@@ -7,6 +7,7 @@
#include <unistd.h>
#include <sys/ioctl.h>
+#include <rte_errno.h>
#include <rte_log.h>
#include <rte_memory.h>
#include <rte_eal_memconfig.h>
@@ -18,6 +19,8 @@
#ifdef VFIO_PRESENT
+#define VFIO_MEM_EVENT_CLB_NAME "vfio_mem_event_clb"
+
/* per-process VFIO config */
static struct vfio_config vfio_cfg;
@@ -53,6 +56,42 @@ static const struct vfio_iommu_type iommu_types[] = {
},
};
+/* for sPAPR IOMMU, we will need to walk memseg list, but we cannot use
+ * rte_memseg_walk() because by the time we enter callback we will be holding a
+ * write lock, so regular rte-memseg_walk will deadlock. copying the same
+ * iteration code everywhere is not ideal as well. so, use a lockless copy of
+ * memseg walk here.
+ */
+static int
+memseg_walk_thread_unsafe(rte_memseg_walk_t func, void *arg)
+{
+ struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
+ int i, ms_idx, ret = 0;
+
+ for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
+ struct rte_memseg_list *msl = &mcfg->memsegs[i];
+ const struct rte_memseg *ms;
+ struct rte_fbarray *arr;
+
+ if (msl->memseg_arr.count == 0)
+ continue;
+
+ arr = &msl->memseg_arr;
+
+ ms_idx = rte_fbarray_find_next_used(arr, 0);
+ while (ms_idx >= 0) {
+ ms = rte_fbarray_get(arr, ms_idx);
+ ret = func(msl, ms, arg);
+ if (ret < 0)
+ return -1;
+ if (ret > 0)
+ return 1;
+ ms_idx = rte_fbarray_find_next_used(arr, ms_idx + 1);
+ }
+ }
+ return 0;
+}
+
int
vfio_get_group_fd(int iommu_group_no)
{
@@ -214,6 +253,38 @@ vfio_group_device_count(int vfio_group_fd)
return vfio_cfg.vfio_groups[i].devices;
}
+static void
+vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len)
+{
+ struct rte_memseg_list *msl;
+ struct rte_memseg *ms;
+ size_t cur_len = 0;
+
+ msl = rte_mem_virt2memseg_list(addr);
+
+ /* for IOVA as VA mode, no need to care for IOVA addresses */
+ if (rte_eal_iova_mode() == RTE_IOVA_VA) {
+ uint64_t vfio_va = (uint64_t)(uintptr_t)addr;
+ if (type == RTE_MEM_EVENT_ALLOC)
+ rte_vfio_dma_map(vfio_va, vfio_va, len);
+ else
+ rte_vfio_dma_unmap(vfio_va, vfio_va, len);
+ return;
+ }
+
+ /* memsegs are contiguous in memory */
+ ms = rte_mem_virt2memseg(addr, msl);
+ while (cur_len < len) {
+ if (type == RTE_MEM_EVENT_ALLOC)
+ rte_vfio_dma_map(ms->addr_64, ms->iova, ms->len);
+ else
+ rte_vfio_dma_unmap(ms->addr_64, ms->iova, ms->len);
+
+ cur_len += ms->len;
+ ++ms;
+ }
+}
+
int
rte_vfio_clear_group(int vfio_group_fd)
{
@@ -276,6 +347,8 @@ int
rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
int *vfio_dev_fd, struct vfio_device_info *device_info)
{
+ struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
+ rte_rwlock_t *mem_lock = &mcfg->memory_hotplug_lock;
struct vfio_group_status group_status = {
.argsz = sizeof(group_status)
};
@@ -363,6 +436,10 @@ rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
rte_vfio_clear_group(vfio_group_fd);
return -1;
}
+ /* lock memory hotplug before mapping and release it
+ * after registering callback, to prevent races
+ */
+ rte_rwlock_read_lock(mem_lock);
ret = t->dma_map_func(vfio_cfg.vfio_container_fd);
if (ret) {
RTE_LOG(ERR, EAL,
@@ -370,10 +447,26 @@ rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
dev_addr, errno, strerror(errno));
close(vfio_group_fd);
rte_vfio_clear_group(vfio_group_fd);
+ rte_rwlock_read_unlock(mem_lock);
return -1;
}
vfio_cfg.vfio_iommu_type = t;
