xref: /openbmc/qemu/hw/vfio/spapr.c (revision 5dd0be53)
1 /*
2  * DMA memory preregistration
3  *
4  * Authors:
5  *  Alexey Kardashevskiy <aik@ozlabs.ru>
6  *
7  * This work is licensed under the terms of the GNU GPL, version 2.  See
8  * the COPYING file in the top-level directory.
9  */
10 
11 #include "qemu/osdep.h"
12 #include <sys/ioctl.h>
13 #include <linux/vfio.h>
14 
15 #include "hw/vfio/vfio-common.h"
16 #include "hw/hw.h"
17 #include "exec/ram_addr.h"
18 #include "qemu/error-report.h"
19 #include "qapi/error.h"
20 #include "trace.h"
21 
22 static bool vfio_prereg_listener_skipped_section(MemoryRegionSection *section)
23 {
24     if (memory_region_is_iommu(section->mr)) {
25         hw_error("Cannot possibly preregister IOMMU memory");
26     }
27 
28     return !memory_region_is_ram(section->mr) ||
29             memory_region_is_ram_device(section->mr);
30 }
31 
32 static void *vfio_prereg_gpa_to_vaddr(MemoryRegionSection *section, hwaddr gpa)
33 {
34     return memory_region_get_ram_ptr(section->mr) +
35         section->offset_within_region +
36         (gpa - section->offset_within_address_space);
37 }
38 
39 static void vfio_prereg_listener_region_add(MemoryListener *listener,
40                                             MemoryRegionSection *section)
41 {
42     VFIOContainer *container = container_of(listener, VFIOContainer,
43                                             prereg_listener);
44     const hwaddr gpa = section->offset_within_address_space;
45     hwaddr end;
46     int ret;
47     hwaddr page_mask = qemu_real_host_page_mask();
48     struct vfio_iommu_spapr_register_memory reg = {
49         .argsz = sizeof(reg),
50         .flags = 0,
51     };
52 
53     if (vfio_prereg_listener_skipped_section(section)) {
54         trace_vfio_prereg_listener_region_add_skip(
55                 section->offset_within_address_space,
56                 section->offset_within_address_space +
57                 int128_get64(int128_sub(section->size, int128_one())));
58         return;
59     }
60 
61     if (unlikely((section->offset_within_address_space & ~page_mask) ||
62                  (section->offset_within_region & ~page_mask) ||
63                  (int128_get64(section->size) & ~page_mask))) {
64         error_report("%s received unaligned region", __func__);
65         return;
66     }
67 
68     end = section->offset_within_address_space + int128_get64(section->size);
69     if (gpa >= end) {
70         return;
71     }
72 
73     memory_region_ref(section->mr);
74 
75     reg.vaddr = (uintptr_t) vfio_prereg_gpa_to_vaddr(section, gpa);
76     reg.size = end - gpa;
77 
78     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_REGISTER_MEMORY, &reg);
79     trace_vfio_prereg_register(reg.vaddr, reg.size, ret ? -errno : 0);
80     if (ret) {
81         /*
82          * On the initfn path, store the first error in the container so we
83          * can gracefully fail.  Runtime, there's not much we can do other
84          * than throw a hardware error.
85          */
86         if (!container->initialized) {
87             if (!container->error) {
88                 error_setg_errno(&container->error, -ret,
89                                  "Memory registering failed");
90             }
91         } else {
92             hw_error("vfio: Memory registering failed, unable to continue");
93         }
94     }
95 }
96 
97 static void vfio_prereg_listener_region_del(MemoryListener *listener,
98                                             MemoryRegionSection *section)
99 {
100     VFIOContainer *container = container_of(listener, VFIOContainer,
101                                             prereg_listener);
102     const hwaddr gpa = section->offset_within_address_space;
103     hwaddr end;
104     int ret;
105     hwaddr page_mask = qemu_real_host_page_mask();
106     struct vfio_iommu_spapr_register_memory reg = {
107         .argsz = sizeof(reg),
108         .flags = 0,
109     };
110 
111     if (vfio_prereg_listener_skipped_section(section)) {
112         trace_vfio_prereg_listener_region_del_skip(
113                 section->offset_within_address_space,
114                 section->offset_within_address_space +
115                 int128_get64(int128_sub(section->size, int128_one())));
116         return;
117     }
118 
119     if (unlikely((section->offset_within_address_space & ~page_mask) ||
120                  (section->offset_within_region & ~page_mask) ||
121                  (int128_get64(section->size) & ~page_mask))) {
122         error_report("%s received unaligned region", __func__);
123         return;
124     }
125 
126     end = section->offset_within_address_space + int128_get64(section->size);
127     if (gpa >= end) {
128         return;
129     }
130 
131     reg.vaddr = (uintptr_t) vfio_prereg_gpa_to_vaddr(section, gpa);
132     reg.size = end - gpa;
133 
134     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, &reg);
135     trace_vfio_prereg_unregister(reg.vaddr, reg.size, ret ? -errno : 0);
136 }
137 
138 const MemoryListener vfio_prereg_listener = {
139     .name = "vfio-pre-reg",
140     .region_add = vfio_prereg_listener_region_add,
141     .region_del = vfio_prereg_listener_region_del,
142 };
143 
144 int vfio_spapr_create_window(VFIOContainer *container,
145                              MemoryRegionSection *section,
146                              hwaddr *pgsize)
147 {
148     int ret = 0;
149     IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
150     uint64_t pagesize = memory_region_iommu_get_min_page_size(iommu_mr), pgmask;
151     unsigned entries, bits_total, bits_per_level, max_levels;
152     struct vfio_iommu_spapr_tce_create create = { .argsz = sizeof(create) };
153     long rampagesize = qemu_minrampagesize();
154 
155     /*
156      * The host might not support the guest supported IOMMU page size,
157      * so we will use smaller physical IOMMU pages to back them.
