I think the code is on the right track. Instead of using vmbus_uio_get_num_subchan() and calling vmbus_uio_get_subchan() on each channel, you can just create a new function vmbus_uio_get_secondary_subchan(). This function goes through all subchannels and map ring buffers to the same addresses used by the primary process.
In the original code, vmbus_uio_map_secondary_subchan() has a check for "if (mapaddr == ring_buf)". If the address is mapped to somewhere else, the address won't work for the secondary process. So we need to keep this check. From: Stephen Hemminger <sthem...@microsoft.com> Sent: Wednesday, August 11, 2021 8:26 AM To: Jonathan Erb <jonathan....@banduracyber.com>; Long Li <lon...@microsoft.com> Cc: dev@dpdk.org; sta...@dpdk.org Subject: RE: [PATCH v2] bus/vmbus: Fix crash when handling packets in secondary process Looks fine, only comments would be to keep to original style and add check if primary channel not found? From: Jonathan Erb <jonathan....@banduracyber.com<mailto:jonathan....@banduracyber.com>> Sent: Monday, August 9, 2021 9:07 AM To: Long Li <lon...@microsoft.com<mailto:lon...@microsoft.com>>; Stephen Hemminger <sthem...@microsoft.com<mailto:sthem...@microsoft.com>> Cc: dev@dpdk.org<mailto:dev@dpdk.org>; sta...@dpdk.org<mailto:sta...@dpdk.org> Subject: Re: [PATCH v2] bus/vmbus: Fix crash when handling packets in secondary process You don't often get email from jonathan....@banduracyber.com<mailto:jonathan....@banduracyber.com>. Learn why this is important<http://aka.ms/LearnAboutSenderIdentification> Would it be possible to resolve the lack of subchannels in secondary processes as follows: 1. Create the following function in vmbus_uio.c to obtain the count of subchannels created by the primary: int vmbus_uio_get_num_subchan(struct vmbus_channel *primary) { const struct rte_vmbus_device *dev = primary->device; char chan_path[PATH_MAX]; struct dirent *ent; DIR *chan_dir; int err; int subchan_cnt = 0; snprintf(chan_path, sizeof(chan_path), "%s/%s/channels", SYSFS_VMBUS_DEVICES, dev->device.name); chan_dir = opendir(chan_path); if (!chan_dir) { VMBUS_LOG(ERR, "cannot open %s: %s", chan_path, strerror(errno)); return 0; } while ((ent = readdir(chan_dir))) { unsigned long relid, subid, monid; char *endp; if (ent->d_name[0] == '.') continue; errno = 0; relid = strtoul(ent->d_name, &endp, 0); if (*endp || errno != 0 || relid > UINT16_MAX) { VMBUS_LOG(NOTICE, "not a valid channel relid: %s", ent->d_name); continue; } subchan_cnt++; } closedir(chan_dir); //Less one for primary channel return subchan_cnt - 1; } 2. Then change the bottom of vmbus_uio_map_secondary() to be simply: /* fd is not needed in secondary process, close it */ close(fd); if (vmbus_chan_create(dev, dev->relid, 0, dev->monitor_id, &dev->primary)) { VMBUS_LOG(ERR, "cannot create primary channel"); return -1; } int subchannels_cnt = vmbus_uio_get_num_subchan(dev->primary); for (i = 0; i < subchannels_cnt; i++) { if(vmbus_uio_get_subchan(dev->primary, &chan)) return -1; STAILQ_INSERT_TAIL(&dev->primary->subchannel_list, chan, next); } return 0; } In this way the secondary processes detect the number of subchannels created by the primary, then perform their own mappings as needed. All this can occur before rte_eal_init has completed. Jonathan On 7/26/21 6:16 PM, Long Li wrote: Subject: [PATCH v2] bus/vmbus: Fix crash when handling packets in secondary process Have secondary processes construct their own copy of primary channel with own mappings. Remove vmbus_channel primary ptr from struct mapped_vmbus_resource as its not used. Populate virtual memory address "addr" in struct rte_mem_resource for secondary processes as netvsc will attempt to reference it thus causing a crash. It was initialized for primary processes but not for secondary. Cc: sta...