xref: /openbmc/qemu/hw/pci/pcie_sriov.c (revision d84ed5d2) 
1 /*
2  * pcie_sriov.c:
3  *
4  * Implementation of SR/IOV emulation support.
5  *
6  * Copyright (c) 2015-2017 Knut Omang <knut.omang@oracle.com>
7  *
8  * This work is licensed under the terms of the GNU GPL, version 2 or later.
9  * See the COPYING file in the top-level directory.
10  *
11  */
12 
13 #include "qemu/osdep.h"
14 #include "hw/pci/pci_device.h"
15 #include "hw/pci/pcie.h"
16 #include "hw/pci/pci_bus.h"
17 #include "hw/qdev-properties.h"
18 #include "qemu/error-report.h"
19 #include "qemu/range.h"
20 #include "qapi/error.h"
21 #include "trace.h"
22 
23 static void unparent_vfs(PCIDevice *dev, uint16_t total_vfs)
24 {
25     for (uint16_t i = 0; i < total_vfs; i++) {
26         PCIDevice *vf = dev->exp.sriov_pf.vf[i];
27         object_unparent(OBJECT(vf));
28         object_unref(OBJECT(vf));
29     }
30     g_free(dev->exp.sriov_pf.vf);
31     dev->exp.sriov_pf.vf = NULL;
32 }
33 
34 bool pcie_sriov_pf_init(PCIDevice *dev, uint16_t offset,
35                         const char *vfname, uint16_t vf_dev_id,
36                         uint16_t init_vfs, uint16_t total_vfs,
37                         uint16_t vf_offset, uint16_t vf_stride,
38                         Error **errp)
39 {
40     BusState *bus = qdev_get_parent_bus(&dev->qdev);
41     int32_t devfn = dev->devfn + vf_offset;
42     uint8_t *cfg = dev->config + offset;
43     uint8_t *wmask;
44 
45     if (total_vfs) {
46         uint16_t ari_cap = pcie_find_capability(dev, PCI_EXT_CAP_ID_ARI);
47         uint16_t first_vf_devfn = dev->devfn + vf_offset;
48         uint16_t last_vf_devfn = first_vf_devfn + vf_stride * (total_vfs - 1);
49 
50         if ((!ari_cap && PCI_SLOT(dev->devfn) != PCI_SLOT(last_vf_devfn)) ||
51             last_vf_devfn >= PCI_DEVFN_MAX) {
52             error_setg(errp, "VF function number overflows");
53             return false;
54         }
55     }
56 
57     pcie_add_capability(dev, PCI_EXT_CAP_ID_SRIOV, 1,
58                         offset, PCI_EXT_CAP_SRIOV_SIZEOF);
59     dev->exp.sriov_cap = offset;
60     dev->exp.sriov_pf.vf = NULL;
61 
62     pci_set_word(cfg + PCI_SRIOV_VF_OFFSET, vf_offset);
63     pci_set_word(cfg + PCI_SRIOV_VF_STRIDE, vf_stride);
64 
65     /*
66      * Mandatory page sizes to support.
67      * Device implementations can call pcie_sriov_pf_add_sup_pgsize()
68      * to set more bits:
69      */
70     pci_set_word(cfg + PCI_SRIOV_SUP_PGSIZE, SRIOV_SUP_PGSIZE_MINREQ);
71 
72     /*
73      * Default is to use 4K pages, software can modify it
74      * to any of the supported bits
75      */
76     pci_set_word(cfg + PCI_SRIOV_SYS_PGSIZE, 0x1);
77 
78     /* Set up device ID and initial/total number of VFs available */
79     pci_set_word(cfg + PCI_SRIOV_VF_DID, vf_dev_id);
80     pci_set_word(cfg + PCI_SRIOV_INITIAL_VF, init_vfs);
81     pci_set_word(cfg + PCI_SRIOV_TOTAL_VF, total_vfs);
82     pci_set_word(cfg + PCI_SRIOV_NUM_VF, 0);
83 
84     /* Write enable control bits */
85     wmask = dev->wmask + offset;
86     pci_set_word(wmask + PCI_SRIOV_CTRL,
87                  PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE | PCI_SRIOV_CTRL_ARI);
88     pci_set_word(wmask + PCI_SRIOV_NUM_VF, 0xffff);
89     pci_set_word(wmask + PCI_SRIOV_SYS_PGSIZE, 0x553);
90 
91     qdev_prop_set_bit(&dev->qdev, "multifunction", true);
92 
93     dev->exp.sriov_pf.vf = g_new(PCIDevice *, total_vfs);
94 
95     for (uint16_t i = 0; i < total_vfs; i++) {
96         PCIDevice *vf = pci_new(devfn, vfname);
97         vf->exp.sriov_vf.pf = dev;
98         vf->exp.sriov_vf.vf_number = i;
99 
100         if (!qdev_realize(&vf->qdev, bus, errp)) {
101             object_unparent(OBJECT(vf));