158      */
159     if (pagesize > rampagesize) {
160         pagesize = rampagesize;
161     }
162     pgmask = container->pgsizes & (pagesize | (pagesize - 1));
163     pagesize = pgmask ? (1ULL << (63 - clz64(pgmask))) : 0;
164     if (!pagesize) {
165         error_report("Host doesn't support page size 0x%"PRIx64
166                      ", the supported mask is 0x%lx",
167                      memory_region_iommu_get_min_page_size(iommu_mr),
168                      container->pgsizes);
169         return -EINVAL;
170     }
171 
172     /*
173      * FIXME: For VFIO iommu types which have KVM acceleration to
174      * avoid bouncing all map/unmaps through qemu this way, this
175      * would be the right place to wire that up (tell the KVM
176      * device emulation the VFIO iommu handles to use).
177      */
178     create.window_size = int128_get64(section->size);
179     create.page_shift = ctz64(pagesize);
180     /*
181      * SPAPR host supports multilevel TCE tables. We try to guess optimal
182      * levels number and if this fails (for example due to the host memory
183      * fragmentation), we increase levels. The DMA address structure is:
184      * rrrrrrrr rxxxxxxx xxxxxxxx xxxxxxxx  xxxxxxxx xxxxxxxx xxxxxxxx iiiiiiii
185      * where:
186      *   r = reserved (bits >= 55 are reserved in the existing hardware)
187      *   i = IOMMU page offset (64K in this example)
188      *   x = bits to index a TCE which can be split to equal chunks to index
189      *      within the level.
190      * The aim is to split "x" to smaller possible number of levels.
191      */
192     entries = create.window_size >> create.page_shift;
193     /* bits_total is number of "x" needed */
194     bits_total = ctz64(entries * sizeof(uint64_t));
195     /*
196      * bits_per_level is a safe guess of how much we can allocate per level:
197      * 8 is the current minimum for CONFIG_FORCE_MAX_ZONEORDER and MAX_ORDER
198      * is usually bigger than that.
199      * Below we look at qemu_real_host_page_size as TCEs are allocated from
200      * system pages.
201      */
202     bits_per_level = ctz64(qemu_real_host_page_size()) + 8;
203     create.levels = bits_total / bits_per_level;
204     if (bits_total % bits_per_level) {
205         ++create.levels;
206     }
207     max_levels = (64 - create.page_shift) / ctz64(qemu_real_host_page_size());
208     for ( ; create.levels <= max_levels; ++create.levels) {
209         ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create);
210         if (!ret) {
211             break;
212         }
213     }
214     if (ret) {
215         error_report("Failed to create a window, ret = %d (%m)", ret);
216         return -errno;
217     }
218 
219     if (create.start_addr != section->offset_within_address_space) {
220         vfio_spapr_remove_window(container, create.start_addr);
221 
222         error_report("Host doesn't support DMA window at %"HWADDR_PRIx", must be %"PRIx64,
223                      section->offset_within_address_space,
224                      (uint64_t)create.start_addr);
225         return -EINVAL;
226     }
227     trace_vfio_spapr_create_window(create.page_shift,
228                                    create.levels,
229                                    create.window_size,
230                                    create.start_addr);
231     *pgsize = pagesize;
232 
233     return 0;
234 }
235 
236 int vfio_spapr_remove_window(VFIOContainer *container,
237                              hwaddr offset_within_address_space)
238 {
239     struct vfio_iommu_spapr_tce_remove remove = {
240         .argsz = sizeof(remove),
241         .start_addr = offset_within_address_space,
242     };
243     int ret;
244 
245     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
246     if (ret) {
247         error_report("Failed to remove window at %"PRIx64,
248                      (uint64_t)remove.start_addr);
249         return -errno;
250     }
251 
252     trace_vfio_spapr_remove_window(offset_within_address_space);
253 
254     return 0;
255 }
256