@dpdk.org<mailto:sta...@dpdk.org> Signed-off-by: Jonathan Erb <jonathan....@banduracyber.com><mailto:jonathan....@banduracyber.com> --- v2: * Remove unnecessary check for NULL pointer before call to rte_free() per reviwer comment. drivers/bus/vmbus/private.h | 1 - drivers/bus/vmbus/vmbus_channel.c | 4 +--- drivers/bus/vmbus/vmbus_common_uio.c | 14 +++++++++----- 3 files changed, 10 insertions(+), 9 deletions(-) diff --git a/drivers/bus/vmbus/private.h b/drivers/bus/vmbus/private.h index 528d60a42f..746212bd5f 100644 --- a/drivers/bus/vmbus/private.h +++ b/drivers/bus/vmbus/private.h @@ -42,7 +42,6 @@ struct mapped_vmbus_resource { rte_uuid_t id; int nb_maps; - struct vmbus_channel *primary; struct vmbus_map maps[VMBUS_MAX_RESOURCE]; char path[PATH_MAX]; }; diff --git a/drivers/bus/vmbus/vmbus_channel.c b/drivers/bus/vmbus/vmbus_channel.c index f67f1c438a..119b9b367e 100644 --- a/drivers/bus/vmbus/vmbus_channel.c +++ b/drivers/bus/vmbus/vmbus_channel.c @@ -351,10 +351,8 @@ int rte_vmbus_chan_open(struct rte_vmbus_device *device, err = vmbus_chan_create(device, device->relid, 0, device->monitor_id, new_chan); - if (!err) { + if (!err) device->primary = *new_chan; - uio_res->primary = *new_chan; - } return err; } diff --git a/drivers/bus/vmbus/vmbus_common_uio.c b/drivers/bus/vmbus/vmbus_common_uio.c index 8582e32c1d..83c56b6fa2 100644 --- a/drivers/bus/vmbus/vmbus_common_uio.c +++ b/drivers/bus/vmbus/vmbus_common_uio.c @@ -69,8 +69,10 @@ vmbus_uio_map_secondary(struct rte_vmbus_device *dev) fd, offset, uio_res->maps[i].size, 0); - if (mapaddr == uio_res->maps[i].addr) + if (mapaddr == uio_res->maps[i].addr) { + dev->resource[i].addr = mapaddr; continue; /* successful map */ + } if (mapaddr == MAP_FAILED) VMBUS_LOG(ERR, @@ -88,9 +90,9 @@ vmbus_uio_map_secondary(struct rte_vmbus_device *dev) /* fd is not needed in secondary process, close it */ close(fd); - dev->primary = uio_res->primary; - if (!dev->primary) { - VMBUS_LOG(ERR, "missing primary channel"); + if (vmbus_chan_create(dev, dev->relid, 0, + dev->monitor_id, &dev->primary)) { + VMBUS_LOG(ERR, "cannot create primary channel"); return -1; } Looking at this closer, I don't think it will work for subchannels in the secondary process. The code after is: STAILQ_FOREACH(chan, &dev->primary->subchannel_list, next) { if (vmbus_uio_map_secondary_subchan(dev, chan) != 0) { VMBUS_LOG(ERR, "cannot map secondary subchan"); return -1; } } Because at this time, the primary channel is just created, "subchannel_list" should be NULL. The secondary process ends up running without subchannels. In the primary process, "subchannel_list" are populated when it calls hn_dev_configure(), so it is good. Sorry I didn't spot this earlier. The best way to fix this is to use rte memory functions for allocating vmbus device. In this way they can be properly mapped to the secondary process. But rte memory functions are not available at the time vmbus device is probed. I haven't been able to find a good way to fix this. Will keep looking. Thanks, Long @@ -211,8 +213,10 @@ vmbus_uio_unmap_resource(struct rte_vmbus_device *dev) return; /* secondary processes - just free maps */ - if (rte_eal_process_type() != RTE_PROC_PRIMARY) + if (rte_eal_process_type() != RTE_PROC_PRIMARY) { + rte_free(dev->primary); return vmbus_uio_unmap(uio_res); + } TAILQ_REMOVE(uio_res_list, uio_res, next); -- 2.17.1
