.msg_type = code;
172 hdr->resp.msg_version = GDMA_MESSAGE_V1;
173 hdr->resp.msg_size = resp_size;
176 /* The 16-byte struct is part of the GDMA work queue entry (WQE). */
183 struct gdma_wqe_request {
184 struct gdma_sge *sgl;
188 const void *inline_oob_data;
191 u32 client_data_unit;
194 enum gdma_page_type {
198 #define GDMA_INVALID_DMA_REGION 0
200 struct gdma_mem_info {
203 dma_addr_t dma_handle;
207 /* Allocated by the PF driver */
208 u64 dma_region_handle;
211 #define REGISTER_ATB_MST_MKEY_LOWER_SIZE 8
214 struct gdma_context *gdma_context;
216 struct gdma_dev_id dev_id;
222 /* GDMA driver specific pointer */
225 struct auxiliary_device *adev;
228 /* MANA_PAGE_SIZE is the DMA unit */
229 #define MANA_PAGE_SHIFT 12
230 #define MANA_PAGE_SIZE BIT(MANA_PAGE_SHIFT)
231 #define MANA_PAGE_ALIGN(x) ALIGN((x), MANA_PAGE_SIZE)
232 #define MANA_PAGE_ALIGNED(addr) IS_ALIGNED((unsigned long)(addr), MANA_PAGE_SIZE)
233 #define MANA_PFN(a) ((a) >> MANA_PAGE_SHIFT)
236 #define MANA_MIN_QSIZE MANA_PAGE_SIZE
238 #define GDMA_CQE_SIZE 64
239 #define GDMA_EQE_SIZE 16
240 #define GDMA_MAX_SQE_SIZE 512
241 #define GDMA_MAX_RQE_SIZE 256
243 #define GDMA_COMP_DATA_SIZE 0x3C
245 #define GDMA_EVENT_DATA_SIZE 0xC
247 /* The WQE size must be a multiple of the Basic Unit, which is 32 bytes. */
248 #define GDMA_WQE_BU_SIZE 32
250 #define INVALID_PDID UINT_MAX
251 #define INVALID_DOORBELL UINT_MAX
252 #define INVALID_MEM_KEY UINT_MAX
253 #define INVALID_QUEUE_ID UINT_MAX
254 #define INVALID_PCI_MSIX_INDEX UINT_MAX
257 u32 cqe_data[GDMA_COMP_DATA_SIZE / 4];
263 u32 details[GDMA_EVENT_DATA_SIZE / 4];
270 struct gdma_queue *eq;
271 struct dentry *mana_eq_debugfs;
274 typedef void gdma_eq_callback(void *context, struct gdma_queue *q,
275 struct gdma_event *e);
277 typedef void gdma_cq_callback(void *context, struct gdma_queue *q);
279 /* The 'head' is the producer index. For SQ/RQ, when the driver posts a WQE
280 * (Note: the WQE size must be a multiple of the 32-byte Basic Unit), the
281 * driver increases the 'head' in BUs rather than in bytes, and notifies
282 * the HW of the updated head. For EQ/CQ, the driver uses the 'head' to track
283 * the HW head, and increases the 'head' by 1 for every processed EQE/CQE.
285 * The 'tail' is the consumer index for SQ/RQ. After the CQE of the SQ/RQ is
286 * processed, the driver increases the 'tail' to indicate that WQEs have
287 * been consumed by the HW, so the driver can post new WQEs into the SQ/RQ.
289 * The driver doesn't use the 'tail' for EQ/CQ, because the driver ensures
290 * that the EQ/CQ is big enough so they can't overflow, and the driver uses
291 * the owner bits mechanism to detect if the queue has become empty.
