Comments inline.
Internal Use - Confidential > -----Original Message----- > From: fengchengwen <[email protected]> > Sent: Tuesday, November 11, 2025 5:21 PM > To: Bathija, Pravin <[email protected]>; [email protected] > Cc: [email protected] > Subject: Re: [PATCH v3 4/5] vhost_user: support function defines for back-end > > > [EXTERNAL EMAIL] > > On 11/11/2025 7:31 PM, Bathija, Pravin wrote: > > Responses inline. > > > > > > Internal Use - Confidential > >> -----Original Message----- > >> From: fengchengwen <[email protected]> > >> Sent: Tuesday, November 4, 2025 12:06 AM > >> To: Bathija, Pravin <[email protected]>; [email protected] > >> Cc: [email protected] > >> Subject: Re: [PATCH v3 4/5] vhost_user: support function defines for > >> back-end > >> > >> > >> [EXTERNAL EMAIL] > >> > >> On 11/4/2025 12:21 PM, Pravin M Bathija wrote: > >>> Here we define support functions which are called from the various > >>> vhost-user back-end message functions like set memory table, get > >>> memory slots, add memory region, remove memory region. These are > >>> essetially common functions to initialize memory, unmap a set of > >>> memory regions, perform register copy and align memory addresses. > >>> > >>> Signed-off-by: Pravin M Bathija <[email protected]> > >>> --- > >>> lib/vhost/vhost_user.c | 80 > >>> +++++++++++++++++++++++++++++++++++------- > >>> 1 file changed, 68 insertions(+), 12 deletions(-) > >>> > >>> diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c index > >>> 168432e7d1..9a85f2fc92 100644 > >>> --- a/lib/vhost/vhost_user.c > >>> +++ b/lib/vhost/vhost_user.c > >>> @@ -228,7 +228,17 @@ async_dma_map(struct virtio_net *dev, bool > >>> do_map) } > >>> > >>> static void > >>> -free_mem_region(struct virtio_net *dev) > >>> +free_mem_region(struct rte_vhost_mem_region *reg) { > >>> + if (reg != NULL && reg->host_user_addr) { > >>> + munmap(reg->mmap_addr, reg->mmap_size); > >>> + close(reg->fd); > >>> + memset(reg, 0, sizeof(struct rte_vhost_mem_region)); > >>> + } > >>> +} > >>> + > >>> +static void > >>> +free_all_mem_regions(struct virtio_net *dev) > >>> { > >>> uint32_t i; > >>> struct rte_vhost_mem_region *reg; @@ -239,12 +249,10 @@ > >>> free_mem_region(struct virtio_net *dev) > >>> if (dev->async_copy && rte_vfio_is_enabled("vfio")) > >>> async_dma_map(dev, false); > >>> > >>> - for (i = 0; i < dev->mem->nregions; i++) { > >>> + for (i = 0; i < VHOST_MEMORY_MAX_NREGIONS; i++) { > >>> reg = &dev->mem->regions[i]; > >>> - if (reg->host_user_addr) { > >>> - munmap(reg->mmap_addr, reg->mmap_size); > >>> - close(reg->fd); > >>> - } > >>> + if (reg->mmap_addr) > >>> + free_mem_region(reg); > >>> } > >>> } > >>> > >>> @@ -258,7 +266,7 @@ vhost_backend_cleanup(struct virtio_net *dev) > >>> vdpa_dev->ops->dev_cleanup(dev->vid); > >>> > >>> if (dev->mem) { > >>> - free_mem_region(dev); > >>> + free_all_mem_regions(dev); > >>> rte_free(dev->mem); > >>> dev->mem = NULL; > >>> } > >>> @@ -707,7 +715,7 @@ numa_realloc(struct virtio_net **pdev, struct > >> vhost_virtqueue **pvq) > >>> vhost_devices[dev->vid] = dev; > >>> > >>> mem_size = sizeof(struct rte_vhost_memory) + > >>> - sizeof(struct rte_vhost_mem_region) * dev->mem->nregions; > >>> + sizeof(struct rte_vhost_mem_region) * > >> VHOST_MEMORY_MAX_NREGIONS; > >>> mem = rte_realloc_socket(dev->mem, mem_size, 0, node); > >>> if (!mem) { > >>> VHOST_CONFIG_LOG(dev->ifname, ERR, @@ -811,8 +819,10 @@ > >>> hua_to_alignment(struct rte_vhost_memory > >> *mem, void *ptr) > >>> uint32_t i; > >>> uintptr_t hua = (uintptr_t)ptr; > >>> > >>> - for (i = 0; i < mem->nregions; i++) { > >>> + for (i = 0; i < VHOST_MEMORY_MAX_NREGIONS; i++) { > >>> r = &mem->regions[i]; > >>> + if (r->host_user_addr == 0) > >>> + continue; > >>> if (hua >= r->host_user_addr && > >>> hua < r->host_user_addr + r->size) { > >>> return get_blk_size(r->fd); @@ -1250,9 +1260,13 > >>> @@ vhost_user_postcopy_register(struct virtio_net > >> *dev, int main_fd, > >>> * retrieve the region offset when handling userfaults. > >>> */ > >>> memory = &ctx->msg.payload.memory; > >>> - for (i = 0; i < memory->nregions; i++) { > >>> + for (i = 0; i < VHOST_MEMORY_MAX_NREGIONS; i++) { > >> > >> I think the using MAX_NREGIONS are most for convienent, but it will > >> impact the performance, because the rte_vhost_va_from_guest_pa() > >> should iter the entire array. > >> > > > > Replaced VHOST_MEMORY_MAX_NREGIONS with memory->nregions. > Please review v4. > > > >> I think we should keep the original impl: make sure the nregions > >> entry of memory-region is always valid. > >> > >> Beside, where is the modification for rte_vhost_va_from_guest_pa()??? > > > > Could you please provide more detail ? rte_vhost_va_from_guest_pa() was > never called from here before. > > Because rte_vhost_va_from_guest_pa() use mem-nregions as for uplimit: > for (i = 0; i < mem->nregions; i++) { > > this function was never called, but I think we need to make each commit > complete. >From what I gather, both of your latest comments, for patches 4 and 5 are >related to performance impact because of 128 memory regions. My thoughts on >this are, I will work on an algorithm to optimize as you suggested. Meanwhile we could approve patches 1-4. The rest of the implementation does not need 128 memory regions. The existing 8 regions are fine. I put in this change as I felt it makes the system scale more and was also suggested by Maxime. I think patches 1-4 stand by themselves and have been thoroughly tested. They could be merged in unless you have some other recommendations or changes you would like me to consider. Thoughts ? > > > > >> > >>> + int reg_msg_index = 0; > >>> reg = &dev->mem->regions[i]; > >>> - memory->regions[i].userspace_addr = reg->host_user_addr; > >>> + if (reg->host_user_addr == 0) > >>> + continue; > >>> + memory->regions[reg_msg_index].userspace_addr = reg- > >>> host_user_addr; > >>> + reg_msg_index++; > >>> } > >>> > >>> /* Send the addresses back to qemu */ @@ -1279,8 +1293,10 @@ > >>> vhost_user_postcopy_register(struct virtio_net *dev, int main_fd, > >>> } > >>> > >>> /* Now userfault register and we can use the memory */ > >>> - for (i = 0; i < memory->nregions; i++) { > >>> + for (i = 0; i < VHOST_MEMORY_MAX_NREGIONS; i++) { > >>> reg = &dev->mem->regions[i]; > >>> + if (reg->host_user_addr == 0) > >>> + continue; > >>> if (vhost_user_postcopy_region_register(dev, reg) < 0) > >>> return -1; > >>> } > >>> @@ -1385,6 +1401,46 @@ vhost_user_mmap_region(struct virtio_net > *dev, > >>> return 0; > >>> } > >>> > >>> +static int > >>> +vhost_user_initialize_memory(struct virtio_net **pdev) > >> > >> This function should be part of 3/5, else the 3/5 will compile fail > > > > I have moved support functions to patch-3. Please review v4. > > > >> > >>> +{ > >>> + struct virtio_net *dev = *pdev; > >>> + int numa_node = SOCKET_ID_ANY; > >>> + > >>> + /* > >>> + * If VQ 0 has already been allocated, try to allocate on the same > >>> + * NUMA node. It can be reallocated later in numa_realloc(). > >>> + */ > >>> + if (dev->nr_vring > 0) > >>> + numa_node = dev->virtqueue[0]->numa_node; > >>> + > >>> + dev->nr_guest_pages = 0; > >>> + if (dev->guest_pages == NULL) { > >>> + dev->max_guest_pages = 8; > >> > >> It should be VHOST_MEMORY_MAX_NREGIONS > > > > Done. Please review v4. > > > >> > >>> + dev->guest_pages = rte_zmalloc_socket(NULL, > >>> + dev->max_guest_pages * > >>> + sizeof(struct guest_page), > >>> + RTE_CACHE_LINE_SIZE, > >>> + numa_node); > >>> + if (dev->guest_pages == NULL) { > >>> + VHOST_CONFIG_LOG(dev->ifname, ERR, > >>> + "failed to allocate memory for dev- > >>> guest_pages"); > >>> + return -1; > >>> + } > >>> + } > >>> + > >>> + dev->mem = rte_zmalloc_socket("vhost-mem-table", sizeof(struct > >> rte_vhost_memory) + > >>> + sizeof(struct rte_vhost_mem_region) * > >> VHOST_MEMORY_MAX_NREGIONS, 0, numa_node); > >>> + if (dev->mem == NULL) { > >>> + VHOST_CONFIG_LOG(dev->ifname, ERR, "failed to allocate > >> memory for dev->mem"); > >>> + rte_free(dev->guest_pages); > >>> + dev->guest_pages = NULL; > >>> + return -1; > >>> + } > >>> + > >>> + return 0; > >>> +} > >>> + > >>> static int > >>> vhost_user_set_mem_table(struct virtio_net **pdev, > >>> struct vhu_msg_context *ctx, > > > >