+
+ /* register callback for mem events */
+ ret = rte_mem_event_callback_register(
+ VFIO_MEM_EVENT_CLB_NAME,
+ vfio_mem_event_callback);
+ /* unlock memory hotplug */
+ rte_rwlock_read_unlock(mem_lock);
+ if (ret && rte_errno == ENOTSUP) {
+ RTE_LOG(DEBUG, EAL, "Memory event callbacks not supported\n");
+ } else if (ret) {
+ RTE_LOG(ERR, EAL, "Could not install memory event callback for VFIO\n");
+ return -1;
+ } else {
+ RTE_LOG(DEBUG, EAL, "Installed memory event callback for VFIO\n");
+ }
}
}
@@ -411,6 +504,8 @@ int
rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
int vfio_dev_fd)
{
+ struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
+ rte_rwlock_t *mem_lock = &mcfg->memory_hotplug_lock;
struct vfio_group_status group_status = {
.argsz = sizeof(group_status)
};
@@ -418,13 +513,20 @@ rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
int iommu_group_no;
int ret;
+ /* we don't want any DMA mapping messages to come while we're detaching
+ * VFIO device, because this might be the last device and we might need
+ * to unregister the callback.
+ */
+ rte_rwlock_read_lock(mem_lock);
+
/* get group number */
ret = vfio_get_group_no(sysfs_base, dev_addr, &iommu_group_no);
if (ret <= 0) {
RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver\n",
dev_addr);
/* This is an error at this point. */
- return -1;
+ ret = -1;
+ goto out;
}
/* get the actual group fd */
@@ -432,7 +534,8 @@ rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
if (vfio_group_fd <= 0) {
RTE_LOG(INFO, EAL, "vfio_get_group_fd failed for %s\n",
dev_addr);
- return -1;
+ ret = -1;
+ goto out;
}
/* At this point we got an active group. Closing it will make the
@@ -444,7 +547,8 @@ rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
if (close(vfio_dev_fd) < 0) {
RTE_LOG(INFO, EAL, "Error when closing vfio_dev_fd for %s\n",
dev_addr);
- return -1;
+ ret = -1;
+ goto out;
}
/* An VFIO group can have several devices attached. Just when there is
@@ -456,17 +560,30 @@ rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
if (close(vfio_group_fd) < 0) {
RTE_LOG(INFO, EAL, "Error when closing vfio_group_fd for %s\n",
dev_addr);
- return -1;
+ ret = -1;
+ goto out;
}
if (rte_vfio_clear_group(vfio_group_fd) < 0) {
RTE_LOG(INFO, EAL, "Error when clearing group for %s\n",
dev_addr);
- return -1;
+ ret = -1;
+ goto out;
}
}
- return 0;
+ /* if there are no active device groups, unregister the callback to
+ * avoid spurious attempts to map/unmap memory from VFIO.
+ */
+ if (vfio_cfg.vfio_active_groups == 0)
+ rte_mem_event_callback_unregister(VFIO_MEM_EVENT_CLB_NAME);
+
+ /* success */
+ ret = 0;
+
+out:
+ rte_rwlock_read_unlock(mem_lock);
+ return ret;
}
int
@@ -884,7 +1001,7 @@ vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
/* check if window size needs to be adjusted */
memset(¶m, 0, sizeof(param));
- if (rte_memseg_walk(vfio_spapr_window_size_walk, ¶m) < 0) {
+ if (memseg_walk_thread_unsafe(vfio_spapr_window_size_walk, ¶m) < 0) {
RTE_LOG(ERR, EAL, "Could not get window size\n");
return -1;
}
@@ -903,7 +1020,7 @@ vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
return -1;
}
- if (rte_memseg_walk(vfio_spapr_map_walk,
+ if (memseg_walk_thread_unsafe(vfio_spapr_map_walk,
&vfio_container_fd) < 0) {
RTE_LOG(ERR, EAL, "Could not recreate DMA maps\n");
return -1;