1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef _X86_POSTED_INTR_H
3#define _X86_POSTED_INTR_H
4
5#include <asm/cmpxchg.h>
6#include <asm/rwonce.h>
7#include <asm/irq_vectors.h>
8
9#include <linux/bitmap.h>
10
11#define POSTED_INTR_ON 0
12#define POSTED_INTR_SN 1
13
14#define PID_TABLE_ENTRY_VALID 1
15
16#define NR_PIR_VECTORS 256
17#define NR_PIR_WORDS (NR_PIR_VECTORS / BITS_PER_LONG)
18
19/* Posted-Interrupt Descriptor */
20struct pi_desc {
21 unsigned long pir[NR_PIR_WORDS]; /* Posted interrupt requested */
22 union {
23 struct {
24 u16 notifications; /* Suppress and outstanding bits */
25 u8 nv;
26 u8 rsvd_2;
27 u32 ndst;
28 };
29 u64 control;
30 };
31 u32 rsvd[6];
32} __aligned(64);
33
34/*
35 * De-multiplexing posted interrupts is on the performance path, the code
36 * below is written to optimize the cache performance based on the following
37 * considerations:
38 * 1.Posted interrupt descriptor (PID) fits in a cache line that is frequently
39 * accessed by both CPU and IOMMU.
40 * 2.During software processing of posted interrupts, the CPU needs to do
41 * natural width read and xchg for checking and clearing posted interrupt
42 * request (PIR), a 256 bit field within the PID.
43 * 3.On the other side, the IOMMU does atomic swaps of the entire PID cache
44 * line when posting interrupts and setting control bits.
45 * 4.The CPU can access the cache line a magnitude faster than the IOMMU.
46 * 5.Each time the IOMMU does interrupt posting to the PIR will evict the PID
47 * cache line. The cache line states after each operation are as follows,
48 * assuming a 64-bit kernel:
49 * CPU IOMMU PID Cache line state
50 * ---------------------------------------------------------------
51 *...read64 exclusive
52 *...lock xchg64 modified
53 *... post/atomic swap invalid
54 *...-------------------------------------------------------------
55 *
56 * To reduce L1 data cache miss, it is important to avoid contention with
57 * IOMMU's interrupt posting/atomic swap. Therefore, a copy of PIR is used
58 * when processing posted interrupts in software, e.g. to dispatch interrupt
59 * handlers for posted MSIs, or to move interrupts from the PIR to the vIRR
60 * in KVM.
61 *
62 * In addition, the code is trying to keep the cache line state consistent
63 * as much as possible. e.g. when making a copy and clearing the PIR
64 * (assuming non-zero PIR bits are present in the entire PIR), it does:
65 * read, read, read, read, xchg, xchg, xchg, xchg
66 * instead of:
67 * read, xchg, read, xchg, read, xchg, read, xchg
68 */
69static __always_inline bool pi_harvest_pir(unsigned long *pir,
70 unsigned long *pir_vals)
71{
72 unsigned long pending = 0;
73 int i;
74
75 for (i = 0; i < NR_PIR_WORDS; i++) {
76 pir_vals[i] = READ_ONCE(pir[i]);
77 pending |= pir_vals[i];
78 }
79
80 if (!pending)
81 return false;
82
83 for (i = 0; i < NR_PIR_WORDS; i++) {
84 if (!pir_vals[i])
85 continue;
86
87 pir_vals[i] = arch_xchg(&pir[i], 0);
88 }
89
90 return true;
91}
92
93static inline bool pi_test_and_set_on(struct pi_desc *pi_desc)
94{
95 return test_and_set_bit(POSTED_INTR_ON, addr: (unsigned long *)&pi_desc->control);
96}
97
98static inline bool pi_test_and_clear_on(struct pi_desc *pi_desc)
99{
100 return test_and_clear_bit(POSTED_INTR_ON, addr: (unsigned long *)&pi_desc->control);
101}
102
103static inline bool pi_test_and_clear_sn(struct pi_desc *pi_desc)
104{
105 return test_and_clear_bit(POSTED_INTR_SN, addr: (unsigned long *)&pi_desc->control);
106}
107
108static inline bool pi_test_and_set_pir(int vector, struct pi_desc *pi_desc)
109{
110 return test_and_set_bit(nr: vector, addr: pi_desc->pir);
111}
112
113static inline bool pi_is_pir_empty(struct pi_desc *pi_desc)
114{
115 return bitmap_empty(src: pi_desc->pir, NR_VECTORS);
116}
117
118static inline void pi_set_sn(struct pi_desc *pi_desc)
119{
120 set_bit(POSTED_INTR_SN, addr: (unsigned long *)&pi_desc->control);
121}
122
123static inline void pi_set_on(struct pi_desc *pi_desc)
124{
125 set_bit(POSTED_INTR_ON, addr: (unsigned long *)&pi_desc->control);
126}
127
128static inline void pi_clear_on(struct pi_desc *pi_desc)
129{
130 clear_bit(POSTED_INTR_ON, addr: (unsigned long *)&pi_desc->control);
131}
132
133static inline void pi_clear_sn(struct pi_desc *pi_desc)
134{
135 clear_bit(POSTED_INTR_SN, addr: (unsigned long *)&pi_desc->control);
136}
137
138static inline bool pi_test_on(struct pi_desc *pi_desc)
139{
140 return test_bit(POSTED_INTR_ON, (unsigned long *)&pi_desc->control);
141}
142
143static inline bool pi_test_sn(struct pi_desc *pi_desc)
144{
145 return test_bit(POSTED_INTR_SN, (unsigned long *)&pi_desc->control);
146}
147
148static inline bool pi_test_pir(int vector, struct pi_desc *pi_desc)
149{
150 return test_bit(vector, (unsigned long *)pi_desc->pir);
151}
152
153/* Non-atomic helpers */
154static inline void __pi_set_sn(struct pi_desc *pi_desc)
155{
156 pi_desc->notifications |= BIT(POSTED_INTR_SN);
157}
158
159static inline void __pi_clear_sn(struct pi_desc *pi_desc)
160{
161 pi_desc->notifications &= ~BIT(POSTED_INTR_SN);
162}
163
164#ifdef CONFIG_X86_POSTED_MSI
165/*
166 * Not all external vectors are subject to interrupt remapping, e.g. IOMMU's
167 * own interrupts. Here we do not distinguish them since those vector bits in
168 * PIR will always be zero.
169 */
170static inline bool pi_pending_this_cpu(unsigned int vector)
171{
172 struct pi_desc *pid = this_cpu_ptr(&posted_msi_pi_desc);
173
174 if (WARN_ON_ONCE(vector > NR_VECTORS || vector < FIRST_EXTERNAL_VECTOR))
175 return false;
176
177 return test_bit(vector, pid->pir);
178}
179
180extern void intel_posted_msi_init(void);
181#else
182static inline bool pi_pending_this_cpu(unsigned int vector) { return false; }
183
184static inline void intel_posted_msi_init(void) {};
185#endif /* X86_POSTED_MSI */
186
187#endif /* _X86_POSTED_INTR_H */
188