294 struct gdma_dev *gdma_dev;
296 enum gdma_queue_type type;
299 struct gdma_mem_info mem_info;
304 bool monitor_avl_buf;
308 struct list_head entry;
310 /* Extra fields specific to EQ/CQ. */
315 gdma_eq_callback *callback;
318 unsigned int msix_index;
320 u32 log2_throttle_limit;
324 gdma_cq_callback *callback;
327 struct gdma_queue *parent; /* For CQ/EQ relationship */
332 struct gdma_queue_spec {
333 enum gdma_queue_type type;
334 bool monitor_avl_buf;
335 unsigned int queue_size;
337 /* Extra fields specific to EQ/CQ. */
340 gdma_eq_callback *callback;
343 unsigned long log2_throttle_limit;
344 unsigned int msix_index;
348 gdma_cq_callback *callback;
351 struct gdma_queue *parent_eq;
357 #define MANA_IRQ_NAME_SZ 32
359 struct gdma_irq_context {
360 void (*handler)(void *arg);
361 /* Protect the eq_list */
363 struct list_head eq_list;
364 char name[MANA_IRQ_NAME_SZ];
367 struct gdma_context {
369 struct dentry *mana_pci_debugfs;
371 /* Per-vPort max number of queues */
372 unsigned int max_num_queues;
373 unsigned int max_num_msix;
374 unsigned int num_msix_usable;
375 struct gdma_irq_context *irq_contexts;
380 /* This maps a CQ index to the queue structure. */
381 unsigned int max_num_cqs;
382 struct gdma_queue **cq_table;
384 /* Protect eq_test_event and test_event_eq_id */
385 struct mutex eq_test_event_mutex;
386 struct completion eq_test_event;
387 u32 test_event_eq_id;
391 void __iomem *bar0_va;
392 void __iomem *shm_base;
393 void __iomem *db_page_base;
394 phys_addr_t phys_db_page_base;
398 /* Shared memory chanenl (used to bootstrap HWC) */
399 struct shm_channel shm_channel;
401 /* Hardware communication channel (HWC) */
404 /* Azure network adapter */
405 struct gdma_dev mana;
407 /* Azure RDMA adapter */
408 struct gdma_dev mana_ib;
411 #define MAX_NUM_GDMA_DEVICES 4
413 static inline bool mana_gd_is_mana(struct gdma_dev *gd)
415 return gd->dev_id.type == GDMA_DEVICE_MANA;
418 static inline bool mana_gd_is_hwc(struct gdma_dev *gd)
420 return gd->dev_id.type == GDMA_DEVICE_HWC;
423 u8 *mana_gd_get_wqe_ptr(const struct gdma_queue *wq, u32 wqe_offset);
424 u32 mana_gd_wq_avail_space(struct gdma_queue *wq);
426 int mana_gd_test_eq(struct gdma_context *gc, struct gdma_queue *eq);
428 int mana_gd_create_hwc_queue(struct gdma_dev *gd,
429 const struct gdma_queue_spec *spec,
430 struct gdma_queue **queue_ptr);
432 int mana_gd_create_mana_eq(struct gdma_dev *gd,
433 const struct gdma_queue_spec *spec,
434 struct gdma_queue **queue_ptr);
436 int mana_gd_create_mana_wq_cq(struct gdma_dev *gd,
437 const struct gdma_queue_spec *spec,
438 struct gdma_queue **queue_ptr);
440 void mana_gd_destroy_queue(struct gdma_context *gc, struct gdma_queue *queue);
442 int mana_gd_poll_cq(struct gdma_queue *cq, struct gdma_comp *comp, int num_cqe);
444 void mana_gd_ring_cq(struct gdma_queue *cq, u8 arm_bit);
455 u32 inline_oob_size_div4:3;
456 u32 client_oob_in_sgl :1;
458 u32 client_data_unit :14;
464 #define INLINE_OOB_SMALL_SIZE 8
465 #define INLINE_OOB_LARGE_SIZE 24
467 #define MAX_TX_WQE_SIZE 512
468 #define MAX_RX_WQE_SIZE 256
470 #define MAX_TX_WQE_SGL_ENTRIES ((GDMA_MAX_SQE_SIZE - \
471 sizeof(struct gdma_sge) - INLINE_OOB_SMALL_SIZE) / \
472 sizeof(struct gdma_sge))
474 #define MAX_RX_WQE_SGL_ENTRIES ((GDMA_MAX_RQE_SIZE - \
475 sizeof(struct gdma_sge)) / sizeof(struct gdma_sge))
478 u32 cqe_data[GDMA_COMP_DATA_SIZE / 4];
492 #define GDMA_CQE_OWNER_BITS 3
494 #define GDMA_CQE_OWNER_MASK ((1 << GDMA_CQE_OWNER_BITS) - 1)
496 #define SET_ARM_BIT 1
498 #define GDMA_EQE_OWNER_BITS 3
500 union gdma_eqe_info {
512 #define GDMA_EQE_OWNER_MASK ((1 << GDMA_EQE_OWNER_BITS) - 1)
513 #define INITIALIZED_OWNER_BIT(log2_num_entries) (1UL << (log2_num_entries))
516 u32 details[GDMA_EVENT_DATA_SIZE / 4];
520 #define GDMA_REG_DB_PAGE_OFFSET 8
521 #define GDMA_REG_DB_PAGE_SIZE 0x10
522 #define GDMA_REG_SHM_OFFSET 0x18
524 #define GDMA_PF_REG_DB_PAGE_SIZE 0xD0
525 #define GDMA_PF_REG_DB_PAGE_OFF 0xC8
526 #define GDMA_PF_REG_SHM_OFF 0x70
528 #define GDMA_SRIOV_REG_CFG_BASE_OFF 0x108
530 #define MANA_PF_DEVICE_ID 0x00B9
531 #define MANA_VF_DEVICE_ID 0x00BA
533 struct gdma_posted_wqe_info {
537 /* GDMA_GENERATE_TEST_EQE */
538 struct gdma_generate_test_event_req {
539 struct gdma_req_hdr hdr;
543 /* GDMA_VERIFY_VF_DRIVER_VERSION */
545 GDMA_PROTOCOL_V1 = 1,
546 GDMA_PROTOCOL_FIRST = GDMA_PROTOCOL_V1,
547 GDMA_PROTOCOL_LAST = GDMA_PROTOCOL_V1,
550 #define GDMA_DRV_CAP_FLAG_1_EQ_SHARING_MULTI_VPORT BIT(0)
552 /* Advertise to the NIC firmware: the NAPI work_done variable race is fixed,
553 * so the driver is able to reliably support features like busy_poll.