102             object_unref(vf);
103             unparent_vfs(dev, i);
104             return false;
105         }
106 
107         /* set vid/did according to sr/iov spec - they are not used */
108         pci_config_set_vendor_id(vf->config, 0xffff);
109         pci_config_set_device_id(vf->config, 0xffff);
110 
111         dev->exp.sriov_pf.vf[i] = vf;
112         devfn += vf_stride;
113     }
114 
115     return true;
116 }
117 
118 void pcie_sriov_pf_exit(PCIDevice *dev)
119 {
120     uint8_t *cfg = dev->config + dev->exp.sriov_cap;
121 
122     unparent_vfs(dev, pci_get_word(cfg + PCI_SRIOV_TOTAL_VF));
123 }
124 
125 void pcie_sriov_pf_init_vf_bar(PCIDevice *dev, int region_num,
126                                uint8_t type, dma_addr_t size)
127 {
128     uint32_t addr;
129     uint64_t wmask;
130     uint16_t sriov_cap = dev->exp.sriov_cap;
131 
132     assert(sriov_cap > 0);
133     assert(region_num >= 0);
134     assert(region_num < PCI_NUM_REGIONS);
135     assert(region_num != PCI_ROM_SLOT);
136 
137     wmask = ~(size - 1);
138     addr = sriov_cap + PCI_SRIOV_BAR + region_num * 4;
139 
140     pci_set_long(dev->config + addr, type);
141     if (!(type & PCI_BASE_ADDRESS_SPACE_IO) &&
142         type & PCI_BASE_ADDRESS_MEM_TYPE_64) {
143         pci_set_quad(dev->wmask + addr, wmask);
144         pci_set_quad(dev->cmask + addr, ~0ULL);
145     } else {
146         pci_set_long(dev->wmask + addr, wmask & 0xffffffff);
147         pci_set_long(dev->cmask + addr, 0xffffffff);
148     }
149     dev->exp.sriov_pf.vf_bar_type[region_num] = type;
150 }
151 
152 void pcie_sriov_vf_register_bar(PCIDevice *dev, int region_num,
153                                 MemoryRegion *memory)
154 {
155     PCIIORegion *r;
156     PCIBus *bus = pci_get_bus(dev);
157     uint8_t type;
158     pcibus_t size = memory_region_size(memory);
159 
160     assert(pci_is_vf(dev)); /* PFs must use pci_register_bar */
161     assert(region_num >= 0);
162     assert(region_num < PCI_NUM_REGIONS);
163     type = dev->exp.sriov_vf.pf->exp.sriov_pf.vf_bar_type[region_num];
164 
165     if (!is_power_of_2(size)) {
166         error_report("%s: PCI region size must be a power"
167                      " of two - type=0x%x, size=0x%"FMT_PCIBUS,
168                      __func__, type, size);
169         exit(1);
170     }
171 
172     r = &dev->io_regions[region_num];
173     r->memory = memory;
174     r->address_space =
175         type & PCI_BASE_ADDRESS_SPACE_IO
176         ? bus->address_space_io
177         : bus->address_space_mem;
178     r->size = size;
179     r->type = type;
180 
181     r->addr = pci_bar_address(dev, region_num, r->type, r->size);
182     if (r->addr != PCI_BAR_UNMAPPED) {
183         memory_region_add_subregion_overlap(r->address_space,
184                                             r->addr, r->memory, 1);
185     }
186 }
187 
188 static void clear_ctrl_vfe(PCIDevice *dev)
189 {
190     uint8_t *ctrl = dev->config + dev->exp.sriov_cap + PCI_SRIOV_CTRL;
191     pci_set_word(ctrl, pci_get_word(ctrl) & ~PCI_SRIOV_CTRL_VFE);
192 }
193 
194 static void register_vfs(PCIDevice *dev)
195 {
196     uint16_t num_vfs;
197     uint16_t i;
198     uint16_t sriov_cap = dev->exp.sriov_cap;
199 
200     assert(sriov_cap > 0);
201     num_vfs = pci_get_word(dev->config + sriov_cap + PCI_SRIOV_NUM_VF);
202     if (num_vfs > pci_get_word(dev->config + sriov_cap + PCI_SRIOV_TOTAL_VF)) {
203         clear_ctrl_vfe(dev);
204         return;
205     }
206 
207     trace_sriov_register_vfs(dev->name, PCI_SLOT(dev->devfn),
208                              PCI_FUNC(dev->devfn), num_vfs);
209     for (i = 0; i < num_vfs; i++) {
210         pci_set_enabled(dev->exp.sriov_pf.vf[i], true);
211     }
212 }
213 