555 #define GDMA_DRV_CAP_FLAG_1_NAPI_WKDONE_FIX BIT(2)
556 #define GDMA_DRV_CAP_FLAG_1_HWC_TIMEOUT_RECONFIG BIT(3)
557 #define GDMA_DRV_CAP_FLAG_1_VARIABLE_INDIRECTION_TABLE_SUPPORT BIT(5)
559 #define GDMA_DRV_CAP_FLAGS1 \
560 (GDMA_DRV_CAP_FLAG_1_EQ_SHARING_MULTI_VPORT | \
561 GDMA_DRV_CAP_FLAG_1_NAPI_WKDONE_FIX | \
562 GDMA_DRV_CAP_FLAG_1_HWC_TIMEOUT_RECONFIG | \
563 GDMA_DRV_CAP_FLAG_1_VARIABLE_INDIRECTION_TABLE_SUPPORT)
565 #define GDMA_DRV_CAP_FLAGS2 0
567 #define GDMA_DRV_CAP_FLAGS3 0
569 #define GDMA_DRV_CAP_FLAGS4 0
571 struct gdma_verify_ver_req {
572 struct gdma_req_hdr hdr;
574 /* Mandatory fields required for protocol establishment */
575 u64 protocol_ver_min;
576 u64 protocol_ver_max;
578 /* Gdma Driver Capability Flags */
579 u64 gd_drv_cap_flags1;
580 u64 gd_drv_cap_flags2;
581 u64 gd_drv_cap_flags3;
582 u64 gd_drv_cap_flags4;
584 /* Advisory fields */
586 u32 os_type; /* Linux = 0x10; Windows = 0x20; Other = 0x30 */
599 struct gdma_verify_ver_resp {
600 struct gdma_resp_hdr hdr;
601 u64 gdma_protocol_ver;
608 /* GDMA_QUERY_MAX_RESOURCES */
609 struct gdma_query_max_resources_resp {
610 struct gdma_resp_hdr hdr;
623 /* GDMA_LIST_DEVICES */
624 struct gdma_list_devices_resp {
625 struct gdma_resp_hdr hdr;
628 struct gdma_dev_id devs[64];
631 /* GDMA_REGISTER_DEVICE */
632 struct gdma_register_device_resp {
633 struct gdma_resp_hdr hdr;
639 struct gdma_allocate_resource_range_req {
640 struct gdma_req_hdr hdr;
644 u32 allocated_resources;
647 struct gdma_allocate_resource_range_resp {
648 struct gdma_resp_hdr hdr;
649 u32 allocated_resources;
652 struct gdma_destroy_resource_range_req {
653 struct gdma_req_hdr hdr;
656 u32 allocated_resources;
659 /* GDMA_CREATE_QUEUE */
660 struct gdma_create_queue_req {
661 struct gdma_req_hdr hdr;
669 u32 log2_throttle_limit;
670 u32 eq_pci_msix_index;
673 u8 rq_drop_on_overrun;
674 u8 rq_err_on_wqe_overflow;
675 u8 rq_chain_rec_wqes;
680 struct gdma_create_queue_resp {
681 struct gdma_resp_hdr hdr;
685 /* GDMA_DISABLE_QUEUE */
686 struct gdma_disable_queue_req {
687 struct gdma_req_hdr hdr;
690 u32 alloc_res_id_on_creation;
693 /* GDMA_QUERY_HWC_TIMEOUT */
694 struct gdma_query_hwc_timeout_req {
695 struct gdma_req_hdr hdr;
700 struct gdma_query_hwc_timeout_resp {
701 struct gdma_resp_hdr hdr;
720 enum gdma_mr_access_flags {
721 GDMA_ACCESS_FLAG_LOCAL_READ = BIT_ULL(0),
722 GDMA_ACCESS_FLAG_LOCAL_WRITE = BIT_ULL(1),
723 GDMA_ACCESS_FLAG_REMOTE_READ = BIT_ULL(2),
724 GDMA_ACCESS_FLAG_REMOTE_WRITE = BIT_ULL(3),
725 GDMA_ACCESS_FLAG_REMOTE_ATOMIC = BIT_ULL(4),
728 /* GDMA_CREATE_DMA_REGION */
729 struct gdma_create_dma_region_req {
730 struct gdma_req_hdr hdr;
732 /* The total size of the DMA region */
735 /* The offset in the first page */
738 /* enum gdma_page_type */
741 /* The total number of pages */
744 /* If page_addr_list_len is smaller than page_count,
745 * the remaining page addresses will be added via the
746 * message GDMA_DMA_REGION_ADD_PAGES.
748 u32 page_addr_list_len;
749 u64 page_addr_list[];
752 struct gdma_create_dma_region_resp {
753 struct gdma_resp_hdr hdr;
754 u64 dma_region_handle;
757 /* GDMA_DMA_REGION_ADD_PAGES */
758 struct gdma_dma_region_add_pages_req {
759 struct gdma_req_hdr hdr;
761 u64 dma_region_handle;
763 u32 page_addr_list_len;
766 u64 page_addr_list[];
769 /* GDMA_DESTROY_DMA_REGION */
770 struct gdma_destroy_dma_region_req {
771 struct gdma_req_hdr hdr;
773 u64 dma_region_handle;
777 GDMA_PD_FLAG_INVALID = 0,
780 struct gdma_create_pd_req {
781 struct gdma_req_hdr hdr;
782 enum gdma_pd_flags flags;
786 struct gdma_create_pd_resp {
787 struct gdma_resp_hdr hdr;
793 struct gdma_destroy_pd_req {
794 struct gdma_req_hdr hdr;
798 struct gdma_destory_pd_resp {
799 struct gdma_resp_hdr hdr;
803 /* Guest Virtual Address - MRs of this type allow access
804 * to memory mapped by PTEs associated with this MR using a virtual
805 * address that is set up in the MST
807 GDMA_MR_TYPE_GVA = 2,
810 struct gdma_create_mr_params {
812 enum gdma_mr_type mr_type;
815 u64 dma_region_handle;
817 enum gdma_mr_access_flags access_flags;
822 struct gdma_create_mr_request {
823 struct gdma_req_hdr hdr;
825 enum gdma_mr_type mr_type;
830 u64 dma_region_handle;
832 enum gdma_mr_access_flags access_flags;
839 struct gdma_create_mr_response {
840 struct gdma_resp_hdr hdr;
846 struct gdma_destroy_mr_request {
847 struct gdma_req_hdr hdr;
851 struct gdma_destroy_mr_response {
852 struct gdma_resp_hdr hdr;
855 int mana_gd_verify_vf_version(struct pci_dev *pdev);
857 int mana_gd_register_device(struct gdma_dev *gd);
858 int mana_gd_deregister_device(struct gdma_dev *gd);
860 int mana_gd_post_work_request(struct gdma_queue *wq,
861 const struct gdma_wqe_request *wqe_req,
862 struct gdma_posted_wqe_info *wqe_info);
864 int mana_gd_post_and_ring(struct gdma_queue *queue,
865 const struct gdma_wqe_request *wqe,
866 struct gdma_posted_wqe_info *wqe_info);
868 int mana_gd_alloc_res_map(u32 res_avail, struct gdma_resource *r);
869 void mana_gd_free_res_map(struct gdma_resource *r);
871 void mana_gd_wq_ring_doorbell(struct gdma_context *gc,
872 struct gdma_queue *queue);
874 int mana_gd_alloc_memory(struct gdma_context *gc, unsigned int length,
875 struct gdma_mem_info *gmi);
877 void mana_gd_free_memory(struct gdma_mem_info *gmi);
879 int mana_gd_send_request(struct gdma_context *gc, u32 req_len, const void *req,
880 u32 resp_len, void *resp);
882 int mana_gd_destroy_dma_region(struct gdma_context *gc, u64 dma_region_handle);
883 void mana_register_debugfs(void);
884 void mana_unregister_debugfs(void);