214 static void unregister_vfs(PCIDevice *dev)
215 {
216     uint16_t i;
217     uint8_t *cfg = dev->config + dev->exp.sriov_cap;
218 
219     trace_sriov_unregister_vfs(dev->name, PCI_SLOT(dev->devfn),
220                                PCI_FUNC(dev->devfn));
221     for (i = 0; i < pci_get_word(cfg + PCI_SRIOV_TOTAL_VF); i++) {
222         pci_set_enabled(dev->exp.sriov_pf.vf[i], false);
223     }
224 }
225 
226 void pcie_sriov_config_write(PCIDevice *dev, uint32_t address,
227                              uint32_t val, int len)
228 {
229     uint32_t off;
230     uint16_t sriov_cap = dev->exp.sriov_cap;
231 
232     if (!sriov_cap || address < sriov_cap) {
233         return;
234     }
235     off = address - sriov_cap;
236     if (off >= PCI_EXT_CAP_SRIOV_SIZEOF) {
237         return;
238     }
239 
240     trace_sriov_config_write(dev->name, PCI_SLOT(dev->devfn),
241                              PCI_FUNC(dev->devfn), off, val, len);
242 
243     if (range_covers_byte(off, len, PCI_SRIOV_CTRL)) {
244         if (val & PCI_SRIOV_CTRL_VFE) {
245             register_vfs(dev);
246         } else {
247             unregister_vfs(dev);
248         }
249     } else if (range_covers_byte(off, len, PCI_SRIOV_NUM_VF)) {
250         clear_ctrl_vfe(dev);
251         unregister_vfs(dev);
252     }
253 }
254 
255 void pcie_sriov_pf_post_load(PCIDevice *dev)
256 {
257     if (dev->exp.sriov_cap) {
258         register_vfs(dev);
259     }
260 }
261 
262 
263 /* Reset SR/IOV */
264 void pcie_sriov_pf_reset(PCIDevice *dev)
265 {
266     uint16_t sriov_cap = dev->exp.sriov_cap;
267     if (!sriov_cap) {
268         return;
269     }
270 
271     pci_set_word(dev->config + sriov_cap + PCI_SRIOV_CTRL, 0);
272     unregister_vfs(dev);
273 
274     pci_set_word(dev->config + sriov_cap + PCI_SRIOV_NUM_VF, 0);
275 
276     /*
277      * Default is to use 4K pages, software can modify it
278      * to any of the supported bits
279      */
280     pci_set_word(dev->config + sriov_cap + PCI_SRIOV_SYS_PGSIZE, 0x1);
281 
282     for (uint16_t i = 0; i < PCI_NUM_REGIONS; i++) {
283         pci_set_quad(dev->config + sriov_cap + PCI_SRIOV_BAR + i * 4,
284                      dev->exp.sriov_pf.vf_bar_type[i]);
285     }
286 }
287 
288 /* Add optional supported page sizes to the mask of supported page sizes */
289 void pcie_sriov_pf_add_sup_pgsize(PCIDevice *dev, uint16_t opt_sup_pgsize)
290 {
291     uint8_t *cfg = dev->config + dev->exp.sriov_cap;
292     uint8_t *wmask = dev->wmask + dev->exp.sriov_cap;
293 
294     uint16_t sup_pgsize = pci_get_word(cfg + PCI_SRIOV_SUP_PGSIZE);
295 
296     sup_pgsize |= opt_sup_pgsize;
297 
298     /*
299      * Make sure the new bits are set, and that system page size
300      * also can be set to any of the new values according to spec:
301      */
302     pci_set_word(cfg + PCI_SRIOV_SUP_PGSIZE, sup_pgsize);
303     pci_set_word(wmask + PCI_SRIOV_SYS_PGSIZE, sup_pgsize);
304 }
305 
306 
307 uint16_t pcie_sriov_vf_number(PCIDevice *dev)
308 {
309     assert(pci_is_vf(dev));
310     return dev->exp.sriov_vf.vf_number;
311 }
312 
313 PCIDevice *pcie_sriov_get_pf(PCIDevice *dev)
314 {
315     return dev->exp.sriov_vf.pf;
316 }
317 
318 PCIDevice *pcie_sriov_get_vf_at_index(PCIDevice *dev, int n)
319 {
320     assert(!pci_is_vf(dev));
321     if (n < pcie_sriov_num_vfs(dev)) {
322         return dev->exp.sriov_pf.vf[n];
323     }
324     return NULL;
325 }
326 
327 uint16_t pcie_sriov_num_vfs(PCIDevice *dev)
328 {
329     uint16_t sriov_cap = dev->exp.sriov_cap;
330     uint8_t *cfg = dev->config + sriov_cap;
331 
332     return sriov_cap &&
333            (pci_get_word(cfg + PCI_SRIOV_CTRL) & PCI_SRIOV_CTRL_VFE) ?
334            pci_get_word(cfg + PCI_SRIOV_NUM_VF) : 0;
335 }
336